A Neural Network‐Based Scale‐Adaptive Cloud‐Fraction Scheme for GCMs

Cloud fraction (CF) significantly affects the short- and long-wave radiation. Its realistic representation in general circulation models (GCMs) still poses great challenges modeling atmosphere. Here, we present a neural network-based (NN-based) diagnostic scheme that uses grid-mean temperature, pressure, liquid ice water mixing ratios, relative humidity to simulate sub-grid CF. The scheme, trained using CloudSat data with explicit consideration of grid sizes, realistically simulates observed CF correlation coefficient >0.9 for liquid-, mixed-, ice-phase clouds. also captures non-monotonic relationship between is computationally efficient, robust GCMs variety horizontal vertical resolutions. For illustrative purposes, conducted comparative analyses 2006–2019 climatological-mean cloud fractions among CloudSat, simulations from NN-based Xu-Randall (optimized same way as scheme). improves not only spatial distribution total but structure. example, biases too-many high-level clouds over tropics low-level regions around 60°S 60°N are reduced. These improvements found be insensitive spatio-temporal variability large-scale meteorology conditions, implying can used different climate regimes.