Fine-Tuning Unifies Foundational Machine-Learned Interatomic Potential Architectures at ab initio Accuracy

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Machine-learned interatomic potentials (MLIPs) offer efficient atomistic simulations.
• General-purpose MLIPs show architecture-dependent deviations from high-accuracy ab initio methods.
• There is a need for universally accurate and computationally efficient MLIPs.

Purpose of the Study:

• To demonstrate that fine-tuning transforms foundational MLIPs to achieve near-ab initio accuracy.
• To benchmark the effectiveness of fine-tuning across diverse MLIP architectures and chemical compounds.
• To introduce a toolkit for reproducible fine-tuning workflows.

Main Methods:

• Benchmarking five leading MLIP frameworks (MACE, GRACE, SevenNet, MatterSim, ORB) on seven diverse compounds.
• Utilizing datasets from ab initio molecular dynamics trajectories for fine-tuning.
• Evaluating force and energy predictions against ab initio reference data.

Main Results:

• Fine-tuning universally enhances force predictions by 5-15 times and energy accuracy by 2-4 orders of magnitude.
• Specialized system-specific fine-tuning eliminates architecture-dependent deviations.
• Fine-tuning reduces force errors by an order of magnitude and harmonizes performance across architectures.

Conclusions:

• Fine-tuning is a universal method for achieving system-specific accuracy in MLIPs.
• This approach preserves the computational efficiency of MLIPs.
• The aMACEing Toolkit facilitates widespread adoption of fine-tuning workflows.