Fast and Fourier features for transfer learning of interatomic potentials.

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Atomistic simulations

Background:

• Training machine learning interatomic potentials (MLIPs) is computationally intensive and data-hungry.
• Current methods limit the routine application of MLIPs in large-scale simulations.

Purpose of the Study:

• Introduce a scalable and lightweight transfer learning framework, named franken.
• Enable computationally and data-efficient training of MLIPs for diverse systems.

Main Methods:

• Extract atomic descriptors from pre-trained graph neural networks.
• Utilize random Fourier features for efficient kernel approximation.
• Implement a closed-form fine-tuning strategy for rapid adaptation of potentials.

Main Results:

• Franken outperforms kernel-based methods in training time and accuracy on transition metals.
• Reduced model training time from hours to minutes on a single GPU.
• Achieved stable and accurate potentials for water and interfaces with minimal data.

Conclusions:

• Franken offers a fast and practical solution for training and deploying MLIPs.
• Enables efficient atomistic simulations across various systems and simulation levels.