Firefly and PC GAMESS-related discussion club

 
Learn how to ask questions correctly  
 
 

 

olgerdovich
olgerdovich@yandex.ru

First, what is(are) the trait(s) of my species that makes its slow convergence "more or less expected"?

Second, what are the rules of adjusting the shift value? Since this powerful remedy secures covergence at the cost of its speed, how should one search the happy medium to get a steady convergence in reasonable time?

You use the value of 0.5 and it works for the given case, but for isomeric system it does not work. There is a problem of oscillating of energy during geometry optimization after several steps, so that two adjacent steps differ significantly in energy while odd and even steps coverge apart rather fast, having essentially constant difference between even and odd steps all the time. Is such a case (which appears to need a damping) a subject for "fshift" remedy?

Finally, each successful step in geometry optimization prints "energy converged" or "density converged" or "diis converged" (or something like this). What does each of these statements point at and how one could use it struggling against convergence problems?

Best regards,
Vladimir

On Mon Jan 26 '15 3:25am, Alex Granovsky wrote
----------------------------------------------
>I'm sorry I forgot to attach files.

>Here they are.

>Kind regards,
>Alex Granovsky
>
>
>
>
>On Mon Jan 26 '15 3:17am, Alex Granovsky wrote
>----------------------------------------------
>>Dear Vladimir,

>>I'm sorry for the delay with my reply.

>>I found these systems converges very slowly and requires large shift
>>values to converge. This is what could be more or less expected for
>>your  systems. I'm attaching inputs and outputs for your problematic
>>testcases.

>>Hope this help.

>>Kind regards,
>>Alex Granovsky
>>
>>
>>
>>On Sun Jan 25 '15 3:43am, olgerdovich wrote
>>-------------------------------------------
>>>Dear prof. Granovsky,

>>>I'm  very sorry to trouble you again in connection with my failure to obtain a consistent result with geometry optimization of nitro anion-radiacals with incorporation of electron correlation at RO-B3LYP-level.

>>>To recall problem briefly, there was no problem when just ROHF was applied for any tested anion-radical system with nitro group, but conergence problem appeared even for H-NO2 anion-radical system when B3LYP functional was applied.

>>>Would you be so kind, please, just to make clear how serious the problem is and what sort of workaround could work for the given case?
>>>Alternatively, what method could you kindly recommend for optimization of the geometry of anion-radical of nitro compound with reasonable electron correlation correction providing that similar reasonable degree of electron correlation is thus corrected for during geometry optimization in the case of parent neutral singlet particle (nitrocompound)?
>>>
>>>
>>>Best regards,
>>>Vladimir

>>>On Sat Jan 10 '15 10:50pm, Alex Granovsky wrote
>>>-----------------------------------------------
>>>>Dear Vladimir,
>>>>Dear Thomas,

>>>>I'm currently investigating this testcase. This may take some time.

>>>>All the best,
>>>>Alex
>>>>
>>>>
>>>>
>>>>On Sat Jan 10 '15 3:00pm, Thomas Pijper wrote
>>>>---------------------------------------------
>>>>>Dear Vladimir,

>>>>>I've experimented a bit with your case. You're right that the problem most likely is not due to partial linear dependence. In retrospect, a case with a smallest eigenvalue of 4.34786E-05 typically should not cause any convergence problems. I'm sorry for leading you down the wrong path.

>>>>>I tried some other things (additional $SCF options, using converged ROHF starting orbitals) but was not able to make your case converge. Maybe someone else (Alex?) has any further suggestions.
>>>>>
>>>>>
>>>>>Kind regards,
>>>>>Thom
>>>>>
>>>>>
>>>>>
>>>>>On Sat Jan 10 '15 4:38am, olgerdovich wrote
>>>>>-------------------------------------------
>>>>>>Dear Thomas,

>>>>>>Unfortunately, your advice does not work and the origin(s) of the problem is not (only) partial linear dependence.

>>>>>>Inspection of out-file revealed that partial linear dependence (not very strong: "THE OVERLAP MATRIX HAS    1 EIGENVALUES BELOW 1.5E-04. THE SMALLEST OF THESE IS 4.34786E-05.") was due to diffuse (?) L-shells (L-shell with exponent of 0.0438) of carbons; of one carbon with independency degree 0.6607E-04 and, possibly, of two other carbons with the same independency degree of 0.1947E-03.

>>>>>>When I removed $BASIS group from input-file and replaced previous $DATA group with $DATA group from punch-file with deleted problematic l-shell of one carbon, without affecting other settings of input-file, warning about linear dependency disappeared; however, there was still no convergence during the first SCF cycle after 99 iterations, just as it was when I posted my problem. Still no convergence was after first 99 iterations when I additionally removed two other suspicious (in the sence of linear dependence) carbon L-shells.
>>>>>>RO-B3LYP also crashes in 6-31+G** (not 6-31++G**) basis.

>>>>>>What should I try next?

>>>>>>Best regards,
>>>>>>Vladimir

>>>>>>On Fri Jan 9 '15 9:35pm, Thomas Pijper wrote
>>>>>>--------------------------------------------
>>>>>>>Dear Vladimir,

>>>>>>>If the tightening of various accuracy-related parameters does not work, the problem can most likely only be resolved by deleting one of the overlapping functions. The manual has a section dedicated to partial linear dependence (pages 73-76) which provides instructions on how to do this.
>>>>>>>
>>>>>>>
>>>>>>>Kind regards,
>>>>>>>Thom
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>On Fri Jan 9 '15 9:29pm, olgerdovich wrote
>>>>>>>------------------------------------------
>>>>>>>>Dear Thomas,

>>>>>>>>I'm sorry, in the previous reply I made a mistake - I used DIIS=.T. SOSCF=.F. in input too (see attach to the first message), so I applied all the recomendations from the warning about linear dependence in advance, but it didn't help

>>>>>>>>What should I do else?
>>>>>>>>
>>>>>>>>
>>>>>>>>Best regards,
>>>>>>>>Vladimir

>>>>>>>>On Fri Jan 9 '15 8:54pm, Thomas Pijper wrote
>>>>>>>>--------------------------------------------
>>>>>>>>>Dear olgerdovich,

>>>>>>>>>The problem of the SCF not converging is due to partial linear dependence in your basis. This is a very common problem when diffuse functions are employed. Partial linear dependence is indicated by the following lines in your output:

>>>>>>>>>

                    * * * WARNING * * *
 ------------------------------------------------------------------------------
 THE OVERLAP MATRIX HAS    1 EIGENVALUES BELOW 1.5E-04.
 THE SMALLEST OF THESE IS 4.34786E-05.
 THIS INDICATES A PARTIAL LINEAR DEPENDENCE IN YOUR ATOMIC BASIS.

 TO OBTAIN SCF CONVERGENCE MAY REQUIRE
 MORE ACCURATE INTEGRAL EVALUATION (INTTYP=HONDO, ICUT=11, ITOL=30 IN $CONTRL),
 MORE ACCURATE DIRECT SCF FOCK MATRIX FORMATION (FDIFF=.FALSE. IN $SCF), OR
 CHANGING CONVERGERS (DIIS=.T. SOSCF=.F. IN $SCF).

 EIGENVALUES BELOW 1.0D-07 PROBABLY WON'T CONVERGE.
 EIGENVALUES BETWEEN 1.0D-07 AND 1.0D-06 MAY REQUIRE
 TIGHTENING OF -NCONV- DENSITY CONVERGENCE IN $SCF.

>>>>>>>>>With respect to your second question, the B3LYP functional contains terms describing electron correlation so this should be accounted for. Note though that DFT may not always be accurate, for example when considering long-range electron-electron interactions.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>Hope this helps.

>>>>>>>>>Kind regards,
>>>>>>>>>Thom
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>On Fri Jan 9 '15 1:57am, olgerdovich wrote
>>>>>>>>>------------------------------------------
>>>>>>>>>>Dear colleagues,

>>>>>>>>>>I'm a newbie in computational chemistry, but with some diligence I started learning it in relation to quite simple chemical tasks from the field of nitrocompounds and I've stumbled upon several problems. First of them is the subject of this post.

>>>>>>>>>>When I ran geometry oprimization with hessian computing for 2-nitropropane, first at HF/6-31++G** and then at B3LYP/6-31++G** level (with starting geometry form HF-calculation), I found no problem, there were convergences to ground states.

>>>>>>>>>>Next I ran geometry oprimization with hessian computing for 2-nitropropane radical anion at ROHF/6-31++G** and once again found no trouble. It was surprising to me that some vibrations were very intense in IR to the contrary to their analogues for neutral particle, but I assumed that it was rather reasonable taking into account concomitantly changed spin and charge. The resulting species was in ground state again.

>>>>>>>>>>But when I started geometry oprimization (with hessian computing) for 2-nitropropane radical anion at B3LYP/6-31++G** level (with starting geometry form ROHF-calculation), run crashed on the first iteration cycle for SCF even when I indicate "maxit=99" (at higher values compilator returned error, it seemed like maxit should nor be larger then NSTEP=100 in $STATPT or just larger then 100): Energy changes started to oscillate and resulted in no convergence.

>>>>>>>>>>Output-file is attached

>>>>>>>>>>What may I adjust in my input to get result?
>>>>>>>>>>Hoe can I take into account electron correlation for anion-radical like the above in general?

>>>>>>>>>>Thank you in advance