Machine-learning correction to density-functional crystal structure optimization

TL;DR

This paper introduces a machine learning approach to improve the accuracy of density functional theory predictions of crystal structures, specifically lattice parameters and volumes, by correcting PBE and PBEsol calculations using experimental data.

Contribution

The study presents a simple, explainable machine learning correction method that enhances DFT predictions of crystal lattice parameters and volumes, reducing errors significantly.

Findings

Machine learning improves PBE volume predictions to match PBEsol accuracy.

Error in PBEsol calculations reduced by 35% with ML correction.

Lattice constant errors decreased by a factor of 3 to 5.

Abstract

Density functional theory is routinely applied to predict crystal structures. The most common exchange-correlation functionals used to this end are the Perdew-Burke-Ernzerhof (PBE) approximation and its variant PBEsol. We investigate the performance of these functionals for the prediction of lattice parameters and show how to enhance their accuracy using machine learning. Our dataset is constituted by experimental crystal structures of the Inorganic Crystal Structure Database matched with PBE-optmized structures stored in the materials project database. We complement these data with PBEsol calculations. We demonstrate that the accuracy and precision of PBE/PBEsol volume predictions can be noticeably improved a posteriori by employing simple, explainable machine learning models. These models can improve PBE unit cell volumes to match the accuracy of PBEsol calculations, and reduce the error of the latter with respect to experiment by 35%. Further, the error of PBE lattice constants is reduced by a factor of 3--5. A further benefit of our approach is the implicit correction of finite temperature effects without performing phonon calculations.