Chen, Yi‐ChangYi‐ChangChenWu, Chien‐MingChien‐MingWu2025-12-312025-12-312025-07https://www.scopus.com/pages/publications/105009787679https://scholars.lib.ntu.edu.tw/handle/123456789/734847This study explores the potential of deep learning as a subgrid parameterization for global storm-resolving models (GSRMs) by employing Large-Eddy Simulation (LES) to generate high-resolution cold pools under various convective structures. The high-resolution data is coarsened to 0.8, 1.6, 3.2, and 6.4 km to mimic the horizontal resolutions of GSRMs. U-Net deep learning models are developed to predict the high-resolution distribution of cold pools using coarsened near-surface (at height of 100 m) physical variables, including horizontal winds, potential temperature, and relative humidity. Results show that the U-Net models effectively capture cold pool characteristics, particularly their edges and intensity distribution at coarser scales. Additionally, high-resolution predictions provide enhanced information on horizontal heterogeneity that is not fully captured by low-resolution fields across different convective regimes. Sensitivity experiments indicate that U-Net prediction from input that includes wind fields outperforms those with thermodynamic variables only, highlighting the importance of accurately simulating dynamical variability in GSRMs. These findings can contribute to the advancement of improved subgrid machine-learning based parameterizations for next-generation atmospheric models.cold pooldeep learningLarge-Eddy simulationsU-Net[SDGs]SDG7[SDGs]SDG13journal article10.1002/asl.1309