From owner-chemistry@ccl.net Fri Aug 19 13:33:01 2011
From: "Steve Williams willsd(_)appstate.edu" <owner-chemistry:_:server.ccl.net>
To: CCL
Subject: CCL:G: What is wrong with M06 vibrational frequencies
Message-Id: <-45297-110819103147-24279-HszkON+GZhAiC4DVQgaY2A:_:server.ccl.net>
X-Original-From: Steve Williams <willsd[#]appstate.edu>
Content-Transfer-Encoding: 7bit
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Date: Fri, 19 Aug 2011 10:31:25 -0400
MIME-Version: 1.0


Sent to CCL by: Steve Williams [willsd() appstate.edu]
On 08/19/2011 07:55 AM, Gerald Knizia knizia]*[theochem.uni-stuttgart.de 
wrote:
>
> Sent to CCL by: Gerald Knizia [knizia\a/theochem.uni-stuttgart.de]
> Adel El-Azhary azhary---ksu.edu.sa wrote:
>> I am not sure that anyone of you experienced this problem before or 
>> knows a solution for it. I am calculating vibrational frequencies at 
>> the HF, B3LYP, M06, M06L, M062x, M06HF and MP2 levels of a molecule. 
>> All vibrational frequencies are real at the HF, B3LYP and MP2 levels 
>> but at the M06, M06L, M062x and M06HF levels, one or more vibrational 
>> frequency is imaginary.
>
> The M06 (and the other M* functionals) are basically fitted to 
> death[1] on a large training set in order to give good thermochemistry 
> values. They are not fitted on potential energy surfaces. While in 
> general this apparently works well, you have to be wary. It's only to 
> be expected that they break for some molecules and applications which 
> are sufficiently different from what they are fitted for[2]. If you 
> are sure that your geometries with the different methods converged to 
> equivalent structures, you maybe shouldn't worry about that too much 
> and simply use some other theoretical method.
>
> [1] They have over 30 fitting parameters, additionally to the choice 
> of the functional form.
> [2] ...or which in general cannot be reliably represented by DFT 
> methods, like dispersive interactions or strong correlations. Maybe HF 
> and B3LYP anr just as wrong, but in a different way, which causes them 
> to erroneously report positive PES curvature. To be absolutely sure 
> about that you would need to repeat your calculation with a more 
> accurate wave function method, but this might be unaffordable.
This is a good point.  If these M* functionals are "broken" for anything 
but thermochemistry at single points, you might be able to have 
gradients and second derivatives "broken" in the same way.  I think you 
are using gaussian for these calculations, so (if you can afford it), 
you might want to try opt=(calcall,tight) with your ultrafine grid.  
This will use the M* analytic Hessian to guide the optimization at every 
step.  This is guaranteed to be really slow, but will take fewer steps 
in the optimization.  There is no need for a subsequent freq calculation 
since this automatically reported on the converged geometry.
Steve Williams