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

Laboratoire/Entreprise : LSCE/Lab-STICC
Durée : 36 mois
Contact : ronan.fablet@imt-atlantique.fr
Date limite de publication : 2020-12-31

Contexte :
Extremes are crucial features of geophysical processes and can play a fundamental role in terms of societal impacts, e.g. major floods. By definition, extreme events are rare, but they happen and records are made to be beaten. In terms of machine learning algorithms, it is difficult to learn from very few examples even in a large learning database. In addition, the probability distribution of extreme events cannot be well captured by measures based solely
on deviations from the mean. These two issues clearly challenge the classic learning paradigm. From an uncertainty point of view, there exists a probability theory tailored to model extremal behavior, the so-called Extreme Value Theory
(EVT).

Sujet :
The main task of the PhD student will be to build bridges between physics-informed neural networks (NN), and multivariate EVT used in environmental statistics. A major bottleneck to couple both NN and EVT techniques is the question of metrics for rare events, and how to assess predictive distributions from forecast models. This two aspects will be studied in detail during the PhD.

This PhD will be part of the ANR Melody. This implies that the main application domain will be the field of ocean dynamics and consequently, all algorithms will be tested on low–dimensional toy models (Lorenz models) or intermediate size models (1D-Burgers equations, 2D-QG models, etc).

The PhD will be co-supervised by Dr. P. Naveau (CNRS, LSCE), Dr. A. Sabourin (IMT, LTCI) and Dr. R. Fablet (IMT, Lab-STICC). The PhD could take place in Paris and/or Brest.

Profil du candidat :
MSc. and/or Engineer degree in Data science, statistical learning and/or geosciences (ocean dynamics)

Formation et compétences requises :
Programming skills (Python)
Machine Learning and Deep learning skills (e.g., scikit-learn, pytorch, keras…)

Adresse d’emploi :
LSCE (Paris, Saclay) and/or IMT Atlantique (Brest)