Fast and Accurate Artificial Neural Network Potential Model for MAPbI₃ Perovskite Materials

Area of Science:

• Materials Science
• Computational Chemistry
• Solid-State Physics

Background:

• Hybrid organic-inorganic perovskites are crucial for photovoltaic and optoelectronic applications.
• Developing computationally efficient potential models for perovskite atomistic simulations is challenging due to their chemical complexity.
• High-fidelity simulations require accurate potential energy evaluation, often relying on computationally expensive ab initio methods.

Purpose of the Study:

• To demonstrate the efficacy of artificial neural network (ANN) models for accurate and efficient potential energy evaluation of methylammonium lead iodide (MAPbI3) perovskite.
• To assess the performance of ANN models in predicting structural and dynamic properties of MAPbI3.
• To highlight the potential of ANN models for accelerating atomistic simulations of complex materials.

Main Methods:

• Trained artificial neural network (ANN) models using extensive datasets of tetragonal MAPbI3 crystal structures, energies, and atomic forces derived from ab initio calculations.
• Validated the trained ANN models by predicting lattice parameters, energies, and atomic forces for cubic MAPbI3 perovskite.
• Incorporated atomic forces into the ANN training process to enable phonon mode extraction.

Main Results:

• ANN models accurately predicted lattice parameters and energies/atomic forces for cubic MAPbI3, showing excellent agreement with ab initio calculations.
• Phonon modes extracted using the trained ANN model closely matched results from ab initio calculations.
• ANN models achieved 10^4 to 10^5 times faster energy evaluations compared to Vienna Ab initio Simulation Package (VASP).

Conclusions:

• Artificial neural network models provide a computationally efficient and accurate approach for potential energy evaluation in MAPbI3 perovskite.
• ANN models significantly accelerate atomistic simulations, enabling exhaustive sampling of configuration spaces for property predictions.
• This methodology holds great promise for advancing the study of thermodynamic properties and phase stabilities in complex perovskite materials.