Physics-Informed Neural Networks for the safety analysis of nuclear reactors

Abstract

This work explores the development of surrogate models for estimating the evolution of quantities of interest during nuclear reactor accident scenarios. Physics-Informed Neural Networks (PINNs) offer a promising surrogate modelling approach because they allow integrating laws of physics and domain knowledge into traditional Neural Network (NN) surrogates. Specifically, the proposed solution incorporates an additional term in the PINN loss function to enforce physics-based constraints in correspondence of allocation points, which are randomly sampled points whose corresponding target output is not known. As a result, accuracy of the estimation of the quantities of interest and their adherence to the laws of physics are improved. Applications to a synthetic case study and to the response of a nuclear microreactor system during a Loss of Heat Sink scenario confirm that the developed surrogate model based on PINN with allocation points improves the estimation accuracy with respect to other state-of-the-art methods.