Applied Mathematics Seminar——Finding collective variables of metastable stochastic dynamics for the study of kinetics on large timescales
报告人：Dr. Wei Zhang (Zuse Institute Berlin)
时间：2023-04-27 14:00-15:00
地点：Room 1560, Sciences Building No. 1
Abstract: Collective variables (CVs) are often used to gain insights into the complex kinetics of high-dimensional dynamical systems. In recent years, thanks to the development of advanced sampling methods for trajectory data generation and the rapid emergence of novel machine learning techniques, various data-based learning approaches are becoming available for CV identification. In this talk, we focus on the use of the leading eigenfunctions of the system's infinitesimal generator as CVs. First, we present a motivation from model reduction, which allows us to argue that the leading eigenfunctions are the optimal CVs of the underlying metastable stochastic dynamics for the study of kinetics on large timescales. Next, we propose a deep learning-based algorithm for computing the leading eigenfunctions by training artificial neural networks. Finally, we present some numerical examples to demonstrate our findings.
