Machine-learned approximations to Density Functional Theory Hamiltonians

TL;DR

This paper introduces a machine learning approach to rapidly and accurately predict DFT Hamiltonians, significantly reducing computational costs in electronic structure calculations while maintaining high accuracy and transferability across materials.

Contribution

The authors develop a machine learning method using Kernel Ridge Regression to predict DFT Hamiltonians, enabling automated, scalable, and transferable electronic structure predictions.

Findings

Predicted Hamiltonians yield electronic spectra close to DFT results

Method is basis-independent and applicable to various materials

Achieves high accuracy with reduced computational effort

Abstract

Large scale Density Functional Theory (DFT) based electronic structure calculations are highly time consuming and scale poorly with system size. While semi-empirical approximations to DFT result in a reduction in computational time versus ab initio DFT, creating such approximations involves significant manual intervention and is highly inefficient for high-throughput electronic structure screening calculations. In this letter, we propose the use of machine-learning for fast and accurate prediction of DFT Hamiltonians. Using suitable representations of atomic neighborhoods and Kernel Ridge Regression, we show that an accurate and transferable prediction of DFT Hamiltonians for a variety of material environments can be achieved. Electronic transmission spectra computed using predicted Hamiltonians compare accurately with their DFT counterparts. The method is independent of the specifics of the DFT basis or material system used and can easily be automated and scaled for predicting Hamiltonians of any material system of interest.