Accurate temperature projections are critical for climate adaptation and policymaking, yet substantial uncertainties remain in model results, particularly at regional scales. In this study, we develop a deep learning model to predict the timing of temperature thresholds in China based on historical datasets and use observational data to constrain future projections. We demonstrate that historical annual mean temperature can highly predict future temperature changes for China. Using this model, we estimate that the country will reach a 2°C temperature increase before the 2030s. Our explainable model reveals that the Southern Ocean, especially the Southeastern Pacific, is a key driver for these projections. This region exhibits a slow response of sea surface temperature to greenhouse gases, reflecting the pace and signal of global warming. Moreover, green's function perturbation experiments with numerical climate model further indicate this region as an optimal forcing area for East Asian temperature variations via dynamic pathways. Our findings underscore that deep learning models can not only extract the emerging global warming signal from annual mean temperature data but also account for dynamic interactions between regions, allowing historical data to effectively constrain future projections.

人工智能 气候变化