Offre en lien avec l’Action/le Réseau : MACLEAN/– — –

Laboratoire/Entreprise : Lab-STICC
Durée : 36 mois
Contact : ronan.fablet@imt-atlantique.fr
Date limite de publication : 2021-05-31

Contexte :
This PhD position is opened in the framework of AI Chair OceaniX on physics-informed deep learning for ocean monitoring and surveillance.

The design of observing systems for ocean monitoring and surveillance is a key issue for a wide range of applications and scientific challenges. Multi-platform observing systems naturally arise as appealing solution to best exploit different observation modalities (e.g., satellite vs. in situ observation, surface vs. interior observation, passive vs. active sensors,….). In this context, the ability to design optimal monitoring and sampling strategies is a key challenge. The emergence of deep learning [1], especially end-to-end learning and deep generative models, provides new means to investigate this challenge jointly to the design of the inverse model, that is to say the design of optimal monitoring and sampling strategies so that one can retrieve the best reconstruction of some processes of interest and/or reduce the associated uncertainty.

Sujet :
In the framework of AI Chair OceaniX (https://cia-oceanix.github.io/about/), this PhD will investigate this generic issue and its application to ocean monitoring and surveillance through relevant case-studies (i.e., design of future earth observation missions, optimal synergy control between satellite platforms and in situ sensor networks, data-driven adaptive sampling strategies for in situ networks). From a methodological point of view, the proposed framework will explore synergies between Deep learning planning schemes [4], Bayesian variational setting [3], learning- based data assimilation models [2] and solvers and deep generative models, especially GANs [5]. The expected outcome is the ability to jointly learn observation operators, priors and solvers w.r.t. performance measure for hidden dynamics. Through relevant constraints (e.g., sparsity priors) on the observation operators, we expect to derive the targeted optimal measurement and sampling strategies.
Numerical experiments will be carried out initially on toy examples (e.g., low-dimensional chaotic systems). OSSEs (Observing System Simulation Experiments) generated from realistic numerical simulations will also be considered. Applications to real datasets might also be of interest in the last stage of the PhD.

Profil du candidat :
Applicants must hold a M.Sc degree in signal processing, mathematics or physics.

Formation et compétences requises :
Background in machine learning, deep learning and applied statistics expected.
Knowledge and experience with deep learning frameworks would be a plus.

Adresse d’emploi :
Brest

Document attaché : 202104161258_offre_these_en.pdf