Goals of the chair
Accurate predictions of future weather conditions are essential for the safety of people and goods, and for the management of a wide range of economic activities.
Since the mid-20th century, weather forecasting has relied on the physical modeling of atmospheric dynamics. Significant and steady improvements in the quality of these forecasts have been achieved, particularly thanks to increased computational resources and the use of new observations. Nevertheless, the accurate forecasting of local, high-impact phenomena remains difficult and costly with current modeling tools.
Principal investigator
- Laure Raynaud (Researcher at Météo-France, HDR, CNRM)
- Laurent Risser (IR CNRS, HDR)
Artificial intelligence (AI), and particularly deep learning methods utilizing neural networks, has been successfully applied in a wide range of applications in recent years. Weather forecasting is no exception; the potential uses of AI in this field are numerous and could lead to major methodological advancements, coupled with significant gains in performance and quality.
This chair will focus on one of the most promising perspectives for using deep learning in weather forecasting: the development of new models entirely learned from data. This approach will be leveraged for local-scale probabilistic forecasting, and its combination with existing physical models will also be explored.
Leveraging recent publications on the subject, state-of-the-art architectures such as Vision Transformers, Graph Neural Networks, and Physics-Informed Neural Networks will be deployed. The results will provide, on one hand, a new model combining deep learning and physics for better anticipation of extreme events. On the other hand, there will be a thorough evaluation of the forecast quality obtained with current models and with the new deep learning models, along with recommendations for their use in operational conditions.
Co-chairs
-
- Valérie Masson (Research Director Météo-France)
- Christophe Bovalo (R&D Engineer)
- Luciano Drozda (Researcher Cerfacs)
Beyond the potential methodological and technological breakthroughs resulting from this work, another objective of the chair is to qualify the robustness and explainability of deep learning-based forecasting models. This will involve, in particular, applying certain standard Explainable AI (XAI) methods to the developed models, adapting them to the specificities of meteorological data. The integration of physical constraints into the construction and training of deep learning models will be another way to ensure their quality and strengthen user confidence in these new tools.
This chair project is fully aligned with adaptation strategies in the context of a changing climate and increased needs for precise, local forecasts. Improved weather forecasting will have a direct impact on the anticipation and real-time monitoring of extreme weather events. The chair's results will allow for better prediction of the location, intensity, and timing of these events, which will contribute to enhancing alert systems and emergency preparedness capabilities.
Improved weather forecasts will also enable better prediction of indirect impacts such as floods, landslides, air pollution, and wildfires. Furthermore, vulnerable sectors like agriculture, energy, and transportation will directly benefit from very high-quality forecasts for managing their activities.
To achieve these ambitious objectives, the chair proposes an integrated approach, supported by a multidisciplinary consortium bringing together complementary expertise in atmospheric modeling, deep learning, software engineering, and mathematics.
- Explore deep learning methods to produce local-scale weather forecasts.
- Optimization strategies and parallelization to train large ML models on massive datasets
- Develop physics-informed learning models.
- Extend and apply existing explainability diagnostics to analyze the relationships learned by deep learning models.
- ATOS-Eviden : Development of Machine Learning emulators for weather forecasting, collaboration coordinated by L. Raynaud.
- MAGELLIUM : Hybrid Physics-DNN models in remote sensing, Cifre PhD thesis supervised by L. Risser and X. Briottet.
- Berger-Levrault : Extraction of regulatory information from unstructured documents, Cifre PhD supervised by C. Trojahn, M. Chevallier and B. Billami, 2021.
- Akkodis : Supervision by L. Risser of an Engineer made available by Akkodis to work on bias and explainability issues in high-dimensional machine learning.
- Thales Alenia Space : a semantic approach to characterizing change in the context of Earth observation using image processing and contextual data. Cifre PhD supervised by C. Trojahn, N. Aussenac-Gilles and Romain Hugues.
- New generation of data-based weather forecasting models with physical constrains.
- Trustworthy and explainable ML models on very-high dimensional data with physical constrains
English 
French