Discovering Reaction Pathways, Slow Variables, and Committor Probabilities with Machine Learning.

Area of Science:

• Computational Chemistry and Physics
• Machine Learning in Molecular Dynamics

Background:

• Atomistic simulations struggle to model transitions between metastable states in slow molecular processes.
• Accurate simulation requires defining low-dimensional reaction coordinates (RCs) using collective variables (CVs).
• Traditional RC discovery relies on intuition, but machine learning (ML) offers data-driven alternatives.

Purpose of the Study:

• To compare two variational data-driven ML methods for discovering meaningful RC models.
• To identify the slowest decorrelating CV and the committor probability for molecular processes.
• To assess the effectiveness of these methods in capturing slow molecular dynamics.

Main Methods:

• Comparison of two Siamese neural network-based ML methods: VAMPnets (SRVs) and VCNs.
• VAMPnets discover the slowest decorrelating CV.
• VCNs, inspired by transition path theory, identify the committor probability.

Main Results:

• Both VAMPnets and VCNs successfully discover relevant descriptors for slow molecular processes.
• The methodologies demonstrate their ability to capture the dynamics of slowest degrees of freedom.
• Both methods are compatible with importance-sampling schemes via reweighting for kinetic property approximation.

Conclusions:

• Variational data-driven ML methods provide powerful tools for discovering reaction coordinates in molecular simulations.
• VAMPnets and VCNs offer complementary approaches to understanding and accelerating the simulation of slow molecular dynamics.
• These ML strategies enhance the applicability of importance-sampling techniques for studying complex molecular transitions.