Instability-Avoiding Active Learning for Cluster Expansions in Complex Multielement Materials

Abstract

The high efficiency of cluster expansions make them appealing for studying chemical disorder in complex composition spaces such as in multi-principal element alloys (MPEAs). Several works have attempted to address the rapidly growing training cost with number of chemical species through active learning, transfer learning, and chemical embedding. However, many composition spaces have large regions where the host lattice becomes dynamical unstable, which are often avoided a priori so as to not generate expensive but inapplicable training data. Here, we demonstrate a procedure for integrating stability classification within an active learning workflow to autonomously avoid calculations for unstable structures. Our workflow augments the stability classification procedure with Mahalanobis distance-based structure selection to ensure model robustness by training set diversification. We benchmark our methods by training a cluster expansion for the complex FCC MPEA spanning the Ni-Fe-Cr-Al-Ti-Si alloy space, in which only Ni and Al are thermodynamically stable as FCC