In the last few days, I ran a few simulations both in 2d and 3d and found a few interesting phenomena during the flow calculations under high Reynolds number. With a higher Re, there existed some place needy to be taken care of, such as the value of ac-zeta, different means of anti-aliasing… And I would like to make some discussions following what @WillT shared, please correct me if I was wrong at somewhere.
-
Mesh and order
It’s highly recommended adopting a higher order to solve high Re problem, for you can use a coarser mesh and also accelerate the simulation with P-multigrid. But due to the restriction of yplus in ILES, the mesh resolution shall be determined corresponding to the number of orders you plan to use. Usually, a higher order could produce a more accurate solution, is that applicable to all situations? I ran a square cylinder simulation both in 2d and 3d from p1 to p3, the result shows a decreasing drag coefficient from 2.2 to 2.0 with the increase of the polynomial order. 2d and 3d simulations both met the same situation, so which one is more close to the actual value? -
Anti-aliasing in config file
It’s very interesting to decide whether to take measurements to avoid aliasing. I tested different cases with or without anti-alising, and found that when the flow was complicated with different vortex intertwining, the measurement was always a must. About the two different methods, Park’s paper AIAA Aerospace Research Central is really helpful to better understand them. L2 projection is more accurate but setting the right value seems to be a trick to me, high value is time-costing while a low one produce NaNs. The modal filtering is less accurate but you don’t need to change the dt and dtau, so more convenient. Actually, in my 2d simulation of a bridge girder, I could barely choose a reasonable value of quad-deg (either results in NaNs, or limits my dtau to 1e-8). Also, turning to modal filtering raised aRuntimeError: NaNs detected at t = 50.840000000517. Although it’s a 2d case under high Reynolds number, I think the dimension shall not be the vital cause. -
The stiffness of the pressure residuals and ac-zata
Just as other post tells, pressure residuals should not be a factor in measuring the convergence of the ac-ns case. The simulation can even continue for a long time with a high but stable pressure residuals. Except for switching to 3d or adopting a p-multigrid method, is changing ac-zeta will do any help? The parameter is the only value for us to adjust, and it’s suggested to be in the range of 2~10 times the velocity. For a simulation, you can obtain the desired Reynolds number by simply changing the nu, which means the ref velocity remains the ‘same’ under different cases. Maybe an increase of ac-zeta with the increase of Re would be more in line with reality?
Pardon me for being so wordy, but these are the ones I could not find more details about in the former discussion of the forum. Hope someone can share his/her opinions.
Regards, Thatcher