Spectral Maps for Learning Reduced Representations of Molecular Systems

Investigating processes in complex molecular systems, which are characterized by many variables, is a crucial problem in computational physics. These systems can be reduced to a few meaningful degrees of freedom known as collective variables (CVs). However, identifying these CVs is a significant challenge, especially for systems with long-lived metastable states. This is because the information about the slow kinetics of rare transitions needs to be encoded in CVs. In this talk, we review recent advances in learning slow CVs and focus mainly on our spectral map technique, a promising deep-learning method that learns CVs based on the slowest timescales. By maximizing the spectral gap between slow and fast eigenvalues of a Markov transition matrix constructed from simulation data, our method effectively captures a simplified representation of alanine dipeptide in solvent. This practical application of our method demonstrates its ability to extract slow CVs, making it a valuable tool for analyzing complex systems.
Comment: Presented at 34th IUPAP Conference on Computational Physics (CCP2023)