Proceedings of International Conference on Hybrid and Organic Photovoltaics (HOPV23)
Publication date: 30th March 2023
Hybrid halide perovskite (HHP) materials have shown extraordinary performance and attracted great interest in the solar energy conversion community.[1] However, there are still several major challenges to be tackled before their commercialisation, the first of which is its phase instability under ambient temperature with an undesired phase transition into low-efficiency phases. Molecular dynamics (MD) is a natural choice for simulating these phase transition behaviours at the atomic scale, and the recent emergence of machine learning potential (MLP) methods enables large-scale MD simulation with density functional theory level accuracy.[2] This work utilised an existing kernel-based machine learning force field method [3] to produce MD trajectories and developed a quantitative analysis method to study the dynamic properties of HHP. The dynamic analysis covers a wide variety of properties, including octahedral tilting and distortion, pseudo-cubic lattice parameters, molecular orientation and displacement, as well as the spatial correlation of these properties. Six different HHP materials (A = Cs, MA, FA and X = I, Br) were selected as the example compounds for demonstration as they are well-studied and of experimental interest. The analysis method is implemented for the validation of force field training and interpretation of dynamic properties from large-scale MD trajectories. Moreover, it provides a machine learning featurization pathway for perovskite materials focusing on their dynamic behaviours. The constituent black phases of all listed materials were successfully reproduced and analyzed in detail. The corresponding phase transition temperatures were found with good agreement to experimental results. This combination of MLP methods with dynamic property analysis has shown great potential in studying realistic behaviours of perovskite materials and is easily transferable to more complicated structures.
We thank A. Ganose, S. Kavanagh, J. Klarbring, K. Tolborg, M. Dubajic, A. Iqbal, W. Baldwin and K. Morita for helpful discussions. This work used the ARCHER2 UK National Supercomputing Service.