
[Sponsors] 
February 2, 2016, 05:26 
Grid independence study problem

#1 
Member
Kamal Bisht
Join Date: Jun 2015
Location: Germany
Posts: 57
Rep Power: 11 
Hi, during the grid independence study I am getting the results closer to the experimental values as expected. But if I refine my mesh too much, the results start varying and shows an error more than the coarsest mesh. What is the reason for this behavior?
I have attached a graph showing different results. 

February 2, 2016, 10:37 

#2 
Senior Member
Matt
Join Date: Aug 2014
Posts: 947
Rep Power: 17 
Hard to say without knowing what you are modeling or how you modeled it. My initial guess would be that it's related to turbulence model and the its assumptions, but I can't say for sure.


February 2, 2016, 11:57 

#3 
Member
Kamal Bisht
Join Date: Jun 2015
Location: Germany
Posts: 57
Rep Power: 11 
I am doing multiphase simulation (water and air) in a pipe with VOF method and realizable kepsilon method


February 2, 2016, 12:02 

#4 
Senior Member
Matt
Join Date: Aug 2014
Posts: 947
Rep Power: 17 
Ok... can you show a plot that has cell count of your meshes for each of your mesh dependency models along X axis and error with respect to experimental data along the Y axis? The plot you show is too busy to interpret properly.
I might also help to see shots of the mesh for each level of the mesh dependency study to make sure there isn't something funky going on elsewhere. 

February 2, 2016, 12:28 

#5 
Member
Kamal Bisht
Join Date: Jun 2015
Location: Germany
Posts: 57
Rep Power: 11 
Sorry but I don't get what do you want me to plot. What do you mean by error wrt to experimental on Y axis? Does it mean that I should plot the difference between the experimental and simulation result for each mesh by a line graph, which mean 5 different lines for each type of mesh?


February 2, 2016, 12:42 

#6 
Senior Member
Matt
Join Date: Aug 2014
Posts: 947
Rep Power: 17 
You tell me what error you meant by:
"But if I refine my mesh too much, the results start varying and shows an error more than the coarsest mesh." Your plot doesn't show an error, just overlaid curves so it is hard to quantify and visualize the error you are talking about. What I do see is a VERY chaotic experimental curve and then a bunch of CFD curves that are very close to each other. It is very difficult to read. What I would suggest doing is calculating the error (difference between CFD1 and EXP, then CFD2 and EXP), and so on for each instant t in your time domain. Then calculate the RMS error for the whole time domain and use this single value as your estimate of error for each CFD model. Then you will have a plot or bar graph that has mesh count along the X axis and RMS error along the Y axis consisting of 5 points. That will be much easier to read and interpret. Another question that comes to mind... since this is transient. Are each of your time steps converging? 

February 2, 2016, 13:19 

#7 
Member
Kamal Bisht
Join Date: Jun 2015
Location: Germany
Posts: 57
Rep Power: 11 
Actually lines in my graph are pressure values. The minimum pressure observed during experiments is 92 Pa. As I am refining the mesh, the pressure drop predicted by the simulation getting better but for the finest mesh it's around 46 Pa. however after 0.1 s, pressure is stabilized around 0 pa which is mostly accurate for all simulations.
Yeah my time steps are converging as the residuals fall to 10e5 and below. 

February 2, 2016, 13:22 

#8 
Senior Member
Matt
Join Date: Aug 2014
Posts: 947
Rep Power: 17 
I am not sure that calculating the error based on lowest pressure is a good approach since your experimental values are very erratic. I stand by my suggestion of RMS error.


February 2, 2016, 13:28 

#9 
Senior Member
Matt
Join Date: Aug 2014
Posts: 947
Rep Power: 17 
To be 100% clear, let;
CFD(t) represent the pressure value as a function of time as calculated by your CFD analysis. EXP(t) represent the pressure value as a function of time for your experimental results. ERR(t) = (CFD(t)EXP(t))/EXP(t) x 100 % (this is the instantaneous error of your CFD solution with respect to your experimental results.) Then the RMS error (assuming you have a constant time step) is sqrt(sum(ERR^2)/n), where n is the number of samples in ERR. This will give you a single value that represents the average error for a given CFD case across the whole time domain. 

Thread Tools  Search this Thread 
Display Modes  


Similar Threads  
Thread  Thread Starter  Forum  Replies  Last Post 
Problems with convergence and grid independence study  nikn  CFX  0  August 26, 2015 06:32 
What is meant by Grid Independence Study?  Khan  FLUENT  10  July 2, 2015 22:40 
grid study, important question  hamid1  FLUENT  1  August 4, 2013 00:14 
[ANSYS Meshing] grid study, important questions  hamid1  ANSYS Meshing & Geometry  2  February 10, 2012 13:28 
Grid independence check  Mattis Voss  FLUENT  1  November 23, 2005 10:30 