(9c) Machine Learning the Thermochemistry of All Inorganic Crystalline Solids

High-throughput, theory-driven computational materials design and discovery has become an essential tool in materials science and engineering. The data-driven and cost-effective framework for accelerated discovery introduced by the Materials Genome Initiative (MGI) has transformed the scale and rate of materials development by exploiting the predictive ability of quantum chemical computational methods. In turn, 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 primarily by low-cost density functional theory (DFT) total energy calculations (inherently performed at 0 K). 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, limiting the success of MGI approaches for existing and emerging high temperature applications. In this report, 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 (e.g., using PBE+U). Implementing our approach on the entirety of open materials databases (e.g., MaterialsProject) reveals interesting trends with respect to the stability (and metastability) of all inorganic crystalline solids. Specifically, we quantify the fraction of compounds in the Inorganic Crystal Structure Database (ICSD) predicted to be thermodynamically stable, quantify the magnitude of metastability for thermodynamically unstable structures, and assess chemistry-specific trends in these predictions.