Performance assessment of high-throughput Gibbs free energy predictions of crystalline solids

Abstract

Crystalline solids are integral to a broad range of natural science and engineering applications and their Gibbs free energy $G$ is an important parameter in modeling their thermodynamics. However, predicting $G$ for solids remains a difficult task and an under-explored field in high-throughput thermochemistry. In this work, we benchmark the performance of the newest generation of machine learning (ML) predictions, machine learning interatomic potentials (MLIPs), and density functional theory in predicting $G$ within the harmonic and quasi-harmonic approximations against experimental data from 100–2500 K and for up to 784 compounds. Furthermore, these calculations are fed to a reaction network (RN) from which experimentally informed predictions can be made. We find that predictions of $G$ made by MLIPs display promising performance but with the help of the RN, simpler methods show similar or better performance. Nonetheless, we show that much of the calculated and experimental data for $G$ still lack the accuracy and precision required for some thermodynamic modeling applications. Finally, we apply the RN to predict the room temperature Gibbs free energy of formation and find that it performs satisfactorily but that improvements need to be made before these predictions can be used as reliable indicators of thermodynamic stability in general applications.