Performance Assessment of Universal Machine Learning Interatomic Potentials: Challenges and Directions for Materials'...

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine learning interatomic potentials (MLIPs) offer a balance between high accuracy and computational efficiency for materials simulations.
• Advanced universal MLIPs (UIPs) utilize equivariant representations and deep graph neural networks, aiming for broad applicability across the periodic table.
• Current UIPs are primarily trained on density functional theory (DFT) calculations of bulk materials.

Purpose of the Study:

• To evaluate the generalization capabilities of existing universal MLIPs (MACE, CHGNet, M3GNet) for calculating surface energies.
• To identify shortcomings of out-of-the-box foundation models in tasks beyond their primary training data.

Main Methods:

• Assessed the performance of openly available universal MLIPs (MACE, CHGNet, M3GNet).
• Focused on the representative generalization task of calculating surface energies.
• Analyzed errors in relation to total energy and out-of-domain distance from training data.

Main Results:

• Out-of-the-box universal MLIPs exhibit significant errors when calculating surface energies.
• Model errors correlate with the total energy of surface simulations.
• Performance issues stem from the models being out-of-domain compared to their bulk-material-centric training datasets.

Conclusions:

• Universal MLIPs serve as effective starting points for developing specialized models through fine-tuning.
• Expanding training datasets to cover a wider range of material configurations is crucial for true universality.
• Further development is needed to enhance the predictive power of MLIPs for diverse material systems and properties.