Adversarial-residual-coarse-graining: Applying machine learning theory to systematic molecular coarse-graining

Area of Science:

• Computational chemistry
• Machine learning
• Molecular modeling

Background:

• Coarse-graining (CG) simplifies complex molecular systems for simulations.
• Generative models offer powerful tools for data representation and generation.

Purpose of the Study:

• To develop a systematic framework for molecular coarse-graining (CG) by integrating CG approaches with implicit generative models.
• To explore the capabilities of generative adversarial networks (GANs) within this new CG framework.
• To enable rigorous parameterization of CG models, including those with virtual sites.

Main Methods:

• Utilizing connections between molecular coarse-graining (CG) and implicit generative models.
• Focusing on the formalism of generative adversarial networks (GANs).
• Demonstrating parameterization strategies, including those with virtual sites.

Main Results:

• The framework allows for various parameterization strategies, some similar to existing CG methods.
• It rigorously parameterizes CG models with virtual sites, which lack direct reference atomistic connections.
• A limitation is the absence of a closed-form expression for the CG Hamiltonian after integrating over virtual sites.

Conclusions:

• The developed framework provides a rigorous method for molecular coarse-graining (CG) using machine learning generative models.
• It successfully parameterizes CG models with virtual sites, expanding CG capabilities.
• The approach is applicable in cases where traditional methods like relative entropy minimization are not suitable.