We have developed a methodology for a machine-learning acceleration of high-throughput computational alloy discovery in collaboration with the research group of Gus Hart from BYU. The approach relies on our machine-learning interatomic potentials, MTP, which are trained using our active learning approach. This allows for predicting new stable alloys with geometrically new structures, in contrast to the commonly used on-lattice models (such as the cluster expansion approach) that assume a fixed host lattice. We achieve a 1000-fold speed-up in comparison to the ab initio-based screening (e.g., with density functional theory) .
The corresponding article “Accelerating high-throughput searches for new alloys with active learning of interatomic potentials” was recently accepted to Computational Materials Science and now is available at arXiv.