Text

Session 4 - Development of Standards for Artificial Intelligence Systems
	ML23251A060
Person / Time
Issue date:	09/13/2023
From:	Robert Roche-Rivera; NRC/RES/DE/RGDB
To:	;
References
	Download: ML23251A060 (1)
	v • d • e

Development of Standards for Artificial Intelligence Systems Opening Remarks by Victor Hall, Deputy Division Director, RES/DSA Session Chair: Luis Betancourt, Branch Chief, RES/DSA/AAB Panelists/Speakers:

Elham Tabassi (NIST)

Jonnie Bradley (DOE)

Ahmad Al Rashdan (INL)

Trey Hathaway (NRC)

Thiago Seuaciuc-Osorio (EPRI)

AI RISK MANAGEMENT PLAYBOOK DOE/NRC AI/ML Meeting September 13, 2023 Ms. Jonnie Bradley, Sr. Program Manager/

Responsible AI Official, Artificial Intelligence and Technology Office 2

AGENDA

Overview

AI RMP Features

Pathways

Future Developments

Walkthrough 3

DELIVERING ON AITO'S MISSION AND VISION 1.

From AI inventory data tracking to AI Portfolio and Program Optimization & Impacts

Strategic portfolio analysis and alignment of AI/ML investments, ensuring alignment with national security priorities, facilitates department-wide responsible and trustworthy AI across program offices (ex. Federated Learning) 2.

From management of AI projects to orchestration of AI Strategy and Partnership Development

Builds robust AI partnerships and customer excellence across internal, external, and international boundaries and addresses strategic communications for the Department.

4

DELIVERING ON AITO'S MISSION AND VISION Visionaries and practitioners of Modern-Day Governance for AI and Autonomous Innovations

Responsible AI Official

Convene

Coordinate

Facilitate

Orchestrate

Advocate

Integrate

Risk Management 5

OVERVIEW: AI Integrated Ecosystem 5

AI RMP AIX System AI Pathways AI Strategy AI Advancement Council 1

2 3

4 5

Guidance for leaders of new AI/ML projects and a reference of risks and mitigation techniques for technical users A space for the DOE AI community to post questions, share knowledge, and connect with SMEs Certificate upskilling program with two tracks: Management and Technical A space to test and validate algorithms and models using the latest techniques Enables better coordination across DOE programs and alignment with strategic priorities in AI DOE AI Community Hub Active or underway In planning stages 6

AI RISK MANAGEMENT PLAYBOOK 7

AI RMP VALUE The AI Risk Management Playbook (AI RMP) is a comprehensive reference guide for AI risk identification with recommended mitigations to support responsible and trustworthy (R&T) AI use and development.

8

1. April 2020 - AI Asset Risk Management Framework TIMELINE: AI RISK MANAGEMENT PLAYBOOK September 2022 -

AI RMP Public Q&A April 2020 - AI Asset Risk Management Framework August 2020 - Internal AI Risk Management Playbook (AI Ethics, Data Compliance, Asset Management, Secure Model Usage) 2021 - DOE review, edit, comment, &

concurrence, and intergovernmental and industry review/comment August 2022 - Public Release

- energy.gov/ai/airmp 9

AI RMP: FEATURES Essential Guidance Dynamic system featuring 141 unique risks and mitigation techniques with ability to easily and continuously expand Intelligent Search Ability to filter according to lifecycle stage, assets, as well as mapping to project roles and direct keyword searching Trustworthy AI Integration with EO 13960:

Promoting the Use of Trustworthy AI, including ability to filter by principle 10

USING AI RMP AI Use Case Example Risk Using data for AI systems to determine peak charging hours for EV owners Unintended biases in datasets used to build AI systems; disadvantaged communities are not included in the datasets Using data for AI systems to determine which neighborhoods to install EV supercharging stations Using data for AI systems to predict power outages and which communities will be serviced after an outage 11

AI RMP: PATHWAYS Additional filters designed for project managers and AI novices Provides understanding of risks according to development stage Connects risks with Trustworthy AI principles Ability to directly search via keyword according to issues encountered and mitigation techniques Quickest pathway for technical experts to find relevant risk or mitigation recommendation MANAGEMENT TECHNICAL 12

AI RMP: FUTURE DEVELOPMENT Additional Content Continuously add new risks and mitigation techniques; Eventual integration with NISTs AI Risk Framework Enhanced Searching New search capabilities co-developed with playbook users including the ability to search by project role Increased Engagement Additional features to drive community engagement including contributor badges and enhanced editing capabilities Additional Content Continuously add new risks and mitigation techniques; Eventual integration with NISTs AI Risk Framework Enhanced Searching New search capabilities co-developed with playbook users including the ability to search by project role 13

Q&A/

Discussion 14

ADDITIONAL RESOURCES

Executive Order 13960 - Promoting the Use of Trustworthy Artificial Intelligence in the Federal Government

NIST AI Technical Standards

Blueprint for an AI Bill of Rights

National Artificial Intelligence Initiative

Algorithmic Discrimination Protections

AI Now Institutes Annual Reports

Partnership on AI

Alan Turing Institutes Fairness, Transparency, Privacy group

Harvard's Embedded Ethics Modules

DAIR Institute

Fairness Tutorial 15

THANK YOU Artificial Intelligence and Technology Office U.S. Department of Energy September 13, 2023 Website: energy.gov/ai AIRMP: energy.gov/ai/airmp Inquiries, comments may be sent to:

doeaimailbox@hq.doe.gov 16

DOE DRAFT RESPONSIBLE AND TRUSTWORTHY AI PRINCIPLES DOE AI Principles Draft Equitable Traceable Reliable Governable Accountable 17

September 13, 2023 NRC Standards Forum Ahmad Al Rashdan, INL Considerations for the Development of a Standards-compatible AI Acknowledgment: Ted Quinn and Roman Shaffer 18

Ahmad Al Rashdan, Ph.D.

Idaho National Laboratory, USA A senior scientist with 17 years of industrial and research experience in automation, artificial intelligence/machine learning (AI/ML), and instrumentation and controls (I&C).

Lead of several efforts under the U.S. Department of Energy (DOE) focusing on plant modernization using AI methods and advanced analytics.

Author of more than 80 reports and journal papers, 9 patent applications and awards, and developer/co-developer of 8 Copyrighted software packages.

Recipient of several funding and recognition awards including a recent 2022 R&D 100 and the prestigious INLs Director Excellent Engineering Achievement Award.

Holder of several professional leadership positions Standards involvement:

Member of the IEC SC45A WG12 to create a standard for AI Application for Nuclear Installations.

Member of the ANS Large Light Water Reactor Consensus Standards Committee.

Vice Chair to the ANS Simulators, Instrumentation, Control Systems, Software & Testing standards Subcommittee.

Senior Research and Development Scientist E-mail = ahmad.alrashdan@inl.gov Phone = +1 979 422 4264 19

Computer Vision Machine Learning (CVML) in Fire Watch

Automatically identifying a fire in a video stream to eliminate/reduce the need for fire watches.

Sixteen different models were evaluated.

An ensemble of models is developed to improve accuracy

(>99% achieved)

Smoke is being integrated.

Temporal effects are being considered to eliminate false positives (mist, fog, steam, etc.)

20

CVML in Gauge Reading

Automating manual logging of analog gauges (i.e., a method to recognize gauges in oblique angles and read their values)

Automated gauge reading impacts a wide spectrum of activities in a plant including operator rounds, gauges calibration, and peer verification, and improves data fidelity for online monitoring.

21

The Need for Compatible AI and What Does that Mean?

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.

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

Regulations frequently cite standards.

Can AI be customized to meet those standards requirements?

Do we need new AI-focused standards?

22

Example of Considerations CVML 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 CVML models:

Causes the CVML system to be susceptible to common cause failure

Overestimates the software verification results

Introduces a cybersecurity concern Methods to create independent datasets may be needed (e.g., GANs)

Methods to quantify independence may be needed 23

CVML Compatibility with the Current Requirements Characteristic/Consideration Independence Defense in Depth CCF V&V QA Configuration Control Cyber Security CGD Maintainability Traceability Design Control Repeatability Deterministic Nature Explinability Reliability FMEA Simplicity Justification Trustworthiness 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 24

Conclusions 25

lwrs.inl.gov 26

IEC/SC45A/WGA12 Artificial Intelligence for Nuclear Facilities Trey Hathaway, Ph.D.

Reactor Systems Engineer U.S. Nuclear Regulatory Commission RES/DSA/AAB 2023 NRC Standards Forum September 13, 2023 27

AI Standardization Efforts

ISO/IEC JTC 1/SC 42 - Artificial Intelligence (AI)

- Created in 2017

- Secretariat: ANSI

- Scope: Standardization encompassing Artificial Intelligence

Provide guidance to JTC 1, IEC, and ISO committees developing artificial intelligence applications

IEC - International Electrotechnical Commission

- IEC/SC 45A - Instrumentation, control and electric power systems of nuclear facilities

Scope:

Prepare standards applicable to electronic and electrical functions and associated systems and equipment used in nuclear energy generation facilities to improve efficiency, safety and security of nuclear energy generation.

28

WGA 12

WGA 12 - AI for nuclear facilities

- First meeting held on August 22, 2023

Composed of a multi-disciplinary international team

- Currently 35 experts

NRC has four staff involved

- Ismael Garcia, Kim Lawson-Jenkins, Tanvir Siddiky - NSIR

- Trey Hathaway - RES 29

WGA 12

Tasks

- Develop and maintain standards and reports for AI applications in nuclear facilities

- Provide guidance to stakeholders developing, deploying, and overseeing AI applications for nuclear facilities

- Cover fundamental characteristics of AI of nuclear facility applications

- Applicable to the entire nuclear facility life cycle 30

WGA 12

Provide overview of AI from a nuclear perspective

- Discuss concepts, applications, and challenges

Explore definition of AI for nuclear facility applications

Discuss AI applications with focus on IEC/SC45A cross-cutting areas

- Instrumentation and control

- Annex to discuss applications of AI to other aspects of nuclear facility uses beyond IEC/SC45A scope 31

WGA 12 Working Group Working Group Titles WGA2 Sensors and measurement techniques WGA3 Instrumentation and control systems: architecture and system specific aspects WGA5 Special process measurements and radiation monitoring WGA7 Functional and safety fundamentals of instrumentations, control and electrical power systems WGA8 Control rooms WGA9 System performance and robustness toward external stress stems WGA10 Ageing management of instrumentation, control and electric power systems in NPP WGA11 Electrical power systems: architecture and system specific aspects Cross-cutting Topics with other Working Groups in SC45A 32

WGA 12

Three levels of documents to be produced

- Level 2 document on cross-cutting areas

General Requirements

Horizontal Standard

- e.g., Trustworthiness & Risk Management, Data Processing & Management, Testing and V&V, Performance Assessment

- Level 3 document on AI specific applications in nuclear facilities

Vertical Standard

Standards for specific applications

- e.g., Virtual sensors, anomaly detection

- Technical reports that support Level 2 and Level 3 standards

Documents created through IAEA Consultancy meetings will be leveraged in the IEC work 33

Status and Next Steps

Participants will notify working group convener of areas where they would like to offer input - by September 9, 2023

Meetings with other working groups at IEC General Meeting in October 2023 to discuss cross-cutting areas

Develop the Level 2 Standards first, then begin to explore application specific standards 34

w w w. e p r i. c o m Thiago Seuaciuc-Osorio Principal Technical Leader 2023 NRC Standards Forum September 13, 2023 AI-Assisted Ultrasonic Inspections in the Nuclear Power Industry 35

AI-Assisted Analysis of UT Inspections UT inspections are an important part of the scope of an NDE program Some inspections are challenging or have large volumes of data Machine learning tools can potentially assist in the analysis of the data

- Increase reliability

- Decrease analysis time Assist means AI flags regions for review: final decisions still rest with the qualified inspector.

Goal: Develop auto-analysis tools to assist in UT inspections 36

How Would AI Assist in UT Inspections?

Current Inspection

- Examiners distribute their energy across a high volume of (mostly benign) data AI Assisted Inspection

- Examiners focus their energy on the regions that require more careful review, while AI takes care of the more monotonous portion 37

Coming Soon Work in Progress Complete Development Status

Two successful field trials (2022-23)

Positive qualification assessment Reactor Vessel Upper Head Penetrations

Field trial now (Sep 2023)

Qualification assessment underway Dissimilar Metal Welds

Starting in 2024 Core Barrel + Core Shroud 38

False Calls Personnel Data Analysis Current Qualification Framework Develop AI-Assisted Analysis that fits this overall framework Detection Procedure Data Collection Blind Tests Dual Qualification Separate Stages 39

What initial qualification may look like Procedure would be updated to include a defined process for the AI evaluation AI algorithms for qualification would be developed and provided to the qualification body: model is frozen at this stage Yes Are all flawed regions flagged?

Qualification Body Determination Fail No Pre-recorded data AI flags areas for review Only flagged regions are presented to candidate, who reviews and makes calls Pass/Fail Based on typical detection

+ false call criteria AI model is applied 40

Elements that do NOT Change Qualification framework is maintained False Calls Personnel Data Analysis Detection Procedure Data Collection Same blind tests with same criteria Same data enables:

Ease of implementation

Parallel deployment 41

Elements that DO Change Modifications are typical of any new procedure False Calls Personnel Data Analysis Detection Procedure Data Collection Procedure adds AI analysis stage:

Freezes model 100%

detection required Personnel only reviews flagged regions:

Same test

More restrictive

Initial screening potentially on tool

Further analysis remain the same 42

Parallel & gradual deployment:

AI-Assisted as one of the two required independent reviews Gain operational experience Different failure modes Going Forward

Utilities can use for oversight & planning

Vendors can seek qualification Ready for Use

Assessment of reliability of AI-assisted analysis o UT, VT

Qualification protocol o Re-qualifications Supporting Research (2024 onwards) 43

TogetherShaping the Future of Energy 44

ML23251A060

Text

Navigation menu

Search