DESCRIPTION :
Recent advances in generative modeling have opened new possibilities for representing complex geophysical fields. In weather forecasting, modern diffusion, flow-matching, and flow-based models have demonstrated strong performance, but current approaches struggle with multiscale atmospheric structure and with the fine-to-coarse interactions typical of physical fields. Improving the spectral and dynamical fidelity of these generative models is an important step toward reliable data-driven forecasting systems.
The master thesis will investigate how to adapt and refine generative modeling frameworks-particularly diffusion and flow-matching methods-to atmospheric data. The focus is on understanding how different scales of motion are represented during the generation process, how multiscale structures propagate through the model dynamics, and how model design choices influence numerical stability and physical realism. The work will contribute to the development of next-generation AI weather models capable of representing both global-scale patterns and fine-scale variability.
Principales activités
* Study existing generative modeling frameworks used in scientific data (diffusion, flow matching, deterministic flows) and analyse how they behave on multiscale atmospheric fields.
* Explore strategies for representing scale interactions during the generation process, such as alternative noise designs, interpolation schemes, or multiscale parameterizations.
* Develop and test model adaptations tailored to weather data-e.g., architecture choices, training procedures, or sampling methods that improve spectral accuracy and physical consistency.
* Contribute to prototype implementations and small-scale experiments within the ARCHES weather modeling effort.
Code d'emploi : Mannequin Photo (h/f)
Domaine professionnel actuel : Employés du Service de la Promotion des Ventes
Niveau de formation : Bac+3
Temps partiel / Temps plein : Plein temps
Type de contrat : Contrat à durée indéterminée (CDI)
Compétences : Intelligence Artificielle, Vision par Ordinateur, Data Mining, Python (Langage de Programmation), Machine Learning, Pytorch, Deep Learning, Technologies Informatiques, Anglais, Français, Recherche, Architecture, Expérimentation, Elaboration des Prévisions, Traitement d'Image, Mathématiques, Réalisation de Prototypes, Recherche Scientifique, Transformateurs (Électrique), Etudes et Statistiques, Prévision Météo, Interpolation
Courriel :
bezenac@inria.fr
Téléphone :
0139635511
Type d'annonceur : Employeur direct