MACE foundation models for lattice dynamics: a benchmark study on double halide perovskites

Area of Science:

• Materials Informatics
• Computational Materials Science
• Artificial Intelligence in Chemistry

Background:

• Foundation models, such as MACE, represent a breakthrough in universal potentials for inorganic solids.
• Benchmarking computational methods is crucial for understanding model limitations and driving theoretical advancements.

Purpose of the Study:

• To benchmark MACE foundation models for screening dynamic stabilities of inorganic solids.
• To evaluate model performance using a DFT database of halide double perovskites.

Main Methods:

• Utilized a DFT database of room-temperature dynamic stability and vibrational anharmonicity for ~2000 cubic halide double perovskites.
• Benchmarked four MACE foundation model variants for predicting dynamic stabilities.
• Analyzed error sources, including atomic force errors and configurational space sampling.

Main Results:

• Model accuracy correlates positively with increased training data.
• MACE models more accurately predict dynamic stabilities of weakly anharmonic materials compared to highly anharmonic ones.
• Error amplification in atomic forces during Hessian matrix computation is the primary source of inaccuracy.

Conclusions:

• MACE foundation models demonstrate potential for materials screening but require further refinement for complex systems.
• Physics-inspired approaches are needed to enhance the accuracy and reliability of foundation models in atomistic modeling.
• Findings guide future development towards more robust and accurate universal potentials.