[Pw_forum] NEB Convergence Problem

[Tone Kokalj]

On Sun, 2011-11-06 at 18:33 -0500, Vic Bermudez wrote:
> Greetings All,
> 
> 	I'm struggling with a NEB calculation, and I wonder if someone can advise
> me. I first converged the path using NEB without CI. This took many
> iterations (810 to be exact), but it got there in the end. I then restarted
> with the optimized path and CI set to 'auto'.

No don't do that. First use the no-CI, just to "stabilize" the path
trajectory to some extent, and to get a rough idea on the reaction path.

So there is a question: "how do I know when it is stabilized enough". At
the beginning of NEB run, the highest energy image can change from
iteration to iteration, but when this gets stable, i.e., a particular
image displays the highest energy over several number of iterations, say
10-15, then you can stop the no-CI run, and start with the CI run.

Just as a rough guide, from my experience, I would say that when the
error is on the order of 0.5 eV/A, this is already good enough to plug
in the CI.


>  As you can see from the
> attached output, the convergence is extremely slow. In this case I know what
> the "right" answer is because I'm reproducing results in the literature. The
> forward activation energy should be about 1.65 eV, and at the current rate
> at which the activation energy is increasing it will take many thousands of
> iterations to get there. 

Well I don't think this is the reason. Your convergence is already quite
good, with almost all the images below 0.05 eV/A. Hence I don't think it
will ever climb to 1.65 eV. There must be some difference with respect
to what you are comparing to: either the initial state (IS) is different
or you located a new TS with lower energy.

However in your energy path, I see a lot of problems. As for
intermediate images, two of them appear as minima (3 and 6) and two of
them as TS (5 and 8). Hence your IS is not really the best minimum (if
you calculate the barrier wrt image 3, you get a value of 1.50 eV, which
is already closer to your reference value). 

This is also the primary reason for bad convergence: the NEB convergence
is best at describing only one elementary reaction, i.e., IS-->TS-->FS,
involving only one saddle point. Please decompose your path into several
separate calculations:

1. first do two optimizations by taking the structure 3 and 6.
Structure-3 will definitely result in lower energy structure than your
IS.

2. when you understand #1, then construct a new NEB run so that it
consists of only one elementary step (i.e. one saddle point).

Regards, Tone
-- 
Anton Kokalj
J. Stefan Institute, Jamova 39, 1000 Ljubljana, Slovenia 
(tel: +386-1-477-3523 // fax:+386-1-477-3822)

Please, if possible, avoid sending me Word or PowerPoint attachments.
See:  http://www.gnu.org/philosophy/no-word-attachments.html