Speed up Multi-Scale Force-Field Parameter Optimization by Substituting Molecular Dynamics Calculations with a...

Area of Science:

• Computational chemistry and materials science
• Application of machine learning in scientific modeling

Background:

• Molecular modeling relies on accurate force fields (FFs) for predicting system properties.
• Force-field parameter (FFParam) optimization is crucial for enhancing FF accuracy and applicability.
• Traditional FF optimization involves time-consuming molecular dynamics (MD) simulations.

Purpose of the Study:

• To accelerate the multiscale force-field parameter optimization process.
• To substitute computationally expensive MD simulations with a machine learning (ML) surrogate model.
• To optimize Lennard-Jones parameters for carbon and hydrogen in molecular modeling.

Main Methods:

• Development and implementation of a machine learning surrogate model.
• Substitution of MD simulations with the ML surrogate in a multiscale FFParam optimization workflow.
• Optimization focused on Lennard-Jones parameters for n-octane, targeting conformational energies and bulk density.

Main Results:

• Achieved a speed-up factor of approximately 20 by replacing MD simulations with the ML surrogate.
• Maintained the quality of the optimized force fields comparable to traditional methods.
• Presented a comprehensive workflow for acquiring and preparing data for ML surrogate model training.

Conclusions:

• Machine learning surrogate models offer a significant acceleration for force-field parameter optimization.
• This data-driven approach maintains the accuracy of force fields while drastically reducing computational cost.
• The presented methodology enables more efficient development and application of molecular modeling tools.