A large number of inorganic Zintl phases have been discovered by solid-state chemists driven by chemical intuition and through serendipitous accidents. These discoveries have only scratched the surface, given the vast compositional and structural diversity that Zintl phases can accommodate. The large chemical space of Zintl phases, as well as intermetallic compounds in general, remain under-explored. Data-driven computational chemistry has made great strides in charting large chemical spaces that cannot be tractably explored with first-principles methods. Here, we use graph neural networks and the upper bound energy minimization approach to efficiently scan a chemical space of >90,000 hypothetical Zintl phases and discover 1812 new thermodynamically stable phases with 90% accuracy, as validated with DFT. Some of the predicted Zintl phases have been synthesized and reported in independent studies in the literature. We show that our approach is more than 2X more accurate than M3GNet (40% accuracy) on the same dataset. Using a random forest model and SHAP analysis, we demonstrate the critical role of ionic bonding in stabilizing Zintl phases. Finally, we showcase an example that highlights the need to carefully consider cation disorder in computational materials discovery workflows.
