Computational stability investigations of perovskite alloys with machine learning

Perovskite solar cells are a promising alternative to silicon-based photovoltaics, but their adoption is hindered by toxicity concerns and insufficient long-term stability. Although compositional engineering has mitigated these problems to some degree, identifying optimal materials within complex alloy spaces has proven challenging. Computational materials science can help guide compositional engineering, yet accurate simulations of perovskite alloys are prohibitively expensive with conventional quantum-mechanical methods. In this thesis, I accelerated computations with machine learning (ML) to enable stability investigations of perovskite alloys.

In our computational framework, an ML interatomic potential (MLIP) is fitted to data from density functional theory (DFT) simulations of perovskite alloys. MLIPs enable rapid ML evaluations of large numbers of mixing configurations, providing insight into alloy stability through convex hull and free energy analyses. To model complex materials, including ternary alloys and hybrid perovskites, our work introduces an efficient data generation and model training scheme based on clustering analysis and active learning. In addition, a two-stage ML approach was developed to predict relaxation energies directly from unrelaxed structures, substantially reducing computational cost.

The methodology was applied to the inorganic CsPb(Cl/Br)3 alloy space and the hybrid systems (Cs/FA)Pb(Br/I)3 and (Cs/FA)Sn(Br/I)3, where FA stands for formamidinium. The fitted MLIPs achieve predictions at quantum-mechanical accuracy. Our efficient data generation scheme was validated on CsPb(Cl/Br)3 data, demonstrating considerable reduction in the number of DFT calculations necessary for model training. The two-level ML workflow that was applied for the relaxation effects in hybrid perovskite alloys accelerated computations substantially, providing a total speedup of approximately nine orders of magnitude in comparison to DFT.

For CsPb(Cl/Br)3, convex hull analysis predicts two stable alloy compositions at 0 K, while free energy calculations in the same alloy space suggest stabilization across the entire composition range at operational temperatures due to entropy effects. For (Cs/FA)Pb(Br/I)3, computed free energy landscapes are in qualitative agreement with existing experimental evidence, showing a high degree of stability for low-Cs, low-Br quaternary alloy compositions. The stable composition regions are narrower for the Sn-based material space, providing fewer opportunities for compositional engineering. Overall, the computational framework developed in this thesis enables accurate simulations of complex mixed perovskites, supporting the discovery and optimization of better photovoltaic materials.