Machine learning-empowered study of metastable γ-CsPbI3 under pressure and strain

Abstract

Metastable γ-CsPbI₃ is a promising solar cell material due to its suitable band gap and chemical stability. While this metastable perovskite structure can be achieved via introducing external pressure or strain, experimenting with this material is still challenging due to its phase instability. In this work, we present the first instance of exploiting various machine learning (ML) models to efficiently predict the band gap and enthalpy of metastable γ-CsPbI₃ under pressure or strain while identifying key structural features that determine these properties. ML models trained on experimentally benchmarked, first-principles calculation datasets exhibit excellent performance in predicting the behavior of tuned systems, comparable to predictions made for ambient material databases. In particular, graph neural networks (GNNs) that explicitly include a graph encoding the bond angle information outperform other ML models in most scenarios. The pressure-tuned system demonstrates a strong linear relationship between structural features and properties, effectively captured by global structural features using linear regression models. In contrast, the strain-tuned system shows a non-linear relationship, exhibiting superior prediction performance using GNNs trained on local environments. This study opens up opportunities to apply and develop ML models for understanding and designing materials under extreme conditions.