Enhancing uncertainty analysis: POD-DNNs for reduced order modeling of neutronic transient behavior

Abstract

In reactor safety analysis, sensitivity analyses on critical parameters are essential for ensuring the reliability of safety conclusions, particularly regarding transient behavior, which often requires time-consuming computations. Developing surrogate models presents a promising solution. This paper extends the Proper Orthogonal Decomposition-Radial Basis Function (POD-RBF) framework to the 3D Light Water Reactor Core Transient Benchmark (3DLWRCT) for control rod ejection accidents. The primary aim is to simulate transient behavior under random perturbations in the macroscopic neutronic cross-sections of fuel assemblies.

Our results indicate that the traditional POD-RBF approach struggles to accurately reconstruct the highly nonlinear transient system, whether through spatiotemporal folding or spatial data reduction. To overcome these challenges, we enhance the model by integrating Deep Neural Networks (DNNs) and employing the Tree-structured Parzen Estimator for optimal neural network architecture selection. This improved approach significantly increases the accuracy of the surrogate models, demonstrating its feasibility and effectiveness. The integration of DNNs offers a deeper understanding of complex interactions within the reactor core, effectively capturing nonlinearities and yielding reliable predictions even under uncertainty.