The Designer's Guide Community Forum
Modeling >> Semiconductor Devices >> Lack of systematic approach. Frustration. [RANT]

Message started by physixer on Mar 16th, 2017, 5:31pm

Title: Lack of systematic approach. Frustration. [RANT]
Post by physixer on Mar 16th, 2017, 5:31pm

I am trying to perform AlGaN/GaN HEMT 2D simulation up to breakdown voltages (1000 V to 2000 V). I face various convergence failure messages from Sentaurus TCAD. Sometimes it works and sometimes it doesn't.

The problem is: I have not been able to systematically diagnose the cause of convergence failures.

  • It is well-known that we need a high-resolution mesh around the 2DEG. First of all, why? where does this folk knowledge come from and where is the systematic justification for it? You might say, "why don't you try the coarse mesh and find out for yourself". Well guess what, a lot of the times, if the simulation finishes (successfully) with a high-res mesh, it finishes with a low-res mesh as well! And if it fails for low-res mesh, it fails for high-res mesh as well! (note: I'm not saying I don't know high-res mesh is good from the numerical POV. I know that. What I'm wondering is, when is low-res the cause of convergence failure and when it isn't and what are the scales of hi vs lo-res, 100 per micron? 1000? 10k? 100k?)

  • I read one of the TCAD "guidelines" that said "Always use the keyword Extrapolate in Math section (unless case A, B, C etc)". Well guess what, that's a garbage advice, from the official source no less. I spent a month dealing with one convergence failure after another only to find out removing the 'Extrapolate' flag fixed the failures for a huge number of configurations. And no, my configs did not fall into the A, B, C cases which are the documented exceptions.

  • In reverse bias, the pre-breakdown current is really low. Is that the cause of convergence failure? as opposed to a bad mesh? How do I find that out I'm completely wasting my time mis-diagnosing the convergence problems?

  • Don't even get me started on time-step adjustment. Nothing could be more ad-hoc than that.

The whole thing appears to be based on trial-and-error, and once things get working, we are none-the-wiser about why things failed, and how to avoid failure in future simulations.

Title: Re: Lack of systematic approach. Frustration. [RANT]
Post by physixer on Apr 3rd, 2017, 8:40am

An example for point 4 above:

If we set MaxStep larger than your final Vbias, TCAD selects its own bias-ramp-up strategy (it still doesn't go from 0 to vbias in one solve). No problem. Let's call it "auto ramp-up". Sometimes it works, other times we select a ramp-up strategy using small values for InitStep, MaxStep, and so on, and then it works. Call it manual ramp-up.

But get a load of this. When it works for auto, and we still decide to do another run with manual ramp-up, with each step equal or smaller than the smallest step from the auto ramp-up, it still fails sometimes (fails to reach final vbias), even though we didn't change the mesh, didn't change the physical models, didn't change anything else about the system! Another way to put it: if we're failing at manual ramp-up no matter how small of a MaxStep we pick, what might work is to not do manual ramp-up at all, but instead set MaxStep larger than final Vbias and not specify any value for InitStep and MinStep.

Point being: if larger steps work, smaller steps should always work (it should be impossible to fail in a run with smaller steps). But that's not the case observed here. TCAD is quite a beauty isn't she?

The Designer's Guide Community Forum » Powered by YaBB 2.2.2!
YaBB © 2000-2008. All Rights Reserved.