Lean CNNs for mapping electron charge density fields to material properties

TL;DR

This paper presents a lean CNN framework with significantly fewer parameters for predicting material properties from electron charge density fields, utilizing feature engineering of spatial correlations to improve efficiency and robustness.

Contribution

The study introduces a novel lean CNN architecture combined with feature engineering of spatial correlations, reducing model complexity while accurately predicting material properties from ECD fields.

Findings

Achieved robust structure-property predictions with less than 81K parameters.

Demonstrated effectiveness on a diverse dataset of crystalline systems.

Reduced computational cost compared to existing models.

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.