Stable climate simulations using a realistic GCM with neural network parameterizations for atmospheric moist physics and radiation processes
Abstract. In climate models, subgrid parameterizations of convection and cloud are one of the main reasons for the biases in precipitation and atmospheric circulation simulations. In recent years, due to the rapid development of data science, Machine learning (ML) parameterizations for convection and clouds have been proven the potential to perform better than conventional parameterizations. At present, most of the existing studies are on aqua-planet and idealized models, and the problems of simulated instability and climate drift still exist. In realistic configurated models, developing a machine learning parameterization scheme remains a challenging task. In this study, a group of deep residual multilayer perceptrons with strong nonlinear fitting ability is designed to learn a parameterization scheme from cloud-resolving model outputs. Multi-target training is achieved to best balance the fits across diverse neural network outputs. The optimal machine learning parameterization, named NN-Parameterization, is further chosen among feasible candidates for both high performance and long-term simulation. The results show that NN-Parameterization performs well in multi-year climate simulations and reproduces reasonable climatology and climate variability in a general circulation model (GCM), with a running speed of about 30 times faster than the cloud-resolving model embedded Superparameterizated GCM. Under real geographical boundary conditions, the hybrid ML-physical GCM well simulates the spatial distribution of precipitation and significantly improves the frequency of precipitation extremes, which is largely underestimated in the Community Atmospheric Model version 5 (CAM5) with the horizontal resolution of 1.9° × 2.5°. Furthermore, the hybrid ML-physical GCM simulates a stronger signal of the Madden-Julian oscillation with a more reasonable propagation speed, which is too weak and propagates too fast in CAM5. This study is a pioneer to achieve multi-year stable climate simulations using a hybrid ML-physical GCM in actual land-ocean boundary conditions. It demonstrates the emerging potential for using machine learning parameterizations in climate simulations.