Transferability of Data Sets between Machine-Learned Interatomic Potential Algorithms

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Foundational Machine Learning Interatomic Potential (FMLIP) models require extensive data for training.
• Transferring data between different machine learning (ML) architectures can accelerate model fine-tuning.
• Optimizing training data for one ML method and reusing it for another can reduce costly iterative training.

Purpose of the Study:

• To examine the reusability of training data between different ML architectures for FMLIPs.
• To assess the impact of training data configurations on model performance across various ML algorithms.
• To provide principles for enhancing training sets for molecular liquid models with minimal computational effort.

Main Methods:

• Compared training data transferability between feedforward neural networks (Deep Potential model) and message-passing networks (MACE).
• Utilized a common battery electrolyte solvent as a test case.
• Proposed and applied a simple metric to assess model performance and generalization.

Main Results:

• MACE models demonstrated good performance even with simple training sets, unlike simpler architectures requiring iterative training.
• Human-intuitive data configurations transferred effectively between algorithms, whereas automatically generated configurations did not.
• System-specific training data proved necessary for realistic model performance compared to pretrained FMLIPs.
• Model stability was maintained for minor molecular shape changes but not for alterations in functional chemistry.

Conclusions:

• Training data properties significantly influence the behavior of MLIPs.
• Strategic data set enhancement can accelerate the simulation of new chemical systems using FMLIPs.
• Careful consideration of data transferability and system specificity is crucial for efficient and accurate MLIP development.