Evaluating mechanical property prediction across material classes using molecular dynamics simulations with universal...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Accurate simulation of material properties requires precise interatomic interactions.
• Quantum-mechanical methods are computationally expensive for large-scale simulations.
• Universal machine-learned interatomic potentials (MLIPs) offer a computationally efficient alternative.

Purpose of the Study:

• To assess the accuracy of six universal MLIPs for predicting temperature and pressure responses in diverse materials.
• To evaluate MLIPs' performance in calculating bulk modulus, thermal expansion, and thermal decomposition.
• To identify key factors influencing MLIP accuracy for materials simulation.

Main Methods:

• Evaluated six universal MLIPs across three architectures (GNNs, GNS, Graph Transformers).
• Tested models on 13 diverse materials, including metal-organic frameworks and inorganic compounds.
• Calculated bulk modulus, thermal expansion, and thermal decomposition for each material and model.

Main Results:

• Observed qualitative agreement with experimental data, outperforming UFF4MOF.
• Systematic underestimation of bulk modulus and overestimation of thermal expansion were noted across models.
• 'MACE-MP-0a', 'fairchem_OMAT', and 'Orb-v3' emerged as top performers with average errors around 41-43%.
• Dataset homogeneity and structural representation significantly impact model accuracy.

Conclusions:

• Universal MLIPs provide a viable alternative to quantum-mechanical methods for materials simulation.
• Model performance is sensitive to training data quality and representation.
• Further development is needed to achieve truly universal and highly accurate MLIPs for diverse material dynamics.