Description du sujet de thèse

PINNACLE: Physics-Informed Neural Networks for Accelerated Cloud Light-Scattering Emulation

Artificial intelligence is profoundly transforming atmospheric sciences, particularly in the domain of numerical weather prediction. On the one hand, state-of-the-art models such as GraphCast have demonstrated outstanding predictive skill. On the other hand, physics-informed deep learning is rapidly advancing, integrating artificial intelligence with the governing physical laws to achieve more faithful representations of atmospheric processes.

In the field of remote sensing, the continuous increase in the spatial resolution of satellite sensors considerably amplifies the computational requirements for analyzing ever more detailed observations. Physics-informed deep learning offers a promising avenue to accelerate the realistic modeling of atmospheric radiative signals while maintaining physical consistency.

In this project, we propose to investigate the synergistic use of convolutional neural networks, graph neural networks, and attention-based architectures, with the attention mechanisms explicitly guided by the physical principles and intrinsic properties of the atmosphere. This approach aims to enable the development of computationally efficient models for simulating radiative transfer in three-dimensional cloudy atmospheres.

Contexte de travail

The successful candidate will be appointed at the Laboratoire d'Optique Atmosphérique (LOA – UMR 8518) in Villeneuve-d'Ascq (Université de Lille and CNRS), within the Radiation–Cloud Interaction research group, and will work in close collaboration with the FOX team of the CRISTAL laboratory.

The LOA team has internationally recognized expertise in the field of radiative transfer and remote sensing. The CRISTAL laboratory team is likewise highly recognized for its research in computer vision and neuro-inspired artificial learning. Both teams have been collaborating for four years on projects at the interface between atmospheric sciences and artificial intelligence (two PhD theses, one completed and one ongoing), and have already established a common framework ensuring effective collaboration.

The appointed doctoral candidate will be responsible for the full range of activities related to the development, testing, and evaluation of the learning models. They will also receive support from LOA experts for the execution of the Monte Carlo simulations required to generate the training datasets.

At LOA, the candidate will have access to appropriate research infrastructure.

This position falls within the scope of the Protection of Scientific and Technical Potential (PPST). In accordance with applicable regulations, the candidate's appointment is therefore subject to prior authorization by the competent authority of the French Ministry of Higher Education and Research (MESR).

Le poste se situe dans un secteur relevant de la protection du potentiel scientifique et technique (PPST), et nécessite donc, conformément à la réglementation, que votre arrivée soit autorisée par l'autorité compétente du MESR.

Contraintes et risques

International travel to attend major scientific conferences will be required.
This is an interdisciplinary PhD project at the interface of Atmospheric Sciences and Computer Science.