Data-driven analysis of the rotational energy landscapes of an organic cation in a substituted alloy perovskite

Abstract

Lead-free hybrid organic–inorganic perovskites have recently emerged as excellent materials particularly in highly potential yet low-cost photovoltaic technologies. Calculations have previously suggested that CH₃NH₃BiSeI₂ can be used as an alternative material for the highly studied CH₃NH₃PbI₃ due to its eco-friendliness and comparable performance. Herein, with the aid of Euler angles, the interplay between the organic CH₃NH₃ (MA) cation and the inorganic BiSeI₂ framework, obtained from first-principles calculations, is thoroughly scrutinised by means of the multidimensional total energy landscape. The highest peak of 17.9 meV per atom, protruding from the average plateau of 9 meV per atom within the four-dimensional topography, is equivalent to 208 K, the temperature at which the MA cations freely reorient. Moreover, the complexity of the angle–energy relationship is mitigated by exploiting a high-fidelity simulation based on deep learning. The deep artificial neural network of five hidden layers with 500 neurons, each fed by ten descriptors – three Euler angles and seven various bond lengths – predicts the total energies with an accuracy within the root mean square error of 0.39 ± 0.03 meV per atom for the test dataset. This novel statistical model in turn offers a tantalising promise to provide an accurate prediction of this material's energies, while diminishing the need for costly first-principles calculations.