Towards robust operational neural network parameterizations of convection in climate models — advances in stability, credibility and software

Abstract:

Machine learning based parameterization of moist convection for climate modeling has been proved in concept since 2018 by several research groups in the limit of an idealized aquaplanet but unsolved challenges remain. I will review some of the interesting findings from follow-on tests during the past two years that have clarified the potential for such an approach to work in more realistic settings focusing on the engineering lessons learned and the challenges that remain. Along the way I will discuss emerging diagnostics for testing the physical credibility of prototype neural network emulators, and their relationship to prognostic test results, the growing importance of formal hyperparameter tuning, some strategies that are helping incorporate physical constraints and generalizability into hybrid machine learning models, and new software that makes prognostic testing simpler. I will also share our latest measurements of skill when emulating global cloud superparameterization using deep neural networks trained on large-ensemble hyperparameter searches, within a modern version of the superparameterized Community Earth System Model that includes real geography, seasons, and diurnal cycles.