Machine learning prediction of electronic density of states using Lorentzian basis expansion

Abstract

In this work we propose an approach for fast and reliable prediction of electronic density of states (DOS), which provides valuable information about physical properties of crystals and particularly important for discovery and screening of new materials. Presented method is based on graph neural networks (GNN) and represents target DOS in form of lorentzian basis expansion, what makes it compact and convenient in practice. Its algorithm requires only atomic numbers and positions as input and very fast computationally even for extended unit cells, while traditional approach, based on density functional theory (DFT) can be prohibitively expensive in such cases. Developed method predicts contributions of s , p , d and f atomic shells to the total DOS; such contributions provide valuable information for interpretation of x-ray spectroscopic results. Presented method is tested on several classes of crystals: metals, ionic crystals, molecular crystals and covalent atomic crystals (like diamond) and in all cases good agreement with ground-truth results is observed. So far, developed method can be a helpful tool for material study. The developed GRADE software is available for free on https://scd.chem.msu.ru/ .