ClimSim2: Coupling Global Climate Models with Neural Network Cloud Emulators with Sungduk Yu

Abstract: In this presentation, Yu discusses ongoing progresses and challenges encountered in ClimSim2, the next phase of the ClimSim project. ClimSim2 extends beyond the initial focus on data generation and curation, targeting to dynamically couple neural network (NN) emulators with operational global climate models. Using the Fortran-Keras Bridge, the MLP (multi-layer perceptron) baselines have been tested in various configurations: including single and ensemble model inference, as well as a novel partial coupling scheme, in which only a chosen subset of NN output variables is coupled. Echoing their poor skills in offline evaluations, results indicate that cloud condensate tendencies are identified as the most critical factor leading to early model crashes—guiding where we should put our future efforts. Additionally, Yu shares the current status of collaborative software engineering efforts to integrate PyTorch within the Fortran code base of climate models, aiming for seamless integration of more sophisticated models beyond MLP.

Bio: Yu is a project scientist at the University of California, Irvine, specializing in developing and integrating machine learning algorithms for earth system models. His Ph.D. from the University of California, Irvine, explored multi-scale climate modeling frameworks and the global energetics of the coupled climate system. Following this, his postdoctoral work at Yale University delved into the predictability of extreme El Niño events and tropical climate dynamics, broadening his expertise in climate science.