Fast and Accurate Artificial Neural Network Potential Model for Organometal Halide Perovskite
Hsin-An Chen1*, Chun-Wei Pao1
1Research Center for Applied Sciences, Academia Sinica, Taipei, Taiwan
* Presenter:Hsin-An Chen,
In the last decade, hybrid organometal halide perovskite materials are one of the promising materials for photovoltaic and optoelectronic applications because of its high absorption coefficient, high power conversion efficiency (>20%, increasing rapidly), and low fabrication cost. The thermodynamic properties as well as morphologies are vital for performance of perovskite materials; however, these calculations require exhaustive exploration of configuration space, which are infeasible for ab initio calculations due to system size limitations and (~103 atoms) and slow energy/force evaluation. In this work, we demonstrate that the artificial neural network (ANN) model can be applied to construct an efficient potential in evaluating the energy and atomic forces of organometal halide perovskite. The ANN potential was trained by training sets comprised of thousands of atomic images containing both atomic coordinations and energies from ab initio calculations and we demonstrate that the trained ANN model can successfully predict the energies of perovskite materials. Furthermore, we demonstrate that a up to 105 speedup in energy evaluation relative to ab initio calculations using VASP can be obtained by using the trained ANN model, thereby allowing the possibilities to exhaustively sample the configuration space of chemically-complex perovskite materials to predict structural properties such as phase stabilities and phase diagrams.

Keywords: hybrid organometal halide perovskite, artificial neural network