Kernel-Based Machine Learning for Efficient Simulations of Molecular Liquids.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Current machine learning (ML) models for force fields are computationally intensive.
• High computational cost hinders their application in molecular dynamics simulations.

Purpose of the Study:

• To develop practical and computationally efficient strategies for parametrizing traditional force fields using ML.
• To reduce the computational cost of ML-based force fields for molecular liquids.

Main Methods:

• Particle decomposition ansatz for two- and three-body force fields.
• Kernel-based ML models incorporating physical symmetries.
• Use of switching functions and covariant meshing to enhance training data and efficiency.
• Implementation of many-body representations, decomposition, and kernel regression in VOTCA software.

Main Results:

• Demonstrated efficient parametrization of force fields for model molecular liquids (Lennard-Jones, Stillinger-Weber, coarse-grained water).
• Covariant meshing proved effective for learning instantaneously averaged forces.
• ML potentials enabled computationally efficient molecular dynamics simulations.
• Simulations accurately reproduced two- and three-body distribution functions.

Conclusions:

• The developed strategies significantly improve the computational efficiency of ML force fields.
• These methods enable accurate molecular dynamics simulations of liquids.
• The open-source VOTCA package facilitates the application of these advanced ML techniques.