Discussion Forum

Posted: 1 decade ago Jan 10, 2012, 1:24 a.m. EST

Hi That is proposing "manual scaling" it comes from the very large difference in absolute values that you get with different physics and different dependent variables. From the numerical simulation (mathematics of binary number theory) the rule is keep all values close to "1" because there you have the best "resolution" (binary numbers are just a (poor) approximation of the Real numbers). Normally COMSOL has some build in rules to set the scaling values for the dependent variables, but when you add new ones, or in given cases COMSOL might choose a poor value and the results, or even the convergence might be poor or even fail. Take FSI with pressures p in 1E5 [Pa] displacements u in 1E-6 [m] and velocities U in 1 [m/s], if you divide the P values by 1E5, and multiply the u values by 1E6 you get all three unknowns to start close to "1", this to ease the matrix inversions and factorisations. BUT you need to correct the results by adjusting back the 1e5 and 1e-6 ! Normally COMSOL knows how to do that internally to keep the results coherent ;) -- Good luck Ivar

Posted: 1 decade ago Jan 10, 2012, 1:57 p.m. EST

Hi Ivar: Thank you very much. I am still not sure if I can choose the right manual scaling when running my simulation, is there any tutorials or discussions on that topic?

Posted: 1 decade ago Jan 10, 2012, 3:07 p.m. EST

Hi The "right value", from my understanding, is approximately the average value you expect to get, a factor 10 above or below seldom gives problems, it not an exact science, as often the solver scans orders of magnitude when solving. As I said mostly COMSOL looks after the scaling for you (it tells you the values used in the solver log), but when you add specific ODE or PDE variables, then it's often usful to give a manual scaling, as then you mostly (should) know better than COMSOL what to expect ;) -- Good luck Ivar

Posted: 1 decade ago Mar 1, 2012, 8:57 p.m. EST

In the FSI example that you gave, what happens if you have a case where the one of your variables has a wide range of values across your geometry. If I have a 2D flow situation where there is high flow (1-10 range) in some portions and low flow (0.1-0.001 range), what "scale value" do you choose?

Posted: 1 decade ago Mar 2, 2012, 1:52 a.m. EST

Hi in some cases you need to test and try a litle, starting somewhre "in between". Wat I have not done so far, is to try to run twice the same study, by using the the frst results as initial conditions for the second study and see if COMSOL uses the same scaling by default, or if it is clever enough to analyse the initial conditions to get a "better" guess. But even then COMSOL might get fooled, nothing better than looking over its shoulder and and tweaking a little by yourself, so long you understad what is happening -- Good luck Ivar

Posted: 1 decade ago May 3, 2012, 3:33 a.m. EDT

[QUOTE] ... As I said mostly COMSOL looks after the scaling for you (it tells you the values used in the solver log) ... Ivar [/QUOTE] Hi Ivar, Could you give me a hint about where to find this (in which log). The log I identify as the solver log is the one under the "Time-Dependent Solver 1" node. For example, I just ran a simulation, and it said: Time-Dependent Solver 1 in Solver 1 started at 3-May-2012 17:27:32. Time-dependent solver (Generalized-alpha) Number of degrees of freedom solved for: 205. Nonsymmetric matrix found. Step Time Stepsize Res Jac Sol Order Tfail NLfail 0 0 out 4 3 4 0 1 1e-009 1e-009 out 5 4 5 2 0 0 2 2e-009 1e-009 out 6 4 6 2 0 0 3 3e-009 1e-009 out 7 4 7 2 0 0 4 4e-009 1e-009 out 8 4 8 2 0 0 5 5e-009 1e-009 out 9 4 9 2 0 0 6 6e-009 1e-009 out 10 4 10 2 0 0 7 7e-009 1e-009 out 11 4 11 2 0 0 8 8e-009 1e-009 out 12 4 12 2 0 0 9 9e-009 1e-009 out 13 4 13 2 0 0 1e-008 - out 10 1e-008 1e-009 14 4 14 2 0 0 Time-Dependent Solver 1 in Solver 1: Solution time: 1 s. Where I don't think the information about variable scaling appears. Is this not the right log? Or do I need to enable something? Evgeni

Posted: 1 decade ago May 3, 2012, 5:56 a.m. EDT

Hi indeed I do not see it there, there is a "log detailed" option in one of the solver nodes, try that, but I thought that in 4.2a it was always there, maybe not ;) -- Good luck Ivar

Posted: 1 decade ago May 3, 2012, 8:53 p.m. EDT

I found it! I searched for "log detailed", which gave me the page "COMSOL Multiphysics > Solver API > Solution Object Data", and the corresponding section was the "Advanced" node under the "Time-Dependent Solver 1" (in my case). It has a "Solver log" field. I'm sure I've looked there a few times, but never noticed it. Many thanks!