Olivier Gicquel, Ronan Vicquelin, Jean Taine Phd Students : Y. Zhang, J. Kim, G. Wang, R. Goncalves dos Santos

Transcription

1 Développement d un code Monte- Carlo pour le calcul intensif utilisant les spécificités de la méthode ERM Olivier Gicquel, Ronan Vicquelin, Jean Taine Phd Students : Y. Zhang, J. Kim, G. Wang, R. Goncalves dos Santos!

2 RANS versus LES simulations 2 Experimental data (turbulent premixed propane / air V-shape flame) Flame visualisation using PLIF on OH radical Extracted flame front Knikker et al. (2000, 2004, 2006)

3 RANS versus LES simulations RANS framework Instantaneous flame front Mean flame front (RANS) 3 P(c) β c α Mean temperature = probability to be in burnt gases! time

4 RANS versus LES simulations RANS framework 4 Mean flame front Optical path α P(c) β Averaging of radiative heat transfers along instantaneous optical paths Models for cross-correlations!

5 RANS versus LES simulations LES framework 5 Local (weighted) average over a small volume Fresh and burnt gas locations identified at the resolved scale level Instantaneous filtered flame front (LES)

6 A dedicated Monte Carlo solver Classical Monte Carlo approach 6

7 A dedicated Monte Carlo solver Classical Monte Carlo approach 7 Can not estimate the radiative power in a single point of the domain without performing a full simulation The convergence is a global converge The convergence is hard to optimize Problem of load balancing in massively large simulation Need a large amount of memory

8 A dedicated Monte Carlo solver Reciprocal Monte Carlo approach ERM (L. Tessé et al) 8

9 A dedicated Monte Carlo solver Reciprocal Monte Carlo approach ERM 9 Estimation of the radiative power in a single point of the domain without performing a full simulation Possibility of a local convergence Very good scalability and load balancing Need only few memory Very accurate when there are large isothermal regions

10 Few words about parallelisation Monte Carlo code structure 10

11 Speed up Few words about parallelisation Monte Carlo code structure cpu_time nb_proc Nb_proc

12 Few words about parallelisation Monte Carlo code structure % task assignment analysis of Rainier code 90% 80% 70% 60% 50% 40% 30% 20% overhead iprobe gather bcast barrier irecv isend wait init Comput 10% 0%

13 Few words about parallelisation Hybrid parallelisation 13

14 Few words about parallelisation Hybrid parallelisation 14 14

15 Few words about parallelisation Hybrid parallelisation 15

16 Load balancing with GPU 16

17 Optimization of the local number of realization Typical results 17

18 Comparison with DOM 18 radiative power at x = 0.1m radiative power at x = 0.15m! DOM MC! DOM MC diff = 5.1 % diff = 7.7 % Mesh : 3.4million cells Method Nb_procs optical paths spectral bands cpu time memory Domasium min 2G monte carlo _max min 0.48G domasium_bis min > 2G

19 Monte-Carlo optimisation (Level of importance) Avoiding the draw-backs of ERM 19 Gas : K - Wall : 500 K

20 Monte-Carlo optimisation (Level of importance) Avoiding the draw-backs of ERM 20 Gas : 500 K - Wall : K

21 Monte-Carlo optimisation (Level of importance) Avoiding the draw-backs of ERM 21

22 Monte-Carlo optimisation (Level of importance) Avoiding the draw-backs of ERM 22 Gas : K - Wall : 500 K

23 Monte-Carlo optimisation (Level of importance) Avoiding the draw-backs of ERM 23 Gas : 500 K - Wall : K

24 Fully coupled AVBP - Monte Carlo Simulation Texte du sous-titre 24

25 Final conclusion 25 LES is a good framework to investigate Combustion - Radiation interaction Fully coupled simulations are available on gaz turbine geometries Dedicated solver are mandatory Radiation can change the conductive heat fluxes Crucial need of experimental setups dedicated for both combustion and radiation! Accurate radiative boundary conditions

26 15th International Conference on Numerical Combustion Palais des Papes, Avignon, France April, Organized by EM2C-CNRS and Ecole Centrale Paris (Dead line 31 Oct) Copyright Yann de Fareins / Noir d Ivoire