Deep learning for variational multiscale molecular modeling.

Area of Science:

• Computational Chemistry
• Biophysics
• Machine Learning

Background:

• Atomistic molecular simulations are crucial for chemical and biophysical studies.
• Limited accessible timescales and challenges in rare event analysis hinder equilibrium property extraction.
• Current strategies like enhanced sampling or coarse-grained (CG) models have individual limitations.

Purpose of the Study:

• To develop a novel machine learning approach integrating fine-grained (FG) and CG simulations.
• To enable mutual benefits and crosstalk between simulations at different scales.
• To overcome limitations of individual simulation strategies for complex systems.

Main Methods:

• Utilizing deep generative learning to infer accurate CG models from FG simulations.
• Employing deep reinforcement learning to guide FG simulations using CG models.
• Implementing a variational and adaptive training objective for end-to-end training of parametric molecular models via deep neural networks.

Main Results:

• Demonstrated efficiency and flexibility of the proposed machine learning method.
• Successfully applied the approach to challenging chemical and biomolecular systems.
• Showcased the synergistic benefits of integrating FG and CG simulations.

Conclusions:

• The developed machine learning framework effectively bridges FG and CG simulations.
• This integrated approach enhances the study of rare events and equilibrium properties in molecular systems.
• The method offers a powerful and versatile tool for advancing computational chemistry and biophysics.