Learning Compound Climate Extremes: Generative AI for Hot–Dry Event Risk

Learning Compound Climate Extremes: Generative AI for Hot–Dry Event Risk

 

In recent years, there has been a growing interest in the applications of machine learning methods to multi-hazard events, mainly due to their ability to ingest large amounts of data and capturing the relationships between variables. Compound weather and climate events (CEs) are of significant societal importance, as they present greater risks, and better understanding their response to climate change is crucial. This response has mostly been explored through dynamical climate model ensemble methods. However, accurately estimating the uncertainty of climate scenarios often requires very large ensemble simulations to be conducted, which can be computationally costly.

Generative deep learning methods offer a cost-effective alternative by enabling the generation of large sets of synthetic events which follow the joint distribution of high-dimensional data.

This work builds on the HazGAN framework, an ML framework which generates synthetic event sets for risk analysis of specific CEs in defined regions, capturing the dependence structure among variables. We utilise a HadGEM3 Large Ensemble to further condition the model on the large-scale climate background state, allowing the characteristics of the synthetic events to vary with different climate regimes. This approach aims to account for non-stationarity in compound event behaviour and to provide a physically consistent framework for exploring changes in hot–dry extremes under a changing climate. We also address uncertainties associated with using machine learning for extrapolation by rigorously testing out-of-distribution predictions. This work enhances the understanding of compound events, their risks, and future impacts under climate change scenarios