Different results with aoforce and NumForce for b3lyp!!!(bug?)

[evgeniy]

Dear Developers and Users of TM,

I got quite different results, regarding the IR intensities, with
aoforce and NumForce. The calculations were performed for the ground
state of C60 with b3lyp. The difference is very substancial for the three modes,
with aoforce results being:

       mode                  52       53       54

     frequency         588.54   588.54   588.54

     symmetry              t1u      t1u      t1u

        IR                      YES      YES      YES
|dDIP/dQ|   (a.u.)   0.0382   0.0382   0.0382
intensity (km/mol)  2599.24  2599.24  2599.24
intensity (  %   )     100.00   100.00   100.00


while the NumForce gave:


       mode              52       53       54

     frequency         588.42   588.42   588.42

     symmetry           t1u      t1u      t1u

        IR                      YES      YES      YES
|dDIP/dQ|   (a.u.)     0.0033   0.0033   0.0033
intensity (km/mol)    19.14    19.14    19.14
intensity (  %   )       75.83    75.83    75.83

I wonder what can be the reason.
Many thanks!

Best regards,
Evgeniy

[antti_karttunen]

Hi,

Interesting results, the difference is indeed rather large. Was this TURBOMOLE 6.6? Was RI on? Maybe it would be helpful if you could post the control files and aoforce outputs for both cases. I did a quick test with plain B3LYP/def2-SVP (TURBOMOLE 6.5) and the differences were not nearly as large as for your case. I'm attaching the vibspectrum files. There are some differences, but then again my convergence criteria were not very high.

Best,
Antti

[evgeniy]

Hi Antti,

Thanks for your response. Indeed with B3LYP/def2-SVP (TM 6.5) everything is OK.
The problems begin with cc-pVDZ! Hm, I am stumped with this.  Can it be a bug?

Best,
Evgeniy

[antti_karttunen]

Hi Evgeniy,

With B3LYP/cc-pVDZ I see bit larger differences between aoforce and NumForce, but still nothing like what you get with aoforce for your example (vibspectrum files attached, TURBOMOLE 6.5). What integration grid are you using for DFT? I used m4, not the default m3. And did I understand correctly that you are not using RI or DFT-D3 corrections?

Best,
Antti

[evgeniy]

Hi Antti,

Thanks for your response. Well, I used the deault m3 (no RI or D3) and this seems
to be the culprit. Could you maybe give some hint regarding the different DFT integration
grids. Is there some more or less universal and robust grid? Thanks!

Best,
Evgeniy

[antti_karttunen]

Hi Evgeniy,

I think the old rule of thumb was to use m4 for for systems with more than 50 atoms and it seems that the same advice is still in the manual (http://www.turbomole-gmbh.com/manuals/version_6_6/Documentation_html/DOKse13.html#x21-670004.4). Since the same grid should be used when energy comparisons are done between molecules, I decided pretty early on to always use the m4 grid (to avoid the recalculation of smaller systems when a comparison to a larger one is necessary). I don't think the speed difference is that large, anyway.

I have never encountered a situation where the m4 grid would lead into some strange problems, so I consider it a robust one. Definitely for some very non-conventional situations even m4 can fail. It's interesting to see that m3 gives rise to some issues already for C60 with cc-pVDZ, but then again TURBOMOLE has not really been tuned for the highly contracted cc-bases.

Best,
Antti

[Hauke]

I tried to calculate it with B3LYP/cc-pVDZ TM 6.5, only default grid m3, no RI, no D3 (like you) but  scfconv 8 and jobex -gcart 4 as recommended for vibrational spectra in the manual (I don't know if you followed this procedure)

The results I got are pretty close to each other
NumForce
    52        tu            588.11        15.53242       YES     NO
aoforce
    52        tu            588.90        11.91478       YES     NO

So I think it is not the grid m3 to blame (although of course a grid of m4 is not a bad idea)

Which convergence settings did you use (what happens if you try  scfconv 8 and jobex -gcart 4)?

[christof.haettig]

So aoforcewas far from convergence? For large conjugated systems the RI could be a problem, since it is used in a 100% consistent manner.