Unified Graph-Based Interatomic Potential for Perovskite Structure Optimization.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Halide perovskites (HaPs) show promise for optoelectronics and catalysis.
• Their complex compositional space (alloys, defects, surfaces) hinders optimization.
• Efficient exploration of HaP potential energy surfaces is challenging.

Purpose of the Study:

• To develop a unified graph-based deep learned interatomic potential for HaPs.
• To enable efficient optimization and prediction of energetics across diverse HaP structures.
• To navigate the complex potential energy surface (PES) of HaPs.

Main Methods:

• Trained a M3GNet-based machine learning interatomic potential (IAP) on a comprehensive DFT data set of ~12,000 HaP structures.
• Included bulk alloys, native/impurity defects, and surface slabs in the training data.
• The IAP framework was trained on energies, forces, and stresses for gradient-based optimization.

Main Results:

• The M3GNet-IAP demonstrated robust generalizability across the complex HaP PES.
• Achieved low prediction errors: energies (3.7 meV/atom), forces (16.5 meV/Å), and stresses (5.5 MPa).
• Accurately predicted formation, decomposition, defect, and surface energies for HaPs.

Conclusions:

• The unified surrogate model offers a holistic approach to HaP geometry optimization.
• This method facilitates efficient exploration of diverse structural variations in HaPs.
• The model is transformative for discovering new HaP compositions, defects, dopants, and surface properties.