Author Topic: Difference between aoforce and NumForce  (Read 2593 times)

kkyyllee

  • Newbie
  • *
  • Posts: 3
  • Karma: +0/-0
Difference between aoforce and NumForce
« on: August 25, 2016, 04:29:48 pm »
Hi all,

I'd like to see the performance of NumForce, so I did a harmonic vibrational frequency analysis on NH_3. In the setting, I used default basis set and eht for the guess. I first optimized it with "jobex -level scf -gcart 5", and then did "aoforce" and "NumForce -central -d 0.001889726". The absolute deviations between aoforce and NumFroce are 3.98, 0.59, 2.94, 1.03, 2.24, and 2.29 cm^(-1). From my experience, the error is quite large for such small molecule. After I reviewed all the articles in the topic of "Aoforce and Numforce", I confirmed that the result I got is weird w.r.t. the difference between aoforce and numforce. Do you guys have any idea?

p.s. 0.001889726 equals to 0.001 angstrom which is the default value for most other QM program.

Thank you for your feedback,
Kuan-Yu

gunnar.schmitz

  • Developers
  • Jr. Member
  • *
  • Posts: 10
  • Karma: +0/-0
Re: Difference between aoforce and NumForce
« Reply #1 on: August 26, 2016, 09:25:33 am »
Dear Kuan-Yu,

maybe you can post a bit more details of your calculation. How do you choose for example your $scfconv? For numerical derivatives you need even more tightly converged SCF calculations than usual. As default the threshold is at least  "$scfconf 7" but the manual encourages the user to use at least "$scfconf 8" in combination with NumForce. Play a bit around with this threshold. I guess the agreement will become better. If you are doing a DFT calculation there is also the grid, which effects the accuracy. Default is m3, but m4 or m5 might be better for high accuracy in combination with numerical derivatives.

Cheers,
Gunnar 

uwe

  • Global Moderator
  • Sr. Member
  • *****
  • Posts: 425
  • Karma: +0/-0
Re: Difference between aoforce and NumForce
« Reply #2 on: August 26, 2016, 09:58:59 am »
Hi,

in addition to what Gunnar wrote and in case of a DFT input, add weight derivatives to the $dft section since those are included by default in aoforce but not in NumForce.

This hint and also those of Gunnar are given in the Turbomole manual, the prerequisites section of the chapter 'Calculation of Vibrational Frequencies and Vibrational Spectra'. See e.g.:

http://www.cosmologic.de/files/downloads/manuals/TURBOMOLE-html-manual/DOK.html

or directly:

http://www.cosmologic.de/files/downloads/manuals/TURBOMOLE-html-manual/DOKli15.html#x91-28100014

I just tried NH_3 with Hartree-Fock and your settings - increasing $scfconv (or $denconv) a little bit gives numerical deviations lower than 1 cm^-1 (for modes which have > 1000 cm^-1 as total value).

For most applications (check for minimum, transition state search, thermochemistry, ...) the accuracy of the default settings is sufficient, though. But it doesn't hurt to follow the advice of the manual either...

Regards,

Uwe

kkyyllee

  • Newbie
  • *
  • Posts: 3
  • Karma: +0/-0
Re: Difference between aoforce and NumForce
« Reply #3 on: August 26, 2016, 03:57:59 pm »
Thanks. I did use "scfconv 8". I also played around with "jobex -gcart" and "NumForce -d". The best result I can get so far is "0.24, 0.39, 0.05, 0.65, 0.89, 1.23" with -gcart 5 and -d 0.01889726 in HF calculation. I still worry about the errors for such small molecule though.
 
Dear Kuan-Yu,

maybe you can post a bit more details of your calculation. How do you choose for example your $scfconv? For numerical derivatives you need even more tightly converged SCF calculations than usual. As default the threshold is at least  "$scfconf 7" but the manual encourages the user to use at least "$scfconf 8" in combination with NumForce. Play a bit around with this threshold. I guess the agreement will become better. If you are doing a DFT calculation there is also the grid, which effects the accuracy. Default is m3, but m4 or m5 might be better for high accuracy in combination with numerical derivatives.

Cheers,
Gunnar

kkyyllee

  • Newbie
  • *
  • Posts: 3
  • Karma: +0/-0
Re: Difference between aoforce and NumForce
« Reply #4 on: August 26, 2016, 04:02:44 pm »
I have not tried $denconv yet. Thank you for the advice.
Hi,

in addition to what Gunnar wrote and in case of a DFT input, add weight derivatives to the $dft section since those are included by default in aoforce but not in NumForce.

This hint and also those of Gunnar are given in the Turbomole manual, the prerequisites section of the chapter 'Calculation of Vibrational Frequencies and Vibrational Spectra'. See e.g.:

http://www.cosmologic.de/files/downloads/manuals/TURBOMOLE-html-manual/DOK.html

or directly:

http://www.cosmologic.de/files/downloads/manuals/TURBOMOLE-html-manual/DOKli15.html#x91-28100014

I just tried NH_3 with Hartree-Fock and your settings - increasing $scfconv (or $denconv) a little bit gives numerical deviations lower than 1 cm^-1 (for modes which have > 1000 cm^-1 as total value).

For most applications (check for minimum, transition state search, thermochemistry, ...) the accuracy of the default settings is sufficient, though. But it doesn't hurt to follow the advice of the manual either...

Regards,

Uwe