Adversarial training for dynamics matching in coarse-grained models.

Area of Science:

• Computational chemistry
• Molecular dynamics simulations
• Biophysics

Background:

• Molecular dynamics (MD) simulations are crucial for understanding molecular systems but are computationally expensive.
• Coarse-grained (CG) models simplify systems to reduce computational cost, but often lack dynamic consistency, limiting their use in kinetics studies.
• Existing CG methods struggle to accurately reproduce the dynamics of complex molecular systems.

Purpose of the Study:

• To develop a novel adversarial training framework for bottom-up coarse-grained modeling.
• To ensure both thermodynamic and kinetic fidelity in coarse-grained simulations by aligning with all-atom (AA) reference dynamics.
• To provide a systematic and principled approach for preserving dynamic consistency in complex molecular systems.

Main Methods:

• An adversarial learning framework combining a physics-based generator and a neural network discriminator was developed.
• The framework trains the CG model to generate trajectories indistinguishable from AA reference trajectories.
• CG parameters are optimized adversarially, removing the need for predefined kinetic features.

Main Results:

• The method successfully reproduced key properties of liquid water, including radial and angular distribution functions.
• Accurate prediction of dynamical mean squared displacement was achieved.
• The framework demonstrated the ability to extrapolate long-timescale dynamics from short training trajectories.

Conclusions:

• The proposed adversarial training framework offers a robust method for developing accurate coarse-grained models.
• This approach enhances the reliability of CG simulations for kinetics-driven processes.
• It represents a significant advancement in bottom-up CG modeling, ensuring dynamic consistency.