Description
Date depot: 3 juillet 2023
Titre: Machine learning for extreme weather forecasting
Directrice de thèse:
Claire MONTELEONI (Inria-Paris (ED-130))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: Extreme weather events, such as heat waves and extreme storms, can have outsized impacts on society, especially considering the resulting hazards, such as drought, wildfire, and flooding. Machine learning can help to improve the forecasting of such extreme events, which will be critical in helping communities adapt to a changing climate.
The current practice of weather forecasting is primarily based on numerical weather prediction (NWP) which involves physics-driven simulations running on supercomputers. Running these NWP models is quite computationally expensive, and thus the forecasts are not updated in real-time. There's currently a race among researchers in the tech industry to outperform NWP at weather forecasting, by training deep learning models on reanalysis data products (e.g. ERA5). After training (which requires significant GPU compute time), these data-driven models output forecasts much faster than standard NWP. While these data-driven models generally outperform NWP on average forecasts, they are not explicitly trained to forecast extreme events, and thus they tend to under-predict them.
This thesis project will focus on improving machine learning-based forecasting of extreme weather events, addressing the current limitations of data-driven models for weather:
(1) Training to optimize average predictions results in less forecast skill for extreme events;
(2) Lack of Interpretablity / Explainability of the models and forecasts.
To address these challenges, this research will entail both algorithm design and applications to different sorts of extreme weather events. On the algorithmic side, we will experiment with generating quantile forecasts and other distributional forecasting approaches. This entails considering different evaluation metrics, as well as how to quantify forecast uncertainty. We will also explore algorithmic pipelines designed to adapt machine learning models learned for average forecasting to focus more on the tails of the distribution (extremes).
On the applications side, decision-makers need models that are particularly skillful at forecasting specific types of extreme events. Many of the recent data-driven models for weather claim to be general (or "foundation") models. For any particular extreme event type, there is much work to be done to improve its forecasting, whether by fine-tuning such foundation models, or by training data-driven models specifically for forecasting the extreme event type in question. This thesis will explore both of these approaches.
Finally, we will experiment with the models for forecasting extreme events, and use post-hoc interpretability techniques (a variety of which have recently emerged in the machine learning literature) to determine their suitability for use by human decision-makers and other stakeholders.
Résumé dans une autre langue: Anglais
Doctorant.e: Dauvilliers Clément