RadHull1—A rough but More Complete Convex Hull of Stability

By combining machine learning energy models trained on large, chemically diverse datasets with combinatorial enumeration of structure prototypes and all possible charge-balance-allowed assignments of elements to crystal sites, Radical AI built perhaps the most complete convex hull of thermodynamic stability to date. While below density functional theory in accuracy, it spans the entire space of plausible inorganic compounds and can be used as a guide to explore energetically favorable regions of phase space with more accurate methods. We use this hull internally to guide explorative synthesis with Radical AI's self-driving lab and show initial results on the new kinds of materials this may yield in practice. The complete convex hull is made freely available to the community in the hopes that it enables more far-flung exploration of uncharted chemical space in future research.