Active Learning Guided Optimization of Perovskites Under Extreme Conditions

Halide perovskites have emerged as promising materials for solar cells, demonstrating a dramatic increase in power conversion efficiencies over the past decade. However, organic-inorganic halide perovskites face chemical instability under heat and moisture conditions due to their volatile organic A-site cations. In contrast, inorganic halide perovskites exhibit better chemical stability but suffer from phase instability at ambient conditions due to the small size of the cesium ion at the A-site. The perovskite phase, characterized by corner-sharing octahedra, spontaneously transitions to a thermodynamically more stable but non-functional phase with edge- or face-sharing octahedra under ambient conditions. This phase instability has hindered the commercialization of these materials. One effective strategy to maintain the functional perovskite phase is to achieve a metastable phase by applying pressure or strain. The functional perovskite phase can be metastably preserved when quenched to ambient conditions. This underscores the importance of understanding the relationship between material structure and property changes under extreme conditions. Due to the experimental challenges in accessing metastable halide perovskites, first-principles calculations and machine learning (ML) techniques have been used to simulate extreme conditions and corresponding material properties. Our prior work demonstrated that various ML models achieved highly accurate predictions of the properties of halide perovskites based on their tuned structures under pressure or strain conditions [1]. Motivated by these findings, in this study, we applied an active learning approach to ab initio calculations to accelerate the prediction of the optimal synthetic conditions for achieving metastable halide perovskites with high stability and desirable properties. We then used these prediction results to design experimental validation. Our research provides valuable insights into the development of an efficient workflow for data-driven materials discovery under extreme conditions.<br/><br/>[1] Han, Minkyung, et al. "Machine learning-empowered study of metastable γ-CsPbI3 under pressure and strain." Journal of Materials Chemistry A 12.18 (2024): 11082-11089.