Description
Date depot: 12 avril 2019
Titre: Data Assimilation for hybrid physical-machine learning models. Application on the nowcasting of precipitations
Directeur de thèse:
Dominique BÉREZIAT (LIP6)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
The objective of this thesis is to explore the emergent field that combines the disciplines of Machine Learning and Data Assimilation. In the literature, it has been shown equivalences between a traditional regression using machine learning and a data assimilation framework.
We aim at determining if we can make use data assimilation in re-analyzing numerical model runs and observations, to constrain neural networks under similar constraints, as well as the learning capabilities of Machine Learning in learning numerically unresolved parts of dynamic systems such as sub-grid parametrization. The idea of this work is to train the neural network to estimate only the part of prediction which is not already predicted by the physical model. We will combine physical modeling with a neural network to improve the nowcasting of precipitation. The training set needed to train the neural net will be produced by a data assimilation scheme.
Doctorant.e: Filoche Arthur