Deep Learning Reveals Thermodynamic Control of Global Warming on Precipitation

Global warming exerts profound impacts on precipitation through thermodynamic and dynamic processes. However, disentangling the respective contributions of these two processes remains challenging due to the lack of effective approaches to characterize the nonlinear interactions within the climate system. Here, we present a deep learning (DL) framework based on ERA5 to distinguish the contributions of different processes to long-term linear trend of annual precipitation during 1961–2010 and the precipitation variability across different global warming levels. The results reveal that thermodynamic processes account for 41% of global annual precipitation trend variation over land, while dynamic processes contribute 59%, integrating both spatial patterns and temporal changes during 1961-2010. Thermodynamic processes drive increases in annual mean precipitation globally due to the rising atmospheric vapor, except part of Africa, whereas dynamic processes govern the spatial heterogeneity of precipitation trends. Notably, under extreme warming scenarios, thermodynamic terms amplify precipitation variability, increasing the frequency of wet months by 15.4% and decreasing drought frequency by 5.6% globally. This study highlights that the nonlinear processes in climate system can be decomposed through DL climate attribution.