Machine Learning Directed Optimization of Classical Molecular Modeling Force Fields.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Accurate molecular force fields are crucial for predictive molecular simulations.
• Developing force fields that precisely replicate experimental properties presents a significant challenge.
• Existing methods often require extensive computational resources for parameterization.

Purpose of the Study:

• To develop an efficient, machine learning-directed workflow for optimizing molecular force field parameters.
• To significantly reduce the number of molecular simulations needed for accurate force field development.
• To demonstrate the workflow's applicability across different chemical systems.

Main Methods:

• A multiobjective optimization strategy guided by machine learning was employed.
• Millions of potential force field parameter sets were evaluated computationally.
• A small subset of promising parameter sets was validated using molecular simulations.
• The workflow was tested on hydrofluorocarbon vapor-liquid equilibrium and ammonium perchlorate crystal phase simulations.

Main Results:

• The machine learning workflow successfully identified multiple low-error force field parameter sets.
• The approach proved generalizable across distinct chemical systems and properties.
• Efficient exploration of the parameter space was achieved, minimizing simulation requirements.
• Accurate prediction of vapor-liquid equilibrium for HFCs and crystal phase behavior for AP was demonstrated.

Conclusions:

• The presented machine learning-directed workflow offers a powerful and efficient method for developing accurate molecular force fields.
• This approach significantly accelerates the force field parametrization process.
• The identified parameter sets provide reliable tools for future molecular simulations in materials science and chemistry.