Invited Talk
Application of Deep Learning to Atmospheric Model Physics Parameterization
Guang-Jun ZHANG, University of California San Diego (AS58-A006)

Biography

Dr. Guang Jun Zhang is a Distinguished Researcher at Scripps Institution of Oceanography, University of California San Diego. He obtained his Ph.D. from University of Toronto in Canada in atmospheric physics and worked at the Canadian Climate Centre afterwards before joining Scripps Institution of Oceanography in 1992. He is the author of the well-known Zhang-McFarlane convection parameterization scheme, which has been used widely in major climate modeling centers in the world, including the US NCAR Community Earth System Model and its predecessors in the last 30 years, the US DOE Energy Exascale Earth System Model (E3SM), the Canadian Centre for Climate modeling and analysis (CCCma) CanESM, and the Norwegian Earth System Model NorESM. In the past several years, he has been actively involved in the E3SM model development. His research interests are in the areas of global climate model development, convection and cloud parameterization and the simulation and understanding of tropical climate systems including ITCZ and Madden-Julian oscillation. Recently, he became interested in machine learning and its application to climate science. His team has developed an advanced neural network emulator for subgrid processes in climate models.

Data-driven approaches using machine learning to representing subgrid model physical processes such as convection and clouds in global climate models (GCM) have been actively explored in recent years. Most of the studies so far are for offline tests of trained neural network emulators or for aquaplanet simulations. A few years ago, we developed a deep-learning-based moist physics parameterization using a convolutional residual neural network (NN) with training data from a superparameterized GCM with real-world geography. The NN uses both current environmental variables and history information of convection as input 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 parameterization scheme has very high prediction accuracy for all output variables considered. Furthermore, although trained on data from the current climate, it generalizes well to a warmer climate in an offline test. It is found that convective memory plays a dominant role in both model prediction accuracy and generalization. The neural net design also plays an important role that cannot be ignored. In this talk, in addition to briefly summarizing our previous work, I will present results from our recent work on hybrid GCM simulations for both the current climate and a warmer future climate. We performed multi-year stable online integrations using the NN-based parameterization that replaces conventional convection and cloud parameterization schemes in the host GCM. The simulations capture the global precipitation distribution well but has significant temperature and moisture biases in high latitudes. Various approaches are explored to improve the model simulation and to understand factors that contribute to stable model integrations and generalization to warm climates.