RAFFLE: Active learning accelerated interface structure prediction

Abstract

Interfaces between materials play a crucial role in the performance of most devices. However, predicting the structure of a material interface is computationally demanding due to the vast configuration space, which requires evaluating an unfeasibly large number of highly complex structures. We introduce RAFFLE, a software package designed to efficiently explore low-energy interface configurations between any two crystals. RAFFLE leverages physical insights and genetic algorithms to intelligently sample the configuration space, using dynamically evolving 2-, 3-, and 4-body distribution functions as generalised structural descriptors. These descriptors are iteratively updated through active learning, which inform atom placement strategies. RAFFLE's effectiveness is demonstrated across a diverse set of systems, including bulk materials, intercalation structures, and interfaces. When tested on bulk aluminium and MoS$_2$, it successfully identifies known ground-state and high-pressure phases. Applied to intercalation systems, it predicts stable intercalant phases. For Si|Ge interfaces, RAFFLE identifies intermixing as a strain compensation mechanism, generating reconstructions that are more stable than abrupt interfaces. By accelerating interface structure prediction, RAFFLE offers a powerful tool for materials discovery, enabling efficient exploration of complex configuration spaces.