Efficient Gaussian Process Regression for prediction of molecular crystals harmonic free energies

TL;DR

This paper introduces an efficient Gaussian Process Regression framework for accurately predicting harmonic free energies of molecular crystals, enabling high-throughput thermodynamic calculations with minimal training data.

Contribution

The authors develop a transfer learning approach that reduces computational costs and achieves high accuracy in predicting free energies of molecular crystals using GPR models trained on empirical and DFT potentials.

Findings

Achieves mean absolute deviation below 0.04 kJ/mol/atom at 300 K.

Requires only 60 crystal structures for high accuracy.

Successfully predicts thermal lattice expansion and polymorph stability.

Abstract

We present a method to accurately predict the Helmholtz harmonic free energies of molecular crystals in high-throughput settings. This is achieved by devising a computationally efficient framework that employs a Gaussian Process Regression model based on local atomic environments. The cost to train the model with ab initio potentials is reduced by starting the optimisation of the framework parameters, as well as the training and validation sets, with an empirical potential. This is then transferred to train the model based on density-functional theory potentials, including dispersion-corrections. We benchmarked our framework on a set of 444 hydrocarbon crystal structures, comprising 38 polymorphs, and 406 crystal structures either measured in different conditions or derived from them. Superior performance and high prediction accuracy, with mean absolute deviation below 0.04 kJ/mol/atom at 300 K is achieved by training on as little as 60 crystal structures. Furthermore, we demonstrate the predictive efficiency and accuracy of the developed framework by successfully calculating the thermal lattice expansion of aromatic hydrocarbon crystals within the quasi-harmonic approximation, and predict how lattice expansion affects the polymorph stability ranking.