Machine Learning the Gibbs Energies of Inorganic Crystalline Solids

The drive to rapidly predict, screen, and optimize materials using first-principles calculations has led to large datasets and the construction of open-source databases that are populated almost exclusively by temperature-independent density functional theory (DFT) calculations. Although these datasets are an invaluable resource, their predictive ability at finite temperatures is limited and current methods for evaluating the missing temperature dependencies are computationally prohibitive. In this work, we show that compound Gibbs energies of formation can be predicted with accuracy comparable to the quasiharmonic approximation at finite temperatures of up to at least 1800 K and at the same computational expense as a DFT total energy calculation. Implementing our approach on the Materials Project database reveals interesting trends with respect to the stability (and metastability) of inorganic crystalline solids. Specifically, we quantify as a function of temperature the fraction of compounds in the Inorganic Crystal Structure Database (ICSD) predicted to be thermodynamically stable and the magnitude of metastability for thermodynamically unstable structures and identify chemical spaces which are found to have anomalous ranges of accessible metastability.