I will presume that the TS has a quite small imaginary frequency at wb97xd/3-21G, i.e. the potential energy surface is indeed quite flat in the region near the TS.  With the larger 6-31G(d) basis set, the surface may get even flatter and the optimizer just falls off the saddle point region and can't find its way back.

In such cases, I've had success sometimes from reducing the step size including IOP(1/8) in the keyword line. The G09 default is IOP(1/8=10); troublesome cases have been solved sometimes with IOP(1/8=5) or so; I know I have had to use IOP(1/8=2) in a few cases. Many, many additional optimization steps may be required but if it works, hey so what?---cpu cycles are abundant nowadays. Also, recalculating the FC's with the new basis set at the start of the reoptimization run is often essential. Do a FREQ calculation with the 6-31G(d) basis set and check that the imag freq is still there; then use opt=(ts,rcfc,noeigentest) to restart the opt with the larger basis set.

It is also essential that the TS transition vector is being expressed in the 'right' internal coordinates. If the TS transition vector definitely (your best opinion) must involve, for example, a particular distance between two atoms, then it is essential to have that distance in the calc as an internal coordinate. Sometimes, G09 uses a lot of angles and dihedrals because it 'misses' seeing the 'obvious' large contributor reaction coordinate. You can fix that using 'modredundant'......build the essential internal coordinate and, perhaps, give that coordinate a negative diagonal force constant in the (initial) Hessian. That way, the TS search will definitely include the 'important' coordinate from the start.

For argument's sake, let's assume that the (often long and therefore 'missed' in the G09 modified internal coordinate competition) distance between atoms 10 and 15 should contribute significantly to the TS transition vector.  The keywords etc could look like:

#wb97xd/6-31g(d) opt=(ts,rcfc, noeigentest,modredundant) iop(1/8=5) geom =check etc. etc.

title

0 1

blank

10 14 B

10 14 H -0.05

blank

Long answer, sorry----hopefully, some of it may be of use to you.

kk-j

TS search is always critical and it may depend on basis set. It is not always true that program will able to locate TS from reactant and product geometry. Improvement of your guess geometries in TS input might help.  

Thanks for your response. I appreciate any thoughts. As I said originally, I'm not after help in particular just some opinions. 

I've got a 16 step mechanism to chart in Gaussian - it's hard work! I've been able to find each TS step using 6-31G(d) from two optimised 3-21G guessed reactant and product geometries. However, I've come across a very troublesome transition state. I can find the TS using 3-21G, but if I use the force constants and/or geometry from that low level and plug them into 6-31G(d), the TS disappears! 

I also follow a similar protocol for minima and TS optimization and have not experienced this issue ..what else are you adding in your route section?

have you tried using the geometry and starting afresh with the FC?

Hi,

I have essentially tried a few routes:

wb97xd/6-31G(d) opt=ts,readfc geom=allcheck

Dear Anthony,

It can happen that you find a TS at low basis set while with a higher one, you can not. The TS with DFT method can be an "artifact of the method". If you check a computational studies about TS, you will see that it is usually that the TS are tested, with other methods such MP2 and/or with higher basis set in order to be sure that the TS found it is "real".

On the other hand, I think that it is extremely important (in general) to add diffusion and polarization to the basis set. In some imaginative way, the TS structure has a higher delocalized cloud of electrons, the bonds are weaker, etc, so, by using a "atom-located" orbitals, you are not defining correctly all the space that you are interested in.

Thank you for that insight, Joaquim. 

I will be honest with you, I haven't seen anyone check their TS results using methods such as MP2. But then that maybe because I am looking at a very specific field of study. How would you go about checking your TS results using MP2? Is it simply a case of reperforming a TS search? Surely if that is the case there would never be the need of using DFT in the first place. My other concern is that such an approach would double my workload. This wouldn't be a problem if my mechanism was only two or three steps, but it is probably around 16 to 20 TS steps. 

Thanks

I will presume that the TS has a quite small imaginary frequency at wb97xd/3-21G, i.e. the potential energy surface is indeed quite flat in the region near the TS.  With the larger 6-31G(d) basis set, the surface may get even flatter and the optimizer just falls off the saddle point region and can't find its way back.

In such cases, I've had success sometimes from reducing the step size including IOP(1/8) in the keyword line. The G09 default is IOP(1/8=10); troublesome cases have been solved sometimes with IOP(1/8=5) or so; I know I have had to use IOP(1/8=2) in a few cases. Many, many additional optimization steps may be required but if it works, hey so what?---cpu cycles are abundant nowadays. Also, recalculating the FC's with the new basis set at the start of the reoptimization run is often essential. Do a FREQ calculation with the 6-31G(d) basis set and check that the imag freq is still there; then use opt=(ts,rcfc,noeigentest) to restart the opt with the larger basis set.

It is also essential that the TS transition vector is being expressed in the 'right' internal coordinates. If the TS transition vector definitely (your best opinion) must involve, for example, a particular distance between two atoms, then it is essential to have that distance in the calc as an internal coordinate. Sometimes, G09 uses a lot of angles and dihedrals because it 'misses' seeing the 'obvious' large contributor reaction coordinate. You can fix that using 'modredundant'......build the essential internal coordinate and, perhaps, give that coordinate a negative diagonal force constant in the (initial) Hessian. That way, the TS search will definitely include the 'important' coordinate from the start.

For argument's sake, let's assume that the (often long and therefore 'missed' in the G09 modified internal coordinate competition) distance between atoms 10 and 15 should contribute significantly to the TS transition vector.  The keywords etc could look like:

#wb97xd/6-31g(d) opt=(ts,rcfc, noeigentest,modredundant) iop(1/8=5) geom =check etc. etc.

title

0 1

blank

10 14 B

10 14 H -0.05

blank

Long answer, sorry----hopefully, some of it may be of use to you.

kk-j

Dear Karsten,

That was a fantastic reply. Thank you so much for that information on introducing a negative contribution to the vector corresponding to the internal coordinate involved in the reaction. I *think*, I've just found the TS I wanted (I'm performing an IRC as we speak), but if that fails I will have a go at introducing the negative mod red contribution.

Many thanks again. 

Dear Anthony,

I am agree with you, MP2 can be prohibitive with very large systems, and depending of the size, the approximations are very different (taking into account that I usually work with small systems). Even so, and referring also to the first paragraph of the Dr. Karsten, some reaction are not trivial and the methods can produce artifacts, highlighting metals and semimetals compounds. There are papers where they have found TS structures with DFT but not with MP2, and conversely. Sometimes, it can happen with the basis set, where with a small basis set you can detect the TS but, by increasing basis set, the grid, etc... you finally can find it.

As an example, in a document I made time ago, here in researchgate, I calculated the TS of a Sn2 reaction with different methods at aug-cc-pvDz and MP2 gave 63.7 kJ/mol, while B3LYP gave 42.7 kJ/mol. Of course, in this case it do not care because the objective was to show that we must be very critic with methods, but in a real investigation, where you do not know how the system will proceed or without reference values nor studies to support your results, it is not trivial to be sure about your TS, even with IRC.

Of course, it is my opinion and I like to be critic with each results that I obtain, but each problem requires different solutions with different sacrifices.

Dear Joaquim,

Many thanks for that information. It is very good to have these kinds of discussions. After three years of using Gaussian I am very much past "how do I do XYZ...", and more at the stage of discussing the methodology. 

Thanks again.