Contexte et atouts du poste

Context

This PhD thesis will be co-supervised by Emmanuel De Bézenac and Claire Monteleoni (ARCHES team, Inria Paris), in the EDITE doctoral program, Sorbonne University, and Xavier Renard at AXA Research, Paris. The position will be funded by AXA.

Mission confiée

Recent advancements in AI weather models have shown promising capabilities in weather prediction from short to medium-term (0 to 10 days lead time), comparable to physic-based state-of-the-art models. Notable examples include FourCastNet, SFNO, GraphCast, GenCast, Pangu-Weather, which have adopted a trend observed in other fields (e.g., language models, vision models, multi-modal models): they leverage large architectures with an extensive pre-training on very large datasets. This type of large models (also called foundation models) facilitates the development of downstream applications.

In the context of weather forecasting, these AI models are typically trained on the gold standard global reanalysis dataset ERA-5, developed by the European Centre for Medium-Range Weather Forecasts. ERA-5 is widely used for weather and climate studies due to its high resolution, comprehensive coverage, and accurate representation of atmospheric conditions from 1979 to the present day. AI models, such as FourCastNet or GraphCast, are trained on ERA5 data: they take a snapshot of the atmosphere state at time (t) and they are trained to predict the atmospheric conditions at time (t+1) (typically 6 hours later).

This PhD aims to assess and develop the capabilities of AI weather models in simulating extreme events, a promising avenue for risk estimation, prevention, etc. However, potential issues may prevent accurate simulations: for instance, model biases towards average weather conditions (e.g., reliance on RMSE metric) and the underrepresentation of extreme events in the ERA5 dataset.

Principales activités

The PhD (under a CIFRE scheme) will be jointly hosted by the AI research team at AXA in collaboration with the Natural Catastrophe R&D modelling team at AXA and the ARCHES team at INRIA. The primary objective is to develop knowledge and methodologies to better understand climate risks using AI weather models. The PhD student will evaluate and enhance the capabilities of AI weather models to simulate plausible and unobserved extreme events, focusing on identifying or developing datasets and metrics to assess model performances. The PhD student will explore solutions to improve existing models or create new approaches to address the limitations in extreme event simulation. The downscaling of the simulations to a higher resolution than the native ERA-5 resolution will also be addressed during the PhD. A key component of this work will involve estimating climate risks, particularly to determine the optimal set of initial conditions to input AI weather models to get, in fine, an accurate estimation of the risk of extreme weather event across various geographical locations. While this research is open to any major climate risks (tropical cyclones, extra-tropical cyclones, windstorms, hail, floods, heatwaves, etc.), the student may focus on a subset of perils.

Compétences

Technical skills and level required :

• Bachelors and Master's degrees in Computer Science (Informatique), statistics, mathematics, or a related field
• Machine learning, data mining, statistics, and/or AI coursework and/or projects
• Familiarity with modern machine learning / deep learning software, tools, pipelines

Languages :

• Written competency in English
• Oral competency in English or French