Deep learning based physical-statistics modeling of ocean dynamics
The modeling of dynamical phenomena in geophysics and climate is based on a deep understanding of the underlying physics, described in the form of PDEs, and on their resolution by numerical models. The ever-increasing number of observations of physical systems, the recent rise of deep learning and t...
Main Author: | |
---|---|
Other Authors: | , , , , |
Format: | Doctoral or Postdoctoral Thesis |
Language: | English |
Published: |
HAL CCSD
2023
|
Subjects: | |
Online Access: | https://theses.hal.science/tel-04166816 https://theses.hal.science/tel-04166816/document https://theses.hal.science/tel-04166816/file/DECHELLE_MARQUET_Marie_these_2023.pdf |
id |
ftsorbonneuniv:oai:HAL:tel-04166816v1 |
---|---|
record_format |
openpolar |
spelling |
ftsorbonneuniv:oai:HAL:tel-04166816v1 2024-09-15T18:24:29+00:00 Deep learning based physical-statistics modeling of ocean dynamics Modélisation physico-statistique de la dynamique des océans basée sur l'apprentissage profond Déchelle-Marquet, Marie Institut des Systèmes Intelligents et de Robotique (ISIR) Sorbonne Université (SU)-Centre National de la Recherche Scientifique (CNRS) Sorbonne Université Patrick Gallinari Marina Lévy 2023-05-10 https://theses.hal.science/tel-04166816 https://theses.hal.science/tel-04166816/document https://theses.hal.science/tel-04166816/file/DECHELLE_MARQUET_Marie_these_2023.pdf en eng HAL CCSD NNT: 2023SORUS170 tel-04166816 https://theses.hal.science/tel-04166816 https://theses.hal.science/tel-04166816/document https://theses.hal.science/tel-04166816/file/DECHELLE_MARQUET_Marie_these_2023.pdf info:eu-repo/semantics/OpenAccess https://theses.hal.science/tel-04166816 Machine Learning [cs.LG]. Sorbonne Université, 2023. English. ⟨NNT : 2023SORUS170⟩ Neural networks Physics-guided Hybrid modeling Apprentissage profond Réseaux de neurones Modélisation physique Modèles hybrides Données océanographiques Données réelles [INFO.INFO-LG]Computer Science [cs]/Machine Learning [cs.LG] [INFO.INFO-NE]Computer Science [cs]/Neural and Evolutionary Computing [cs.NE] [PHYS.PHYS.PHYS-AO-PH]Physics [physics]/Physics [physics]/Atmospheric and Oceanic Physics [physics.ao-ph] info:eu-repo/semantics/doctoralThesis Theses 2023 ftsorbonneuniv 2024-07-25T23:47:41Z The modeling of dynamical phenomena in geophysics and climate is based on a deep understanding of the underlying physics, described in the form of PDEs, and on their resolution by numerical models. The ever-increasing number of observations of physical systems, the recent rise of deep learning and the huge computational power required by numerical solvers, which hinders the resolution of existing models, suggest that the future of physical models could be data-driven. But for this prognosis to come true, deep learning must tackle several challenges, such as the interpretability and physical consistency of deep models, still largely under-addressed by the deep learning community.In this thesis, we address both challenges: we study the prediction of sea surface temperature (SST) using hybrid models combining a data-driven and a physical model. Ensuring the physical plausibility of hybrid models necessitates well-posing their learning: otherwise, the high versatility of neural networks may lead the data-driven part to bypass the physical part.Our study is divided into two parts: a theoretical study on hybrid models, and a practical confrontation of our model on simulations of real data. First, we propose a new generic well- posed learning framework based on the optimization of an upper-bound of a prediction error. Second, we study real-like ocean observations of SST and velocity fields from the Gulf Stream current in the North Atlantic (from the NATL60 model). This application highlights the challenges raised by confronting physics aware learning to the complexity of real-world physics. It also raises issues such as model generalization, which we discuss as a possible perspective. La modélisation des phénomènes dynamiques en géophysique repose sur une compréhension de la physique sous-jacente, décrite sous la forme d'EDP, et sur leur résolution par des modèles numériques. Le nombre croissant d'observations de systèmes physiques, l'essor récent de l'apprentissage profond et l'énorme puissance de calcul requise par ... Doctoral or Postdoctoral Thesis North Atlantic HAL Sorbonne Université |
institution |
Open Polar |
collection |
HAL Sorbonne Université |
op_collection_id |
ftsorbonneuniv |
language |
English |
topic |
Neural networks Physics-guided Hybrid modeling Apprentissage profond Réseaux de neurones Modélisation physique Modèles hybrides Données océanographiques Données réelles [INFO.INFO-LG]Computer Science [cs]/Machine Learning [cs.LG] [INFO.INFO-NE]Computer Science [cs]/Neural and Evolutionary Computing [cs.NE] [PHYS.PHYS.PHYS-AO-PH]Physics [physics]/Physics [physics]/Atmospheric and Oceanic Physics [physics.ao-ph] |
spellingShingle |
Neural networks Physics-guided Hybrid modeling Apprentissage profond Réseaux de neurones Modélisation physique Modèles hybrides Données océanographiques Données réelles [INFO.INFO-LG]Computer Science [cs]/Machine Learning [cs.LG] [INFO.INFO-NE]Computer Science [cs]/Neural and Evolutionary Computing [cs.NE] [PHYS.PHYS.PHYS-AO-PH]Physics [physics]/Physics [physics]/Atmospheric and Oceanic Physics [physics.ao-ph] Déchelle-Marquet, Marie Deep learning based physical-statistics modeling of ocean dynamics |
topic_facet |
Neural networks Physics-guided Hybrid modeling Apprentissage profond Réseaux de neurones Modélisation physique Modèles hybrides Données océanographiques Données réelles [INFO.INFO-LG]Computer Science [cs]/Machine Learning [cs.LG] [INFO.INFO-NE]Computer Science [cs]/Neural and Evolutionary Computing [cs.NE] [PHYS.PHYS.PHYS-AO-PH]Physics [physics]/Physics [physics]/Atmospheric and Oceanic Physics [physics.ao-ph] |
description |
The modeling of dynamical phenomena in geophysics and climate is based on a deep understanding of the underlying physics, described in the form of PDEs, and on their resolution by numerical models. The ever-increasing number of observations of physical systems, the recent rise of deep learning and the huge computational power required by numerical solvers, which hinders the resolution of existing models, suggest that the future of physical models could be data-driven. But for this prognosis to come true, deep learning must tackle several challenges, such as the interpretability and physical consistency of deep models, still largely under-addressed by the deep learning community.In this thesis, we address both challenges: we study the prediction of sea surface temperature (SST) using hybrid models combining a data-driven and a physical model. Ensuring the physical plausibility of hybrid models necessitates well-posing their learning: otherwise, the high versatility of neural networks may lead the data-driven part to bypass the physical part.Our study is divided into two parts: a theoretical study on hybrid models, and a practical confrontation of our model on simulations of real data. First, we propose a new generic well- posed learning framework based on the optimization of an upper-bound of a prediction error. Second, we study real-like ocean observations of SST and velocity fields from the Gulf Stream current in the North Atlantic (from the NATL60 model). This application highlights the challenges raised by confronting physics aware learning to the complexity of real-world physics. It also raises issues such as model generalization, which we discuss as a possible perspective. La modélisation des phénomènes dynamiques en géophysique repose sur une compréhension de la physique sous-jacente, décrite sous la forme d'EDP, et sur leur résolution par des modèles numériques. Le nombre croissant d'observations de systèmes physiques, l'essor récent de l'apprentissage profond et l'énorme puissance de calcul requise par ... |
author2 |
Institut des Systèmes Intelligents et de Robotique (ISIR) Sorbonne Université (SU)-Centre National de la Recherche Scientifique (CNRS) Sorbonne Université Patrick Gallinari Marina Lévy |
format |
Doctoral or Postdoctoral Thesis |
author |
Déchelle-Marquet, Marie |
author_facet |
Déchelle-Marquet, Marie |
author_sort |
Déchelle-Marquet, Marie |
title |
Deep learning based physical-statistics modeling of ocean dynamics |
title_short |
Deep learning based physical-statistics modeling of ocean dynamics |
title_full |
Deep learning based physical-statistics modeling of ocean dynamics |
title_fullStr |
Deep learning based physical-statistics modeling of ocean dynamics |
title_full_unstemmed |
Deep learning based physical-statistics modeling of ocean dynamics |
title_sort |
deep learning based physical-statistics modeling of ocean dynamics |
publisher |
HAL CCSD |
publishDate |
2023 |
url |
https://theses.hal.science/tel-04166816 https://theses.hal.science/tel-04166816/document https://theses.hal.science/tel-04166816/file/DECHELLE_MARQUET_Marie_these_2023.pdf |
genre |
North Atlantic |
genre_facet |
North Atlantic |
op_source |
https://theses.hal.science/tel-04166816 Machine Learning [cs.LG]. Sorbonne Université, 2023. English. ⟨NNT : 2023SORUS170⟩ |
op_relation |
NNT: 2023SORUS170 tel-04166816 https://theses.hal.science/tel-04166816 https://theses.hal.science/tel-04166816/document https://theses.hal.science/tel-04166816/file/DECHELLE_MARQUET_Marie_these_2023.pdf |
op_rights |
info:eu-repo/semantics/OpenAccess |
_version_ |
1810464844648808448 |