Spectral Map: Embedding Slow Kinetics in Collective Variables.

Area of Science:

• Physical Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Complex physical systems often require high-dimensional data representation.
• Identifying meaningful collective variables (CVs) is crucial but challenging.
• Characterizing slow kinetics and rare transitions between metastable states is particularly difficult.

Purpose of the Study:

• To propose an unsupervised deep-learning method for constructing slow collective variables (CVs).
• To address the challenge of identifying CVs that capture slow dynamics in physical systems.
• To improve the understanding of rare transitions in high-dimensional systems.

Main Methods:

• Developed an unsupervised deep-learning method named spectral map.
• Constructed slow CVs by maximizing the spectral gap of a transition matrix.
• Estimated the transition matrix using an anisotropic diffusion kernel.

Main Results:

• Successfully identified slow collective variables (CVs) in high-dimensional systems.
• Demonstrated the method's effectiveness in capturing slow kinetics.
• Applied the spectral map method to reversible folding processes.

Conclusions:

• The spectral map method provides an effective approach for identifying slow CVs.
• This deep-learning technique enhances the analysis of complex physical system dynamics.
• The method shows promise for studying rare events and transitions in chemistry and physics.