Data Reduction in Deterministic Neutron Transport Calculations Using Machine Learning

TL;DR

This paper introduces a machine learning approach using autoencoders and DJINN to significantly reduce data storage needs in neutron transport calculations, while maintaining accuracy and efficiency.

Contribution

It presents a novel ML-based method that replaces large cross section matrices with compressed neural network models, reducing data storage by 94% in high-group problems.

Findings

94% data reduction achieved with autoencoders and DJINN

Scalar flux preservation demonstrated across various conditions

Decreased computational wall clock times

Abstract

Neutron cross section matrices for fission and scattering data are required for each material, temperature, and enrichment level to calculate the neutron transport equation accurately. This information can be a limiting factor when using the multigroup discrete ordinates (SN) method when the number of energy groups is large. Machine Learning (ML) can be used to replace the need for the cross section matrices by reproducing the function that maps the scalar flux to the scattering and fission sources. Through the use of autoencoders and Deep Jointly-Informed Neural Networks (DJINN), the data storage requirements are reduced by 94% of the original data for a 618 group problem. This is accomplished while preserving the scalar flux, maintaining generality, and decreasing wall clock times.