Dear all,
I would like to ask a question about the CI-NEB code. If there are several maxima and minima in the pathway, how does the CI method work? Will it move the closest image to the extremum for all extrema, or only for the one highest in energy, or how is this choice made?
Thanks a lot, Berit
Several maxima and minima with CI-NEB
Moderator: moderators
The climbing image is only used for the highest energy image along the path. In principle, the algorithm could be applied to any local maximum (and/or minimum) along the path, but we have done this with the implementation in vasp. Usually when an intermediate minimum is found, it's easier to minimize the geometry as a separate calculation, and construct a new neb for the second reaction barrier.
However, if it would help to rigorously converge to all extrema along the minimum energy path, we would be happy to add a flag for it.
However, if it would help to rigorously converge to all extrema along the minimum energy path, we would be happy to add a flag for it.
Dear Andri and Graeme,
thanks for your answers. I had already thought about splitting up the NEB pathway and do each piece separately and this is probably the best solution. I was just curious about how the implementation is in VASP.
So do I understand it correctly that the extrema with the highest energy is determined and then the image closest to it in reaction coordinate is moved there?
You are probably right that if there are local minima, one should look at the path more closely instead of generically converge to all extrema, but I am not sure.
Another question, if one wants to refine a particular piece of the NEB path (around the saddle point), does it make sense to do a NEB calculation only between two points which are not local minima? I am not sure.
Thanks a lot, Berit
thanks for your answers. I had already thought about splitting up the NEB pathway and do each piece separately and this is probably the best solution. I was just curious about how the implementation is in VASP.
So do I understand it correctly that the extrema with the highest energy is determined and then the image closest to it in reaction coordinate is moved there?
You are probably right that if there are local minima, one should look at the path more closely instead of generically converge to all extrema, but I am not sure.
Another question, if one wants to refine a particular piece of the NEB path (around the saddle point), does it make sense to do a NEB calculation only between two points which are not local minima? I am not sure.
Thanks a lot, Berit
It's actually the highest energy image which is moved up in energy to the saddle point. This would typically be the image which is closest to the extrema of highest energy, but the two definitions are not the same. (I think they are the same only for a harmonic potential at the saddle).
I have never been a fan of doing short bands between points that are not minima. It seems dangerous to me, and not the best way to find the saddle. That said, several groups do use this approach, and I think it was even being added to dacapo.
The only reason I can see someone wanting to start a short neb is to find the saddle with fewer images. The climbing image takes away the need to increase the density of images near the saddle -- there only needs to be enough points in the path to resolve the reaction coordinate at the saddle.
In order to start a short neb between two non-extrema points, you must have some confidence that these points are on the minimum energy path, and define the reaction coordinate. If this is the case, you could use a drag type method (or a single climbing image between the endpoints) to find the saddle. But the problem is that you don't usually know that you can trust images along the band until the neb converges. If the band is draped over a shoulder in the potential, running a short neb will prevent the neb from discovering the true minimum energy path. I think we have all run NEBs that appear to be converging to a high saddle, only to break up into two mechanisms. This is a strength of the method which should not be eliminated by jumping to short bands.
Also, as Andri already mentioned, if you have a guess at the saddle point from a roughly converged neb, it is better to use a minimum-mode following method (such as the dimer or lanczos methods) in which the lowest curvature mode is found and followed. These methods also use fewer images, and there is no bias in the reaction coordinate -- it is fully optimized during convergence by finding the lowest curvature direction.
I have never been a fan of doing short bands between points that are not minima. It seems dangerous to me, and not the best way to find the saddle. That said, several groups do use this approach, and I think it was even being added to dacapo.
The only reason I can see someone wanting to start a short neb is to find the saddle with fewer images. The climbing image takes away the need to increase the density of images near the saddle -- there only needs to be enough points in the path to resolve the reaction coordinate at the saddle.
In order to start a short neb between two non-extrema points, you must have some confidence that these points are on the minimum energy path, and define the reaction coordinate. If this is the case, you could use a drag type method (or a single climbing image between the endpoints) to find the saddle. But the problem is that you don't usually know that you can trust images along the band until the neb converges. If the band is draped over a shoulder in the potential, running a short neb will prevent the neb from discovering the true minimum energy path. I think we have all run NEBs that appear to be converging to a high saddle, only to break up into two mechanisms. This is a strength of the method which should not be eliminated by jumping to short bands.
Also, as Andri already mentioned, if you have a guess at the saddle point from a roughly converged neb, it is better to use a minimum-mode following method (such as the dimer or lanczos methods) in which the lowest curvature mode is found and followed. These methods also use fewer images, and there is no bias in the reaction coordinate -- it is fully optimized during convergence by finding the lowest curvature direction.