Discussion Forum

But if I don't stop it, it might be at 2% for more than a day.

Moreover, tomorrow it might be 20% as well :)

Rules for convergence is really governed by your problem. If you think 5% error is tolerable for your task, just set this limit. But if you want to get 0.1% you got to either wait or somehow optimize your model.

I am running highly non-linear 3D time-dependent (upto 30days with 1[h] increments) heat transfer models. It takes 13 hours in one of my computers and 21 hours in another. However, when I start solving, the overall progress stays at 0% to 4% for the first half time (6 to 10 hours of runtime) and then after that, solving speed increases as well as the solver time increments.

Using a probe, I checked the solver time steps. At the very beginning, time increments are 3 to 24 seconds for the first hour of solver time and then they gradually increase and reach 1800 to 3600 seconds (which is my limit) in around 30 hours of solver time and then stay in these steps. This is around 4% of entire solution progress and after this point it takes much less time for the solver to continue.

P.S. Runtime is the actual time which the solution takes and solver time is the time I designate in the solver settings. I use 1e-3 as tolerance in my solver settings and update Jacobian in each solver time step.

Hope this helps,
--
Tolga

Thanks everyone for the responses. They really helped!

Sylvana

Hi dear All,
I have the same problem, could anybody please let me know what is the meaning of "Reciprocal of time step size"?
And what is the meaning when it shows a horizontal line, and when it shows fluctuation?
Thank you,
Giti

Hi Giti,

The reciprocal of the time step size is one divided by the time step size. The bigger the steps the smaller the value. In most default transient solvers COMSOL has the flexibility to increase or decrease the time step size as it sees fit. It increases the step size when the analysis is going well (I am sure there is a more technical description for that), and decreases it when the solver either fails to converge or converges but the time integration error is large.

Nagi Elabbasi
Veryst Engineering

Hi

the choice of time steps and time step progression, can change drastically the time to solution. If you have a diffusion type problem, a logarithmic or power series is often faster than a linear time step

examples: a linear:

range(1,1,1000) range(2000,1000,100000)

a power series;
0 2^{ range(-4,1,12) range(14,2,18) }

or any combination

--
Good luck
Ivar

Hi Ivar,
I was wondering if the time step could have some effect on the transient model's results?
Thank you,
Giti

Hi Nagi,
Thank you so much. Your explanation was so helpful.
Regards,
Giti

Hi

yes it can in particularaly if the time step "jumps over" some changes in the boundary conditions. This happens typically with harmonic type time dependent BC's and automatic time stepping, as COMSOL tries to double regularly the step duration, and then you might get aliasing. Else my impression is "just" that the steps must be small enough for the convergence criteria to control the algorithm that is why COMSOl often makes many, even hundred of intermediaqte time steps between each value it send "out" and stores as your demanded time value

--
Good luck
Ivar

Thank you Ivar.

Hi everyone,

I am also having a similar problem. My question is can we not obtain a simple error v/s time step-type convergence plot and how does one interpret this kind of convergence plot i.e. the reciprocal of time step v/s time step. How does it inform us whether the solution has converged or not?

Thanks,
Mahvash

PS: I am using Comsol 5.2