Synthetic Force-Field Database for Training Machine Learning Models to Predict Mobility-Preserving Coarse-Grained...

Area of Science:

• Computational chemistry
• Materials science
• Machine learning applications

Background:

• Molecular simulation faces a trade-off between accuracy (all-atom) and efficiency (coarse-grained).
• Developing accurate coarse-grained models that capture both structure and dynamics remains a challenge.
• Data-driven approaches, particularly machine learning, offer potential but require robust datasets.

Purpose of the Study:

• To develop a data-driven approach for constructing coarse-grained models that accurately represent both structure and dynamics.
• To address the critical need for high-quality datasets in machine learning for molecular simulations.
• To improve the efficiency and accuracy of coarse-grained molecular dynamics simulations.

Main Methods:

• Constructed a synthetic database of coarse-grained potentials from unphysical all-atom models.
• Trained a neural network using the generated database to predict coarse-grained potentials for real liquids.
• Evaluated model quality by measuring combined structural and dynamical accuracy loss during coarse-graining.

Main Results:

• The machine learning-based coarse-grained potentials outperformed iterative Boltzmann inversion for all eight studied hydrocarbon liquids.
• The neural network demonstrated superior performance even for nonspherical all-atom surfaces where both methods degraded.
• The developed synthetic database and machine learning models are publicly available.

Conclusions:

• The proposed data-driven, machine learning approach enables the efficient derivation of accurate coarse-grained models for liquids.
• This method effectively balances accuracy and efficiency in molecular simulations.
• The approach shows promise for advancing the field of coarse-grained molecular dynamics.