ESAC Student Seminar – Milan Klöwer

Building climate models that learn from data

Abstract
Climate models provide crucial information for climate change mitigation and adaptation. But they also lack necessary accuracy with respect to many atmospheric variables. Conventional atmospheric models are built on physical laws for the resolved dynamics but so-called parameterizations of processes that are not explicitly resolvable including cloud formation and precipitation. However, decades of observational data are available which a model could automatically correct itself to. In this talk, I want to highlight recent advances in data-driven atmospheric models, such as GraphCast and NeuralGCM, and what is needed to step from weather to climate predictions. Especially NeuralGCM is the first hybrid model with a conventional dynamical core but machine-learned parameterizations. Trained on 3-day weather forecasts it can be run for decades, for example, reproduce statistically the tropical cyclones of the last 40 years. Online learning allows the best of both worlds: A combination of existing physics knowledge for model components that are more certain, but learned parameterizations to represent uncertain microphysics and constraining them automatically with data. Motivated by the success of NeuralGCM, we are currently developing SpeedyWeater.jl, an atmospheric general circulation model that’s interactive and extensible using the Julia programming language. We want to use SpeedyWeather.jl to better understand how models can be trained towards present-day climate data even though the predictions are for future climates not existent in the training data.

ESAC Student Seminar

A forum for students and postdocs to share recent research, hone presentation skills, and build community among peers, sponsored by the EAPS Student Advisory Committee. Open to current EAPS graduate and undergraduate students and postdocs. Typically hosted on Thursdays during the semester, including pizza lunch.

Contact: esac.officers@gmail.com