<scp>DecGAN</scp> : A bias correction and downscaling deep‐learning model for numerical predictions with applications to sea surface temperature in the <scp>Pacific</scp> and precipitation in <scp>Asia</scp>

Abstract Climate modeling efforts often grapple with two overarching categories of forecast errors: systematic errors, arising from a model's inherent inability to fully capture complex physical processes, and accumulated errors, which emerge as inaccuracies compound over iterative predictions. In this paper, we correct both error types by introducing DecGAN, a novel multistep generative adversarial network designed for end‐to‐end bias correction and high‐resolution downscaling of numerical models. Leveraging the Community Earth System Model (CESM), we employ observational and reanalysis data from Hadley Centre Sea Ice and Sea Surface Temperature (HadiSST) ( https://www.metoffice.gov.uk ) and European Centre for Medium‐Range Weather Forecasts Reanalysis v5 (ERA5) ( https://cds.climate.copernicus.eu ) as reference labels to train our model and apply it to the Pacific sea surface temperature (SST) and Asian precipitation for evaluation. Comprehensive results demonstrate that DecGAN substantially reduces the average prediction errors over 12 consecutive months by 77.5% for Pacific SST and 18.8% for Asian precipitation, while significantly enhancing the structural fidelity of the forecasts as evidenced by Structural Similarity Index (SSIM) increases of 17.63% and 2.41%, respectively. Beyond correcting biases, DecGAN also downscales predictions from a coarse spatial resolution of 1° × 1° to 0.25° × 0.25° for SST, and 2.5° × 2.5° to 0.25° × 0.25° for precipitation, yielding high‐fidelity maps crucial for regional impact assessments. Our error analysis reveals that accumulated errors are strongly influenced by the model's ability to simulate internal variability, whereas systematic biases primarily stem from external forcings that the base model struggles to represent accurately. This distinction underscores the need to advance both the physics‐based core of numerical modeling and the data‐driven corrective strategies to improve SST forecasts. Furthermore, given El Niño–Southern Oscillation (ENSO)'s far‐reaching impacts on Asian precipitation via atmospheric circulation anomalies, incorporating rigorous bias correction and downscaling is pivotal for more reliable long‐term precipitation projections. By bridging numerical model outputs with deep‐learning innovations, DecGAN paves the way for a new generation of climate prediction systems that are both more accurate and spatially refined.