Machine learning density functionals from the random-phase approximation

TL;DR

This paper develops a machine learning-based density functional that approximates the random-phase approximation, enabling more accurate and scalable first-principles calculations in chemistry and materials science.

Contribution

It introduces a novel ML-derived functional that extends RPA accuracy to larger systems and broader chemical spaces, surpassing standard gradient functionals.

Findings

ML-RPA outperforms standard functionals in accuracy for diamond surfaces and water

ML-RPA extends RPA applicability to larger systems and longer time scales

The approach combines supervised and unsupervised learning techniques

Abstract

Kohn-Sham density functional theory (DFT) is the standard method for first-principles calculations in computational chemistry and materials science. More accurate theories such as the random-phase approximation (RPA) are limited in application due to their large computational cost. Here, we construct a DFT substitute functional for the RPA using supervised and unsupervised machine learning (ML) techniques. Our ML-RPA model can be interpreted as a non-local extension to the standard gradient approximation. We train an ML-RPA functional for diamond surfaces and liquid water and show that ML-RPA can outperform the standard gradient functionals in terms of accuracy. Our work demonstrates how ML-RPA can extend the applicability of the RPA to larger system sizes, time scales and chemical spaces.