Using physical property surrogate models to perform accelerated multi-fidelity optimization of force field parameters

Area of Science:

• Computational chemistry
• Molecular modeling
• Physical chemistry

Background:

• Accurate van der Waals interactions are crucial for molecular dynamics simulations.
• Training Lennard Jones (LJ) potential parameters is computationally expensive, limiting optimization.
• Current methods often restrict parameter optimization to local regions.

Purpose of the Study:

• To introduce a multi-fidelity optimization technique for global LJ parameter optimization.
• To accelerate the search for optimal LJ parameters using surrogate modeling.
• To improve the quality and transferability of force field parameters.

Main Methods:

• Utilized Gaussian process surrogate modeling to create inexpensive models of physical properties.
• Implemented an iterative framework combining differential evolution with simulation-based validation.
• Refitted LJ parameters for the OpenFF 1.0.0 (Parsley) force field using up to 195 physical property targets.

Main Results:

• The multi-fidelity technique found improved LJ parameter sets compared to simulation-based optimization.
• The method enabled broader searches and escape from local minima.
• New parameter sets demonstrated comparable accuracy and improved transferability to similar molecules.

Conclusions:

• The multi-fidelity optimization technique offers a rapid and more global approach for refining molecular models.
• This method enhances the optimization of force field parameters against physical properties.
• The technique provides a foundation for further advancements in molecular modeling.