Machine Learning Force Field Parameters from Ab Initio Data.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Developing accurate polarizable force fields is crucial for molecular simulations.
• Traditional methods often rely on empirical fitting to experimental data, which can be time-consuming and may not capture underlying physics.
• Quantum mechanics (QM) calculations offer high-fidelity data but are computationally expensive for large systems.

Purpose of the Study:

• To develop a novel machine learning (ML) and genetic algorithm (GA) strategy for optimizing polarizable force field parameters.
• To utilize only ab initio quantum mechanics (QM) data for force field parameterization.
• To validate the performance of the ML/GA approach by predicting experimental condensed phase properties.

Main Methods:

• Applied ML techniques combined with GA to determine polarizable force field parameters.
• Used ab initio QM data from various levels of theory (MP2, DFMP2) for molecular clusters.
• Trained and validated the model using dimer electrostatic potentials and cluster interaction energies for methanol.
• Introduced an offset factor to improve energy surface representation.

Main Results:

• Achieved excellent agreement between QM training data and the optimized force field model.
• The ML/GA optimized force field demonstrated superior performance in predicting density and heat of vaporization compared to the empirical AMOEBA force field.
• The model successfully predicted experimental condensed phase properties without direct use of experimental data in optimization.

Conclusions:

• Demonstrated the feasibility of developing accurate polarizable force fields using solely QM data through ML/GA.
• The developed ML/GA strategy offers a promising alternative to traditional empirical force field optimization.
• The methodology is extensible to other molecular systems, advancing computational chemistry simulations.