Physics guided discrepancy learning for neutron induced cross sections

Nuclear data, such as neutron inelastic scattering cross sections, are fundamental for numerous applications, including reactor design, astrophysics, and nuclear security. However, reaction model codes are limited in their ability to calculate cross sections with high accuracy due to model approximations and uncertainties in nuclear input parameters. This work aims to enhance the predictive capability of reaction modeling codes through discrepancy learning, a machine learning approach that models the difference between theoretical predictions and experimental data. Several machine learning architectures—including a feed-forward neural network, a one-dimensional convolutional neural network, Gaussian process regression, and a Bayesian multilayer perceptron—were evaluated for their ability to learn the discrepancies between theoretical cross sections calculated using a reaction modeling code and the “true” cross sections inferred from experimental data, i.e., the measurement–model mismatch. By systematically comparing these architectures, this work aims to determine which models most effectively capture the measurement–model mismatch and elucidate the underlying reasons for their relative performance. The expected outcome is a framework for physics-guided machine learning of neutron-induced reaction cross sections important for a wide range of applications.