Ceilidh ... Re^2: Slow convergence of geometry optimization using PCM

Firefly and PC GAMESS-related discussion club

Learn how to ask questions correctly

We are NATO-free zone

Re^2: Slow convergence of geometry optimization using PCM

Dear Alex and All,
Thank you for your answer.
Sorry for delay. I had some ideas to check before answering.
Unfortunately, with relaxed OPTTOL of 3.0D-4 and increased precision I remained in a saddle point of the 2nd order - the point with the minimal E, that the optimizer had ever come across before (imaginary frequencies of ~200I and 400I cm-1 :( ). Restart from
1) THIS POINT with tightened OPTTOL (not to get stuck in this point) and tightened DXMAX=0.01 TRMAX=0.07 (not to run far away from it) shows that the optimizer jumps out from this point (although HESS=CALC), and never (at 100 steps) finds another point with the energy beyond or equal the starting one.
2) FROM OFFSET POINTS along the largest imaginary frequency’s mode (+/-100%; +/-10%,+/-1%) also didn’t find a minimum.

In all cases
$contrl inttyp=hondo icut=11 itol=30 $END
$scf nconv=6 $end
were used.

INTRESTINGLY, Energies in the offset points along the largest imaginary mode +/-1% are higher than in the saddle point. Why then this frequency is imaginary? As far as I can imagine this can happen only in case when the saddle point is surrounded with two very-very close (<1% along the imaginary mode) minima. And they are unlikely much lower in energy than the saddle point (otherwise the gradients around the saddle point along imaginary mode should be very high).
Taking into consideration that changes in energy during above mentioned optimization runs were within 10-4 au, the idea is disturbing me that to seek further for mathematical minimum in this situation has no sense (bearing in mind the accuracy of methods). But how can I count then vibrational frequencies (let alone to use such kind of results for publications)?

Could it be helpful in this situation:
1) Try other level of theory and/or basis;
2) Use BSSE correction (until now no BSSE correction is used)?

Could it be that the procedure to account for polarization of a solvent by charges on the cavity made by spheres around nuclei produces very unsmooth potential field in the cavity? Say, PES is like big water waves covered with many small ripples in different directions, producing abundance of local minima that in fact have no physical sense and are artefacts of such mathematical model of the solvent cavity surface. In connection with this does it make sense:

3) To experiment with increase in displacement sizes VIBSIZ (say on 1-2 orders of magnitude) for hessian calculation to neglect these ripples in this calculation;
4) Smooth out cavity surface using OMEGA, FRO and RET parameters?

By the way, I’m afraid there is an error in Firefly_input_rev002. It is stated that the default value for RET is 0,2. But judging from output files the default value is 100.

I’ll be grateful for any ideas.
Tatiana.

On Wed Feb 19 '14 11:49pm, Alex Granovsky wrote
-----------------------------------------------
>Dear Tatiana,

>Sorry for the delayed reply.

>The current status of PCM gradient code in Firefly is as follows.
>The PCM derivatives are partly analytical partly numerical.
>In most cases, the semi-numeric nature of gradients does not
>allow efficient geometry optimizations with tight gradient
>convergence criteria. I'd recommend to relax OPTTOL to at least
>3.0D-4 or above. In addition, one needs to increase the overall
>precision of computations, e.g.:

 $contrl inttyp=hondo icut=11 itol=30 $END
 $scf nconv=6 $end

>Hope this helps.

>Kind regards,
>Alex Granovsky

>P.S. It is impossible to change the default number of number
>of tesserae per sphere using input file.
>
>
>On Wed Feb 19 '14 1:19am, Tatiana Alieva wrote
>----------------------------------------------
>>Dear All,

>>First, Thank you to all Firefly developers for Firefly package!
>>Now about the issue. I’ve encountered a slow convergence of geometry optimization while using PCM. Water dimer (2 water molecules connected with an H-bond) were first optimized in gas (RHF 6-31G(d,p)). The stationary point was checked for minimum. Then the gas optimized dimer was optimized using PCM (either D-PCM or COSMO) and the same basis and level of theory. Reoptimization in PCM required 58 and over 200 iterations, correspondingly, (and in both cases ended in saddle points :( ). Usage HESS=CALC for D-PCM increased number of steps to >100. Attempts to optimize the complex first with a simpler basis (or smaller values of solvent epsilon– as an intermediate step between vacuum and water epsilon) and then do it in the target basis, level of theory, and epsilon showed to be also inefficient.

>>Since water dimer is the simplest system in my study, and I’m planning to calculate much bigger complexes in water, I’m getting worried about PCM geometry optimization for them.

>>Slow convergence of the optimization process and oscillations using PCM as far as I see is a well-known phenomenon. I’ve learned that Gaussian 98 enables users to specify the number of surface elements (tesserea) for each sphere and the area of the surface elements. Here
>>
>>is shown that these parameters can be of crucial importance for the geometry optimization efficiency.
>>Unfortunately, I haven’t found corresponding keywords in the Firefly documentation. Are there such tools in Firefly? If not, is it difficult to provide them to make Firefly more flexible?
>>Are there other suggestions how to speed up convergence in PCM? I will be grateful for any piece of advice.
>>Tatiana

[ This message was edited on Mon Feb 24 '14 at 9:23pm by the author ]

Mon Feb 24 '14 9:23pm

This message read 936 times