Generative spatiotemporal downscaling model improves projections of climate extremes

Climate change is amplifying extreme events, posing escalating risks to biodiversity, human health, and food security. Global climate models (GCMs) are essential for projecting future climate, yet their coarse resolution and high computational costs constrain their ability to represent extremes. Here, we introduce FuXi-CMIPAlign, a generative deep learning framework for downscaling Coupled Model Intercomparison Project (CMIP) outputs. The model integrates Flow Matching for generative modeling with domain adaptation via Maximum Mean Discrepancy loss to align feature distributions between training data (ERA5 reanalysis) and inference data (European Consortium-Earth), thereby mitigating input discrepancies and improving accuracy, stability, and generalization across emission scenarios. FuXi-CMIPAlign performs spatial, temporal, and multivariate downscaling, enabling more realistic simulation of compound extremes such as tropical cyclones (TCs). Applied to the historical period (2005–2014), it reduces global 99th-percentile mean absolute errors by 26%, 42%, and 33% for high temperature, extreme precipitation, and strong wind, respectively, and reproduces TC activity better aligned with ERA5. Under future scenarios (2015–2100), FuXi-CMIPAlign projects pronounced increases in land area affected by high temperature and frequency of extreme precipitation under high-emission scenarios, along with up to 60% rise in TC intensity and frequency over the Northwest and Northeast Pacific. In contrast, strong wind events over land shows a counterintuitive weakening trend. These results demonstrate that FuXi-CMIPAlign substantially improves CMIP6 projections of climate extremes, providing a robust generative framework for advancing climate risk assessment, mitigation and adaptation.