An Ensemble of Neural Networks for Moist Physics Processes, Its Generalizability and Stable Integration.

With the recent advances in data science, machine learning has been increasingly applied to convection and cloud parameterizations in global climate models (GCMs). This study extends the work of Han et al. (2020, https://doi.org/10.1029/2020MS002076) and uses an ensemble of 32‐layer deep convolutional residual neural networks, referred to as ResCu‐en, to emulate convection and cloud processes simulated by a superparameterized GCM, SPCAM. ResCu‐en predicts GCM grid‐scale temperature and moisture tendencies, and cloud liquid and ice water contents from moist physics processes. The surface rainfall is derived from the column‐integrated moisture tendency. The prediction uncertainty inherent in deep learning algorithms in emulating the moist physics is reduced by ensemble averaging. Results in 1‐year independent offline validation show that ResCu‐en has high prediction accuracy for all output variables, both in the current climate and in a warmer climate with +4K sea surface temperature. The analysis of different neural net configurations shows that the success to generalize in a warmer climate is attributed to convective memory and the 1‐dimensional convolution layers incorporated into ResCu‐en. We further implement a member of ResCu‐en into CAM5 with real world geography and run the neural‐network‐enabled CAM5 (NCAM) for 5 years without encountering any numerical integration instability. The simulation generally captures the global distribution of the mean precipitation, with a better simulation of precipitation intensity and diurnal cycle. However, there are large biases in temperature and moisture in high latitudes. These results highlight the importance of convective memory and demonstrate the potential for machine learning to enhance climate modeling. Plain Language Summary: The representation of storms and clouds through empirical algorithms known as parameterizations in global climate models (GCMs) is one of the main sources of biases in the simulation of rainfall and atmospheric circulation. Here an ensemble of 8 deep neural networks are used to replace the conventional parameterization of atmospheric moist physics processes. They are trained on data sampled from 1‐year present‐day climate simulation by a "superparameterized" climate model, which uses a two‐dimensional cloud‐scale model to explicitly simulate convection and clouds inside each GCM grid box. On ensemble averaging, the neural nets produce highly accurate predictions of precipitation characteristics including global distribution and intensity. Furthermore, the machine‐learned emulator trained on data in the current climate also represents convection and precipitation extremely well in a warmer climate. A member of the ensemble of the neural nets is implemented into a GCM. The model is then integrated for 5 years, producing reasonable results. Key Points: An ensemble of deep convolutional residual neural networks is used to reduce the uncertainty in moist physics emulationsThe ensemble of the neural networks trained on data from a present‐day climate simulation generalizes well to a +4K warm climate offlineA multi‐year stable online integration is achieved in a real‐geography GCM with reasonable results [ABSTRACT FROM AUTHOR]