Firefly and PC GAMESS-related discussion club


 
Learn how to ask questions correctly  
 
 
We are NATO-free zone
 



Re^4: Question on various recommended thresholds for SA-MCSCF gradient calculations

GrEv
grevgen13@gmail.com


Hi Thom,

Ok, I see your point. Many thanks!

Best regards,
Evgeniy

On Mon Apr 18 '16 8:35pm, Thomas Pijper wrote
---------------------------------------------
>Hi Evgeniy,

>That is indeed what I meant. There is typically no need to tighten the 'remaining stuff' if your calculations converge nicely and your results make sense. To make sure you're using the right amount of precision, you could perhaps increase the overall precision to a much higher degree (using all relevant thresholds) for a typical calculation you're performing, just to see how this affects your results. I also recommended doing this whenever you have doubts about the validity of certain results.
>
>
>Kind regards,
>Thom
>
>
>
>On Sun Apr 17 '16 3:39pm, GrEv wrote
>------------------------------------
>>Hi Thom,

>>Many thanks for your reply. INTTYP, ICUT, $MCSCF ACURCY and $MOORTH are already present in my input, so I was particularly interested in the remaining stuff ($TRANS CUTTRF, $GUGDIA CVGTOL, $GUGEM CUTOFF, $GUGDM2 CUTOFF, ). As I understood from your reply one has to tighten those additional thresholds if something (in particular, geometry optimization) does not converge / goes wrong. Otherwise, one can leave them at the defaults. Is that what you meant?

>>Best regards,
>>Evgeniy

>>On Sun Apr 17 '16 11:59am, Thomas Pijper wrote
>>----------------------------------------------
>>>Hello,

>>>There are indeed many parameters that affect MCSCF precision. In addition to the ones you mention, $MCSCF ACURCY as well as options of $MOORTH can be an important as well. Which settings are optimal differs from case to case.

>>>In the case of SA-MCSCF gradient calculations, leaving these parameters at their defaults may lead to inaccuracies in the calculated gradients. (Firefly calculates gradients for SA-MCSCF orbitals through a semi-numerical approach; therefore, small inaccuracies in the calculated energies will have a relatively large impact on the accuracy of the gradient.) This in turn will cause problems like problematic convergence in geometry optimizations.

>>>Whenever increasing accuracy, I always start with the INTTYP and ICUT keywords. ITOL can be increased as well, but typically has little effect. ICUT=11 is often enough, but you can increase further to 13 or even beyond. For SA-MCSCF, I would also always tighten the $MCSCF ACURCY parameter to at least 5.0D-07. In addition, I would also tighten the various options of the $MOORTH group (important for SA-MCSCF gradients!); see the manual for examples.

>>>When you find that a higher degree of precision is needed, I would start with tightening $TRANS CUTTRF as well as either $DET CVGTOL (for ALDET code) or $GUGEM CUTOFF, $GUGDIA CVGTOL, $GUGDM2 CUTOFF (for GUGA code).

>>>This is only a general approach, so you may need find out what works best with your system.

>>>How adjusting these parameters affects performance is difficult to say as this also depends on your system. In general, CVGTOL has little impact (at least for my systems) as the orbital optimization often takes much more time than the preceding CI coefficient optimization. Apart from that, there is little advice I can give you here.

>>>Hope this helps!
>>>
>>>
>>>Kind regards,
>>>Thom
>>>
>>>
>>>
>>>On Fri Apr 15 '16 8:26pm, GrEv wrote
>>>------------------------------------
>>>>Hello,

>>>>In the case of SA-MCSCF gradient calculations, there are various recommended more tight thresholds in the manual (p. 144), namely INTTYP, ICUT, ITOL, CUTTRF, CUTOFF and CVGTOL. I wonder:
>>>>1) Which of them are most crucial?
>>>>2) How they would affect the computation time?
>>>>3) What should be expected if they are left by default?

>>>>I noticed some inconsistency in the manual, p. 179, input example 3, concerning the above thresholds. In the example, $GUGEM CUTOFF=1D-12, $GUGDM2 CUTOFF=1D-12, while the recommended values are 1.0D-20 and 1.0D-15, respectively. Are these thresholds system dependent?

>>>>Many thanks!
>>>>


[ Previous ] [ Next ] [ Index ]           Wed Apr 20 '16 12:47pm
[ Reply ] [ Edit ] [ Delete ]           This message read 289 times