Dynamic Training Enhances Machine Learning Potentials for Long-Lasting Molecular Dynamics.

Area of Science:

• Computational physics and chemistry
• Materials science
• Machine learning

Background:

• Molecular dynamics (MD) simulations are crucial for studying complex systems.
• Machine learning (ML) potentials reduce computational cost but often lack accuracy in long simulations.

Purpose of the Study:

• To introduce dynamic training (DT), a novel method for improving ML model accuracy in extended MD simulations.
• To demonstrate the effectiveness of DT in enhancing prediction accuracy for challenging chemical systems.

Main Methods:

• Developed and applied dynamic training (DT) to an equivariant graph neural network (EGNN).
• Tested the DT-enhanced EGNN on a hydrogen molecule interacting with a palladium cluster on graphene.

Main Results:

• DT significantly improved prediction accuracy compared to conventional training methods.
• The DT approach demonstrated superior performance on a complex material system.

Conclusions:

• Dynamic training (DT) is an effective, architecture-independent method for enhancing ML potential accuracy in long MD simulations.
• DT offers a practical solution for advancing computational simulations in physics and chemistry.