Fine-tuning foundation models of materials interatomic potentials with frozen transfer learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Machine-learned interatomic potentials (MLIPs) enable accurate atomistic simulations but require extensive training data.
• Foundation models offer broad applicability but lack the precision for critical material properties.
• Generating high-quality training datasets for MLIPs is computationally expensive and challenging.

Purpose of the Study:

• To demonstrate that foundation model potentials can achieve chemical accuracy through transfer learning.
• To reduce the data requirements for training accurate MLIPs.
• To develop a more efficient simulation workflow for materials discovery.

Main Methods:

• Fine-tuning foundation model potentials using transfer learning with partially frozen weights and biases.
• Applying frozen transfer learning to datasets for reactive chemistry at surfaces and tertiary alloy properties.
• Developing a surrogate model using the transfer-learned potential as ground truth.

Main Results:

• Frozen transfer learning with 10-20% of the data achieved accuracy comparable to models trained from scratch on thousands of data points.
• The fine-tuned foundation models reached chemical accuracy for challenging datasets.
• An equally accurate, yet more computationally efficient, surrogate model was created.

Conclusions:

• Transfer learning with partially frozen weights is an effective method to achieve high accuracy in foundation model potentials.
• This approach significantly improves data efficiency and computational efficiency in MLIPs.
• The proposed workflow accelerates materials simulations and discovery.