Convergence criteria based on temporal statistics

Rojj · November 6, 2016, 04:32

Hi,

In Openfoam+ it is possible to end the simulation when relevant statistics have converged within a certain tolerance. (Case Termination Controls)

For example (see attached) I would like to stop the simulation when the residuals have not changed considerably during the last n iterations.

Is this possible with OF 3.x or 4.x?

Thanks

Rojj · December 1, 2016, 03:10

Hi there

I assume this is not possible then :-)

Lorenzo92 · December 1, 2016, 03:57

Well it depends on the kind of simulation : are you using RANS or LES? Indeed if you are using RANS velocity is already a statistics since it is defined as temporal average, if you are using LES it is surely a good idea to monitor the life of first and second order moments ( keep an eye also on variance: pPrime2Mean, UPrime2Mean ecc ecc).
In the latter case I assume you have a stationary boundary condition , not varying with time, otherwise the flow won't reach by nature a steady state .
I other words look at moments varying with time advancing to know how far the transient phase is( if there is one of course!)

Lorenzo92

Rojj · December 1, 2016, 04:33

Thanks Lorenzo,

I am running a RANS simulation (External flow over buildings).

I did not know that the residual was a temporal statistics. However, sometimes the residuals tend to oscillate around a value that is very close to the convergence limit, but never get there. This would not be a problem if I could monitor the solution visually.

I am running 8 to 16 wind directions (on 8 to 16 AWS machine) at the same time and the whole process is managed programmatically via scripts. I have tried to output the solution every (say) 100 time steps and you can see that it stops changing after 300-400 (varies with case) time steps. Nevertheless the script will wait until 2000 times teps (endTime).

One solution could be to create logs, monitor the last 20 residuals with a background script and alter the controlDict endTime once some statistics is satisfied. I was hoping to find a more elegant solution like in OpenFOAM 3+.

Thanks for reading.

Lorenzo92 · December 1, 2016, 14:07

Residuals are a measure of how much numerical solution fits the governing equations, so they are not statistics like mean or variance.

I got your issue eventually : I know that in fvSolution there is a section called "residual controls" (assuming for example you are using simpleFoam) .
You can there set the residual tolerance for each field: once it has been reached the simulation should stop automatically if I remember correctly, ignoring the endTime in controlDict.

Rojj · December 1, 2016, 14:29

Yes, it is what I actually use. My residuals are all 1e-4 for Ux, Uy, Uz, epsilon, k and p. See the image I attached to the first post.

The fact is that often they become lower than 1e-3, but then get very close to 1e-4 without ever reaching it and I have to wait till 1000 iteration to exit simpleFoam. Usually it is Uz and/or p that manifest this issue.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Statistics for adaptive convergence	whykickamoocow_7	ANSYS	0	October 1, 2012 22:28
Convergence Criteria in FLUENT	mahi007	FLUENT	9	March 1, 2012 07:46
SIMPLE algorithm, convergence criteria	bbasal	Main CFD Forum	3	December 3, 2011 00:24
convergence criteria	iuaniu	CFX	3	September 24, 2009 20:20
convergence criteria	swarley	FLUENT	3	June 24, 2009 12:18

December 1, 2016, 03:10		#2
Rojj Member Ruggiero Guida Join Date: Apr 2013 Location: World Posts: 46 Rep Power: 13	Hi there I assume this is not possible then :-)

December 1, 2016, 03:57		#3
Lorenzo92 Member Lorenzo Join Date: Oct 2015 Location: Graz Posts: 49 Rep Power: 11	Well it depends on the kind of simulation : are you using RANS or LES? Indeed if you are using RANS velocity is already a statistics since it is defined as temporal average, if you are using LES it is surely a good idea to monitor the life of first and second order moments ( keep an eye also on variance: pPrime2Mean, UPrime2Mean ecc ecc). In the latter case I assume you have a stationary boundary condition , not varying with time, otherwise the flow won't reach by nature a steady state . I other words look at moments varying with time advancing to know how far the transient phase is( if there is one of course!) Lorenzo92

December 1, 2016, 04:33		#4
Rojj Member Ruggiero Guida Join Date: Apr 2013 Location: World Posts: 46 Rep Power: 13	Thanks Lorenzo, I am running a RANS simulation (External flow over buildings). I did not know that the residual was a temporal statistics. However, sometimes the residuals tend to oscillate around a value that is very close to the convergence limit, but never get there. This would not be a problem if I could monitor the solution visually. I am running 8 to 16 wind directions (on 8 to 16 AWS machine) at the same time and the whole process is managed programmatically via scripts. I have tried to output the solution every (say) 100 time steps and you can see that it stops changing after 300-400 (varies with case) time steps. Nevertheless the script will wait until 2000 times teps (endTime). One solution could be to create logs, monitor the last 20 residuals with a background script and alter the controlDict endTime once some statistics is satisfied. I was hoping to find a more elegant solution like in OpenFOAM 3+. Thanks for reading.

December 1, 2016, 14:07		#5
Lorenzo92 Member Lorenzo Join Date: Oct 2015 Location: Graz Posts: 49 Rep Power: 11	Residuals are a measure of how much numerical solution fits the governing equations, so they are not statistics like mean or variance. I got your issue eventually : I know that in fvSolution there is a section called "residual controls" (assuming for example you are using simpleFoam) . You can there set the residual tolerance for each field: once it has been reached the simulation should stop automatically if I remember correctly, ignoring the endTime in controlDict.

December 1, 2016, 14:29		#6
Rojj Member Ruggiero Guida Join Date: Apr 2013 Location: World Posts: 46 Rep Power: 13	Yes, it is what I actually use. My residuals are all 1e-4 for Ux, Uy, Uz, epsilon, k and p. See the image I attached to the first post. The fact is that often they become lower than 1e-3, but then get very close to 1e-4 without ever reaching it and I have to wait till 1000 iteration to exit simpleFoam. Usually it is Uz and/or p that manifest this issue.