(A) study on feasibility of functionalising neutron cross-sections using artificial neural networks

Abstract

This study has been aimed at testing an unconventional way of storing and reproducing temperature dependent cross-section data with the ultimate goal of improving computational times of large scale Monte Carlo calculations . Artificial neural networks(ANNs) have been trained, using ACE-format cross-section data, to provide a continuous cross-section output based on any energy and temperature input within the trained region. The study has been focused on the absorption and scattering cross-section of $^{238}$U due to the complexity and importance of that particular isotope. Based on current results, final file sizes are projected to end up in the 5 to 7 megabyte range for absorption and scattering respectively. Cross-sections reproduced using ANNs generally show relative errors of less than 0.1%, in particular for scattering and in the resonance peaks for absorption. Between absorption resonances the relative error is generally less than 1.0% but reaches up to 5.0% in a few cases.