B1043
Title: Using spatial extreme-value theory with machine learning to understand compound extremes: A case study on heat
Authors: Jonathan Koh - University of Bern (Switzerland) [presenting]
Abstract: Many extreme weather events in the past decade affected large areas, and the regional to sub-continental spatial scale was important for their impacts. A novel methodology is proposed that combines spatial extreme-value theory with a machine learning (ML) algorithm to model temperature extremes dependent on the large-scale atmospheric flow. The method allows quantifying the probabilities associated with the occurrence, the intensity, and the spatial dependence of summertime heat extremes across Europe, hence offering to assess the likelihood of spatial extreme events beyond observational records. Using new loss functions, a theoretically-motivated spatial process is fitted to extreme positive temperature anomaly fields from 1959-2022, with daily 500-hpa geopotential height fields and local soil moisture across the Euro-Atlantic region as predictors. The generative model reveals the importance of individual circulation features in determining different aspects of heat extremes, thereby enriching our understanding of heatwaves from a data-driven perspective. Subsidence and low soil moisture contribute to the magnitude of the heat extreme, while upstream cyclones and Rossby wave breaking contribute to spatially larger events. The approach offers an attractive alternative to physical model-based techniques, or ML approaches that instead optimises scores focusing on predicting the bulk instead of the tail of the data distribution.