Combinatorial screening for new materials in unconstrained composition space with machine learning

Typically, computational screens for new materials sharply constrain the compositional search space, structural search space, or both, for the sake of tractability. To lift these constraints, we construct a machine learning model from a database of thousands of density functional theory (DFT) calculations. The resulting model can predict the thermodynamic stability of arbitrary compositions without any other input and with six orders of magnitude less computer time than DFT. We use this model to scan roughly 1.6 million candidate compositions for novel ternary compounds (${A}_{x}{B}_{y}{C}_{z}$), and predict 4500 new stable materials. Our method can be readily applied to other descriptors of interest to accelerate domain-specific materials discovery.