PRACTICAL CFD MODELING: Judging Convergence
HTML-код
- Опубликовано: 14 июл 2024
- CFD convergence is not an exact science. The CFD engineer relies on three tools to judge when a simulation finishes: monitors, flow patterns, and residuals. But none of these tools work 100% of the time. The well trained engineer understands how to use these tools and how to combine them into a cohesive picture and reliably judge a converged CFD simulation.
Read full detailed article here: dmsonline.us/practical-cfd-mo...
View more tips and helpful articles at www.dmsonline.us/
6.0 References
[1] TCFD, "Axial Compressor CFD," TCFD, 31 Dec 2018. [Online]. Available: www.cfdsupport.com/axial-comp.... [Accessed 31 Dec 2018].
[2] Atsushi Ueyama, "Progress of Time," Cradle MSC Software Company, 31 Dec 2018. [Online]. Available: www.cradle-cfd.com/tec/column.... [Accessed 31 Dec 2018].
[3] RUclips Author: Holzmann CFD, "Holzmann CFD & OpenFOAM® - Dynamic Meshes in Multiphase Flows #2 (Topology Change, Ship Simulation)," RUclips, 10 Feb 2018. [Online]. Available: • Dynamic Meshes in Mult... . [Accessed 31 Dec 2018].
DISCLAIMER: These videos are intended to provide an overview of CFD. They are not a substitute for a comprehensive education. Certain key concepts were intentionally left out of these videos. Please do not think you can perform CFD after only watching these videos. 
Наука
Excellent....!
Thanks for your simple but elegant explanations Brother......!
you are the best .. this is exactly what i was looking for.... thank you so much
I like the idea of a monitor.
GREAT!
Hi Nick.
Thank you for another quality and well produced clip on CFD.
I have two queries on convergence that are bugging me and perhaps you could help....
1. Does the presence of boundary layers (i.e. assuming all acceptable metrics such as skewness, orthogonality, y+ size for model used AND adequate no. of layers) affect the rate of convergence ?
2. In a case with an oscillating residuals analysis (i.e. golf ball with Re 1e-5), may we consider a solution accurate enough if the Continuity Residual does not drop below 10^-3, but a velocity monitor plot is oscillating within a small range (as you stated)? (key point here being the high 10^-3 continuity residual !)
Excellent questions. To the first part about boundary layers, I can't say with certainty. Most of my simulations are scenarios where boundary layer flow is not a strong component of the overall flow regime. Modeling the boundary layer definitely affects the accuracy of my final simulation, but I never tested to see if there was a strict correlation with the rate of convergence. I can certainly say that including turbulence equations slows the rate of convergence, mainly due to boundary layer interactions.
To the second part, I have worked with simulations that had a Continuity Residual of 10^-3, especially with free surface flow. But that is a very high residual, and I would argue that the simulation requires extra scrutiny to accept that residual. As you said, the velocity monitor plot is excellent for extra information. But I would only extract results based on my monitored values. If you only monitored velocity, don't extract any results for force or pressure. Different physics converge at different rates. It would also be wise to plot the mass balance for your domain. In a perfect simulation, total mass flow in = total mass flow out of the domain. Accounting for errors in machine precision, the total difference of mass flow should be less than 10^-6 (normalized). Otherwise, the high continuity residual may indicate that you lost mass somewhere in the domain.
Hope that helps.
@@DatawaveMarineSolutions Thanks for the replies Nick...I still have much to learn :)
As to your second reply; yes I do always plot the mass balance and where applicable Cd and Cl. In fact these monitors stabled out nicely and that's why I was specifically curious about the Continuity Residual (10^-3) not dropping.
Hi Nick,
First I have just watched a bunch of your videos and they are excellent, it helps a lot when learning CFD !
I had a question concerning iterations and time steps. As you explained in the time discretization video: the number of iteration per step should never be more than 10 (more like 5) and the time step should be reduced if in doubt.
However in the residuals plot of this video, there are approximately 10 000 to 20 000 iterations per time step. What is the reason behind this ?
Thank you in advance, and keep up the good work ;)
I borrowed that particular residuals plot from another website, so I don't know the exact simulation that went with it. That specific image was something I debated when creating the video. But it was such a perfect example of the sawtooth pattern that I decided to include it.
I can take a guess on what happened with that simulation. As you mentioned, there are too many iterations between each spike in the residuals for the simulation to solve in time. One other thing can produce that same pattern is a steady simulation, where the boundary conditions are programmed to change after a certain number of iterations. I see that for very unstable simulations.
For example: imagine you need some inlet velocity of U = 50 m/s, but the simulation is too unstable to initialize with that speed. Instead, program the boundary conditions to start with U = 10 m/s, then after 10,000 iterations, step up to U = 20 m/s, at 20,000 iterations, step up to U = 30 m/s, and so on until you reach the full speed. The programming that allows you to do this changes with each piece of software, but they normally have something to allow that type of advanced boundary condition.
I loved your videos that were about boats. Have you considered making a series where you evaluate classic boats from an engineering perspective? Boats like Spray, Joshua, Gypsy Moth IV, Suhaili. I think you'd find quite the audience for something like that.
A good idea. I will add it to the list. I recognize that the CFD videos don't have a very large audience appeal. I am also creating several videos with a more general focus on boats to intersperse with the CFD videos.
Hi sir,
One of the best video on RUclips now...really love video contain..
I have one question to you. As in many converge plot UX,Uy,Uz,p,k all value are almost close to each other but omega was very less value than other. Why this happened? I am learner of CFD.
Can you explain me? Please.
It all depends on the scenario. The low Omega value indicates that there was very little change in turbulence dissipation. So the Omega equation converged easily, resulting in a lower residual. But remember that turbulence also involves the k equation. Both equations drive turbulence.
@@DatawaveMarineSolutions Thank you for response sir..!🙌🔥 Can you share any link or equation name used for K and omega? 🤔
@@rohitsuryawanshi2229 turbmodels.larc.nasa.gov/index.html
That lists several turbulence models. Any set of residuals involving k and omega would be one of several options: Wilcox K-omega, Menter K-omega BSL, or Menter K-omega SSL
Would you reccomend OpenFOAM in a commercial / consultancy setting ie. not an academic/research environment
yes. I use it in my own consulting. A commercial CFD code charges $20k - $60k US per year in license fees. It can be very hard to get enough dedicated CFD business to justify that license cost. Instead, OpenFOAM requires no maintenance costs. The frees more resources for validation studies. The downside: OpenFOAM is slightly slower to setup, due to the text interface. And the mesher is limited.
If you can afford it: the best combination is to pair OpenFOAM with a commercial meshing program.
While decreasing the residual, should I increase the number of iterations. Any kind of reply is helpful, thanks in advance
Number of iterations doesn't matter. If you set 500 and you hit 500 but you didn't converge, you can increase the number of iterations and then resume the calculation
Hello, for the spikes when they go up, they should not exceed the residuals criteria also like 1e-3 for example ? Thnks
I would not hold that as a rule. I had simulations where the spikes jumped up to 1e-0 on their first inner iteration. And then quickly jumped down to the appropriate criteria. When judging convergence and accuracy of the simulation, the last inner iteration is the only residual number that counts. Instead, the initial spike is an indication of simulation stability. For my example where the first inner iteration jumped to 1e-0, that simulation was extremely unstable. I technically should have gone to a smaller timestep, but I was limited by computing resources.
Yes but i thought that tis is normal for the first iterations of the whole simulation and not per time step and then when the simulation is continued like 50000 iterations the spikes should be lower than the residual criteria. Is this right thanks
Sir, I have a reactor which has an impeller and a shaft. I use steady-state simulation as the initial condition of my transient simulation. In my transient simulation, the residuals converge for each timestep. I am plotting a monitor for 1) the moment about my shaft+impeller and 2) volume avg turbulent kinetic energy. The latter reaches a steady state with almost 0 fluctuations but the moment monitor doesn't reflect the steady-state as good as the kinetic energy. There are also a lot of fluctuations in the moment monitor. Please suggest how to reduce them and get a flat curve at steady state. (I tried to decrease the under relaxation value for body forces until 0.5 and also increased the residual to 1e-06 but still no stable flat curve)
It sounds like you have done everything correctly. I can't say more without a detailed exam of your simulation. Some alternative explanations:. VOF simulations often result in oscillating convergence. And that can be perfectly valid. Check the pattern of oscillation. Is it reasonably small compared to the mean value? If yes, and it shows a consistent pattern, I would say you have a fully converged simulation.
Llllll
Lllllllllllllllllllll
Llllllllololllll
Lllloloolllloololollllllllllll
WHY IS HE WEARING HIS GLASSES LIKE THAT????
I was trying to reduce the glare from the glasses. Unfortunately, it looks worse than glasses with glare.