CSE Community Seminar | October 23, 2026

Abstract

We consider the use of kernel methods for the problems of pattern extraction and forecasting based on spatiotemporal time series data. First, we overview kernel-based pattern extraction methods using the eigenvalue decomposition of kernel integral operators. Second, we consider the use of kernel regression methods for the forecasting of spatiotemporal dynamics. Both problems are formulated in an appropriate Hilbert space of spatiotemporal observables, leading to favorable properties in the presence of spatial symmetries of the underlying dynamics. A common underlying thread of the work is our focus on the scalable numerical implementation of the studied methods, employing low rank approximation techniques and distributed GPU computing. We conclude with numerical results based on the modified Hasegawa-Wakatani equations of plasma dynamics.