DESCRIPTION :
Augmenting physical models with neural networks is an important area of research [1], it allows to improve the forecasting of physical systems when only partial knowledge of their dynamics is known. In this setting, the system is decomposed into a physical part, driven by the known equation(s) of the physical system, and a learned part, where parameters are chosen in a supervised manner, by comparing with observations or outputs of higher fidelity models. This approach has the potential to improve our physical models and increase their efficiency, providing an alternative to physical parameterisations that attempt to approximate unresolved phenomena. Large-scale approaches that successfully combine dynamical models with deep learning, such as the Neural GCM [2], have had a very important impact on the weather and climate prediction community and seem very promising for ocean models.
Most of the existing works [3, 4, 5] have trained these parametrisations in an "offline" way, training the network to compensate the error of the physical model independently at each step. Even if this approach has shown encouraging results, it still has the effect of making the system unstable at test time, causing the simulation to stop after a few numerical integration steps. Several papers [6, 7, 8] have investigated this problem and shown that training parametrisations "online" has a major positive impact on the results, mostly mitigating stability issues. However, training parametrisations online is a major challenge, as the backpropagation through time can be unstable and requires large computational and memory load during training. Several strategies have been proposed to mitigate these difficulties, and the deep learning literature is rich in approaches to deal with the problem associated with long-term backpropagation through time. With the emergence
of ocean models implemented in modern auto-differentiable languages [9, 10], applying these kinds of approaches to ocean models becomes possible and seems like a great opportunity.
The aim of the postdoc is to benchmark different online learning approaches in simplified Navier-Stokes systems (shallow water, quasi geostrophic models) and to explore different solutions to improve their results and reduce their training computational cost. It is also expected to compare these methods with the diffusion-based autogressive models presented in [11], which do not rely on physical models. By providing a benchmark and a set of techniques to facilitate the training of these approaches, the aim is to pave the way for the application of such approaches in larger, more complex ocean general circulation models.
[1] Yin, et. al. "Augmenting Physical Models with Deep Networks for Complex Dynamics Forecasting". Journal of Statistical Mechanics: Theory and Experiment - 2021
[2] Kochkov, et. al. "Neural General Circulation Models for Weather and Climate". Nature - 2024
[3] Zanna, et. al. "Data-Driven Equation Discovery of Ocean Mesoscale Closures". Geophysical Research Letters 47 - 2020
[4] Ross, et. al. "Benchmarking of Machine Learning Ocean Subgrid Parameterizations in an Idealized Model". Journal of Advances in Modeling Earth Systems 15 - 2023
[5] Pedersen, et. al. "Reliable Coarse-Grained Turbulent Simulations through Combined Offline Learning and Neural Emulation". arXiv - 2023
[6] List, et. al. "How Temporal Unrolling Supports Neural Physics Simulators". arXiv - 2024
[7] Ramadhan, et. al. "Capturing Missing Physics in Climate Model Parameterizations Using Neural Differential Equations". essoar - 2022.
[8] Frezat, et. al. "A Posteriori Learning for Quasi-Geostrophic Turbulence Parametrization". Journal of Advances in Modeling Earth Systems 14 - 2022
[9] Ramadhan, et. al. "Oceananigans.Jl: Fast and Friendly Geophysical Fluid Dynamics on GPUs". Journal of Open Source Software 5 - 2020
[10] Häfner, et. al. "Veros v0.1 - a Fast and Versatile Ocean Simulator in Pure Python". Geoscientific Model Development 11 - 2018
[11] Kohl, et. al. "Benchmarking Autoregressive Conditional Diffusion Models for Turbulent Flow Simulation". arXiv - 2024
Principales activités
Principales activés : étude mathématique d'une hiérarchie de modèles pour l'océanographie, codage d'un coeur numérique associé, application à la simulation océanographique.
Activités complémentaires : participation aux activités régulières de l'équipe (séminaires, groupe de travail, participation occasionnelle à l'encadrement doctoral)
Compétences
Compétences techniques et niveau requis : thèse en calcul scientifique ou dans un domaine voisin
Langues : français, anglais
Compétences relationnelles : capacités de collaborations
Avantages
* Restauration subventionnée
* Transports publics remboursés partiellement
* Congés: 7 semaines de congés annuels + 10 jours de RTT (base temps plein) + possibilité d'autorisations d'absence exceptionnelle (ex : enfants malades, déménagement)
* Possibilité de télétravail et aménagement du temps de travail
* Équipements professionnels à disposition (visioconférence, prêts de matériels informatiques, etc.)
* Prestations sociales, culturelles et sportives (Association de gestion des œuvres sociales d'Inria)
* Accès à la formation professionnelle
* Sécurité sociale
Code d'emploi : Déménageur (h/f)
Domaine professionnel actuel : Employés Chargement et Déchargement
Temps partiel / Temps plein : Plein temps
Type de contrat : Contrat à durée indéterminée (CDI)
Compétences : Réseaux de Neurones Artificiels, Python (Langage de Programmation), Machine Learning, Graphics Processing Unit (GPU), Deep Learning, Free and Open Source Software, Anglais, Convivialité, Analyse Comparative (Benchmark), Équations Différentielles, Expérimentation, Dynamique des Fluides, Elaboration des Prévisions, Géologie, Parameterized, Sciences Physiques, Simulations, Etudes et Statistiques, Gestion des Transports
Courriel :
julien.salomon@inria.fr
Téléphone :
0139635511
Type d'annonceur : Employeur direct
