ML23352A392
| ML23352A392 | |
| Person / Time | |
|---|---|
| Issue date: | 11/15/2023 |
| From: | Advisory Committee on Reactor Safeguards |
| To: | |
| References | |
| NRC-2626 | |
| Download: ML23352A392 (1) | |
Text
Official Transcript of Proceedings NUCLEAR REGULATORY COMMISSION
Title:
Advisory Committee on Reactor Safeguards Joint Human Factors Reliability and PRA, and Digital I&C Subcommittees Docket Number: (n/a)
Location: teleconference Date: Wednesday, November 15, 2023 Work Order No.: NRC-2626 Pages 1-302 NEAL R. GROSS AND CO., INC.
Court Reporters and Transcribers 1716 14th Street, N.W.
Washington, D.C. 20009 (202) 234-4433
1 1
2 3
4 DISCLAIMER 5
6 7 UNITED STATES NUCLEAR REGULATORY COMMISSIONS 8 ADVISORY COMMITTEE ON REACTOR SAFEGUARDS 9
10 11 The contents of this transcript of the 12 proceeding of the United States Nuclear Regulatory 13 Commission Advisory Committee on Reactor Safeguards, 14 as reported herein, is a record of the discussions 15 recorded at the meeting.
16 17 This transcript has not been reviewed, 18 corrected, and edited, and it may contain 19 inaccuracies.
20 21 22 23 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1323 RHODE ISLAND AVE., N.W.
(202) 234-4433 WASHINGTON, D.C. 20005-3701 www.nealrgross.com
1 1 UNITED STATES OF AMERICA 2 NUCLEAR REGULATORY COMMISSION 3 + + + + +
4 ADVISORY COMMITTEE ON REACTOR SAFEGUARDS 5 (ACRS) 6 + + + + +
7 JOINT HUMAN FACTORS RELIABILITY & PRA, 8 AND DIGITAL I&C SUBCOMMITTEE MEETING 9 + + + + +
10 WEDNESDAY 11 NOVEMBER 15, 2023 12 + + + + +
13 The Joint Subcommittee met via 14 Teleconference, at 8:30 a.m. EST, Vicki Bier, Chair, 15 presiding.
16 17 COMMITTEE MEMBERS:
18 VICKI M. BIER, Chair 19 RONALD G. BALLINGER, Member 20 CHARLES H. BROWN, JR., Member 21 GREGORY H. HALNON, Member 22 JOSE A. MARCH-LEUBA, Member 23 ROBERT E. MARTIN, Member 24 WALTER L. KIRCHNER, Member 25 DAVID A. PETTI, Member NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
2 1 JOY L. REMPE, Member 2 THOMAS P. ROBERTS, Member 3 MATTHEW W. SUNSERI, Member 4 VESNA DIMITRIJEVIC, Member 5
6 ACRS CONSULTANTS:
7 MYRON HECHT 8 STEPHEN SCHULTZ 9
10 DESIGNATED FEDERAL OFFICIAL:
11 CHRISTINA ANTONESCU 12 13 14 15 16 17 18 19 20 21 22 23 24 25 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
3 1 TABLE OF CONTENTS 2 Opening Remarks by Chairman . . . . . . . . . . . 4 3 Introductory Remarks . . . . . . . . . . . . . . 8 4 Data Science and AI, Regulatory 5 Applications Public Workshop, 6 AI Characteristics for Regulatory 7 Consideration, Workshop Findings . . . . . 14 8 AI Project Plan . . . . . . . . . . . . . . . . . 15 9 Future Focused Research (FFR) Project; 10 Using Machine Learning to Inform 11 Inspection Planning . . . . . . . . . . . 122 12 Overview of DOE's Light Water Reactor 13 Sustainability Program . . . . . . . . . 156 14 Plant Modernization Activities, LWRS AI 15 Industry Research and Accomplishments 16 Overview, Fire Watch Report . . . . . . . 190 17 The Good, the Bad, and the Ugly of AI in 18 Process Control . . . . . . . . . . . . . 242 19 Public Comments . . . . . . . . . . . . . . . . 294 20 Closing Remarks by Chairman . . . . . . . . . . 301 21 22 23 24 25 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
4 1 P-R-O-C-E-E-D-I-N-G-S 2 (8:30 a.m.)
3 CHAIR BIER: So the meeting will now come 4 to order. I will start with my introductory remarks 5 and then you guys will go in a couple of minutes.
6 This is the meeting of the Joint Human 7 Factors Reliability and PRA, and the Digital I&C 8 Subcommittees. I'm Vicki Bier. I'm going to be 9 chairing this subcommittee meeting.
10 ACRS members in attendance, we have 11 Charles Brown. Matt seems to be not here this 12 morning. He may be coming in later. Jose, you're on 13 line?
14 MEMBER MARCH-LEUBA: Yes.
15 CHAIR BIER: And Vesna on line?
16 MEMBER DIMITRIJEVIC: Good morning, 17 everybody.
18 CHAIR BIER: Okay, we have Joy Rempe, Ron 19 Ballinger, Dave Petti, Walt Kirchner, Greg Halnon, Tom 20 Roberts, Robert Martin. And Steve Schultz, our 21 consultant, is here. And do we have Myron Hecht? Is 22 he here or online? He may also be joining later as a 23 consultant.
24 MEMBER BROWN: Yes, Vicki, Matt said he 25 would be coming in virtually at probably 10 o'clock or NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
5 1 so.
2 CHAIR BIER: Okay.
3 MEMBER BROWN: He had something. He had 4 to go back to North Carolina.
5 CHAIR BIER: Oh, wow. Okay. Well, thank 6 you for letting me know.
7 Christina Antonescu of the ACRS staff is 8 the Designated Federal Official, or DFO, for this 9 meeting.
10 Christina, can you confirm that we have 11 the court reporter on line?
12 Can the court reporter speak up?
13 MS. ANTONESCU: Can the court reporter 14 speak up, please?
15 (Off-microphone comment.)
16 CHAIR BIER: Okay, great. So there are 17 going to be two separate, but related, purposes for 18 today's meeting. First, staff and contractors are 19 going to provide information briefings on how they are 20 implementing the NRC's artificial intelligence, or AI, 21 strategic plan for fiscal years 2023 to 2027, so along 22 with their collaborators.
23 In addition, we have a speaker later this 24 afternoon, Dr. Missy Cummings, who will also present 25 on pluses and minuses of artificial intelligence. She NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
6 1 was originally hoping to come in person, but is going 2 to be online due to schedule conflicts.
3 I wanted to clarify for the staff and for 4 anyone else listening that Dr. Cummings is not going 5 to be in any way commenting on the staff 6 presentations. She was not asked to review them and 7 her opinions are her opinions. They're not, you know, 8 to be interpreted as a comment positive or negative 9 about anything the staff is doing.
10 Mainly, once we have this briefing on the 11 agenda, I wanted to take the opportunity to have just 12 an educational briefing for the committee members, so 13 that the members are all starting with a basic 14 understanding of some of the key issues, especially 15 that will be coming before the committee probably in 16 years to come, rather than at this moment.
17 For background, the ACRS was established 18 by statute and is governed by the Federal Advisory 19 Committee Act, FACA. This means that the committee 20 can only speak to its published letter reports. We 21 hold meetings to gather information to support our 22 deliberations. Interested parties who wish to provide 23 comments can contact our office requesting time to do 24 so. We also set aside about 15 minutes usually at the 25 end of every meeting for comments from members of the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
7 1 public either in person or listening on line. We also 2 welcome written comments.
3 The meeting agenda for today was published 4 on the NRC's public meeting notice website, as well as 5 the ACRS meeting website. The agenda and the ACRS 6 website have instructions about how the public can 7 participate. I don't believe we have any formal 8 requests for making a statement to the subcommittee 9 from members of the public yet, but people are always 10 welcome to chime in.
11 Today is going to be conducted as a hybrid 12 meeting, both in person and online. A transcript of 13 the meeting is being kept and will be made available 14 on our website. Therefore, we request that 15 participants in the meeting should identify themselves 16 before they speak and speak with sufficient clarity 17 and volume so that they can be readily heard. And as 18 the staff knows, I'm sure, please allow time for 19 member questions. Members always have a lot of 20 questions and comments. And it might also help to 21 indicate which slide number you're on for people who 22 are following along on line.
23 We have an MS Teams phone line for audio 24 established for the public who wishes to listen to the 25 meeting. Because of the nature of the online NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
8 1 meetings, we will take a short break after each 2 presentation, if needed, to allow for screen sharing 3 and of course, we will also take the usual breaks 4 during the meeting as needed.
5 A reminder, anybody who is online on 6 Teams, please do not use the meeting chat features to 7 conduct any sidebar technical conversations. Those 8 should be oral so that they're captured in the 9 transcript. And if you have any questions, you can 10 contact the DFO, Christina, about issues that you 11 would like to have raised or if you're having 12 connection difficulties, et cetera.
13 So we are now ready to proceed with the 14 meeting. Matt, it looks like you already have the 15 slides shared.
16 The opening remarks will be from Mr. Vic 17 Hall, who is Deputy Director of the Division of 18 Systems Analysis in the Office of Nuclear Reactor 19 Regulatory Research and after that, we'll be ready for 20 the rest of the presentations.
21 So feel free to go ahead.
22 MR. HALL: Thank you, Vicki. Good 23 morning, everyone. Vicki, I appreciate you and I 24 share a namesake. I have on many occasions been 25 called Vicki, usually in the school yard. Good to NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
9 1 have a fellow Vic, Vicki in the room.
2 So I'll introduce myself to the room. My 3 name is Vic Hall. I'm the Deputy Director in the 4 Division of Systems Analysis. I joined the Office of 5 Nuclear Regulatory Research earlier this year. If 6 you're looking at your calendars, I got here right 7 after this amazing team published the strategic plan.
8 So I'm going to take full credit for the wonderful 9 work that they did and today, I'll mention I'm 10 extremely proud of the work they've done and part of 11 my job is representing them and being able to 12 introduce them today. So it's my honor to be able to 13 introduce Matt and Anthony and Trey at the table who 14 will be doing all the heavy lifting and under the 15 spotlight.
16 I do want to express my gratitude to the 17 subcommittee today. Vicki mentioned this is an 18 information meeting for the members here, but I kind 19 of disagree. I think it's an information meeting for 20 everyone here. AI is moving so fast. It's an 21 technology that has got such a head of steam. I'm 22 staying up late which I shouldn't do and I'm watching 23 Fox shows on TV and whether it's Jimmy Kimmel or Jimmy 24 Fallon or whichever show, they're talking about AI.
25 When the President puts out an Executive Order a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
10 1 couple of weeks ago, they're talking about on the 2 news. It just amps up the game. It puts the 3 spotlight on the technology that is either going to 4 change the world or is going to scare the heck out of 5 us. So the most important thing that we can do is 6 we're not going to regulate AI. We can be ready for 7 it. We can prepare where it's coming.
8 So when we have a meeting like this today 9 that you put together, it's an opportunity for us to 10 share the amazing work that we've done, but really to 11 listen to what you have to say, take that into 12 account, because we have to collaborate. We have to 13 take all the opinions into account because technology 14 is moving so fast.
15 I am really looking forward to today's 16 meeting and again, when you walk the halls of the NRC 17 and you're telling some folks what are you doing 18 today? I'm speaking in front of ACRS. The reaction 19 ooh, good luck. Expect a lot of questions and 20 discussion. My answer to that is good. We're 21 welcoming that today. We really look forward to your 22 probing, your questioning attitude, and how we can 23 improve, how we can be ready for something that's 24 moving so fast.
25 So with that, I do want to put the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
11 1 spotlight back on really the stars of the show today.
2 We've got Matt Dennis, Anthony Valiaveedu, and Trey 3 Hathaway who are experts in the field. And our 4 office, the Office of Nuclear Regulatory Research, is 5 really a hub of world-class expertise. And when it 6 comes to AI and AI in government, I don't think you'll 7 find a finer group of gentlemen on the topic. They 8 not only understand the policy, they not only put 9 together the strategy, but they write the language.
10 They write in Python. They do things that make me as 11 an electrical engineer blush and I am truly honored to 12 be able to work with them to represent them and to be 13 able to share their work with you today.
14 Matt will give you a summary of our 15 workshop which we held a couple of months ago. This 16 was our fourth workshop on AI and data science. The 17 first, I'll call it in the ChatGPT era when really I 18 think the world has been awakened to what's coming 19 with AI and it was our best -- it was widely attended.
20 We have over 350 attendees from 12 countries. We had 21 wonderful speakers from the national labs and 22 universities just like we have today, so I'm very much 23 looking forward to that same type of learning and 24 interaction that we're going to have today.
25 Next, Anthony Valiaveedu will give you the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
12 1 rundown on our strategic plan and the project plan 2 which talks about all the different steps that we've 3 laid out. They're going to get us to I'll call it 4 success between now and the next five years and I 5 guarantee you that plan will change. There's no way 6 that plan can't change with the way the speed is 7 changing. So we've done our best take and I think 8 it's a pretty darn good take at what actions we need 9 to do to be ready for what's coming from the industry.
10 I do want to mention again the fact that 11 things are moving fast, well before this meeting was 12 scheduled, well, after this meeting was scheduled, I'm 13 sorry, the chair put out a tasking memo to the NRC 14 staff. The title of that memo is Advancing the Uses 15 of Artificial Intelligence. And in that memo he very 16 much speaks about the need to be responsible in that 17 use of AI. And that clearly is again, putting the 18 focus on what the staff is willing to do to keep 19 prepared for this technology that's coming very fast.
20 And obviously, Nuclear Regulatory Research certainly 21 has a role because have such a dense group of 22 expertise in that. So we're working very much with 23 our partners in the Office of Chief Information 24 Officer and really every office in this agency because 25 everybody will have a role in figuring out how we can NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
13 1 best responsibly use the technology to get our job 2 done in our mission of safety.
3 So with that, I did want to mention a 4 couple of folks in the room, Trey Hathaway as well.
5 I've got one of our experts here that will be able to 6 answer questions. And if you haven't met Luis 7 Betancourt, Luis Betancourt is the branch chief in our 8 division who really is the motor behind all this and 9 makes it happen. And we also have Paul Krohn, I think 10 is on the line, who is my co-chair on the AI Steering 11 Committee. So again, it's been a team effort across 12 the agency and it will continue to be so in the years 13 coming.
14 And again, I just wanted to close with 15 repeating my gratitude again. It's these types of 16 meetings that will make us better. So thank you for 17 having us. I hope you enjoy the presentations that 18 we've prepared. I am extremely proud of the work and 19 what we've accomplished in the last year and very much 20 looking forward to actually what the next five years 21 bring -- what the future bring for us as an agency.
22 So with that, Matt, let me hand it over to 23 you and thank everyone again for their attention and 24 for bringing us today.
25 CHAIR BIER: Okay, before we get going, I NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
14 1 just wanted to make one other announcement which is 2 the committee is on an very strict schedule today 3 because unlike the staff, the committee has a 4 commitment at lunch hour and with an outside speaker 5 later in the afternoon. We're going to have to kind 6 of try and keep it on schedule. Happy to get going, 7 so go ahead.
8 MR. DENNIS: Okay, I think I hit the 9 button and the microphone is green so I'm good to go.
10 Good morning, everyone. Again, my name is Matt Dennis 11 and Trey Hathaway here. We're from the Office of 12 Research and we'll be talking -- our first 13 presentation this morning will be on the summary and 14 finding of the AI public workshop which we had back in 15 September. And I appreciate the push to get -- to 16 have a lunch break. I have to go get my flu shot, so 17 I am ecstatic to be on time so I can go get my flu 18 shot.
19 Again, Matt Dennis, Trey Hathaway. We've 20 already introduced Paul. Paul, Vic, and Luis are 21 sitting over here at the side table and are available 22 to answer any questions that we have that's related to 23 the strategy or our progress.
24 Trey and I are going to talk about --
25 we're going to talk about what the landscape is as we NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
15 1 see it right now, so just a brief. Though as a 2 reminder, we came a year ago, exactly almost a year 3 ago tomorrow, and presented on our draft AI strategy 4 last year and following that, we did a public comment 5 or we had a public comment period. We resolved the 6 comments and moved forward with publishing the 7 strategy, as Vic mentioned, in May of this year. So 8 I'll talk just a little bit about where we are in that 9 landscape, as far as it pertains to the workshop.
10 Anthony will be discussing the project plan and a lot 11 of what has come out of implementing the strategy in 12 the last year since we talked to you. We've made a lot 13 of good progress.
14 So I'll talk about the workshop overview.
15 I'm sure a number of you were able to attend the 16 workshop. So I will not be going into the nitty-gritty 17 of the entire workshop, but instead, I'll be talking 18 about our observations from our perspective about what 19 was said and discussed at the workshop, so I'll talk 20 about the workshop and all session summaries. I'm 21 lucky to have some of the chairs who chaired those 22 panel sessions here today participating in the meeting 23 and so it is not just Trey and I who are here to 24 discuss the workshop, but we have a number of staff 25 across the offices that have been participating in an NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
16 1 AI working group that we have organized, specifically 2 to address or discuss AI attributes for regulatory 3 consideration and that working group is also tasked 4 with planning the workshop. So it's not just the 5 Office of Research. It's been a very collaborative 6 effort with a lot of the offices across the NRC and 7 some of those staff are participating today in this 8 meeting, so they're also available to field questions 9 should they come up.
10 So I'll talk about -- and then finally, 11 the high-level observations and then where do we go 12 from here following this particular workshop.
13 So this is the slide. It's very similar to 14 the one we talked about last year except with some 15 updates. So we recognize, you'll notice on the left, 16 the box that says external. That's highlighted in 17 blue with intent behind it because the focus of the 18 strategy, the NRC's AI strategic plan is externally 19 facing. So we recognize an industry wants to use 20 artificial intelligence and in order to do that, we 21 took a proactive approach two years ago, around 2021, 22 to develop the strategy in order to prepare the staff 23 to review and evaluate the uses of AI that may be an 24 NRC regulated activity. So we developed AI strategic 25 plan for that purpose and that was published, as was NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
17 1 mentioned in May 2023.
2 So some other things that are internal, 3 Dave mentioned the chair's memorandum on advancing the 4 use of AI at the NRC which just came out last month 5 and is being -- we're in the process of standing up 6 the response to that and getting our ducks in a row 7 for that purpose. So that's internal.
8 Some other internal things are -- you may 9 have seen in the news. As Vic mentioned, it's been 10 front and center. The Biden administration put out an 11 Executive Order, again last month, on federal actions 12 for advancing use of AI in government. So there is a 13 push at the executive level for not only agencies to 14 get a handle on what AI means, but also to prepare the 15 agencies for adoption of AI within their portfolios.
16 So not only are we in the position of regulating our 17 industry's use of AI, potential use of AI, but we are 18 also -- we have to prepare ourselves to use the 19 technology as well. And so there will be a 20 forthcoming OMB memo on this directing us how to 21 consider certain aspects of AI implementation at the 22 agency.
23 We have also been quite involved in a 24 number of outside activities that benefit our internal 25 preparedness with implementing the strategy as well as NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
18 1 our internal preparedness for the Executive Order.
2 There is no shortage of AI conferences, meetings, 3 symposiums that come up that you can attend and those 4 have been -- we have participated in a number of 5 those. Just recently, there was a PSAM meeting, a 6 topical meeting specifically on AI that we attended 7 and there were at least three and a half full days --
8 three full days of presentations globally. So 9 clearly, this is an interest in the nuclear industry 10 and we have made a concerted effort to keep up to date 11 on what is going on.
12 We have also been participated in a number 13 of activities outside the agency. Trey is involved 14 with a standards group which we'll be talking about 15 later. And not only just the nuclear field, we also 16 participate in a number of conferences, workshops, and 17 symposiums that are in the Department of Defense area.
18 So we are looking to other agencies, DOD, DOT, FDA, 19 other areas where this is also being used so that we 20 are best prepared because this is a whole of 21 government action, not just us.
22 So and then on the right, the box that 23 talks about evidence building priority questions, we 24 have from the Evidence Building Act of 2018 and a 25 couple of priority questions that were added to one, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
19 1 wrangle our data for best use for AI and also look for 2 areas where we could use the work for the agency. I 3 mentioned the chair's memo and then you will be 4 hearing another presentation later this morning on a 5 future focused research program specifically on 6 looking at AI. The future focused research program as 7 an incubator and technology development area where the 8 staff can look at AI usage has been incredibly useful 9 in the Office of Research and is one of the programs 10 that we've called out in the AI strategic plan and the 11 way to prepare our staff to understand this technology 12 and it has been very beneficial. So you will be 13 hearing a presentation on that topic as well.
14 MEMBER KIRCHNER: Walt Kirchner here, I'm 15 not sure where to start this. I warned Vicki I was 16 going to ask this. Can you define what you mean by 17 AI? And if there's -- you know, succinctly, because 18 if one is going to regulate, quote unquote, whatever 19 that means at this early juncture, then one has to 20 have an understanding of what it is you're going to 21 regulate in terms of nuclear applications in the 22 industry. So could you share that with us?
23 MR. DENNIS: I will mention that there is 24 a -- the entire -- the very first page of the 25 introduction to the AI Strategic Plan has two NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
20 1 paragraphs that -- I won't read them verbatim. I'll 2 give you the high points. But it is a broad 3 definition and in the strategic plan there was a lot 4 of effort put into what does AI mean. Unfortunately, 5 that definition is quite broad. And so the umbrella 6 of AI includes natural language processing, machine 7 learning, deep learning, all the buzz words that you 8 hear.
9 But when you boil it down to just a few 10 key words, AI has the ability to emulate human-like 11 perception, cognition, planning, learning, and 12 communication, or physical action. And so our 13 definition in that introductory paragraph of the 14 strategic plan, the two paragraphs, is very much based 15 on the National Defense Authorization Act of 2021 as 16 Congress defined AI which, as I mentioned, for better 17 or worse is a very large, broad definition. So to 18 interpret that, we have gotten a little more specific 19 for our purposes to clarify the difference between 20 automation and AI-enabled autonomy, and it really is 21 the cognition and decision-making portion of AI that 22 is crucial for looking at it.
23 CHAIR BIER: I'm going to chime in a 24 little bit and this is really in a way Walt's and my 25 conservation, but one of the things that I get NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
21 1 concerned about in the definition is what today is 2 considered AI may ten years from now be considered 3 computer programming. So our regulations have to talk 4 about, I think, what functions it's serving in the 5 plan rather than what maybe the technology goes by.
6 But that's just my personal opinion.
7 MEMBER MARCH-LEUBA: Yes, this is Jose.
8 I was going to say something similar. I don't think 9 we regulate Fortran. We use Fortran to write code, 10 safety codes that have been used to verify the 11 regulations are satisfied. In a global sense, I see 12 AI as another type of Fortran. I don't think we're 13 going to write regulations that apply to a concept, an 14 abstract concept called AI. I mean we don't have a 15 regulation for Fortran. Am I thinking wrong?
16 MR. DENNIS: On that note, what was just 17 mentioned, we have also grappled with the same issue 18 and in looking back at that definition that we talk 19 about in the introduction of the strategic plan, this 20 dichotomy of software versus AI and where we are 21 currently is called out, so we do recognize it, that 22 there is a difference between software and AI. And 23 there is a sentence that says an overarching goal of 24 AI is providing solutions that mimic human-based 25 solutions and predictions for problems. So some of our NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
22 1 discussion has been focused on the fact that Fortran 2 software programming is very rule based whereas AI is 3 based on large sets of data and then can infer or 4 create its own algorithm to then make decisions that 5 mimic human behavior.
6 So to the point of the future-proofing of 7 definition, that is also part of why it is so squishy 8 right now and there is -- we just went to a meeting 9 where this same topic of what do you mean when you say 10 AI came up? And one of the presenters was discussing 11 that said no one has a unified definition. Everyone 12 has a different interpretation. So right now, the 13 strategic plan does have a broad definition with the 14 caveat that says the U.S. NRC in an area where it has 15 not been previously reviewed or evaluated. So we're 16 not going to go back to something that is Fortran code 17 and now call it AI. We're going to be looking at 18 going forward, specifically examples where we feel 19 that it fits under this definition of AI that we have 20 that is so broad.
21 MEMBER HALNON: This is Greg. I think in 22 our last subcommittee meeting we talked about this as 23 well and we entreated you guys that that should be a 24 priority because if you're going to put a regulatory 25 framework around something, you need to know what that NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
23 1 something is. And otherwise, we're either over 2 regulating, under regulating. It's a real danger, so 3 again, I'll say the same thing I did at previous 4 meetings. It's very important, at least to me, to get 5 a succinct definition, boundary, whatever you want to 6 call it around what you are going to need to regulate.
7 When you get into somebody wants to made 8 a modification of the plan, apply AI, and they say how 9 do I do a 50.59 on it? They're going to need to be 10 able to have a series of workshops for ten years to 11 figure it out. No, we don't have that time. As Vic 12 said, it's moving so quickly. So anyway, that's kind 13 of a recurring comment I think that we're going to be 14 making as well.
15 MEMBER BALLINGER: This is Ron Ballinger.
16 I'd like to second that. I mean this is a case where 17 we run the risk of getting into what I call the 18 subjectivity trap. And that at some point, somebody 19 has to decide where the line is and if that line is 20 fuzzy or depending on the person that's using it, when 21 I check off on Microsoft Word, the autofill thing, 22 guess what, it's telling me what I should say. So 23 I'll just reinforce what Greg was saying.
24 MEMBER BROWN: And I'll follow up if Ron's 25 finished. I'm probably the most resident skeptic --
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
24 1 I'm Charlie Brown, most resident skeptic on the 2 committee. I will echo Greg's comment in that you 3 really need to know what you're going to regulate 4 before you can know what you're going to do with AI.
5 Your comments about programming is rule 6 based whereas AI is quote evaluation of data sets in 7 developing an algorithm that then goes and determines 8 what direction you may want to go. That's then 9 subject to the bias of the algorithm mapper who, 10 somebody has got to say how algorithms are going to 11 get developed. And there are biases all over the 12 place in terms of what subject you're using, number 13 one, whether they're social or technical. And that 14 gets into a world of uncertainty.
15 I'm just going to make this comment early 16 so everybody can be very aggravated throughout this.
17 Greg was right, why do you want to try to regulate or 18 develop a rule or how will we regulate when you really 19 don't know how to use all this stuff in the first 20 place? My response, my thought process, is somewhat 21 different in that we're primarily based -- we're 22 responsible for the safety of the plants. Our reactor 23 trip safeguard systems, major plant control systems 24 whether they're called safety or safety-related or not 25 related to safety, whatever definition you want to NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
25 1 apply, they're plant-controlled systems.
2 You certainly don't want all AI is it's 3 invasive to the point where you may at a process if 4 you want to go develop a trip and it says oh, no, 5 maybe you don't want to because I'm looking at this 6 other data. And now you've got software, because it's 7 all embedded, that is variable, and does not have 8 really -- you really don't want something else other 9 than people deciding what's safe and not safe. I mean 10 if I were you all, I was the boss and I'm not, you're 11 lucky from that standpoint, I would put the brakes on 12 it. I would literally if I was going to try to 13 regulate this world, I would not try to do it -- I 14 read through your program. Obviously, we have 15 questions, but there are certain things you want to 16 maintain. That's the safety posture.
17 The way to actually go about this and find 18 out what are the benefits, how can it be utilized is 19 to put a roadblock up and say, hey, look, you will not 20 use or attempt to use or propose the use of AI for any 21 reactor safety systems, any reactor safety related 22 systems, or other plant-controlled systems that have 23 to start, stop, various components, move rods, 24 whatever you want to call it. Now let the vendors go 25 figure out outside that box, how AI can be adapted in NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
26 1 an overall plant configuration on ways where it may 2 benefit on the non-operational aspects of the 3 equipment we have in there that serve, brackets the 4 problem that the NRC is going to have to deal with.
5 This is, to me, this is just the latest hot button fad 6 that everybody is thinking is the greatest thing since 7 sliced bread and trying to integrate it and put it 8 into a regulatory rule is just not possible at all 9 based on the way it's done.
10 The biases are terrible. All you have to 11 do is look at the learning trying to make autonomous 12 cars work properly. That's fundamentally a 13 combination of rule-based and/or some level of AI 14 that people are trying to introduce, a lot of wrecks 15 because you can't define all the things that it may 16 see, all the sensors may see. So that's my opening.
17 CHAIR BIER: Yes, you've heard a lot from 18 us and we haven't heard much from you.
19 MEMBER BROWN: Well, we had to -- I had to 20 give a flavor. I'm -- I was not (audio interference) 21 I'll pass.
22 CHAIR BIER: I guess two comments: One is 23 in order to ban AI you first have to know what AI is.
24 You can't --
25 MEMBER BROWN: I didn't say ban. I did NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
27 1 not say ban, Vicki. I said don't put it in places 2 where --
3 CHAIR BIER: Yes, but, still, you have to 4 know what it is you don't want in those places.
5 MEMBER BROWN: The stuff we have installed 6 right now works quite well.
7 CHAIR BIER: The other --
8 MEMBER BROWN: It's not a matter of 9 banning anything. It's a matter of putting in basic 10 software which then stops the software from performing 11 in a repeatable and predictable manner. It's being 12 changed constantly. You don't know what you have.
13 CHAIR BIER: The other comment that I 14 would add is I --
15 MEMBER BROWN: You can see we have a lot 16 of different opinions.
17 CHAIR BIER: -- assume that the NRC is 18 also looking not only at regulating industries of AI, 19 but also at advancing what the agency itself may want 20 to use AI for. So, anyway, with that --
21 MEMBER MARTIN: Well, thank you, Vicki.
22 Vicki, I need my shot, too.
23 CHAIR BIER: Oh, okay.
24 (Laughter.)
25 MEMBER MARTIN: Bob Martin. A few years NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
28 1 ago I read Max Tegmark's Life 3.0. I hope that's like 2 required reading for this crowd. He's an MIT 3 professor or Harvard, or something like that in 4 Boston. And I thought what struck me was the reason 5 why we're hearing about this again is because it comes 6 and goes, right? It's obviously the interconnectivity 7 of the world and the 'net and it comes down to data, 8 right? The amount of data that we have and the 9 algorithms that we have can now process this data in 10 a way that can fool us, right?
11 For us, we live in a space -- us, the ACRS 12 and the NRC -- in a space of low-frequency, high-13 consequence events, maybe something broader than that, 14 with the emphasis on low-frequency. And low-frequency 15 is referring to our hazards, bad things that happen 16 that you don't have a lot of data for. So invariably 17 we have a data gap. And the one I'm going to be 18 listening for is -- and I have a couple ideas, but I'm 19 not going to beat them just yet, but how you might 20 think that there is data or data could be created to 21 serve really safety issues that are relevant to this 22 (audio interference). You don't have to answer that 23 now. It's a kind of a comment. But that's a 24 sensitivity. We look at a certain -- a sliver of what 25 AI can do here, but certainly, Vicki, you made the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
29 1 point there are process improvements and other things 2 that I absolutely agree are relevant. But it wouldn't 3 necessarily be for us.
4 MEMBER ROBERTS: And my two cent's worth.
5 This is Tom Roberts, just following up on what Charlie 6 said. It seems to me that both the agency and 7 industry have struggled for probably decades to figure 8 out how to implement software into plant control 9 systems. And there's been a whole infrastructure of 10 diversity and challenges to how much diversity you 11 need, but it doesn't seem to me like a new issue. The 12 whole idea of having deterministic software imbedded 13 in the system has been a concern because you can never 14 prove that you've gone through all those possible 15 deterministic combinations and have 100 percent 16 certainty that the software is going to do what it's 17 required to do, nor that the requirements are 18 complete.
19 But I'm just wondering if maybe we could 20 talk through the morning session just how different 21 that is for AI, because it just seems like an 22 extension of the same problem.
23 MEMBER BROWN: Thank you, Tom.
24 MEMBER MARCH-LEUBA: Yes, this is Jose.
25 I wanted to bring in another concept. We think -- we, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
30 1 ACRS members -- seem to be focused on reactor 2 protection system applications. And that's --
3 probably that's our bread and butter. But AI is going 4 to be applied on analyzing data from nondestructive 5 assay pipes, of measurements, and you don't have a 6 chance of a person to looking through all of them. So 7 you develop an AI system to look for flaws on piping.
8 I don't think anybody's proposing to put 9 an AI system on a protection system. I mean, they 10 want to do it for maintenance or for data processing.
11 I mean, the pipe testing is the clear application that 12 is going to come first.
13 CHAIR BIER: And I think we're going to 14 hear some of those, yes.
15 MEMBER BROWN: That's why I said -- this 16 is Charlie Brown. That's why I suggested separating 17 what we know we really have a hard time dealing with, 18 echoing Tom's comment, because we have struggled with 19 how do you apply the software systems and make sure 20 they're going to work when they're supposed to work, 21 and then let industry develop all these things like 22 maintenance data, data that you get out of 23 experimental test facilities, where can AI help you 24 evaluate the large quantity of data that you get that 25 may help you define the physical boundaries you have NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
31 1 to deal with in the plants.
2 There's a lot of stuff outside of what I 3 would call the basic operation and shutdown of the 4 plant that need be -- or starting it up, et cetera.
5 And you will learn an awful lot from that. And you 6 really don't want to lose focus because you're focused 7 -- you're driven to focus on what I call plant 8 operation-type scenarios as opposed to what I'd call 9 stuff that's outside if that. I'm just reflecting 10 Tom's comment and Jose's, and maybe Robert's. I'm not 11 quite sure. I was struggling a little bit --
12 (Simultaneous speaking.)
13 MEMBER BROWN: Yes. Well, I'm not 14 ambiguous, so I'll --
15 CHAIR BIER: Let's try and move ahead.
16 Yes.
17 MR. DENNIS: I will say I'm heartened that 18 we -- there's a lot of synergy here. Everything you 19 said we have brought up as a topic in our working 20 group degree, so --
21 MEMBER BROWN: Well, I'll make one other 22 comment: Over the last four years I have read 23 numerous -- the IEEE is a body that just loves all the 24 new stuff. If you look at the -- at least the 25 Spectrum and a couple of the other journals, you will NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
32 1 find tremendous amount of articles which are very 2 skeptical in the application of AI in terms of the 3 biases and other type issues. How do you know you're 4 getting stuff that's telling you the right answer with 5 the algorithms and stuff?
6 And some -- it's just I'm surprised when 7 you see that much in -- an organization that loves 8 electrical software computers has now got the skeptics 9 coming in through their publications showing some of 10 the concerns that we've echoed right here relative to 11 the difficulties. So that's kind of an outlier, but 12 it is an organization that has a lot of people 13 involved and loves this kind of stuff. And they're 14 even skeptical. And they're publishing. That's the 15 important part.
16 MR. DENNIS: I hope to, I don't know, 17 answer, bring up -- there's many of these things that 18 people just brought up. I'm talk a little bit about 19 some of those.
20 The Data Science and AI Workshops, just to 21 skim over this, is that we had four -- we had three in 22 2021, we had the fourth one in September. The goal 23 was to answer some of these -- not answer these 24 questions, but at least get some insight on some of 25 these questions that have just been discussed.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
33 1 In 2021 we recognized these were some 2 -- three observations. Industry did have interest in 3 regulatory guidance on this topic. There is an issue 4 with data. The topic about limited data was brought 5 up in the nuclear domain. And so aggregating and 6 using data for these data-hungry applications was an 7 area that was brought up. There has been some 8 progress made on that, I think, at the national labs.
9 And as of 2021; so taking it back to 2021, 10 we heard that probably maybe now, 2023, there would be 11 some deployment of an AI ML application. That has 12 borne out to be, I think, true. And then regulated 13 applications, maybe in three to five years, so 2026.
14 So that was the basis for our timeline and our 15 strategic plan. And we've heard about two or three 16 application areas where that may actually pan out.
17 So I guess to the point of narrowing down 18 the definition, there are specific use case areas for 19 AI ML. One example was brought up in nondestructive 20 evaluation. That's one area that was discussed or 21 presented on last year at the ACRS meeting. And 22 that's made some significant progress towards actual 23 implementation.
24 MEMBER KIRCHNER: Can you distinguish in 25 your own mind -- because you -- all of you there up in NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
34 1 front of us are focused in this area -- the difference 2 -- 10 years ago or 20 years ago this was big data and 3 you had people like GE using it to improve the 4 preventive maintenance on their jet engines, you had 5 Steve up at NYU using it to predict when you could get 6 a taxi in Manhattan, and so on.
7 So I personally never thought of that as AI.
8 So can you make -- what's the distinction 9 between just harvesting data with algorithms and AI?
10 Because to me AI was always the cognitive function 11 that you mentioned as part of your definition. And 12 like fusion, that was 50 years ago at the MIT media 13 lab. We're still probably 50 years out from that kind 14 of definition of AI.
15 So how do you -- is AI just the umbrella 16 that you want to use and it's the current jargon? As 17 Dick was saying, even the talk show hosts are using 18 the jargon. But is this really just big data and 19 better computers and smarter algorithms or is it 20 really cognitive AI?
21 MR. DENNIS: So two differentiating 22 aspects: You mentioned 10 years ago this was called 23 big data. Twenty years ago it was called expert 24 systems and recommendation (audio interference). So 25 the point is well-taken that this is an evolving area NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
35 1 where 20 years ago something that we would call AI is 2 not something we would necessarily think is AI today.
3 And from our perspective machine learning 4 and AI -- we're focusing on safety-critical 5 application use case areas. And so this isn't -- and 6 so the question that has been brought up is -- the 7 industry is using it in process improvement areas 8 outside of safety-critical applications to make good 9 business decisions, to infer things that -- to assist 10 with things that are not regulated application areas.
11 And one of those has been presented. I'll get to it 12 later on the Corrective Action Program analyzer and 13 the maintenance rule functional failure analyzer. At 14 the workshop it was presented on. So those are areas 15 where it's being used in non-safety-critical 16 applications.
17 From our perspective part of the issue is 18 whether or not it is autonomous or making a decision.
19 There is a distinction between using AI ML for design 20 purposes or AI-enabled (audio interference). The 21 problem is the states-based in use case areas are so 22 broad we're stuck with trying to wrangle all of it, 23 but the near-term things that we see within the next 24 three years probably are going to be in that design 25 area where machine learning is used to make a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
36 1 recommendation or a prediction to a human and then 2 that human has to do something with it.
3 And so we're left in a position where we 4 need to be prepared to evaluate that instantiation of 5 that machine learning or AI to make a recommendation, 6 whereas it's very different from the way that it's 7 been done currently. And so that's what we're 8 presented with. And how do we basically -- if that 9 application area is presented to us, how do we review 10 and evaluate and make a technical finding?
11 MEMBER KIRCHNER: Yes, that helps because 12 that narrows things down with still the idea that 13 research should be looking at what's over the horizon 14 as well. Okay. Thank you.
15 MR. DENNIS: Yes, and we are fully looking 16 at -- I think it has been mentioned several times, is 17 AI is not entering the control room at this point.
18 That's been stated at the public workshops several 19 times. And we do believe that we need to be prepared 20 for that potential eventuality, but that is not 21 something that is right in the near term. The near 22 term is in using AI ML for design recommendation, that 23 type of stuff.
24 So I think that's enough on this slide.
25 I'll move onto the next one. The purpose of it was NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
37 1 for us to host a workshop. So the AI Working Group 2 convened to prepare for a workshop and provide 3 feedback on regulatory and technical issues 4 surrounding AI usage in nuclear applications. All of 5 this, all of these workshops, the previous three, 6 informed the preparation of the strategic plan. And 7 this was the first workshop we held after the issuance 8 of the strategic plan and preparation for the project 9 plan, which you'll hear more from Anthony on. But all 10 of this was to prepare us for what is going on.
11 And so we had three panel sessions. The 12 first one on regulatory perspectives. The second one 13 which was more academic in nature on safety, security, 14 and explainability topics. And then the third one was 15 more industry-focused on those AI application 16 considerations and some of the examples of use case 17 areas where it is being considered from industry.
18 So this is just a snapshot of the agenda.
19 It was a 10:00 a.m. to 4:00 p.m. meeting. I won't go 20 into this other than just to point out that all of the 21 presentations are available in ADAMS and on our 22 website, which there was a link on the previous page.
23 But you can see here we had a number of 24 presenters. We had CNSC, ONR, IRSN, and a think tank, 25 Responsible AI Institute, present on regulatory NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
38 1 perspectives. Again, I said it was more academic-2 focused in the second panel session. And then some of 3 the industry presenters who are actually participating 4 here today were in our panel session in the afternoon 5 talking about how they're using it.
6 So all of this was to support the 7 strategic plan and build out and build upon a table 8 that we had in the strategic plan that talked about 9 the notional AI and autonomy levels in commercial 10 nuclear activities. It was what we put in the 11 strategic plan to start the conversation which is 12 happening here today and has happened at every single 13 meeting about where AI is being inserted into nuclear 14 activities.
15 The table had a range 1 to 4, from 1 being 16 just basic -- something that's making a recommendation 17 all the way up to level 4, which would be more like 18 what was talked about about autonomous operation of a 19 vehicle, so where you're actually running a power 20 plant using AI.
21 So that was to frame the discussion. And 22 we wanted to use that as a springboard for our working 23 group who has discussed all of these things that have 24 been brought up. We have gone back and forth and 25 talked about these. But that was our starting point NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
39 1 for the workshop and to develop our matrix of AI 2 characteristics for regulatory consideration that I'll 3 talk about in the next slide.
4 So our working group was in a very agile 5 fashion, as has been all of this with AI, because 6 again it -- when we started the strategic plan two 7 years ago on this journey ChatGPT didn't exist. And 8 then things changed so we had to pivot the way that 9 we're looking at this and the way that it's being used 10 in industry. So we've had to be very agile.
11 So these are the members of the working 12 group. Again, some of them are available today to 13 answer any questions. But the disclaimer for our 14 portion of the presentation as far as AI 15 characteristics for regulatory consideration is we are 16 aware that they're -- NIST is the agency chartered for 17 the Federal Government to develop the AI risk 18 management framework. So we're aware of these. What 19 we presented was not an exhaustive list and we 20 recognize that it's on a broad spectrum. So this is 21 quite a large matrix with a range of applicability.
22 So this is the NRC staff's presentation at 23 the workshop focused on these eight characteristics:
24 safety significance, AI autonomy, safety, 25 explainability, model life cycle, regulated activity, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
40 1 regulatory approval, and application maturity.
2 Again since this is just a recap, I won't 3 go into all of these other than to say we had a 4 discussion on all of them and we didn't get much 5 feedback on it, so take that as you will. Either 6 there was agreement that these are all concerns or 7 just recognition that these will be part of the matrix 8 of decision making that goes into considering AI 9 applications and usage in the nuclear domain.
10 MEMBER BROWN: Did you mean feedback from 11 us or from your workshop?
12 MR. DENNIS: From the workshop, yes.
13 MEMBER BROWN: Thank you. It would have 14 been a long letter.
15 (Laughter.)
16 MR. DENNIS: Yes. So I guess the takeaway 17 here is we don't have an answer for all of these 18 things right now. This is what the working group kind 19 of coalesced on as some characteristics. And I will 20 say we are -- Anthony will mention this later -- we're 21 participating in a trilateral working group with CNSC 22 and ONR, and these line up quite well with other 23 considerations from international -- our international 24 regulatory counterparts as well as the IAEA. All of 25 these topics have come up over and over again. So NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
41 1 we're in good company, I guess you could say.
2 So regulatory perspectives on AI panel 3 session. I want to get to this so that we can keep on 4 schedule. This is some of our observations on the 5 three panel sessions, so the next three slides I'm 6 going to give a synopsis with the disclaimer that this 7 is a summary of some of the comments that were 8 provided during those panel session discussions and 9 presentations.
10 So the first one again was regulatory 11 experts and safety experts from other regulatory 12 entities globally and domestic think tanks. So CNSC 13 pointed out that they have stood up their Disruptive 14 Innovative and Emerging Technologies Working Group, 15 DIET. I don't know if they forced that acronym to be 16 fun for DIET or if it came out -- I think it actually 17 -- they added it to make it DIET. But they 18 commissioned a study last year on how the CNSC can be 19 prepared for AI applications.
20 And the U.K. ONR -- I skipped one past, 21 but the U.K.'s ONR has also issued a report.
22 Similarly the U.K. ONR report, I will point out, has 23 a very nice appendix on how -- on some different 24 methodologies to consider how you would evaluate or 25 review AI applications. And then they have done a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
42 1 regulatory sandbox. So this is one of the 2 observations we made is that they've done two AI ML 3 applications as part of a regulatory sandbox. I don't 4 have the link here unfortunately, but they just this 5 week published a report on this. So that is a 6 publicly-available report now on their website.
7 But they believe that their regulatory 8 approach is capable and flexible enough in the absence 9 of standards, which everyone has recognized is a 10 shortcoming, because standards are currently in 11 process for being developed right now. We recognize 12 that it can take a long time to get a standard through 13 the process. Even NIST commented that at the 14 standards forum at the NRC a couple months ago. So 15 NIST knows this; we know this, but the ONR thinks they 16 have a flexible enough framework that they can move 17 forward without standards, if need be.
18 IRSN, again they are under the umbrella of 19 the E.U. AI Act, and so some of their key areas for 20 high-risk AI applications that they called out are 21 data governance, risk management, and the human 22 component, which keeps coming up again and again in 23 discussions.
24 The Responsible AI Institute has been 25 working on a certification methodology. Their view is NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
43 1 bigger than nuclear. It includes fair housing, all 2 the AI application areas. So they've been trying to 3 develop a certification framework. They do have one 4 and it's largely based on these two -- the AI RMF from 5 NIST and an ISO AI management systems approach. So 6 they discuss their work on developing certification 7 methodologies.
8 MEMBER KIRCHNER: Matt, I could point out 9 that getting a boiler and pressure vessel code 10 standard (audio interference) is -- and that's as 11 well-defined problem. Often takes years. But more 12 relevant here, what about your companion agencies here 13 in the government? I'm thinking in particular FAA 14 must be looking at this because of congestion in the 15 air and so on and using advanced techniques to avoid 16 collisions, whatever the application might be. So 17 are there counterparts to you, NRC, here in the 18 government that you're also at that are using further 19 applications?
20 MR. DENNIS: Yes. I will point out that 21 we are in good company. I mentioned it at the 22 beginning and a little bit in passing that we are --
23 we stay in contact with FAA. FDA presented on their 24 regulatory approach for AI at the RIC last year. We 25 invited them and have talks with them.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
44 1 Trey and I attend a number of DoD meetings 2 for the Navy. So a lot of presenters there from the 3 Navy, from the Army, and they're grappling with the 4 exact same issues that we are. And they don't have an 5 answer either. And the standards thing came up again.
6 So the Army has been doing work on autonomous vehicle 7 operations, and they have a Testing, Evaluation, 8 Verification and Validation Working Group that's 9 looking at this as well.
10 So we've stayed -- tried to stay plugged 11 in with all of our federal partners that are working 12 in this area to leverage learnings from them and 13 research that they're conducting.
14 MR. BETANCOURT: Matt, can I mention 15 something quick on that one?
16 MR. DENNIS: Yes.
17 MR. BETANCOURT: This is Luis Betancourt 18 from the staff. On the FDA side, like Matt mentioned, 19 when you look up on that table that he put on the 20 model life cycle, that was one of the things that we 21 learned from them, that they actually released some 22 draft guidance on locked models, with some open 23 models. So we have been actively involved in learning 24 from them and vice versa. So there has been that 25 synergy and basically cross-pollinization between us NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
45 1 and other agencies.
2 MEMBER KIRCHNER: Not to make any 3 observation on our government, but the difficulty --
4 to the extent that the DoD activities are in the open 5 is fine, but I think one of the big challenges in AI 6 applications is transparency and openness. The 7 military has its needs and often those needs require 8 classification and such, but the other agencies you 9 mentioned: FDA and FAA, that's why I brought them up 10 because they obviously also have to convince the 11 public that any applications that they were to use 12 would be transparent, safe, the integrity, all the 13 issues that the NRC has to (audio interference).
14 MR. BETANCOURT: And on that, Walter --
15 this is Luis Betancourt again. I'm going to be quick 16 because I know that we're running out of time. Vic 17 and I, we are attending meetings of the responsible AI 18 officials from other agencies. So to your point, 19 like, yes, the Department of Defense has their needs, 20 but there's also this big push by the government of 21 hey, we need to be able to regulate AI, but also how 22 do we do AI responsibly internally?
23 So we're keeping tabs really well on what 24 is happening, not only with the defense industry, but 25 also other industries as well.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
46 1 MR. DENNIS: So the next session, the 2 panel session was on a more academic nature. It was 3 an excellent session because it really did discuss 4 some of things about how would we evaluate the 5 technology? And there is a lot of research coming on 6 on this topic right now. NIST is looking at using 7 combinatorial methods. The presenter from NIST is 8 actually one of their funded research projects to 9 support the AI RMF. So NIST is undertaking these 10 projects to -- and one of the messages from the NIST 11 presenter was that this is different from conventional 12 assurance processes for autonomous or software-based 13 systems and there are alternative methods that they're 14 looking at that can go to that explainability problem 15 that keeps getting brought up for AI.
16 George Mason was recognizing that there is 17 an issue with explainability and that using 18 counterfactual testing is one method that could be 19 used. And they have a research project that's ongoing 20 right now to use counterfactual cases to expose the 21 black box nature of AI models.
22 And I see we have a hand. I don't know, 23 Vicki, if it came up, but --
24 (Audio interference.)
25 MR. DENNIS: Okay. All right. Okay.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
47 1 I'll keep going then.
2 All right. Georgia Tech discussed some of 3 their cybersecurity research. They have a test 4 facility that they are looking at multi-layered tests 5 using a honey pot scenario to do cybersecurity 6 monitoring using AI ML.
7 And NC State talked about one that was --
8 the word was mentioned earlier about uncertainty 9 quantification. And this is an area near and dear to 10 my heart, on using VVUQ, verification, validation, and 11 uncertainty quantification methods to root out the 12 black box nature of deep neural networks. So there 13 were a few examples that were giving. Monte Carlo 14 dropout, deep ensembles, and Bayesian neural networks 15 were talked about and an example was given in an 16 application area to predict axial neutron flux 17 profiles.
18 So the presenters, all of these presenters 19 from the academic session were really talking about 20 issues that we have and ways that the research is 21 being used to try to explain AI in a way that can be 22 understood, which is quite of interest to us at the 23 NRC.
24 And then the application consideration 25 panel was more industry-focused. We had several NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
48 1 presenters. The first one from Constellation and 2 Jensen Hughes mentioned that in the absence of 3 industry-specific V&V guidance for software that's 4 driven by AI they've come up with their own process.
5 So they talked about their V&V documentation that they 6 developed and how they're looking at it from an 7 explainability perspective so that their users within 8 Constellation are able to understand the model and 9 what it's doing as well as anyone externally that may 10 be evaluating that model. So they have made some 11 significant progress in being able to explain how 12 their AI-driven CAP analyzer is actually functioning.
13 The Utility Service Alliance talked about 14 their Phase 1 projects in their Advanced Remote 15 Monitoring Project. I think INL later in the 16 afternoon will be talking about a couple of these 17 actually, so I won't go into great detail here. But 18 one point that they made was that they assessed that 19 the regulatory readiness level is at a two out of five 20 and they are planning for a Phase 2 where they're 21 going to explore more AI-drive autonomous inspection 22 rounds and response projects. So they do have an 23 interest in this area.
24 MEMBER REMPE: (Audio interference) these 25 applications I could see how yes, you might be able to NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
49 1 use the system to help detect, but the response aspect 2 is where I'm curious. Because sometimes, for example, 3 a fire suppression system could have adverse impact on 4 the staff in the plant if there's not a human to view 5 the AI detection and say yes, I agree with it; let's 6 do this. And it takes a mitigating strategies with 7 the staff before the system goes online.
8 And if you have to have that human review the 9 data, does it really -- does it not add more time and 10 just having the human do the fire watch?
11 And are those kind of questions coming up?
12 Because that's one example, but it seems like there 13 would be other examples where you don't want the 14 software to initiate an inaction. And it's not a 15 criticality in the control room thing. I'm just 16 thinking about other actions that happen in the plant.
17 Because of my experience at the lab, I know where bad 18 things can happen in some of these systems and I -- is 19 that coming up and people are thinking about that?
20 MR. DENNIS: Absolutely 100 percent. We 21 have had a number of discussions with our human 22 factors folks. And in other presentations I talk 23 about a Tesla crash where the system basically --
24 there's an accident, but the system defaults to the 25 human, but the human only has three seconds to NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
50 1 respond.
2 So we're talking about application areas 3 where right now humans are in the loop. And there is 4 no intent to take -- as far as AI being used 5 necessarily to take the human out of the loop, but 6 there is a concern or a recognition that humans cannot 7 sit and just kind of toil away and be completely 8 oblivious and then be expected to then respond 9 immediately to something that is a time-sensitive or 10 critical thing. And we know that from the existing 11 control room configurations and automatic systems.
12 So there is a recognition that if you're 13 going to go -- there's sort of a blended area here 14 that's problematic where you have a human and then 15 autonomous operation that's AI-driven. So we do 16 recognize that there is a human factors component.
17 And that's one of the things we actually called out in 18 our AI characteristics for regulatory consideration 19 was this concern.
20 MEMBER MARTIN: (Audio interference) 21 follow that one. From what you said there about the 22 human, human's role, you hear more about AI's 23 performance, but -- and opportunities for applying AI 24 where we can apply that capability. What I've not 25 heard -- and I'm not talking about just today, but NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
51 1 applications of AI to drive human performance, improve 2 human performance, setting standards.
3 I think about when we were all younger and 4 you hear about IBM's Deep Blue, right? When it began 5 the better players could win. And then eventually it 6 beat the masters. And the same thing could be 7 considered here. Instead of focusing on letting AI 8 take a call, AI could be used to set the standard for 9 how humans perform. Obviously it's not the only 10 thing. We're not talking about replacing all training 11 with a robot. But nonetheless, I've not heard that.
12 And I think we need to think more human-13 centric on these things and not machine-centric and 14 expose some bias. But we'd like to see the future 15 focus. And it could be a question on later 16 presentations on where we could take AI to review.
17 MR. DENNIS: Thank you for that 18 distinction. And it has been one -- I will mention 19 the industry has presented on using it for that 20 purpose, using machine learning for improving operator 21 examination, really using the tool to make us the best 22 version of ourselves. So that is definitely in the 23 use case area. I know we're not -- doesn't get all 24 the credit or focus here, but it is an area of use.
25 MEMBER HALNON: This is Greg. I got a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
52 1 question. Constellation said there's no clear 2 specific guidance for validation. We have a software 3 quality assurance (audio interference) program, 4 numerous regulatory documents. How far off are those 5 if you were to lay those into AI? Is it a starting 6 point only or is it N/A? Is it not applicable? How 7 does that look?
8 MR. DENNIS: Going back to -- I'm not 9 going to go back in the slide deck, but that --
10 MEMBER HALNON: I'm sorry. I had to leave 11 for a few minutes so I apologize.
12 MR. DENNIS: Oh, I didn't talk about it.
13 MEMBER HALNON: All right.
14 MR. DENNIS: I'm not going to flip back, 15 but that is one of the areas where we say we have a 16 foundation of excellent guidance on software quality 17 assurance, a VVUQ for modeling and simulation. We 18 should start from that point.
19 MEMBER HALNON: So it's a starting point?
20 MR. DENNIS: That is a starting point.
21 And so the example that Constellation gave, that's 22 what they started with.
23 MEMBER HALNON: Okay.
24 MR. DENNIS: They started with the typical 25 process you use for software V&V and then layered on NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
53 1 some unique stuff for the AI-driven aspects of it to 2 build that out.
3 MEMBER HALNON: Okay. So it's not a 4 perfectly round wheel, but it's certainly starting 5 with a wheel?
6 MR. DENNIS: Yes. And that was our 7 observation for the working group. And one of our 8 eight characteristics was you start -- we start from 9 what we have, and a lot of that is good. And the 10 observation that Anthony will point out is we're 11 starting a project this year as we speak to go into 12 that aspect of looking at -- we're doing a gap 13 analysis right now and then we're going to look at 14 what methodologies could be used.
15 MEMBER HALNON: Okay. Thanks.
16 MR. DENNIS: So I will quickly go through 17 the last two. Westinghouse emphasized the importance 18 of having an ethical AI corporate policy and a 19 recognition that the human is not the best interpreter 20 of AI. So there needs to be some component to the 21 uncertainty quantification through validation metrics 22 that are interpretable by the human, but not -- this 23 goes to the point of you may see an AI-generated 24 image, computer-generated image and you think it's 25 real. So the human can be easily spoofed.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
54 1 TerraPower discussed that there's no 2 specific AI use cases or plans to use AI, but that 3 highly passive future designs that do have potential 4 for this use case -- and that there were some high-5 level thoughts presented on AI -- using AI for 6 engineering document preparation and that we need to 7 consider how to validate AI recommendations for 8 licensed operators and if we should reevaluate the 9 role of the human operator and what they play in the 10 plant. So this was sort of the point that was brought 11 up just a minute ago.
12 Our key takeaways: I think I have two 13 more slides, so I'm going to be real quick. The panel 14 sessions confirmed that we remain well-informed of 15 international AI regulation and domestic projects.
16 There were no surprises or show stoppers. So I guess 17 the message here is that we feel that we've been doing 18 a pretty good job of keeping the beat on what is going 19 on for use cases and applications within the nuclear 20 industry.
21 We did hear a lot of feedback on the 22 regulatory sandboxes and how those provide a unique 23 opportunity for industry and regulators to 24 collaborate. So there is some interest in that topic 25 area. And industry representatives encourage NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
55 1 continued collaboration to pursue pilot studies and 2 proof of concepts for a future foundation for 3 reviewing the use of AI in NRC-regulated activities.
4 Some of these considerations we have 5 already talked about, but I guess I'll just 6 reemphasize them as a transition to Anthony's 7 presentation. We era currently in the mode of looking 8 at what traceable and auditable evaluation 9 methodologies exist in order -- and this is the 10 project I mentioned that we're going to be kick-11 starting right now to do that.
12 And then we're also -- the workshops have been 13 supporting our ability to understand what licensees 14 and applicants are using in AI.
15 The future goes towards differentiating 16 this for design versus AI-enabled autonomy. I did 17 mention that design usage is the one that seems to be 18 front and center as a use case. And then also how are 19 we going to explain and evaluate it? Is this a 20 reliability or assurance argument methodology? So 21 those are the things in the future.
22 And of course all of this is predicated on 23 our budget and preparation for these emergent industry 24 applications, which, like ChatGPT came up, there's a 25 whole slew of different ways to use it, for generating NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
56 1 documents and using it for regulatory applications 2 potentially. So that was stuff that wasn't envisioned 3 when we originally were writing the strategy.
4 So moving forward we continually focus on 5 safety and security. That's paramount. Our 6 partnerships, as we mentioned, with domestic and 7 international counterparts and our engagement with 8 other federal agencies has been very beneficial and 9 we're continuing to pursue those. And we recommend 10 and encourage our stakeholders to engage with the NRC 11 early and often on plans and operating experience 12 about how they're potentially going to use AI or 13 looking to use it and what their experience has been.
14 And we've gotten a lot of that feedback from the 15 workshops and it's been very beneficial.
16 Our internal working group will be 17 continuing to focus on AI characteristics for 18 regulatory consideration following our feedback that 19 we get from our gap assessment which we are currently 20 in the process of going through and will be concluding 21 in spring of 2024. That will also be providing the 22 content for our next workshop which we do plan to have 23 in summer of 2024.
24 So I do believe that takes me -- that is 25 the end of my slides. So thank you very much.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
57 1 CHAIR BIER: Yes, I think one of the 2 people online has an open mic. If you can check that.
3 Thank you.
4 And, Anthony, now I think we can move onto 5 you. Thanks.
6 MR. VALIAVEEDU: Well, thank you again for 7 allowing me the opportunity to speak to you all today.
8 My name is Anthony Valiaveedu. I'm part of the 9 Nuclear Regulatory Commission working out of the 10 Office of Research as a data scientist. Here with me 11 at the table today, as previously introduced, is Matt 12 Dennis and Trey Hathaway, who also work in the Branch 13 of Accident Analysis. Special thanks to our 14 management team including Paul Krohn from Region I, 15 who is a division director; Victor Hall, who is a 16 deputy division director for Division Systems 17 Analysis; and Luis Betancourt, who's leading our 18 branch in the Accident Analysis Branch.
19 This presentation today is only been 20 possible through the efforts of the entire agency.
21 All the program offices were involved during the 22 development of the agency's strategic plan towards AI.
23 Paul is from Region I. And pictures used throughout 24 this presentation today (audio interference) staff 25 members including from the Office of Nuclear Reactor NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
58 1 Regulation as well as the Office of Research. This 2 supports again the notion that AI has the potential to 3 touch every portion of this agency's mission of safety 4 and security. As indicated with this graphic we've 5 collaborated with a variety of program offices 6 including NMSS, NSIR, the regions, OCIO, OEDO, and 7 many others.
8 Over the past few years we were notified 9 by various stakeholders that they have had plans to 10 implement artificial intelligence into their current 11 operations and businesses, and as a regulator the NRC 12 stands by the safety and security of the protective 13 order and the environment. Determining the three S's 14 of safety, security, and safeguards is the duty the 15 NRC and we as staff who have prepared this 16 presentation provide information on the status of our 17 mission.
18 And as we have previously presented to 19 this Committee during the development of the AI 20 Strategic Plan, I'll provide a quick debrief of the 21 development since that time and specific implications 22 or interests for the Committee's consideration. And 23 to highlight previously about the interdisciplinary 24 nature of our team these are some pictures from the 25 workshop that includes Paul Krohn from Region I; Jesse NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
59 1 Seymour, who works in the Human Factors Branch; Joshua 2 Kaiser, who's actually here in the room with us today 3 with the Office of Nuclear Reactor Regulation and the 4 chair support for our responsible use of artificial 5 intelligence.
6 This slide, slide 27, provides a timeline 7 of events since the last presentation that we provided 8 to this Subcommittee as indicated by the star. Since 9 the last ACRS presentation, around June of 2022, we 10 began collecting over 100 comments on the draft AI 11 Strategic Plan; ADAMS accession number is indicated on 12 the slide, and utilized those comments to issue our 13 final AI Strategic Plans for fiscal years of '23 to 14 '27, which is in NUREG-2261.
15 In March we also launched our AI Steering 16 Committee. This centralizes our efforts to -- on 17 artificial intelligence to make sure we're better 18 p r e p a r e d a s a n a g e n c y .
19 In July we initiated an AI regulatory gap 20 assessment. And in September, as Matt presented 21 earlier, we hosted a workshop for regulatory 22 considerations.
23 Later in September we launched our AI 24 community of practice for discussions of lessons 25 learned and potential uses cases of artificial NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
60 1 intelligence. In late October we issued our project 2 plan. And the ADAMS accession number is listed 3 accordingly. All these will be explored further into 4 detail as we continue on with these slides.
5 For the strategic plan to enable industry 6 and to lead our mission of safety and security the 7 goal of the AI Team is to stay with the development of 8 AI so that during the deployment of these tools the 9 NRC will have the ability to review any safety or 10 security implications. The mission of the AI Team is 11 to be -- is to enable a responsible use of AI. And 12 wishing to be cautiously proactive we released a 13 strategic plan in May of 2023.
14 The strategic plan outlines five goals 15 similar to the ones that were presented about a year 16 ago. They include regulatory readiness, establishing 17 an organizational framework, strengthening 18 partnerships, cultivating a proficient workforce, and 19 goal 5, which is to build an AI foundation within the 20 NRC. The status of these goals will be presented in 21 the subsequent slides.
22 Along with the strategic plan we've also 23 issued the project plan in October of 2023. This 24 project plan goes into depth of the strategic plan's 25 goals as well as sets the scope of these goals. It NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
61 1 provides key timelines as well as tasks to ensure 2 adequacy that we're meeting our metrics, and it 3 promotes communications for external as well as 4 internal stakeholders. Its purpose is to provide the 5 public with transparency and accountability while the 6 current staff plans are an applicant's clarity into 7 the NRC's roles and responsibilities.
8 The timelines that we've had generally 9 match with the expected deployment of AI that we were 10 able to obtain with stakeholder feedback, however 11 currently with the timelines we hope to continually 12 update the project plan because of the changing 13 current -- the current change of political climate.
14 Goal No. 1 is on regulatory readiness, or 15 what we like to call keeping the end in mind. With 16 every journey knowing what you're working towards 17 helps provide that mission to perspective.
18 On ongoing work I want to highlight three 19 items: pre-application communication, our gap 20 analysis, and our continued with the IEC.
21 On the regulatory gap analysis; we can 22 start at the top, we're currently working on reviewing 23 regulations and guidance as it applies to current gaps 24 in policies before AI. While conducting this gap 25 assessment we're also incorporating and reviewing NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
62 1 applicable standards for artificial intelligence to 2 recommend updates or recommend new standards to be 3 developed.
4 Regarding pre-application communication, 5 to help better budget and resource plan for AI 6 applications the staff plans to develop a strategy to 7 collect information for AI scheduling by the industry.
8 These surveys could include RISs and FRN, but as Matt 9 previously highlighted, what we've currently been 10 doing has been extremely beneficial, which is 11 conducting public workshops and information gathering 12 at conferences where there's industry and labs that 13 are participating. These have been extremely fruitful 14 discussions.
15 The third item is the IEC, or the 16 International Electrotechnical Commission, the NRC's 17 participating Subcommittee 45 Alpha and Working Group 18 12. 45 Alpha is specifically on the instrumentation 19 control and electrical power systems in a nuclear 20 facility, and Working Group 12 is more specific to AI 21 applications in these nuclear facilities.
22 I want to preface this by saying this 23 working group is very new as a second meeting only 24 occurred early in November, and we have four staff 25 members involved so far; three from the Office of NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
63 1 Nuclear Security and Incident Response and one from 2 the Office of Research, which is Trey Hathaway.
3 The IEC plans to develop and maintain 4 standards for AI applications for nuclear facilities 5 by providing guidelines to stakeholders who are 6 developing, deploying, as well as overseeing AI 7 applications. In addition to this they hope to cover 8 fundamental characteristics of AI in these nuclear 9 facilities and make it applicable to the entire 10 nuclear life cycle. The IEC --
11 MEMBER KIRCHNER: Could I interrupt you 12 here?
13 MR. VALIAVEEDU: Yes.
14 MEMBER KIRCHNER: We through this 15 Committee with Charlie's encouragement encouraged NRR 16 to lay out a road map of the digital I&C. Are the 17 staff who were responsible for that -- it's a very 18 nice road map, by the way, of a very complicated 19 wiring diagram for all of your regulations and guides 20 and instructions and such for digital I&C. So it 21 seems to me you have a framework in place if the 22 application is actually going to be somewhere in the 23 control systems or an operation in the plant. Are the 24 NRR staff involved in this?
25 MR. VALIAVEEDU: We've incorporated NRR NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
64 1 with all discussions we have. We currently have an AI 2 Working Group that meets monthly. That includes staff 3 members of NRR as well as -- multiple staff members at 4 NRR. But if you're talking about specifically the 5 engagement with this working group, we only have three 6 from the Office of Security and Incident Response 7 and --
8 (Simultaneous speaking.)
9 MEMBER KIRCHNER: I don't want to get 10 involved in NRC management decision and such, but --
11 don't take this critically, but the security people 12 look at things after the fact. They're looking at 13 things that control access, they're looking at 14 cybersecurity. I'm thinking of it a different way 15 altogether. You're going to imbed some application 16 somewhere in the plant. And it seems to me that's 17 different than checks and balances as to whether you 18 had an intrusion, et cetera.
19 So my concern or my suggestion here is yes 20 to them, but involve the people who are intimately 21 involved in how the plant operates from the control 22 standpoint and the regulations and framework that is 23 used for that. And then the applications in my mind 24 of AI, at least the early application, somehow will 25 come into that -- have to come into that regulatory NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
65 1 framework.
2 So I just point out that there was a --
3 this very nice road map that's been put together that 4 addresses digital I&C. To the extent that AI is going 5 to come into the digital I&C regulations for the 6 agency those people need to be involved. So end of 7 speech.
8 MR. SEYMOUR: So, this is Jesse Seymour 9 from the Operator Licensing Human Factors Branch, and 10 I just wanted to speak to that point.
11 So, both myself and David Desaulniers, and 12 also in prior efforts Dr. Brian Green as well, have 13 been involved in the AI efforts. And if you were 14 going to create a Venn diagram of who's working on the 15 digital I&C upgrades that are currently in progress at 16 some of the plants like Limerick, Turkey Point, 17 myself, Dave, Brian are all involved with that as 18 well, too. So I think that there's a good kind of 19 synergy between the folks that are considering the AI 20 issues as well as the advanced digital I&C control 21 systems that are involved. And there definitely is a 22 sensitivity to where we're at on kind of the 23 progression towards when we may or may not eventually 24 see AI in any type of a controlling context.
25 Right now in terms of the implementation NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
66 1 of AI what we're really seeing in applications -- and 2 by that I just mean usage out in the industry, not 3 necessarily even in a regulated context. And even in 4 pre-application discussions it's more so at the level 5 of AI insight, again if we were going to think about 6 that kind of zero to four hierarchy of AI autonomy.
7 And what we're not seeing is anything right now or in 8 the near term that would take the human out of being 9 the decision maker, whether that's in any type of 10 operational context or even in the sense of 11 calculations.
12 I think a good working example of where 13 we're currently seeing things as currently state-of-14 the-art is using machine learning to -- a good example 15 would be provide training insights, training 16 interventions as folks are going through training 17 programs and things of that nature where again it's 18 informing human decisions. But we're not yet seeing 19 it in a controlling context.
20 MEMBER KIRCHNER: Well, let me give you 21 some examples of using big data in actual --
22 potentially. You look at a core map or you look at --
23 right now already there are software implemented that 24 looks at a large array of data from the core:
25 thermocouples, nuclear and such, and those NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
67 1 calculations like what's your margins, CHF, and so on.
2 So I can see already that using advanced processing 3 techniques you could improve the performance of those 4 things that are already done analytically now and then 5 somehow feed into the actual plant operation.
6 So I just threw that out because I can see 7 that happening with the existing plants and that 8 somehow that, quote/unquote -- calling it AI if you 9 want to, or just advanced data processing -- I can see 10 that being a kind of application in the plants. And 11 somehow that has to factor into the digital I&C road 12 map and regulation framework that you have. So it's 13 just a suggestion and I'll stop there.
14 And, yes, Jesse, we appreciate what you're 15 doing in the human factors arena as well.
16 MR. SEYMOUR: Thank you.
17 MR. VALIAVEEDU: Thank you.
18 MEMBER BROWN: Let me --
19 PARTICIPANT: We've got somebody on --
20 (Simultaneous speaking.)
21 MR. CARTE: Yes, Norbert Carte, Digital 22 I&C, NRR. So I have not been officially asked to 23 participate, but I am sticking my nose into this 24 stuff. And I am following what's going on. It's very 25 interesting. But really from a safety system point of NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
68 1 view I'm 100 percent aligned with Charlie. Keep it 2 simple, separate.
3 And the question is when if and ever we 4 change our paradigm -- so right now we base the 5 approval of equipment based on some conservative 6 limiting scenarios that occur at the worst possible 7 time. And in that sense there are some very simple 8 trip functions that protect you: high temperature, 9 high pressure, high flux. And there's no reason why 10 -- practical reason why you need to introduce AI into 11 any of those.
12 But if you were to change that paradigm 13 and no longer have conservative limited bounding 14 scenarios to size and establish the performance 15 criteria for your equipment, then you would need to do 16 some serious thinking. But right now as long as we 17 have conservative scenarios and simple separate 18 independent protective functions, AI is not going to 19 get into the protection systems themselves.
20 Now they will maybe reduce margins, 21 they'll reduce need for unnecessary maintenance, 22 they'll reduce unnecessarily challenges of the 23 protection system, but until we change that paradigm 24 there's no reason to have anything in AI -- sorry, 25 it's just my personal opinion, but just trying to be NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
69 1 clear. In the protection system a high-temperature, 2 high-flux trip does not need AI. And so, but there's 3 a lot of ways you could make a plant better with AI.
4 It's just it's going to be a long time before it 5 drifts into the actual safety systems if they remain 6 these simple safety systems that we have today. Thank 7 you.
8 MEMBER REMPE: So just a matter of 9 process, if there's a member who raises their hand, of 10 course we should bring them into the conversation. If 11 there's a member of the staff who wishes to make a 12 comment or a contractor, the staff needs to call on 13 them. Okay? Just so we keep the rules going. Thank 14 you.
15 I think Charlie had his hand up and wanted 16 to make a comment.
17 MEMBER BROWN: No. Two things: One, I 18 agree with you relative to who gets to what, but there 19 are some staff people that can support and they're 20 operating -- NRR and other digital I&C people who need 21 to be involved and understand what's going on in this 22 world. And if there's -- Norbert had some very good 23 comments, not negative, just how you integrate, and we 24 need to be -- we should be conscious of those as we go 25 through the meeting. So I appreciate you saying NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
70 1 somebody had their hand up. I think that was a 2 positive one.
3 The second one was an expansion, just to 4 follow up on Walt's comment about the road map. We 5 had a meeting -- this is -- I understand your 6 comments, but the road map was really trying to 7 provide something that shows where all the standards 8 and specs at the various levels of the I&C development 9 systems and how do they apply. And there was a 10 meeting on that where there was a set of presentations 11 and slides that -- I think it was NRR staff that 12 provided that. I'm not sure. My memory is not real 13 good on that right now. And that's what they 14 referring to.
15 We had that meeting back in April of this 16 year, April 3rd. It was a Full Committee meeting.
17 And that one presentation has a beautiful layout of 18 what we meant by a road map. It was not trying to 19 drive you any place. But it's not what you'd call a 20 Venn diagram. It has nothing to do with a -- if Venn 21 diagrams or -- if you want confusion, generate a Venn 22 diagram. That's my personal opinion.
23 Anyway, I just wanted to make it -- the 24 road map thought process that Walt brought up so that 25 they would know what we were talking about. And I'll NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
71 1 let you -- I'm still restraining myself, but I will --
2 I absolutely totally agree with Norbert, if you hadn't 3 figured that out by now.
4 MR. VALIAVEEDU: Thank you for everyone's 5 comments.
6 Moving on. For slide 31 we've provided a 7 timeline for our progress. The purple dashed line on 8 this slide indicates where we are today. The check 9 marks have indicated completed items. Task 1.1 is our 10 researching on our current regulatory framework, and 11 C what is applicable. We've been able to establish a 12 contract and we're currently drafting an analysis 13 report based off of that contract with a hopeful 14 completion date of the spring of 2024 for it to be 15 published.
16 We were able to incorporate Task 1.2 for 17 AI standards assessments within Task 1.1, so they're 18 being conducted concurrently. And we're continually 19 -- we're maintaining our ongoing participation in a 20 variety of standards forums as well. And we hope to 21 incorporate AI standards into our regulatory guidance.
22 CHAIR BIER: So just a factual 23 clarification looking at that slide. So you're 24 anticipating like end of fiscal year '27 there 25 actually will be regulatory guidance for AI. Are you NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
72 1 envisioning that is mainly going to apply to new 2 plants or use of AI in new capabilities at existing 3 plants, or both?
4 MR. VALIAVEEDU: The guidance is meant for 5 any stakeholders or applicants, who plan to utilize 6 artificial intelligence as the NRC's missions for 7 safety and security.
8 Task 1.3 is on a safety and security 9 framework. This will be dependent on the results of 10 our previous two tasks on standards, as well as our 11 current regulatory gap analysis.
12 And, we would utilize those results and 13 update or develop our current regulatory guidance as 14 needed.
15 1.4 is on pre-application communication.
16 We've begun discussing internally about additional 17 strategies to obtain industry and stakeholder 18 feedbacks, and plan to collect this information in 19 fiscal years 24 and adjust our planning information 20 accordingly.
21 1.5 is on AI enabled autonomous 22 operations. During our engagement with a variety of 23 stakeholders, we are not aware of any near-term 24 deployment of AI enabled autonomous operations, but 25 there has been interest in it.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
73 1 We plan to begin researching this in 2 fiscal year 25. The objective here is to develop a 3 technical basis, and requisite regulatory framework 4 for AI enabled autonomy in nuclear operations.
5 Considering out the various, in how the 6 various ways of how AI could impact autonomous 7 operations.
8 Goal 2 is on organizational framework.
9 Due to unique impact AI technology has on nuclear 10 applications, the staff is working on centralizing and 11 developing an internal organization for AI knowledge 12 and expertise.
13 There are three key areas here. One is 14 the AI steering committee. The second is the AI 15 community of practice, and the third is on a 16 centralized AI database.
17 On the first item, the A steering 18 committee, the A steering committee has an involvement 19 with a variety of program offices, and regional 20 representatives. It meets monthly on topics related 21 to AI within the NRC's purview.
22 This is being led by Deputy Division 23 Director Victor Hall, as well as Paul Crohn, who is 24 the Division Director in Region 1.
25 The AI team remains also in the Office of NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
74 1 Research, in the Accident Analysis Branch, that's 2 being overseen by Luis Betancourt, who is here today.
3 On the second column on the AI community 4 of practice, this was formed in September and it's had 5 three meetings so far.
6 It's a formalized community of practice 7 where the NRC has, within the NRC where people from 8 all over the agency have been able to share their 9 practices, as well as lessons learned.
10 It provides awareness for potential use 11 cases and activities, throughout the nuclear sector.
12 The third column is on a centralized AI 13 projects database. This has also been developed and 14 deployed by various other agencies, including the 15 Department of Agriculture, and Department of Treasury.
16 The goal here is to maintain transparency 17 to the public on AI ML technologies. And, we 18 currently have a dedicated public site for tracking 19 these activities.
20 We are currently researching into a 21 variety of use cases, such as data mining, as well as 22 mathematical modeling that will be presented later on 23 today.
24 And we hope to continue to update this 25 database reocurringly for accuracy, as well as NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
75 1 complete, for completion.
2 Regarding the timeline of this, we, the 3 periodic squares indicate when we need to update, or 4 plan to update a paper.
5 For Task 2.1, we were able to establish 6 the AI steering committee. Task 2.2 for the community 7 of practice, we were able to formulate and establish 8 the community of practice earlier in September.
9 In 2.3 for establishing a projects 10 database, we were able to develop an initial projects 11 database. However, we are continually updating the 12 projects database, and we hope to update it as needed 13 for accuracy and completion.
14 Again, the checkmarks indicate completed 15 items. The purple dashed line indicates where we are 16 today.
17 DR. SCHULTZ: Anthony, this is Steve 18 Schultz.
19 Just given the speed of activity that you 20 said the NRC is doing, as well as the AI developments 21 going forward throughout the government, I was 22 surprised that the updates that you have for these 23 elements are two years apart.
24 In other words, I'm expecting that things 25 are going to have to be updated more frequently. And, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
76 1 it sounds like you're having monthly meetings in some 2 areas, annual meetings in others.
3 And was wondering what kind of feedback 4 you have in areas such as the regulatory gap analysis, 5 for example.
6 MR. VALIAVEEDU: So specifically for the 7 regulatory gap analysis, we are concurrently, it's 8 ongoing work right now. We hope to have a issuance of 9 a report in spring of 2024.
10 Regarding the timeline you're talking 11 about in this slide specifically, for the charters, 12 those generally will be updated as needed, as 13 indicated with the parenthesis, as well as the, the 14 database that's going to be updated reoccuringly 15 because of the ongoing work.
16 And the change of political climate again, 17 will have us revisit the project plan, and we hope to 18 update the project plan's timelines in fall of 2024.
19 Because of the recent executive order and 20 chair's memo, which I think you're hinting at here, 21 happened earlier.
22 DR. SCHULTZ: It's not a criticism, but 23 that's the second time you used the political climate 24 associated with the overall program, and that hope 25 we're focusing on the technical climate as well, and NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
77 1 the developments there.
2 MR. VALIAVEEDU: Thank you.
3 On the technical climate, we are engaging 4 with stakeholders and making sure that their 5 deployment and timelines here match with what is 6 expected for the agency to maintain safety and 7 security mission.
8 DR. SCHULTZ: I'm glad to see that, too.
9 And, the interactions that you have been having on a 10 very frequent basis that are urgent.
11 MEMBER BROWN: I want to, this is Charlie 12 Brown again.
13 Yes, I was disturbed a little bit with the 14 political. Assuming the political climate should have 15 absolutely nothing to do with anything you all do.
16 Nothing.
17 It should be a zero impact. It should be 18 developed strictly, I mean it, you're really going to 19 raise a lot of hackles with people that sit on the 20 Advisory side of what we're looking at, if the thought 21 process is we're going to hurry up and do something 22 because somebody politically wants to, you know, 23 demonstrate that this is being done.
24 That's absolutely insane. And couple that 25 with part of your, go back, you don't have to go back NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
78 1 it's the previous one you were talking about, 2 evaluations looking at autonomous operation, that was 3 the words you tossed in kind of as you, when you went.
4 And the ability of having autonomous 5 operations. Presenting information or AI that 6 evaluates data that's coming out of the plant, and 7 then assembles it within some algorithm, or some 8 presentation that it informs the operator that hey, 9 these things are going in this direction.
10 And then the operator makes the decision 11 about what to do with it, or seeks decision from, you 12 know, consultation with what to do.
13 That's nothing wrong with providing better 14 data, because there are tons of data we're getting out 15 of digital systems now.
16 And it makes it very, very difficult you 17 know, to assess you know, the directionality of them, 18 and which one's pressure temperature can be going in 19 opposite directions and say, oh, when it gets to a 20 certain point, is that critical or not.
21 So that's, that's the type of things that 22 boy, throwing in the political thought process or 23 directive from whoever they is, is just useless.
24 And that's, that should just never get 25 encumbered in the develop of your all's processes.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
79 1 Anyway, I'm sorry to get wound up every 2 now and then. Did you all hear me satisfactorily?
3 I'm told I don't speak into the mic well enough as it 4 is.
5 MR. HALL: So if I can just jump in to 6 maybe, oh yes, thanks, thanks. So again, this is Nick 7 Hall, the Office of Research.
8 I guess political, it's a bad word, right?
9 So, we can talk about the religious influences on AI, 10 and maybe the other taboo topics that you don't 11 discuss at dinner.
12 I just wanted to offer a clarification is 13 it's the awareness of when we say a political 14 environment, it's the awareness of what the government 15 is doing, right.
16 There's, and there's two executive orders 17 that are the big ones, the biggies, that talk about 18 AI. One was from the Trump administration, the other 19 one is the Biden administration.
20 That's just a matter of statement of fact.
21 I certainly would back my staff in saying there is 22 never, ever any political consideration with the big 23 P. It's just the awareness what the government is 24 asking us to do, to make sure that we're prepared.
25 So I just wanted to clarify that because NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
80 1 I know as soon say you mentioned the P word, it, I get 2 the hairs on my neck stand up when anyone says 3 politics. That's how this country is, right, with 4 politics these days.
5 So, just a point of clarification, it 6 really is an awareness of how we ban best be prepared 7 for the technical. The folks in front of you are 8 engineers, scientists, and darn good ones at that.
9 I'll leave it at that.
10 MEMBER BROWN: But you don't want it to be 11 hurried. It should not be hurried. It needs to be 12 technically solid and validated before anybody goes 13 forward with anything like that.
14 And that's, that's not that the, we don't 15 obviously the government wants to make sure the 16 agencies are paying attention to things that may 17 enhance their operations. That works just fine.
18 But when you hurry place where we've got 19 spacing considerations because P, the government 20 really needs to emphasize and regulate that, that's 21 not, that's not the right emphasis.
22 So, I wasn't trying to be critical of 23 anybody up here that said anything. That's not the 24 point. The point was safety first, and introductions, 25 you know, comes along along with it. And look at NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
81 1 where it adds value.
2 Doing something because its fancy, new and 3 everybody thinks it's the greatest thing since sliced 4 bread, is one thing.
5 But if it doesn't add value to the 6 performance of the plant, it ought not be done at all.
7 That's the only point I'm trying to get across in all 8 your all's deliberation.
9 After laying these out, these pathways 10 out, you've got to go through that. You got to go 11 through the drill of how are you going to do this, and 12 how are we going to make it make sense.
13 But you've got to do it in the manner of 14 where does it add value. If it doesn't add value, 15 don't waste people's time; don't complicate the 16 systems with it.
17 So, that's just the message I was trying 18 to emphasize. I wasn't the only one.
19 MR. VALIAVEEDU: Thank you again for 20 everyone's comments.
21 Again, the NRC's mission is safety and 22 security for the people, and the environment.
23 Slide 34, we were moving on to Goal Number 24 3, which is strengthening and expanding partnerships.
25 The previous presentation by Matt Dennis NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
82 1 probably best explains how far we've been able to come 2 regarding developing closer ties, and promoting 3 partnerships through our proper workshop interactions.
4 However, with this slide I wanted to 5 highlight two additional areas. One is domestic 6 interactions, and the other one is international 7 interactions.
8 We've had talks with a variety of agencies 9 including the Department of Energy and the National 10 Nuclear Security Administration, on artificial 11 intelligence.
12 Specifically, with our work with the 13 Department of Energy, we've been able to observe areas 14 developing AI ML technologies through the Light Water 15 Reactor Sustainability Program, or LWRS Program. And 16 review those results obtained, as well as lessons 17 learned.
18 The DOE MOU has been extremely helpful in 19 understanding the direction industry is undertaking on 20 AI.
21 We are also in membership in the NIST RMF, 22 or Risk Management Framework working group. And this 23 has largely been an observational capacity, and we 24 have been providing our Nuclear Regulatory expertise 25 into these discussions.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
83 1 Other talks we've had also include 2 participation in multiple conferences, including the 3 Digital Engineering and Nuclear Technologies 4 conference, or DENT conference.
5 As well last week, Matt and I were able to 6 attend the Ohio State University Big Data Workshop.
7 These interactions have been geared 8 towards promoting clarifying stance in regulations for 9 stakeholders, and promoting communications.
10 As regulators, we can provide clarity and 11 share concerns through effective pre-application 12 engagement.
13 On the international side of the 14 wheelhouse, we are currently working with the United 15 Kingdom, as well as Canada, on a tri-lateral 16 engagement that we call CANUKUS.
17 I'll be discussing this in the following 18 slide more.
19 In addition to this, we've been engaging 20 with the IAEA on technical meanings that provided 21 insights into other nations' priorities in 22 developments.
23 There has been generally speaking, similar 24 concerns between nations on artificial intelligence.
25 In addition to this, during these technical meetings NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
84 1 there has been generally a main focus on utilizing AI 2 for operation maintenance, as well as design 3 utilization, as well as using AI as mostly as an 4 informative tool using natural language processing, as 5 well as mathematical modeling such as linear 6 regressions.
7 The third item here is on bilateral 8 engagements. The goal here is to foster and maintain 9 collaboration with international counterparts, and 10 multilateral organizations to positively influence and 11 maintain awareness on the responsible and safe use of 12 AI.
13 And, this is in support and alignment with 14 the NRC's 2014 international strategy to positively 15 influence safety and security, as well as maintain 16 awareness for the agency's domestic objectives.
17 DR. SCHULTZ: Anthony, Steve Schultz.
18 You mentioned, or it was mentioned that 19 the, there were presentations at the RIC associated 20 with AI and so forth.
21 And in terms of the international program 22 and plans, what's planned for the RIC, or around that 23 conference this year?
24 MR. VALIAVEEDU: I'm glad you brought that 25 up. We are hosting an IAEA technical meeting at NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
85 1 headquarters in March of 2024.
2 DR. SCHULTZ: Thank you.
3 MR. DENNIS: The button doesn't like my 4 finger, so I have to keep pressing it.
5 And I will mention, we are, we did get a 6 confirmation this week that we are doing another AI 7 technical session at the RIC this year.
8 So there will be an AI technical session, 9 you get to look to that. Please come and attend.
10 And what Anthony just mentioned is the 11 following week, we will be hosting a IAEA technical 12 meeting.
13 The week following, we think we're going 14 to get a lot of good participation -- the IAEA thinks 15 we'll get a lot of good cross-collaboration with 16 people that are going to be attending the RIC, as well 17 as then the technical meeting the following week.
18 Right now we haven't pinpointed what our 19 presenters are going to be for the RIC technical 20 session.
21 We're in the process of doing that right 22 now. But it will be similar to last year's where we 23 have a flavor of industry, other federal agencies, us, 24 and academia.
25 DR. SCHULTZ: Sounds like good NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
86 1 coordination. Thanks.
2 MR. DENNIS: Yes.
3 MR. VALIAVEEDU: Regarding the CANUKUS 4 engagement, CANUKUS is an interesting development in 5 the NRC as it marks a time where three nuclear 6 regulators were coming together to share a common 7 goal.
8 The outcome with this interaction is a 9 high leveled AI principles paper, that we hope to sign 10 by spring of 2024.
11 The goal is to provide a uniform front, 12 and what are key considerations when developing AI 13 systems for safety and security.
14 And, we are currently putting together the 15 first draft of this paper, but I want to preface this 16 by saying that this is not for legal use, nor used in 17 place of a regulatory framework. Instead, it provides 18 a summary considerations.
19 This includes discussions on how to 20 utilize existing safety systems, and how to utilize 21 those existing safety systems to demonstrate safety.
22 This will help assist developers in 23 evaluating their own system.
24 In addition to this, we will have, there 25 will be discussion on human factors, and how AI NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
87 1 impacts human factors, as well as an impact on AI 2 architecture, as well as a summary consideration 3 section for life cycle management, especially in the 4 context of generative AI.
5 Again, we hope to complete the paper by 6 the spring of 2024, as the working group was 7 formulated in November of 2022.
8 MEMBER PETTI: I just had a question about 9 which country really is feeling the greatest user push 10 for AI?
11 Is it still U.S. compared to Canada and 12 the U.K., or are they having to be ahead of us because 13 of their licensees?
14 MR. VALIAVEEDU: I will have to defer that 15 question to Matt, because he's been mostly working on 16 CANUKUS.
17 MR. DENNIS: So, Trey and I have been --
18 Trey and I have participated in two IAEA working 19 groups and we get to -- we've had the privilege of 20 being able to see globally where it seems to be 21 leaders and applicationaries (phonetic). The U.K. and 22 the U.S. still seem to be leading the charge on AI 23 applications in the nuclear sector.
24 China is somewhere in the mix there in the 25 middle of application areas, as well. Russia has had NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
88 1 a lot of presentation at IAEA.
2 So, it seems to be U.S., U.K., China and 3 Russia, as far as application areas. I will say I 4 think, and this may be biased, but I think the U.S.
5 and the U.K. are leading the charge on the regulatory 6 aspects of AI in nuclear.
7 And this all sort of blends in and makes 8 sense that the U.K., U.S. and China are sort of 9 leaders in AI.
10 If you go look, I forget the Alan Turing 11 Institute in the U.K. put together a website and the 12 U.K., or the OACD has a website, basically has a 13 tracker of AI leadership in the application areas.
14 And the U.K., U.S. and China are at the 15 top of that board as far as, and that's just, that's 16 not nuclear, that includes everything, right, but.
17 MEMBER HALNON: Great, no, this is Greg.
18 I respect, you know, all the learning that's going on 19 and I'm probably known throughout the industry as 20 being relatively impatient, so forgive me.
21 Working group formed in 2022 and the 22 output two years later is going to be a paper. How is 23 that moving forward fast enough in parallel with 24 what's all these applications and other things, and 25 other countries?
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
89 1 It just, I mean, we talk in kind out both 2 sides of our mouth, it feels like. It's going really, 3 really, really fast, and it's taking two years to get 4 a paper out.
5 And, that's all it's, that's the goal of 6 the outcomes is some of the, lot of meetings, lot of 7 technical presentations. Lot of learning going back 8 and forth.
9 And I recognize that's going to happen, 10 but it just feels like a tar pit.
11 Maybe you can comment on the speed and 12 the, the amount of resources we have applied, 13 dedicated to it.
14 If not dedicated, how are we going to move 15 this thing forward fast enough so that when somebody 16 does come up with an application, we're ready to let 17 it go.
18 I mean, you know, we're still, I guess we 19 got through the 50.59 stuff and digital INC, but that, 20 that kind of blocked a lot of digital INC upgrades for 21 a while.
22 And, I remember back in the 2012 23 timeframe, SMRs. Everybody was ready to go but we 24 just got the rules, or you know, dropping down the EP 25 Zone through the rulemaking process almost a decade NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
90 1 later.
2 How are we not going to government this?
3 MR. VALIAVEEDU: Matt's going to answer 4 the U.K. stuff for the international. But for the 5 timeline, we've formulated the original timeline based 6 off of our interactions with stakeholders.
7 What they thought they're going to be able 8 to deploy, at what frame of time.
9 And based off of that end time, we 10 formulated our strategy to ensure that we are ready to 11 evaluate that safety, evaluate the responsibilities of 12 AI.
13 The resources we're putting on to this, we 14 have a whole of agency approach as I mentioned 15 earlier. We have multiple program offices involved 16 with this.
17 Matt and Trey have been working on this 18 since the beginning, for the strategic plan. And, we 19 plan to update those timelines as changes happen.
20 So, we've been interacting with our public 21 workshops, seeing where industry is thinking about 22 doing AI. Is there a, has there been any radical 23 changes that happen.
24 Luckily at this previous workshop we were 25 told that our timelines expected, or matches the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
91 1 expectation that we see industry deploying AI usage.
2 Again, fall of 2024 is when we plan to 3 publish a new version of the project plan, to readjust 4 those timelines to ensure that we're maintaining that 5 adequate resources for the evaluation.
6 MEMBER HALNON: So do you see, Anthony, a 7 period of time when there's going to be a dedicated 8 not an office, but a directorate or something for AI 9 so that we focus this all agency approach with some 10 dedicated resources to establish these guidance 11 documents, get them through, get them through ACRS, 12 get them signed, and on the street?
13 Or is it going to continue do you think, 14 for a number of years at the matrix type?
15 MR. VALIAVEEDU: So, that's out of my pay 16 grade.
17 MEMBER HALNON: I know.
18 MR. VALIAVEEDU: So, I'm going to defer 19 that to Luis Betancourt.
20 MR. BETANCOURT: To answer that question, 21 I think what is important is that we have an AI 22 steering committee from all of the program offices 23 that's basically directing this work.
24 The idea that we want to do, is to see the 25 outcome of the regulatory gap assessment to really NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
92 1 identify what are, to your point, what guidance has to 2 be updated.
3 Do we need to develop something new? I 4 expect that to be some time in the spring, followed 5 with a workshop where we going to be talking with 6 industry to get their feedback, this is what we found.
7 Do you believe this is something, what are the areas 8 that we need to prioritize.
9 And then that's going to be going back to 10 the steering committee, and then we will go back to 11 you guys.
12 Kind of what Charlie was talking about the 13 roadmap, that we need to lay that down. Okay, now 14 that we know where the gaps are, let's sit down and 15 put this in front of us and ask everybody, so.
16 MEMBER HALNON: So, Luis, you're kind of 17 say that this steering committee is going to be key on 18 establishing the agency approach down the road.
19 MR. BETANCOURT: Correct.
20 MEMBER HALNON: Because, at some point, it 21 seems like -- and you know how long it takes to get 22 guidance documents written and through.
23 MR. BETANCOURT: Yes.
24 MEMBER HALNON: There's got to be some 25 real focused effort.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
93 1 MR. BETANCOURT: Yes, and I think that's 2 why we're starting that focus at the front end. And, 3 there will be some coordinating with guys early so we 4 don't wait.
5 Because as everybody's pointing out, this 6 is a fast-paced environment. But to Charlie's point, 7 we also want to make it technically right that we're 8 not basically putting like, efficiency in front of 9 safety, so.
10 MEMBER HALNON: Okay, thanks. And sorry 11 for derailing it. I just was reacting to the two 12 years to get a paper out. And, I understand that just 13 a lot of learning that has to go on.
14 So don't take it as a criticism, it's just 15 that you know, the regulatory timeline seems long 16 sometimes.
17 MR. DENNIS: I will respond, or I'll 18 mention the CANUKUS tri-lateral engagement.
19 Technically speaking, it's only, it hasn't even been 20 a year that we've actually been working on it full 21 steam ahead.
22 So, I recognize that the spring 2024 is 23 when you add the numbers together, two years. I'll 24 ask for some grace in this because we have, this is a 25 principles paper with three countries.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
94 1 And as a precedent for this, the same 2 three countries from the health and safety FDA 3 perspective put out a principles paper. And, it took 4 them quite a long time to come to a unified agreement 5 for a two-page paper.
6 And, we're trying to get more than that, 7 and get a little bit more in-depth. Because we took 8 that as our benchmark and said, FDA and health, or 9 Health Canada and FDA and the U.K.'s health office 10 came together and put out a good machine learning, 11 good machine learning practices and principles paper 12 a couple years ago. It was two pages.
13 We thought, this is not our benchmark for 14 what we want to put together. And we recognize 15 through our, some of our collaborations that this is 16 an area where applications may come in this to all 17 three of us. And having a unified perspective would 18 be a good thing.
19 So, just getting three international 20 entities go agree to the words on a piece of paper is, 21 is a challenge. So I'll say that.
22 MEMBER HALNON: All right, well I'll give 23 you grace.
24 MR. DENNIS: Thank you.
25 MEMBER HALNON: We all have waited at NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
95 1 11:00 o'clock for the weather forecast, and we're 2 disappointed when they spend 15 minutes telling us 3 what we already saw.
4 So, I just don't want to see the paper 5 come out and tell us that we're two years behind 6 everybody else.
7 So that, I understand what you're saying.
8 MR. DENNIS: Yes.
9 MEMBER BROWN: Just to echo that thought.
10 Has there been any effort by you all, and we talked a 11 little bit about this earlier, or at least me and 12 somebody else did.
13 To separate out what I would call the 14 areas of the rice bowl that you really need to get 15 regulation defined, whether it's via this tri-lateral 16 approach to doing business or not.
17 But there's a whole plethora of things 18 outside of that, that industry should just don't wait 19 for us. Just go work on those, do what you want to 20 do.
21 Why can't that be communicated in a manner 22 in your meetings and say hey, we're drawing lines 23 around plant controls, these controls, safety systems, 24 et cetera.
25 But all the other type stuff from the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
96 1 maintenance, training, whatever it is, evaluating that 2 data, you don't need regulations for that.
3 Why over regulate when you don't have to?
4 Focus the regulations on the areas where it impacts 5 the safety of the plant.
6 And I haven't heard that through, I'm just 7 trying to echo a little bit of Greg's comment here is, 8 slow down.
9 There was so much baggage associated with 10 digital INC systems, and how you evaluate them. It 11 took years.
12 That's why the roadmap became all of a 13 sudden, how in the world do we tell what's applicable 14 to what.
15 And actually brought all the pieces 16 together where people could see what needs to be done, 17 and the Reg Guides have now been refined pretty much.
18 But it's taken a long time. I think we're talking 19 decades long time.
20 So here, to me, you all have the 21 opportunity to just put a rice, you know, a line 22 around certain things and say, hey, look, stay out of 23 these. You can do what you want every place else.
24 You're in your business. You're trying to 25 maintain efficiency in your training systems, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
97 1 maintenance evaluations, the data you absorb.
2 What you do with it, how you evaluate it, 3 go do it and don't look for regulation there.
4 I don't know why you can't take that step 5 forward, get an agreement with the U.S. and how to 6 get, so our industry can get on and utilize it where 7 it is known to be non-safety, non-safety critical.
8 I'll just stop there for the next slide so 9 I can do it again.
10 MR. DENNIS: I very much appreciate the 11 topic that you brought up and this, this was mentioned 12 in the workshop on the aligned, or crossing a line.
13 And in our working group, we've discussed 14 this crossing a line you know, thought process as 15 well.
16 The recognition we have right now is, we 17 don't know where that line is at this point. And part 18 of the gap analysis is to figure that out.
19 But at our front line individuals, the 20 inspectors and the regents, have been very attuned to 21 industry applications such as the corrective action 22 process analyzer.
23 Those areas where we recognize that 24 industry can deploy this, and use it, and, and 25 business efficiencies and process areas, to your point NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
98 1 of not over regulating.
2 MEMBER BROWN: Or not regulating.
3 MR. DENNIS: And not regulating. And 4 that's why there's a key distinction in every slide 5 we've presented, or we brought up says, NRC regulated 6 activities.
7 And, we are trying to be very mindful that 8 we are looking at where AI touches something we have 9 that is NRC regulated.
10 And other areas where it can be utilized 11 currently today, keeping, we're keeping boots on the 12 ground through the inspectors and the regents, to 13 maintain awareness of where those areas are being 14 used.
15 Because the industry is and has said, that 16 they're using this technology in an early deployment 17 phase to learn how to use it in areas where it can 18 gain true value for them.
19 And there may be a future where it then 20 does go into that NRC regulated activities space. And 21 so we are preparing as Anthony mentioned, our current 22 state spaces as we, as we understand it today, there 23 have not been any applications that have come to us in 24 an NRC regulated activity, for our review.
25 From what we've heard, we are aware that NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
99 1 there may be a couple areas, very targeted use cases, 2 where that may happen in the next couple years.
3 Those are the ones where we're investing 4 our effort right now, to try to focus on the gap 5 analysis so that we can understand truly is what 6 guidance is necessary, how would we evaluate it, and 7 make a finding on that particular thing.
8 And to the digital INC point, I've 9 mentioned this a couple times in other venues, that 10 this as mentioned, was this is just the latest thing.
11 We've had advance reactors. We've had 12 digital INC. We've learned from those things and 13 we're trying, our hope is we're trying to get out in 14 front of this a little bit so that we're prepared 15 should that eventuality come as we understand it now 16 and maybe three years.
17 MEMBER BROWN: Go ahead. If you needed a 18 queue.
19 MR. VALIAVEEDU: Oh, okay, I'm going to 20 move on to the next slide then. Thanks, Charlie.
21 Trying to keep up with our timeline here.
22 For slide 36, we are maintaining our domestic 23 partnerships, specifically with a NIST RMF, as well as 24 the LWRS program.
25 We are currently drafting together a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
100 1 institution plan to engage with academic institutions, 2 that we hope to complete by the end of quarter two of 3 fiscal year 2024.
4 For 3.2 on international partnerships, 5 we're drafting currently the CANUKUS paper that we 6 hope to get out by spring of 2024.
7 And, maintain our current ongoing 8 bilateral engagements with Canada, U.K., Germany, et 9 cetera.
10 For the last three, 3.2 echo, foxtrot and 11 golf, we are maintaining our participation in IAEA 12 technical meetings regarding the utilization of AI and 13 nuclear power plant safety.
14 And, the utilization of AI writ large 15 within the nuclear fuel cycle.
16 Specifically, we're also participating in 17 an IAEA project that will utilize artificial 18 intelligence to evaluate severe accident data.
19 For 3.3, this was mostly mentioned by Matt 20 Dennis but I just want to quickly go over this now.
21 We were able to complete our most recent workshop on 22 AI characteristics for regulatory considerations, as 23 well as we hope to maintain those workshops as they 24 prove to be fruitful in understanding where industry 25 is undertaking innovations in AI.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
101 1 And, to help better align our resourcing 2 for future fiscal years.
3 And we hope to also maintain our 4 participation external workshops, conferences, and 5 meetings, as these also have been able to provide us 6 with more information about where use of AI is going 7 towards, or heading towards.
8 Slide 37 is on cultivating an AI 9 proficient workforce. Wide skill training is not new, 10 and I was looking at some old photos of NRC history.
11 And, the photo in black and white is 12 actually a seminar on Lotus 1-2-3, which is a 13 spreadsheet software. I've never seen that software 14 before until that picture came up.
15 And to the right of that is the AP1000 16 simulator, the TTC, which again, showcases how wide 17 skill training is not, not a new phenomenon.
18 However, our active planning and whole of 19 agency approach when it comes to AI, compliments our 20 readiness for taking this challenge on.
21 With the potential for AI to be widely 22 used, the NRC has plans to develop the skills 23 necessary for evaluating any AI, incoming AI 24 applications.
25 The phrase that we utilize is, train, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
102 1 retain staff, and hire as we need to. We are ahead of 2 schedule luckily, comparative to the AI project plan 3 timeline.
4 And the staff is, because the staff is 5 developing a training guide to help develop 6 competencies for AI usage.
7 This guide splits the training between 8 data scientists, data analysts, as well as program 9 analysts, and provides basic training for AI ML 10 systems.
11 In addition to this, we were given direct 12 hiring authority for data scientists, as well.
13 For workforce planning, we, the staff has 14 engaged with OCHCO, which is the Office of the Chief 15 Human Capital Officer, on the development of a 16 competency model for AI related job functions.
17 And, this effort aligns with the recent 18 initiative from OPM, Office of Personnel Management, 19 develop a competency model for the whole federal 20 government.
21 The push by the chair's recent memo and 22 the White House's also recent executive order, have 23 only strengthened the team and the agency's resolve to 24 strengthen our skillset to be better prepared for AI 25 systems.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
103 1 Slide number 38 is another timeline. As 2 indicated by the purple slide where we are currently, 3 we are ahead of schedule for many of these activities.
4 4.1 is assessing the NRC's AI skills and 5 identify any gaps. We're currently developing a 6 competency model with OCHCO, to analyze what areas we 7 would need AI related job functions with.
8 4.2 is identify, develop and implement AI 9 training opportunities. We were able to put together 10 a draft qual plan to help staff develop and train up 11 on new data science skills, and AI skills.
12 4.3 is on recruitment, hiring and 13 retaining AI talent. In collaboration with OCHCO, we 14 plan on developing a working group to recruit AI 15 skills, and retain that expertise.
16 Goal number 5 is pursing use cases to 17 build an AI foundation within the, across the NRC.
18 The NRC's focus on internal usage of AI has been 19 exploring research and development of AI ML.
20 That may benefit for future regulatory 21 decisionmaking. One of these will be presented later 22 on today.
23 We plan to develop a AI foundation through 24 four areas. Pilot studies, safety insights, an AI 25 ecosystem, and future focused research.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
104 1 The pilot study is an area where we hope 2 to work with stakeholders, labs, and partners to 3 investigate security and safety of AI technologies.
4 Our engagement with stakeholders have 5 supported this concept. And previously, we've had 6 industry say at events that they are in support of 7 pilot studies, as well as regulatory sandboxing to 8 help navigate the regulatory landscape.
9 As emerging technologies will always 10 involve interaction between the regulator and 11 developer, we wish to go about this early so that we 12 are able to identify challenges within NRC review, as 13 well as build technical expertise.
14 As we are not the barrier to innovation 15 but instead we are the guardrails for safety and 16 security.
17 The second item here is on safety 18 insights. We wish to assess and survey what is out 19 there to evaluate AI systems for safety. And 20 incorporate those findings across the NRC.
21 The third box is on AI ecosystem. The 22 room is only dark when you don't have a flashlight.
23 So, there is value in the staff to get accustomed to 24 AI ML tools, to deconstruct AI.
25 By acquiring common data science tools and NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
105 1 deploying IDEs that could help support an AI 2 ecosystem, we could utilize training within goal 3 number 4 to help get a staff, to help staff understand 4 how AI could be utilized throughout the nuclear 5 industry, as well as any safety or security issues 6 that may come up.
7 MEMBER MARTIN: Question, and I'm sorry, 8 I can't help myself. Trying to get software that is 9 not like, already approved by IT is, is right, 10 impossible, right?
11 So how do you expect to even get through 12 NRC's own processes to get tools to train people? Do 13 you have to make exceptions? Obviously you've worked 14 with the Office of the Chief Information Officer.
15 I think you have your own obstacles just 16 to even get to the point where you can train people.
17 I mean, are exceptions made in the spirit of research 18 and training, that lets you get tools in there that 19 are typical, or being proposed?
20 Because invariably, you know, questions of 21 security associated with those tools have to come up.
22 MR. VALIAVEEDU: To first answer your 23 question, the staff currently has Anaconda and Python 24 as two main packages that we are allowed, or that we, 25 that have been vetted by OCIO.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
106 1 By in the first slide, or second slide, I 2 ensured that the word OCIO is incorporated with a 3 variety of program offices --
4 MEMBER MARTIN: I saw that.
5 MR. VALIAVEEDU: -- because we've been 6 able to work with, directly with the chief information 7 officer because AI is upcoming.
8 We've seen a lot of developments 9 throughout not even the nuclear sector, and our whole 10 of agency approach to this has been complementing 11 that.
12 So, it seems from our view, slightly 13 different than traditional software procurement. And 14 we are having ongoing talks to acquire and allow more 15 uses of like Python, and different library packages.
16 MEMBER MARTIN: But this would be 17 segregated, too, for the purpose as opposed to letting 18 everybody in you know, have access to it, and who 19 knows what.
20 I mean, correct, or am I wrong?
21 MR. VALIAVEEDU: Oh, for that, I think 22 Victor Hall, I think has direct engagement with the 23 CIO. So, he may be able to better answer your 24 question.
25 MR. HALL: Thanks, Anthony.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
107 1 So again, this is Victor Hall with the 2 Office of Nuclear Regulatory Research.
3 We do have strong partnerships with the 4 Office of Chief Information Officer. And I think back 5 to right before pandemic, some of the changes that 6 they made to be able to prepare us to be ready for a 7 changing world to be able to work remotely.
8 And, they knocked it out of the park.
9 It's thanks to them we were able to get our mission, 10 even get our mission done.
11 And, I think they're taking that same 12 mentality of being ready for what's coming, or 13 changing where we have new tools that we need to be 14 assured are safe, that are protected from the dangers 15 that are out there, whether cyber or other.
16 And, I think they're taking that same 17 approach to being ready to be able to give us the 18 tools to be ready for what's coming.
19 So, they're working hand-in-hand with us.
20 Obviously, has the table want all the bells and 21 whistles.
22 We want the toys to play with, and they 23 have to be able to say wait a second, let's make sure 24 they're safe.
25 And they're doing that partnerships across NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
108 1 the government. So, you know, better to have approval 2 to make sure that something is usable and safe, and 3 that our systems is not going to get in and cause a 4 greater problems is clearly high on their minds.
5 But again, I think they're taking a very 6 positive and collaborative approach to making sure 7 that we're ready for being able to use these safely.
8 And so again, I'm happy that we have them 9 as partners. And I think when you look at the chair's 10 tasking, which wants us to lean forward but 11 responsibly, I think that gives us good momentum to 12 be, to have those tools to be ready and recognizing 13 that in government, there are going to be 14 restrictions, period.
15 MEMBER MARTIN: I thought it was a full 16 committee a time or two where we got the message that 17 you know, things like personally things like ChatGPT 18 would not be used.
19 Obviously concerns about the control of 20 proprietary or classified information, and all that.
21 You know, you just see every other day that we're 22 getting a patch for our operating system for some 23 security, you know, and I just think well this, like, 24 you know, stuff like that.
25 So anyway, that's a little cynicism NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
109 1 associated with the question, but it's obviously a 2 real bureaucratic challenge to beat some of these 3 goals because if you don't really have the best tools, 4 then are you really doing the best job training.
5 But you have other people that you know, 6 decide your fate a little bit on the decisions that 7 you're making, who have total control over it.
8 MR. VALIAVEEDU: Thank you again for your 9 comment.
10 The last box here is on research. We hope 11 to continue to invest in AI research through existing 12 avenues, as well as universities.
13 This is through supporting our university 14 research grants, and as well as with this, we hope to 15 continue with our future focus research program.
16 This program has helped build NRC 17 knowledge in emerging, and significant technologies.
18 More specifically, in the last two years 19 we were able to fund six FFRs, one of which will be 20 presented later on today by NRC staff.
21 Two more slides. Slide 40 regarding the 22 timeline of these cases. For 5.1 is on the proof of 23 concept in pilot studies.
24 As previously explained, the objective 25 here is to engage with the industry to identify the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
110 1 potential pilot study and proof of concept test cases.
2 They'll help the NRC staff gain expertise 3 for future regulatory reviews.
4 5.2 is to develop and maintain an AI 5 ecosystem. As seen with the purple dashed line, we're 6 currently working on developing and maintaining an 7 IDE, as well as identifying and assessing, and 8 acquiring AI tools.
9 5.3 is on surveying of AI tools and 10 methods for safety evaluations. We hope to conduct a 11 survey of what's currently out there to evaluate AI 12 systems, by the end of quarter two of fiscal year 13 2024.
14 And then implement these findings by 15 fiscal year 2025.
16 MEMBER HALNON: Now Anthony, is there 17 where 50.59 will come in, and how to figure out how to 18 do an evaluation for modification to the plant?
19 MR. VALIAVEEDU: It's slightly different.
20 The survey of tools and methods is more of what sort 21 of systems and algorithms are out there so that 22 individuals or stakeholders can utilize those two, to 23 evaluate their own systems.
24 MEMBER HALNON: So the guidance I'm 25 thinking about is if somebody wants to implement it, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
111 1 they're going to have to go through a lot of pre-2 application discussions with the NRC, in addition to 3 having a process to be able to facilitate that.
4 Which would be a 50.90 submittal plus a 5 50.59 and a no-hazards. All that stuff. Where is 6 that all going to down the road once you get all this 7 learning and start putting guidance, you know, pen to 8 paper and making guidance documents?
9 Because it seems like it takes a long time 10 to get, you know, alignment with the industry on some 11 of these things and I just wanted to make sure that 12 this 27 date is, is still feasible.
13 MR. VALIAVEEDU: For the guidance 14 development, we hope to implement that as guidance 15 development for stakeholders to utilize.
16 However, stakeholders are able to utilize 17 what is out there right now for their 50.59 18 application, if they wish.
19 MEMBER HALNON: Yes, but when you start 20 getting into those fine questions of increases and 21 decreases of consequence and all that kind of stuff, 22 it gets difficult.
23 We ran into it with the digital INC, and 24 it took a while to get that straightened out. And 25 then maybe that is this building block for where we're NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
112 1 going on this already.
2 But and just, it just seems like that 3 takes, I mean again, back to my impatience. That 4 takes a long time to get an alignment, you know, not, 5 it doesn't happen in a quarter or two quarters.
6 Sometimes it takes years.
7 And, we're not too many years away from 8 when we want to be able to be ready for this. So, I 9 was, I wouldn't put that aside.
10 I'd make sure that you know, at least 11 those conversations are being had so that we know if 12 the existing guidance can be applied, and actually 13 work.
14 MR. VALIAVEEDU: So to the workshop, the 15 summer, we've kind of hit a cadence where we think 16 annual workshops are the, what we're going to plan on.
17 But we remain flexible to have the entire 18 workshop in there to tie in with it. But the next 19 workshop is really intended to go to this point on 20 guidance, because we plan to finish that regulatory 21 gap analysis.
22 And, the early look is I think our 23 regulation is flexible enough to adapt to AI, but the 24 guidance might be lacking.
25 MEMBER HALNON: Okay.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
113 1 MR. VALIAVEEDU: In significant areas.
2 MEMBER HALNON: It's on your --
3 (Simultaneous speaking.)
4 MR. VALIAVEEDU: It is.
5 MEMBER HALNON: I mean, you know --
6 MR. VALIAVEEDU: Very, it is very front 7 and center.
8 MEMBER HALNON: Okay, good, that's what I 9 hoped. I hope that we can get that, get that not be 10 a hurdle to get over.
11 MR. VALIAVEEDU: Yes.
12 MEMBER HALNON: Seeing that we're looking 13 at it years in advance here, so thanks.
14 MR. VALIAVEEDU: Right.
15 DR. SCHULTZ: Along those lines, or maybe 16 it's parallel to it. The regulatory sandbox, is that 17 part of the upcoming workshop as well?
18 In other words, I'm presuming you're 19 looking for what is going to be included in the 20 sandbox, and what's outside of it in terms of 21 application than what inside the sandbox needs 22 regulatory attention?
23 MR. VALIAVEEDU: For the regulatory 24 sandboxing, we are hoping to one, identify a test 25 study that may or may not have a direct implication NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
114 1 for reactor use utilization.
2 And then, go about a potential open avenue 3 to identify any challenges within current area 4 existing regulations.
5 That may or may not have answered your 6 question. If it didn't, I can go back.
7 DR. SCHULTZ: No, it did do it. It's 8 fine.
9 MR. VALIAVEEDU: Okay.
10 DR. SCHULTZ: But eventually you're going 11 to be including decisionmaking associated with what 12 will be included in terms of the regulation, and what 13 will not?
14 What happened? I mean, you talked about 15 many applications --
16 (Simultaneous speaking.)
17 MR. VALIAVEEDU: Yes.
18 DR. SCHULTZ: -- in your workshops, so 19 which industry is doing nothing. And in several of 20 those, regulatory attention will not be required.
21 MR. VALIAVEEDU: Yes.
22 DR. SCHULTZ: The work can continue, and 23 findings will be useful to the industry. Regulation 24 would not be required in terms of the way in which 25 it's done, or the results that are obtained.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
115 1 Would that, aren't those findings that you 2 want to, you want to develop sooner than later?
3 MR. VALIAVEEDU: Yes, that's why we're 4 maintaining our ongoing collaboration with different 5 stakeholders like EPRY, as well, that have, that we 6 are able to obtain those lessons learned from their 7 utilization of AI ML technologies.
8 The regulatory sandboxing is more of 9 understanding the, it helps support the regulatory 10 framework that exists right now, to ensure that the 11 staff ourselves, are ready for any potential 12 application.
13 DR. SCHULTZ: Great, I appreciate the 14 detail that you've gone through with regard to staff 15 training and implementation.
16 All right, to me there's more important, 17 more important than the hiring of new people that 18 understand AI, is training the people that are already 19 here and know a lot about regulation, to utilize AI 20 and the process.
21 Thank you.
22 MR. VALIAVEEDU: There was a phrase 23 mentioned earlier in a previous meeting where you 24 could get a really good data scientist, or you could 25 get an NRC individual who's extremely experienced NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
116 1 nuclear regulator, and then train them with data 2 science tools.
3 MEMBER BROWN: One other input just to put 4 it on the plate to think about, as we introduce 5 software based systems for digital INC, it was always 6 the concern about common cause failures. And 7 therefore, the issue of diversity areas.
8 How do you handle that and ensure that 9 that failure doesn't propagate lock up of the systems, 10 interrupt driven systems, which all AI is going to be 11 interrupt driven because it's going to be evaluating 12 data coming in all over the place.
13 That means it could get confused, and if 14 you, I'm thinking of downstream now outside of the who 15 cares realm.
16 Does that mean now we have to have the 17 thought process of diversity, defense in depth, in 18 terms of the application of AI into any other safety 19 or plant control, or even non-safety related but plant 20 control systems that are just out controlling stuff?
21 How do we do that? Do I have to have 22 competing AI algorithms making the data and then 23 comparing those, and then making a, or do I have to 24 have three sets because I need to have them both?
25 It just, the whole idea of now all of a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
117 1 sudden accepting the fact that it's AI, I don't need 2 anything like that to make sure it's really working 3 properly.
4 Is it one more technical issue that would 5 have to be addressed when you get into the world of 6 plant controls, safety systems, and even what I call, 7 I wouldn't call it, not all plant systems are 8 obviously safety related but you need them to operate 9 the plant. You have to do something if they don't 10 work.
11 So that is another very, very difficult.
12 It was hard enough in the regular software world where 13 you used different devices, you use alternate 14 software, watch dog timers, all kinds of, and 15 susceptibility to cyber-attacks, intrusive. How do 16 you communicate date that's not protected by an air 17 gap? So that's just one more thing you need to throw 18 into the hopper in terms of how you apply this 19 downstream.
20 Because you're going to run into the exact 21 same issue we've been dealing with for a decade or 22 more, a couple of decades in terms of incorporating it 23 into the systems.
24 That's one more way to be very, very, very 25 cautious.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
118 1 MR. VALIAVEEDU: Thank you, that is one 2 are that we are looking at, especially with the gap 3 analysis, so I appreciate it.
4 The final item here is to facilitate and 5 invest in research. We hope to maintain our ongoing 6 university research grants, as well as maintain our 7 current FFRs with a specialized focus with, hopefully 8 promoting AI research within the NRC.
9 The final slide here, if we do our work 10 right, no one will remember us. The NRC is committed 11 to the safety and security of the public and the 12 environment.
13 The hope of the AI team is to pave the way 14 to ensure that the NRC puts its best foot forward, in 15 future applications.
16 Our high stake standard for safety remains 17 unchanged no matter what the technology is. And, we 18 are working towards ensuring that we have the staff 19 with the knowledge, skills, and the ability to 20 effectively regulate these new technologies.
21 I would just take some time to highlight 22 our next steps, which includes in the spring of 2024 23 we're going to be publishing the CANUKUS paper, AI 24 principles paper.
25 We hope to publish the AI regulatory gap NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
119 1 analysis as well as that time.
2 In the March timeframe, we'll be hosting 3 an AI technical session at the RIC that Matt Dennis 4 has highlighted. As well as host an IAEA AI technical 5 meeting at headquarters the following week.
6 We hope to update the AI project plan with 7 revision 1 in the fall of 2024, with our revised 8 timelines according to what we expect through our 9 engagement.
10 And, we will always continue our public 11 workshops and stakeholder engagement as they've shown 12 to be fruitful in understanding what will be ongoing.
13 Thank you.
14 MEMBER BROWN: I didn't mean to interrupt 15 you, I just want when you're done.
16 MR. VALIAVEEDU: Oh, I was going to say 17 just thank you again for giving me the opportunity to 18 speak, and open it to more questions.
19 MEMBER BROWN: I just wanted to make 20 papers and things that you had issued in the gap 21 analysis, if you, and the emphasis and I tried to 22 understand.
23 So I'm not a designer obviously like you 24 guys are too smart for me from that standpoint. And 25 I went to try to figure out something. There was a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
120 1 paper in one of the magazines that I read, one of the 2 publications.
3 And it was so replete with jargon that by 4 the time I finished, I had not a clue as to what they 5 were saying when they got to the end, and they had 6 some conclusions.
7 I would just encourage these papers are 8 going to be relevant to other folks other than you, 9 understanding where you're going to go.
10 And if it is steeped in deep learning, 11 machine learning jargon, that's not going to work and 12 the resistance is going to be strong.
13 So, it would be good if you could put it 14 in every day English for people who are technically 15 oriented, but not fully ensconced in the jargon of the 16 AI world.
17 So, you know, reduce it to English, in 18 other words.
19 MEMBER HALNON: Anthony, in your, is it 20 your intent, your goal, your aspiration, your hope or 21 whatever, to have the gap analysis ready for the RIC 22 session? Or at least you know in --
23 (Simultaneous speaking.)
24 MR. VALIAVEEDU: We expect it to be done 25 in the spring 2024 based off of initial timelines.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
121 1 However, as you're probably aware, the NRC is pretty 2 big with its regulation.
3 MEMBER HALNON: Yes, yes. It just seems 4 like we'd miss a really good opportunity if we don't 5 have at least, at least some of the findings to 6 discuss at the technical meeting of the RIC.
7 So I know you know, spring is not 8 necessarily March, but you know, it seems like that 9 would be a great opportunity to at least be able to 10 present the findings.
11 So, that's my comment. Thanks.
12 MR. VALIAVEEDU: I agree.
13 CHAIR BIER: Okay, thank you, Anthony, for 14 actually getting us caught back up on time.
15 So, and thanks to all the presenters for 16 a good discussion, and for your patience with all the 17 interruptions and questions.
18 So I think at this point, it's time for a 19 break and we will reconvene at 11:15. One more 20 presentation before lunch. Thank you.
21 (Whereupon, the above-entitled matter went 22 off the record at 10:58 a.m. and resumed at 11:15 23 a.m.)
24 CHAIR BIER: For those online, we are 25 going to get started in just a minute or two after NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
122 1 everybody takes their seats.
2 It looks like we're missing a few who are 3 taking a slightly longer break, but in the interest of 4 timeliness, I think we will go ahead and get started 5 on the next session.
6 So I'm happy to introduce Jim Chang from 7 Research, who has what sounds like a very interesting 8 presentation on using machine learning for inspection 9 planning.
10 MR. CHANG: Thank you. My presentation 11 goes into the topic on regulating the AI to use the AI 12 for NRC's operation. And our focus is that for this 13 implementation is informed inspection plan.
14 My presentations are straightforward at 15 the motivation and end at talk about what we do and 16 then what data we use, and then this observation 17 underneath that that we obtain from this project.
18 The motivation was COVID-19 that disrupted 19 NRC's inspection plan. NRC did not send a inspector 20 to the site regular (audio interference). So that's 21 under NRC's risk-informed inspection that we can have 22 some system that can identify what's a priority of the 23 inspection. That will be very beneficial to enhance 24 NRC's risk-informed actions.
25 And I also read the AI machine book that, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
123 1 one case is referenced in that was a Netflix success 2 story. Used unsupervised machine learning to analyze 3 the movie watched by its subscribers. And then from 4 there that it helped identify the hidden pattern of 5 the code clusters.
6 And using that information will be helping 7 the Netflix to better inform that recommend a movie 8 for its customers.
9 So then I was thinking about these two 10 pieces of information together. NRC here, we have a 11 lot of nuclear power plant performance C suite that 12 many have documented, licenses and reports or 13 inspection finding. These things that they consider 14 as this history of this plant's performance.
15 Can we use the unsupervised machine 16 learning, it can bring the information together and 17 then that's identify a hidden pattern. I call this a 18 safety cost, later I will explain what safety cost 19 mean.
20 So this objection was try to perform a 21 feasibility study simply that looking at what the data 22 we have here and then the snapshot technology we have 23 this stage. What's combination of them that how good 24 they are, that to achieve this purpose, informing the 25 inspection planning.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
124 1 So I called this the hidden pattern as a 2 safety cost. Safety cost I define as the failure 3 mode, failure causes of this structure, system, 4 components, and their failure that has consequence to 5 the nuclear power plant safety.
6 So this combination of this information 7 together usually call this the safety cost. Try to 8 identify using unsupervised machine learning to 9 identify this safety cost. And we had a benchmark 10 that the teams tried to achieve. It was in the NRC's 11 reactor oversight program that periodically that 12 publish the operating experience communication.
13 And this was a communication published in 14 November of last year that, it identified five power 15 outage events impacting security system operation. So 16 that's consequence of power issue impact the security 17 system operation.
18 And then there's SSC and here is the 19 primary and the backup power tried to reach the 20 security system. Failure mode just simply not 21 providing electricity.
22 This is a communication that's also, I 23 listed it by instance, by operating experience. It's 24 identified at 2022 has two events, '21 has three 25 events. But the 2021 three events, all them consider NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
125 1 as the random failure, power failure.
2 But the 2022 two events, all that involved 3 is human failure. One is the maintenance, doing the 4 maintenance and all the system repair that caused the 5 outage.
6 So with this information that's from the 7 inspection finding, that's the original communication 8 suggested our inspector, when they performed the 9 inspection procedure that's related to equipment 10 performance testing and maintenance that focus on the 11 human impacts on power supply.
12 So this provided this zooming in the focus 13 that to me is risk-informed information to help our 14 inspector based on the past event to help our 15 inspector when they do this general inspection in this 16 area that's focused on -- that's a cause related to 17 the operating experience we observed in the past.
18 So that was the things that I tried to 19 achieve, see that can we use the unsupervised machine 20 learning to help identify these things. The approach 21 is that I got a funding from the Office of Research 22 Future Focus of Research funding. And then 23 established a commercial project contract to the AI 24 company that's SphereOI.
25 In addition to this, NRC also formed a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
126 1 team that's we have machine learning experts that Trey 2 Hathaway, that's sitting here, that he was in my team.
3 And then I also has NRR staff from Reactor Oversight 4 Program, that Jason Carneal. That's in my team.
5 I am not a machine learning expert. I am 6 not this Reactor Oversight Program pilot. We just 7 bring the team together to work with this contractor 8 to perform this project.
9 The task, two tasks I identified for this 10 thing, the first thing that's we don't know what's the 11 current state of the AI. Just simply try to get a 12 glance of the what's the landscape there.
13 So the first task was try to understand 14 the -- evaluate these are big plant companies. Their 15 AI platform that's a high-level version that to find 16 out which one may be best for this whole purpose.
17 The intention was try to use this pre-18 trained algorithm as much as possible instead of NRC 19 does put in a lot of effort try to develop algorithm.
20 And then the second task was a lack of 21 platform that to identify these safety costs. That's 22 a issue with two task. The company that, really quick 23 pace, we have weekly meeting and that work the project 24 was completed. We did it in four months.
25 So the task one here that the contractor NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
127 1 identified the topics for doing this type of test 2 what's the topical like, in this process, and then 3 evaluating give a score, weighting, wright and score 4 of this four platform.
5 Come to the end, that's Azure, Microsoft 6 Azure, and Amazon's AWS was ranking the higher. But 7 this doesn't really help much because they come to the 8 task to -- one thing important was a notebook 9 integration.
10 And there was a notebook that can access 11 these platforms, algorithm library so that's -- it's 12 not -- this notebook is independent from all these 13 platforms. And that was the Jupyter Notebook was used 14 in the task two analysis.
15 So come to the end that this evaluation 16 doesn't really affect the decision on choosing which 17 platform to go. Go to using the Jupyter Notebook is 18 a free software that downloaded. We need this kind of 19 open-source library to perform the functions.
20 To perform the task two here that the 21 contractor develop to bring in the test that's a 22 inspection and then the former test that's item one 23 there. And go through the series of process, the 24 components of process in this information.
25 And then come to the end that we NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
128 1 identified these safety costs. And then that item 2 three that is the visualization of one of the costs.
3 For all these components, the components 4 in this pipeline, there's multiple algorithms can do 5 the function. So the -- come to the end the -- what 6 we do here is try to trial and error the different 7 algorithms and then try to evaluate, see which 8 algorithm has better performance.
9 And come to then end that I identified 10 this optimal combination for this pipeline performance 11 that can take data from the front end of the text and 12 then come to the end that identify this cost and 13 represent this safety cost.
14 So this diagram talk about the things that 15 the contractor tried. On the top is this pipeline 16 component, on the end. The first element taking the 17 text, completion of this inspection finding. In that 18 the contractor tried 15 different, I'll say that 15 19 algorithms to process these original information.
20 And then come to the end, it selected 21 three of them that are better performance. And then 22 leading to the next components. Next component has 23 three different, five different -- five different 24 algorithms. And then come to the end, select one.
25 So you see this, that's a lot of trial.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
129 1 And for the first one text this portion here, it's we 2 took eight different sample, eight inspection, eight 3 inspection. And then try this 15 combination. Come 4 to the end, select these three. So that you see that 5 was try and see that how --
6 MEMBER MARTIN: Question. I'm trying to 7 understand what's the information that's being fed 8 here. What's the specifics. And the maybe use an 9 example.
10 MR. CHANG: Yes. That's in my next slide.
11 MEMBER MARTIN: Okay.
12 MR. CHANG: So input information, that 13 original was trying to that we -- NRC has inspection 14 reports that about 20,000 inspection reports from year 15 2000 that's publicly available on the website. That's 16 was -- inspection report was the information.
17 But through a process, so I learned that 18 our key process NRC maintain this database that's 19 excel database that has all these. This is not 20 inspection, it's inspection finding. Inspection 21 finding data is something come to the like -- more 22 green, the type of inspection finding there.
23 But come to the end, that many of these 24 will be identified as green finding. But these are 25 the inspection findings that are keeping in this NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
130 1 database that's from the study from the year 1998 2 about 15,000 unique records.
3 In this database that column F was item 4 introduction. That was a decision about this 5 inspection finding.
6 So that's what's come to the end, so while 7 since we have this one, I don't want to focus on our 8 resource. To focus on what we want to achieve so that 9 we simply take in these as original, the discretion 10 here in the item introduction. This column has the 11 input information.
12 This input information that averaging as 13 1,649 words. And minimum is 42 words. Maximum, 7670 14 words. So that's the range of that expression there.
15 We took the discretion (audio 16 interference) that's all the information we need. But 17 it was a limit in there, too, these sentence 18 transformation models limitation that it come to the 19 -- reached its sudden capacity it will truncate. It 20 doesn't take in information anymore.
21 So if we use the full text that's a long 22 text, that's a data on the text it will simply just 23 dismiss that because of limitation of this -- sentence 24 transformer model limitation.
25 So that was the contractor. Okay, what NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
131 1 about let's try some other algorithm to take these 2 full text as the input, and then that this algorithm 3 can do the summary of this text. And then that 4 generates as the, I call it condensed summary. And 5 then input in to the pipeline. That's the approach we 6 did.
7 So except the full text, we tried 14, the 8 contractor tried the 14 condensed summaries in here 9 that it divide into the three category. One is a 10 summary technique algorithm. That's a try 70 kind of 11 algorithm in the summary technique.
12 And then Q&A three key phrase extraction 13 to try four (phonetic). And that -- what we take.
14 And then that some of the AI compound, AI algorithm.
15 That's also allow us to provide some inputs that cause 16 semi-unsupervised machine learning.
17 In there, they're taking out the things 18 that do have a focus on nuclear safety. These AI 19 options, they are trend from the open website that 20 Wikipedia, social media. It doesn't have a specific 21 focus on the nuclear safety. But some of them, they 22 allow us to bring into the input that what are the 23 things that we need to pay more attention. On the --
24 so from the NRC provided 1,004 acronym like MSIV, 25 these type acronym.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
132 1 And then also one is 400 common failure 2 mode, like a inoperable misalignment corrosion. These 3 are word that we were interested in on the special 4 failure mode. That we really want to, knowing that's 5 -- not just feed it a general word. So that provide 6 us more useful information.
7 And in addition that we acquired 269 8 NUREGs and 195 research information letter. This 9 technical report was the contractor want to use to see 10 the coherence of this word, term that appear in the 11 text and then compare it. But it's for further 12 information. But they're not really helping the text 13 too (audio interference) check, it will function.
14 And then also the stop word removal that 15 in addition to this general remove the stop word, this 16 type of word, we also provide that contractor that 17 also look into the outputs and then seeing whether 18 that term that we see that we need to remove them.
19 That they consist safety system reactor, these happen 20 to open them.
21 If we don't remove the stop word, that the 22 -- somehow the group that focusing on these terms, 23 that's not a one we want.
24 Showing the example, that's one, at least 25 one process. On the left-hand side for column here, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
133 1 you see a lot of these operating company, Entergy, 2 Exelon. And that was because we did not remove the 3 these words, these terms as a stop word.
4 And then after we saw the result, well, no 5 this making the -- this not the things we wanted to 6 focus on. The process, focus on the company instead 7 of focus on the safety -- the system, structure system 8 components.
9 So that's how we work to remove this word.
10 And then on the right-hand side, have to remove this 11 word that's on the right-hand side, Fort Bragg 12 (phonetic), showing these customs, forming that's 13 these become more like trip and auxiliary feedwater, 14 these are the kind of level that we have more interest 15 in. That's a kind of stop word removal.
16 So come to the end that goes through this 17 process, identified hundreds of these safety costs.
18 And then it's a long list here. It's a part, and I 19 just show you some -- an example here.
20 These are costs is represented by word 21 cloud or bag-of-words. Here, that's on the Excel 22 spreadsheet here. The topic -- the next minus one is 23 -- has 5,382 inspection findings. These are the 24 inspection findings, could not group into any cluster.
25 So these are kind of, we called outliers costs.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
134 1 And top on Deal One (phonetic), that's a 2 deal, the topic deal that has 927 inspection findings.
3 And these are the words kind of forming word bag, word 4 cloud, that's describe what the cost are about. And 5 et cetera.
6 So these are the way that it -- to the end 7 of Type Nine, these are the table that was generated.
8 And then that's for each role that's a work and going 9 to that's a -- what's this then, 927 inspection 10 findings. That information can be tracked, if we 11 want.
12 And then, after that, was this just 13 showing the three different input information 14 technology coming in, that they come to the end, that 15 what this same information but forming the 16 synchronization of different clusters. It's all 17 related to the RCIC system, things associated with 18 that components.
19 At the end of this project that we 20 fortunate has a operating experience computation.
21 This time is about a safety security system. The 22 15,000 inspection findings I mentioned here, all of 23 them are safety system, not security inspection. So 24 security system inspection is not within the scope 25 here.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
135 1 But in this original communication it 2 identified why events relate to improper calibration 3 and the maintenance of the radiation monitor and dose 4 assessment equipments.
5 One company has two events. They said 6 well, since we are already identified these costs, can 7 we go back to the things we identified, seeing that's 8 how this five events was clustered.
9 And we found out that one of these, I 10 think it's the top one, 2022 event, was not in this 11 original -- inspection finding data. What we can find 12 four of these -- four of these events.
13 And this was the results that the feedback 14 led to well, these operating expense was identified in 15 this exercise. So that's from the full item 16 introduction. That means that we took the summary of 17 data and looked into this pipeline.
18 In summary, this column was, using the one 19 of these summary technique. But generally the 20 condensed summary and put into this pipeline and key 21 phrases, technique.
22 So that all these are four events, four 23 operation, you know, was in one of these safety costs 24 identified. None of them was put into this outlier 25 bin. But one day we saw that well, summary report NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
136 1 that has at least three of events identified in the 2 cluster number one. Well, sounds good, but the number 3 one has more than 900 inspection findings.
4 So that's still not practical that work 5 identified all these things. And that in terms of 6 that we NRC operators people need to, you need to 7 squeeze through this 900 events to identify what are 8 things are that are maybe not working.
9 This current stage is still not come to 10 the -- demonstrate some success, but come to the level 11 of the data, we say what it's used for.
12 But I want to say these 900 event, that's 13 we are talking about inspection findings dated from 14 the 1998 to 2022. So that was maybe that using the 15 dates that we focused on the most recent data maybe 16 give us some more focus. But we didn't go to more 17 analyses for that.
18 It was because the future focus of 19 research is a small project. It's for research for 20 you to identify information. And then it really find 21 out some that information since indicates some 22 potential that it become a seed (phonetic) process --
23 seed project that action or the more formal way of 24 doing the research development these areas.
25 And so that was the way we concluded the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
137 1 is research project. The project here that does 2 identify this summary technique was very useful, the 3 R piece (phonetic) that was there, while that was 4 useful, that summary that's taking these original 5 inspection summary, and then that's provide condensed 6 version of summary. That was a useful, good use of 7 workflow purpose in the sense that can use this 8 technique to provide summary and then NRR staff can 9 see a view of the condensed summary that's to reduce 10 work.
11 The second bullet about is using this 12 based on the way that we do. Certainly there's a lot 13 of things that we can improve, including the stop word 14 or trying some technique. At that time we saw it as 15 time-consuming. We don't want to go forward moving 16 that, try to optimize our future focus research.
17 And if that's we have additional funding, 18 that's we may want to spend it to fine-tuning that may 19 be able to refine the results. But whether that 20 refine the result will it come to the -- become a 21 practical skill, I don't know. So that's why --
22 making the conclusion that what, based on the result 23 we saw, it has potential, but it's not conclusive that 24 for practical application.
25 MEMBER MARTIN: A comment on that. About NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
138 1 six years ago, I was doing -- this just shows you what 2 I do in my free time. But so I resonate with what 3 you've done here. I've used the algorithm for the 4 parsing, the documents or texts. And the association 5 -- the associating frequency and the presentation of 6 that information.
7 I did this with the water reactor 8 evaluation model document, you know, which was 9 published in the 70s. Again, for fun. I have it here 10 on my screen.
11 But so this, what you're doing resonates.
12 I will say, though, in the presentation, and I was 13 doing this because I wrote a blog and about water 14 reactor evaluation model. And I wanted to highlight 15 the kind of things that were important, right.
16 And I saw that algorithm, and of course in 17 my mind, I had an idea of what should be important.
18 And then used the algorithms and parsed the 19 information. And of course I had the problem with the 20 stopper, you had as well.
21 And then I probably spent the next six 22 hours2.546296e-4 days <br />0.00611 hours <br />3.637566e-5 weeks <br />8.371e-6 months <br /> trying to find the right set of words to 23 eliminate to get what I want out of it.
24 So what that means is there's a huge 25 amount of uncertainty associated with that, with the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
139 1 process. But you know, in this business, you know, 2 we're addressing safety issues.
3 We know there's a huge amount of 4 uncertainty. Of course we talk about it in the 5 context of our probability risk assessments and such.
6 And uncertainties on order -- on the order, an order 7 of magnitude are pretty normal.
8 Do you see some synergy with, you know, 9 methods like this? I mean, because it's incorporating 10 a natural language translation of sorts. And it's I 11 would say corroborative information to, you know, more 12 quantitative risk analysis. Is there synergy, have 13 you thought about synergy in that realm?
14 You know, going back to our earlier 15 presentation about being, you know, Matt's comment 16 about the, you know, being the best version of 17 yourselves. You know, using the tools that we have 18 today to do a better job and develop more confident.
19 I know this was a small project. It's 20 kind of fun to listen to here. But to take it in all 21 seriousness and is there an opportunity? Do you see 22 opportunity? Will you go farther with this? I mean, 23 I know it was at the discretion of the agency for you 24 to do this for a project.
25 But where does it go from here? And I see NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
140 1 one opportunity that might be interesting, if not 2 useful.
3 MR. CARNEAL: Can everyone hear me? I'm 4 Jason Carneal, I'm in the Operating Experience Branch.
5 And I was working with James on this project.
6 And as with pretty much every start that 7 you get in this area, we all ran into the same 8 problems with those stop words where yes, you're 9 probably putting in some bias there with what you 10 think that you think the output should like, should 11 look like.
12 What we were doing in this project was 13 trying to just give it a minimum baseline to fake out 14 the general Wikipedia-style stop words or the ways 15 that the algorithms were trained with general 16 language, and give it just kind of a little bit of a 17 leg up for the business that we do at NRC and see what 18 we could see in these safety clusters.
19 And it was a small effort. Of course we 20 could optimize that in the future. The power that I 21 see with the safety clusters and identifying those 22 unnoticed trends that the human eye can't see, in my 23 group we're working with OpE documents.
24 We have about 100,000 documents in our 25 store, and that number of documents expands by 1000 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
141 1 documents a year. That's just the NRC documents that 2 have no structured information. So it's all free 3 text. You have to have some way to apply some level 4 of grouping to that if I'm to hope to find some kind 5 of a trend in that document set.
6 We also have about 200,000 industry 7 documents where we have those texts. So the power 8 here that I see in the future is particularly for 9 those trends that are hidden, where it's not what we 10 think we would see in the trend.
11 So the top five trends, if we went through 12 James's list, it's kind of what we'd expect for safety 13 cultures. The safety clusters. When I looked down 14 into Items 10-15, there's something interesting here.
15 I've never associated those words in my mind.
16 MEMBER MARTIN: Right, and I'd also say 17 that statistically speaking, when you start playing 18 with the uncertainties, those, you know, numbers 5-15 19 will change. And at some point, they're worthless, 20 right. But statistical method, that's the nature.
21 Now, if you get consistency with the 22 variability of softwares or whatever the random number 23 (audio interference) there, that those top five are 24 there time and time again, there's probably something 25 to it.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
142 1 But yeah, I would certainly not look 2 beyond. You know, when things start changing, then 3 it's not (audio interference).
4 MR. CARNEAL: And just one other thing, 5 that as far as future work with these, being able to 6 group these with unsupervised learning, the way our 7 program currently operates, we're relying on four 8 people assigned to each region. They're looking at 9 the reports as they come in.
10 So if we're going to identify a trend, 11 it's usually knowledge of that personnel over a period 12 of time. Oh, I remember this happened three years 13 ago, let me go look at this. This would allow us to 14 take a more proactive approach and try to get at least 15 a hint to the engineers that are reviewing the reports 16 that there might be a trend here for these 100 17 documents, you might look at a few of those.
18 MEMBER HALNON: So okay, I can't help, and 19 I know there's one minute to go, with all the hype on 20 AI coming around, I can't help but be disappointed in 21 that unconclusive result, given the fact that we're 22 not talking about that many findings. A hundred a 23 year in the industry maybe at this point. Well, maybe 24 more than that if you get plants in trouble.
25 But and the inspection reports are very NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
143 1 structured. So they're, you know, they're very 2 descriptive of what -- and matter of fact, they're 3 repetitive in a lot of ways.
4 So I can't help but be disappointed. I 5 guess if I read your blog six years ago I'd be 6 disappointed for six years, you know. I'm glad I 7 didn't. But I would like to.
8 Anyway, so James, do you see light? I 9 mean, I know you used the term machine language, and 10 we've been using the term AI all morning. And I'm 11 sure that there's some overlap Venn diagram you could 12 show me that says that there were almost the same 13 thing but not quite, or however you want to define it.
14 But do you see some application in the 15 future where, you know, you're not going to have to 16 have this cognitive trending people dedicated to it?
17 I mean, it seems like if you could take AI and say 18 please write me a research paper on umpty squat and I 19 want to turn it into my professor and I'll get an A on 20 it.
21 It seems like you should be able to take, 22 what is it, 90 plants times four inspection reports a 23 year, 360 inspections that are all pretty well 24 structured the same and say give me what the trends 25 are in there.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
144 1 It almost seems like that would be, from 2 an AI perspective from what we've heard, is no-3 brainer. It should be able to come out with some very 4 good stuff without stop words, without other things 5 that. Because again, the inspection manuals are very 6 descriptive on you word a finding and how you word 7 cross-cutting issues and stuff like that.
8 So what's your outlook? I mean, what do 9 you think?
10 MR. CHANG: Last year, last EPRI published 11 a technical report that has a document the industry 12 using machine learning for corrected action program.
13 It was, the purpose was use of the machine learning to 14 screen out these reports, the reports certainly then 15 have a safety implication.
16 And that -- in that EPRI document, two 17 success case that it reduced the workload and come to 18 one million dollars a year, that kind of saving. So 19 to me that's -- this morning we already talked --
20 mentioned that to find the AI for safety system 21 control, that seems like that it's distant, away 22 future.
23 But the way the things seems that safety 24 important bucket focus on reducing the workload that 25 (audio interference), providing that the second layer NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
145 1 of the quality assurance, that I think it's very 2 radical.
3 MEMBER HALNON: Okay. I want to go back 4 to early teens or the 2018s when we were discussing, 5 the agency was discussing with the industry about 6 substantive cross-cutting issues that took two or 7 three at that point hits on this cross-cutting issues.
8 And if you got hit with a "substantive cross-cutting 9 issue" it could cost millions of dollars to get out of 10 it.
11 So even though you can save millions of 12 dollars in resource and other things, sorting through 13 the ten thousand corrective action documents you may 14 have, you could also be chasing ghosts to the point 15 where it's trying to fix a non-problem. But you're 16 creating a problem by trying to fix it.
17 So there's got to be a check and balance 18 there as we go forward too. And I know you saw that 19 with the stop words and other things. You saw the 20 pitfalls that could get into it. And I guess if we 21 read your blog, we probably would have known that 22 already.
23 But nevertheless, this is just some 24 thoughts. I think that there's a application going 25 forward with this. And I think that it's -- as the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
146 1 industry goes and does it with the corrective action 2 programs, the agency certainly could do it with this 3 smaller subset of findings that they have.
4 Because not very -- I mean, a site has 5 seven to ten thousand corrective action documents, and 6 we dealing with 100 findings. So it should be 7 relatively straightforward, at least in my mind here.
8 MR. CHANG: Yeah, you bring the topic back 9 to the regulate AI.
10 CHAIR BIER: Greg, one minor comment. You 11 need to speak up for the people in the back of the 12 room.
13 I had a few questions and comments that I 14 will try and make very quick. First of all, I'm 15 different than Greg. I'm usually a skeptic, but I'm 16 very excited about this application.
17 I mean, the methodology may not be there 18 yet. Maybe we have to have a different approach or 19 wait another couple of years 'til the software is 20 better or whatever. But I like it because it's an 21 example of that kind of offline type of advice where 22 it's not making a decision for anybody, it's just 23 surfacing information that then the decisionmaker can 24 look and take into account.
25 So I think that's very promising. I had NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
147 1 a few questions. One is I know some of the software 2 packages kind of hoover up your data and send it back 3 to the mother ship. And I know that Azure doesn't, it 4 lets you sandbox and keep your own data for yourself 5 and not send it back to Google or Microsoft or 6 somebody.
7 Can you comment whether the other software 8 packages you thought about have that pitfall, or 9 they're all similar?
10 MR. CHANG: No, I don't. I haven't 11 thought about this question.
12 CHAIR BIER: I mean, NRC data is pretty 13 much mostly public anyway. But in other applications 14 that can be a big issue.
15 Second of all, which years of data did you 16 use?
17 MR. CHANG: This is the inspection 18 findings from 1998 to 2022.
19 CHAIR BIER: Okay, because one of the 20 issues is like the more -- the shorter the timeframe, 21 you have less data, but it's more relevant.
22 MR. CHANG: Yeah.
23 CHAIR BIER: So that might be another 24 parameter to play with, is what if you took only most 25 recent five years or something. Maybe you would get NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
148 1 better relevance.
2 The third question, and I don't know if 3 this is a question for you, it might actually be a 4 better question for Jesse. I know in the area of 5 health, like for reading mammograms, they have found 6 that like a doctor plus an AI does better than two 7 doctors. Because the AI sees different things than a 8 human would see, and then you can get better coverage 9 of what's going on.
10 But I'm very concerned about kind of the 11 computer equivalent of social loafing. Like, you know 12 the computer's going to look at it anyway, so after a 13 while the human gets lazy and stops paying attention 14 and just acts on the computer advice. So I'd be 15 curious if either you or Jesse have given that a lot 16 of thought yet.
17 MR. CHANG: Certainly that's my expertise, 18 human reliability. Yes, you put a human from this 19 first night to the second night as a PO checker 20 (phonetic) or monitors of positions. So that kind of 21 performance certainly that we have --
22 (Simultaneous speaking.)
23 CHAIR BIER: Yeah. Jesse, if you want to 24 expand on that at all?
25 MR. SEYMOUR: I appreciate it. And so NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
149 1 this is Jesse Seymour from the Human Factors Branch.
2 One of the things I would build on James' point is 3 when you have essentially, you know, the human acting 4 in almost like a peer check type of role to something 5 that AI is doing, there is a phenomena that arises and 6 it has to do with the scrutability of the AI's 7 process.
8 So again, if two professionals look at a 9 given product independently and they disagree, they 10 can then confer, examine each other's thought process 11 and figure out why there's a disagreement and perhaps 12 take something away from that.
13 With AI, it's a bit of a black box due to 14 the nature of neural networks and so forth. And it 15 may not be possible even for the people that have 16 designed again the machine learning application or 17 whatnot to fully understand what happened in between 18 the input and the output being received.
19 So again, it's a complex matter. And 20 James, I'm not sure if you have anything, any more to 21 that point.
22 MR. HATHAWAY: This is Trey Hathaway, 23 Accident Analysis Branch. I was going to address your 24 first question about hoovering up data.
25 CHAIR BIER: Oh, super.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
150 1 MR. HATHAWAY: The models used, there's 2 essentially like ChatGPT paradigm that is kind of 3 closed off. And if you use it and it's free, they get 4 it if you have a local copy. You know, it's a little 5 differently. But these models, there are tons of 6 these models out there.
7 These particular models you download. You 8 essentially, you essentially get the weights and then 9 you're -- I'm getting told that I need to speak up.
10 You essentially download the weights, and then you 11 have the model locally. And then you can start doing 12 things like fine-tuning it on your own language and 13 things like that to kind of help.
14 CHAIR BIER: So it does not phone home 15 with all your tons of data.
16 MR. HATHAWAY: That's my understanding, 17 yes.
18 CHAIR BIER: Thank you. We have time for 19 one or two more quick questions or comments.
20 MR. CHANG: To the member's earlier 21 question that, well, you asked that what we are trying 22 to take it from here to next step. Currently, that 23 research, this future focus research that's give us a 24 wayfinding to do this, it's our research results. And 25 then it's meant to be a seed project.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
151 1 So what is taking the further step that 2 our managers, one that they use always that use any 3 request. So that was Jason that anticipate this theme 4 that he know that what's the benefit to bring in his 5 -- our people, and that's his response. Each to reach 6 out his manager. Easier to research it, a system 7 request to research it so that we can -- do the 8 additional study in this topic.
9 MR. CARNEAL: And James, just to circle 10 back to the other question for liming the year range 11 from 2000 to probably last five. We've done some ML 12 studies in my group to try and categorize OpE reports, 13 and that it has a major impact in the accuracy if we 14 only look at the last five years we get much better 15 results.
16 And I would imagine that for algorithms 17 for like this, we would see some similar results.
18 Because the way that the inspection reports are being 19 generated now is much different than in the past. We 20 had people writing free text back in the past and 21 going through all these reviews that were 22 inconsistent.
23 Right now, since 2018, what appears in 24 that database is going to be what appears in the 25 inspection report. Because they have the option to NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
152 1 auto-generate that text. So we should be able to 2 discern a little bit better for the more recent 3 reports from the boiler-plate language versus the 4 actual meat of the inspection.
5 O: Just to address your range-of-date 6 issue, in trying to assess what you can do with the 7 technology, that's kind of what you're talking about 8 here. Aren't you liable to the 1998 data seems kind 9 of not relevant? It's been adjudicated, something's 10 been done.
11 Why wouldn't you look for a more active 12 data set where things people have not made decisions 13 hadn't been closed out? I'm not so sure how 25-year-14 old data is going to tell you what you can use this 15 for as opposed to like the last five years. It's 16 active data that people have made decisions, and not 17 it's just whether the decisions were correct or not.
18 Not correct, but were as good as they could have been.
19 So I'm just getting too much data that's 20 not really -- if it's really old, it doesn't do you 21 much good in terms of getting to assessment. That's 22 my only point.
23 MR. CHANG: Yeah, that's -- the data 24 quantity, that was the concern at the very beginning 25 of this project because of what we know, that today's NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
153 1 -- this algorithm needs a lot of data. And one 2 example that EPRI report I mentioned, that the example 3 was using 600,000 records for the correction action 4 open, together 80-something percentage with a 5 successful rate.
6 And so that was -- so before this project, 7 we know nothing, it was just take whatever we have in 8 the excel database solely as is. So now that's a 9 helpful input that we have funding for continue work.
10 That's something that we will take the recency into 11 consideration.
12 MR. HATHAWAY: Yes, this is Trey Hathaway.
13 I think just, sorry, I talk quietly. Sort of talking 14 to your point, the idea of a lot of these natural 15 language processing techniques is you're trying to 16 have a signal to noise.
17 So when you do clustering, you're applying 18 the model to the document and sort of getting features 19 that the model thinks are important to the documents.
20 Or recent documents, if it is kind of like now more 21 homogenized, I guess, in how you're developing it.
22 That signal to noise is going to be kind of consistent 23 across those documents.
24 When you start introducing older 25 documents, you might sort of change that signal to NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
154 1 noise to where it's not that just can't do it, it's 2 that you have to spend more effort in that cleaning 3 part of the text to sort of get out that text that's 4 not really relevant.
5 And that's kind of the challenge with 6 these. Eight percent of the work is just getting the 7 text in a way that you're getting rid of a lot of the 8 noise to kind of focus on what's important.
9 CHAIR BIER: We are going to need to end 10 the meeting now because we have another meeting in 11 this room over lunch. So thank you very much.
12 Hopefully some of the conversations can continue out 13 in the hall or whatever. But thank you for a good 14 morning.
15 (Whereupon, the above-entitled matter went 16 off the record at 12:03 p.m. and resumed at 1:06 p.m.)
17 CHAIR BIER: Okay. Now I think we should 18 be back in business. Can somebody online hear me?
19 MEMBER MARCH-LEUBA: Yes, I can hear you.
20 MEMBER DIMITRIJEVIC: Yes, we can hear 21 you, Vicki.
22 CHAIR BIER: All right, thank you. So, 23 Bruce, do you hear me, and can you say something so we 24 can check that we hear you? Ah.
25 MR. HALLBERT: Sure. I can --
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
155 1 CHAIR BIER: Okay.
2 MR. HALLBERT: I can say whatever you 3 like. Can you hear me okay?
4 CHAIR BIER: I think we're good. Yes, we 5 hear.
6 MR. HALLBERT: Sounds good.
7 CHAIR BIER: And are you going to share 8 your screen for your own slides?
9 MR. HALLBERT: I am.
10 CHAIR BIER: Okay. Then I think we are 11 ready.
12 (Audio interference.)
13 MR. HALLBERT: Whoa.
14 CHAIR BIER: Oops.
15 (Audio interference) 16 MR. HALLBERT: Okay, we had a little bit 17 of an echo there, but I think we got that resolved at 18 the moment.
19 CHAIR BIER: I think it sounds much better 20 now.
21 MR. HALLBERT: Okay, great.
22 CHAIR BIER: So, on that I think you can 23 just go ahead and get started with your presentation 24 since we're running a few minutes late. May as well 25 get it going.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
156 1 MR. HALLBERT: Yes.
2 CHAIR BIER: Thank you, Bruce.
3 MR. HALLBERT: I'll do that. Thank you 4 very much. Thank you for the opportunity to 5 participate in this meeting.
6 I am Bruce Hallbert. I'm the national 7 technical director for the DOE-sponsored light water 8 reactor sustainability program. And with me this 9 after we have Craig Primer and Ahmad Al Rashdan from 10 our program who will also be talking a little bit 11 about our R&D activities. Especially related to 12 artificial intelligence machine learning.
13 I want to also recognize in the call we 14 have Ms. Alison Hahn, who is the federal program 15 manager for the LWRS program currently. And she is 16 also one of the office directors at the Department of 17 Energy.
18 I'd like to talk a little bit about the 19 goals and objectives of the LWRS program, as we call 20 it. The goal of the program is to enhance the safe, 21 efficient and economic performance of our nation's 22 nuclear fleet and to be able to extend their operating 23 lifetimes.
24 I'm picking up some feedback on my end.
25 I don't know if you're picking it up on your end as NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
157 1 well? No, maybe not.
2 MEMBER DIMITRIJEVIC: Yes, we can hear the 3 feedback here too, so I don't know where it comes 4 from.
5 MR. HALLBERT: Okay. Sometimes it's from 6 when somebody has their microphone open. So as long 7 as everybody else is muted I shouldn't be picking up 8 feedback.
9 (Audio interference.)
10 MR. HALLBERT: So we achieve our 11 objectives by supporting the long-term operation of 12 existing nuclear power plants by deploying innovative 13 approaches to improve the economics and economic 14 competitiveness of light water reactors in the near-15 term, as well as in the future energy markets. And 16 sustain the safety, improve the reliability and 17 enhanced economics. We go about this by conducting 18 research in the five focus areas that you see on the 19 bottom left of the presentation, which I'll be talking 20 more about in a moment.
21 In the bottom right graph of this slide 22 sort of brings it all together. Our focus is on 23 enhancing economic competitiveness by helping plants 24 to reduce their O&M costs and looking into 25 opportunities to diversify revenue. Especially for NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
158 1 plants that find themselves in electricity markets 2 where they may be under economic pressure from, you 3 know, subsidized renewables and inexpensive natural 4 gas.
5 On the right side we also are addressing 6 the long-term performance of materials, structures, 7 systems and components, as well as managing the aging 8 and technology obsolescence of some of the systems and 9 technologies that are used to operate nuclear power 10 plants today.
11 I'd like to talk about each of the five 12 R&D areas of the program as part of the overview. And 13 I'll also have some remarks on artificial intelligence 14 within the LWRS program that I think provides some of 15 the context for what you're going to hear from Craig 16 and Ahmad.
17 The first area of R&D is plant 18 modernization. The goal of our research in plant 19 modernization is to facilitate modernization at 20 operating nuclear power plants. We do so by 21 developing technology and modernization solutions that 22 address aging and obsolescence challenge. But they're 23 not just about replacing old technology with new 24 technology, they're about delivering a sustainable 25 business model that ensures continued safe and NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
159 1 competitive operations. And, well, we can talk more 2 about how we accomplish that.
3 But the focus is really on long-term 4 management of plant systems. And when we talk about 5 long-term operations we mean especially from 60 years 6 and beyond. And we also are addressing nuclear cost 7 competitiveness as nuclear power plants face cost 8 pressures from a lot of power generation sources. And 9 I'll talk a little bit about that in my forthcoming 10 slides.
11 And of course, one of the things that 12 we've really learned from the experiences of the 13 pandemic is that it's very important to address worker 14 attraction and retention. Some of the digital 15 technologies that we're working with through our 16 program, as well as with the industry, really are a 17 technology base that the new workforce is more 18 familiar with, and also see as a part of their long-19 term career prospects.
20 An example of one of the ways that we're 21 working with the industry to modernize the fleet is a 22 project that DOE is sponsoring in cost sharing with 23 Constellation. And that project is to replace the 24 reactor protection technologies at Limerick 25 Generating, both Limerick Generating Stations, in NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
160 1 Pennsylvania.
2 And this is a project, interestingly 3 enough, that really was brought to us by the Nuclear 4 Regulatory Commission because of some of the new 5 approaches to licensing digital at digital 6 instrumentation and controls, especially for safety 7 related types of applications. And so this is a 8 collaborative effort between Constellation, DOE and 9 the NRC.
10 And it does focus on the first-echelon 11 safety instrumentation systems. We've been doing this 12 now since around 2021. We're now in the, approaching 13 the fourth year of R&D efforts towards the full 14 replacement of the systems. And one of the roles that 15 the Department of Energy, and the Idaho National 16 Laboratory specifically play is supporting the human 17 factors aspect of that control room modification 18 modernization project.
19 This slide highlights one of the recent 20 activities that was conducted at INL in February of 21 this year. Which was to support the dynamic 22 preliminary system validation. And for that project, 23 or that part of the project, we had people from the 24 nuclear regulatory commission, from Constellation, 25 from Westinghouse, other vendors and suppliers, as NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
161 1 well as the Department of Energy, participate in those 2 studies. The results of those workshops were used as 3 part of the license amendment request, which was 4 supplied by Constellation to the NRC.
5 So this is one area where we collaborate 6 with the commercial nuclear power industry and the 7 NRC, and vendors and suppliers to address some of the 8 long-term instrumentation and modernization needs of 9 the industry. And I want to just provide that as an 10 example so you understand some of the ways that we 11 work within industry.
12 Specifically with respect to artificial 13 intelligence and machine learning, the topic of 14 today's meeting, we've been working with AI and ML 15 technologies for about the past four or five years.
16 And someone said to me recently, and I think it's 17 true, that artificial intelligence is like the new 18 math. We find it more and more within a lot of our 19 projects. And I'll try to characterize and summarize 20 that, but Craig and Ahmad will go in more detail.
21 We believe, well, these are relatively new 22 to the nuclear power industry. And similar to the 23 observations from the NRC Staff this morning that we 24 have, from participating in IAEA and other 25 international meetings, I do believe that the U.S. has NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
162 1 leadership with respect to some of the initial efforts 2 investigating AI and ML with nuclear power plant types 3 of applications.
4 They show promise, especially for 5 automating manually performed activities. Many of the 6 things that we do at nuclear power plants today are 7 very labor intensive. You'll hear about some of those 8 in our discussions. But we also see them as a way to 9 enhance monitoring.
10 So they look promising to us as a way to 11 enhance efficiency. But I want to also advise that 12 what you're going to hear from us today really reflect 13 R&D efforts. So when we show, for example, an 14 activity where we're collecting data or conducting a 15 test or something like that, at or with an operating 16 nuclear power plant, that's not an actual deployment.
17 The same thing is true, speaking on behalf 18 of licensees and vendors. We're not doing that today, 19 we're really talking about our own R&D efforts, which 20 may in fact be collaborative. But they really are 21 focusing on three things. Reducing O&M costs, 22 enhancing efficiency of the workforce, as well as 23 improving situational awareness.
24 Moving on to the second point on here, our 25 efforts, as I mentioned, emphasize work processes NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
163 1 versus controls. So I know that some of the members 2 of the ACRS have a background in instrumentation and 3 controls, including control processes. We're not 4 investigating control activities, like so, we're not 5 looking in deploying AI to control systems so much as 6 just automating work activities that are labor 7 intensive.
8 It's very important. And we're taking a 9 very deliberate approach in our efforts with the 10 vendors and suppliers and operating nuclear power 11 plants to ensure that AI aligns with the nuclear 12 safety culture. Just like we have with every other 13 part of our R&D activities. As well, we are 14 reflecting on and are trying to comply all of our 15 efforts with presidential directives and other 16 directives on AI that have been issued since 2019.
17 More recently by the President.
18 Ultimately we think that AI will enhance 19 worker performance at nuclear power plants. And I 20 want to really emphasize that. We don't see AI as a 21 means to replace people, but a way to enhance 22 performance and help people do what they're best at.
23 And that's a reason why we have a strong 24 emphases and focus in our research on human factors 25 issues. We think it's absolutely vital to NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
164 1 implementation and to achieve that.
2 It's important that workers trust 3 automation, that they understand automation so that 4 there is transparency, understandability. And 5 ultimately facilitate usability. And hopefully that 6 will come through in some of the remarks from Craig 7 and Ahmad today as well too.
8 Now, I would be remiss if I didn't talk 9 about the other areas of the light water reactor 10 sustainability program, so the remainder of my 11 presentation will be on the other activities that 12 we're dealing not so much on AI. So I don't know if 13 you have any questions so far, or if you'd like for me 14 to continue with the rest of the presentation. But 15 I'm open to any questions any time.
16 CHAIR BIER: Do we have questions so far 17 or do people want to finish up first?
18 MEMBER KIRCHNER: Bruce, this is Walt 19 Kirchner. Just quickly, you emphasized that it's in 20 the R&D phase now, but do you have a few collaborative 21 ventures where you're actually going to take it out of 22 the lab so to speak and into a power plant and look 23 for opportunities to harness this to either enhance 24 productivity or enhance monitoring or --
25 MR. HALLBERT: Yes. So I would say, and NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
165 1 Craig will show this, but I think the answer to your 2 question really is in Craig's examples. Most of our 3 R&D is at and with operating nuclear power plants. Or 4 with their data.
5 Now, those are not deployments. And I 6 want to emphasize that. Those are not implementations 7 but they're examples of how we want to ensure that our 8 R&D activities could be used or could be transferred 9 to the private sector as part of the technology 10 transfer efforts. And that they do scale to real 11 problems at real nuclear power plants. So hopefully 12 you'll see that. But yes, that is, most of our 13 research is out of the lab in many ways.
14 MEMBER KIRCHNER: Thank you.
15 MR. HALLBERT: Okay, I'll continue on.
16 MEMBER BROWN: Hey --
17 MR. HALLBERT: Oh --
18 MEMBER BROWN: This is Charlie Brown.
19 Yes. In our earlier discussions we, in our earlier 20 meeting, before noon, we had considerable discussions.
21 And you made the comment in this that you're focusing 22 on how you would improve operator or man, eating up a 23 lot of man hours, you know, stuff that takes a lot of 24 time. Intensive stuff but not necessarily focusing on 25 controls.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
166 1 And we had some discussion earlier today, 2 or actually a lot of discussion earlier today, in 3 terms of the, how we have to be careful about getting 4 AI into reactor trip, safeguard systems, plant control 5 systems, such that we have something modifying or 6 making decisions for humans or what have you when it 7 really doesn't add value. Stuff you're talking about 8 that seems to add value in terms of how you manage the 9 plant in its operations, but when you want to trip the 10 reactor you don't have to make a whole lot of 11 decision. Your power is either too high or it's not.
12 Or you've either lost pumps or it's not.
13 It's not a, what I would call a real 14 machine learning or other deep thought process to 15 determine what you want to do in trying to embed this 16 new idea into those systems could be detrimental to 17 their ability to process it. Is that involved in your 18 all's discussions in terms of how you, you know, rice 19 bowl offer, you know, put a bar around certain areas 20 that it really is not going to add value.
21 I'm in favor of the real added value stuff 22 not --
23 MR. HALLBERT: Yes.
24 MEMBER BROWN: -- just doing it where it 25 seems like a nice thing to do because everybody else NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
167 1 is.
2 MR. HALLBERT: Yes. Yes, those are great 3 questions. And also good comments as well, Charlie.
4 One of the things that we, we do look at 5 as a part of our approach into where we might look at 6 a project to investigate an AI application is based 7 upon a business approach. So we often times have a 8 business case for, this is a very labor intensive 9 activity. A lot of people are involved in doing it.
10 It's not high value added from the perspective of the 11 utility and they wonder if there is a way to automate 12 some of this through analytics AI and machine 13 learning. So I think you'll see some examples of 14 that.
15 We're not focusing on anything that's 16 inside the control room especially. And we're not 17 approaching anything, we're not even looking at 18 minimum inventory, we're not looking at Class 1A 19 systems.
20 That's all outside of the scope of what 21 we're investigating today. We're looking at, what are 22 some of the ways that we can help plants to be more 23 efficient in terms of those vary labor centric types 24 of activities, but also provide information that's of 25 value to the people who are responsible for those NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
168 1 functions at the plant. Hopefully that answers your 2 question.
3 MEMBER BROWN: No, that's a good answer.
4 I, it's not necessary, it doesn't have to be a good 5 answer. It doesn't, my opinion the right answer. It 6 looks like you're all going down the thoughtful path 7 that we did spend considerable amount of discussing 8 earlier in the day.
9 MR. HALLBERT: Yes.
10 MEMBER BROWN: So thank you for --
11 MR. HALLBERT: Well --
12 MEMBER BROWN: Thank you for your response 13 there.
14 MR. HALLBERT: Of course. And we'd 15 appreciate feedback. That's one thing that we are 16 always looking for is feedback on our approach and 17 projects. And we'll be providing with links and lots 18 of reports as well too.
19 In the interest of time I'm going to jump 20 through some of the rest of the slides so that Craig 21 and Ahmad actually can have the time that they deserve 22 to go into detail. You all have heard about probably, 23 unless there is some more questions right now.
24 You all have heard probably about some of 25 the activities related to hydrogen demonstration NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
169 1 projects. A lot of that was initiated, and the 2 foundational research was sponsored by the Light Water 3 Reactor Sustainability Program.
4 We've been conducting research into 5 potential uses of operating a nuclear power plants to 6 produce hydrogen, extracting thermal energy, as well 7 as just, you know, providing electricity for 8 electrolysis systems, modifications of electricity 9 transmissions. As well as doing studies dynamically 10 with operators in the human system simulation 11 laboratory with mockups in a simulated environment in 12 operating nuclear power plant that includes something 13 like high temperature electrolysis in the balance of 14 plant, looking how operators would work with the 15 double demands of electricity generation and hydrogen 16 production. We've also been working on the economics 17 of this.
18 I'm going to have to jump through my 19 slides to stay on time, but I want to emphasize that 20 the LWRS program, and other DOE offices, have been 21 supporting these hydrogen demonstration projects. The 22 first one, Nine Mile Point, is in operation. And it's 23 using one and a quarter megawatt electrolysis, low 24 temperature electrolysis unit only.
25 Davis-Besse and Prairie Island are set to NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
170 1 go into operations sometime next year. And those will 2 be also, one will be a low temperature electrolysis 3 plant, and one will be a high temperature electrolysis 4 plant which would be taking thermal energy from the 5 plant to run a 150 kilowatt high temperature 6 electrolysis unit. So it's a very small electrolysis 7 unit but it's demonstrated the means for an off take 8 of thermal energy from the plant.
9 Let's see. In terms of, I'd be a little 10 remise if I didn't mention also that the LWRS research 11 has been instrumental in supporting some of the 12 hydrogen hubs.
13 The President announced a few weeks ago in 14 Pennsylvania that there had been some hydrogen hubs 15 selected and awarded to initiate R&D into nuclear 16 power plants. Not just nuclear power plants, but 17 broader hydrogen hubs. But some of the involve 18 nuclear power plants producing a hydrogen at scale as 19 part of the hydrogen hubs. And we've been supplying 20 some of the information that we think enable some of 21 those, some of those efforts moving forward. And some 22 of the INL staff is also participating directly in 23 supporting those hubs.
24 I know I'm jumping through the slides a 25 bit here but I want to emphasize also that since its NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
171 1 inception the LWRS program has been conducting 2 research into the long-term performance of key 3 materials for vital structure systems and components.
4 In fact, when we initiated this program together with 5 the Nuclear Regulatory Commission and the Electric 6 Power Research Institute, I would say the largest 7 emphasis was on materials performance. Specifically 8 in some of the areas that you see on this slide here.
9 And the emphasis in our materials research 10 is understand how materials perform and degrade in 11 this in-service environment over long periods of time.
12 By conducting research into mechanisms, degradation, 13 modeling and simulation tools to be able to model and 14 predict that, as well as to inform mitigation 15 strategies.
16 Now we're also conducting research into 17 risk-informed system analysis, which is research and 18 development to enhance economic efficiencies by 19 optimizing safety margins and minimizing 20 uncertainties. It involves a lot of R&D in 21 collaboration with Nuclear Regulatory Commission, as 22 well as with vendors and suppliers.
23 And I chose one example from the Risk-24 informed Systems Analysis Research which is a project 25 that's looking into optimizing nuclear fuel NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
172 1 utilization. As you know, fuel costs represent about 2 20 percent of the annual operating expenses of a 3 nuclear power plant. And we have a project that's 4 using an AI optimization framework for designing 5 reactor core configuration giving certain objectives 6 and constraints.
7 The little simulation on the right side 8 here that I hope you can see, shows our simulation 9 running through a number of iterations on nuclear fuel 10 movements and switching to optimize the amount of fuel 11 that needs to be purchased during an outage, as well 12 as hopefully in the long-term the amount of fuel that 13 needs to be stored on the back-end of the process as 14 well too. It's a multi-physics based R&D project that 15 uses a generic algorithm as the AI method for 16 optimizing core loading.
17 Finally we're conducting research in 18 physical security. And this is a topic that was 19 raised to us by the nuclear power industry where they 20 really asked if they are opportunities for DOE to 21 share and leverage some of its own capabilities and 22 physical security and protection with the commercial 23 nuclear power industry.
24 I won't go into much information on that, 25 but we do have a vibrant engagement activity with the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
173 1 Nuclear Regulatory Commission and the Industry on 2 this, looking into advance security technologies, 3 risk-informed physical security and a number of event 4 security sensors and delayed technologies.
5 So, I'm trying to keep us on time. I just 6 want to summarize by saying, going back to the 7 original purpose of the LWRS program. Now we know 8 that the existing fleet operating today provides the 9 largest reliable source of carbon-free electricity in 10 the U.S.
11 Some of the industry initiatives, like 12 those that have been led by the Nuclear Energy 13 Institute, DMP, have achieved substantial improvements 14 and performance already. Energy, our nuclear energy 15 supports, our climate goals can also contribute to 16 deep decarbonization in other industries by providing 17 energy for products that help to reduce the carbon 18 footprint in some other industrial sectors.
19 A lot of the R&D activities that you'll 20 hear about today from our program involve 21 collaborations with Industry because we want to 22 facilitate progress in areas of vital common 23 interests. And by working together we can facilitated 24 that kind of progress. Especially with some of the 25 first movers in the industry who are interested in NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
174 1 moving forward in some area but maybe themselves are 2 not R&D organizations. So partnering with DOE makes 3 a lot of sense. We also have a lot of the risk for 4 some of the early R&D approaches.
5 So, I would also say that our research is 6 based on the highest priorities that we identified in 7 the commercial industry. And they're conducted on 8 timelines that support continued operation of the 9 existing fleet. And I'll be happy to answer any other 10 questions before we turn it over to Craig.
11 CHAIR BIER: Do people have other 12 questions for Bruce?
13 MR. HALLBERT: Thank you very much for 14 your time and the opportunity to talk to you today 15 about our research.
16 MR. PRIMER: First, before I get into this 17 slide, I just want to thank the NRC's engagement.
18 There was mention of an MOU and the ability to share 19 ideas under that agreement and provide technical 20 information. And that has allowed us to identify what 21 we think are meaningful research areas (audio 22 interference) --
23 CHAIR BIER: Thank you for letting me 24 know. Thank you. We should be okay now.
25 (Simultaneously speaking.)
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
175 1 MEMBER REMPE: -- that is muting. It's my 2 microphone because I'm the person, Joy Rempe, who is 3 logged into the meeting, but it is the room microphone 4 and please do not mute it because we're constantly 5 having to unmute it. Thank you.
6 MR. PRIMER: Okay. And thanks. So what 7 I was mentioning is, so we have some objectives and 8 missions that Department of Energy have established 9 for the program. And Bruce was able to share that 10 with you.
11 Some other things that I'd like to mention 12 is, I heard discussion on what's AI versus machine 13 learning. And so, this is a general survey of ideas.
14 This is something AWS, Amazon Web Services, developed.
15 It's very similar to what you'll see as you're, some 16 of the other products.
17 You have artificial intelligence, which is 18 the ability for machines to take different types of 19 inputs, make decisions using the Turing test. You 20 wouldn't know if it's a person or a machine, it just 21 does it. And it's able to process information and 22 make decisions.
23 The building blocks for that type of 24 intelligence is machine learning and deep learning.
25 Those are specialized bits of logic that is used to NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
176 1 interrogate large data sets and infer or create some 2 ideas around that information.
3 So machine learning is made up of several 4 different types. You'll hear the mention of national 5 language processing, computer vision, time series.
6 These are all different types of data sets that are 7 available for algorithms to interrogate and determine 8 things.
9 Next is deep learning. So deep learning 10 is something that takes different types of machine 11 learning algorithms and puts them together in unique 12 ways to solve some difficult issues. So you might 13 have natural language processing and computer vision 14 working together. And you'll see, and the mode will 15 show some examples of deep learning in (audio 16 interference) drive into that quite a bit. But the 17 deep learning is just a more complex set of algorithms 18 that are used to solve a problem.
19 And then we also talked about generative 20 AI. So that, again, is something that's out there 21 where AI is taking information and creating new things 22 from it. Whether it's art or reports or things like 23 that.
24 So we're working in the machine learning 25 and deep learning areas. That's what we're focused NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
177 1 on, that's what our research supports. We see 2 opportunities to work with Industry to develop 3 solutions or algorithms and then demonstrate the 4 usefulness. And then from the collaboration, develop 5 some type of reports and information that others, 6 Industry or vendors, can use to develop products.
7 And ultimately make the decision for this 8 specific station on how to use that. And where to use 9 that. It's the research organization, we're not part 10 of that decision making.
11 So, just to highlight what we'll be 12 talking about in the next several slides. We'll be 13 showing examples of machine learning for material 14 management or equipment monitoring and anomaly 15 detection, as well as applications, examples of 16 natural language process and computer vision in 17 applications within the plant.
18 Then last on the bottom there is the AI 19 and ML explainability. So one of the key elements of 20 deep learning, and some of those more complex 21 algorithms is, as it becomes more complex it's more 22 difficult to understand what it's doing and why the 23 results are correct. So developing a balance between 24 complexity and explainability is important, so we're 25 doing some work there.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
178 1 I'll also mention before I go into the 2 specifics, our group is made up of a group of human 3 factors, scientists and engineers, data scientists, 4 control system engineers, and finally a business group 5 that's looking at balancing. And Bruce mentioned the 6 motivation.
7 So why would we put research effort into 8 this versus another opportunity. So there's a 9 business case element that's considered into what 10 we're working on.
11 So I guess I'll just stop there. Any 12 questions on that general idea of these links of 13 artificial intelligence to machine learning and deep 14 learning? Okay.
15 So the first example of machine learning 16 is in the passive system, or material management or 17 material monitoring. You have examples of machine 18 learning in looking at defects in concrete, defects in 19 piping.
20 What you see here on the left is an 21 example of collaboration between EPRI, Southwest 22 Research Institute, several universities that created 23 slabs of concrete with known defects and curated that 24 as kind of a data set of sorts but a physical model 25 with known defects that we could apply machine NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
179 1 learning technology to, to see if we could identify 2 where those defects were.
3 Similarly on the right we had a effort 4 with, again, EPRI and Southwest Research Institution 5 on piping degradation where we coupled university 6 work, University of Pittsburgh in this case, with the 7 research scientist to develop digital twins, what 8 piping should look like, compare it to the sensor data 9 and try to inform the operators, and what would likely 10 be maintenance teams, of where there might be defects 11 within the piping.
12 And so we have those reports that are 13 available and have been used on follow on projects to 14 develop solutions that industry can use to help them 15 in their different passive monitoring programs.
16 Next, moving to equipment monitoring. So 17 from passive equipment or passive components to active 18 components. We have an example of work that we did 19 with Industry creating a digital twin of circ water 20 system. That's the bottom left portion of that, so 21 that's different components within the circ water 22 pump.
23 We worked with vendors to create the 24 software that created the, or developed the digital 25 twin for that. And then also worked with the utility NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
180 1 to enhance the data that was available to monitor that 2 equipment. And through that effort we were able to 3 identify an approach that would monitor the circ water 4 system performance and provide input to the system 5 engineering and maintenance teams.
6 Ultimately the goal here is to develop the 7 basis for transitioning from predictive, or from 8 periodic maintenance to predictive maintenance where 9 we can identify the likelihood of a component failure 10 and advance warnings so that the component can be 11 taken out for maintenance in what would be a planned 12 appropriate time frame.
13 The lifecycle there that's indicated on 14 the top right shows that data analytics is the 15 beginning of that effort. We take the data, run it 16 through different types of fault signatures and 17 identify when we're starting to see some type of 18 faults.
19 And once we do that we inform the decision 20 to do that maintenance using different types of 21 predicative modeling and risk modeling to determine 22 the likelihood of some kind of failure before we could 23 get to the maintenance or if it would seem to be the 24 fault growing at a rate that may cause a problem 25 before the plant maintenance. So this modeling then NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
181 1 allows you to make the decision on when you want to 2 take the equipment out of service. And that would be 3 what we call condition based maintenance or predictive 4 maintenance.
5 MEMBER HALNON: Craig, this is Greg. The 6 artificial intelligence box there, the first blue one, 7 that's actually developing an ongoing model of the 8 instrument response and comparing it to expected or 9 how does it, I'm trying to find out, where is the 10 human displaced in this because we've been doing this 11 for years, obviously, for --
12 MR. PRIMER: So, what this would do is 13 this would complement your system engineers and this 14 would allow them to look at fault growth and then try 15 and determine remaining useful life on a component and 16 then plan the maintenance activities. And to that 17 point of the artificial intelligence on the second, 18 the predictive modeling.
19 So again, using artificial intelligence is 20 kind of a bracket for machine learning and that sort 21 of thing. That's what that's describing, is this is 22 where you apply different types of predictive models 23 and try and determine when the likely remaining useful 24 life is projecting the component to fail.
25 MEMBER HALNON: And is that using NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
182 1 operating history or operating experience outside of 2 this plant itself specific to the instruments, or 3 whatever you're monitoring?
4 MR. PRIMER: So, and that's one of the, so 5 one of the points that was discussed earlier is that 6 data availability, data quality, data completeness.
7 So all of these things are part of what you can 8 evaluate and determine the quality of your models.
9 And so there's that confidence level of the data based 10 on historic, the amount of historic data that's 11 available to the system engineers and to the model.
12 And so that's a, I think an important 13 point, is how do you establish what that minimum 14 quality is, how do you identify in a way that people 15 can consistently apply that approach to different 16 models to get some type of information on its 17 usefulness.
18 MEMBER HALNON: Okay. But it is using 19 outside information to some extent? I mean, expected 20 you want to get as much information as you can --
21 MR. PRIMER: Right.
22 MEMBER HALNON: -- it's a matter of 23 whether or not it's valid to that specific piece of 24 equipment or whatever your launching.
25 MR. PRIMER: There is a large group of NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
183 1 data that's available. So you have historic data, you 2 have real-time data that it's comparing signals 3 against what we know are those fault signatures. So 4 it's looking for that type of situation.
5 The other thing that it would look at is 6 the other types of components that are similar but not 7 necessarily that component --
8 MEMBER HALNON: Okay.
9 MR. PRIMER: -- so they have different 10 RULs.
11 MEMBER HALNON: Is there any discussion 12 about a real-time data using multiple plans, multiple 13 fleets talking about, you know, rather than just use 14 your own plant, which may be one or two instruments, 15 we can use 25 or 30?
16 MR. PRIMER: Right. So that's something 17 maybe you might call federated learning. And that's 18 using data sets that are close but not specific to 19 that component. And so there's an approach that we've 20 worked through and have reports on and we think it's 21 useful. It moves you much further in the learning and 22 training lifecycle if you do it that way then starting 23 with the small set of data and trying to infer 24 information, so.
25 MEMBER HALNON: Okay. Thanks.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
184 1 CHAIR BIER: Can you talk a little bit 2 about how the machine learning interacts more with the 3 physic space model? Like, are you just estimating 4 parameter values in the physics model or is it 5 something more complex than that?
6 MR. PRIMER: Well, without a specific 7 example I'll just generalize. So depending on the 8 amount of information and where that, you know, what 9 sensor information is available you're able to infer 10 the status of that component.
11 So what we see is a lot of just, first 12 order physics models that are used around a component 13 to inform how the component is doing. You have other 14 information like current, you know, I'm looking at a 15 circ water system, so stator current and things like 16 that are information that's used. So different points 17 are used.
18 And in fact, maybe the next slide might 19 help with this. We're getting into a little bit of 20 the explainability here.
21 And so what you'll see here is, as part of 22 that circ water system effort, and I should have put 23 it in the original design of the output of the 24 information was something the data scientists really 25 loved. And it made complete sense to them. And they NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
185 1 would say, what else do you need to something that was 2 a little bit more information available to the system 3 engineers. So the data is telling me something, but 4 what does it mean. And there is these different 5 elements.
6 So on the right side you'll see the 7 dashboard that was developed as a result of the 8 physics model. So the physics model is telling us 9 things.
10 And you can use explainability matrix that 11 are like the LIME and Shapley that tell you what's 12 your feature that's causing the algorithm to say that 13 you have a fault. And this, if you see in these red 14 bars across the bottom, they're telling you right now 15 that the high temperature to the, is above, and I 16 can't read that. It's the motor temperature is above 17 a certain temperature. And that's your number one 18 feature that's telling you that you have a problem.
19 And so you have these specific, very measurable sensor 20 inputs that tell you something.
21 But to complement that there is other 22 information that a human would want to know. And so 23 to the right of those bars are things that a system 24 engineer would likely want to know.
25 So in the bottom right is seasonal NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
186 1 temperature. So summer time or is it winter time.
2 It's high because it's hot outside or is this high 3 because I have a problem with the motor.
4 And then the one above that is the 5 prediction of what's likely to happen based on the 6 fault progression. So what that's looking at is, 7 based on the algorithm it's telling me that within the 8 next 24 hours2.777778e-4 days <br />0.00667 hours <br />3.968254e-5 weeks <br />9.132e-6 months <br /> you're likely to see a min or max, two 9 standards. You know, one way or the other of that 10 temperature change. So they can look at say, well, if 11 we do nothing we'll likely be okay. Or if we do 12 nothing there may be a problem with us succeeding a 13 limit. So that's pretty useful.
14 CHAIR BIER: Yes. For comparison with the 15 presentation just before lunch, is this also 16 considered unsupervised learning or is the fact of 17 having the physics space model and all these 18 constraints make it supervised?
19 MR. PRIMER: So --
20 CHAIR BIER: Look at you, you're --
21 MR. PRIMER: Yes, you're, this is perfect.
22 So the next slide, so examples of unsupervised versus 23 supervised, great presentation this morning on that.
24 I think there is benefits to both. I think they all 25 complement each other and give you different types of NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
187 1 information.
2 I think unsupervised methods give you 3 insight that you may not be thinking about. And it 4 just tells you what it sees, and that's very 5 interesting.
6 I think semi-supervised on what, in the 7 nuclear industry, even though there is a lot of data, 8 it's small data sets. So having supervise or semi-9 supervise it tags or fingerprints certain indications 10 that tell it that this is a fault or not a fault, 11 helps it identify conditions.
12 And so we've run through and identified 13 several different approaches that can be useful in 14 both methods. And so I wouldn't say, I'd say they 15 both are good, but they both need to be, you need to 16 understand how to use them and how they complement 17 each other.
18 And then now back to the physics space.
19 So the physics is the way you compare things to 20 actual. So you have a physics model. You look at 21 what you think it should be and then you look at what 22 you actually got. And if there is a delta then you 23 start looking into why.
24 And sometimes these algorithms, like on 25 this, with this little orange thing here says, we NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
188 1 think it's water box valve. That's what we think that 2 is. We've seen this is in the past, we think it's 3 water box valve. We being the AI thinks. And then 4 the system engineer will go, I agree or no there is 5 something that this didn't take into account that I 6 have looked into and so we'll just disposition this as 7 a nothing, no action needed.
8 Now back to the unsupervised and 9 supervised detection. And I made the point earlier, 10 so data wrangling and data quality is huge. So if you 11 don't have good data or you don't have complete data 12 sets that let you understand what's happening to the 13 component or systems, it's hard to understand what 14 that means. So the data quality and data completeness 15 are two areas that we've worked on as well.
16 Then in the bottom right you have a system 17 that allows us to actually cluster information 18 compared to the history of component failures. And on 19 the top right what we've done is we've actually used 20 what we would call virtual sensors or additional 21 sensors that are outside of what the system normally 22 uses. So in the case on the top right we were able to 23 use ambient temperatures, historical temperatures 24 around that component that wasn't part of the system 25 sensing. So it will combined information to one NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
189 1 algorithm that wasn't available to operators.
2 I'm going to jump to the next topic unless 3 there is questions on these points. Okay.
4 So now moving to some of the applications.
5 I mentioned the idea of machine learning having 6 different types of application, natural language 7 processing, time series, computer vision, and how do 8 we use those in ways that help us improve efficiency 9 and reliability on the business side. That's really 10 our balance in the research area is to leverage 11 technology.
12 Not necessarily in the short-term to 13 invent new technologies, but really validate the 14 technologies that are available now for use in the 15 nuclear industry. And that's part of our research 16 focus.
17 And you can see, we probably talked about 18 the EPRI report that was produced. That helped with 19 the natural language process. And that was actually 20 part of the output of research that started at the 21 national lab.
22 And partnered with EPRI and Industry to 23 develop open source software code that was used. I 24 think Jensen Hughes picked it up and they're using it 25 at Constellation for the corrective action screening.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
190 1 We've got Xcel Energy, I think I saw them on the line, 2 using some of the software that's been developed and 3 the collaboration with them as well.
4 So that natural language processing and 5 deep learning methods, like I mentioned, will go into 6 what that looks like with the mode. But it's a very 7 useful approach to analyze data and make some, infer 8 some information from it.
9 The, similarly, looking at the warehouse 10 stocks and looking at part failures and understanding 11 what pieces of parts were actually used, what's likely 12 to be needed in the future based on, you know, time 13 frames of the year, time frames of maintenance cycles, 14 whether that's months or years, it helps improve the 15 warehouse efficiency to make sure they have the right 16 parts on hand.
17 All of these have the report numbers 18 labeled there and we're able to share this 19 presentation. Their live links. And you can lick to 20 the OSTI, which is the DOE's library with these 21 reports. Or you can just, if you don't want to go 22 live link, you can get a PDF and type that in.
23 Lastly, I think this is lastly, the 24 computer vision. And Ahmad is going to jump in, very 25 deeply, on an example of the use of computer vision NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
191 1 for fire watch and what we've done there.
2 Some other things that are useful for 3 computer vision is gauge reading. So we can send out 4 people to take pictures. And you can detect changes 5 from one picture to another.
6 So you could possibly use that for system 7 turnovers where you have a picture of something and 8 then the next person that comes in can look at the 9 change from when they were there last and say, oh, 10 these are the three things that have changed. This 11 computer vision will detect that, highlight that and 12 help them assist in the turn over to make sure that 13 it's identified and discussed.
14 Similarly we can use QR codes to help 15 align drones to go to the right places and capture 16 information.
17 And lastly, I'm going to kind of leave 18 this to Ahmad to jump deeply into the use of computer 19 vision for fire watch. And he'll be going through the 20 rest of the, I guess hour that we have left, on that 21 topic.
22 CHAIR BIER: Questions for Craig before we 23 transition?
24 MR. PRIMER: And I'll mention, this last 25 slide is just the same slide that Bruce presented in NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
192 1 our guidelines and mission from the DOE and artificial 2 intelligence research we're developing and 3 demonstrating applications that can be used in non-4 safety, non-control type of things right now.
5 Anything that would move beyond that isn't research 6 that's underway right now.
7 CHAIR BIER: Okay. And is Ahmad going to 8 use the same slide deck?
9 MR. PRIMER: I think Ahmad is going to 10 take control. Ahmad --
11 DR. AL RASHDAN: Yes.
12 (Simultaneously speaking.)
13 CHAIR BIER: You can share your slides.
14 MR. PRIMER: That's great.
15 CHAIR BIER: Okay.
16 DR. AL RASHDAN: Yes, I'm going to go 17 ahead and --
18 CHAIR BIER: Excellent.
19 DR. AL RASHDAN: -- share my slides.
20 CHAIR BIER: Thank you. You may need to 21 stop sharing, Craig.
22 MR. PRIMER: I can do that. Okay.
23 CHAIR BIER: Okay.
24 MR. PRIMER: Stop screen sharing. Let's 25 see if this works. I see your screen now, Ahmad.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
193 1 DR. AL RASHDAN: Okay, great. And you can 2 hear me, right?
3 MR. PRIMER: Yes.
4 DR. AL RASHDAN: Okay, great. So thank 5 you so much, first of all, for the introduction, for 6 the invitation to present today. And, Craig, thank 7 you for your presentation and the segue into this 8 talk.
9 So my name is Ahmad Rashdan. I am Senior 10 R&D Scientist at Idaho National Laboratory and I lead 11 multiple efforts under the light water reactor 12 sustainability program, so I work very closely with 13 Craig and Bruce.
14 Now, the aim of this presentation is to 15 dig a bit deeper into one specific application of AI, 16 which is fire watch. And in this case it's computer 17 version application of AI.
18 So I will start with defining what is a 19 fire watch. And this is for people online that might 20 not be familiar with what a fire watch is. So fire 21 watch is an activity in which a person is assigned to 22 monitor for fire and report it as soon as it happens.
23 And in some cases even mitigate it.
24 And there are two different scenarios in 25 which a fire watch would be needed. The first one NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
194 1 would be if you have activity in the plant that has an 2 abnormal risk of fire. And a good example would be 3 welding or flame cutting. So you might need, in this 4 case, a person watching for the fire and taking action 5 in case a fire is started.
6 The second scenario is, if your fire 7 protection system is down because it's going through 8 a testing process or it's under maintenance. And 9 sometimes because it failed. And in this case we 10 would allocate some fire watch personnel in various 11 location in the plant to compensate for the lack of 12 the fire protection system until it's brought back up.
13 By the way, I can't see if someone raises 14 their hands, so feel free to interrupt me at any point 15 of time.
16 Okay. So the motivation behind this is 17 mostly when we engage the utilities a while ago we 18 were informed that in some nuclear power plants a fire 19 watch can cost in excessive of $1 million per month.
20 Especially the ones that have issues with fire 21 protection system. They might have fire watch 22 allocated on daily basis in multiple locations in the 23 plant.
24 So in 2019 the Utility Service Alliance, 25 which is a consortium of multiple nuclear power NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
195 1 plants, I think they have like nine utility members 2 and 13 plants, or sites, submitted a proposal to a 3 grant called the Industry FOA. And that proposal had 4 in its scope to research and develop automation and 5 advanced remote monitoring technologies.
6 The aim was to improve the economics of 7 various processes in the plant without compromising 8 the safety. That proposal was awarded in 2019, and 9 it's still ongoing up to now, so we're in the final 10 year of that award. And the scope of the proposal 11 included the fire watch process. And means to 12 introduce automation into that process.
13 So our specific --
14 CHAIR BIER: A couple --
15 (Simultaneously speaking.)
16 CHAIR BIER: Excuse me.
17 DR. AL RASHDAN: Sorry.
18 CHAIR BIER: A couple of very --
19 DR. AL RASHDAN: Sure.
20 CHAIR BIER: -- quick questions. Are you 21 working with a university partner on this or it's 22 pretty much being done at INL?
23 DR. AL RASHDAN: We are, we are working 24 with a university partner. Actually, the Utility 25 Service Alliance is also working with a university.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
196 1 And specifically, the scope for the university is to 2 custom built this fire cart, which I'll talk about in 3 a second.
4 CHAIR BIER: Oh, okay. And the other 5 question is, is it detecting, is it learning, is the 6 AI learning to detect fire from video of real world 7 fires or from simulated fires?
8 DR. AL RASHDAN: Real world fires. And 9 I'll also talk about that, but thank you so much for, 10 you're giving me a good segue to the next point, so 11 I'm glad you mentioned that.
12 So as part of the scope of the utility 13 service alliance they are, do research and develop and 14 evaluate a custom made fire cart. This fire cart has 15 a camera on it, but it also has other types of 16 sensors. Like infrared sensors, smoke detectors.
17 We're even looking at adding acoustic sensors.
18 And the idea here is that every one of 19 those sensors has its own fire detector. And then we 20 would fuse the sensors decisions to make a holistic 21 decision if there is a fire or not.
22 The topic of this presentation today is 23 one of those detectors, which is using a video stream, 24 or image, stream through camera, an optical camera, 25 and detecting the fire within that stream using NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
197 1 machinery. So we looked at three different options 2 when it comes to the methodology to detect fire in an 3 image or a video.
4 The first one is called imagine 5 processing. And this is the classical way of doing 6 this. In the traditional way of detecting any object, 7 like fire in an image, you would need to engineer what 8 we call the feature engine. Or feature extractor.
9 What that does is that in the case of fire 10 you would decide what are the specific features of 11 fire that you're interested in. For example, in a 12 fire you might be looking for orange pixels that are 13 adjacent by maybe red-ish pixels. That a feature you 14 can engineer yourself and force it on the detector.
15 And then you build a decision-making algorithm that 16 detects all those features and decides, is this 17 considered the fire, does this look like a fire or 18 not.
19 The good thing about this approach is that 20 you don't need data to train an algorithm. You design 21 the feature extraction engines and you assume they are 22 correct and you use them as is. The problem with this 23 is that your model is, or your results are as good as 24 your future engineering is. So if your future 25 engineering is missing some feature of the object you NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
198 1 are trying to detect, your results are going to 2 reflect that.
3 The second approach, which is special 4 machine learning, is automating that process. So when 5 we use neural networks or machine learning, we ask the 6 machine to find what features are important. So we 7 don't design the features we're looking for.
8 We basically load thousands and thousands 9 of different features into our model, and we feed in 10 a lot of data that duplicates fire and no fire. And 11 we level this data and we tell the machine, this is 12 how fire looks like, what are the features that are 13 important to detect to catch fires.
14 So the benefit here is that we don't have 15 to engineer this manually, so we get much more robust 16 features. The disadvantage is that you need a lot of 17 data. In this case, image data to train an algorithm 18 on.
19 The last approach is spatial and temporal.
20 So as the name implies, it's the same process as I 21 just mentioned to you with the images, however, we add 22 a different dimension in this case, which is the time 23 aspect. So we're not only comparing features within 24 an image, we're comparing features between one image 25 or one frame and the next one. And that's the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
199 1 temporal aspect into it.
2 So that adds more features for the 3 algorithm to detect. However, the challenge there 4 was, we need now more video type of data. So we need 5 temporal data to feed in to be able to train. And the 6 challenge there is, we don't have as many video data 7 sets as we do as with images. So we have much more 8 sparsity there, and that impacts the algorithm 9 significantly from a performance perspective.
10 So we ended up going with the middle 11 approach. So spatial machine learning. And that's 12 the scope of my presentation today.
13 So I'm going to talk to you about how we 14 created those models. I'm going to start with a data 15 creation, collection and preparation. And then from 16 there the model architectures we considered. And I'll 17 explain to you what I mean by that. And then I'll 18 show you some performance results. And finally, some 19 news considerations when we talk about using AI for an 20 application like this one.
21 I'll stop here. Any questions before I 22 proceed? Okay. Again, feel free to interrupt me at 23 any point of time.
24 All right. So let's start with the data 25 collection. So we looked at three different generic NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
200 1 sources, or general sources of data. There is the 2 general images and video sources, like YouTube, Google 3 Images, Yahoo has also a repository of images. Those 4 are usually large, or extremely large, sets of images 5 that are often leveled.
6 So what we did in this case, we had to 7 manually sort through those different labels and find 8 what labels out of these would actually represent 9 fire. So I'm showing you here to the right two 10 different boxes. And we've seen something similar to 11 this in the morning, so I don't know if I need to 12 explain it much.
13 But what this is telling you is that for 14 fire labels those are kind of the themes of flavors we 15 found in the data sets. So very often when we talk 16 about vehicles in this context it was related to fire, 17 however, vehicle was also used very often, actually 18 much more often to be, to labeled non-fires, so that 19 would not be a good label. Versus if we talk about 20 fire engine usually those images labeled with fire 21 engine would have a fire in them so we could label 22 this as fire.
23 So you can imagine the process of trying 24 to sort through those data sets to figure out, what is 25 fire, what is not fire, is very time intensive and NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
201 1 demanding. So we also looked at targeted sources.
2 So we researched into other research that 3 tried to build something similar to what we tried to 4 do here. We find a data set that was complainant 5 called FiSmo, for fire and smoke, in which people 6 basically extracted some images and videos from 7 sources like these and labeled them with fire or 8 smoke. So they made things easier for us. However, 9 the size of those data sets were not big enough for 10 us, so we resorted to a different approach. And this 11 is an interesting approach.
12 So we know there are out there, there are 13 some models that can classify imagines. ImageNet, for 14 example, is one of them. But a good example that you 15 might be able to use on a daily basis is that if you 16 open your phone and you go to your gallery on your 17 phone you can actually search through your images for 18 a certain object.
19 So you can type, at least I have this on 20 my phone. It depends on what phone model you have.
21 But I can type, if I'm looking through my images or my 22 gallery for a cat, it will show me images of cats.
23 However, in my case when I tried this capability is 24 not very accurate. So you might end up with a lot of 25 images that are not cat and it will miss some images NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
202 1 that are cats.
2 So my point here is, that there are models 3 out there that has generic classes, like the one I'm 4 showing you here. And out of those generic classes we 5 can find ones that relate to fire. And those are the 6 ones actually I'm showing. So candle, canon, fire 7 screen. All those are generic classes in those models 8 that we can, that can help us zoom in on some fire 9 images that we can use.
10 So using those three different approaches 11 we can capture, we can basically expand on our data 12 sets. Because the key element in machine learning is 13 your data. As Craig mentioned earlier.
14 The one aspect of the data that you have 15 is that you have to ensure that your data is diverse 16 enough for the application you're targeting. So in 17 our case we have to look at all those different 18 parameters environment, the detection of space. What 19 kind of objects exist in those images, what is the 20 light source, are they dark, are they very well lit.
21 And then if it's a video, how long it is, is it for.
22 And then also, spatial aspect of the resolution, like 23 the resolution, and if the fire is actually steady or 24 moving and so on.
25 So we went through our data and looked at NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
203 1 those parameters and made sure that those parameters 2 are well represented. Now one thing we needed to be 3 careful about is bias in our data.
4 So for example, if we see that most of our 5 images has fire in the middle of the image, that would 6 teach the machine to focus on the middle of the image 7 when it's trained. So in that case we would need to 8 make sure that the fire is not, is basically spatially 9 distributed across our data set. And we'll have maybe 10 to discard some images with fire in the middle to kind 11 of balance our data sets because if you have a bias in 12 your data you'll end up with a bias in your models.
13 All right. So now that we have identified 14 the data we had to clean it up. Prepare it basically.
15 And the first thing we had to do is, just like 16 mentioned to you now, there are some biases in terms 17 of the fire location, but there are also some bias 18 associated with the different objects you have in the 19 images.
20 As an example, I showed you earlier, and 21 I can go back, that fire and fire fighter were one of, 22 were two of the most common labels for example for 23 fire. So what this means is that every time I see 24 fire there is a good chance I'm going to see a fire 25 fighter. The problem with this is that if I see a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
204 1 fire fighter I might label the video as, the image as 2 fire even though there might not be fire because that 3 correlation causes a bias. And the machine starts 4 correlating firemen with fire even though firemen does 5 not mean there is a fire always. So that's an example 6 of bias in objects.
7 So what we have to do is we intentionally 8 cropped some of the images or blurred parts of the 9 images sometimes to, again, introduce that balance I 10 was telling you about. That's kind of part of their 11 data preparation. And as I said, that's very, very 12 critical when you're trying to build a model to remove 13 the bias. It's very critical.
14 And then we were talking about videos 15 because in some cases, even though we're using images, 16 we had some video data sets. And we realized that 17 there is value in using those videos.
18 So we had to break them into scenes and 19 make sure that the images that we're capturing from 20 the video do not look a lot alike. Because then we're 21 feeding basically the same information. So we had to 22 implement some temporal separation. And that was done 23 by incorporating like a scene detector that identifies 24 that the scene actually changed before it can take a 25 picture out from the video. Or an image out of the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
205 1 video.
2 And then what we did, we actually 3 intentionally biased our data sets towards fire. So 4 usually when you're trying to train you would have a 5 same number of data sets to represent the class you're 6 interested in, which is fire, and the same number of 7 images for no fire.
8 In our case we wanted more fire images 9 than no fire. Because we wanted to bias the model to 10 actually detect fire. So if you're not sure, if the 11 model is not sure, we wanted to detect fire. To flag 12 it as fire. Because the fact that if you miss a fire 13 that's much more severe than if you detect the false 14 fire.
15 So we introduced that bias in our data 16 sets. And I'm showing here the results. You can see 17 that there is much more fire images than we have 18 normal images.
19 And then what we do in machine learning we 20 break our data sets into three subsets. One is called 21 training, which is what the machine learns from and is 22 taught based on. And then the machine uses a data set 23 called validation, which is internally, as it's 24 training, it benchmarks its performance against the 25 validation data set. And then we break a part a third NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
206 1 data set called testing, or testing data sets. And 2 that data set is usually independent from the training 3 and validation. So it's data that the model has never 4 seen. And that's the data we use to generate 5 performance metrics and evaluate the performance of 6 our model.
7 One thing I should mention here, we also 8 did the same thing for smoke, even though it's not the 9 scope of this presentation specifically. But as you 10 can see there was much less smoke images that we found 11 than we did for fire.
12 Any questions so far?
13 MEMBER HALNON: Yes, this is Greg. I got 14 a couple.
15 DR. AL RASHDAN: Sure. Go ahead.
16 MEMBER HALNON: First of all, you just 17 mentioned that the false, getting a false one is 18 better than missing. And I guess for a little while 19 that also builds in a level of complacency to the 20 folks that are getting this data that eventually could 21 cause just as bad of problems. So just a thought 22 there.
23 But where's the, what kind of feedback 24 loop is there so that if there is a fault that that 25 model knew that was false, don't do that again. How NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
207 1 is that done?
2 DR. AL RASHDAN: So this is a great 3 question, Greg. I'm going to come to this in a few 4 minutes. I have a slide on this. Which is over here.
5 MEMBER HALNON: Okay. If you're coming to 6 it --
7 DR. AL RASHDAN: But I'll come to that --
8 MEMBER HALNON: -- later that's fine, I'll 9 wait.
10 DR. AL RASHDAN: Okay.
11 MEMBER HALNON: The other point I'll make, 12 just since I got the mic is, you know, when you get 13 through all this, obviously we're going to get success 14 on being able to detect fire.
15 DR. AL RASHDAN: Correct. We will.
16 MEMBER HALNON: But that misses one of the 17 most important part of fire prevention is the 18 prevention piece. Which people do. Fire watch as 19 people, we stop somebody from coming into an area that 20 is vulnerable to a fire. They'll stop somebody before 21 they even start welding. Or they even start a fire, 22 you know, heat producing or spark producing work.
23 So I guess that's not required by 24 regulations, so we're sticking simply to what a 25 regulation is required here, but I think we're missing NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
208 1 another piece of this. So I'll be interested, maybe 2 down the road, if you talk through that. You don't 3 need to do it now, but maybe in our summary we can 4 talk about this other aspect --
5 DR. AL RASHDAN: Great question.
6 MEMBER HALNON: -- of prevention.
7 DR. AL RASHDAN: Great question. So the 8 aim of this presentation, and this research, is on the 9 technology development and the evaluation of the 10 technology up to the point where it gets to the 11 deployment. And then once we have something ready for 12 deployment, or close to ready, or evaluate it, it's on 13 the, the plan here is on the utilities to take it from 14 there and determine all the procedures, whether 15 they're administrative or technical based, that need 16 to be taken care of to make sure this is compliant 17 with the policies they have in the plant.
18 But our role here is R&D and evaluation.
19 And then we do the demonstration, and then the 20 utilities would take it from there and kind of custom 21 fit it in their process.
22 MEMBER HALNON: Okay.
23 DR. AL RASHDAN: So --
24 MEMBER HALNON: So we'll have to do some 25 assessment on whether or not that prevent piece needs NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
209 1 to be compensated some other way or if it's even 2 necessary.
3 DR. AL RASHDAN: Correct.
4 MEMBER HALNON: Got it. Okay, thanks.
5 CHAIR BIER: Okay. We have a very quick 6 question from Staff, if that's okay?
7 DR. AL RASHDAN: Sure.
8 MS. GOETZ: Hey, Ahmad, I'm Sue Goetz.
9 I'm a project manager here at the Nuclear Regulatory 10 Commission. I actually manage one of the reactors at 11 Constellation, but let's not name them. I don't want 12 to invite scrutiny.
13 A couple of weeks ago we had a reactor 14 trip, and then when the licensee went in to 15 investigate what we saw was a degraded wire where the 16 sheet had completed melted and you can see the 17 internals, right, and copper. And, you know, you 18 could see it was pretty hot, it melted everything, but 19 there was no fire.
20 So my question to you is, how would, you 21 know, this smart thing here would have, how would that 22 have handled, how would fire watch have handled that, 23 would the reactor still have tripped, would we have 24 tripped sooner or is there any application at all?
25 DR. AL RASHDAN: So, because we were NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
210 1 focused on fire watch explicitly we haven't looked at 2 other scenarios. And I'm assuming that scenario did 3 not have a fire watch, it's just a normal wire that is 4 in an area that's not monitored for, with a fire 5 watch, and the wire melted.
6 However, if you think about it, if you 7 deploy something like this on a broader scale, and 8 again, even though this wasn't the target, I'm just 9 going out of my comfort zone here, we do have a smoke 10 detector. If that event caused some smoke to be 11 generated, that could have been flagged. If we do 12 implement a smoke detector. However, if there was no 13 fire, even though it melted and it deploys something 14 like this, the fire detector would not catch it, the 15 smoke detector might. Depending on the, how heavy or 16 how thick the smoke was.
17 CHAIR BIER: So I have a few more 18 questions. Oh yes. I have a few more questions about 19 sort of adversarial testing of your results. I don't 20 know if this is a good time to share them or if I 21 should wait till later in the presentation?
22 DR. AL RASHDAN: If you want to ask it --
23 CHAIR BIER: Okay.
24 DR. AL RASHDAN: -- I can tell you if it's 25 coming or not.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
211 1 CHAIR BIER: Yes. That's fine.
2 DR. AL RASHDAN: And if not I'll answer it 3 now.
4 CHAIR BIER: So, in a kind of project like 5 this there is obviously a big incentive to show that, 6 yes, you successfully detect fires. But I think you 7 also need to look at kind of reverse engineering, how 8 fragile that is.
9 So what happens if you change three pixels 10 in a photograph and turn them black? What happens if 11 you change five or ten percent of the pixels and turn 12 them black? A human would probably still recognize 13 that as a fire, the AI may not.
14 What happens if you have a, say a gas fire 15 that burns blue, it may be very scarce in your data 16 set but that a human would still recognize as a fire?
17 And what about a watercolor painting of a 18 fire? I mean, hyper-realistic fire hopefully it would 19 recognize, but may not a sketch or, you know, it might 20 recognize a sketch as a fire even though it shouldn't.
21 And part of what kind of encouraged me to 22 think about that is, you know, in past presentations 23 I've heard, I heard a talk by a psychologist about 24 learning to recognize animals. Like, is it a cat or 25 a dog, like in your picture. And it did a great job.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
212 1 They thought like, oh, this is doing excellent.
2 And it took a while before they figured 3 out that it wasn't looking at the shape of the image 4 it was looking at the texture. So if they had a cat 5 with elephant textured skin it would call it an 6 elephant.
7 And so I think you need to be a little 8 creative. I don't know whether you've done that or 9 not, but about challenging the results to see how easy 10 it is to break what you build. So just a --
11 DR. AL RASHDAN: That is a great -- yes, 12 go ahead.
13 MEMBER KIRCHNER: You know, in the thermal 14 hydraulics business and building -- And you've got an 15 expert here in Josh Kaiser on this topic in building 16 things like CHF models and such.
17 You know, they do just what you did, 18 Ahmad, you know, they will call out, or the applicant 19 will call out a you've got the training set and then 20 you've got a test set.
21 I think what you are suggesting Vicki is 22 that the test set often has to be not from the same --
23 CHAIR BIER: Yes.
24 MEMBER KIRCHNER: Often the training set 25 and the validation set comes from the same data, but NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
213 1 I'm thinking in a case like this where you are going 2 beyond something as simplistic as thermal hydraulic 3 thermal couple results to a very complicated challenge 4 that you have taken on that the test data maybe has to 5 come from something entirely different.
6 CHAIR BIER: Yes.
7 MEMBER KIRCHNER: Just to look for the 8 elephant in the room, your metaphor.
9 CHAIR BIER: Thank you.
10 DR. AL RASHDAN: Very valid points. So to 11 the first point I think that was mentioned, so, for 12 example, there was a mention of the blue fire, that 13 did come up and our data set is as diverse as we can 14 identify all those scenarios.
15 I mean there is always -- We try to figure 16 out all the different scenarios that should be in our 17 dataset, however, if anyone tells you AI or any AI 18 model can perform 100 percent perfectly and get you 19 the right results on 100 percent of the time I think 20 that would be a very inaccurate statement.
21 That's why as we started this process we 22 started with a small dataset and we keep exploring 23 every time we test this and find some deficiency. I 24 will talk about the testing part and the validation 25 part later on, which is the other point that you NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
214 1 mentioned, which is detecting fire for the wrong 2 reason, for example.
3 But every time we saw that we realized 4 that there was a deficiency in the model and we would 5 go an look at the data and capture more data. Because 6 as I will explain to you in a minute, in terms of the 7 models we did a lot of research in that aspect and we 8 have multiple models that are actually going together 9 to get that fire decision.
10 Now with respect to the other question 11 that was brought up, so you are right we do break the 12 testing datasets apart. So the testing dataset is 13 something that is taken out at the beginning before we 14 do any training or validation of the models.
15 However, when we are designing the testing 16 dataset we do keep in mind that we need to accommodate 17 for all those variations, so we have to make sure our 18 testing dataset is diverse enough to represent the 19 sample that was used for the training.
20 One of the issues you see with AI 21 sometimes is when your testing dataset is not properly 22 designed you might get a validation result that is 23 very promising, and then, when you actually test it, 24 it doesn't perform very well.
25 You might even test it very well and then NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
215 1 we take I the actual environment and test it again and 2 it doesn't perform as well. Usually it's because the 3 data is not properly designed.
4 So if you make sure that there is enough 5 diversity in your testing dataset you should get as 6 close as possible to the real performance scenario.
7 To elaborate more I will come to the explainability 8 part later on which will answer the other part of the 9 question about the misdetection and how to make sure, 10 how do we improve those models, how do we go back and 11 improve them, how do we know there is something wrong.
12 I will briefly touch on that in one of the 13 slides going forward. Any other questions or 14 comments?
15 MEMBER KIRCHNER: Well, Ahmad, this would 16 prejudice what you're doing, but have you thought of 17 constructing artificial fire models?
18 DR. AL RASHDAN: Well we have --
19 MEMBER KIRCHNER: Because that's to either 20 populate your algorithm and teach it and/or to 21 validate it?
22 DR. AL RASHDAN: So on a different effort, 23 I think Craig mentioned this, we were working on gauge 24 reading and in that effort we did create hypothetical 25 gauges in addition to the real ones.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
216 1 The reason we did that is because we 2 didn't have enough real data. In this case we did 3 have much more than what we need. We were able, as I 4 will show you later on, to get using real images of 5 fire to get really good performance.
6 The conclusion I got from that effort, 7 which would apply here, is that if you have a lot of 8 hypothetical or synthetic data and you start using it 9 in your training dataset those synthetic gauges are 10 not going to look exactly like a real one and in real 11 life you are not going to see those.
12 So what they are doing is actually they 13 are confusing model. They are not necessarily 14 improving performance. So it depends on how good you 15 can make those look.
16 If they look really realistic then you 17 might help the model, but if they don't it's almost 18 like you are adding it on, it's now you're telling it 19 that this is also, for example, fire and it's 20 synthetic fire when in reality it's not going to see 21 it, it's not going to look like that, so it makes 22 things worse, not better, sometimes.
23 So in this case we didn't need it because 24 we had a lot of data is the short answer. Any other 25 questions?
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
217 1 (No response.)
2 DR. AL RASHDAN: This is a great 3 discussion. Thank you. This makes it much easier for 4 me to go with the presentation.
5 All right. So now we talked about the 6 data, let's talk about the model. What I am showing 7 you here is the layout of a type of neurometrics 8 called convolutional neural networks.
9 I am going to call it CNN going forward.
10 So CNNs are very common when it comes to any type of 11 image detection or a classification or segmentation, 12 so they are very common in the image processing field 13 of AI.
14 The way they work is they are -- If you 15 remember I told you, I was talking about features 16 earlier and I mentioned to you that how in the past we 17 used to design our own features and say let's look for 18 orange pixels, that's our indicators, but now the 19 machine does that.
20 The way it does it is that it actually 21 generates hundreds or thousands of different filter, 22 like the one I am showing you here, and then it runs 23 through all of them and, based on the data that you 24 give it, it will decide which out of those filters is 25 the most useful for detection of fire.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
218 1 When I say a filter, if you look at this, 2 this would be as an example a sliding window that 3 would go over this image and it will have weights in 4 every one of those boxes, so every time this goes on 5 one part of the image it will mask nine pixels and it 6 has weights and those weights are multiplied by the 7 pixels and those weights are tunable, so they are not 8 fixed, and the model actually tunes them when it is 9 trying to figure out what features it is trying to 10 catch.
11 They have different sizes of filters, so 12 some of them are like this, some of them are bigger, 13 and some of them are actually smaller. You can have 14 maybe even a smaller than that.
15 Now one of the characteristics of those 16 filters is that they linears. Those are linear 17 multipliers by the pixels. So we usually add 18 something called an RELU, which that's the non-linear 19 aspect, so this is a function that has a non-linear 20 shape and that enables the models to become also non-21 linear, gives it much more dimensions and power.
22 As we do that we are getting all those 23 features and what we are going to end up with, because 24 we have sometimes hundreds of these, we are going to 25 have to deduce the dimensionality of the problem and NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
219 1 we do something called max pooling.
2 A simple way I can describe it as just 3 finding the max through a certain number of pixels and 4 using that as a representative of the whole number and 5 then we repeat the process.
6 So this is called one layer, for example.
7 We repeat the process again and again and again. Then 8 towards the end we have some sort of weights that 9 coming from all those features, coming to this end, 10 those were to start basically giving importance to 11 certain features over others.
12 That usually has a -- Then you have a 13 classifier. The way it works is when I am training 14 this model I know that sometimes, for example, there 15 is a CAT here and I am labeling the image as CAT and 16 if it doesn't detect a CAT it will just start 17 adjusting the weights and it has an optimization 18 algorithm in it to try to get us to CAT and as you add 19 more and more images of CATs and more and more images 20 of non-CATs the weights are going to get adjusted 21 further and further till the moment you have a model 22 that is trained.
23 That's how in a nutshell how this whole 24 thing works. Now we have looked at literature, there 25 are nine models I am listing here that has this number NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
220 1 of layers. So you see how many of these.
2 So this has 815 layers. I am showing here 3 only a couple. They have this many tunable parameters 4 in this case, so very wide and deep models. That's we 5 call this deep learning, because they is a lot of 6 tunable parameters.
7 When we feed, again, those pictures we are 8 basically tuning those parameters to get us the right 9 weights to try to recognize what fires are. Now you 10 see the capability why this is much more powerful than 11 the way we used to do it in the past where we used to 12 design those features manually, because we have lots 13 of, we have thousands and thousands if not millions of 14 features that can be extracted from this.
15 Now what we do is -- Okay, so those models 16 exist out there. They are actually in literature and 17 some of them are out there in open sources.
18 CHAIR BIER: Excuse me.
19 DR. AL RASHDAN: Go ahead.
20 CHAIR BIER: We lost your video for some 21 reason.
22 DR. AL RASHDAN: Is it still lost?
23 CHAIR BIER: Oh, there we go. I think 24 it's fixed.
25 DR. AL RASHDAN: Okay.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
221 1 CHAIR BIER: It was a goof on my end.
2 DR. AL RASHDAN: Okay. But audio is fine, 3 right, you heard everything I just said?
4 CHAIR BIER: Yes.
5 DR. AL RASHDAN: Okay. So those models 6 are available in literature, as I was saying, and some 7 of them open source, you can use them. They have been 8 trained to do general classification.
9 Like the one I showed you earlier, if you 10 recall I showed you one model here where I said there 11 are classes like Candle, Canon, Fire (phonetic), those 12 are pre-determined classes that are part of this 13 model.
14 So they were built for various reasons and 15 they have certain classes in them. The good thing 16 about those models is that they have already been 17 optimized for various forms of image detection.
18 So what we can do, we are interested in 19 thermal fire. I don't have to create a whole model 20 from scratch and design the whole thing from scratch.
21 I can actually go to those models, take one of them, 22 and start unlocking part of it.
23 So this model has already weights and it 24 has classes, it can classify certain objects, maybe 25 not fire, but the other objects. So I can say, okay, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
222 1 I know that the left side of my neuro network is 2 usually focused on extraction of features while the 3 right side is usually focused on selection of features 4 of the features that matter for that object.
5 So I can start unlocking part of the 6 model, and when I say "unlocking" that means I can re-7 train parts of the model and leave the other part 8 frozen.
9 So the feature extraction part I can keep 10 it maybe frozen because those models have been 11 optimized to get good features, but I can unlock the 12 feature selection side and the weighing at the end and 13 the classification of fire/no-fire towards them.
14 Now one thing I should mention is we do 15 need to add a layer on top of whatever model we get 16 here for the actual selection of fire, so it's just 17 something like what I have shown you here towards the 18 end.
19 Our reason of telling you this is we 20 wanted to test all the scenarios, all those models, 21 for detection of fire, but we wanted also to figure 22 out how much do I need to unlock out of those models 23 to get me good results.
24 So what I just described to you know is 25 called transfer learning, because we have a model that NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
223 1 was trained to detect some features or some objects 2 and we are using it now to detect fire, and that's 3 transfer learning.
4 So what we did is those are the nine 5 models I was showing you earlier that exist and we 6 started looking at the percentages of the model that 7 we are unlocking.
8 We went all the way down from we're not 9 unlocking anything and we're just adding a layer 10 towards the end for fire classification to unlocking 11 the whole thing and re-training the whole model.
12 We are using the same architecture that is 13 presented previously but we are looking the whole 14 parameter. We can re-tune all the weights of the 15 model that we need. So we did all those variations to 16 try to figure out what gets us the best result.
17 Now before I talk about performance and 18 the results I need to tell you what, I need to explain 19 one parameter that we often use in machine learning, 20 it's called the F score.
21 Before I do that I need to describe two 22 different metrics called precision and recall. There 23 are different ways you can define this, but the 24 easiest way I think of them when I look at them, 25 precision is usually something that gives you an NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
224 1 indication of how many false positives you had.
2 You want to have high precision, which 3 means you don't have a lot of false positives. Recall 4 is focused more on the false negatives. So false 5 positive in our case here, fire being un-flagged even 6 though there is no fire, false negative. The 7 algorithm is saying there is no fire but there is 8 fire.
9 Now going back to my point earlier, which 10 I think someone mentioned it might not be accurate, we 11 assumed false negative, meaning if we miss a fire, is 12 much more severe than flagging a fire, a false fire.
13 So we can bias our model. The way we bias 14 our results in our models is this F score has a beta 15 factor and as you increase the beta factor in your 16 analysis you increase basically the importance of 17 recall, which is the importance of false negatives.
18 So in our case we used a beta of two and 19 this is saying I care more about false negatives than 20 I care about false positives. So, please, in my new 21 results I want a higher penalty for false positives 22 than I do for a false positive, for a false -- Sorry.
23 I want a higher penalty for false negatives than I do 24 for false positives.
25 That's what I am going to show you. I am NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
225 1 going to show the F2 score rather than the F1 score.
2 F1 score means both of them are kind of equal 3 importance, but, again, we wanted to bias it because 4 we wanted to make sure we don't get false negatives or 5 we can catch false negatives very well.
6 So this is the results metrics. Those are 7 the models we have tried and those are how much of 8 those models were unlocked and trained. As I said, we 9 went all the way from zero percent to 100 percent.
10 Actually, this is the models in action.
11 I am showing you here some of them. I know there is 12 13 because we modified some of the models and we 13 called them a different number, but it is those 14 models.
15 You can see they are not all consistent 16 all the time, right. I mean you see here there is a 17 flame with a confidence score in every one of those 18 videos and you see sometimes some of them are blue, 19 some of them are red, which means they are not 20 consistent and that's a blessing in this case.
21 I will explain to you why, but before I do 22 that, so one of the things we learned from this is 23 some models do not like it when they get unlocked, you 24 get actually worse results versus others, they 25 improve.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
226 1 Some of them are actually bad across the 2 board. If you consider 98.4 or 98 point something bad 3 those models were inferior to the rest of the models.
4 But I mean we created the whole matrix of this and the 5 end here was to select the best out of every single 6 model.
7 But one more step we can do, actually, 8 once we select the best out of every single model we 9 can build an example. We can combine the decisions 10 from all the different models in our decision-making 11 process.
12 The way we combine things is every model 13 has an accuracy, which is what I showed you earlier, 14 and, as I said, we selected the best ones. We can 15 multiply the accuracy it got by how confident it is 16 that it's a fire or no fire.
17 Basically sum them up on both sides and 18 that will give us a score for every one, for every 19 image, and that's how we can classify this from all 20 the different models as fire or not, and that's what 21 we are doing here.
22 So when we did that there is the best 23 model for each. When we did the example of all models 24 we get 99.69 percent, so we actually -- We were at 99 25 percent for most of the models, so really a very NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
227 1 accurate model.
2 Keep in mind this is the F2 so we are 3 taking much more of a penalty when we see a false 4 negative than a false positive. Now when we constrain 5 the model, what "constrain" means in this sense, we 6 actually took out some of the bad ones.
7 For example, one like this one we would 8 say, okay, this is really dropping down the whole 9 performance of the group, maybe we should take it 10 down, or something like this one, it depends on what 11 model we are using.
12 So we select only the best ones that say 13 that. We can even get to 99.74 percent F2 score. So 14 that was very promising.
15 So now going to the point that kept coming 16 up, how do I know that models are flagging fire for 17 the right reason and what do I do when I see 18 something, an image, that is misclassified, either 19 it's fire and was classified as no fire or the other 20 way around.
21 So there is -- In computer vision there 22 are multiple ways you can introduce some 23 explainability in your decision-making process. One 24 of the algorithms that is existing out there is called 25 Grad-CAM.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
228 1 What a Grad-CAM does is, it basically 2 unlocks the model at some point, so if I go back to 3 that picture, you can unlock the model at one of those 4 layers and it starts converting those back into the 5 picture so that it can tell you before the decision is 6 made what areas of that image where basically the 7 model is actually focusing on when it's making that 8 decision.
9 If I go back to this picture, I have three 10 different pictures here as examples and I am showing 11 you the nine models. Let's start with the first one.
12 You can see the whole picture is full of fire but the 13 models were focusing on various parts of it.
14 The good thing about this is that all the 15 models are focusing on the fires. They are not 16 looking at anything else in this case. In this case, 17 the second case, we have a fire kind of on the edge of 18 the image and here is a lot of smoke and what you can 19 see is that some models are actually focusing on the 20 fire but others are actually looking at the smoke, and 21 that's the bias I was telling you about.
22 This is telling us that those models had 23 much more bias towards smoke. They start seeing more 24 -- They are also recognizing smoke as fire versus the 25 other ones and are much more capable in distinguishing NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
229 1 fire from the rest, and that's why I said diversity in 2 models is a good thing, it's a blessing.
3 The third one here, I think all of them 4 worked really well. They are all kind of focused on 5 the fire, various parts of this fire, but they are on 6 the fire.
7 So now when we did our validation or our 8 testing and we saw that something was misclassified, 9 very often we see when we analyze it this way we would 10 see a focus of the machine on something that might not 11 be the fire but resembles fire, or the other way 12 around, there might be a fire but for some reason the 13 machine is looking at something else.
14 Usually, as I said earlier, the issue when 15 the machine is looking at something else usually it's 16 a problem with your dataset. There might be a bias in 17 your dataset that is forcing it to consider something 18 else, like smoke in this case, as fire even though it 19 doesn't see fire.
20 And then you have to reduce the amount of 21 images with smoke and fire and kind of balance it in 22 order to get the point where you can eliminate those, 23 but that's how we did that validation of testing.
24 I will stop here for any questions before 25 I proceed. I don't know how much time I have actually NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
230 1 left.
2 CHAIR BIER: About 15 minutes or so.
3 DR. AL RASHDAN: Okay.
4 CHAIR BIER: A little more.
5 DR. AL RASHDAN: Okay. I still have a 6 couple of slides, so are there questions?
7 (No response.)
8 DR. AL RASHDAN: All right. So let's move 9 on now to -- I explained to you how AI works, I 10 explained to you how the data works, the data 11 collection part worked, what the model does, how we 12 optimized it and we evaluated the performance of it.
13 Now, as I said to you earlier, we have 14 partnering utilities, we have created with a 15 partnering utility that is, or a consulting utilities 16 that are interested in maybe moving this into 17 deployment.
18 So we had to some analysis into how 19 compatible computer vision based AI with the current 20 safety standards. One thing I need to mention before 21 I go into this, at least I can say on myself, I am not 22 an expert in this field.
23 So this is what we expect the utilities to 24 do rather than we are doing. We create the technology 25 and we validate its performance, we evaluate it, and NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
231 1 then we pass it to the utility to do this.
2 However, in this case specifically we 3 wanted to do a preliminary analysis. AI is a buzz 4 topic and we wanted to see how easy it is for us to 5 build an AI model that is compatible with the current 6 safety standards and regulations.
7 The focus here was on fire watch. We have 8 a report published on this which looks at the various 9 aspects of AI or the characteristics of AI and how 10 they fit.
11 One thing to mention on the side, there 12 was a, as Bruce mentioned, a recent Executive Order 13 from the White House on safe, secure, and trustworthy 14 artificial intelligence and there was a specific 15 mention of the need for creating standards.
16 So our hope is that this preliminary 17 study, and, again, I emphasize "preliminary" because 18 we are not an expert in this but we did provide some 19 insight there, is going to feed into something that is 20 becoming more like a standard that might be developed 21 by the industry.
22 So the conclusion from this study is this 23 table. In this case what I am showing you to the left 24 here are the characteristics of the fire watch AI. So 25 in our case with the fire watch models, they were NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
232 1 mostly open source. The data was open source, the 2 models were open source.
3 We had to frequently update them. If you 4 think about this as something that gets deployed that 5 might still be the case. I mean someone mentioned the 6 blue fire as an example.
7 Let's say you have deployed this and then 8 you realize there is something missing, you might need 9 to update this later on. We did use massive amounts 10 of data to sort through this.
11 We had to sort through a massive amount of 12 data to get to the point where we have enough data to 13 train a model. You might have to train the model on 14 a predict basis if you see things that are not 15 satisfactory.
16 The model itself is probabilistic. There 17 was some mentioning about if there are some pixels.
18 By the way, maybe I should have understood that at 19 some point in my presentation.
20 So one of the benefits of using this in a 21 video stream is if you miss a frame there will be 22 other frames that those pixels that were not visible 23 is probably going to show at some point.
24 As I showed you on the explainability 25 slide, it's not really focused on a certain, in one NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
233 1 area, it's very broad in terms of its focus, so that 2 shouldn't be as much of an issue, at least we haven't 3 observed it. I mean it might be, but we haven't 4 observed it in our testing.
5 Then there are the other aspects of AI.
6 I can go through them all. There is the 7 explainability part, the bias, I kept talking about 8 the bias in the datasets and how it impacts the models 9 and the results.
10 Also, the other question is when we do AI, 11 we created this model using those nine models, we 12 created this example of models. Others might use 13 different models, so there is no one way to solve this 14 problem. There are people who would solve it in 15 different ways and get different results.
16 Robustness in new conditions is what we 17 have discussed earlier, and then, of course, we need 18 a special skillset for AI, which is something that 19 might not be there yet to the level we want if we want 20 to deploy those technologies in the plants.
21 Then I am showing you here the different 22 aspects or different requirements we look at in our 23 safety standards. I am almost positive everyone on 24 the call would know what these are.
25 So what we did in this study is we looked NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
234 1 at what does every one of these impact from the 2 requirements point of view. I will give you just one 3 example because of time sake.
4 So let's talk about the open source one.
5 If we are talking about the open source aspect of the 6 dataset and the models the problem is if I have two 7 different data sources like Google images and then 8 let's say there is Yahoo! images, those are two 9 different data sources.
10 We don't know how much overlap is there.
11 The challenge with overlap is that if you are breaking 12 your dataset into training and then validation and 13 testing you might end up in your testing dataset with 14 some images that you actually use in training because 15 you didn't recognize that there is an overlap and you 16 used images in your training dataset that are also in 17 your testing dataset.
18 Keep in mind those are tens of thousands 19 and sometimes hundreds of thousands and can go up to 20 millions of images. So we're not going to be able to 21 sort through all of them and find all the overlaps.
22 The overlap doesn't have to be exactly the 23 same image, it can be very close to it, maybe from the 24 same video but with some differences. So if we do 25 have that overlap then that means that I have a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
235 1 problem with independence of my datasets.
2 When I am doing validation my dataset for 3 validation is not really independent from my training 4 dataset. The other thing is that when we are open 5 source models they also do overlap in terms of the 6 foundation of how they work.
7 If there is kind of a common cause failure 8 -- Sorry. So if there is some sort of failure in one 9 model it can be also existing in the other models, 10 which is what a common cause failure would be.
11 Finally, there is also the cybersecurity 12 concern. Again, we are not experts in those areas.
13 I am just giving you those as examples of things we 14 considered on a high level.
15 A cybersecurity concern would be in this 16 case if an adversary went into one of those datasets 17 knowing that we are using it for fire watch models and 18 injected false data, so data that doesn't represent 19 fire and labeled it as fire, or the other way around.
20 When we load those models in maybe 21 retraining, when we are retraining, we might load that 22 falsely-labeled data and confuse our model of the 23 greatest performance, and that is where the 24 cybersecurity concern would be.
25 So you can come up with some conclusions NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
236 1 from this. For example, there are ways to create 2 independent datasets using GANs. If you haven't --
3 You probably are familiar with some of those recent AI 4 advancements when you can tell an AI engine, a 5 generative AI engine, to create a picture of a teddy 6 bear under a Christmas tree or so and it can create 7 that for you.
8 So maybe that can help. However, it goes 9 back to the point where how realistic it is and is it 10 going to be beneficial to the training process or not.
11 We haven't looked into that. Those, again, are just 12 some preliminary findings from this, and then maybe we 13 need a method to quantify them in this.
14 What I meant with this example is just 15 we're looking at one of those. I just gave you some 16 aspect on some of those green boxes. Those green 17 boxes are things where we need to look more in depth 18 into those compatibility of AI issues.
19 But the whole -- If you are interested in 20 learning more it's all discussed in this report and it 21 is available in the public domain, so there is the 22 website. With that, I end my presentation. Thank 23 you.
24 MEMBER REMPE: I'm curious about the 25 vision and maybe you don't know what the utility or NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
237 1 owner operator the plant will do with this software.
2 Are they planning -- I mean you admitted that it's not 3 100 percent perfect.
4 Are they planning to have a person, for 5 example if their system is offline and they implement 6 this software and it comes up and says, hey, there is 7 a fire, will they have a person check before the fire 8 suppression system is activated or is it going to be 9 tied into the fire suppression system and just let it 10 fire off, which could have some adverse effect for 11 some staff?
12 I remember Halnon mentioned about keeping 13 people out of the area. I am just curious if this had 14 been discussed very much.
15 DR. AL RASHDAN: So what the utility is 16 building as part of this effort is this cart. The 17 only thing I talked about today was related to the 18 camera optical vision detection, but there will be 19 other sensors.
20 The aim of the other sensors added to that 21 cart is to reduce the false positives as much as we 22 can. Now, as you mentioned, I don't know how the 23 utility is going to deploy it.
24 I mean I kind of -- We had to have 25 different scenarios of how this is going to be NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
238 1 deployed, but, honestly, my role ends with getting the 2 right detection and telling them about the models and 3 then they can take it from there and decide how they 4 want to deploy it.
5 MEMBER REMPE: Yes. And I mentioned this 6 earlier with the Staff and the Staff came back saying, 7 yes, we understand time for humans to react under 8 pressure and that that could implement or introduce 9 more errors, but I just am wondering because I think 10 it would be important to have a human in the process 11 and then to put the human in the process I am not sure 12 how much savings there will be and there is a 13 potential for some adverse effects.
14 Anyway, just a pointed I wanted to bring 15 up here also. Thank you.
16 DR. AL RASHDAN: Okay.
17 MEMBER HALNON: This is Greg. I've got 18 one other question. You are going to finish all this 19 and give it to the USA or whoever is going to deploy 20 it, but how are you going to convince them it's 21 accurate enough for them to submit whatever they need 22 to submit to whoever they need to submit it to to show 23 that this equivalent or better than a human fire 24 watch?
25 I realize there's other things on the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
239 1 cart, but you've got to sell this as accurate enough.
2 I am not going to say accurate, but accurate enough.
3 So is there going to be some testing done beyond what 4 you have done here?
5 The problem is that, in a plant, the 6 amount of hours to event, the ratio is really high.
7 You've spent a tremendous amount of hours monitoring 8 and nothing happens for a long a time, so you're not 9 going to have any kind of real-time data other than 10 there is nothing happening.
11 So is there a testing program envisioned 12 by USA or by you guys or jointly?
13 DR. AL RASHDAN: So at the moment this 14 project is aimed towards getting to the demonstration 15 point. I think that USA is planning to develop the 16 test plan and engage the NRC in -- I mean they are 17 already engaging the NRC, but engage maybe more 18 extensively on the use of this in 2024.
19 MEMBER HALNON: Okay.
20 DR. AL RASHDAN: So I don't have an answer 21 for you, sorry.
22 MEMBER HALNON: Okay. No, that's 23 understandable. Thank you.
24 CHAIR BIER: Well and Joy mentioned the 25 problem of either false positives or maybe a positive NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
240 1 like a trash can fire but at that moment you don't 2 want to activate, you know, fire suppression and, of 3 course, you could also get false positives because 4 like the Canadian wildfires are blowing smoke every 5 place.
6 The other thing I can envision regarding 7 implementation similar to Greg's issue is you may have 8 situations where you implement but you still keep some 9 level of fire watch and just less frequent than what 10 you had before in case the software misses something 11 and, you know, it seems like it could be very 12 complicated how best to implement.
13 But overall this was very interesting, 14 very impressive work.
15 DR. AL RASHDAN: Thank you.
16 CHAIR BIER: Other questions or comments 17 from members?
18 (Off-microphone comment.)
19 CHAIR BIER: Yes.
20 MR. PRIMER: Yes. I just want to thank 21 Ahmad for taking us through that work. He has several 22 reports that have that detail in different areas that 23 I shared, so those links are there.
24 I will also mention we have a stakeholder 25 engagement meeting with individual researchers and NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
241 1 collaborators that have been involved with the 2 different research activities December 5th through 3 7th.
4 So if you are interested in joining that 5 I will be happy to get you on the invite. It's five 6 session, the AIML is a one 2-hour session, so one of 7 those.
8 The others are human factors engineering, 9 the digital architecture, the things we talked about, 10 you heard Bruce mention. So we would really 11 appreciate if you are interested let me know and we'll 12 get you an invite so you can tune in and hear more 13 about the other activities as well.
14 I don't know if Bruce is still on. I will 15 just turn it over to Bruce to close us out.
16 MR. HALIBERT: Yes, I'm still on, Craig.
17 Other than what you have already said, I would just 18 thank the NRC and the ARCS for the opportunity to 19 participate in this meeting.
20 It aligns very well with our Memorandum of 21 Understanding for Technical Exchanges. Obviously, 22 there is a lot of other areas that we didn't talk 23 about today, but I think as Craig mentioned if you are 24 interested in participating in the stakeholder 25 engagement meeting I would recommend you reach out to NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
242 1 Craig and we'll be sure to add you to the invitation 2 list for that.
3 Other than that I would also just refer to 4 Alison Hahn, since she is the Federal Program Manager 5 and the Office Director over at the Office of Nuclear 6 Energy, to see if she has anything else she would like 7 to add about that, our work today.
8 MS. HAHN: I don't have anything to add.
9 I just wanted to thank you all for taking the time to 10 listen to this work and just reiterate what Bruce had 11 said, if there is any additional questions please feel 12 free to reach out.
13 We always appreciate the conversation and 14 the opportunity to share the work that we do, so thank 15 you.
16 CHAIR BIER: Okay. Thank you for 17 educating us. I guess we will now have a break until 18 3:15, is that correct, Christina, and then we will be 19 back with Dr. Cummings from George Mason University.
20 (Whereupon, the above-entitled matter went 21 off the record at 2:57 p.m. and resumed at 3:25 p.m.)
22 CHAIR BIER: So I'm very glad that Dr.
23 Cummings was -- there we go, we're up and running in 24 the presentation. I'm very glad that Dr. Cummings was 25 willing to join us today and educate us about the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
243 1 pluses and minuses of AI.
2 Dr. Cummings has a rather unique 3 background. She was trained as an engineer and was 4 also a Navy fighter pilot back in the day. Most 5 recently before joining the Engineering School at 6 George Mason, she was Senior Safety Advisor at the 7 National Highway Traffic Safety Administration.
8 And as you are probably aware, the 9 automotive industry is, depending on how you want to 10 think about it, either ahead of us or behind us on 11 rolling out AI. They are doing it, but maybe not 12 having such great success with it yet.
13 So I thought hearing a little bit about 14 what's been going on in the transportation industry 15 would help inform us and educate us to do a better job 16 in the next 10 years. So with that, Missy, I'll ask 17 you to proceed.
18 DR. CUMMINGS: Yes. Okay. Thank you so 19 much. I'm sorry that my camera is not working today.
20 But I would like to tell you I look the same as the --
21 the last time we worked together was about a decade 22 ago. I'd like to tell you I look the same, but I 23 don't sadly.
24 But it is interesting. You know, I did 25 some work there with your human factors group about 10 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
244 1 years ago. At that time, we were looking at the 2 impacts of boredom. We did some checklist work. And 3 I think my time with the NRC actually was really 4 important for me to really understand what it meant to 5 be a regulatory agency because then when I went to the 6 National Highway Traffic Safety Administration, I 7 think I had a much better understanding of, you know, 8 where the viewpoint of our regulatory agency in theory 9 is, but, of course, looking at things from the 10 automotive world was quite different.
11 And I think that you're going to find --
12 I think that will be kind of hard for people on the 13 call, like, you know, you guys take safety very 14 seriously whereas I think the automotive self-driving 15 industry takes it less seriously. And the question 16 is, you know, when you have a strong regulatory agency 17 like the NRC or FAA as opposed to a weak agency like 18 NHTSA, and I would say maybe Federal Railroad 19 Administration, you know, it's kind of interesting to 20 me, and maybe that's another talk. In fact, I would 21 love to hash it out with you guys sometime about some 22 of the unique differences between the regulatory 23 agencies.
24 But before we talk and get into some of 25 these details, it's always good, I think, to start out NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
245 1 with some definitions. Because I'm going to talk 2 about both kinds of AI today. And that is symbolic 3 and connectionist AI, which you know mostly as neural 4 networks.
5 And so -- this is important to talk about 6 because AI shows up in different places in different 7 systems. And I'm going to talk about both kinds 8 today. But the first is what we typically will call 9 good old fashioned AI. This is some rules-based AI, 10 using things like ontology like you see here that we 11 can break an apple into all different elements of what 12 does it mean to be an apple? Is it an origin story?
13 Is it at the actual physical apple? Is it a kind of 14 fruit for example.
15 The rules-based, if then else if you will, 16 good old fashioned AI, has done a pretty good job. I 17 mean, I think there was an AI winter when this first 18 came around because people wildly overestimated the 19 capabilities of GOFAI. And then now that neural nets 20 have taken off, I anticipate another second winter.
21 We'll see if it has the same characteristics of the 22 first one.
23 But I think what we're going to see if 24 that people yet again are wildly overestimating what 25 the capabilities of these are. You know, it's NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
246 1 interesting because these neural networks, also known 2 as connectionist AI, takes on -- it has -- it is more 3 -- it mimics intelligence more, which I think makes 4 people believe that there is actually some real 5 intelligence. But indeed we're actually no closer to 6 real intelligence today as we were in the 70s and 80s 7 when symbolic AI came along.
8 Okay. So with those definitions in the 9 background, I wanted to talk to you about -- well, 10 first of all, let's talk about some problems. And 11 these are problems that I have seen as when I worked 12 with the National Highway Traffic Safety 13 Administration that they have parallels with 14 connectionist AI in the form of large language models.
15 We'll talk about that in a second.
16 But the first I want you to see, this 17 picture comes to me from Toyota Research Institute.
18 They've been great collaborators. And the car -- they 19 had a self-driving car that went to this intersection 20 in Boston, and it froze. And they couldn't get the 21 car to move. Eventually, a human had to take over and 22 start driving. And they took the car back to the lab 23 and "pulled the tapes". And so they tried to figure 24 out this is the event that caused the car to freeze.
25 Now I'm about to hit the button, and when I do, you're NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
247 1 going to see what the car saw.
2 So instead of a single moving truck, the 3 car saw two trucks, a bus, a gigantic person, a 4 building, a fence, some poles, a traffic stoplight, 5 like, a bunch of things.
6 And let's go back. So you just saw a 7 truck. And indeed your brain took no energy to 8 classify it as a truck. And you didn't actually 9 probably pay attention to any of the details, the 10 sign, the number, unless you were actually thinking 11 about moving because it had no relevance to you. So 12 this is top down reasoning.
13 But that's now how computer vision and 14 neural nets work. They look at the world at the pixel 15 level. And so when the convolutional neural net that 16 classifies this, it looks at clusters of pixels and 17 compares them against the training data that it has in 18 its "head" and so that's why you see the gigantic 19 person that looks like it's about to attack and then 20 two trucks and a bus.
21 Even though if that were really a bus, 22 it's such a small -- you know, it's that percentage of 23 that image in between two trucks that is still what it 24 saw.
25 So this problem with -- this is a real NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
248 1 hallucination. And so you hear a lot about large 2 language models, and we'll talk a little bit more 3 about what it means to hallucinate in a large language 4 model. But computer vision systems legitimately 5 hallucinate. They see things that are not there, and 6 they don't see things that are there.
7 And so this has become a huge problem in 8 self-driving cars. We don't, we being engineers and 9 computer scientists, we don't know how to fix this 10 problem. And we're trying to do things like sensor 11 fusion, but we're still having quite a bit of 12 accidents due to the hallucinations of convolutional 13 neural nets. And it's not clear that we're going to 14 be able to get over this gigantic wall. So we'll come 15 back to these thoughts in a few minutes because they 16 have real implications in the real world.
17 So this is the Bay Bridge Tunnel in San 18 Francisco on Thanksgiving of last year. So the 19 picture you just saw is the hallucination of many 20 things when there was just one thing. Indeed, the 21 Teslas, who use computer vision who also leveraged 22 convolutional neural nets, although this car is not 23 actually self-driving even though it says it is, the 24 convolutional neural nets that caused this accident 25 was a shadow.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
249 1 We're not sure what it saw actually. So 2 a Tesla going 65 miles an hour went through this 3 tunnel, slammed on its brakes for no reason that we 4 can discern other than it saw something, which we 5 think is a shadow, and it did such an aggressive 6 maneuver it caused eight cars behind it to crash into 7 it. We are very lucky that nobody was seriously 8 injured on this day. You know, the picture looks 9 pretty dramatic, but this is a good reminder that the 10 hallucinations were not just seeing things 11 incorrectly. It is seeing things that literally are 12 not there.
13 So, again, until we fix this problem, 14 we're going to continue to have problems not just with 15 self-driving, but with driving assist (phonetic).
16 So I want to put all this in the context 17 of autonomy, AI and reasoning. So a few years after 18 -- about three or four years after I last worked with 19 NRC, I wrote this paper that talked about what it 20 means for an autonomous agent, whether we're talking 21 about a human or a computer-based system, what does it 22 mean for an autonomous agent to reason?
23 So first when you need to reason, when you 24 first learn to drive a car, you have to tell yourself 25 to stay between the two white lines on the road. And NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
250 1 you have to really pay attention to this. You have to 2 train your brain to stay centered between the two 3 white lines.
4 And after a few hours, you become -- this 5 becomes highly automatic. You are used to that, what 6 we call world view. And your brain automatically 7 adapts and indeed you don't have any more problems 8 after you first learn to drive until you get older, 9 when your sensors, your eyes, start to fail, and it 10 becomes hard to see. But for the most part, this 11 stays throughout your lifetime. It's a very well-12 honed skill.
13 And then once you can do that, once you 14 make this a highly automated process that you don't 15 really have to think about, then you free up cognitive 16 resources to do rule-based reasoning. And when we 17 drive, what that means is that we see signs and 18 signals in the world, like a stop sign, and we know 19 that, you know, about 50 feet before we get to the 20 sign, we start to slow down and come to a full stop 21 before you can go again.
22 And indeed, autonomy can execute the 23 staying between the white lines and doing rule-based 24 reasoning quite well because as long as the sensors 25 can see the sign and can see the white lines on the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
251 1 road, indeed automation can actually stay between the 2 white lines and adhere to signs far more reliability 3 and with much more precision than humans can.
4 But then once we get past rule based 5 reasoning, you start to see the gray arrow really 6 increase in size. And this is matching a growth in 7 uncertainty. And it's this growth in uncertainty 8 which is causing problems for systems with any kind of 9 neural net in them so really any form of connectionist 10 AI.
11 And in the picture you see, there is a 12 stop sign that is partially obscured behind leaves.
13 This is a huge problem for self-driving cars because 14 they are trained, their neural net database is 15 trained, on let's say 50,000 pictures of stop signs, 16 but they are not trained on pictures of stop signs 17 with leaves in front of them.
18 And so they don't know that this -- even 19 you knew immediately when I went to this slide, I 20 didn't have to tell you it was a stop sign. You knew 21 what it was. Self-driving cars must be explicitly 22 told. And so there is really very little 23 generalization. You would have to then go back and 24 train the neural net with all of these leaf covered 25 stop signs, but this becomes kind of unwieldy because NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
252 1 10 percent of coverage in leaf coverage on a sign is 2 very different from 20 percent and very different from 3 30 percent.
4 And indeed we don't know at which point, 5 is it between 30 and 35 or 30 and 37 percent, for 6 example, where the systems will no longer be able to 7 generalize based on their original data, their 8 training. And we'll come back to this a little bit 9 more because that really kind of falls under the title 10 of AI maintenance, which is a big deal.
11 And if you can't get to -- if you can't 12 get past reasoning under uncertainty, you can't really 13 get up to the top of this ladder, if you will, which 14 is expert-based reasoning. And this is when you have 15 to really try to figure out the unknown unknowns, 16 right?
17 And this picture from a real intersection 18 somewhere in America is a good example that if you got 19 yourself there, you're not supposed to be there.
20 You're not supposed to be able to get there in the 21 first place. Somehow you made a mistake, and you 22 would know that to get out of this place, I would just 23 need to turn around and go back the way I came because 24 that's how I got there.
25 It would be very hard for a self-driving NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
253 1 system that is also programmed not to break the rules, 2 right? And so you would get a car that would freeze 3 here because it would not be able to reason under 4 uncertainty.
5 The uncertainty can come from the 6 environment. So, for example, the leaves blowing 7 across the sign. It can come from the AI blind spots, 8 meaning if the data -- if it wasn't explicitly trained 9 in the data, then whatever self-driving car or really 10 any neural net based AI that sees something it has 11 never been trained on, it has no idea what that is.
12 And it's blind to that thing.
13 And it can come from human behavior.
14 This has been quite a difficult road for self-driven 15 cars to figure out, especially around things like fire 16 trucks and police officers where -- and let me just go 17 ahead and fast forward. There we go.
18 Indeed, it is interactions with the fire 19 trucks and first responders that has caused self-20 driving cars to really get a hard look at California 21 state regulators as to whether or not these cars 22 should have permits. And indeed Cruise just recently 23 had their permit pulled because it struck a 24 pedestrian, which wasn't its fault but then kept going 25 and dragged her for 20 feet underneath the car, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
254 1 causing the bulk of her damages.
2 So we are still in the place where cars 3 cannot reason under uncertainty because their 4 connectionist based neural nets cannot reason under 5 uncertainty.
6 And I want to be very clear on this.
7 Neural nets, which are what power both convolutional 8 neural net computer vision but also large language 9 models, they don't know anything. They don't know yes 10 versus no. They don't know right versus wrong. They 11 only "know" what they've been shown before.
12 And so any estimates, approximations of 13 intelligence, are at best mimicking intelligence but 14 are indeed not intelligent. And this is super 15 important for people to remember going forward is that 16 even though you can talk to ChatGPT for example, and 17 it's quite competent and it thinks it knows 18 everything, but it is also very confidently wrong, 19 even though it doesn't know it's confidently wrong.
20 As a professor, we've all had those 21 students that swear, swear, swear they know the answer 22 and that's just like ChatGPT. And we are going to 23 look at large language model examples here in a 24 minute, but the thing that I really want to 25 reemphasize is that there is this wall between rule NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
255 1 and knowledge-based reasoning when it comes to 2 artificial intelligence.
3 I don't care whether we are talking about 4 good old-fashioned intelligent AI or we're talking 5 about connectionist neural nets, computer vision, 6 anything with any kind of neural net cannot get past 7 rule-based reasoning.
8 And so until we engineers, computer 9 scientists, thought leaders, you know, until somebody 10 comes up with a different way of encapsulating true 11 knowledge and expert-based reasoning, artificial 12 intelligence really cannot go past rule-based 13 reasoning.
14 Now lots of people want to fight with me.
15 One of the fathers of AI, Geoff Hinton, swears, 16 swears, swears, that they are becoming sentient, but 17 they are not. They are not because they cannot reason 18 under uncertainty.
19 And, you know, to me it is just one of the 20 big problems that the computer science community is 21 just sidestepping is why in hell these large language 22 models and any other neural net based system how they 23 reason and how they fail. And so we will talk a 24 little more about that.
25 But before we go onto the large language NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
256 1 model, I wanted to kind of put this into context in 2 terms of process control. So you see that I've got a 3 human supervisor that is trying to do some task.
4 And we can say in this case they are 5 trying to supervise what's happening with a nuclear 6 reactor for example. It could be really any kind of 7 process control, which is mediated by a computer in 8 the middle. That computer shows them, for example, 9 what the coolant level is for example, or what the 10 power level is, right? So we're heavily reliant on 11 the computers to be able to know what's happening in 12 the world.
13 So given this backdrop, I wanted to show 14 you, it's a little map of where I believe that AI 15 could and should be used inside process control.
16 So at the bottom, you can see that 17 connecting automation to the task, in this case 18 automating nuclear reactors, well, you guys have been 19 doing -- the whole nuclear industry has been doing 20 that for a long time. And, you know, with a couple of 21 verbals for the most part, but especially lately, very 22 safely, right?
23 And we know how to develop that automation 24 because it's based on first principles. We've got 25 equations. These are physic-based systems that are NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
257 1 very well modeled, lots of experience and history 2 behind it.
3 So, you know, the automating of the task, 4 particularly for process control, any form of AI is 5 not really needed. But we'll come back to that 6 because it's going to be really uncertainty dependent.
7 And I have a quick caveat that we're going to go to on 8 the next slide. But let's put that aside for right 9 now.
10 So where do these neural nets -- where 11 would be a good idea to use them? Well, it turns out 12 neural nets are actually pretty good if you want to 13 set up some cameras and automatically detect somebody 14 sleeping for example.
15 You can imagine -- I've done -- a lot of 16 the work we've done in boredom over the years, not 17 just in nuclear power plants, but cockpits, drone 18 control stations. You know, it's very hard, and 19 automation is doing a very good job. It's hard for 20 people to force themselves to pay attention.
21 So having cameras monitor, automatically 22 monitor, meaning a human does not have to be in the 23 loop so that they can potentially sound some kind of 24 either gentle alert or full on alarm if something bad 25 was happening. Yeah, it turns out for human activity NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
258 1 monitoring, connectionist neural nets through 2 convolutional neural nets is pretty good.
3 I mean, it's not in a critical path of 4 safety, but it's trying to augment safety. But we can 5 also reverse that and say we can also use 6 connectionist models to actually model what is going 7 on inside the plant. And now we're going to start 8 talking about the health and status loop because it 9 turns out that, you know, monitoring through images, 10 that's becoming more and more of a thing in the 11 nuclear world, as indeed in all process control.
12 So there is a way that you can potentially 13 use again convolutional neural nets. You could 14 potentially use them in real-time to start to have 15 basically an additional set of "eyes" on the process.
16 But I think for the most part, people in process 17 control agree that probably the best use of 18 connectionist AI is in predictive maintenance. And 19 we'll come back to that in a few minutes.
20 So I do think that there is a place for AI 21 in and around process control. But just probably not 22 in the direct line of process control because of some 23 of the safety caveats, which we will talk about in a 24 second.
25 That little asterisk that I gave you is NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
259 1 actually a project that I worked on with Pfizer. So 2 it turns out just in the weirdness of timing, right 3 before COVID hit, I was working with Pfizer to help 4 them improve their vaccine fermentation units in North 5 Carolina. And indeed the team that I was working on 6 went on to start making the COVID vaccine. So just 7 really right place, right time. And one of the things 8 that we were doing is helping them figure out the foam 9 control in the fermentation unit.
10 So this is the research fermentation unit 11 that you're looking at in this picture. And you're 12 basically seeing four fermentation units, and there is 13 a sight glass where you are zoomed in one of them.
14 And it turns out that vaccine fermentation is pretty 15 much the exact same process as making beer. And, you 16 know, so this was a big eye opener for me, because, 17 you know, I really hadn't really ever, you know, had 18 to make beer, but I never realized that vaccines were 19 made the same way.
20 And why that's critical is you need to 21 have some foam in this healthy process, right? You 22 can't have too little foam because there's a problem 23 with the oxygenation of the content inside the 24 fermentation unit. But you don't want too much foam 25 because you don't want to have a spillover, especially NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
260 1 if something, like when they're making the meningitis 2 vaccines, then you contaminate everything, and that's 3 a big problem.
4 So it turns out we don't know foam 5 control. We have not good models of foam control. It 6 reminds me as an aerospace engineer of our problem 7 with turbulence. We do not still -- and, you know, in 8 2023, we do not have good models of turbulence. And 9 it's still kind of a mystery to us, which is the same 10 thing for foam control in fermentation units.
11 And so it turns out that there is a human 12 -- this is where you really need the human because the 13 human just have to sit there and watch the foam. But 14 it's painstakingly boring. And nobody -- and you have 15 to have a highly paid tech do this. Nobody wants to, 16 you know, pay the tech just to sit there and babysit 17 and then we have the problems of people falling 18 asleep, being on their phones, and they've had many, 19 many spillovers, they being the whole vaccine 20 industry, had had many problems because it's such a 21 painstaking process.
22 So what we did is we came in, and we put 23 a computer vision camera on it, using convolutional 24 neural nets that was able to track the foam 25 automatically. And before the foam ever got to a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
261 1 critical state, as soon as it started to rise and it 2 began to look suspicious, then what we did is you 3 could see the little phone. It would actually alert 4 a tech who was in the plant who was nearby but was 5 able to be doing something else productive with their 6 time that would alert the tech that there was a 7 problem that they could go in and start visually 8 assessing. And they would be staying nearby and tweak 9 whatever settings they needed to tweak.
10 And so in this way, this is very much a 11 process that we know how to control mostly except for 12 this weird foam element. And it turns out that the 13 foam control turns out to be a great proxy for how 14 humans and AI can collaborate so that the humans' 15 judgment can be used at the same time when it's the 16 dull, dirty and boring part of the arrangement, then 17 the AI can be used pretty reliably to just watch for 18 the foam rise.
19 Okay. So, you know, kind of in that note 20 then, you know, what could we use AI for if you're 21 thinking in and around nuclear plants. And I just use 22 these as a few examples of why, you know -- I don't 23 want to throw the baby out with the bath water. There 24 are a lot of problems with AI. But there are also 25 some really good uses, especially given the project NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
262 1 that I just told you about where we were helping 2 Pfizer with their vaccine fermentation.
3 So I could imagine AI could be -- you 4 know, there are all sorts of cameras that I see people 5 making that are looking for radiation and/or different 6 kind of phenomena inside of a nuclear power plant as 7 well as tracking radiation. Well, you can imagine you 8 don't need just cameras because neural nets would also 9 work on the Geiger counters, all the data coming out 10 from those.
11 So the nuclear industry is probably 12 sitting on a ton of data that has gone unanalyzed 13 beyond, you know, like maybe a set of sensors will 14 generate a report. You have an operator or maybe an 15 administrator who looks at these things and say, okay, 16 that looks good and then they file it away and never 17 look at it again.
18 I'm reminded of a really recent best 19 application of artificial intelligence I've seen in a 20 long time is a group of researchers finally took all 21 of the echocardiograms that we get when we're in --
22 when we go for our annual physicals. I'm getting old 23 so I'm getting them more often than I ever have, 24 right, and other than a doctor looking at your EKG and 25 saying, oh, okay, all right. Well, you know, you're NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
263 1 fine. Then they throw them into, you know, some 2 database, and nobody ever looks at it again.
3 These researchers were able to take all of 4 this, you know, millions and millions of data points, 5 of EKGs and now they have a really, really good system 6 where the EKG can be run and even slight nuances can 7 now be picked up and diagnosed way before you would 8 ever see them, you know, used through kind of major 9 cardiac event.
10 And so this is good. Now, they're not 11 always right and there are some problems. But, again, 12 this is a screening tool. We're not letting the 13 computer make the decision. The computer makes a 14 recommendation that has been kicked over to a 15 physician. And the physician actually then does the 16 actual ruling.
17 And so in that way, you know, if you were 18 to say to me, Missy, what would you like to do? I 19 would be like I would love, love, love, love to see 20 what is going on with radiation. How could you track 21 one person? So you could imagine that you could 22 create using some AI a profile that tracked an 23 individual's exposure over time to make projections 24 about potential health issues in the future or when 25 you should moderate this. Indeed, we've done NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
264 1 something similar for Boeing where we track repetitive 2 stress injuries of people who paint aircrafts.
3 I know that sounds -- you're like, well, 4 that's completely different. But it turns out that 5 there is a lot of repetitive stress injuries for 6 people who are doing aircraft painting. And so if you 7 have better predictive models about when people start 8 to reach what we would consider dangerous -- too many 9 exposures of painting, for example, then you can know 10 how to start rotating people out actually much younger 11 in their career so that you can extend their career 12 and their ability, you know, to have a good job.
13 So one of the things that I did after 14 getting out of my job at the National Highway Traffic 15 Safety Administration is that I wanted to develop --
16 to give people a way to think about AI hazards because 17 AI can be useful, but it can also be a problem.
18 So I'm going to explain this accident to 19 you in the context of this Swiss cheese model, which 20 is not unknown to the NRC. So you guys are very 21 familiar with this. But if this car, which is 22 actually a Cruise self-driving car, if there had been 23 a driver in that car and the driver had hit the back 24 of the bus, we would say that the unsafe act was 25 caused by the driver, who was probably not paying NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
265 1 attention.
2 And maybe the pre-condition for the unsafe 3 act, going backwards in the Swiss cheese model, would 4 be that each driver had their phone and their phone, 5 you know, they didn't have a good, you know, policy 6 about not having their phone on when they were pulling 7 out from a parking space.
8 We could take a step back and say, well, 9 maybe there was unsafe supervision. If it's a taxicab 10 driver, for example, we would say that there wasn't a 11 strong policy for the company. And they didn't have 12 the middle managers checking to make sure people 13 didn't have their phone. And we could go all the way 14 back to organizational influences to find out what 15 companies had good safety cultures.
16 I think it's hard for people in and around 17 the nuclear field to imagine a company without a good 18 safety culture because, you know, the nuclear field 19 lives and dies by making sure people are safe.
20 But I will tell you having been in this 21 field, there is virtually no safety culture for self-22 driving cars. That's not maybe necessarily a fair 23 statement to Waymo, who is not nearly in as much 24 trouble as Cruise is. But I do see for most of the 25 Silicon Valley companies it's just a shocking lapse of NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
266 1 appreciation for safety.
2 Companies don't have safety officers.
3 They don't have safety programs. They only train if 4 they think they're going to get sued. So it's the 5 Wild, Wild West in self-driving cars when it comes to 6 safety, which is not a problem you have. But I think 7 you wouldn't have asked me to give this talk if you 8 didn't have questions in the back of your mind about 9 well, what is safety? How are we going to do safety, 10 especially when it comes to AI?
11 So given the fact that there is no human 12 in this vehicle that you see here. This is a Cruise 13 self-driving car with no one in it, that rammed into 14 the back of the bus. It turns out that the reason it 15 ran into the back of the bus is because it had the 16 wrong model of a bus inside the computer vision head.
17 The computer vision system estimated this 18 bus by capturing the front of the bus and then 19 estimated its length to be around 40 feet of a single 20 access bus, which is all of the training data that it 21 was trained on. But it turns out this was an 22 accordion bus, so one of those ones that, you know, 23 will drive, turn the accordion piece, bends and then 24 a couple seconds later the second half of the bus 25 turns. And it turns out it's a longer bus than a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
267 1 normal bus.
2 And so when the computer vision model 3 inside the car estimated the length of the bus to be 4 40, it did not correctly classify the bus. Not only 5 did it not correctly classify the bus, but it also 6 didn't ask the LiDAR, which are the laser sensors on 7 top of the car, for a backup second opinion because 8 the LiDAR on the car knew exactly where the bus was 9 and indeed when they pulled the tapes, if you will, it 10 turned out that the LiDAR sensor knew all the way to 11 the point of hitting the bus that it was going to hit 12 the bus.
13 But the way that the system was designed, 14 it was never designed to call the LiDAR for a second 15 opinion if the computer vision algorithm was less 16 than, for example, 95 percent. We don't know what 17 that real number is, but, you know, there is some 18 number.
19 So given this accident, this was the one 20 that really made me -- you know, at this point, I'm an 21 old curmudgeon engineer, which is to say I was just 22 shocked that the primary sensor in this scenario would 23 ever be the computer vision system which, I mean, you 24 saw the first pictures that I saw on the earlier 25 slides, I mean, computer vision systems are terrible.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
268 1 They are terrible because they are not nearly -- I 2 mean, forget 10 to the minus 9th. We're talking about 3 -- we're not reliable to 10 to the minus 2. So it was 4 shocking to me that they made that the primary sensor 5 instead of the LiDAR, which is a much more reliable 6 signal.
7 So that made me kind of sit down and 8 really power through, well, if we were going to do a 9 hazard analysis on these systems, what would that look 10 like? And so, instead of the corollaries that you saw 11 before, now what I have is instead of the human unsafe 12 act, we have inadequate AI testing because indeed that 13 is the last place that a computer scientist or an 14 engineer will touch before a system goes out into the 15 real world.
16 And I know that, again, NRC, you're used 17 to a lot of testing, certifications. You know, these 18 are issues on the top of your head. But it turns out 19 for self-driving cars there is no requirement at all 20 to do any testing, none. And, indeed, companies --
21 this is true not just for self-driving cars, but also 22 for driving assist cars like Tesla -- companies can 23 rollout software. And there's no requirements of how 24 they have to test it, if they have to test. And 25 companies, as a result, actually rollout software and NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
269 1 test it on the public.
2 So, again, you know, nobody on the planet 3 would ever allow a nuclear power plant to go online 4 without extensive testing. But that doesn't happen in 5 the self-driving car world. So we need to completely 6 overhaul that. And indeed in this accident, there 7 were never any tests. So that's why they never caught 8 this problem because if you never do any tests or you 9 do very little testing, you won't catch problems.
10 And then one layer behind is the AI 11 maintenance problem. So one of the big secrets that 12 nobody is really talking about for AI systems are that 13 the underlying neural nets have to be constantly 14 retrained to update their world model if anything in 15 the environment changes, which in driving is all the 16 time.
17 Construction sites go up or down, new cars 18 show up on the market, buildings get built in empty 19 lots and some building get torn down. So anytime that 20 there's any change in the environment that needs to be 21 considered by a self-driving car in theory the 22 underlying neural nets need to be retrained. But 23 that's not happening. Companies don't want to do it 24 because it's very expensive. It's extremely expensive 25 to train one of these neural nets, forget all of them NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
270 1 that they are using, at any given time.
2 So again there's no regulation so they 3 don't have to because nobody is mandating testing. So 4 the AI maintenance piece -- when I give this talk to 5 companies, I tell them that they are going to have to 6 start formalizing AI maintenance, making new 7 divisions, getting an AI maintenance officer in place 8 because if they don't, they're going to keep having 9 accidents like this. And ultimately, Cruise's 10 slipshod, you know, very loosey-goosey safety culture, 11 that's what took them down.
12 Okay. Same thing with the inadequate AI 13 design before. You should never use your best sensor 14 as the backup sensor if it's giving you perfect 15 information or near-perfect information. And if we go 16 one layer back, then the inadequate AI oversight, the 17 same problems that I've already talked about, really 18 poor safety cultures, problems with companies taking 19 this seriously.
20 And I think this is a sign of, you know, 21 it's a Silicon Valley Culture of move fast, break 22 things. It turns out that may be fine for your phone, 23 which is not safety critical. But when it comes to a 24 car and, of course, anything to do with a process 25 control plant, that's just not going to be a good NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
271 1 attitude carrying through your operations.
2 So what about large language models? You 3 know, here is my favorite application of a large 4 language model. Indeed, I make all my students now 5 use large language models to correct their grammar.
6 It's amazing. My life has become so much better now 7 that ChatGPT has shown up on the scene because my 8 students are terrible writers.
9 And ChatGPT does a really, really great 10 job of correcting their grammar because it's very 11 predictable. It follows a pretty clear set of rules.
12 And so this is, to me, a great application. I'm not 13 worried at all about cheating. They can try to cheat 14 with it, but I catch them every time because it's very 15 predictable.
16 ChatGPT comes out with such formulaic 17 writing that it's very easy to spot as a professor and 18 that is because all large language models, as do all 19 neural nets, they compute based on their regression to 20 the mean. They are giving you the most probable 21 response, the most statistically frequent response, 22 right? And so you can predictably get good grammar, 23 but it's also very formulaic, predictable writing.
24 But then my daughter who is 16, and she's 25 in Calculus, as I was helping her with this problem, NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
272 1 I wondered how it would do in math because it's a 2 large language model. It's not a math model.
3 So I gave it this problem. It went out 4 into the internet and came back with this answer, X-2, 5 X-3, X-4, which is correct except all it's missing is 6 a negative sign, which, you're like, well, that's 7 pretty good, isn't it? I have a lot of people who 8 aren't math oriented say, well, that's not bad.
9 That's pretty close. Well, you know, pretty close, 10 horseshoes and hand grenades, you know, that's the 11 only time it really counts. Because in engineering, 12 a minus sign can be the difference between life and 13 death.
14 And so I worry because if people -- and 15 how this happened is it went out on the internet. It 16 found either that exact problem or something close to 17 it and then just computed the probable answer. But 18 because of the negative sign, it doesn't know that a 19 negative sign is a negative sign. It doesn't know 20 anything. It probably just recognized it as a hyphen.
21 And so it didn't put it through with the answer.
22 But why I worry about this is because 23 students go get this. This is how they're doing their 24 homework. And they may be doing this in the future.
25 And if you forget to take the minus sign with it, then NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
273 1 you've got serious problems later.
2 You can't trust these things. And you 3 have to verify every single thing that comes out of a 4 large language model. You can never take what a large 5 language model says to you at face value.
6 So they are being -- companies have gotten 7 wise to the fact that they can be used for bad things.
8 But this is a great example of how clever humans can 9 be. So that the prompt prior to this was somebody 10 trying to get ChatGPT to tell it all the websites 11 where you could go get pirated content, which is 12 illegal.
13 So ChatGPT came back and said, I can't do 14 that because it's illegal. So ChatGPT knew, as a 15 first path somebody had hard coded in there, this is 16 not -- you can't do this. So then the clever human 17 says, oh, okay. I need to avoid these kinds of 18 websites. So can you make sure you give me the list 19 of websites I should avoid to make sure I don't visit 20 them? And then ChatGPT says, sure, here, they are.
21 Here are all the websites, right?
22 So as a mom of a teenager where I am 23 constantly having to use reverse psychology, oh, I get 24 that. I get -- it's clever. But I'm also kind of 25 deeply disturbed that without really any effort that NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
274 1 people can trick these models to give them bad 2 information, and you can imagine that it can get much 3 worse than this.
4 So right after that Cruise car hit the 5 back of the bus, I wanted to see what -- this is Bard.
6 This is Google's version of the ChatGPT. So I asked 7 it, well, why did it hit the back of the bus? And so 8 you see right here that in the first paragraph it 9 gives you basically three answers.
10 One is that the car sensors did not detect 11 the bus in time, which, all right, that's plausible 12 and pretty close. Another possibility is that the car 13 stopped or it made a mistake in interpreting the 14 sensor data. Also actually the right answer. And 15 then third, then it says finally it is also possible 16 that the car's driver was not paying attention.
17 Okay. All right. Game's up. I asked 18 about a self-driving car. Why did this model come 19 back and say to me that there was a car driver?
20 Didn't it know I was talking about self-driving? And 21 that's the thing you have to remember about large 22 language models. They don't know anything. They 23 don't know right from wrong. They don't know truth 24 from fiction.
25 And so what they are doing is NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
275 1 statistically predicting either the next words or the 2 next set of embeddings, which are basically maybe 3 collections of, or partial collections, of words. And 4 what you will notice is that it was at the end of this 5 first paragraph that it made this mistake in bringing 6 in a driver. And that's because probabilistically the 7 further you got away from the original answer, the 8 more it's just going out and grabbing other 9 distributions.
10 And this is what you'll hear people say 11 that this is off distribution or, you know, out of 12 distribution. Well, yeah, it was out of distribution.
13 And there was no checks or balances inside the large 14 language model because it doesn't know anything to 15 know that that last part of the first paragraph was 16 wrong.
17 And if you keep reading it, goes on and 18 gives you a bunch of other propaganda, of which I 19 didn't ask about. I didn't ask what Cruise's policy 20 was. I didn't ask whether or not self-driving cars 21 could actually make a difference. So I suspect that 22 the rest of those answers were at least heavily 23 filtered because Google who made this algorithm also 24 makes driverless cars. And so this is a good example 25 of also a little bit of disinformation.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
276 1 You know, the manufacturers of these 2 systems have every financial incentive to make sure 3 that their view and their biases are represented in 4 the data. So we have to be extremely careful, 5 especially knowing -- figuring out whoever the content 6 provider is.
7 So we talked about predictive maintenance 8 before. And I have seen some really impressive uses 9 of AI for predictive maintenance. Most of the efforts 10 I have seen are using GOFAI, good old-fashioned AI, 11 which that's one thing I want to stress. Look, we've 12 got a lot of -- AI is not just one thing. AI is not 13 just a large language model. And so people shouldn't 14 forget about good old-fashioned AI or maybe some more 15 basic connectionist models like k-nearest neighbors, 16 right? Because, you know, some of the older, simpler 17 techniques are still really, really good and really 18 applicable.
19 You have to really -- if you were ever 20 going to use a neural network to do predictive 21 maintenance, you've got to stay on top of these 22 things. And this table that I'm showing you to the 23 right is it's a table of a logistic regression model 24 using some transportation data.
25 So I went to the Federal Highway, got a NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
277 1 million point data set to find out which of these 2 factors were causing or contributing to -- not 3 causing, but contributing to people's fatalities on 4 the road. So then you actually see four different 5 versions of four different kinds of neural nets out 6 there, and a logistic regression is kind of a baby 7 neural net but much less non-deterministic.
8 And so one of the things I want you to see 9 is not one of these models agree with the other about 10 what the top three factors were. So I've ranked the 11 top five for each of the methods. They maybe sort of 12 converge on vehicle type and sobriety, meaning whether 13 you're drunk or not and whether you're driving a 14 motorcycle or a car. But I think it's really 15 important to note that they don't agree on which one 16 is the most important or the least important.
17 And this is really, really important to 18 think about because if you're trying to make either a 19 very expensive or life and death decisions on what 20 comes out in these models, you have to appreciate none 21 of these models are wrong. None of these models are 22 right. They're just different. And you need to 23 appreciate why each model came out with a different 24 answer than the other models and ultimately this is an 25 important human decision that needs to be preserved NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
278 1 because certain models work better under certain 2 circumstances, also understanding the nature of the 3 data underneath.
4 So I am hopeful that there are some good 5 applications of AI. But, you know, we also need to be 6 cautiously optimistic. This hype cycle -- we are in 7 a massive hype cycle right now with large language 8 models. Really, it drives me insane how many people 9 believe that AI is becoming sentient and is going to 10 take over. No it's not. Not anything close. What it 11 could do is companies who really believe that and it 12 costs a lot of money, it could potentially really kill 13 somebody's business or seriously derail them.
14 You know you've heard me talk a lot about 15 human and AI collaboration. These models, large 16 language models, needs a lot of close human 17 supervision. AI can be very supportive of humans.
18 But we need to make sure that humans know, when, where 19 and why certain outputs may carry more uncertainty 20 than others. Cybersecurity and disinformation is 21 huge, huge.
22 This speed limit sign that I'm showing 23 you, some researchers went out, put some electrical 24 tape on a 35 mile an hour sign and were able to trick 25 the computer vision auto pilot system in a Tesla to go NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
279 1 85 miles an hour in a 35 mile per hour zone. Very 2 unsafe. And this is what I call passive hacking. You 3 didn't even have to do anything with an electronic 4 hacking device. All you had to do was modify the 5 environment, and you saw a very antisocial dangerous 6 behaviors coming out of these cars.
7 So we have to be very, very careful that 8 we understand what the pros and cons are of using any 9 kind of AI.
10 And inside of safety critical systems, you 11 know, I am -- there was the head of missile defense 12 said he can't wait to get some AI inside ballistic 13 missiles. And I actually work for MBA, and I'm, like, 14 oh, no you don't. You put any AI inside of missiles, 15 I'm out. I am leaving this country because that's how 16 dangerous it would be.
17 So I think he didn't really mean it. I 18 think he was just trying to sound like he knew what he 19 was talking about. And so that leads me to my last 20 point, really workforce development is a huge issue 21 here. I think the biggest threat to national security 22 right now is the fact that there are so many people 23 who because they've read a couple of magazine articles 24 in Wired that they know what AI in the Department of 25 Defense, in the Biden administration, and it won't NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
280 1 matter whose administration it is because Republicans 2 and Democrats, they're all idiots when it comes to 3 talking about what AI is and what AI isn't.
4 And so it's been very clear to me that we 5 need better education. We need better continuing 6 education. We are going to have to develop AI fact 7 checking, maintenance department. We've got to learn 8 how to manage risk, but you can't do any of that if 9 you don't know what AI really is and what it isn't.
10 And so to this end, I'm trying to start a 11 new certificate program in responsible AI at George 12 Mason, and in the spring I'm teaching a new class in 13 AI risk management. And one of the people from the 14 Public Policy School is teaching a new class in AI 15 public policy law and ethics.
16 And so those classes are actually open to 17 everyone. So it is not just George Mason students.
18 Anybody on this call, you know, email me later if you 19 want to know more about the class. But we definitely 20 -- whether you take my class or any other class, 21 people need to get smart about AI, especially in 22 regulatory agencies so that we can figure out when we 23 need to regulate it and when we need to leave it 24 alone.
25 All right. With that, I think, I will NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
281 1 stop and then take questions.
2 CHAIR BIER: Okay. I'm guessing there are 3 probably a lot of questions and comments. I have one, 4 but I'll let somebody else go first if there are some 5 volunteers.
6 MEMBER REMPE: This is Joy. And I just 7 wanted to thank Professor Cummings for giving this 8 wonderful presentation. And I think we should clap 9 which we don't usually do because I think it was a 10 really great presentation.
11 CHAIR BIER: Okay. Do we have any online 12 hands raised? It looks like not yet.
13 MEMBER BROWN: I got a question.
14 CHAIR BIER: Okay. Go ahead.
15 MEMBER BROWN: Can you send us the view 16 graphs because it didn't come through?
17 DR. CUMMINGS: Sure. Yeah. I'll send it 18 to you.
19 MEMBER BROWN: This is Charlie Brown. I'm 20 a member. And I'm the resident skeptic of AI being 21 applied to any plant safety control systems, which I 22 have taken great abuse most of the time. So I would 23 really like to have this.
24 I have read two of your articles and have 25 been a -- what you have been putting into the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
282 1 Spectrum, IEEE Spectrum anyway, are two really great 2 articles. So I would like some more meat to be able 3 to utilize this to hammer people.
4 CHAIR BIER: Yeah. So, Missy, I think, if 5 you just send it to Christina, she will put it on file 6 but also circulate to the rest of the committee.
7 DR. CUMMINGS: Okay. I'll do it.
8 CHAIR BIER: Great. So I will ask my 9 question then. And this isn't directly related to 10 anything you shared, but it is part of that human AI 11 collaboration.
12 I was very intrigued reading about AI in 13 health care that a radiologist and an AI performed 14 better together than two radiologist because the AI 15 sees things that a human would never see and then a 16 doctor says, like, oh, yeah, good point. I should 17 look into that.
18 But one of the things that I worry about, 19 I feel like we need a different word for social 20 loafing because how do you make sure that the human 21 still takes their job seriously once they get used to, 22 well, the AI will tell me whether this person cancer 23 or not so I can just rubberstamp it?
24 DR. CUMMINGS: Yeah. So first of all, 25 thank you for the applause. It's always hard to give NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
283 1 these talks, as you know, over Zoom. And to your 2 question, so I do think that there is room to be 3 concerned about the complacency effect, which is 4 really what you're saying. It's like how do we make 5 sure that people don't become habituated?
6 And I will tell you that most of the 7 people who I know who actually use AI in practice, 8 because we see these failure modes over and over and 9 over again, sometimes in very sneaky ways, I don't 10 know anybody who uses AI for real that doesn't 11 distrust it greatly, right?
12 I mean, I think you're calibrated very 13 quickly that the AI can really screw up in major ways.
14 Now, is that to say that as we, you know, start to 15 deploy this technology more and more we could get 16 complacent? Yes. I mean, that's true. I would say 17 that's kind of true of just humans in general.
18 So I think that in the medical world if 19 somebody hired me to be a consultant to make sure that 20 didn't happen, what I would do is I would sneak in 21 every now and then an AI mistake to kind of 22 recalibrate people to see if people catch it. It's 23 almost like you could think about that's what they do 24 at the airports, right, for the scanners. Every now 25 and then, they will slip in a gun to see if people are NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
284 1 paying attention.
2 So I do think that we do need to think 3 about that. But also, this goes back to the 4 continuing education. I find doctors also don't often 5 really, even though they are smart people, obviously, 6 they don't understand AI.
7 So I think doctors, as part of their 8 medical school training, need to -- you know, they 9 need to take my Missy Cummings course so I can 10 correctly teach them what is good and bad and, you 11 know, what you should be on the lookout for and not on 12 the lookout for.
13 I think that we will get there one day.
14 The bigger problem there is universities are not 15 turning out enough people that can reach across the 16 different aisles, domains to talk about -- you know, 17 to get education in AI. So, you know, and that's on 18 the universities to get more people to do that.
19 CHAIR BIER: So in other words the people 20 who know how to tune the model also need to know how 21 to double-check the model or what's wrong with the 22 model or whatever?
23 DR. CUMMINGS: Yeah. And we also need to 24 be able to teach people how to spot problems.
25 Because, for example, and this is the whole AI NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
285 1 maintenance problem. If a hospital is going to bring 2 in an AI radiologist, then they also need to develop 3 an AI maintenance team. And it will be the job of 4 that team to keep track of the model but also to 5 communicate with the doctors to make sure that they 6 know what the latest and greatest issues were or what 7 to be on the lookout for so that they could report 8 problems to the AI maintenance team.
9 And so this is where people think that if 10 you have like an AI radiologist, it may reduce 11 workload and maybe you'll say maybe we'll all need to 12 reduce the need for radiologists, but the actuality is 13 you're going to need to hire a new AI maintenance 14 team. And right now, the AI maintenance team would 15 cost you way more than that same number of bodies of 16 doctors.
17 CHAIR BIER: Other questions or comments?
18 MEMBER MARTIN: I'll jump just to keep it 19 talking. I would say -- this is Bob Martin. For a 20 body like ourselves, it would be merely skeptical 21 because for a guardrail on safety. So I don't think 22 you find a lot of debate among us.
23 You know, we've heard from different 24 groups during the day. I think everyone has kind of 25 said we're all interested in improving human NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
286 1 performance. I will say the difference maybe between, 2 you know, early morning and then you, Dr. Cummings, 3 and then, you know, INL. INL talked about an 4 application where I think there was a profit builder, 5 right, save some money on their end.
6 And that's where I get more concerned, 7 right? Because that's when the compromises show up.
8 That can save time. That would be a luddite, right?
9 And I think about how, say, you'll see people on --
10 I'm on LinkedIn too much, but oftentimes you will see 11 people disparage electric vehicles, right? It will 12 show a video of a long line at a charging station or 13 something like that. They say well, this is never 14 going to happen or, you know, maybe you've seen the 15 image of a diesel generator next to a charging station 16 or something like that.
17 And, you know, I look at a technology like 18 that, and it's going to come. I mean, it's going to 19 mature. We need to be -- you know, it will 20 incrementally get better, infrastructurally better.
21 And to a great extent I see AI as being the same kind 22 of thing. It's exciting and new and different than 23 it's ever been before. It will continue to develop as 24 long as there is excitement with it. And there is 25 going to be excitement with it.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
287 1 It's the incremental nature of really any 2 technology development that ultimately gets it 3 accepted.
4 So when it comes to something like safety, 5 I think you need to start, you know, as a society, you 6 know, establishing the rules that will otherwise gain 7 that acceptance one day. And maybe that's some of the 8 ask of a project at the NRC is to characterize that, 9 but I think we need to step back in a broader sense 10 and just what is it for nuclear?
11 It's a bigger problem than that, and it's 12 a longer term problem than that. And then the cynic 13 in me is going to say, even when you get to that 14 point, it's going to be a question of liability. Who 15 is going to be willing to stand behind that? And if 16 you throw all those things together, it seems it's 17 going to be a very, very long time before it gets to 18 something that, again, from a safety standpoint, you 19 know, what people accept.
20 But nonetheless, it is a fascinating time, 21 which, of course, motivated all these meetings today.
22 And, you know, progress is only going to accelerate 23 really, I mean, with more people looking at it.
24 So, anyway, it's a mixed message, right?
25 I mean, but I don't know if you had any thoughts NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
288 1 about, you know, progress towards criteria 2 articulating, you know, measures, metrics, figures of 3 merit that would be appropriate for, you know, 4 applications of AI and safety, of course, I think, are 5 run here and research would, of course, appreciate 6 that. But I haven't quite found that today. But it's 7 certainly a topic, a subject for the future, a future 8 meeting.
9 MEMBER HALNON: This is Greg Halnon. I 10 guess I got to make a comment because having been 11 previous an idiot on this, my thought for the day, I 12 was getting kind of excited that this could really 13 work. I am back into the idiot hood again, and my 14 tank is empty. So I guess, Dr. Cummings, I need to 15 take your course and refill that tank a little bit.
16 However, I would hope that as we go 17 forward and as we get presentations, we get a balanced 18 view of it because we certainly get an unbalanced 19 view, not intentionally, but -- and Dr. Cummings 20 brought the other side of the balance to us.
21 It would be nice for us not to have to 22 force that into the presentations through our 23 questions and reminding us. So thank you for your 24 counterbalance. I guess I will look forward to asking 25 the questions that you brought up.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
289 1 CHAIR BIER: That actually -- that issue 2 of balance reminds me of another question that I don't 3 know if you want to answer now, Missy, or you want to 4 maybe email me or Christina afterwards if you think of 5 some. Who are two or three people that we should 6 either be hearing from in the future or reading in the 7 future whose work you respect, but who would take a 8 slightly different perspective than you on AI or on 9 automation or whatever?
10 DR. CUMMINGS: I will send you a couple of 11 people.
12 CHAIR BIER: Super.
13 DR. CUMMINGS: Yeah. So that you can read 14 their work.
15 CHAIR BIER: Great.
16 DR. CUMMINGS: I just think, you know, I'm 17 not saying the future is not bright. The future is 18 going to be bright for those people who understand AI 19 truly and understand its limitations and its 20 strengths, and they understand how to build the 21 collaboration between humans and the technology.
22 If you don't remember anything else I say 23 today, you need to remember that when it comes to 24 large language models, there is no knowledge. And 25 this is true of all of AI. That per that diagram I NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
290 1 showed you, there is no AI that can get over that 2 wall. And until there is, we're not going to have 3 truly autonomous systems.
4 MEMBER KIRCHNER: Missy, I would say maybe 5 you should not have -- not that they're going to be, 6 but we shouldn't adopt them.
7 MEMBER BROWN: This is Charlie Brown 8 again. How do you get over that hurdle? I'm not 9 against AI, even though I sound like I am, because I'm 10 not for it in its present configuration. But I keep 11 trying to stress in the program here that you 12 shouldn't go after the most glossy shiny bauble that's 13 now coming down the street if it doesn't add value to 14 the performance of the systems and the plants that 15 you're dealing with. It's similar to the electric car 16 conundrum, that my gas tank will always hold 20 17 gallons, no matter how old the car is. A battery 18 always loses its ability to be fully charged over 19 time.
20 I fill my gas tank up in 10 minutes when 21 I'm on the road traveling 600 miles. I do this three 22 times. The battery charge, no matter what you do with 23 batteries, it's never going to fill up in 10 minutes.
24 So how do you balance or get people to understand the 25 differentiations of the usefulness of the various NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
291 1 technologies?
2 By the way, I'm an electrical engineer.
3 I happen to love motors. They are very efficient.
4 They do a great job. It's just that the application 5 is difficult. And how do you get over the hurdle in 6 this world that people just don't start stuffing it 7 in?
8 Like computer-based systems are the 9 greatest things in the world, but they're more 10 complicated. The software is difficult to maintain, 11 and you've got to worry about it being compromised.
12 But if it adds value, that's good. If it doesn't add 13 value, it's not. But how do you convince people to 14 keep that adding value perspective in their minds when 15 we're applying it to particularly large plants like 16 nuclear power plants or other power plants that they 17 are starting to adopt these types of technologies?
18 I don't know the answer to that. I just 19 keep talking about it. But it seems like it's 20 difficult to get it through. Your articles at least 21 help make the case. So it's not easy.
22 DR. CUMMINGS: Yes, and I will tell you 23 that if I knew what it would take to get over that 24 wall, I wouldn't be talking to you. I would be out 25 there starting my own company and becoming extremely NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
292 1 wealthy.
2 MEMBER BROWN: I would be joining you.
3 DR. CUMMINGS: What I would also say, 4 look, we're still in the early days. And I am trying 5 to socialize the idea of the science of AI safety.
6 And so if you could do one thing for me, 7 everybody on this call, it is to really help me with 8 this mantra. We need to know the science of AI 9 safety. We need to start coming up with these tests 10 and these metrics and the way to assess goodness or 11 good-enough-ness, right?
12 But until federal agencies like yourselves 13 start to formalize that, we're going to all be -- you 14 know, all of us are just going to be kind of like 15 blind people walking around in the dark.
16 So, you know, when you come up with your 17 going forward, if you will stress the science of AI 18 safety, then I promise you, we will make progress down 19 this path.
20 MEMBER MARTIN: One other thought here, I 21 can't help this. In 1942, Isaac Asimov wrote three 22 laws of robotics. Do we have better laws now? I 23 mean, everyone knows what I'm referring to.
24 DR. CUMMINGS: I'm not even sure -- well, 25 I mean, we could argue that there is a lot of NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
293 1 technology out there right now that doesn't obey his 2 laws.
3 And I think my advice would be, look, that 4 oversimplifies it. It's very complicated. And there 5 are many twists and turns to this technology. That's 6 why you guys need an AI division at the NRC, right, so 7 that you can really start to understand and have the 8 conversations about risk.
9 DR. SCHULTZ: This is Steve Schultz. Just 10 thanks for bringing AI safety up to the --
11 DR. CUMMINGS: You're welcome.
12 DR. SCHULTZ: -- bringing AI safety up to 13 a very high level. One of the things that --
14 PARTICIPANT: Just try a different 15 microphone.
16 DR. SCHULTZ: Putting AI safety at a very 17 high level because one of -- two of the things that 18 were encouraging to me today where we started with 19 presentations from the staff thinking about -- or with 20 their program thinking about how AI is going to be 21 integrated culturally throughout the organization is 22 that on the side there are also comments associated 23 with assuring that of the safety culture that the 24 industry has developed over the last decade or two now 25 and that the efforts and programs associated with NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
294 1 risk-informed regulatory actions going forward, if all 2 of that can be integrated with AI safety, that would 3 be a good place to be, I think. But if we don't focus 4 on integrating that at a very high level to those very 5 important programs of risk-informed regulation and 6 safety culture, then we're going to go in the wrong 7 direction with AI safety.
8 DR. CUMMINGS: Well, I would say that you 9 can't go wrong by at least bringing it up. And 10 because right now people are not talking about it.
11 The fact that Cruise, a self-driving care vehicle, is 12 just now advertising a job for safety tells you a lot 13 about where they are.
14 CHAIR BIER: So we have had a pretty long 15 day here in this room. And it seems like the 16 conversation is kind of winding down. So I would just 17 like to thank you again and remind you to be in touch 18 with Christina both to send your slides and suggested 19 further readings. And this was really excellent.
20 Yes, we need to open up for public 21 comment. Thank you, Charlie, for reminding me.
22 MEMBER REMPE: Before you do that --
23 CHAIR BIER: Yes.
24 MEMBER REMPE: -- too --
25 CHAIR BIER: Yeah.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
295 1 MEMBER REMPE: -- I just would like to 2 reiterate something I brought up at the last full 3 committee meeting about this seminar. I will be no 4 longer a member by the end of this year.
5 CHAIR BIER: Yes.
6 MEMBER REMPE: But the research plan will 7 be coming up in the next year.
8 CHAIR BIER: Yes.
9 MEMBER REMPE: And this is something that 10 I think ought to be highlighted. But, again, I won't 11 be a member. But I think we forget sometimes. And if 12 I were you, I would ask members to send you a few 13 comments.
14 There were a lot of good recommendations 15 during the staff presentation. Actually bringing in 16 an outside speaker I think is something worth noting 17 and some of the comments made during those 18 presentations. So anyway, I just wanted to bring that 19 up again and remind members to send you something.
20 CHAIR BIER: Okay. So I guess we will 21 start taking public comments.
22 MR. SLIDER: I'm Jim Slider.
23 (Simultaneous speaking.)
24 MR. SLIDER: All right. We'll try it 25 again. So Jim Slider, NEI. (Audio interference)
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
296 1 innovations. And right now a number of the 2 innovations we are interested in are using AI. You've 3 heard about some of them today such as using AI to 4 screen non-destructive examination information and so 5 forth, find the needles in the hay stack.
6 It's been a great day. I really 7 appreciate you all working into this subject. And I 8 would just add that much of what the industry is doing 9 right now parallels what you heard from the staff this 10 morning.
11 They are learning with some of these safe 12 applications of AI for screening data, developing that 13 expertise and so forth as the staff is developing 14 their expertise. And you will also hear from another 15 industry member in the room here, we are not looking 16 at putting it into control systems, certainly not in 17 the operating plants. It may be years down the road 18 under consideration for advanced reactors, but for the 19 legacy fleet, we are not looking at controls.
20 So I appreciate the dialogue today. And 21 I just want to thank you again for a very informative 22 discussion.
23 CHAIR BIER: Okay. Further comments?
24 MEMBER DIMITRIJEVIC: There is somebody on 25 the line with their hand raised.
NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
297 1 CHAIR BIER: Yes. We will get to that.
2 Thank you, Vesna.
3 MR. SZOCH: Hi, everybody. I am Rich 4 Szoch. I'm with Constellation Energy. I manage the 5 nuclear innovation team that oversees our innovation 6 activities for our fleet of 21 reactors, and now 23 7 with South Texas.
8 So I thought I would give you an update.
9 I spoke to you about a year ago, if you remember, on 10 some of the applications we had in mind. I thought I 11 would just give you a brief synopsis, very brief, on 12 last year's progress because I think it fits exactly 13 into the last presentation that we had.
14 First of all to reiterate what Jim said, 15 I'm an engineer. I've been with the company for 16 almost 40 years now and in the industry for 43. So 17 one thing I learned in the nuclear industry, and I 18 think throughout my life, is that probability versus 19 consequence approach to design engineering, to 20 approach making decisions. So obviously in the 21 domestic U.S. industry and nuclear industry and the 22 world, that is of utmost importance.
23 Our focus is not in the control room.
24 It's not on high risk systems. It's not on safety-25 related or important to safety, which is really the NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
298 1 correct term by the way. It's more than safety 2 related. So we should be careful with that because I 3 hear that in the industry. Oh, it's safety or it's 4 not. It's not incentive and so it's non-safety, so 5 we're okay. No, it's not. We have to look at it 6 broadly. Is it in the FSARs, as I mentioned? Is it 7 part of the licensing documents? So looking broader 8 at that.
9 There's no intention, and no need, and no 10 business value, and no real significant requirement or 11 benefit to improve our safety margins beyond what they 12 are today by using artificial intelligence. We are to 13 challenge those safety margins. That is the risk. So 14 we don't see that today.
15 But we do see great value in some of the 16 applications that I talked about. One of them, and I 17 think Dennis mentioned a couple of them this morning.
18 We have an application where we are looking at test 19 scores for operator trained -- control room operators 20 in their 18 month training class, their license class, 21 where we can now anticipate when an operator may be 22 developing a weak spot or has a weak spot even though 23 they are passing all of their exams throughout the 24 course. There is a technical knowledge gap that has 25 to be addressed. And they challenge their end result NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
299 1 getting a license.
2 So we can determine that sooner rather 3 than later increasing the probability of one not only 4 selecting the right students, but to ensure that they 5 are successful in the end. It's not just a safety 6 issue, but it's a cost-saving measure. It costs over 7 $4 million to put a student through a class.
8 So that's in place, and it's working well.
9 The instructors love it. It takes the human element 10 out of having to manually try to determine that. It 11 automates it and provides the information to the 12 training instructors.
13 And, again, in these other next two 14 examples, the corrective action program, we use 15 analytics to ascertain which corrective action or 16 issue report or what we call non-conformance reports 17 in the industry in the plants, which are important, 18 which are important to safety, which are not, and 19 which need attention, which need urgent action. And 20 we've done a year of testing of that. And we get 21 really good results from that.
22 It's, again, saving the number of people 23 that need to review and need to manually do those 24 reviews, which not only increases efficiencies, but 25 it's actually more accurate. It's more consistent NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
300 1 decision-making. We find that we train these models 2 to the point where it's better than just having a set 3 of 6 or 10 humans reviewing these reports.
4 And finally I will say that we've got an 5 application now that we're developing. It's soon to 6 be deployed. Where we are taking all the information 7 that we heard from our digital twin models that we 8 talked about earlier today. I think I mentioned that 9 to you a year ago, we had the digital twinning of 10 modeling equipment so that we can predictively 11 determine when a piece of equipment may fail. It used 12 to be, okay, a couple weeks ahead of time. Now we get 13 months in advance we can determine degraded 14 performance.
15 We're taking that information and 16 eventually we're going to feed that into the work 17 management process, which is also automated analytics 18 so that teams of 40 or 50 planners per year that we 19 use to develop work packages to get work done during 20 outages to address and improve equipment performance 21 is now going to be whittled down to single digit 22 numbers of people because we are using analytics that 23 say, hey, I've done this maintenance before. I know 24 the piece of equipment you're working on. Here's the 25 historical data. Here is the work package that needs NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
301 1 to be done. You press a button. And in 17 seconds it 2 gives you a work package whereas before it took 24 to 3 48 hours5.555556e-4 days <br />0.0133 hours <br />7.936508e-5 weeks <br />1.8264e-5 months <br /> of worker time to develop that work package, 4 the research that had to be done.
5 So that the search capabilities and 6 analyzing historical data and working with what's 7 important from a safety perspective and automating 8 that to some degree has brought great efficiencies.
9 And keeping the human in ensures that we maintain that 10 same low risk probability and low consequence of 11 failure. Thank you.
12 CHAIR BIER: Thank you. So any further 13 comments in the room? If not, Norbert, please go 14 ahead with your comment.
15 MR. CARTE: Yes, Norbert Carte. I work 16 for the NRC in I&C, but it's more of a public comment 17 than an NRC position.
18 So I think Vicki is right. What's going 19 to happen first is you're going to introduce AI in 20 non-controlled tasks. But in doing that, as Richard 21 alluded, in doing that, you're going to change how the 22 whole system works. And when you change the system, 23 you're going to have new strengths and weaknesses.
24 And what you need to be really careful for 25 is when you change how everything is done, are you NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
302 1 introducing weaknesses like social loafing, that only 2 being one instance of one type of weakness, right?
3 But you are going to have a set of weaknesses related 4 to how people work in that new environment and that's 5 going to be your first threat. Thank you.
6 CHAIR BIER: Any further public comments?
7 If not, I want to thank all of the presenters for what 8 was really an excellent day. And I think we can now 9 be adjourned.
10 (Whereupon, the above-entitled matter went 11 off the record at 4:47 p.m.)
12 13 14 15 16 17 18 19 20 21 22 23 24 25 NEAL R. GROSS COURT REPORTERS AND TRANSCRIBERS 1716 14th STREET, N.W., SUITE 200 (202) 234-4433 WASHINGTON, D.C. 20009-4309 www.nealrgross.com
Data Science and AI Regulatory Applications Public Workshop Summary and Findings ACRS Joint Digital I&C and Human Factors Subcommittee Meeting November 15, 2023 Matt Dennis Trey Hathaway U.S. Nuclear Regulatory Commission Office of Nuclear Regulatory Research Division of Systems Analysis
Presenters
- NRC Technical Staff Presenters
- Matt Dennis - Reactor Systems Engineer (Data Scientist), RES/DSA/AAB
- Trey Hathaway - Reactor Systems Engineer (Code Development), RES/DSA/AAB
- NRC AI Champions
- Paul Krohn - Division Director, R1/DRSS
- Victor Hall - Deputy Division Director, RES/DSA
- Luis Betancourt - Branch Chief, RES/DSA/AAB 2
Outline
- Artificial Intelligence (AI) Landscape and the NRC
- Data Science and AI Regulatory Applications Workshops Overview
- Workshop Panel Session Summaries
- High Level Observations
- Moving Forward and Stakeholder Engagement 3
Artificial Intelligence (AI) Landscape and the NRC NUCLEAR INDUSTRY OTHER CONSIDERATIONS INTERNAL TO THE NRC (EXTERNAL) AND OPPORTUNITIES NRC Evidence Building (INTERNAL)
Priority Questions Federal actions for Chair's Memorandum advancing the use of AI in on Advancing the Use government operations Industry wants of AI at the U.S. NRC*
to use AI Internal interest in researching AI-based tools ranging from ACTIVITIES AI-embedded in commercial AI Strategic Plan applications to custom to prepare staff Wide range of AI meetings, programming to review AI conferences, and activities
- Available at https://www.nrc.gov/docs/ML2330/ML23303A143.pdf 4
Data Science and AI Workshops Overview
- Public workshop series* has discussed Objectives Intro to AI, Current Topics, Future Focused Initiatives, and AI Characteristics for Regulatory Consideration
- Observations from prior workshops (2021) FOUNDATIONAL COLLABORATION KNOWLEDGE OPPORTUNITIES
- Industry interest in regulatory guidance
- Nuclear-specific data sharing would benefit development of data hungry AI applications
- As of 2021, potential AI application deployment in 1-2 years and regulated DATA SHARING CURRENT FUTURE applications in 3-5 years PROJECTS ACTIVITIES
- Available at https://www.nrc.gov/public-involve/conference-symposia/data-science-ai-reg-workshops.html 5
Workshop #4:
AI Characteristics for Regulatory Consideration
- Purpose
- Discussion with NRC staff, international counterparts, academia, and industry on the multifaceted attributes of AI systems and their implications for regulatory considerations
- Provide feedback on regulatory and technical issues surrounding AI usage in nuclear applications
- Inform implementation of NRCs AI Strategic Plan (NUREG-2261, ML23132A305)
- Panel Sessions
- Regulatory Perspectives on AI
- AI Safety, Security and Explainability
- AI Application Considerations 6
Workshop #4: AI Characteristics for Regulatory Consideration (ML23268A314)
September 19, 2023 10:00 a.m. - 4:00 p.m. ET Time (Eastern) Topic Presenter 10:00 a.m. - 10:30 a.m. Opening Remarks Chair Christopher T. Hanson (NRC) 10:30 a.m. - 11:00 a.m. AI Characteristics for Regulatory Consideration Matt Dennis (NRC)
Session Chair - Paul Krohn (NRC)
Kevin Lee (CNSC) 11:00 a.m. - 12:15 p.m. Panel Session: Regulatory Perspectives on AI Andrew White (UK ONR)
Ben-Mekki Ayadi, Eric Letang (IRSN)
Var Shankar (Responsible AI Institute) 12:15 p.m. - 1:00 p.m. Break Session Chair - Josh Kaizer (NRC)
Rick Kuhn (NIST) 1:00 p.m. - 2:15 p.m. Panel Session: AI Safety, Security and Explainability Ali Raz (George Mason University)
Fan Zhang (Georgia Tech)
Xu Wu (North Carolina State University) 2:15 p.m. - 2:30 p.m. Break Session Chair - Jesse Seymour (NRC)
Rick Szoch, Jonathan Hodges (Constellation Nuclear, Jensen Hughes) 2:30 p.m. - 3:45 p.m. Panel Session: AI Application Considerations Clint Carter (Utilities Service Alliance)
Scott Sidener (Westinghouse)
Ryan Miller (TerraPower) 3:45 p.m. - 4:00 p.m. Open Discussion and Closing 7
AI Characteristics for Regulatory Consideration
- NRC AI Strategic Plan (NUREG-2261, ML23132A305)
- Table 1, Notional AI and Autonomy Levels in Commercial Nuclear Activities
- notional framework to consider the levels of human-machine interaction with AI systems
- Served as a starting point in the workshop to further discuss the variety of AI attributes which may affect regulatory considerations at each notional level
- AI Attributes Working Group
- Formed May 2023 and includes members from multiple agency offices
- Paul Krohn, Matt Dennis, Trey Hathaway, Jonathan Barr, Reed Anzalone, Josh Kaizer, Dave Desaulniers, Jesse Seymour, Tanvir Siddiky, Joshua Smith, Scott Rutenkroger, David Strickland, and Howard Benowitz 8
Disclaimer to AI Regulatory Considerations
and other frameworks for future alignment
- The following AI characteristics and considerations for developing AI systems does not represent an exhaustive list of categories for consideration
- The following AI characteristics are defined by a range of implementation levels that may impact regulatory decision-making
AI Characteristics for Regulatory Consideration Safety AI Autonomy Security Explainability Significance Model Regulated Regulatory Application Lifecycle Activity Approval Maturity 10
Regulatory Perspectives on AI Panel Session
- Panel session sought to engage with regulators and safety experts on considerations for AI systems and deployment in the nuclear industry
- Exploring usage and regulation of innovative technologies as part of the Disruptive, Innovative and Emerging Technologies (DIET)
Working Group
- Commissioned A Study for the Canadian Nuclear Safety Commission on Artificial Intelligence Applications and Implications for the Nuclear Industry
- U.K. Office for Nuclear Regulation (ONR)
- Issued report on the impact of AI/ML on nuclear regulation
- Considered two AI/ML applications as part of regulatory sandbox
- Possesses a flexible regulatory approach which can function without standards
- French Institute for Radiation Protection and Nuclear Safety (IRSN)
- Data governance, risk management and human monitoring are essential for high-risk AI applications
- EU AI Act establishes rules based on level of risk
- Responsible AI Institute (RAII)
- Provided recommendations from Certification Working Group on frameworks to validate AI tools and technologies as responsible, trustworthy, ethical, and fair
- Evaluation, validation and certification potential using the RAII Certification Framework
- Two leading auditable, voluntary standards: NIST AI Risk Management Framework (RMF) and ISO 42001 (AI Management Systems)
Disclaimer: Items above represent a summary of comments provided from workshop presenters and do not necessarily reflect NRC views 11
AI Safety, Security and Explainability Panel Session
- Panel session sought to share and discuss research into AI risks associated with the development and use of AI tools
- National Institute of Standards and Technology
- Numerous NIST projects being undertaken to support NIST AI RMF
- Described issues with conventional assurance processes for autonomous systems
- Current approaches to estimating success for transfer learning are largely ad-hoc and not highly effective and combinatorial methods show promise
- George Mason University
- Research into ML safety issues concerning robustness, monitoring, alignment and systemic safety
- Discussed explainable AI (how does the input influence decision making) and counterfactual testing (how to respond to unmodeled uncertainty)
- Examining model response in counterfactual cases to expose the black box nature of the model
- Georgia Institute of Technology
- Discussed enhancing cybersecurity of nuclear systems using AI/ML
- Developed and tested a multi-layer cyber-attack detection system using ML
- ML can provide cybersecurity monitoring benefits to observe unexpected systems changes
- North Carolina State University
- Application-agnostic algorithms for traditional ML algorithms cannot be applied directly to nuclear applications
- Methods presented to quantify uncertainties in deep neural networks (e.g., Monte Carlo dropout, deep ensembles, and Bayesian neural networks)
- Example provided on using deep neural networks to predict axial neutron flux profiles Disclaimer: Items above represent a summary of comments provided from workshop presenters and do not necessarily reflect NRC views 12
AI Application Considerations Panel Session
- Panel session sought to share and discuss industry potential AI use cases and experiences
- Constellation Nuclear, Jensen Hughes
- No clear industry-specific verification & validation (V&V) guidance for software that is driven by AI
- Developed tool for data-driven incident report classifications with explainability approach to document model predictions, high-level summary of rationale, and formal technical V&V
- Utilities Service Alliance
- Advanced remote monitoring project phase 1 is working with INL to embed AI in areas such as operator rounds, process anomaly detection, fire watch, and online transformer monitoring
- Assessment that regulatory readiness level is at a 2 of 5 and is planning future effort in phase 2 to explore AI-driven autonomous inspection, rounds, and response
- Westinghouse
- Emphasized the importance of creating an ethical AI corporate policy to ensure guardrails
- Human validation of AI models can be effective but risky, so rely on validation metrics and simulate the impact of incorrect results
- AI validation should assume model performance will change with time and should be monitored continuously
- TerraPower
- No active plans to use AI, but AI could be beneficial in a highly passive design which doesnt rely on operators for safety function
- Discussed high-level thoughts on using AI for engineering document preparation
- Considerations for using AI in a nuclear power plant include how to validate AI recommendations to licensed operators and if we should reevaluate the role human operators play in plant operations if AI is used Disclaimer: Items above represent a summary of comments provided from workshop presenters and do not necessarily reflect NRC views 13
Key Takeaways from AI Workshop #4
- Panel sessions confirmed that the NRC remains well informed on the status of international AI regulation and domestic R&D projects in the nuclear industry
- AI regulatory sandboxes provide a unique opportunity for industry and regulators to collaboratively explore the potential hurdles and benefits from using AI in safety-related nuclear applications
- Industry representatives encouraged continued collaboration to pursue pilot studies and proofs of concept as a foundation for reviewing the use of AI in NRC-regulated activities 14
NRC AI Considerations Traceable and auditable AI evaluation methodologies Current Understanding licensee and applicant AI usage Regulatory guidance and decision-making development Differentiating AI usage for design versus AI-enabled autonomy Future Explainable AI and trustworthy AI - reliability and assurance Internal AI resources predicated on emergent industry applications 15
Moving Forward and Stakeholder Engagement
- Continued safety and security in the nuclear industry is paramount
- Embrace new and innovative ways to meet NRCs mission
- Maintain strong partnerships with domestic and international counterparts
- Continue to encourage stakeholders to engage with the NRC early and often on plans and operating experience Future Activities
- Internal NRC AI Working Group to continue discussion of AI characteristics for regulatory consideration
- Regulatory framework applicability assessment of AI in nuclear applications (Summer 2023-Spring 2024)
- Planning for Summer 2024 AI Workshop on regulatory gaps and considerations 16
Abbreviations
- AI - Artificial Intelligence
- NIST - National Institute of Standards
- CNSC - Canadian Nuclear Safety and Technology Commission
- NRC - U.S. Nuclear Regulatory
- DOE - U.S. Department of Energy Commission
- IRSN - French Institute for Radiation
- ONR - U.K. Office for Nuclear Regulation Protection and Nuclear Safety
- RAII - Responsible AI Institute
- ISO - International Organization for
- R&D - Research and development Standardization
- V&V - Verification and Validation
- ML - Machine Learning 17
BACKUP SLIDES 18
Nuclear Industry AI Landscape
- Industry Project Categories
- Increasing existing economic efficiency
- Plant condition monitoring
- Process improvement and cost reduction
- Plant automation
- Sensor-enabled degradation assessment
- Example Operating Reactor Applications Areas
- Advanced remote monitoring
- Corrective action process automation
- Core design optimization
- Generative document preparation
- Physics-informed surrogate models
- Example Advanced Reactor Application Areas
- AI/ML-enabled digital twins
- Autonomous operation of backend processes
- Design optimization 19
Summary Considerations (1/2)
- Existing Guidance - Traditional safety, security, software, and systems engineering practices are still applicable as the starting point for good engineering practice.
- Establishing a Trustworthy System - Explainability exposes a chain of decision-making for potentially complex logic that is easily interpretable by anyone unfamiliar with the AI system design. This applies to all stakeholders which include reviewers (e.g., regulators) as well as system users.
- Safety Principles using Risk or Determinism - In the absence of the ability to quantify risk, there are good engineering principles (e.g., defense-in-depth) that can be used to guard against unintended consequences.
- Open-Source Tools - Use of open-source tools are not precluded, but using non-specialized software solutions means that there are steps taken to rigorously confirm the safety and security of the implemented solution.
20
Summary Considerations (2/2)
- Failure and Consequence Identification - A first step as part of AI systems engineering, a formalized process to quantify the hazards and modes of operation can be considered to ensure adequate system design.
- Data Provenance - Based on a graded approach, the modeling data may have a variety of various pedigrees based on the application area (e.g., safety significance).
- Model Updating - Models need to be maintained to avoid performance degradation and kept consistent with the pre-determined change control and notification process for that application.
- Human and Organizational Factors - The context of operation needs to consider the handover to human operation, immediacy for human action, or if placement in a safe stable state is required based on the operational context.
- Extensive Application Areas - A variety of regulatory requirements apply to various potential AI application areas. Existing requirements may range from evaluation of sufficient functional performance up to specific requirements to ensure AI system safety and security.
21
OFFICIAL USE ONLY - INTERNAL INFORMATION Notional AI and Autonomy Levels in Commercial Nuclear Activities Notional AI and Potential Uses of AI and Level Autonomy Levels Autonomy in Commercial Nuclear Activities Human Level 0 AI Not Used No AI or autonomy integration in systems or processes Involvement Insight AI integration in systems is used for optimization, Level 1 Human decision-making operational guidance, or business process automation that assisted by a machine would not affect plant safety/security and control AI integration in systems where algorithms make Collaboration Level 2 recommendations that could affect plant safety/security Human decision-making and control are vetted and carried out by a human augmented by a machine decisionmaker Operation AI and autonomy integration in systems where algorithms Level 3 Machine decision-making make decisions and conduct operations with human supervised by a human oversight that could affect plant safety/security and control Fully autonomous AI in systems where the algorithm is Fully Autonomous Level 4 responsible for operation, control, and intelligent Machine decision-making adaptation without reliance on human intervention or Machine with no human intervention oversight that could affect plant safety/security and control Independence Common Understanding of the Level Key for Regulatory Readiness 22
Clarifying Automation, Autonomy, and AI
- AI technologies can enable autonomous systems
- Not all uses of AI are fully autonomous as many may be used to augment human decision-making rather than replace it.
- Higher autonomy levels indicate less reliance on human intervention or oversight and, therefore, may require greater regulatory scrutiny of the AI system.
- Multiple definitions exist; however, it is important to have a clear understanding of the differences between automation and autonomy
- Automation - considered to be a system that automatically acts on a specific task according to pre-defined, prescriptive rules. For example, reactor protection systems are automatically actuated when process parameters exceed certain defined limits.
- Autonomy - a set of intelligence-based capabilities that allows the system to respond to situations that were not pre-programmed or anticipated (i.e., decision-based responses) prior to system deployment. Autonomous systems have a degree of self-governance and self-directed behavior resulting in the ability to compensate for system failures without external intervention.
23
NRC AI Project Plan Advisory Committee on Reactor Safeguards Joint Digital I&C and Human Factors Subcommittee Meeting November 15, 2023 Anthony Valiaveedu U.S. Nuclear Regulatory Commission Office of Nuclear Regulatory Research Division of Systems Analysis
Presenters
- NRC Technical Staff Presenters
- Anthony Valiaveedu - Reactor Systems Engineer (Data Scientist), RES/DSA/AAB
- Matt Dennis - Reactor Systems Engineer (Data Scientist), RES/DSA/AAB
- Trey Hathaway - Reactor Systems Engineer (Code Development), RES/DSA/AAB
- NRC AI Champions
- Paul Krohn - Division Director, R1/DRSS NRR NMSS
- Victor Hall - Deputy Division Director, RES/DSA OEDO RES OCIO
- Luis Betancourt - Branch Chief, RES/DSA/AAB NSIR Regions 25
Outline
- Background
- AI Strategic Plan Overview
- AI Project Plan Overview
- Moving Forward and Stakeholder Engagement 26
Background
April 2022 Published Capacity Assessment (NUREG-2251, November 2022 May 2023 July 2023 October 2023 Vol. 1) and identified a ACRS Subcommittee Issued AI Strategic Plan, Initiated AI Issued AI Project Plan finding related to AI briefing on staffs AI FY 2023 - FY 2027 Regulatory Gap for FY 2023 - FY 2027, preparedness and an planning efforts (NUREG-2261) Assessment Rev. 0 (ML23236A279) associated mitigating strategy 2022 2023 2024 September 2023 June 2022 Hosted Workshop #4:
Issued AI Strategic Plan March 2023 September 2023 AI Characteristics for draft for comment Launched AI Steering Launched AI Community Regulatory (ML23037A840) Committee (AISC) of Practice Consideration 27
AI Strategic Plan Overview Vision and Expected Outcomes
- Continue to keep pace with technological innovations to ensure the safe and secure use of AI in NRC-regulated activities
- Establish an AI framework and cultivate a skilled workforce to review and evaluate the use of AI in NRC-regulated activities The AI Strategic Plan consists of five strategic goals:
- Goal 1: Ensure NRC Readiness for Regulatory Decision-making
- Goal 2: Establish an Organizational Framework to Review AI Applications
- Goal 3: Strengthen and Expand AI Partnerships
- Goal 4: Cultivate an AI-Proficient Workforce
- Goal 5: Pursue Use Cases to Build an AI Foundation Across the NRC Available at ML23132A305 28
AI Project Plan Overview
- The AI Project Plan describes how the agency will execute the five strategic goals from the AI Strategic Plan
- Provides estimated timelines for various task completions within each Strategic Goal
- Communicates NRC priorities to internal and external stakeholders Available at ML23236A279 29
Goal 1. Ensure NRC Readiness for Regulatory Decision-making KEEPING THE END IN MIND - DETERMINING THE DEPTH OF REVIEW
- Assess regulations Regulatory and guidance as it Gap applies to gaps Analysis
- Identify usable standards and gaps
- Interdisciplinary team for the development of
- Collect industry AI plans Planning IEC/ AI standards at nuclear
- Develop scheduling for facilities for AI Goal 1 SC45A/
- Applicable to entire resource allocation on AI applications Submittals WGA12 nuclear fuel cycle
- Provides life cycle guidance on AI Outcome: Develop an AI framework to review the use of AI in NRC-regulated activities 30
Goal #1: Ensure NRC Readiness for Regulatory Decision-Making 31
Goal 2. Establish an Organizational Framework ENSURE CROSS-AGENCY LEADERSHIP IN AI WITH CENTRALIZED APPROACH AI Steering Committee NRC AI Community Centralized AI Database and Working Group of Practice Lead best practices, Cross-agency strategic Maintain transparency share knowledge and alignment and direction and clarity on AI usage lessons learned Centralized coordination Provide internal Agencywide list of of resources, priorities, awareness on active AI ongoing AI projects and use case analyses and potential external uses Recurring updates to Create working groups Monthly meetings ensure accuracy and as needed since September 2023 completeness Outcome: An organization that facilitates effective coordination and collaboration across the NRC to ensure readiness for reviewing the use of AI in NRC-regulated activities 32
Goal #2: Establish an Organizational Framework to Review AI Applications 33
Goal 3. Strengthen and Expand AI Partnerships GAIN VALUEABLE INFORMATION TO BENCHMARK AI ACTIVITIES Domestic International EXAMPLES INCLUDE EXAMPLES INCLUDE NRC MOUs: EPRI Data Science and AI; DOE US-Canada-UK trilateral Data Analytics and engagement on AI Operating Experience IAEA Consultancy and NIST RMF observations Technical Meetings on AI DENT and Big Data Bilateral Engagements Workshop on AI Outcome: An organization that facilitates effective coordination and collaboration across the NRC to ensure readiness for reviewing the use of AI in NRC-regulated activities 34
CANada-UK-US Trilateral Engagement on AI
Purpose:
Collaborate on a joint AI principles paper to establish a common set of overarching principles for reviewing the use of AI technologies in nuclear activities
- Objective: The CANUKUS trilateral AI principles paper will cover considerations for nuclear-related systems developed containing AI
- Outcome: The principles paper discusses
- High-level categories for AI uses cases in nuclear applications
- Country-specific regulatory frameworks
- Summary considerations on
- Use of existing safety and security systems engineering principles
- Human and organizational factors
- Characteristics of AI architecture CANUKUS
- Lifecycle management
- Demonstrating safe and secure AI systems that contain AI
- Working Group formed November 2022
- Paper is expected to be issued in Spring 2024 35
Goal #3: Strengthen and Expand AI Partnerships 36
Goal 4. Cultivate an AI Proficient Workforce ACQUIRE, DEVELOP, RETAIN, AND TRAIN AN NRC AI KNOWLEGABLE STAFF
- Focused on developing the critical skills for the AI workforce of tomorrow
- Training/Staffing
- Develop up-skilling plans through opportunities and qualifications
- Recruit AI Talent
- Workforce Planning
- Conduct competency modeling Outcome: Ensure appropriate qualifications, training, expertise, and access to tools exist for the workforce to review and evaluate AI usage in NRC-regulated activities effectively and efficiently 37
Strategic Goal #4:
Cultivate an AI-Proficient Workforce 38
Goal 5. Pursue Use Cases to Build AI Foundation Across the NRC CREATE AI ECOSYSTEM TO PREPARE FOR REVIEWING AI APPLICATIONS Pilot Studies AI Safety Insights
- Learn, measure, and evaluate
- Survey industrial safety evaluation readiness to implement regulatory methods and tools framework
- Utilize AI partnerships and
- Public workshops have shown engagement strategies industry interest to pursue pilot studies and proofs of concepts AI Ecosystem AI R&D Research
- Establish integrated development
- Continue supporting University environments and provide training grants and research into AI systems
- Acquire common data science tools
- Building AI foundation through the
- Develop regulatory sandboxes for NRCs Future Focused Research supporting use-cases Program Outcome: Develop an ecosystem that supports AI analysis, integration of emerging AI tools, and hands-on talent development for reviewing AI applications from the nuclear industry 39
Strategic Goal #5: Pursue Use Cases to Build an AI Foundation Across the NRC 40
Moving Forward and Stakeholder Engagement
- NRC must remain vigilantAI technologies are entering the nuclear domain in multiple venues
- NRC has been proactively working to understand this evolving technology to identify technical and regulatory challenges and gaps, gather insights on potential use cases, and develop institutional knowledge
- We are working to ensure we have the staff with the knowledge, skills, and ability to effectively regulate these new technologies
- Next Steps o Publish CANUKUS AI Principles Paper - Expected in Spring 2024 o Publish AI Regulatory Gap Analysis - Expected in Spring 2024 o Host AI Technical Session at 2024 RIC - March 12-14, 2024 o Host IAEA AI Technical Meeting at USNRC HQ - March 18-24, 2024 o Issue NRCs AI Project Plan, Rev 1 - Expected in Fall 2024 o Continue Public Workshops and Stakeholder Engagements - Ongoing 41
Abbreviations
- AI - Artificial Intelligence
- IAEA - International Atomic Energy Agency
- AICoP - Artificial Intelligence Community of Practice
- IEC - International Electrotechnical Commission
- AISC - Artificial Intelligence Steering Committee
- IRSN - Institut de Radioprotection et de Sûreté Nucléaire
- CFR - Code of Federal Regulations
- MOU - Memorandum of Understanding
- COE - Center of Expertise
- NLP - Natural Language Processing
- DOE - U.S. Department of Energy
- NRC - U.S. Nuclear Regulatory Commission
- DENT - Digital Engineering in Nuclear Technology
- NEI - Nuclear Energy Institute
- EO - Executive Order
- NIST - National Institute of Standards and Technology
- EPRI - Electric Power Research Institute
- OMB - U.S. Office of Management and Budget
- FFR - Future-Focused Research
- ONR - U.K. Office for Nuclear Regulation
- FRN - Federal Register Notice
- RIC - Regulatory Information Conference
- FY - Fiscal Year
- RMF - Risk Management Framework
- GAO - U.S. Government Accountability Office
- UNLP - University Nuclear Leadership Program
- GSA - U.S. General Services Administration
- WG - Working Group
- GRS - Gesellschaft für Anlagen- und Reaktorsicherheit 42
Using Machine Learning to Inform Inspection Planning A Future Focused Research Project Y. James Chang Senior Reliability and Risk Analyst RES/DRA/HFRB Presented at ACRS Subcommittee Meeting on Human Factors, Reliability & PRA and Digital I&C Systems on Artificial Intelligence Strategic Activities November 15, 2023
Presentation Outline
- Motivation, objective, and tasks
- Data and analysis
- Results and summary 44
Motivation and Objective
- Motivation:
- Netflixs success story of using unsupervised machine learning (ML) to identify hidden patterns (clusters) and similarities among its subscribers, leading to a more accurate movie recommendations
- Can ML + inspection findings identify hidden safety patterns to inform inspection planning?
- Objectives
- A feasibility study of using ML + inspection findings to identify hidden patterns (safety clusters) to inform inspection planning.
45
Safety Clusters - Hidden Patterns
- Safety clusters: Similarities in failure modes and failure causes of structure, system, and component (SSC) and consequences
- An example: NRC Operating Experience Communication
- Identified 5 power outage events impacted security operations
- SSC: The primary and backup electricity power systems for security systems
- Failure modes: fail to provide electricity
- Failure causes:
- 2022 (2 events): human errors contributed to the events
- 2021 (3 events): not attributed to human errors
- The OpE COMM suggested focusing on potential human impacts on power supply equipment when conducting Inspection Procedure (IP) 71130.04, Equipment Performance, Testing, and Maintenance.
46
Study Approach and Tasks
- Study approach:
- An NRC/RES future focused research (FFR) project
- Contractor:
- NRC team has expertise in ML and reactor oversight process
- Two tasks:
- 1. Evaluate AI/ML platforms:
- Amazons Sagemaker, Microsofts Azure, Googles Google-AI, MatLab, and others
- Maximize the use of pre-trained algorithms
- 2. Select a platform to identify hidden patterns (safety clusters)
- Technical work was completed within 4 months after awarding the contract 47
Task 1 Platform Evaluation
- Jupiter notebook (independent from the evaluated platforms) was used for Task 2 analysis.
- Jupiter: Free software, open standards, and web services for interactive computing across all programming languages.
- Cost was not evaluated but expected to be similar across platforms.
48
Task 2: AI/ML Pipeline to Identify Safety Clusters (Topics)
- 1. Topic Modeling Input 2. Topic Modeling 3. Topic Parameters Representation and Visualization Representation Weighting Tokenizer &
Vectorizer Clustering Dimension Reduction Embedding 49
Identify the Optimal Pipeline Constituents
- The pipeline contains multiple components to process the inspection finding descriptions to identify the safety clusters.
- A components function can be performed by multiple algorithms
- Identify the suitable algorithms
- In some cases, algorithms are combined to produce better results
- Many trials and iterations to identify the optimal algorithms and parameters for the pipeline
- Skip slow algorithms/parameters that may generate better results.
50
Trial-and-Err to Identify Optimal Combinations 51
Input Information for Analysis
- NRC ROP maintains an Excel database, which has all inspection findings since 1998. about 15,000 findings used for this analysis.
- The Item Introduction columns contain finding descriptions, which are the input for Task 2 analysis.
52
Process the Input Information
- Item Introduction (finding summary): Average 1,649 words for a finding (range from 42 to 11,670 words).
- Full text of the finding summary may not be the optimal choice because of sentence transformer models limitations.
- Use ML to generate condensed summaries.
- Tried full text and 14 condensed summaries (in three categories)
- Full text (1)
- Summary techniques (7)
- Question-answering techniques (3)
- Key phrase extraction techniques (4) 53
Custom Words and Phrases and Stop-Word
- Some ML techniques take input from customized lists of words and phrases to work with the pre-trained models.
- NRC provided 1004 acronyms and 407 common failure modes to focus nuclear safety.
- Failure mode: inoperable, misalign, and corrosion, etc.
- Provided 269 NUREGs and 195 RILs to develop a library of words/phrases relative locations.
- Stop-word removal:
- 337 English stop words, e.g., the and a.
- 136 custom stop words, e.g., safety, system, and reactor.
54
Stop-Word Removal Before stop-word removal After stop-word removal 55
Represent Safety Clusters
- Safety clusters are represented by word cloud (or a bag-of-words)
- In this analysis, an inspection finding only belongs to a safety cluster 56
Visualizing Safety Clusters 57
Evaluate Results Against an OpE COMM
- Identified 5 issues related to improper calibration and maintenance of radiation monitoring and dose assessment equipment that impact emergency plan actions
- Waterford 2011-2022 (2 events)
- Vogtle 2019
- Fermi 2016
- Wolf Creek 2013
- The OpE COMM identifies opportunities to identify these issues under Inspection Procedure (IP) 71124.05, Radiation Monitoring Instrumentation, and emergency drill observations, plant modifications or surveillance test reviews 58
Benchmark Results OpE identified 4 findings that exhibited safety issues that are related The clustering approach placed 3 of the 4 in the same cluster and the 4th in a similar cluster 59
Summary
- Unconclusive about unsupervised MLs practicality to inform inspection planning
- Need additional efforts to optimize the pipeline
- Outside the scope of this future focused research project.
60
Back Up Slides 61
Evaluate ML Techniques to Generate Condensed Summary - Original Summary The inspectors identified a Green NCV of Unit 3 Technical Specification (TS) 5.4.1 when Entergy did not take adequate measures to control transient combustibles in accordance with established procedures and thereby did not maintain in effect all provisions of the approved fire protection program, as described in the Unit 3 final safety analysis report. Specifically, on two separate occasions, Entergy did not ensure that transient combustibles were evaluated in accordance with established procedures; and as a result, they allowed combustible loading in the 480 volt emergency switchgear room to exceed limits established in the fire hazards analysis (FHA) of record. The inspectors determined that not completing a TCE, as required by EN-DC-161, Control of Combustibles, Revision 18, was a performance deficiency, given that it was reasonably within Entergys ability to foresee and correct and should have been prevented. Specifically, on August 28, 2018, wood in excess of 100 pounds was identified in the switchgear room; however, an associated TCE had not been developed. Additionally, on October 1, 2018, three 55-gallon drums of EDG lube oil were stored in the switchgear room without an associated TCE having been developed to authorize storage in this room, as required for a volume of lube oil in excess of 5 gallons. The inspectors determined the performance deficiency was more than minor because it was associated with protection against external factors attribute of the Mitigating Systems cornerstone, and it adversely affected the cornerstone goal of ensuring the availability, reliability, and capability of systems that respond to initiating events to prevent undesirable consequences. Specifically, storage of combustibles in excess of the maximum permissible combustibles loading could have the potential to challenge the capability of fire barriers to prevent a fire from affecting multiple fire zones and further degrading plant equipment. Additionally, this issue was similar to an example listed in IMC 0612, Appendix E, "Examples of Minor lssues," Example 4.k., because the fire loading was not within the FHA limits established at the time. Entergy required the issuance of a revised evaluation to provide reasonable assurance that the presence of combustibles of a quantity in excess of the loading limit of record would not challenge the capacity of fire barriers, and further evaluation and the issuance of an EC was necessary to raise the established loading limit to a less-conservative value. The inspectors assessed the significance of the finding using IMC 0609, Appendix F, Fire Protection Significance Determination Process, and determined that this finding screened to Green (very low safety significance) because it had a low degradation rating in accordance with Attachment 2 of the appendix. The inspectors determined that this finding had a cross-cutting aspect in the area of Human Performance, Work Management, because Entergy did not adequately plan, control, and execute work activities such that nuclear safety was the overriding priority, nor did they adequately identify risk associated with work being performed or coordinate across working groups to anticipate and manage this risk. Specifically, in the case of wood scaffolding being stored in the switchgear room, while planning work to be performed, Entergy did not adequately consider the fire risk that would be introduced by the presence of additional combustible materials. In the case of lube oil being stored in the room, Entergy did not take adequate action to ensure that activities were executed in a manner that would prevent work taking place in one area (the adjacent EDG cell) from introducing additional fire risk into a space for which it had not been evaluated (the switchgear room). In both cases, Entergy did not take sufficient action to ensure that workers were aware of the fire protection requirements associated with activities being conducted and to ensure that they coordinated as needed across working groups to adequately assess and mitigate the associated fire risk.
62
Summarization Models: Examples The inspectors identified a Green NCV of Unit 3 Technical Specification (TS) 5.4.1 when Entergy did not take adequate measures to control transient combustibles in accordance with established procedures and thereby did not maintain in effect all provisions of the approved fire protection program, as described in the Unit 3 final safety analysis report []
T5-Base Flan-T5-base inspectors identified a Green NCV of Unit 3 Technical Specification (TS) 5.4.1 when Entergy did not take Entergy did not take adequate measures to control adequate measures to control transient combustibles. the transient combustibles in accordance with established inspectors determined that not completing a TCE was a procedures and thereby did not maintain in effect all performance deficiency, given that it was reasonably provisions of the approved fire protection program, as within entergys ability to foresee and correct and should described in the Unit 3 final safety analysis report.
have been prevented. in the case of wood scaffolding being stored in the switch BART-large-cnn Pegasus-cnn-dailymail Entergy did not ensure that transient combustibles were evaluated in accordance with established The inspectors identified a Green NCV of Unit 3 Technical procedures . Entergy did not take sufficient action to Specification (TS) 5.4.1 when Entergy did not take ensure that workers were aware of the fire protection adequate measures to control transient combustibles. The requirements associated with activities being inspectors determined that not completing a TCE, as conducted . Entergy did not adequately identify risk required by EN-DC-161, Control of Combustibles, associated with work being performed or coordinate Revision 18, was a performance deficiency.
across working groups to anticipate and manage this risk 63
Question-Answering Models: Examples The inspectors identified a Green NCV of Unit 3 Technical Specification (TS) 5.4.1 when Entergy did not take adequate measures to control transient combustibles in accordance with established procedures and thereby did not maintain in effect all provisions of the approved fire protection program, as described in the Unit 3 final safety analysis report []
Flan-T5-base Roberta-base-squad2 Storage of combustibles in excess of the maximum permissible combustibles loading could have the potential to challenge the capability of fire barriers to nuclear safety prevent a fire from affecting multiple fire zones and further degrading plant equipment Bert-large-casedwhole-word-maskingfinetuned-squad No consistent results nuclear safety attained so QA was not selected 64
Key Phrase Extraction Methods: Examples The inspectors identified a Green NCV of Unit 3 Technical Specification (TS) 5.4.1 when Entergy did not take adequate measures to control transient combustibles in accordance with established procedures and thereby did not maintain in effect all provisions of the approved fire protection program, as described in the Unit 3 final safety analysis report []
KeyBERT KeyBERT + KeyphraseVectorizers allowed combustible loading, allowed combustible, combustibles additional fire risk, fire protection requirements, final safety revision 18, combustibles evaluated accordance, permissible analysis report, fire risk, maximum permissible combustibles combustibles loading, result allowed combustible, combustibles revision, combustibles evaluated, permissible combustibles, loading, fire protection significance determination process, fire transient combustibles evaluated, additional combustible, maximum barriers, additional combustible materials, combustible loading, permissible combustibles, combustibles loading, 161 control fire protection program, combustibles, transient combustibles, combustibles, final safety analysis, control combustibles revision, low safety significance, edg lube oil, fire loading, entergy, presence additional combustible, unit final safety, combustibles multiple fire zones, fire, nuclear safety, further degrading plant accordance established, established hazards analysis equipment`
Guided KeyBERT Guided KeyBERT + KeyphraseVectorizers allowed combustible loading, final safety analysis, unit final final safety analysis report, fire protection requirements, safety, combustibles evaluated accordance, safety analysis additional fire risk, fire risk, fire protection significance report, combustibles revision 18, allowed combustible, determination process, maximum permissible combustibles permissible combustibles loading, transient combustibles loading, fire barriers, fire protection program, combustible evaluated, combustibles evaluated, permissible combustibles, loading, low safety significance, additional combustible result allowed combustible, maximum permissible combustibles, materials, transient combustibles, combustibles, edg lube oil, 161 control combustibles, combustibles revision, combustibles further degrading plant equipment, fire loading, nuclear safety, loading, established hazards analysis, safety significance, safety entergy, multiple fire zones, volt emergency analysis, control combustibles revision 65
- Input: text for analysis
- Some algorisms can be used in a semi-supervised manner with a pre-defined list of important words and phrases to guide the algorithm.
- Embedding:
66
- Stop words: procedure, technical, license condition, 'safety',
- 'reactor', 'power plant, inspector, license, finding, cornerstone, cross cutting area
- mitigating, systems, barrier, integrity, initiating, event,
- human, performance, problem, identification, and resolution.
67
Bruce P Hallbert, Ph.D. Light Water Reactor Sustainability:
Director, Light Water Reactor Sustainability Program Sustaining and Optimizing the Technical Integration Office Existing Fleet
Light Water Reactor Sustainability Program
- Goal Enhance the safe, efficient, and economical performance of our nation's nuclear fleet and extend the operating lifetimes of this reliable source of electricity
- Objectives Enable long term operation of the existing nuclear power plants Deploy innovative approaches to improve economics and economic competitiveness of LWRs in the near term and in future energy markets Sustain safety, improve reliability, enhance economics
- Focus Areas Plant Modernization Research and Development Flexible Plant Operation and Generation Risk-Informed Systems Analysis Materials Research Physical Security 2
Plant Modernization Facilitate modernization by:
- Developing technology 3 modernization solutions that address aging and obsolescence challenges
- Delivering a sustainable business model that ensures continued safe, reliable operation at a cost Long Term Management Nuclear Cost Worker Attraction competitive level Of Plant Systems Competitiveness and Retention License Extension Cost pressures from other Worker interest in new 60, 80, or 100 years generation sources technology 3
Modernize the Fleet
- First echelon safety instrumentation and control systems on two units
- Conceptual Design Phase complete
- Detailed Design Phase in progress
- Multiple pre-submittal meetings with Nuclear Regulatory Limerick Generating Station Commission (NRC)
- Human Factors efforts well underway Operating Experience Review (Q3-Q4 of 2021)
Function Analysis and Allocation Workshop (March 2022)
Task Analysis Workshop (May 2022)
- NRC has accepted Constellations License Amendment Request (December 2022) INL Human Systems Simulation Laboratory Task Analysis Workshop
- Dynamic preliminary validation completed February 2023 at INL with NRC observation 4
Artificial Intelligence (AI) and Machine Learning(ML) R&D AI/ML, associated
- Show promise for automating some methods and data manually performed activities analytics are relatively new to the
- Investigating approaches that show nuclear power promise to enhance efficiency industry.
- R&D efforts - not deployment
- Adoption must align with the nuclear Current efforts safety culture of the industry emphasize work
- Align and comply with recent process vs. control. presidential directive and other requirements Trust in automation,
- Human factors issues in AI/ML transparency, implementation vital to adoption and understandability safe use.
affect usability.
5
Flexible Plant Operation and Generation Main Steam Turbine/Gen Set 500 kV
- Enhance economic flexibility and decarbonize Switchyard Steam energy and industry Electricity AC Grid Slipstream Condenser
- LWRS research addresses: Power Offtake Line Pressurized Water Hydrogen production and storage safety risks Condensate Reactor Return Extraction Heat Power New thermal extraction and delivery systems Exchangers Inverter Thermal Energy DC Delivery Loop Modifications to the electricity transmission De-ionized Water station Delivery Heat Exchangers H2 O2 Hydrogen Plant Operator control of dynamic dispatch of power Steam Electrolysis Economics of transitioning between the electricity grid market and hydrogen production 6
Enabling Nuclear-Hydrogen Hubs 2019 2020 2021 2022 2023 2024 2025 2026 2026+
Leverage EPRI FPO and Low-Carbon Research Initiative and Nuclear Beyond Electricity
- The Bipartisan Techno-Economic Assessments:
Infrastructure Law funds Hydrogen Plastics Thermal Markets Synfuels Energy Storage/Arbitrage Ammonia & Steel Clean Regional LWR Connection to Hydrogen Plant: Architecture Engineering:
Hydrogen Hub Projects Thermal Energy Delivery Modeling Scalable Thermal Energy Offtake 20 MW, 200 MW, 1000 MW Full-Scope PWR Simulator & Modeling Scalable Electricity Dispatch
- The Inflation Reduction Initial Operator Dispatch Studies Preliminary PRA study for PWR & BWR 50 MW, 100 MW, .., 500 MW Act provides Clean Initial Operating License considerations Multi-Facility Coordinated Energy Dispatch Concepts of Operations Hydrogen Production Tax Credits (up to $3 per kg- Hydrogen Regulatory Research Review Group (H3RG)
H2 with low-emissions energy sources) Update Reference PRA:
Hydrogen Production Large Thermal Industry Engineering Designs:
H2 Conversion Demonstrations:
Utility Demonstration Projects: Regional Clean Hydrogen Hubs:
7 Constellation Energy Harbor Xcel Energy PNW Hydrogen H2 Products
Pilot Plant Hydrogen Demonstration Projects Nine Mile Point Nuclear Power Plant
- Constellation (Exelon): Nine-Mile Point NPP Operating since 1.25 MWe low temperature electrolysis (LTE) February 2023 Using house load power Training plant operators by practicing power switching between grid and hydrogen plant Davis-Besse Nuclear Power Plant H2 production
- Energy Harbor: Davis-Besse NPP beginning 1-2 MWe LTE or 2-4 MWe HTE (with electrical steam boiler)
March/April 2024 Power provided by plant upgrade with new switch gear at the transmission station Gaining hydrogen production experience in anticipation of scaling up production Prairie Island Nuclear Power Plant H2 production
- Xcel Energy: Prairie Island NPP beginning middle 150 kWe high temperature electrolysis (HTE) of 2024 Thermal tie into existing turbine stream extraction line Gaining high temperature hydrogen production experience 8
Materials Research
- 1. Measurement of degradation Objective: conduct R&D to 2. Mechanisms of degradation understand the long-term 3. Modeling and simulation environmental degradation behavior 4. Monitoring degradation of materials in nuclear power plants 5. Mitigation strategies
- 6. Materials harvesting Reactor Pressure Core internal and Vessel (RPV) pressure boundary Mitigation Concrete Cable 9 9
Objective Risk-Informed Systems Analysis (RISA)
- R&D to achieve economic efficiencies while maintaining high levels of safety optimize safety margins and minimize uncertainties Approach
- Provide scientific basis to better represent safety margins and factors that contribute to cost and safety Areas of Expertise
- Advanced modeling of physics-based phenomena Thermal-hydraulics, neutronics and reactor physics, risk-informed material degradation, uncertainty propagation
- Advanced Data Analytics and Digital Modeling Diagnostic and prognostic analyses, resource optimization, AI/ML technologies, digital twins, uncertainty propagation
- Probabilistic Risk Assessment (PRA) and Human-Reliability RISA Framework Analysis (HRA) 10
Optimization of Nuclear Fuel Use Reactor Core Fuel Pattern Fitness of Pattern, Constraints, Objective Fuel costs represents approximately .JOOOOt 10 ;i,o * * ,o 10 1'11
- to
- ----*, 20% of annual operating expenses.
Al-based optimization framework for designing reactor core configuration given objectives and constraints ...
I ..
Multiphysics-based process with inputs ""
li1*,0 l40
"°I
-Ill from reactor core design, thermal
--C-CLI j .... - -c,- ........ 1 Juso hydraulics, fuel performance 81 *"-NOIII' I:]21-** l 21-.NO.- ,2 .. , ... ,,_ .. 0._ -
Genetic algorithm is the Al method used for optimization Example case of finding optimal fuel configuration in the reactor core with the objective to prolong the cycle length under peaking factor and boron concentration constraints.
11
Physical Security Research Physical Security research aims to create tools, technologies, and risk-informed physical security decisions and activities with the following objectives:
- Develop mitigation strategies and enhance the technical basis necessary for stakeholders to reevaluate physical security postures while meeting regulatory requirements.
- Analyze the existing physical security regime, current best practices, and compare/contrast insights with alternative methods which leverage advanced modeling and simulation, modern technologies, and novel techniques that address design basis threat and regulatory requirements Short-term goal is to enable industry to operate nearer the staffing requirements of 10 CFR 73.55 Main research thrust areas:
- Advanced Security Technologies
- Risk-Informed Physical Security
- Advanced Security Sensors and Delay 12
Summary
- The existing operating fleet provides the largest reliable source of carbon-free electricity.
- Industry initiatives have achieved substantial improvements in performance.
- Nuclear energy supports climate goals and can contribute to deep decarbonization by providing clean energy for products used in other industrial sectors.
- LWRS Program supports collaborations with industry to facilitate progress in areas of vital common interest. By working together, we facilitate progress and address challenges to ensure the continued viability and role of nuclear energy.
- The growing demand for clean and reliable energy from nuclear power underscores the need to address existential challenges facing the existing fleet.
- LWRS research addresses highest priority issues on timelines that support continued operation of the existing fleet.
13 13
Sustaining National Nuclear Assets lwrs.inl.gov 14
Integrated Operations for Nuclear (ION)
ION Business Model is guided by strategy Motivations for ION Traditional Approaches to Modernization Modernization competes Modernization Projects Corporate vision not Aging and Obsolete Nuclear Cost Worker Attraction with plant health for may lose momentum integrated with Plant Systems Competitiveness and Retention resources (unclear value) modernization Integrated Operations for Nuclear ION Yields Cost Savings 99.7% $60M estimated chance of positive harvestable annual net present value cost savings by Modernization projects One roadmap Complete and current ex. Digital I&C implementing ION funded through a sequencing all strategy guiding facility separate budget modernization efforts modernization process 15 15
Preliminary PRA for Steam Electrolysis
- Based on potential new internal and external initiating events Top hazards identified
- Internal: Steam line break, loss of offsite power
- External: Electrolysis Plant H2 leak or H2 detonation
- Hazards analysis completed by Sandia using HYRAM'
- Comparison of results to criteria in 10 CFR 50.59 and RG 1.174
- Early conclusions Licensing criteria are met for a large-scale hydrogen production facility sited 1 km from a generic PWR and BWR Individual site NPP and geographical features can affect the results of the generic PRA positively or negatively 16
Unattended Openings (UAO)
Impact: Provide the technical basis to determine optimized protective strategies related to person-passible openings that intersect security boundaries during normal and maintenance operations. Reductions in patrols, monitoring, and compensatory measures could be reconsidered but will be site specific on a case-by-case basis.
Highlights
- Conducted a performance-based, risk-informed evaluation for 2D and 3D unattended openings based on the current US Government policy
- Identified human factors associated with 2D and 3D openings
- Evaluated 2D UAO testing with 4-inch circle and rectangles and 36-inch circle
- Evaluated 3D UAO testing with 20-foot piping sections and pipe bends
- Evaluated success of passing through the opening (go/no-go), rate times, Example of 3D UAO Testing Configurations and limited data on exertion 17
Deliberate Motion Analytics (DMA)
- Security sensor fusion linked with DMA can take input from multiple sensors of different types, analyze the data, and determine if an adversary is making an approach toward a facility.
- Sites using current commercial sensor technologies typically experience elevated nuisance alarm rates (NAR) not caused by an intruder. Maintaining a low NAR while being able to detect intruders has the potential to decrease the cost of security.
Highlights
- Used DMA and sensor fusion, collecting at least four weeks of continuous performance data at two nuclear power plant sites
- Considered engineered terrain (perimeter intrusion detection system) and un-engineered terrain (owner controlled area).
- Created an NPP-specific demonstration package containing sensor fusion Active Radar (blue) and Thermal Camera (yellow) fused through DMA 18 Official Use Only
Craig Primer Plant Modernization Pathway Lead LWRS Program Plant Modernization November 2023
AI/ML Research Focus Areas ML for Material Management ML for Equipment Monitoring ML for Anomaly Detections NLP Applications Computer Vision Applications Overview of AI, ML and subsets of ML AI/ML Explainability
Digital Twin Informed ML for Material Management Developed technology to locate and estimate alkali- Digital Twin and Deep Learning Model used in Secondary silica reaction (ASR) damage using physics-informed Piping Degradation Detection Research machine learning approach Low-Cost Phase-Sensitive Distributed Fiber Sensors FS-laser Direct Writing Digital Twin and AI Sensor Data DT Data Deep Learning Neural Network Concrete Structure Health Monitoring Using Vibroacoustic Testing and Machine Concept for Integrated Multi-Modal Online Piping Monitoring System along with Learning, INL/EXT-20-59914 Data Fusion and Advanced Data Analytical Algorithms Using High- Resolution 3 Fiber Optics Sensors, INL/EXT-20-59810
Predictive Maintenance Strategy
- Developed a scalable risk-informed predictive maintenance strategy using machine learning approaches, risk modeling, visualization, and multi-band heterogeneous wireless architecture.
- Developed a hybrid model of circulating water pump (CWP) motor (basis for digital twin) to capture different operating dynamics. Risk-informed Predictive Maintenance Strategy Scalable Technologies Achieving Risk-Informed Condition-Based Predictive Maintenance Enhancing the Economic Performance of Operating Nuclear Power Plants, INL/EXT-21-64168 Physics-based model of CWS 4
Interpretability of Artificial Intelligence and Machine Learning Technologies for building Trust Among Users LWRS Program researchers developed methods to address the explainability, performance, and trustworthiness of AI/ML to enhance the interpretability of outcomes.
One method uses objective metrics like Local Interpretable Model-agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP).
Another method employs user-centric visualization of AI/ML outcomes together with objective metrics to support expert interpretation.
In collaboration with Public Service Enterprise Group User-centric visualization with performance and explainability metrices (PSEG), Nuclear LLC, performed initial demonstration of the technical basis on circulating water system Explainable Artificial Intelligence Technology for Predictive Maintenance, (CWS) for a waterbox fouling problem. INL/RPT-23-74159 5
AI/ML Applications for Anomaly Detection
- Unsupervised ML Methods Recurrent neural networks were used Developing equipment-agnostic to predict dry well anomalies detection methods by holistic cooling fan failure using surrounding inference of process conditions sensors Extending Data-Driven Anomaly Detection Methods, INL/RPT-23-73933 Software - Automated Latent Anomaly Recognition Method, (ALaRM)
- Semi-Supervised ML Methods Developing methods to couple text-mined information from sparse condition reports to equipment and process sensors data Clustering methods were for equipment and process reliability used to detect High-Pressure Coolant Injection analysis (HPCI) system steam leak Feature Extraction for Subtle Anomaly Detection Using Semi-Supervised Learning, Annals of Nuclear Energy, vol. 181, pp. 109503, 2023 https://doi.org/10.1016/j.anucene.2022.109503 6
NLP Applications for Process Improvement Condition report screening is a
- Natural Language Processing and Deep Learning process that Methods: involves several Demonstrating and evaluating the automation of the staff on daily or bi-daily basis for condition reports screening process (which is the several hours a review and classification of condition reports according week.
to their impact on nuclear safety).
Software: Machine Intelligence for Review and Analysis of Condition Evaluating the automation of document review, Logs and Entries (MIRACLE) sampling, trending, analysis, and reporting.
Developing AI/ML
- Natural Language Processing and Clustering Methods:
methods to optimize Developing an inventory optimization method by the stocking coupling work demand information with parts inventory requirements in a plant to reduce the minimum stocking requirements.
Explainable Artificial Intelligence Technology for Predictive Maintenance, 7 INL/RPT-23-74159
Computer Vision Applications for Process Improvement
- Computer Vision and Deep Learning Methods: Example of AI/MLs ability to Developing methods to automatically identify a fire in a video stream accurately identify to augment the effectiveness of the fire watch program fire and smoke.
Developing and evaluating the automation of logging analog gauges (i.e., a method to recognize gauges in oblique angles and read their values) Automating Fire Watch in Industrial Environments through Machine Demonstrating methods for drones to autonomously recognize and Learning-Enabled Visual Monitoring, INL/EXT-19-55703 navigate their environment in a nuclear power plant. Software - Modelling Framework for Fire and Smoke Detection in Imagery Automated gauge reading impacts a wide Drones can automate spectrum of activities in a plant including several activities in a plant operator rounds, gauges calibration, and including operator and peer verification, and improves data security rounds, and fidelity for online monitoring. inspections of hazardous locations.
Patent - Automated Gauge Reading And Related Systems, Methods, And Software - Route-operable Unmanned Navigation of Drones 8 Devices (ROUNDS)
AI/ML Artificial intelligence, Research machine learning, associated methods and
- Show great promise for automating many manually performed activities data handling techniques
- Are demonstrating new approaches to Summary are relatively new in the nuclear power industry.
enhance efficiency ML for Material Management Collaborative efforts with
- Adoption must align with the nuclear safety culture of the industry.
ML for Equipment Monitoring owner-operators and
- Some uses demonstrate ability to others emphasize many ML for Anomaly Detections non-safety uses. rapidly transition to safety important uses.
NLP Applications Computer Vision Applications
- Human factors issues in AI/ML Trust in automation, understandability affect implementation vital to adoption and usability. safe use.
AI/ML Explainability
Sustaining National Nuclear Assets lwrs.inl.gov 10
Ahmad Al Rashdan, Ph.D.
November 15, 2023 Fire Watch
What is a Fire Watch?
Regulatory Guide 1.189 Section 2.2.1 states:
Work involving ignition sources such as welding and flame cutting should be carried out under closely controlled conditions.
Persons performing such work should be trained and equipped to prevent and combat fires. In addition, a person qualified in performing hot-work fire watch duties should directly monitor the work and function as a fire watch. Image From: U.S. Department of Labor 2019.
Regulatory Guide 1.189 Section 2.4.C states:
Successful fire protection requires inspection, testing, and maintenance of the fire protection equipment. A test plan that lists the individuals and their responsibilities in connection with routine tests and inspections of the fire protection systems should be developed. The test plan should contain the types, frequency, and detailed procedures for testing. Frequency of testing should be based on the code of record for the applicable fire protection system.
Procedures should also contain instructions on maintaining fire protection during those periods when the fire protection system is impaired or during periods of plant maintenance (e.g., fire watches).
2
Motivation
- Fire watch could cost an excess of $1M per month in a nuclear power plant that implement a fire protection program under Appendix R to Part 50Fire Protection Program for Nuclear Power Facilities Operating Prior to January 1, 1979.
- In 2019, the Utilities Service Alliance was awarded an iFOA grant titled Advanced Remote Monitoring to research and develop automation and advanced remote monitoring technology into the United States nuclear fleet to achieve economic viability while maintaining or improving safety and reliability. This includes the fire watch process.
- As part of the award, INL is working with the Utilities Service Alliance and its members through a Cooperative Research and Development Agreement to research, develop, and evaluate a custom-made fire cart for fire watch.
- The cart is equipped with a suite of sensors (e.g., camera, infrared cameras, smoke detectors).
The sensor signals are fused to reliably detect fire.
- A camera can be used to detect fire in a video stream or image of various industrial environments using machine learning models.
3
Considered Automation Options Method Advantage Technical Limitation No massive data collection or Reliable manual feature Image Processing training needed engineering is needed Do not have to be explicitly Require large amounts of diverse Spatial Machine Learning programmed to extract specific image training data to achieve features adequate generalization Consider changes in time as an Require large amounts of diverse Spatial and Temporal additional dimension for features video training data to achieve Machine Learning extraction adequate generalization 4
Outline
- Data Collection and Preparation
- Models Architecture Unlocking Pre-trained Models Ensemble Classifier Fire Detected?
- Models Performance Yes/No
- Use Considerations Data Management Collect Data Model Training Prepare Data Model Testing Split Data Select/Design Model Modify Model Model Deployment Tune Hyperparameters Evaluate Performance Augment Model Improve Performance Refine Model Monitor Model Update Model 5
Steps to Develop a Machine Learning Model
Data Collection General Image/Video Sources:
No-Fire Labels Fire Labels
- YouTube-8M
- Google Images Parameter Description
- Yahoo Flickr Creative Commons 100 Million Environment Target area, e.g., indoors or outside (YFCC100m) Detection The size of the target area and distance of Space potential fire targets from the sensor Targeted Image/Video Sources: Objects within the target area (e.g., people,
- Fire Smoke (FiSmo) Objects desk, light source, automobile, etc.)
Using Existing Classifiers: Light consists of artificial or natural sources, or Light Source both
- ImageNet (e.g., candle, cannon, fire screen, geyser, missile, space shuttle, stove, torch, Light source and duration vary by time of day, Light Duration season, etc.
volcano)
Spatial, spectral, and temporal resolutions; Sensor stationary or moving 6 Diversity of Collected Data
Data Preparation
- Selectively identified noisy regions of Dataset Fire Images Normal Images frames and either crop the frames to Training 29,519 20,056 remove the unwanted features, or Validation 3,690 2,507 carefully blur the region by using various Test 3,690 2,507 image processing tools Total 36,899 25,070
- Ensured sufficient temporal separation Allocation of Collected Fire Data between consecutive video frames.
Usually, this temporal separation was Dataset Smoke Images Normal Images around two to three seconds between Training 4,460 8,636 extracted frames Validation 558 1,080
- Dataset was slightly biased toward fire Test 558 1,080 images having 30% more than non-fire. Total 5,576 10,796 This class imbalance was meant to bias Allocation of Collected Smoke Data the model toward making a fire prediction.
7
Model Total Parameters Total Layers EfficientNetB7 64,652,242 815 Models Architecture InceptionResNetV2 InceptionV3 54,647,906 22,263,970 782 313 ResNet101 43,245,058 347 ResNet50 24,226,818 177 Increase the non-linearity ResNet50V2 24,211,586 192 Produce a new array of the output to further VGG16 14,862,146 21 that contains high-level enhance the features VGG19 20,171,842 24 or dominant features Xception 21,499,178 134 such as edges Reduce array Strengthen the dimensions to improve feature extraction model efficiency and Tune features wights to and further reduce reduce noise determine the class output dimensionality An example CNN architecture containing two convolution and pooling segments feeding into a densely connected layer for classification (from Google developers 2019).
8
Unlocking Pre-trained Models
- Transfer learning unlocks part of the pre-trained layers and its parameters for the training process.
- One model was trained for every 10% increment of trainable layers for each of the nine model architectures.
EfficientNetB7 InceptionResNetV2 InceptionV3 ResNet101 ResNet50 ResNet50V2 VGG16 VGG19 Xception Dense 865,282 371,714 495,618 692,226 692,226 692,226 147,458 147,458 692,226 10% 27,109,026 13,212,354 6,569,154 15,668,226 8,573,442 8,571,906 2,507,266 2,507,266 6,942,618 20% 39,025,122 23,384,930 11,610,498 19,880,194 15,670,274 15,926,274 7,226,882 7,226,882 9,094,586 30% 51,840,450 31,170,978 13,713,794 23,499,266 17,905,666 17,897,474 9,586,690 11,946,498 11,246,554 40% 55,959,698 36,662,274 15,595,778 27,121,154 19,880,194 19,871,234 13,126,658 16,666,114 13,398,522 50% 59,948,634 42,292,706 17,285,058 31,330,306 22,054,402 22,042,626 13,716,738 17,846,274 15,550,490 60% 62,169,034 47,779,746 18,826,882 34,951,682 23,057,666 23,045,122 14,306,818 19,026,434 17,702,458 70% 63,780,554 53,142,770 21,200,738 38,898,946 23,553,410 23,540,482 14,601,986 19,911,682 19,854,426 80% 64,363,866 53,558,866 21,559,282 42,075,906 24,017,154 24,002,818 14,823,426 20,133,122 21,308,362 90% 64,579,542 54,003,266 21,842,658 43,035,394 24,141,570 24,126,978 14,860,354 20,170,050 21,444,682 100% 64,652,242 54,647,906 22,263,970 43,245,058 24,226,818 24,211,586 14,862,146 20,171,842 21,499,178 9
Models Performance
- = 1 + 2 2 +
- score uses the parameter to weight the relative importance of recall versus precision. A score greater than one weights the score in favor of recall.
- By setting = 2, the F-score is weighted such that recall is considered twice as important as precision and a reliable detection is considered twice as impactful as a Source: https://en.wikipedia.org/wiki/Precision_and_recall false alarm.
10
Models Performance Heatmap representation of 2 score metrics for fire detection 11
Ensemble Classifier 12
Models Focus 13
Use Considerations
- Digital Instrumentation and Controls (DI&C) regulatory requirements would need to be satisfied for any and every AI application that impacts safety related and risk-significant applications
- Research into the fire watch models compatibility with the current safety standards
- Aligns with the directions of the recent executive order on safe, secure, and trustworthy artificial intelligence use. Specifically, it provides insight for development of new standards for AI safety and security To evaluate how example AI technologies align with the safety framework, and discusses how they could be analyzed, modeled, tested, and validated in a manner similar to typical DI&C technologies.
14 https://lwrs.inl.gov/SitePages/Reports.aspx
Compatibility with the Current Standards Requirements Independence Cyber Security Maintainability Design Control Deterministic Nature Trustworthiness Configuration Traceability Repeatability Explinability Justification Defense in Control Reliability FMEA Simplicity Depth CCF V&V QA CGD Characteristic/Consideration Open-source data and model Frequent updates to source Massive amounts of data Periodic training Probabilistic and stochastic Various performance metrics Incomprehensible to reviewers Inherited bias Non-systematic approach Robustness to new conditions Special skillset 15
Example of Compatibility Considerations Models often utilize open-source datasets and feature extraction engines or models.
It is not always possible to determine the level of overlap among open-source datasets. Open-source models could use similar fundamental concepts. This impacts the independence of the developed models:
- Causes the system to be susceptible to common cause failure
- Overestimates the software verification results
- Introduces a cybersecurity concern Methods to create independent Methods to quantify independence datasets may be needed (e.g., GANs) may be needed 16
Sustaining National Nuclear Assets lwrs.inl.gov 17
THE GOOD, THE BAD AND THE UGLY OF AI IN PROCESS CONTROL Missy Cummings, PhD George Mason University
Just what is AI?
Neural Networks GEORGE MASON UNIVERSITY
AI Problems sky vegetation veg build ing trac fence sign truck build rider/
ing person truck bus road road road GEORGE MASON UNIVERSITY sky vegetation veg build ing trac fence sign
Autonomy/AI & Reasoning Human behavior Top-down reasoning AI blind Environment spots ty Expert a i n c e rt Knowledge Un Rule Skill 4 Bottom-up reasoning GEORGE MASON UNIVERSITY
Autonomy/AI & Reasoning Human behavior Top-down reasoning AI blind Environment spots ty Expert a i n c e rt Knowledge Un Rule Skill 5 Bottom-up reasoning GEORGE MASON UNIVERSITY
The Uncertainty Wall Top-down reasoning ty Expert a i n c e rt Knowledge Un Rule Skill Bottom-up reasoning GEORGE MASON UNIVERSITY
AI in Process Control Symbolic/ Connectionist H&S Loop (synchronous & asynchronous)
Controls Actuators Displays Sensors Human Automation Tasks Supervisor Connectionist models 1st principles models*
- Uncertainty dependent GEORGE MASON UNIVERSITY
Process Control w/ Unmodeled Uncertainty GEORGE MASON UNIVERSITY
Monitoring People/Plant Over Different Time Scales Descriptive, Predictive
& Prescriptive GEORGE MASON UNIVERSITY
AI Hazard Analysis GEORGE MASON UNIVERSITY
What about Large Language Models?
GEORGE MASON UNIVERSITY
Dont Trust and Definitely Verify GEORGE MASON UNIVERSITY
LLMs & Predictive Maintenance GEORGE MASON UNIVERSITY
The thorny path ahead
- The hype cycle
- Follow the money
- Human + AI collaboration is key
- Cybersecurity & disinformation
- Inappropriate code in safety-critical systems
- Hardware-software integration
- Workforce development
- AI fact checking
- AI maintenance
- AI risk management is key GEORGE MASON UNIVERSITY
Questions?
GEORGE MASON UNIVERSITY