Characterizing the essential behavior of high-dimensional metastable molecular dynamics along collective variables (CVs) plays an important role in understanding complex molecular systems. In recent years, great advances have been made in developing data-driven numerical methods that allow for automatic identification of CVs and building effective surrogate models of high-dimensional molecular systems. In this talk, I will discuss mathematical aspects of methods in this direction based on learning autoencoders and eigenfunctions of certain operator (e.g. generator and transfer operator) associated to the process. A new approach for jointly learning CVs and effective dynamics will also be presented.
Wei Zhang did his PhD at Peking University in China from 09.2006 to 01.2012. After that, he moved to Berlin and worked at both Free University Berlin and Zuse Institute Berlin. Currently, he works at Zuse Institute Berlin and holds an independent researcher position funded by DFG. His main research interest includes Monte Carlo methods, stochastic processes with metastability or multiple timescales, model reduction of stochastic processes, and applying deep learning techniques to processes from molecular dynamics.