Improving machine learning force fields for molecular dynamics simulations with fine-grained force metrics.

Area of Science:

• Computational Chemistry
• Materials Science
• Biophysics

Background:

• Machine learning force fields (MLFFs) are increasingly used as computationally efficient alternatives to ab initio methods for molecular dynamics (MD) simulations.
• Despite their advantages, MLFFs often exhibit limitations in generalization and robustness, leading to instability during long MD simulations.

Purpose of the Study:

• To develop and validate novel metrics for systematically evaluating the accuracy and reliability of MLFFs across diverse molecular systems and conformations.
• To investigate the correlation between proposed force metrics and the stability of MD simulations generated by MLFFs.
• To provide a framework for improving MLFF performance and MD simulation reliability.

Main Methods:

• Introduced global and fine-grained force metrics, considering elemental and conformational aspects, to assess MLFFs at atomic and conformational levels.
• Evaluated three state-of-the-art MLFFs (ET, NequIP, ViSNet) on aspirin, Ac-Ala3-NHMe, and Chignolin MD datasets (21–166 atoms).
• Performed MD simulations using trained MLFFs from various initial conformations, analyzing force metrics, trajectory stability, and simulation failures.

Main Results:

• Established a relationship between the proposed force metrics and the stability of MD simulation trajectories.
• Identified key factors contributing to simulation collapse in MLFF-driven MD.
• Demonstrated that MLFF performance and MD simulation stability can be enhanced by incorporating these force metrics into the training process.

Conclusions:

• The developed force metrics provide a robust tool for systematically evaluating and improving MLFFs.
• Integrating these metrics into MLFF training, through methods like loss functions, reweighting, or continued training, significantly enhances simulation accuracy and stability.
• This work offers a pathway to more reliable and accurate MLFF-guided molecular simulations.