Vous êtes ici : GDR - Version française > Annonces > Thèses et Post-docs

Appel à candidatures | Recherche, Emploi

PhD Physics Informed Machine Learning for Turbulence Modeling - Nice

Du 1 septembre 2020 au 31 août 2023

Deadline for application is on May 31st, 2019.
CEMEF MINES ParisTech, Sophia Antipolis

For further information, please contact
J. Bec (jeremie.bec@mines-paristech.fr), or
E. Hachem (elie.hachem@mines-paristech.fr).

The proposed research subject deals with the application of deep machine learning in the context of Large-Eddy Simulations (LES). This approach relies on the idea of solving only the largest scales of the flow and modeling the effect of small scales, either by phenomenological techniques, or by directly estimating the solution residue after filtering (as in the case of multi-scale VMS stabilized variational formulations).

For the details click here.

Computational fluid dynamics relies on massive numerical simulations and generates a considerable amount of data. The rise of machine learning techniques has led to major efforts to accelerate and improve fluid solvers. Until now, these methods have been mainly applied to simple, weakly nonlinear cases, described by a small number of degrees of freedom and therefore generally easy to predict. However, flows of interest to applications are most of the time turbulent. They present intermittent multiscale phenomena (in space and time), which are of high dimensionality, and for which statistical fluctuations play a primordial role. The development of reliable and effective models able to cope with such situations remains a real challenge.
The proposed research subject deals with the application of deep machine learning in the context of Large-Eddy Simulations (LES). This approach relies on the idea of solving only the largest scales of the flow and modeling the effect of small scales, either by phenomenological techniques, or by directly estimating the solution residue after filtering (as in the case of multi-scale VMS stabilized variational formulations). The current explosion of data science represents a unique opportunity for a shift of paradigm in turbulent modeling. The proposed strategy consists, on the one hand, to detect new physical constraints for model eligibility, and on the other hand, to generate physically informed local convolutional filters to model the small scales. This will lead to the design of innovative large- scale closures, without any a priori, neither on the discretization operators, nor on the relevant physical quantities. The datasets that will be used for learning and testing corresponds to the state of the art and come from very-high-resolution direct numerical simulations by spectral methods of developed turbulent flows.
This project is strongly transdisciplinary (applied maths, physics, fluid dynamics, statistics, data analysis, high- performance computing applied to engineering), at the edge between basic and applied research. The doctoral student will benefit much from this environment and will diversify his/her expertise through strong scientific interactions with team members. This thesis is timely, particularly relevant to applications, and will lead to publications in international journals of the highest level.

Expected profile

  • Master of Science or equivalent in applied mathematics, physics, or mechanical engineering, with competences in fluid dynamics, statistics, or scientific computing
  • Good experience in programming (C, C++) and in data post-processing and analysis
  • Excellent writing skills, fluent in English
  • Rigorous, autonomous, creative and motivated by working at the edge between basic research and industrial applications

Working conditions

The PhD student will be supervised by:
  • Jérémie Bec (CNRS research director), specialist in the physics of turbulent flow
  • Elie Hachem (Mines ParisTech Professor), specialist in computational fluid dynamics