Offre d’emploi :
Poste de chercheur post-doctorant (H / F) au Laboratoire de LMFA.
Type de recrutement
Type de poste : Formation requise :
Doctorat en mécanique des fluides
CDD Post-doctorat
Durée du contrat : 24 mois
Rémunération :
Selon expérience, sur la base de la grille de
rémunération en vigueur à l’Université de Lyon
Date de prise de fonction : A partir de Janvier 2022
Le labex iMUST vise à encourager une recherche pluridisciplinaire multi échelle, pour aborder et résoudre des problèmes complexes en science des matériaux et des technologies écologiquement viables. Il réunit des compétences et des connaissances développées dans les domaines de la Physique, de la Chimie et de l’Ingénierie.
Le Labex rassemble des chercheurs et enseignants chercheurs dans 21 unités de recherche sur le site de l’Université de Lyon
DESCRIPTION DU POSTE
Positionnement du poste :
Dans le cadre du labex iMUST, un poste postdoctoral est disponible au Laboratoire LMFA.Organisation du travail :
temps completLieu de travail :
Campus Lyon Tech La DouaDépartement Composante Mécanique Bâtiment Oméga
43 Boulevard du 11 novembre 1918 69622 VILLEURBANNE Cedex
Contraintes particulières de travail :
aucuneMission :
Title: Large-scale modeling of liquid sheet atomization using Convolutional Neural Networks.
Presentation of the post-doctoral project, context and objectives:
In the aeronautical combustion chambers, fuel is injected and pulverized in a set of mechanisms known as atomization. This phenomenon involves the break-up of the liquid jet, sometimes in the form of a thin sheet, into a multitude of smaller structures, up to the formation of a spray. The numerical simulation of atomization usually severely challenges even the state-of-the-art numerical methods, mainly in reason of its intrinsic multi-scale aspect (Figure 1).
Figure 1 : Liquid sheet atomization. Breakup into ligaments and droplets.
This project aims at pursuing the development of a multi-scale numerical methodology able to efficiently simulate both the liquid injection and the spray in the same unsteady simulation [1]. This is achieved by dynamically coupling two models on the same simulation: a separated two-phase flow solver (dedicated to the description of the large scales of the sheet and its stretching into ligaments) and a dispersed two-phase flow solver (dedicated to the Lagrangian description of the droplets forming a spray and resulting from ligaments breakup). This coupling allows an optimal resolution of the atomization process at large scales (LES) (Figure 2).
Figure 2 : Multiscale methodology for atomization simulation. Coupling between separated (green sheet) and dispersed (blue droplets) two-phase flow solvers. Reprinted
from [1].
One drawback of this approach is that the space scale transition between the separated and the dispersed two-phase flow solvers leads to a loss of information about the liquid topology. This results in the failure to predict the particle size distribution of the generated droplets during the coupling. This information is of paramount importance for the simulation of the spray evolution. One possible solution is to use subgrid scale modeling for under-resolved liquid structures such as ligaments and smaller droplets in the separated two-phase flow solver. The postdoc project is part of this context and aims at using "deep learning" techniques (artificial intelligence) to enrich the description of the thin liquid structures computed in the separated two-phase flow solver in the framework of large eddy simulation (LES). The project is based on the works proposed in [2] and [3].
In a first step, the training database of the neural network will be obtained from high fidelity direct numerical simulations (DNS) of a liquid sheet atomization in simplified configurations. The simulation tool DYJEAT of ONERA will be used [4] [5]. The liquid volume fraction and the sheet interfacial area density Σ fields will then be extracted.
In a second step, a convolutional neural network (CNN) will be trained from the high fidelity fields to allow the reconstruction of Σ from ( being the only output variable accessible during a LES computation). The Keras API of TensorFlow will be used to create and train deep learning models.
Finally, at the end of the learning step, the neural network will be able to reconstruct on the fly a subgrid information, namely the interfacial area density Σ, from the volume fraction field in the framework of a LES simulation. The ratio /Σ will allow to define locally a droplet diameter D32 (Sauter diameter) characterizing the spray generated by the atomization process. The question of the optimal integration of a neural network in a CFD HPC simulation will also be addressed.
The expected work will be part of each of the three steps described hereinabove namely the DNS of a liquid sheet atomization as a training database, then the learning step for the neural network and finally the use of the network for subgrid scale modeling in the framework of LES of the atomization process.
Bibliography
[1] G. Blanchard, «Modélisation et simulation multi-échelles de l'atomisation d'une nappe liquide cisaillée,» Thèse de doctorat de l'Université de Toulouse, 2014.
[2] C. Lapeyre, A. Misdariis, N. Cazard, D. Veynante et T. Poinsot, «Training convolutional neural networks to estimate turbulent sub-grid scale reaction rates,» Combustion and Flame, vol. 203, pp. 255-264, 2019.
[3] J. Temple-Boyer, «Application de techniques de Deep-Learning pour la modélisation de l’atomisation,» Rapport de stage de fin d'étude, 2020.
[4] D. Zuzio, A. Orazzo, J. Estivalezes et I. Lagrange, «A new efficient momentum preserving Level-Set/VOF method for high density and momentum ratio incompressible two-phase flows.,» J. Comput. Phys., vol. 410, 2020.
[5] T. Xavier, D. Zuzio, M. Averseng et J. Estivalezes, «Toward direct numerical simulation of high speed droplet impact.,» Meccanica, vol. 55, pp. 387-401, 2020.
PROFIL RECHERCHE
Savoirs : Computational fluid dynamics, two-phase flows, artificial intelligence (neural networks).
Savoir-faire : programming (fortran, python), tools for AI (TensorFlow, Keras, PyTorch), writing (LaTeX).
Savoir être : Good written and oral communication, collaborative teamwork.
CANDIDATURES
Renseignements sur le poste :
Mail/Tél. [email protected]
Envoi des candidatures :
Lettre de motivation + CV exclusivement par e-mail à [email protected]