Reaction coordinate flows for model reduction of molecular kinetics

Area of Science:

• Computational chemistry
• Machine learning
• Chemical kinetics

Background:

• Developing accurate low-dimensional kinetic models for complex molecular systems is crucial for understanding their behavior.
• Existing model reduction methods often lack trainability, tractability, or continuous-time representations.

Purpose of the Study:

• To introduce Reaction Coordinate (RC) flow, a novel flow-based machine learning approach for discovering low-dimensional kinetic models.
• To enable data-driven estimation of model parameters for molecular kinetics.

Main Methods:

• Utilizing normalizing flows for coordinate transformation design.
• Employing a Brownian dynamics model to approximate the kinetics of the reaction coordinate.
• Leveraging the invertibility of normalizing flows for a trainable and tractable reduced kinetic model.

Main Results:

• RC flow provides an interpretable and accurate low-dimensional representation of full-state kinetics.
• The Brownian dynamics-based model effectively identifies metastable states in molecular systems.
• Numerical experiments validate the method's ability to discover reduced kinetic models from simulation data.

Conclusions:

• RC flow offers a powerful and versatile tool for molecular kinetics modeling.
• The approach facilitates the discovery of continuous-time, low-dimensional kinetic models with discernible metastable states.
• This data-driven method enhances the understanding and prediction of molecular system dynamics.