Coarse-Grained Modeling Using Neural Networks Trained on Structural Data.

Area of Science:

• Computational Chemistry
• Materials Science
• Statistical Mechanics

Background:

• Coarse-graining simplifies complex molecular systems for large-scale simulations.
• Developing accurate coarse-grained potentials from atomistic data remains a challenge.
• Existing methods often lack transferability across different conditions.

Purpose of the Study:

• To introduce a novel bottom-up coarse-graining method.
• To train artificial neural networks (ANNs) on atomistic simulation data for coarse-grained interactions.
• To demonstrate the transferability of ANN-derived potentials.

Main Methods:

• Utilized a bottom-up coarse-graining approach.
• Employed ANNs trained on radial distribution functions from atomistic simulations.
• Applied the inverse Monte Carlo method to link ANN weights to structural properties.
• Tested on Lennard-Jones systems and methanol-water solutions.

Main Results:

• Successfully modeled a Lennard-Jones system with a linear ANN.
• Developed a nonlinear ANN for methanol-water solutions using atomistic data.
• Demonstrated that the ANN potential is transferable across various methanol concentrations.
• Showcased transferability even for concentrations not present in the training data.

Conclusions:

• The proposed ANN-based coarse-graining method effectively generates transferable potentials.
• This approach enhances the accuracy and applicability of coarse-grained models.
• ANNs provide a powerful tool for deriving accurate inter-site potentials in coarse-grained simulations.