A method for the automatic generation of a minimal basis set of structural templates for material phase-space exploration

Abstract We present a method for efficiently predicting approximate binary convex hulls through the automatic generation of a minimal basis set of representative structural templates. The approach is based on the assumption that chemical space can be partitioned into a limited number of regions with similar bonding characteristics, such that a compact set of templates can capture the key structural and energetic trends across a wide range of compositions. Templates are selected via a data-driven algorithm informed by evolutionary structure searches, and subsequently reused to estimate formation enthalpies across diverse systems using limited DFT relaxations. We validate the method with a comprehensive benchmark dataset of 105 binary element pairs. Convex hulls predicted by our method are compared to evolutionary algorithm predictions using multiple metrics, including composition overlap, hull depth error, structural agreement. When targeting an average enthalpy error of 0.1 eV/atom, sufficient for high-throughput thermodynamic screening—the method achieves a 15-fold reduction in the number of DFT relaxations while maintaining 87% accuracy in identifying stable compositions. The approach is designed for scalability and applications in high-throughput exploration of complex material spaces.