Evaluating a Neural Network‐Based Scale‐Adapted Scheme for Ice Cloud Fraction in the Taiwan Earth System Model (TaiESM1)

Abstract Ice cloud fraction (CF ice ) plays a critical role in the Earth's radiation budget and climate system. However, conventional cloud parameterizations in general circulation models are often inadequate to simulate CF ice and its vertical structure. This study evaluates the Neural Network‐Based Scale‐Adaptive cloud fraction scheme (NSA) in the Taiwan Earth System Model version 1 (TaiESM1), with a focus on its impact on ice cloud simulation and radiative processes. The NSA scheme, trained on CloudSat and ECMWF data, incorporates multiple environmental variables and cloud condensates. Two 10‐year AMIP simulations—one with the default scheme and the other with NSA—were conducted to assess differences in CF ice , cloud radiative forcing, and top‐of‐atmosphere (TOA) radiation. Versus the default scheme, the NSA scheme reduces the global annual mean CF ice , from 29.9% to 25.2%, which is closer to 22.4% based on MODIS‐COSP. The decrease in high‐cloud cover is significant, by 25% globally, and 50% in Antarctica. The NSA scheme also corrects the unrealistic glaciation of low‐level stratus clouds. As a result, the total cloud fraction (CF total ) is reduced from 56.3% to 48.6%, in much better agreement with the CloudSat‐COSP value of 49.8%. The associated changes in cloud radiative forcing and TOA radiation are also calculated to illustrate the effects of these cloud cover changes. Discussion on further improvement of cloud microphysics in TaiESM1 in conjunction with the NSA cloud cover scheme is presented.