Machine learning for prediction of atom-projected properties

Abstract

The application of machine learning (ML) and deep learning techniques in materials science has revolutionized the process of material design and property prediction. ML models have been widely used to accurately predict various material properties, such as formation energy, band gap, and bulk modulus. However, interpreting the individual contribution of each atom to the overall material properties has remained an open challenge. In this paper, we introduce an atom-projected neural network (APNN) model to directly predict the individual contributions of each atom to the material properties. We demonstrate that our model achieves a great performance comparable to that of other graph-based neural networks while providing significantly improved interpretability. The proposed model has great potential for further applications in materials science, such as predicting thermal conductivity or bulk modulus, enabling researchers to understand how individual atoms contribute to the observed properties.