Coarser meshes can produce better results?

asda3D · April 21, 2019, 08:04

Hi everyone,
I am comparing several meshes which utilise coupling CFD and experimental techniques against a fully resolved fine mesh model (19,000 elements using k epsilon model)

The coarse meshes are listed below

CMesh1-630 elements
CMesh2- 960 elements
CMesh3- 2400 elements
CMesh4- 3600 elements

What I have found out is that Cmesh1 produced the most accurate results (temp velocity pressure drop etc) compared to the fine mesh.

I am finding it difficult to explain this phenomenon. What I have learnt from my CFD lectures, is that a denser mesh provides more accurate results because it is more accurately interpolated between two points that are much closer together spatially.

Does anyone know why this could be the case, or point me in the direction of where to look, maybe mesh skewness plays a part.

Thank you in advance

FMDenaro · April 21, 2019, 12:16

Quote:

Originally Posted by asda3D

Hi everyone,
I am comparing several meshes which utilise coupling CFD and experimental techniques against a fully resolved fine mesh model (19,000 elements using k epsilon model)

The coarse meshes are listed below

CMesh1-630 elements
CMesh2- 960 elements
CMesh3- 2400 elements
CMesh4- 3600 elements

What I have found out is that Cmesh1 produced the most accurate results (temp velocity pressure drop etc) compared to the fine mesh.

I am finding it difficult to explain this phenomenon. What I have learnt from my CFD lectures, is that a denser mesh provides more accurate results because it is more accurately interpolated between two points that are much closer together spatially.

Does anyone know why this could be the case, or point me in the direction of where to look, maybe mesh skewness plays a part.

Thank you in advance

What you learned is correct when you consider that the local truncation error of a discretization vanishes for vanishing mesh size.
However,

1) the monotonically decreasing slope of the error appears only after a sufficiently small mesh size, before you could see oscillations of the error.
2) You are using a turbulence model, that means you can have some supplementary effect due to the way the model acts.
3) some bug could be always present in the setting ( that is BC.s or some modelling parameter).
4) 19000 elements is still a very coarse mesh for simulating turbulence, in particular if you have wall turbulence

asda3D · April 21, 2019, 12:51

Thank you for the informative reply,
Firstly this mesh is modelling a fuel bundle inside a reactor core.
In regards to the mesh being too coarse. Using the k epsilon model a y plus of 30 was needed. I started with a 42,000 elemnt model. Any more elements and the domain would satart displaying a large expansion ration which isn't good.

Additionally,through a sensitvity analysis I lost around 4% of accuracy when reducing the mesh density to 19,000

I have been thinking of maybe trying a low reynolds number model which current possesses 390,000 elements (still working on it) and see if that maybe produces better results.

I must also mention i have accounted for roughness aswell. The CFD package had little information about roughness so this could be an issue maybe too.

I will also investigate aspect ratios, skewness and other mesh quality criteria to see of i can find a relationship explaining this problem.

Thank you.

FMDenaro · April 21, 2019, 12:55

Do not forget to check the convergence criteria, are they satisfied at the same level on each grid?

Your flow problem requires a mesh-indipendence study, that is you need to refine more and more the grid until the solution does not change anymore.

Of course, it could be also debated if the reference solution you're using is really accurate...

asda3D · April 21, 2019, 13:00

That was a very quick response that is highly appreciated.

Five lines of interest were created across the domain to ensure mesh convergence was achieved. 18 points of intereest were set for the residual monitoring thay also showed good results.
Even simulating the 42,000 mesh its values lies almost identical to the 19,000 mesh. For example there is a 1.16% difference in pressure gradient values for these two meshes.

I will investigate further to see if i can resolve this.
Thank you FMdenaro for your input

asda3D · April 21, 2019, 17:29

Just an add on the aspect ration is more than double for the finer meshes compared to the coarser meshes. This may be the reason for the coarser meshes possessing greater accuracy.

steve lee · April 21, 2019, 23:00

If your reference data is completely correct.
All of your cases are wrong I think.

4 cases can not simulate correctly.
But your first case was converged the correct solution with wrong way.
Maybe the pressure or velocity fields are not the same comparing reference solution.

And.. you should check not only the number of mesh but also y+ or correct usage of tubulence model

As FMdenaro mentioned mesh should be finer when the solution is not changed.

April 21, 2019, 08:04	Coarser meshes can produce better results?	#1
asda3D New Member anonymous Join Date: Apr 2018 Posts: 26 Rep Power: 8	Hi everyone, I am comparing several meshes which utilise coupling CFD and experimental techniques against a fully resolved fine mesh model (19,000 elements using k epsilon model) The coarse meshes are listed below CMesh1-630 elements CMesh2- 960 elements CMesh3- 2400 elements CMesh4- 3600 elements What I have found out is that Cmesh1 produced the most accurate results (temp velocity pressure drop etc) compared to the fine mesh. I am finding it difficult to explain this phenomenon. What I have learnt from my CFD lectures, is that a denser mesh provides more accurate results because it is more accurately interpolated between two points that are much closer together spatially. Does anyone know why this could be the case, or point me in the direction of where to look, maybe mesh skewness plays a part. Thank you in advance

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
bad momentum source, vortice direction with twoPhaseEulerFoam and non-ortho meshes	cutter	OpenFOAM Running, Solving & CFD	7	November 17, 2016 15:17
compare results from two meshes	sgr	OpenFOAM Post-Processing	4	July 1, 2015 07:38
Different results from similar quality cfx and ICEM meshes	Nick R	CFX	3	January 17, 2011 08:48
Different Results from Fluent 5.5 and Fluent 6.0	Rajeev Kumar Singh	FLUENT	6	December 19, 2010 12:33
Which do you prefer? 6.1 or6.2?	Jen	FLUENT	9	May 31, 2005 04:21

April 21, 2019, 12:51		#3
asda3D New Member anonymous Join Date: Apr 2018 Posts: 26 Rep Power: 8	Thank you for the informative reply, Firstly this mesh is modelling a fuel bundle inside a reactor core. In regards to the mesh being too coarse. Using the k epsilon model a y plus of 30 was needed. I started with a 42,000 elemnt model. Any more elements and the domain would satart displaying a large expansion ration which isn't good. Additionally,through a sensitvity analysis I lost around 4% of accuracy when reducing the mesh density to 19,000 I have been thinking of maybe trying a low reynolds number model which current possesses 390,000 elements (still working on it) and see if that maybe produces better results. I must also mention i have accounted for roughness aswell. The CFD package had little information about roughness so this could be an issue maybe too. I will also investigate aspect ratios, skewness and other mesh quality criteria to see of i can find a relationship explaining this problem. Thank you.

April 21, 2019, 12:55		#4
FMDenaro Senior Member Filippo Maria Denaro Join Date: Jul 2010 Posts: 6,896 Rep Power: 73	Do not forget to check the convergence criteria, are they satisfied at the same level on each grid? Your flow problem requires a mesh-indipendence study, that is you need to refine more and more the grid until the solution does not change anymore. Of course, it could be also debated if the reference solution you're using is really accurate...

April 21, 2019, 13:00		#5
asda3D New Member anonymous Join Date: Apr 2018 Posts: 26 Rep Power: 8	That was a very quick response that is highly appreciated. Five lines of interest were created across the domain to ensure mesh convergence was achieved. 18 points of intereest were set for the residual monitoring thay also showed good results. Even simulating the 42,000 mesh its values lies almost identical to the 19,000 mesh. For example there is a 1.16% difference in pressure gradient values for these two meshes. I will investigate further to see if i can resolve this. Thank you FMdenaro for your input

April 21, 2019, 17:29		#6
asda3D New Member anonymous Join Date: Apr 2018 Posts: 26 Rep Power: 8	Just an add on the aspect ration is more than double for the finer meshes compared to the coarser meshes. This may be the reason for the coarser meshes possessing greater accuracy.

April 21, 2019, 23:00		#7
steve lee Member Steve Join Date: Mar 2018 Location: South Korea Posts: 69 Rep Power: 8	If your reference data is completely correct. All of your cases are wrong I think. 4 cases can not simulate correctly. But your first case was converged the correct solution with wrong way. Maybe the pressure or velocity fields are not the same comparing reference solution. And.. you should check not only the number of mesh but also y+ or correct usage of tubulence model As FMdenaro mentioned mesh should be finer when the solution is not changed.