Targeting Productive Composition Space Through Machine-Learning Directed Inorganic Synthesis

This work presents an approach to aid the discovery of novel inorganic solids by highlighting regions of underexplored, yet likely productive composition space using machine learning. A support vector regression (SVR) algorithm was constructed first to determine a compound’s formation energy (∆𝐸𝑓,SVR) based solely on chemical composition using data from 313,965 high-throughput density functional theory calculations. The resulting predicted formation energies (r² = 0.94; MAE = 8.5 meV/atom) were then used to construct zero-kelvin convex hull diagrams and identify compositions immediately on the hull, as well as +50 meV above the convex hull to capture potential compounds that are considered energetically unfavorable but that are still experimentally accessible. Using this methodology, four ternary composition diagrams, Y−Ag−Tr (Tr = B, Al, Ga, In), were explored owing to the diversity of chemistries as a function of triel element to provide experimental validation for the predictions. A particularly promising but unexplored region in the Y−Ag−In diagram was identified, and the ensuing solid-state high-temperature synthesis produced YAg_0.65In_1.35, which has not been reported. First-principle calculations were finally used to determine the ordering of Ag and In as well as confirm the crystal structure solution. Our combination of machine learning, inorganic synthesis, and computational modeling describes a new avenue where data-centric models and computation play a critical role in supporting the experimental examination of unexplored phase diagrams.