What could cause residual to skyrocket?

[patrick....@gmail.com]

Good morning all,

I'm working with the Phase Field module, and I'm trying to determine what causes my time steps to bottom out during a simulation (i.e. the flat line of Simulaiton Time vs Steps at Step ~140 below).

The only smoking gun I can find is that my linear convergence becomes gradually worse beginning at this time, and the residual for one of my variables (the order parameter) increases significantly. Throughout most of the simulation, the residual is in the neighborhood of 10^-10 to 10^-12, but in the last few time steps it increases dramatically to 10^-4 and the solve soon fails.

Meanwhile, the residuals for my other variables consistently hover around 10^-15 throughout the entire simulation.

Nothing seems to be physically different in the last stages of the simulation, so I was wondering if anyone knew if there were common reasons for a residual to take off suddenly like mine did?

And pragmatically, do you think it would make a difference if I used scaling to reduce that residual througout the simulation and/or made my convergence criteria more lenient?

In case it's useful information: when I made the mesh coarser, the simulation failed in a similar way but at twice the simulaiton time and number of time steps.

Many thanks,

Patrick

[Alex Lindsay]

If you pasted in a picture in your previous email, I can't see it in either thunderbird or gmail. All the numbers that you're giving...are those initial residuals or final residuals after iterations?

--
You received this message because you are subscribed to the Google Groups "moose-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to moose-users...@googlegroups.com.
Visit this group at https://groups.google.com/group/moose-users.
To view this discussion on the web visit https://groups.google.com/d/msgid/moose-users/b9f7309e-1f9d-4e44-b5f7-6c6063e423d7%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Cody Permann]

Don't paste pictures, just copy text to the email. Pictures are searchable when other users try to search the list. In general think carefully about relative tolerances and absolute tolerances. Solvers take and honor both of these types of parameters.

To view this discussion on the web visit https://groups.google.com/d/msgid/moose-users/f81b31af-f43b-2c53-7f6a-7c75bfb55c91%40gmail.com.

[patrick....@gmail.com]

Thanks Cody,

I wasn't sure how to communicate the situation in text. I've attached a .csv of the output data in case anyone wants to look at it.

I appreciate that suggestion. I'll see how the simulation runs when I set a reasonable floor for absolute tolerance.

Thanks again,

Patrick

[patrick....@gmail.com]

Hey Alex,

Thanks for your response. I tried to paste a picture; I probably just did somethine weird with formatting. All it showed was that time steps approached zero.

Those are the final residuals output to the exodus file.

Thanks again,

Patrick

[Alex Lindsay]

I may be a crazy person, but I don't see an attachment

To view this discussion on the web visit https://groups.google.com/d/msgid/moose-users/73e9b3c0-f0f5-4f47-8bf7-4c87dbeca771%40googlegroups.com.

[patrick....@gmail.com]

No, you aren't crazy.

[Cody Permann]

It's really hard to say with this output. In MOOSE you can have several kernels, BCs and many other user-controlled objects directly and indirectly adding contributions to your residual. There must be something going on in your simulation to make the residual jump right at this point in time. Do you know which variable out of your system has largest norm? Are there any objects that are turning on at the time when this occurs? Lots of things to weed out.

To view this discussion on the web visit https://groups.google.com/d/msgid/moose-users/09c33262-4535-499b-abbb-560274909e2b%40googlegroups.com.