Speaker: Erik Thiede
Topic: Hyperparameter Selection in Chemical Koopman Models, or, How I Learned to Stop Worrying and Love the Lag Time
Abstact: There has been considerable interest in the development of model based on Koopman operator, most notably Markov State Models, to analyze chemical simulations. In principle, these techniques can predict chemical rates in the absence of complete reactive trajectories and are able to simplify complicated dynamics to automatically build collective variables. However, in practice these approaches can be highly sensitive to the choice of lag time. Here, I discuss two projects attempting to overcome these difficulties. In the first, we formalize the process of rate estimation as a Galerkin approximation, allowing us to extend MSM rate estimations to more general, and potentially more robust, schemes. In the latter, we analyze the effect of basis choice and lag time on dimensionality reduction for Koopman models, giving new insight into how hyperparameters for MSMs and TICA should be chosen