RAFFLE: active learning accelerated interface structure prediction

Abstract

Interfaces between materials are critical to the performance of many devices, yet predicting their structure is computationally demanding due to the vast configuration space. We introduce RAFFLE, a software package for efficiently exploring low-energy interface configurations between arbitrary crystal pairs, enabling the generation of ensembles of interface structures. RAFFLE leverages physical insights and genetic algorithms to intelligently sample configurations, using dynamically evolving 2-, 3-, and 4-body distribution functions as generalised structural descriptors. These descriptors are refined through active learning to guide atom placement strategies. RAFFLE performs well across diverse systems, including bulk materials, intercalation compounds, and interfaces. It correctly recovers known bulk phases of aluminum and MoS₂, and predicts stable phases in intercalation and grain-boundary systems. For Si∣Ge interfaces, it finds intermixed and abrupt structures to be similarly stable. By accelerating interface structure prediction, RAFFLE offers a powerful tool for materials discovery, enabling efficient exploration of complex configuration spaces.