AI-based leakage prediction with uncertainty quantification and explainable AI for nuclear system

The integrity of the Reactor Coolant System (RCS) is central to nuclear power plant (NPP) safety, particularly under Small Break Loss-of-Coolant Accident (SBLOCA) conditions, where undetected minor leakage can rapidly escalate into severe system failures. Traditional detection techniques—based on deterministic solvers, threshold-based alarms, or first-principles physics—are inadequate for modern reactor diagnostics: they fail to model complex nonlinear interdependencies, cannot generalize beyond predefined fault patterns, and lack the flexibility to produce real-time, uncertainty-aware predictions under variable operating conditions. These limitations are further exacerbated in data-sparse nuclear operating environments where real-time observability is limited, and failure precursors are often masked. This thesis introduces a machine learning-based framework for interpretable and uncertainty-aware leakage prediction in pressurized water reactors (PWRs) to overcome these challenges. Synthetic datasets encompassing 15 SBLOCA scenarios were generated using the Generic Pressurized Water Reactor (GPWR) simulator coupled with the RELAP5 thermohydraulic code. The predictive framework employs three complementary learning models: Fully Connected Neural Networks (FCNNs), Random Forest Regressors (RFR), and Gradient Boosting Regressors (GBR). These were selected for their proven performance in regression tasks and their balanced trade-offs across accuracy, interpretability, data efficiency, and computational tractability—a critical consideration in nuclear safety applications. While more complex deep architectures (e.g., LSTMs, GNNs, transformers) exist in the literature, they often require large training sets, suffer from black-box opacity, and pose significant validation challenges in regulatory environments. In contrast, the chosen models allow rigorous benchmarking and clearer model validation with domain experts. Performance across hot leg, cold leg, and surge line scenarios shows that FCNNs excel at capturing complex nonlinearities, while RFR and GBR offer robustness and greater interpretability. A Bayesian-inspired input perturbation method is employed to quantify prediction confidence, generating uncertainty bounds critical for operator trust and safety margins. LIME-based explainability further reveals key contributing features—such as noble gas count, pressurizer level, and containment humidity—whose influence varies with component and fault severity. This work presents a unified AI pipeline for nuclear leakage prediction and demonstrates that carefully selected machine learning models can enhance early fault detection, reduce false alarms, and support data-driven maintenance in safety-critical nuclear environments.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Nuclear, Plasma, Radiolgc Engr

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Abusultan, Ahmed

Contributors dc:contributor

• Alam, Syed Bahauddin

• ANN
• Artificial Neural Network
• BWR
• Boiling Water Reactor
• CV
• Cross Validation
• FCNN Fully Connected Neural Network
• GPWR Generic Pressurized Water Reactor
• LIME Local Interpretable Model
• Agnostic Explanations
• LOCA
• Loss of Coolant Accident
• MAE Mean Absolute Error
• MSE
• Mean Squared Error
• NPP
• Nuclear Power Plant
• NRC
• Nuclear Regulatory Commission
• PRA
• Probabilistic Risk Assessment
• PWR
• Pressurized Water Reactor
• RELAP5
• Reactor Excursion and Leak Analysis Program
• RMSE
• Root Mean Squared Error
• RCS
• Reactor Coolant System
• SBLOCA
• Small Break Loss of Coolant Accident
• TPE
• Tree Structured Parzen Estimator
• UQ
• Uncertainty Quantification
• XAI
• Explainable Artificial Intelligence

Statement dc:rights

• © 2025 by Ahmed Abusultan. All rights reserved.

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129786

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129786