Using Deep Learning for Convection Parameterization

Data-driven approaches using machine learning to parameterizing model physical processes in Earth System Models have been actively explored in recent years. Deep-learning-based convection parameterization is one such example. While significant progress has been made in emulating convection using neural networks (NN), serious roadblocks remain, including generalization of the NN-based scheme trained on model data from current climate to future climate and integration instability when it is implemented into the model for long-term integrations. This study uses a deep residual convolutional network to emulate convection simulated by a superparameterized global climate model (GCM). The NN uses the current environmental state variables and advection tendencies, as well as the history of convection to predict the GCM grid-scale temperature and moisture tendencies, cloud liquid and ice water contents from moist physics processes. Independent offline tests show that the NN-based scheme has extremely high prediction accuracy for all output variables considered. In addition, the scheme trained on data in the current climate generalizes well to a warmer climate with +4K sea surface temperature in an offline test, with high prediction accuracy as well. Further tests on different aspects of the NN architecture are performed to understand what factors are responsible for its generalization ability to a warmer climate. We are also able to perform multi-year integrations, without encountering any integration instability, when the scheme is implemented into the NCAR CAM5. The details will be presented at the meeting.