End-to-end learning of geophysically-sound CNN representations from satellite-derived observation datasets

When:
30/04/2020 – 01/05/2020 all-day
2020-04-30T02:00:00+02:00
2020-05-01T02:00:00+02:00

Annonce en lien avec l’Action/le Réseau : aucun

Laboratoire/Entreprise : IMT Atlantique, Lab-STICC
Durée : 36 months
Contact : ronan.fablet@imt-atlantique.fr
Date limite de publication : 2020-04-30

Contexte :
Understanding, modeling, forecasting and reconstructing fine-scale and large-scale processes and their interactions are among the key scientific challenges in ocean-atmosphere science. State-of-the-art approaches strongly rely on joint research effort in observing systems (e.g., in situ monitoring, satellite observations) and numerical simulations [e.g., 30]. The ability to relate models and observation data, though significant advances in data assimilation, remain open questions for numerous processes (e.g., small-scale parameterization, ocean-atmosphere interactions, biogeochemical ocean dynamics, climate-scale dynamics). Artificial Intelligence (AI) technologies, models and strategies open new paradigms to address these questions from the in-depth exploration of the existing observation and simulation big data [1,3,12,20-26].

Among others, the recent breakthrough in the resolution of fine-scale cloud processes in climate models [26] is a striking illustration. It further illustrates the typical learning-based paradigm for ocean-atmosphere processes, where a model or representation is learnt from simulation data. However, for numerous processes, on the one hand, the ability of model simulations to be fully representative of real dynamics is questionable and, on the other hand, one would expect to benefit from the existing observation datasets to extract computational representations. The sampling patterns of these observation datasets (e.g., irregular space time-sampling, partially-observed system,…) raise issues which remain to be addressed to develop fully observation-driven and learning-based frameworks for earth science, including space oceanography.

This PhD scholarship is open in the framework of AI Chair OceaniX (Physics-informed AI for Observation-driven Ocean AnalytiX) (https://rfablet.github.io/projects/2019-oceanix).

Sujet :
In this context, the general goal of this project is to address the following topical questions :
Can we develop fully-observation-driven learning-based paradigms from satellite-derived observation dataset, including synergies with other observation data (e.g., ARGO floats, buoys,…) ?
Can learning-based paradigms better inform past and future dynamics from HR satellite-derived observations of the sea surface ?
The methodological backbone underlying these topical questions is the definition and identification of learning-based representations of geophysical dynamics. In the framework of ANR project Melody (2020-2023) and SWOT ST DIEGO project, the proposed methodological framework will be demonstrated and implemented in the context of incoming SWOT mission towards informing past and future sea surface dynamics from HR SWOT snapshots. Case-studies will be designed based on OSSEs (Observing System Simulation Experiment) and real SWOT data.

The PhD candidat will be involved within interdisciplinary scientific collaboration at the interface of Machine Learning, Data Science and Ocean Science including strong collaborations with space oceanography teams (Dr. B. Chapron, Ifremer/LOPS; J. Le Sommer, CNRS/IGE, A. Pascual, CSIC/IMEDEA) and industrial partners (OceanNext, OceanDataLab).

Link to the detailed PhD project: https://www.imt-atlantique.fr/sites/default/files/rfablet/phd_proposal_rfablet_CNESMelody_2020.pdf

Application by mail to ronan.fablet AT imt-atlantique.fr

Profil du candidat :
The targeted PhD candidate shall have a MSc and/or engineer degree in Data Science or Artificial Intelligence with a strong interest in environmental sciences, possibly acknowledged by previous activities or experience, or a degree in Ocean Science with a significant mathematical and programming background.

Formation et compétences requises :
Besides a strong theoretical background, computer skills, including first experience in using state-of-the-art deep learning frameworks (e.g., tensorflow, pytorch) and programming environment (e.g., python, git server), will be particularly expected.

Adresse d’emploi :
IMT atlantique, technopôle Brest-Iroise, Brest

Document attaché :