Linear Jacobi-Legendre expansion of the charge density for machine learning-accelerated electronic structure calculations

Abstract

As the go-to method to solve the electronic structure problem, Kohn-Sham density functional theory (KS-DFT) can be used to obtain the ground-state charge density, total energy, and several other key materials' properties. Unfortunately, the solution of the Kohn-Sham equations is found iteratively. This is a numerically intensive task, limiting the possible size and complexity of the systems to be treated. Machine-learning (ML) models for the charge density can then be used as surrogates to generate the converged charge density and reduce the computational cost of solving the electronic structure problem. We derive a powerful grid-centred structural representation based on the Jacobi and Legendre polynomials that, combined with a linear regression built on a dataefficient workflow, can accurately learn the charge density. Then, we design a machine-learning pipeline that can return energy and forces at the quality of a converged DFT calculation but at a fraction of the computational cost. This can be used as a tool for the fast scanning of the energy landscape and as a starting point to the DFT self-consistent cycle, in both cases maintaining a low computational cost.