Convergence problems - beyond MODFLOW Online Guide FAQ item K

[Ulysses Hillard]

Hi,

Is there a place with guidance and suggestions for how to adjust a model that is not converging beyond what is in the MODFLOW Online Guide FAQ? 

Most of the suggestions in the FAQ concern wetting and re-wetting and the model scenario I am running already has wetting turned off in all layers. It still ends up with convergence problems in certain cells that area adjacent to where the solver creates islands of dry cells. The stress period is ten days long and fails on the second or third day. From this I concluded that the initial heads are not the issue though many are below the bottoms of the cells within which they are defined. The model fails to converge both in steady state and transient runs.

I have already tried a variety of different solvers. PCG (what I started with) gets further than any of the other solvers I tried. I also varied the PCG solver parameters to no avail.

I used LST files and ModelMuse to determine that the cells where the solver fails to reach convergence are always adjacent to dry cell islands. I further identified that the cells in question always have discontinuously high flow values in one or more of the three cell face directions and discontinuously high or low storage relative to their neighbors. I just don't know what to do with this information as far as what to change about the model (hydraulic conductivity? layer top and bottom definition?). 

The convergence failure cells are all within areas of uniform hydraulic conductivity. The model is set to have no horizontal or vertical anisotropy. 

The model includes nine layers with substantial vertically variability, based on 3D surfaces developed from well bore and bedrock investigations.

The model includes GHBs and DRNs. The model runs substantially longer before failing without the DRNs but I was not able to find anything that stood out in the cells adjacent to or near the DRNs

I have this notion that somebody, somewhere has written up further discussion of approaches for getting a model to converge. I searched the web and this forum's threads and found nothing, though. Does anyone know of anything?

Any guidance would be appreciated.

-Ulysses Hillard

[Richard B. Winston]

You could try using shorter time steps. It also might be useful to look
at the heads for time steps prior to the non-convergence to see if they
are reasonable. You could also try using just confined layers until you
have a better handle on what is going on and only then convert them to
unconfined.

> --
> You received this message because you are subscribed to the Google
> Groups "MODFLOW Users Group" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to modflow+u...@googlegroups.com.
> To post to this group, send email to mod...@googlegroups.com.
> Visit this group at http://groups.google.com/group/modflow.
> For more options, visit https://groups.google.com/groups/opt_out.

[Ulysses Hillard]

Thank you for that suggestions. I tried shorter time steps. This had no effect on model convergence.

I did not see anything remarkable in the heads in the time step before the convergence failure.

I can certainly run the model with confined layers. What should I be seeking to learn, though? (Other than using the resulting simulated heads as initial heads)

-Ulysses

[Richard B. Winston]

Assuming that the model is able to run with confined layers, take a look
at the simulated heads and identify locations where there are unusual or
unrealistic heads and see if you can figure out what is causing them.

> > an email to modflow+u...@googlegroups.com <javascript:>.

> > To post to this group, send email to mod...@googlegroups.com

> <javascript:>.

> > Visit this group at http://groups.google.com/group/modflow

> <http://groups.google.com/group/modflow>.

> > For more options, visit https://groups.google.com/groups/opt_out

> <https://groups.google.com/groups/opt_out>.

[Fabian Nick]

Hi Ulysses,

is it the linear solver (i.e. PCG etc) that's not converging or the non-linear solver (Newton?!) that's not converging?

- Fabian

[Ulysses Hillard]

Hi Fabian,

It is PCG that fails to converge at time step 7, with the layers set to be convertible. The model does run through its entire set of time steps when run with all layers confined.

I did attempt to run the model using the NWT version of MODFLOW. It failed at an even earlier time step. I admit to not knowing what parameters to adjust with the NWT solver, though. Do you have any suggestions? I had thought that the preferable approach was to adjust the model parameters and geometry until it converges consistently but I would be delighted to be wrong about that.

-Ulysses

[Ulysses Hillard]

Thank you Dr. Winston.

The model runs all the way to the end of its specified number of time steps without any convergence problems when running with confined layers.

I am looking at a rainbow shaded plot of simulated heads in the final timestep. I do not see any locations where there are unusual or unrealistic heads. The field of values appears to be a continuous surface from a uniform high head in the reservoir (represented by GHBs) to the low heads in the flow face through with the aquifer drains (also represented by GHBs).

Are there other tell tales that it would be advisable to look for?

I will volunteer to assemble these suggestions you (and hopefully others) give if that would be useful.

-Ulysses

[Fabian Nick]

Hi Ulysses,

What you've tried seems reasonable, choosing a different non-linear solver gives a different linear system to be solved which might help some times. I've also observed that problems with confined layers are usually easier to solve for the linear solver than those with unconfined layers. Your model seems to fit into that as well.

Have you tried increasing the maximum number of iterations for the PCG solver? Maybe its convergence is just really really bad and it takes a lot of iterations to solve the system. If this is the case, increasing the number of maximum iterations might help solving the problem although it might be pretty slow and thus not feasible for day-to-day use.

If the data is not confidential, I would be happy to take a look at the model. I work at Fraunhofer SCAI, where we develop an algebraic multigrid solver for linear systems. I could check for you if that solves the problem. However if the problem lies within the physics of the model, that won't help of course...

Regards,

Fabian

[Ulysses Hillard]

Thank you greatly for your offer, Fabian.

I will check regarding confidentiality and get back to you.

I am running PCG with maximum iterations = 100 and threshold head and residual = 1000 ft. I did actually try even larger values but they did not help.

I am pretty sure, however, that the problem lies with the model physical definition. The model was developed by a consultant using a commercial GUI and the resulting creature is a creature with three dimensional features that are defined to much greater detail than is necessary for the problem at hand. It is these features - great vertical variations in seven separate layers - that I am hoping to simplify until the model can run without crashing.

I am looking for indications from the model results and error reporting as to what about those features to change.

-Ulysses

[Fabian Nick]

Hi Ulysses,

any news on this topic? I think I should have some spare time next week!

Regards,
Fabian

[Ulysses Hillard]

Hi Fabian,

I received approval from my supervisor to pass the model to you.

Thank you again for your offer of help!

I attempted to attach the files to this post asa zip archive file but the file was too large for Google Groups, apparently. Is there an alternate way to get the file to you? It is a little over 7M.

-Ulysses

[Fabian Nick]

Hi Ulysses,

I think the easiest thing to do would be if you e-mailed me the files to my work e-mail address which is fabian [dott] nick [attt] scai [dotttt] fraunhofer [dottt] de

If you want to, you can encrypt the file and/or the e-mail with my public GPG key which you can find here: http://pgp.mit.edu/pks/lookup?search=Fabian+Nick&op=index (please use the one that's associated with the Fraunhofer address).

Regards,

Fabian