Lean CNNs for mapping electron charge density fields to material properties

Abstract

This work introduces a lean CNN (convolutional neural network) framework, with a drastically reduced number of fittable parameters (<81K) compared to the benchmarks in current literature, to capture the underlying low-computational cost (i.e., surrogate) relationships between the electron charge density (ECD) fields and their associated effective properties. These lean CNNs are made possible by adding a pre-processing step (i.e., a feature engineering step) that involves the computation of the ECD fields' spatial correlations (specifically, 2-point spatial correlations). The viability and benefits of the proposed lean CNN framework are demonstrated by establishing robust structure-property relationships involving the prediction of effective material properties using the feature-engineered ECD fields as the only input. The framework is evaluated on a dataset of crystalline cubic systems consisting of 1410 molecular structures spanning 62 different elemental species and 3 space groups.