An Improved Neural Network‐Based Scale‐Adaptive Cloud Fraction Scheme: Incorporation of Atmospheric Stability

Abstract Cloud fraction, crucial to cloud climate effects, remains challenging to predict in general circulation models (GCMs). A neural network‐based scale‐adaptive cloud fraction scheme, predicting cloud area fraction that affects cloud radiative effects, was developed previously using CloudSat observations and demonstrated good performance in capturing cloud spatial distribution and vertical structure. This study further improves the scheme by considering in the inputs the atmospheric stability (affecting cloud formation and evolution) and introducing cloud volume fraction (important for cloud microphysics) as an additional output. Results show that including atmospheric stability significantly enhances both prediction accuracy and scale adaptability for cloud area fraction compared to the original scheme. On the test data set, root‐mean square errors (RMSEs) for liquid, mixed‐phase, and ice clouds are reduced by 15.5%, 11.6%, and 6.67%, respectively. These leads to better predictions of cloud area‐fraction vertical structure and spatial distribution, with RMSE reduced by 33% and 17% in those aspects. Notably, improvements are most pronounced for low‐level clouds, particularly in simulating the seasonal variation of stratocumulus over the Southeast Pacific and the cloud regime transitions along the Pacific Cross‐Section Intercomparison transect of the Global Energy and Water Cycle Experiment Cloud System Study/Working Group on Numerical Experimentation. The new scheme also performs well in predicting cloud volume fraction, showing accuracy and scale adaptability comparable to that for area fraction. Offline evaluations using simulation data from the FGOALS‐f3‐L GCM confirm the superiority of the updated scheme, highlighting its potential to improve cloud fraction representation and associated climate effects in GCMs.