LEGI - UMR 5519

Financial Support : CNRS 80|PRIME

Collaboration : J. Le Sommer (IGE, Grenoble) et R. Fablet (Lab-STICC, Brest)

Sub-grid scale closures are key ingredients for turbulent flows simulations. Such closures are needed to account for the impact of unresolved fine scale variables over resolved larger scales variables because of the nonlinearity of fluid dynamics (Sagaut et al. 2006). In practice, these models are usually deterministically asserved to large scale resolved quantities and obtained through the combination of theoretical and empirical considerations. Over recent years, progress have been made in LES and RANS modelling by applying machine learning (ML) techniques for calibrating models on the basis of databases of DNS simulations (Kutz 2017, Vollant et al. 2017). Recent advances in ML algorithms and in their software implementation are therefore expected to yield breakthrough in LES modelling over coming years. Convolutional neural networks (CNN, a particular class of neural networks that rely on local convolution operations) are in particular expected to be well adapted to LES modelling, because they naturally encode the filtering operations involved in designing LES closures (Bolton and Zanna et al. 2018) and can be trained over large databases. In this project, we intend to define a new robust modeling strategy to be able to model a large range of sub-grid scales quantities. The project will first focus on the modeling of scalar quantities as needed for various applications (heat transfer, combustion, oceanography, etc). A particular attention will be devoted to environmental fluid flows simulations and to their applications in operational systems, as for instance ocean prediction systems. The ultimate goal will be to take into account the modeling error propagation, and its interaction with numerical errors in the ML techniques to insure the accuracy of the model for long-time evolution.

Publications

Miscellaneous

2022

Frezat, H., Fablet, R., Le Sommer, J., & Balarac, G. (2022). Learning quasi-geostrophic turbulence parametrizations from a posteriori metrics.

  

Currently in Preparation or Submitted

2022

Frezat, H., Le Sommer, J., Fablet, R., Balarac, G., & Lguensat, R. (2022). A posteriori learning of quasi-geostrophic turbulence parametrization: an experiment on integration steps.

  

Peer-reviewed Publications

2022

Frezat, H., Le Sommer, J., Fablet, R., Balarac, G., & Lguensat, R. (2022). A posteriori learning for quasi-geostrophic turbulence parametrization. Journal of Advances in Modeling Earth Systems, , 1–35.

  

2021

Frezat, H., Balarac, G., Sommer, J. L., Fablet, R., & Lguensat, R. (2021). Physical invariance in neural networks for subgrid-scale scalar flux modeling. Physical Review Fluids, 6(2).

  

Ph.D. Theses

2022

Frezat, H. (2022). Learning sub-grid dynamics in idealized turbulent systems. Ph.D. thesis, Université Grenoble Alpes [2020-....], .