Deterministic and probabilistic deep learning in predicting reactor physics of a source-driven subcritical system

Authors

Ronald Daryll Gatchalian ⋅ PH Nuclear Reactor Operations Section, Department of Science and Technology – Philippine Nuclear Research Institute

Abstract

Effects in subcritical regime pose challenges for assessing integral physics parameters in source-driven subcritical systems. Standard experimental techniques based on Point Reactor Kinetics (PRK), primarily developed for critical systems, fail to account for these biasing effects. Departure from PRK assumptions results in non-ideal response of standard methods. It is crucial to accurately determine k_eff and other kinetics and subcritical parameters from system observables to ensure nuclear safety in Subcritical Assemblies (SCA) and research reactors.
A data-driven methodology based on Deep Learning (DL) was developed to predict the reactor physics parameters of an SCA by mapping from directly measurable properties like core arrangement, reaction rates, and detector response. Deterministic and Probabilistic DL were configured through supervised learning approach using simulation data from neutronics modelling of the Philippine Research Reactor-1 Subcritical Assembly for Training, Education, and Research (PRR-1 SATER). Uncertainty quantification of DL was performed using Monte Carlo (MC) Dropout and Bayesian neural network (BNN).
Test metrics showed accurate DL predictions with R² \geq 0.99 for k_eff, Λ_eff, l, k_s, α, and S_eff that surpassed baseline performance derived from statistical and criticality safety considerations. Comparison to Amplified Source Method (ASM), a standard reactivity measurement technique, indicated that ASM showed a reactivity bias of up to –3.59% Δk/k (–4.86). In contrast, DL had a maximum bias of only 0.38% Δk/k (0.5). Underestimation by ASM represents a nonconservative scenario in criticality safety, while DL proved robust against spatial effects influenced by source location and subcriticality.
Other physics parameters with no equivalent experimental techniques were also accurately predicted. These advantages extended to probabilistic DL capable also of modelling aleatoric and epistemic uncertainties; however, results favored BNN over MC Dropout, showing greater improvement with increasing training set size. Overall, the novel application of DL in subcritical physics evaluation shows promise in an operational setting, addressing current experimental challenges which can enhance the safety and performance of SCAs, and emerging subcritical nuclear systems for waste transmutation.

About the Speaker

Ronald Daryll Gatchalian, Nuclear Reactor Operations Section, Department of Science and Technology – Philippine Nuclear Research Institute

Ronald Daryll Gatchalian is a Science Research Specialist II at the Nuclear Reactor Operations Section of the Department of Science and Technology – Philippine Nuclear Research Institute (PNRI). He joined the Institute in 2016 where he built his experience in the field of nuclear science and technology particularly on nuclear reactor analysis, radiation transport modeling, radiation detection/measurement, radiological hazards assessment, and shielding design.
He holds a bachelor's degree in Applied Physics (Health Physics) from the University of the Philippines Manila, and a PhD in Nuclear Engineering from Texas A&M University, the largest and third top-ranking nuclear engineering department in the USA. His doctoral program was supported by the DOST – Science Education Institute through its Foreign Graduate Scholarships Program alongside PNRI as sending institution.
Dr. Gatchalian has published in peer-reviewed scientific journals such as Annals of Nuclear Energy and Nuclear Science and Engineering. Moreover, he has contributed conference papers at the Winter Conference and Expo of the American Nuclear Society and to the International Conference on Mathematics and Computational Methods Applied to Nuclear Science and Engineering. Dr. Gatchalian is also a member of the American Nuclear Society. Throughout his career, he completed various training courses and professional development programs sponsored by the Los Alamos National Laboratory, International Atomic Energy Agency, Japan Atomic Energy Agency, and Czech Technical University.

Downloads

Issue

2024: Proceedings of the 42nd Samahang Pisika ng Pilipinas Physics Conference

Article ID

SPP-2024-INV-3B-01

Section

Invited Presentations

Published

2024-07-04

How to Cite

[1]

RD Gatchalian, Deterministic and probabilistic deep learning in predicting reactor physics of a source-driven subcritical system, Proceedings of the Samahang Pisika ng Pilipinas 42, SPP-2024-INV-3B-01 (2024). URL: https://proceedings.spp-online.org/article/view/SPP-2024-INV-3B-01.