Interpretable Feature Extraction for the Numerical Particle System

Abstract

Particle system analysis is important but costly in solving many physical problems since particles are the basic composition of almost everything, and machine learning is increasingly used in optimization for numerical simulations. The most important and difficult job is to make the particle system understandable by the machine learning model, which usually called feature extraction. In this paper, a novel method and an accurate physical interpretation for feature extraction of cascade defects data, an important particle system in reactor pressure vessel analysis, was proposed. Four strategies were designed to extract features from cascade defects data in which the DAPP shows the best performance. This study shows that feature extraction based on physical information has a positive significance for the analysis of particle systems and provide a theoretical support for improving the physical interpretability of machine learning models on particle systems.