Machine Learning Approaches for Performance Assessment of Nuclear Fuel Assemblies Subject to Seismic-Induced Impacts

Abstract

In pressurized water nuclear reactors, the seismic performance of fuel assemblies is governed by their spacer grids (SGs) which may experience impacts with neighboring fuel assembly SGs or with the core barrel, depending on the intensity of the seismic event. Nonlinear dynamic analysis aiming at computing the maximum permanent deformation in a statistic framework is computationally demanding due to the different possible core configurations and the dimension of the dataset of seismic excitations. Hence, surrogate models trained by the physics-based dynamic model are proposed to analyze different scenarios, i.e., explore the space of potential core configurations and seismic excitations. Starting from ground motion records corresponding to six levels of seismic hazard, the dynamic excitation at the elevation of the reactor pressure vessel is obtained via transfer functions. Correlation between different seismic intensity measures and the maximum permanent deformation is evaluated. The performance of two well-established surrogate models, namely, artificial neural networks (ANN) and Gaussian process (GP) for regression problems is analyzed and discussed. Bayesian techniques are adopted to enhance the robustness of the trained surrogate models by training sets of neural networks and estimating the hyper-parameter of the GP.