PhD defence of Lianfa Wang

9 July 2024

Improving the confidence of CFD results by deep learning

Lianfa Wang conducted his PhD work under the supervision of Youssef Mesri. He will defend his PhD in Computational Mathematics, High Performance Computing and Data, on July 18, 2024 in front of the following jury:

M. Fabien CASENAVE Safran Tech | Digital Sciences and Technologies, Rapporteur
M. Guillaume HOUZEAUX Barcelona Supercomputing Center-Centro Nacional de Supercomputación, RapporteurM. Yvan FOURNIER Département MFEE, EDF Lab Chatou, Examinateur
M. Hervé GUILLARD Inria Sophia Antipolis Méditerranée, Examinateur
M. Damien TROMEUR-DERVOUT Claude Bernard University Lyon 1 | UCBL · Institut Camille Jordan, Examinateur
Mme Luisa SILVA RESEARCH INSTITUTE IN CIVIL AND MECHANICAL ENGINEERING (GEM) Centrale Nantes, Examinatrice
M. Charles MOULINEC UK Reasearch and Innovation – Science and Technology Facilities Council, Examinateur
M. Youssef MESRI Cemef Mines Paris-PSL, Examinateur
M. Jean-François WALD Département MFEE, EDF Lab Chatou, Invité

Abstract:

The accuracy of computational fluid dynamics (CFD) simulations of a specific flow phenomenon is highly influenced by the case simulation configuration. In order to help users to better select simulation options and increase trust in obtained results for target flow phenomenon, deep learning algorithms are explored to identify flow phenomena resided on the CFD meshes. Convolutional neural networks (CNN) are employed firstly to identify flow phenomena, such as vortex and thermal stratification, on 2D structured meshes. To overcome the inability of CNN consuming unstructured data, a computational efficient graph neural network (GNN) framework which utilizes U-Net architecture and a successive coarsened graph hierarchy generated from the algebraic multigrid method (AMG) of the open source CFD solver code\_saturne is proposed. After a thorough comparison of the currently available kernel functions in terms of identification accuracy and training efficiency, a fast Gaussian mixture model kernel function is proposed which is better suited for identifying flow phenomena on unstructured CFD meshes. The superiority of the proposed architecture and kernel function is demonstrated by numerical experiments on identifying 2D vortices along with its adaptability to identifying 3D flow features.

A computational efficient graph neural network (GNN) framework which utilizes U-Net architecture and a successive coarsened graph hierarchy generated from the algebraic multigrid method (AMG) of the open source CFD solver code\_saturne is proposed for flow phenomena identification on unstructured CFD meshes

 

Keywords: Flow phenomena identification, Convolutional neural networks, Graph neural networks, Algebraic multigrid method

 

See more related news

Soutenance de thèse de Daniel Irmer
Effect of prior plastic deformation on precipitation in Al-based alloys Doctoral project of Daniel Irmer is in the framework of a cooperation between Centre des Materiaux and CEMEF. Daniel […]
Prix de la meilleur communication orale à Sacha Perocheau
Sacha Pérocheau Arnaud is the recipient of the Best Oral Communication Award at the ‘Polymers2024’ international conference held in Athens from 28 to 31 May 2024. Sacha Pérocheau Arnaud is a […]
A contribution to the material design of polymer-based photovoltaic modules: thermomechanical studies, finite element models and prototyping Dylan Habans conducted his PhD work under the […]
Marc Bernacki médaille Portevin SF2M
We are very proud and thrilled by this new distinction, which rewards the work of a researcher who is totally committed to his missions and passionate about his scientific research […]
Pauline Hahn reçoit le prix Esaform de la meilleure communication
Congratulations Pauline Hahn for this great […]