Learning the effective dynamics of complex processes. Theory and application to molecular dynamics. – Talk by Prof. Dr. Christof Schütte
Monday, September 9, 2019 at 15.15 h, ETH Zürich, LEE E 101
Learning the effective dynamics of complex processes. Theory and application to molecular dynamics.
Prof. Dr. Christof Schütte (Freie Universität Berlin and Zuse Institute Berlin)
We present a novel machine learning algorithm for identifying the low-dimensional geometry of the effective dynamics of high-dimensional multiscale stochastic systems. We show that this geometry can be described by a low-dimensional transition manifold in an appropriate function space. Algorithms for embedding and learning the transition manifold are based on reproducing kernel Hilbert spaces and can be extended into a deep learning setting. It will also be discussed how to learn the effective dynamics on the manifold.