UT theoretical chemistry code forum

Dear Sir,

I am currently starting up to use the CI-NEB method for the first time, where I'm looking at vacancy diffusion around substitutional defects in bcc Fe. I looked through previous threads in the forum and found that the guidelines to start with 3-5 images for simple reaction paths. So I simply selected the lower limit and used 3 images, since I believe the path to be quite straightforward (although I might of course be mistaken). I then run a CI-NEB calculation and my band seems to converge to something that at least looks reasonable. However, I would now like to investigate if my reaction path is sufficiently resolved, so I would like to ask two questions:


1. Is there any general guidelines to confirm that the path is sufficiently resolved? Should I always add more images in order to confirm this?

2. In my cases the images are separated by 0.5-0.7 Angstrom when I use 3 images. If I would like to add more images after the 3 images band has converged (with the 2nd image at the saddle point), would it be better to

a) linearly interpolate a new image at each spacing thereby going from 3 to 7 images,
or
b) add a single new image on each side of the saddle and thereby going from 3 to 5 images?

The issue I'm thinking of is that if I go with option b), then the two images on each side of the saddle must be re-optimized to provide an equal spacing between the images. While if I go with option a) I still need to optimize two images (and possibly also the preconverged image in the middle), but the spacing between the images are equidistant from the start. Can I conclude that option a) will give less computational effort per image?

---


In addition to the two above question, I also have a question connected to the available optimizers. I started with the QuickMin algorithm in VASP (IBRION = 3, POTIM = 0.1) to decrease the forces down to 1 eV/Angstrom. After this stage I'm now thinking about switching to LBFGS with the default settings. However, under the information for the optimizers I found the following

''We recommend using CG or LBFGS when accurate forces are available. This is essential for evaluating curvatures. For high forces (far from the minimum) or inaccurate forces (close to the minimum) the quick-min or FIRE methods are recommended.''

which I do not quite understand. What is meant with that the ''inaccurate forces (close to the minium)'' statement? Will the forces always be inaccurate when the band is approaching the MEP? Or is it referring to the forces acting on the images next to the initial and final state?

So does the above statement mean that I should converge my band with LBFGS until I'm very close to the MEP and then switch to say the FIRE method for the final part? Or can I expect the LBFGS to behave well even when the band is close to the MEP?

I have tried to make the evaluation of the forces as accurate as possible by using the following INCAR (for the QuickMin case)

-------------

SYSTEM = Fe126X

ENCUT = 700 eV
ENAUG = 700 eV
EDIFF = 1E-8
EDIFFG = -0.01
LREAL = A
PREC = Accurate
ROPT = -0.0001 -0.0001
ADDGRID = .TRUE.
ISMEAR = 1
SIGMA = 0.1

POTIM = 0.1
IBRION = 3

ISIF = 2
NSW = 100
MAGMOM = 126*4 0
ISPIN = 2
ALGO = Fast

IMAGES = 3
ICLIMB = 0
SPRING = -5
LCLIMB = .TRUE.
LTANGENTOLD = .FALSE.
LDNEB = .FALSE.

-------------



I appreciate any comments that you have time to spare. It would be really helpful so that I can get properly started.


Best regards,
/Dan Fors

When you use the climbing image, the highest energy image will try to climb right to the saddle point. As long as you have enough images to resolve the reaction pathway near the saddle, the climbing image will rigorously converge to the saddle. Then there is no need to add more images. You do not need to check for convergence of the saddle with respect to the number of images.

It is possible that if you have too few images, the climbing image will not converge. Only then do you need to add more.

One reason that you might want to have more images is to better resolve the entire pathway. For example, you may not notice an intermediate minimum along a complex pathway if you have only 3 images. You could, as you suggest, double the number of images. (Adding images only near the saddle is not a good strategy for resolving the entire path.) A cheaper way is to take your converged saddle and minimize after making a small displacement forward and backwards along the path. You can then check if these minimizations end up in your initial and final states, or in an intermediate minimum.

------

The second question is about the optimizers. Quickmin is a first-order optimizer which is relatively stable at any point during optimization. Near the minimum where the potential can be expanded using a harmonic approximation, it is not as efficient as second-order optimizers such as CG or LBFGS.

There are a few caveats. CG and LBFGS are not particularly stable far from the minimum where forces are high. So using quick-min until the forces drop below 1 eV/Ang is a sensible way to start. You can also make sure that the path at this point makes physical sense. Then it will be more efficient to switch to a second order method.

When I talk about forces being accurate enough near the minimum, I'm referring to the errors in the forces which are controlled by the electronic convergence criteria, ediff. When the forces are larger than the errors, curvatures can be determined with sufficient accuracy. Near the minimum, the errors are a higher portion of the force. At this point, you can either go back to quick-min (or fire), or decrease ediff. Your value of ediff=1e-8 is extremely tight. Using a value of 1e-5 will generally speed up your calculations without causing problems unless you are trying to get forces down to ~0.05 eV/Ang.

Also, as a side note, your cutoff is very high and I have never seen a case where you need to use prec=accurate for the calculation of a barrier.

Dear Sir,

Thank you for your quick and detailed reply. It was really helpful and cleared all my question marks. I for now know how to proceed. I will also take your advice and loosen my accuracy parameters (ediff, prec, encut) and (pre)converge the band with these less computational demanding settings. If necessary I will instead increase them again after the first convergence round.

Thank you for taking the time to comment on this matter.

Best regards,
/Dan Fors