Online Learning of Generative AI models for turbulent flow inpainting

Il y a 9 heures

Toulouse, Occitania, France Groupe Calcul Temps plein
GROUPE CALCUL
• Le groupe Présentation Le Réseau métier Le RT
• Paysage du calcul
• Actions RT
• Actions réseau
• Nos cafés
• Offres d'emploi
• Suivez-nous Offres d'emploi Online Learning of Generative AI models for turbulent flow inpainting Proposée par Luciano DROZDA, le 16 novembre 2026 Stage Toulouse, France CERFACS 6 mois Email Web Fichier Suivez la liste Context The Algo team at CERFACS has conducted research in hybridizing solvers of Partial Differential Equations (PDEs) with AI models in order to counter theoretical limitations of classical PDE solvers (e.g., regular sampling of space and time domains, solution regularity) towards practical applications of relevance to Cerfacs industrial shareholders (see [1] and [2], for instance). Beyond hybridization, AI-based surrogate modeling presents itself as a much less expensive potential alternative to numerical simulations of interest to Cerfacs industrial shareholders (e.g., turbomachinery flows, flows over airfoils). The main technical challenge to training these AI-based surrogate models is the prohibitive high-quality training data storage requirements. For instance, in [3], an industrial Direct Numerical Simulation (DNS) of compressible flow over a low-pressure turbine blade required over 11 billion degrees of freedom (DoFs) per equation. One can then estimate the size of a dataset composed of 1,000 snapshots of such a DNS to require around 1.1e10 (number of mesh points) x 5 (number of conserved variables per point) x 8 (number of bytes per DoF considering double precision) x 1,000 (number of snapshots)  440 TB for a single operational point of this blade. To circumvent the issue of prohibitive storage requirements, a training procedure referred to as online learning can be employed, where training data is generated on-the-fly as several numerical simulations run concurrently (see [4]). In this context, we aim to develop Generative AI models for turbulent flow inpainting via online learning. Inpainting consists here in filling in masked flow fields output by running instances of high-fidelity physical solvers conditioned by the rest of the flow. It allows to train AI models on large volumes of data in a self-supervised way (i.e., without the need of human labeling). In [5], a Variational Auto-Encoder (VAE) was (pre-)trained to reconstruct parcels of physical fields stemming from a myriad of PDE solvers via online learning. Reconstruction distinguishes itself from inpainting through the lack of physical field conditioning in the former. Given that turbulent flow modeling essentially revolves around predicting the effects of unresolved scales conditioned by coarse-grained information, inpainting presents itself as a (pre-)training task more suited towards practical applications of AI in Computational Fluid Dynamics (CFD) such as subgrid-scale and wall modeling, super-resolution, and sparse-sensor reconstruction. References [1] Drozda, L., Mohanamuraly, P., Cheng, L., Lapeyre, C., Daviller, G., Realpe, Y., Adler, A., Staffelbach, G., & Poinsot, T. (2023). Learning an optimised stable Taylor-Galerkin convection scheme based on a local spectral model for the numerical error dynamics. Journal of Computational Physics, 493, 112430. https://doi.org/10.1016/j.jcp.2023.112430 [2] Gonzalez, F., Larroque, A., Theret, A., Drozda, L., Odier, N., Legaux, J., & Teste, O. (2026, March 4-6). Scalable graph network for robust sub-grid scale modelling for compressible flows [Conference abstract]. 3rd ERCOFTAC Workshop on Machine Learning for Fluid Dynamics, Amsterdam, Netherlands. [3] Iyer, A. S., Abe, Y., Vermeire, B. C., Bechlars, P., Baier, R. D., Jameson, A., Witherden, F. D., & Vincent, P. E. (2021). High-order accurate direct numerical simulation of flow over a MTU-T161 low pressure turbine blade. Computers & Fluids, 226, 104989. https://doi.org/10.1016/j.compfluid.2021.104989 [4] Schouler, M., Caulk, R. A., Meyer, L., Terraz, T., Conrads, C., Friedemann, S., Agarwal, A., Baldonado, J. M., Pogodziński, B., Sekuła, A., Ribes, A., & Raffin, B. (2023). Melissa: Coordinating large-scale ensemble runs for deep learning and sensitivity analyses. Journal of Open Source Software, 8(86), 5291. https://doi.org/10.21105/joss.05291 [5] Liu, Q., Koehler, F., Holzschuh, B., & Thuerey, N. (2026). Tadpole: Autoencoders as foundation models for 3d pdes with online learning. arXiv. https://doi.org/10.48550/ARXIV.2605.15284 Mission The goal of this internship is to explore the potential of Generative AI architectures to perform turbulent flow inpainting. We will focus on flow matching techniques and backbone architectures suited for handling unstructured data such as Graph Neural Networks (GNNs) and Transformers. We will target statistically converged flow cases to avoid distribution shift problems. These cases are: forced Homogeneous Isotropic Turbulence (HIT), turbulent channel flow. Numerical experiments will be conducted using the Julia programming language. The AI models and the physical solvers employed to generate training data will be differentiated and/or ported to run on Graphics Processing Units (GPUs) using STADE.jl (https://github.com/cerfacs/STADE.jl), the Cerfacs Automatic Differentiation engine based on source transformation. Desired Profile
• Final-year engineering school student or a Master 2 (M2) student
• Knowledge of fundamentals of Fluid Dynamics is required
• Knowledge of fundamentals of Deep Learning is required
• Experience with parallel programming (e.g., MPI, CUDA) is a plus
• Scientific proactivity and rigor are required
• Professional working proficiency in English is required for report writing and work presentation

How to apply
? To apply, please send your CV and covering letter to drozda@cerfacs.fr, applications are open until 31/12/2026. See you soon at CERFACS
• Suivez-nous
• Contactez-nous
• Mentions légales Développé avec Pelican / Creative commons Graphisme par pierredigonnet.com