Abstract

The Machine Integration and Learning for Earth Systems (MILES) group at the US NSF National Center for Atmospheric Research collaborates with groups across the weather/climate spectrum to develop physics-informed machine learning systems that integrate closely with existing community physics-based models. One of our first emulation projects focused on the development of a spectral bin warm-rain microphysics emulator for the Community Atmospheric Model. In recent work, we have reduced our original 7 neural network model to 1 and 3 neural network versions that we have integrated into CAM with a fortran-based neural network inference framework. We have also developed a machine learning surface layer parameterization based on observed near surface atmospheric conditions and surface fluxes. We have tested this model within WRF and the FastEddy GPU LES model to understand some of the sensitivities and important variables. Finally, I will discuss some preliminary results from our new full atmospheric model emulator, the Community Runnable Earth Digital Intelligence Twin (CREDIT). The model is trained on ERA5 model level output and can perform stable 1-hour rollouts to 10 days with a single model. We intend the framework to support community training and inference of AI weather models.

Hybrid participants can join using the following zoom link: https://cam-ac-uk.zoom.us/j/88550555680?pwd=VHgwMW8zRXJ4UTBnUExSSW10NXNjUT09