Domain-aligned generative downscaling enhances projections of extreme climate events

Climate change is exacerbating extreme weather events globally, including high temperatures, extreme precipitation, strong winds, and tropical cyclones, posing severe threats to human health, infrastructure, food security, and socio-economic systems. Although existing global climate models (GCMs) provide essential tools for climate prediction, they face limitations such as insufficient resolution and high computational costs when simulating extreme events. To address these issues, this study proposes a spatiotemporal downscaling model based on generative machine learning-the Domain Aligned Climate Downscaling model (DACD), designed to enhance the simulation capabilities for extreme weather events. The proposed model employs domain adaptation tricks and a Flow Matching training framework to transform global low-resolution climate data into high-resolution local-scale climate information while achieving precise simulation of multivariable and temporal scales. The results show that during the historical period (2005-2014), our model outperformed existing methods in simulating high temperatures, extreme precipitation, strong wind, and tropical cyclone tracks, significantly reducing errors and improving the ability to capture extreme events. Under different future scenarios (2015-2100), the model reveals a significant increasing trend in the frequency and intensity of extreme events, particularly under the high-emission scenario (SSP585). Compared to traditional methods, our model more accurately simulates the spatial distribution and dynamic changes of extreme events, providing an essential tool for understanding the impacts of climate change. This study offers a new technological pathway for high-resolution climate analysis and extreme event prediction, providing scientific support for addressing future climate change and formulating adaptation strategies.