Looking Inside Halide Perovskites Using On-The-Fly Machine Learning Force Fields

Recent developments in the field of high efficiency perovskite solar cells are based on stabilization of the perovskite crystal structure of FAPbI₃ while preserving its excellent optoelectronic properties. Compositional engineering of, for example, MA or Br mixed into FAPbI₃ results in the desired effects, but detailed knowledge of local structural features, such as local (dis)order or cation interactions of formamidinium (FA) and methylammonium (MA), is still limited. This knowledge is, however, crucial for their further development. In this talk, I will present how recent developments in Machine Learning accelerated ab-initio Molecular Dynamics (MD) enable the required length and time scale simulations that can assist in solving some of the open questions. For example, the generated large isothermal-isobaric ensembles can be used to calculate the ¹H-¹H dipolar coupling coefficients that can be compared with those obtained by nuclear magnetic resonance spectroscopy measurements[1]. Furthermore, we can calculate the momentum resolved phonon spectrum by ‘measuring’ their frequencies and lifetimes directly from velocity auto-correlation functions[2]. We observe that the lifetime of many phonon modes is of the order of the inverse of its frequency. This shows the strong anharmonicity of the potential energy surface of halide perovskites, and the necessity of MD approaches. Overall, there is qualitative and quantitative agreement between experiment and theory on several experimental observables, which leads us to think that the simulated crystals structure trajectories provide a realistic insight into the ionic dynamics of the studied samples.<br/><br/>[1] Helen Grüninger et al. , J. Phys. Chem. C, 125, 3, 1742–1753 (2021)<br/>[2] Jonathan Lahnsteiner and Menno Bokdam, Phys. Rev. B 105, 024302 (2022)