Thermodynamic Transferability in Coarse-Grained Force Fields Using Graph Neural Networks.

Area of Science:

• Molecular Modeling
• Computational Chemistry
• Machine Learning in Science

Background:

• Coarse-graining simplifies complex atomistic systems for molecular simulations, enabling larger scales.
• Developing transferable force fields that maintain atomistic accuracy across conditions is a key challenge.
• Existing methods often lack transferability due to averaging at specific thermodynamic states.

Purpose of the Study:

• To develop a highly automated pipeline for training coarse-grained force fields.
• To enhance the transferability of coarse-grained models across various thermodynamic conditions.
• To leverage machine learning, specifically graph neural networks, for improved force field construction.

Main Methods:

• Utilized a Hierarchically Interacting Particle Neural Network with Tensor Sensitivity (HIP-NN-TS) architecture.
• Implemented a force-matching approach within a machine learning framework.
• Developed a highly automated training pipeline for coarse-grained force fields.

Main Results:

• Achieved highly accurate coarse-grained force fields.
• Demonstrated significantly improved transferability of these force fields across diverse thermodynamic conditions.
• Validated the effectiveness of the HIP-NN-TS architecture for this task.

Conclusions:

• Machine learning, particularly graph neural networks, offers a powerful approach to constructing transferable coarse-grained force fields.
• The developed automated training pipeline and HIP-NN-TS model overcome limitations of traditional methods.
• This work paves the way for more robust and versatile molecular simulations.