Regularized by Physics: Graph Neural Network Parametrized Potentials for the Description of Intermolecular...

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Machine Learning

Background:

• Electronic structure methods are computationally expensive for large biological systems.
• Empirical force fields (FF) are widely used but their parameterization is time-consuming and often relies on experimental data.
• Automating FF parameterization and developing ML-based potentials are active research areas.

Purpose of the Study:

• To propose an alternative, automated strategy for force field parameterization using machine learning.
• To retain a physics-based functional form for interpretability, robustness, and efficient simulations.
• To demonstrate the ability to learn FF parameters from scratch using only elemental information, topology, and reference potential energies.

Main Methods:

• Utilizing machine learning (ML) and gradient-descent optimization for FF parameterization.
• Employing a predefined, physics-based functional form for the force field.
• Training fixed-charge and polarizable FF models on *ab initio* potential-energy surfaces.
• Validating learned parameters against experimental and computational data.

Main Results:

• The proposed ML method successfully learns atom types and FF parameters from basic molecular information and reference energies.
• Trained models demonstrate robustness and enable efficient, long-timescale simulations.
• Validated parameters accurately predict properties of dimers, pure liquids, and molecular crystals.

Conclusions:

• This ML-driven approach offers an efficient and automated alternative for force field parameterization.
• The combination of ML with physics-based functional forms yields interpretable and reliable molecular simulations.
• The method shows significant promise for accelerating molecular modeling in various chemical and biological applications.