Enhancing Machine Learning Potentials through Transfer Learning across Chemical Elements.

Area of Science:

• Computational Materials Science
• Machine Learning in Physics and Chemistry

Background:

• Machine learning potentials (MLPs) offer ab initio simulation accuracy at reduced computational cost.
• Effective MLP generalization requires extensive datasets, which are often labor-intensive to generate.
• Data scarcity poses a significant challenge for developing robust MLPs.

Purpose of the Study:

• To introduce and evaluate transfer learning for potential energy surfaces between chemically similar elements.
• To demonstrate the efficacy of initializing a germanium MLP using a pre-trained silicon MLP.
• To address the challenges of data scarcity in MLP training.

Main Methods:

• Implemented transfer learning by initializing a germanium MLP with parameters from a trained silicon MLP.
• Utilized both classical force field and ab initio datasets for training and validation.
• Compared transfer learning performance against traditional training from scratch.

Main Results:

• Transfer learning significantly improved force prediction accuracy compared to training from scratch.
• MLPs trained via transfer learning exhibited enhanced simulation stability and temperature transferability.
• These benefits were amplified in data-scarce scenarios and extended to most out-of-target properties.

Conclusions:

• Transfer learning across chemically similar elements is a viable strategy for developing accurate MLPs.
• This approach is particularly effective in data-scarce regimes, reducing training time and data requirements.
• The findings highlight transfer learning as a promising technique for creating numerically stable and efficient MLPs.