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

Laboratoire/Entreprise : Sorbonne Universite – Institut des Systèmes intell
Durée : 36 mois
Contact : patrick.gallinari@sorbonne-universite.fr
Date limite de publication : 2022-09-30

Contexte :
Deep Learning is beginning to be explored for scientific computing in domains traditionally dominated by physics models (first principles) like earth science, climate science, biological science, etc. It is particularly promising in problems involving processes that are not completely understood, or computationally too complex to solve by running the physics inspired model. The direct use of pure machine learning approaches has however met limited successes for scientific computing. Hence, researchers from different communities have started to explore (i) how to integrate physics knowledge and data, and (ii) how to push the limits of current ML methods and theory; two challenging directions.

Sujet :
The global objective of the thesis is the development of new models leveraging observation or simulation data for the modeling of complex spatio-temporal dynamics characterizing physical phenomena such as those underlying earth-science and climate observations. The classical modeling tools for such dynamics in physics and applied mathematics rely on partial differential equations (PDE). Despite their successes in different areas, current ML based approaches are notably insufficient for such problems. Using ML for physics raises new challenging problems and requires rethinking fundamental ML ideas.

Three main research directions will be explored:

– Hybrid systems – Integrating Physics and Deep Learning. In many situations, there is available some prior physical knowledge provided by PDEs for characterizing the underlying phenomenon. A key issue is then how to combine this prior knowledge with information extracted from the data.

– Domain generalization for deep learning as dynamical models. Explicit physical models come with guarantees and can be used in any context (also called domain or environment) where the model is valid. This is not the case for DNNs, and we have no guarantee that they can be extrapolated to new physical environments.

– Learning at Multiple Scales. Modeling dynamical physical processes often requires considering multiple spatio-temporal scales. For example in climate, global phenomena are influenced by dynamics operating at a smaller scale. Similar problems occur e.g. in computational fluid dynamics. Learning at different scales is an open issue in ML.

Motivating problems and applications will come from climate science.

Profil du candidat :
Computer science or applied mathematics, Good technical skills in programming.

Formation et compétences requises :
Master in computer science or applied mathematics, Engineering school. Background and experience in machine learning. Good technical skills in programming.

Adresse d’emploi :
Sorbonne Université (S.U.), Pierre et Marie Campus in the center of Paris. The candidate will integrate the Machine Learning and Deep Learning for Information Accesss team at S.U. at the ISIR lab. (Institut des Systèmes Intelligents et de Robotique).

Document attaché : 202203010947_2022-PhD-Description-ML-Physics.pdf