On-the-fly fine-tuning of foundational neural network potentials: a Bayesian neural network approach

Key Points

• The aim is to create a method for efficiently fine-tuning neural network potentials while ensuring trustworthy predictions.
• Developed a workflow for fine-tuning neural networks on-the-fly.
• Utilized Bayesian transfer learning for enhanced predictive uncertainty estimation.
• Implemented data-efficient strategies to improve model reliability.
• Achieved improved trustworthiness in predictions through uncertainty estimation.
• Demonstrated effective fine-tuning with minimal data requirements.

Abstract