Kleinskalige Prozesse in der Atmosphäre verstehen und modellieren mit Deep Learning
Zusammenfassung der Projektergebnisse
The goal of this project was to improve the application of machine learning to modeling clouds in climate models. This emergent research direction is made possible by the recent availability to run global cloud-resolving simulations for short periods of time. Such short-term, highresolution simulations, while too short to assess the impact of climate change, could be useful as a training dataset for machine learning approaches that aim to represent or parameterize the subgrid clouds in coarser models. In this project specifically, several issues with existing approaches were tackled. First, initial machine learning parameterization did not follow the laws of physics, in particular energy and mass conservation. In this project, together with collaborators I developed approaches to explicitly force the neural networks to obey physical constraints with essentially similar predictive performance. Second, one major issue in building hybrid physics-machine learning models is that coupled simulations often turn out to be unstable. After extensive discussion with several researchers, I proposed a method to combat these issues. This method works on a simplified toy model but applications to full climate models are technically challenging. The challenges of the approach described above motivated me to explore other research directions. First, in a collaboration we explored focusing on a much smaller subproblem using machine learning, specifically the formation and growth of cloud droplets. Using a betterdefined problem simplifies the machine learning setting and reduces the number of unwanted behaviors. The resulting machine learning parameterization shows impressive skill but, once again, fails once run for longer periods. This actually allowed us to gain some insight into the physical problem at hand. Second, I explored a completely data-driven approach. While this does not allow us to take advantage of the physical knowledge we have, it greatly simplifies the machine learning approach. In particular I focused on medium-range, global weather prediction. As a first step I designed, in collaboration with other researchers, a benchmark dataset that enables easy comparison between different approaches. The benchmark dataset was well received and was covered in the popular science magazine EOS. Then, we trained a state-of-the-art deep learning model to predict pressure and temperature several days ahead. Our model is similarly good as a physical model at similar resolution. However, we hypothesize that it is difficult to scale this approach to the resolutions of current operational weather models because note enough training data is available. Finally, the results from the project show that, while machine learning approaches certainly have great potential for some areas in weather and climate, not all problems lend themselves easily to the application of machine learning. For cloud parameterizations, for example, fundamental problems persist and much research will still be needed.
Projektbezogene Publikationen (Auswahl)
- 2020. Coupled online learning as a way to tackle instabilities and biases in neural network parameterizations: general algorithms and Lorenz 96 case study (v1. 0). Geoscientific Model Development
Rasp, S.
(Siehe online unter https://doi.org/10.5194/gmd-13-2185-2020) - 2020. Enforcing Analytic Constraints in Neural-Networks Emulating Physical Systems
Beucler, T., Pritchard, M., Rasp, S., Ott, J., Baldi, P. and Gentine, P.
(Siehe online unter https://doi.org/10.1103/PhysRevLett.126.098302) - 2020. Potential and Limitations of Machine Learning for Modeling Warm-Rain Cloud Microphysical Processes. Journal of Advances in Earth System Modeling
Seifert, A. and Rasp, S.
(Siehe online unter https://doi.org/10.1029/2020MS002301)
