40m QIL Cryo_Lab CTN SUS_Lab CAML OMC_Lab CRIME_Lab FEA ENG_Labs OptContFac Mariner WBEEShop
  COMSOL elog  Not logged in ELOG logo
Message ID: 28     Entry time: Sun Jun 23 14:00:05 2013
Author: Emory Brown 
Type: General 
Category: General 
Subject: Manipulating the Relative Tolerance 

  We have been seeing an error when attempting to use a stationary solver in conjunction with a set of boundary conditions which does not fix the face of the mirror opposite the applied force.  I have tried a number of settings changes and tweaks in order to attempt to get the stationary solver to converge on our model.  I have found that by switching to the PARDISO solver and increasing the relative tolerance to 500, the solution will converge (Umax=1.49866*10^-10 J).  This does require increasing the relative tolerance greatly, which is concerning.  After doing this, I also found that by increasing the relative tolerance to 700 and using the SPOOLES solver it also converged giving Umax=1.498653*10^-10 J.  The fact that these agree quite well indicates that the increase in relative tolerance may not have harmed our values.  If this were a more complicated system which we would expect to have behaviour which could cause our solvers to get stuck on a set of values I would be more concerned, but I think that this may be a workable fix in this case.  These values of Umax give Sx(100 Hz)=1.97586*10^-40 m^2/Hz which differs by about 5% from our previous value of Sx(100 Hz)=2.08291*10^-40 m^2/Hz.  Unfortunately, this seems to indicate that the difference in results between the COMSOL model and our direct computation is not due to a difference in boundary conditions.  I will spend some time looking for more useful documentation on the relative residual, relative tolerance, and LU factorization: out of memory (despite having more RAM availiable) error, then I will work through our COMSOL model again, possibly remaking it, and check it for any errors which could result in the disparity between our simulated and directly computed results.

After doing this, I was able to find more information on the relative tolerance in the COMSOL release.book (page ~30 https://www.tuegrid-doc.uni-tuebingen.de/dokuwiki/lib/exe/fetch.php?id=hpc-uni%3Asoftware-docs%3Acomsol%3Acomsol&cache=cache&media=hpc-uni:software-docs:comsol:release.pdf).  It appears that the relative tolerance and the Factor in error estimate values jointly determine the maximum allowed difference between successive estimates when using an iterative solver.  I decided to try to get a better idea for this behaviour, so I increased the Relative tolerance to 10000 and surprisingly obtained the same result as before.  I think I am going to recreate our COMSOL model without any unnessesary things implemented and only include a stationary solver, then run this test again as it seems like this should have had some affect on the output.
ELOG V3.1.3-