Deterministic and Probabilistic Deep Learning in Predicting Reactor Physics of a Source-Driven Subcritical System

Gatchalian, Ronald Daryll Espino

Deterministic and Probabilistic Deep Learning in Predicting Reactor Physics of a Source-Driven Subcritical System

Date

2023-11-15

Authors

Gatchalian, Ronald Daryll Espino

Abstract

Spatial, spectral, and modal effects prominent in subcritical regime led to challenges in evaluating integral physics parameters of source-driven subcritical systems. These effects cannot be natively accounted by standard experimental techniques based on Point Reactor Kinetics (PRK) primarily developed for critical systems. Deviation from PRK assumptions led to non-ideal response such as dependence to detector position of evaluated reactivity coefficient k-effective, presence of multiple-Alpha modes, and increasing reactivity coefficient k-effective bias as system becomes deeply subcritical. Capability to determine reactivity coefficient k-effective and other kinetics and subcritical parameters from system observables is paramount for ensuring nuclear safety by maintaining reactivity margins in Subcritical Assemblies (SCA), and safe approach to critical state in research reactors. This dissertation introduces a data-driven methodology based on Deep Learning (DL) for predicting the reactor physics parameters of an SCA by mapping from directly measurable properties like core arrangement, reaction rates, and detector response. Deterministic and Probabilistic deep neural networks were configured through supervised learning approach using simulation data from physics-based neutronics calculations covering both stationary state in source-equilibrium, and in dynamic state in Pulsed Neutron Source experiment. Optimized hyperparameters, architecture priors, preprocessing technique, and input data modality were assessed by inference on a withheld Test set. Test metrics showed accurate DL predictions with coefficient for determination greater than or equal to 0.99 for reactivity coefficient k-effective, lambda effective, orbital angular momentum of the unpaired nucleon, ks, Alpha, and theoretical intensity-independent effective similarity parameter that surpassed baseline performance derived from statistical and criticality safety considerations with superior accuracy and scalability than traditional Machine Learning approaches. Compared to standard reactivity measurement techniques, such as Amplified Source, Area-ratio, and Slope-fit methods, DL consistently provided accurate predictions well within plus-minus1$ of true values regardless of source location and subcriticality, eliminating the need for auxiliary corrective measures. Other physics parameters with no equivalent experimental techniques were also accurately predicted. This capability is unique from neutronics evaluations as measurable quantities are leveraged as predictors. These advantages extended to probabilistic DL capable of modelling aleatoric and epistemic uncertainties thus providing confidence bounds. Overall, the novel application of DL in subcritical physics evaluation shows promise in an operational setting, addressing current analytical challenges which can enhance the safety and performance of SCAs, and emerging subcritical nuclear systems for waste transmutation.

Keywords

Subcritical Assembly, Subcritical Multiplication Factor, Neutron Source Efficiency, Reactivity Determination, MCNP, Machine Learning, Deep Learning, Bayesian Neural Network, Monte Carlo Dropout, Uncertainty Quantification