posted on 2019-06-24, 15:05authored byHsin-An Chen, Chun-Wei Pao
Hybrid organic–inorganic perovskite
materials are promising materials for photovoltaic and optoelectronic
applications. Nevertheless, the construction of a computationally
efficient potential model for atomistic simulations of perovskite
with high fidelity to ab initio calculations is not a trivial task
given the chemically complex nature of perovskite in terms of its
chemical components and interatomic interactions. In the present study,
we demonstrate that artificial neural network (ANN) models can be
employed for efficient and accurate potential energy evaluation of
MAPbI3 perovskite materials. The ANN models were trained
using training sets composed of thousands of atomic images of tetragonal
MAPbI3 crystals, with their respective energies and atomic
forces obtained from ab initio calculations. The trained ANN models
were validated by predicting the lattice parameters and energies/atomic
forces of cubic MAPbI3 perovskite and had excellent agreement
with ab initio calculations. The phonon modes could also be extracted
using the trained ANN model with good agreement with ab initio calculations,
provided that the atomic forces were incorporated into the training
processes. Finally, we demonstrate that for a given system size, the
trained ANN model offers 104 to 105 faster time
consumption per energy evaluation relative to ab initio calculations
using Vienna Ab initio Simulation Package, demonstrating the potential
of the ANN model for exhaustively sampling the configuration spaces
of chemically complex materials for predictions of thermodynamic properties
and phase stabilities.