A Neural‐Network‐Based Scheme for Improving Wegener–Bergeron–Findeisen Process and Its Impact on Mixed Phase Cloud

ABSTRACT Many general circulation models (GCMs) underestimate the liquid cloud and overestimate the ice cloud on mixed‐phase levels (temperature between −40°C and 0°C). The liquid‐to‐ice conversion by WBF (Wegener–Bergeron–Findeisen) process is unrealistically strong due to a hypothesis that liquid and ice is a homogeneous mixture in widely used microphysical scheme. Leveraging machine learning advances, we use a neural network (NN) with a physically constrained output activation function to parameterize the degree of heterogeneous distribution (DHD) between liquid and ice clouds. It is trained and independently evaluated with 2‐year simulations from a superparameterized GCM, SPCAM6. The 10 key predicted features selected by NN include quality of cloud water, ice and snow, temperature and pressure with their vertical gradients. The NN can accurately capture the most of DHD variability, achieving a high correlation coefficient exceeding 0.9. The NN‐based scheme is applied to constrain the WBF process in an independent GCM, Community Integrated Earth System Model (CIESM). Evaluation of two decadal experiments shows that the constrained schemes can improve the distribution of simulated mixed‐phase clouds, mainly in mid‐high latitudes. The constraint on WBF process enhances the amount of liquid cloud and reduces ice cloud on mixed‐phase levels. More ice clouds occur on the upper levels with temperature below −40°C due to the enhanced freezing processes, leading to the reduction of latent heating rate and increase of atmospheric stability. Compared to constant‐value schemes, the NN‐based scheme provides stronger improvement for ice clouds in mid‐high latitudes, while also avoiding the overestimation near −20°C at high latitudes that occurs with the constant‐value schemes. This scheme provides a more reasonable modulation in liquid and ice clouds, resulting in more realistic cloud radiative forcing and precipitation simulations globally.