Novel deep learning-based evaluation of neutron resonance cross sections

Abstract

Neutron resonance cross sections are essential in many nuclear science fields and applications. However, their evaluation and application are extremely complicated. Additionally, the high-frequency, super-wide spectral range of these cross sections cannot be readily approximated by a deep neural network (DNN). To address this issue, we propose a single phase-shift DNN (SPDNN) in which a phase-shift layer is added to a conventional DNN before the output layer to enable wideband processing. Compared with multinetwork algorithms, SPDNN represents a more compact and efficient network, with far fewer parameters. The proposed SPDNN is used to learn the neutron resonance cross sections of 235U fission from the evaluated and experimental libraries, and the results demonstrate its capability as an easy-to-implement, efficient method for approximating the evaluated resonance cross sections and evaluating the experimental data. This study represents the first application of deep learning to the evaluation of highly complex neutron resonance cross sections by adapting DNN.

Keywords

• Neutron resonance cross section
• Deep neural network
• Nuclear data evaluation