This project targets a difficult problem in weather and climate prediction -- the representation of convection. Accurate representation of convection is important, since a majority of current model predictions depend on it. Unraveling the physics involved in convective conditions, clouds and aerosols may take years of modeling to fully understand; however, a set of machine learning techniques, known as "neural net techniques", may provide enhanced predictability in the interim, and this project explores their potential.
The project develops a Python library enabling the use of machine learning (artificial neural networks) in a broad range of science domains. The focus is on integration of convection and cloud formation within larger-scale climate models, with the Community Earth System Model (CESM) as an initial target. The project develops a new set of machine learning climate model parameterizations to reduce uncertainty in weather and climate predictions. The neural networks will be trained on high-fidelity simulations that explicitly resolve convection. Two types of high-resolution simulations will be used for training the neural networks: 1) an augmented super-parameterized simulation, and 2) a full Global Cloud Resolving Model (GCRM) simulation based on the ICOsahedral Non-hydrostatic (ICON) modelling frameworks provided by the Max Planck Institute, using initial 5km horizontal resolution. The effort has the potential to increase understanding of convection dynamics and processes across scales, and could potentially be implemented to address other scale problems as well, where it is too computationally costly or impractical to represent processes occurring at much finer scales than the main grid resolution.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.