- **Main CFD Forum**
(*http://www.cfd-online.com/Forums/main/*)

- - **Reply to Tim Re Dynamic Smagorinsky model
**
(*http://www.cfd-online.com/Forums/main/4790-reply-tim-re-dynamic-smagorinsky-model.html*)

Reply to Tim Re Dynamic Smagorinsky model
Hi ALL.. I have worked on Dynamic subgrid stress model for the flow past square cylinder at high reynolds number. I also compared our results with experimental results of Lyn et al. and also with simulated result of SP Vanka. We are getting good agreement... if you need any help then plz let me know.
In case of dynamic model the value of C is very critical. So You can apply different averaging technique to get smooth value of C like 5 point aver or 9 point... near wall you have to apply less point averaging tech. Also check your grid spacing near the wall. It should be less that some critical value. Also you have to take care while applying test filtering. plane of test filter should be homogenous which is very rare in case of turbulent flow. Also the denominater part of formula which calculate C is very unstable. sometime it becomes 0/0 case. so avoid such situations. I will be happy if you want to discuss anything about this topic. with best regards Ajay ~ |

Re: Reply to Tim Re Dynamic Smagorinsky model
Questions
1. If you have published papers on this problem, could you send them to me via e-mail ? 2. Computational domain size/number of meshes(x,y,z directions) ? 3. Solver ? 4. CPU time ? 5. What kinds of properties did you compare with Lyn's results ? Standard Smagorinsky model gives reasonable results in this problem. When I compared my data with Lyn's data, both time averaged and phase averaged data, it seems reasonable. Due to the recovery rate difference, not easy to compare my data with Lyn's data in the near wake region (around x/D>5). But in the shear layer formation and base region, my data seems to be o.k. What are you going to see from your data? We can show something new phenomena from numerical data. Just simple comparison with Lyn's data is not enough. Because many other people published on that topic. For example, Sohankar, Murakami,... showed time averaged properties; drag, lift coefficient, Strouhal number, turbulent kinetic energy distribution etc. Lyn's data could not show enough resolution and lateral component of velocity close to the cylinder surface. We can show accurately these properties. Jongdae Kim. |

Re: Reply to Tim Re Dynamic Smagorinsky model
1. We have just communicated not published so its not possible for me to send the script of the paper.
2 22.0 in X-dir 10 in Z-dir and 6 in y-dir (non-dimensional). grid size 178x22x80 3. We have used MAC algorithim and for turbulence modeling Dynamic subgrid scale stress model suggested by Peomilli and Liu. 4.we run our code on SUN-60 and in IBM SP. for 40k iteration it took 20 to 24 days. 5. We computed time avergaed flow field and compared time-averaged velocities (all three components) and time averaged turbulent stresses at different downstream location also recovery rate at centerline.Coefficient of pressure Cp has been calculated over the cylinder surface.moreover with phase averaged data we found random and periodic components of fluctuations. Looking forward to see your comments and further ideas. Ajay |

Re: Reply to Tim Re Dynamic Smagorinsky model
Comments;
2. 22.0 in X-dir 10 in Z-dir and 6 in y-dir (non-dimensional). grid size 178x22x80 ; Maybe 16 in y-dir.(not 6, right?). What is the smallest grid size and how big is the cell size compate to Kolmogorov length scale? 22 cells in z-dir means the size of one cell size in z-dir. is 0.5D. If you want to get somewhat meaningful streamwise structures, I recommend 0.1D. Acutally I'm running laminar and turbulent cases now (20D in x-dir., 14D in y-dir., 2D in z-dir. is the base case). With 0.1D in z-dir. cell, 40 cells are used at Re=22000, and 40 and 80 cells are used at Re=45, 150, 245. 3. We have used MAC algorithim and for turbulence modeling Dynamic subgrid scale stress model suggested by Peomilli and Liu. ; Your methods seems to be similar to my approach except SGS model. I use standard Smagorinsky model with wall function. 4.we run our code on SUN-60 and in IBM SP. for 40k iteration it took 20 to 24 days. ; For phase/time averaged properties, more than 600k iteration (nodimensional time step size = 0.001 or 0.002) which corresponds to 1.8 to 2 seconds ( in physical time). It took about two weeks (of wall clock time) in Cray C90. 5. We computed time avergaed flow field and compared time-averaged velocities (all three components) and time averaged turbulent stresses at different downstream location also recovery rate at centerline.Coefficient of pressure Cp has been calculated over the cylinder surface.moreover with phase averaged data we found random and periodic components of fluctuations. ; Do you have z-dir. velocity component to compare ? What kinds of reference data did you use? How was the recovery rate? Usually LES data is higher than Lyn's data but lower than that of Durao's (1988). The turbulence intensity of Durao's experiment is higher than Lyn's and mixing in downstream is higher than Lyn's flow and LES (I use uniform inflow condition which means zero turb. intensity). I'd like to know what kind of boundary conditions you implemented. Could you give me your time averaged results? In my case, St=0.134, Cd=2.065, root mean square of Cl=1.268. As you know, with some coarser grid, LES produce reasonable results. Why? It's an interesting topic to discuss in the future. One more thing to suggest. Why don't you run your code in laminar flow regime. Actually I could not get the similar number of mean Drag Coefficient at Re=100~200 compared to Sohankar's results. I increased mesh resolution but failed to get the number. One idea is to increase computational domain such as 30*20*8 (x,y,z directin). But I don't have computer resource to run such big cases. This is future topic of my study. Jongdae Kim. |

Re: Reply to Tim Re Dynamic Smagorinsky model
Hi,
Can you tell me more about your averaging techniques. As I understand things averaging is done to stop the denomitor approaching zero. This averaging must be done over a statistically homogenous plane or line. In a channel case this is fairly simple - you average parallel to the walls. However in a more complex case this is not possible. How do you find a homogenous direction. I am considering trying to average along a stream line - do you think this will work. I assume you use a structured serial code -is this correct ~( I have an unstructured parallel code which makes thing more difficult). thanks for your help. Tim P.S. when you publish your paper can you let us know? |

Re: Reply to Tim Re Dynamic Smagorinsky model
Sure. I'll let you know. You can read one of my conference papers (PDF files) in my homepage.(http://publish.uwo.ca/~jdkim).
As you know streamline changes depending on the coordinate system. If the coord. sys. is moving with some constant velocity (that is convection velocity), streamlines look different. In averaging, I use time and phase averaging. In time averaging, sometimes I use spatial averaging together in homogeneous direction. Three dimensional turbulent flow around two dimensional square prism is my problem. So I consider that spanwise direction is homogeneous. But in the case of cube which is totally three dimensional problem, I have no idea. Anyway, if the flow itself is very complex, turbulent, in this case, why don't you try standard Smagorinsky model? I understand if you want high accuracy, high order statistical results, Dynamic model is better. But for engineering application (i.e., your computational domain is very complicate, and just mean, rms, reynolds stress etc. are your results to be estimated), you can get reasonable output with reasonable computational costs. Although I don't have experience of using dynamic model but using smagorinsky model, I'm satisfy with my results (time and phase averaged data). Kim. |

Re: Reply to Tim Re Dynamic Smagorinsky model
Hi,
My coordinate system is (thankfully) not moving and I was considering looking at using time averaged stream function as a guide to a homogenous direction. Using the spanwise direction makes a lot of sense. However I would ideally like to get a model running that will work on (almost?) any type of geometry. I currently do use the standard Smagorinsky model with van Driest damping but would like to improve on it . Thanks, Tim. |

Re: Reply to Tim Re Dynamic Smagorinsky model
Could you describe what is your computational domain and boundary/initial conditions and numerical scheme?
Kim. |

Re: Reply to Tim Re Dynamic Smagorinsky model
My geometry is 3-D Flow past a square cylinder in a channel. So i am taking X-Y plane(along the cylinder height) as homogenous plane. I think you are talking about averaging of C not time averaging of flow field. So we have used 5 point and 9 point spatial averaging for C. What i will suggest while calculating Leonard term and (Alpha)ij and (Beta)ij take care of boundary conditions for them. Otherwise it makes C unstable if it does computations on some imaginary cells.
And i am sure you will get better accuracy by using Dynamic model instead of Smagorinsky. In my work i also compared Smagorinsky's results for different value of C with the dynamic model and DNS. Implimenting Smagorinsky is very easy but if u want to go for accuracy then you have to look on dynamic model. If you want to apply test filtering in all direction ( if u suppose that no plane is homogenous) then look at Ghosal et al JFM paper. |

Re: Reply to Tim Re Dynamic Smagorinsky model
Thank you for your suggestions. I'll consider your idea for my future work(Fluid-structure interaction).
Kim. |

All times are GMT -4. The time now is 15:30. |