Topological spatiotemporal prediction of extreme events

We propose a topology-aware spatiotemporal model for predicting the occurrence of extreme events both in time (“when”) and in space (“where”) in nonlinear physical systems. Specifically, our model adopts a unified topological perspective to bridge temporal dynamics and spatial localization. By representing spatial grids as graph nodes, the model captures spatiotemporal dependencies from two complementary views: functional connections describe global temporal evolution, while structural neighborhood connections characterize local spatial interactions. This design supports joint prediction in time and space by effectively capturing temporal precursors and spatial patterns. The model is validated on a synthetic dataset from the two dimensional complex Ginzburg–Landau equation and on ERA5 wind speeds dataset over the North Atlantic. Experiments supported by graph neural network show that our model achieves great performance in both temporal prediction and spatial localization, highlighting its potential for addressing complex extreme event prediction challenges.