Machine Learning Complex Dynamics

9:00–10:30 am
Zoom Webinar

Organizer: HKUST


Understanding molecular mechanisms requires estimating dynamical statistics such as expected hitting times, reaction rates, and committors. In systems with well-defined metastable states and free energy barriers, these quantities can be estimated using enhanced sampling methods combined with classical rate theories. However, calculating such statistics for more complex processes with rugged landscapes or multiple pathways requires more general numerical methods. In this lecture, the speaker will describe a machine learning framework for calculating dynamical statistics by approximating the dynamical operators of the system through a Galerkin expansion, using statistical estimates from short molecular dynamics trajectories.  It will be demonstrated that this approach gives remarkably accurate results for a well-characterized protein folding reaction with relatively little computational cost.  Finally, the approach will be applied to understanding the dynamics of the protein hormone insulin with a view toward designing improved therapeutics for diabetes.


Professor Aaron R. Dinner
Professor of Chemistry
University of Chicago