ML25343A204
| ML25343A204 | |
| Person / Time | |
|---|---|
| Issue date: | 12/01/2025 |
| From: | Tanya Smith NRC/NSIR/DPR |
| To: | |
| References | |
| Download: ML25343A204 (0) | |
Text
Todd Smith, PhD U.S. Nuclear Regulatory Commission AI in Action:
Revolutionizing Radiological Emergency Management
Emergency Preparedness and Incident Response programs are built around receiving and interpreting data to make informed decisions
AI can enhance decisionmaking AI provides solutions that are actionable interpretable transferable This picture generated by AI
How do you define AI?
NEURAL NETWORKS
f w1 w2 w3 y
x1 x2 x3 Bias Activation Function Output Inputs Input Layer Output Layer Hidden Layer
UNSUPERVISED LEARNING Input Layer Output Layer Hidden Layer
Choose 3 to represent all 63 sites
AI can find patterns in the data 1D 2D 3D Df = 1.6 10 0
10 1
10 2
10 3
10 4
0 0.2 0.4 0.6 0.8 1
1.2 1.4 1.6 1.8 2
r, box size
- d ln n / d ln r, local dimension 2D box-count
Neural network classification Rural Urban Coastal
AI added insight to make 4 models out of 3
SUPERVISED LEARNING Input Layer Output Layer Hidden Layer Comparator Desired Output Weight Adjustment
AI is interpretable 0.2 0.4 0.6 0.8 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Target Output ~= 0.84*Target + 0.039 Training: R=0.92201 Data Fit Y = T 0.2 0.3 0.4 0.5 0.6 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 Target Output ~= 1.2*Target + -0.067 Validation: R=0.96026 Data Fit Y = T 0.2 0.25 0.3 0.35 0.4 0.45 0.2 0.25 0.3 0.35 0.4 0.45 Target Output ~= 0.45*Target + 0.12 Test: R=0.93079 Data Fit Y = T 0.2 0.4 0.6 0.8 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Target Output ~= 0.84*Target + 0.034 All: R=0.91177 Data Fit Y = T Can a neural network predict evacuation time based on site-specific parameters?
population population density road network capacity mobilization time travel time Interpretation of AI model provided:
insights into important parameters for evacuation
input to NRCs ETE Study (NUREG/CR-7269)
enhancement to guidance (NUREG/CR-7002)
AI can make actionable predictions AI prediction of release time into containment Long short-term memory (LSTM) network Z. Dahm, S. Kannappel, W. Kelly. Design of a Machine Learning Tool for Predictive Emergency Response, Transactions of the ANS Winter Meeting 2022, Phoenix, AZ, pp. 56-59. doi.org/10.13182/T127-39694
AI accuracy depends on data and training AI prediction of release time into environment Z. Dahm, S. Kannappel, W. Kelly. Design of a Machine Learning Tool for Predictive Emergency Response, Transactions of the ANS Winter Meeting 2022, Phoenix, AZ, pp. 56-59. doi.org/10.13182/T127-39694 Long short-term memory (LSTM) network
FUZZY LOGIC
Response Technical Tool (RTT)
Response Technical Manual (RTM)
RTM Models Response Digital Twin using Fuzzy Logic (RDT)
AI enables technology
NATURAL LANGUAGE PROCESSING Classifier (sentiment analysis)
Output: Sentiment Transformer Input: Text Transformer based Deep Learning Model for Sentiment Analysis
Message Default Model 1 Model 3 Model 6 You cannot sense radiation negative neutral fear 1 star Radiation can only be detected using specialized instruments negative neutral anticipation 3 stars With the correct instruments, radiation is easily detected positive neutral optimism 4 stars We cannot eliminate radiation in our environment positive negative fear 1 star There is no known safe amount of radiation negative negative fear 2 stars There may be some risk from low levels of radiation negative neutral fear 3 stars It is reasonable to assume that less radiation exposure is better negative neutral optimism 3 stars Contamination occurs when radioactive material settles on a surface negative neutral fear 2 stars Go inside, get inside, stay tuned positive neutral anticipation 5 stars Radioactive fallout particles might be raining down in your area.
negative neutral fear 2 stars Model Number Model Name Base Transformer Training Labels Default distilbert BERT NEGATIVE POSITIVE 1
cardiffnlp/twitter-roBERTa-base for Sentiment Analysis RoBERTa 124 million tweets Negative neutral positive 2
cardiffnlp/twitter-roberta-base-emotion RoBERTa 58 million tweets joy, optimism, anger, sadness 3
cardiffnlp/twitter-roberta-base-emotion-latest RoBERTa
- fear, anticipation, optimism 4
cardiffnlp/twitter-roberta-large-emotion-latest RoBERTa 5
finiteautomata/bertweet
-base-sentiment-analysis.
BERT SemEval-2017 40,000 tweets NEG NEU POS 6
nlptown/bert-base-multilingual-uncased-sentiment BERT product reviews in six languages 1-5 stars 7
bhadresh-savani/bert-base-uncased-emotion BERT Unknown sadness, joy, love, anger, fear, surprise 8
Americo/emergency-text-classification Unknown Unknown
- LABEL_1, LABEL_0 We can fine-tune sentiment analysis tools to create better messages AI is transferable
AI is more than machine AI is technology a tool data people culture knowledge This picture generated by AI
NRC has a strategic plan Vision and Outcomes
Continue to keep pace with technological innovations to ensure the safe and secure use of AI in NRC-regulated activities
AI framework and skilled workforce to review and evaluate the use of AI in NRC-regulated activities Five strategic goals Goal 1: Ensure NRC Readiness for Regulatory Decisionmaking 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
Moving forward Embrace new and innovative ways to apply AI in preparing for and responding to emergencies Make use of performance-based and technology-inclusive EP regulations to adopt AI solutions Develop staff AI knowledge, skills, and abilities Maintain awareness and identify new AI use cases Maintain strong partnerships with domestic and international counterparts This picture generated by AI
Todd Smith, PhD todd.smith@nrc.gov SHARED INTENTIONALITY FOR PUBLIC HEALTH AND SAFETY What can AI not do for you?