Cerfacs Enter the world of high performance ...

PhD Defense : Victor XING : “Deep learning for subgrid-scale modeling in large eddy simulations of turbulent premixed combustion”

  Friday 18 November 2022 at 14h00

  Phd Thesis       Conference room - CERFACS - Toulouse    

Link youtube : https://youtu.be/IaxXKVZ2cPM

Abstract :

In a century defined by climate change and data abundance, combustion is moving towards new opportunities created by the digital revolution. Large eddy simulations (LES) of full-scale practical combustion systems are becoming tractable, but their predictive power hinges on the accuracy of the subgrid-scale (SGS) models that account for unresolved turbulent combustion physics. Deep learning (DL) has recently been used to train data-driven SGS models that achieve excellent accuracy in a priori tests. Yet, there are still hardly any applications of DL SGS models in LES of practical combustion systems. This work investigates three elements that must be addressed to enable the adoption of deep learning in practical LES of turbulent premixed combustion: evaluating DL models on high Reynolds test cases, ensuring their ability to generalize beyond their training configuration, and implementing a computationally efficient integration of DL models to high-performance LES solvers. Three DL models that gradually include each of these elements are developed. They are based on U-Net convolutional neural networks (CNNs) trained on downsampled filtered snapshots of direct numerical simulations. First, a model for the total flame surface density is trained on the R2 high Reynolds turbulent jet flame. Excellent a priori generalization to higher Reynolds numbers and to LES snapshots is observed, and insights on the inner workings of the model are provided. Second, a CNN model for the SGS variance of the progress variable is trained on a statistically planar turbulent flame. With a Pfitzner source term formulation and a beta probability density function closure, it is able to accurately predict a priori the SGS variance and the filtered reaction rate on the R2 jet flame, thus demonstrating its ability to generalize to new configurations. Third, the AVBP-DL coupling strategy is developed to enable DL models to be queried by the AVBP solver with negligible computational overhead. Finally, the Masri vented obstructed explosion test case is used to test a posteriori a CNN model for the SGS wrinkling factor trained on the statistically planar turbulent flame. The model predicts the right peak overpressure, but this results from a compensation between excessive wrinkling in the initial laminar phase and insufficient wrinkling in the critical turbulent propagation stage. Several attempts to correct this behavior are then explored.

Jury :
Denis VEYNANTE – Research director – EM2C – Referee

Antonio ATTILI – Assistant Professor – Univ. Edinburgh – Referee

Michael PFITZNER – Univ.-Prof. Dr. rer. nat. – UniBw – Examiner

Pascale DOMINGO – Research director – CORIA  – Examiner

Corentin LAPEYRE –  Senior researcher – CERFACS – Advisor

Thierry POINSOT – Research director – IMFT – Advisor






Artificial Intelligence for computational physics

From Monday 4 December 2023 to Friday 8 December 2023






Multi-architecture parallelism using Kokkos/C++ library

From Monday 11 December 2023 at 14h00 to Wednesday 13 December 2023 at 17h00






PhD Defense : Aurélien LINÉ : ” Modulation of European near-term climate change by multi-decadal internal variability “

Thursday 21 December 2023 at 15h00

  CERFACS - Toulouse - France     Organized by Nathalie BROUSSET