What is the accuracy of scf=qc when convergence criteria is lowered?

I am trying to run single point calculations with Gaussian 09 for a Mulliken population analysis on large graphene/graphite (940-1411 atoms) sheets/clusters.

I was having difficulty with memory issues (single CPU license) running DFTBA for single point calculations. I have stepped my convergence (beginning with conver=1, and increasing by reading in my previous checkpoint files). I found my results oscillating and failing.

I decided to try scf=(qc,conver=5,maxcycle=500). This worked, which concerned me.

It not only worked, but it worked much faster than my failed sp calculations. I know Gaussian's old default convergence for single point calculations was N=4. As my calculation was done in less than two hours compared to the days it was taking for DIIS, I have become concerned at the accuracy of the QC using conver=5. I cannot find anything indicating my approach is bad, but also I cannot find anything syaing my approach should yield reasonable results.

Can anyone please tell me if this approach is reasonably accurate?

Something that I have discovered is that when Gaussian (or other good ab initio software) fails to converge, there is something sketchy about what one is trying to calculate. Sometimes DFT calculations just aren't appropriate, especially in cases that require multiconfigurational SCF for even a zero-th order description.

In this case, given the sheet-like character, I suspect the problem is that the molecular orbitals are so close together, and so similar, that there are serious linear dependency problems that lead to all sorts of havoc in the linear algebra routines and convergence accelerators.

If you are trying to simulate an infinite sheet, then I suggest that you try plane wave codes like those found in solid state chemistry (Gaussian's PBC code, NWChem, abinit, etc.) These sorts of calculations have gotten easier in recent years, but sometimes they require serious horsepower.

There really does seem to be a problem with SCF convergence, and this could be because you are approaching some kind of band structure limit. I have always used QC as an absolute last resort when I must do something wacky.

(One other thing, you have "capped" your peripheral benzenoid rings with hydrogens?)

• Thank you for your reply. First I must ask, why do use QC as a last resort? I tried as such expecting an error due to memory. Everything I have read suggests it is more memory intensive, but at the same time more reliable. It was because of that I became very concerned about the speed of my single point runs (roughly 1/4 of the time my SP-DIIS calculations). If you avoid QC for other reasons, please let me know. – Linda Jun 1 '13 at 3:28
• Next, yes I have tried different configurations of the substrate. I started with a simple single layer graphene; capped my edges of that graphene with hydrogen; am currently running a double layer capped with hydrogen (so far DIIS only); and have graphane and a pure double graphene to try. My geometry is set up according to accepted theory (although it is completely pristine). – Linda Jun 1 '13 at 3:35
• I think that QC calculates the orbital Hessian w.r.t. the coefficients. That's awfully expensive, when there are other methods which could get the job done faster. My point is that there seems to be something pathological with your molecular system, which is why DIIS doesn't work, and why you have to tweak convergence criteria even on QC -- the most robust convergerer known to humankind. – Eric Brown Jun 1 '13 at 3:39
• another question: is Gaussian complaining about linear dependencies? – Eric Brown Jun 1 '13 at 3:41
• My challenge is to accomplish my calculations using the given software on a single cpu and with the memory limitations. I had read publications where they were quite successful doing something similar, using DFTB and SWNT (albeit much smaller surfaces). Our sheets need to be at least 32x32 Angstroms. When my convergene was oscillating I too suspected an issue with molecular orbital overlap, which is why I began "capping" my ends and set up a fully saturated graphane. I will certainly give the PBC code a try. Thanks again for your reply! – Linda Jun 1 '13 at 3:49