Computational Mathematics and Scientific Computing Seminar

Time and Location:

Speaker:

Link:

We discuss two problems related to the analysis and probabilistic quantification of extreme events in complex dynamical systems.

Topic A [\eta leaning]: Assuming we have a labeled dataset that contain few or even no extreme events, how can machine learning models encode available statistical information, obtained e.g. through plentiful unlabeled data for an observable of the system, or other physics-based arguments that provide statistics for an observable.

Topic B [generate extreme events using nudged trajectories]: Assuming we have a finite time series of a complex dynamical system, how can we use it effectively to, not just “replicate” extreme event statistics, but to “extrapolate” extreme event statistics. To achieve this, it is essential to incorporate dynamics encoded in the sequence of events of the time series.

References:

K. Chang, T. Sapsis, Extreme Event Aware (η-) Learning (2026), Nature Communications, In Press

M. Wang, B. Barthel Sorensen, T. Sapsis, Generative Prediction-Correction Framework for Long-time Emulations of Spatially-Resolved Climate Extremes (2026), arXiv

Stamatelopoulos et al., DySCo: Dynamically consistent data-driven downscaling of extremes in climate projections (2026), arXiv