Description du sujet de thèse

We are recruiting a highly motivated person to work at the intersection of explainable artificial intelligence (XAI) and weather forecasting. The ideal candidate will have solid theoretical and practical experience in neural networks and experience in numerical simulation or remote sensing.

## Project: The successful candidate will work on the definition of new explainability methods dedicated to modern deep neural network architectures for weather forecasting. The research will first aim to understand how these neural network architectures transform physically relevant information into statistically rich latent spaces, and then into weather forecasts. Such analysis is indeed crucial to ensure the robustness of these methods, especially for rare events, and to detect physically irrelevant prediction strategies. Different post-hoc techniques based on the analysis of image sensitivity will first be adapted to this task. Emphasis will be placed on the explanation of decisions taken in extreme or rare weather conditions. The use of recent concept-based explainability strategies for specific tasks will then be developed. Drawing inspiration from hybrid AI/Physics methods to explain weather forecasts is also a research track that we intend to explore.

## Main responsibilities:
1. Perform a comprehensive literature review on DNN-based weather forecasts and explainable AI for neural networks.
2. Adaptation and experimentation of existing XAI techniques for images on meteorological data.
3. Develop concept-based XAI strategies that can explain the impact of non-local and multi-scale information in input data for weather forecasting.
4. Find relevant representations of explanations for different types of users (general public, meteorologists, decision-makers)
5. Present research results at conferences and prepare manuscripts for publication in peer-reviewed journals.

## Qualifications:
- Engineering school or Master 2 in applied mathematics, computational physics, computer science or in a related field
- Solid theoretical and practical training in neural networks
- Experience in managing large data sets
- Mastery of Python and PyTorch
- Excellent communication and collaboration skills
- Advanced level in spoken and written English
- Experience in numerical simulation or atmospheric sciences would be an advantage

Contexte de travail

This thesis will be carried out at the Institute of Mathematics of Toulouse (IMT, UMR CNRS 5219), at the University of Toulouse and at the European Center for Research and Advanced Training in Scientific Computing (CERFACS), in collaboration with the French National Center for Meteorological Research (CNRM - UMR 3589), all gathered in a chair of the Institute of Artificial and Natural Intelligence Toulouse (ANITI). The person recruited will be in contact with AI researchers from ANITI, IMT and CERFACS, as well as with weather forecasting researchers/engineers from the CNRM. Finally, it will have access to the IT resources and internal data sets of these institutions to develop original and impactful explainability solutions in the dynamic field of DNN-based weather forecasts. XAI methods will ideally be developed as extensions of the Anemoi package of the European Centre for Medium-Range Weather Forecasts.

Contraintes et risques

No specific risk