UT theoretical chemistry code forum

Hi,
what is the experience with the FIRE optimizer? Is it as fast as claimed in conjunction with NEB as well?

Is it possible to use your optimizers with a regular VASP run (no TS search just ionic relaxation)?

Thanks

We're in the process of running systematic tests of the optimizers, both with regular minimization and TS searches. If you are adventurous, you can get the code from the anonymous cvs server. See the bottom of page:
http://theory.cm.utexas.edu/vtsttools/downloads/

So far, the FIRE optimizer looks promising. It seems to be systematically faster than quick-min (which it is closely related to). It also has quite a few extra parameters, which makes it harder to know if it is working near optimal. Overall, I think the conjugate gradient optimizer is faster overall, but it does require more accurate forces for the finite-difference line optimizer. Anyways, it would be interesting to hear about other people's experience.

With our code modification, you can use the built-in or new optimizers with regular vasp calculations. You can also use either set with the NEB. The only restriction is that you have to use our optimizers with min-mode following saddle point searches (dimer/Lanczos).

Hi,
I'm learning to use the CINEB method and the VTSTTools for diffusion in bulk system. When using the new optimizers ,QM and FIRE, they didn't behave well and quickly converge (even for IMAGE = 1 and EDIFFG = -1eV/A). Is the default paramters for the optimizers the best? Or there are some tips can be used to test the optimal parameters? Besides, do you agree that the optimal paramters is the same for one system in geometry optimization and NEB calculations?
Thanks for reading.

I recommend starting by looking at your initial path. Simply visualize the initial NEB and see if it makes sense. For example, are atoms running into each other? Are there problems with atoms moving in crazy ways?

Second, if your initial forces are high, start with a conservative optimizer such as QM with a small time step, such as 0.01. The forces should systematically drop to below 1 eV/Ang in tens of iterations. Again, look at the paths as they optimize and make sure that the atoms are doing what you expect. Make sure that the electronic structure is converging within the maximum specified iterations.

Once you have forces below 1 eV/Ang, you can switch to more aggressive optimizers and determine which is most efficient for your systems.

If there are still problems with getting forces below 1 eV/Ang, you can attach a .tar.gz file of the run (removing CHG* WAV*) and we can see what's going on.

Hi graeme:
First, I did visualize my diffusion path, and there were no wierd atoms except that the boundary atoms. But, I think they are the same atoms, if A atom is in position of x=1 and B atom is in Position of x=0, and after the nebmake process, there is no atom of x=0.5 in the middle image. Is that Ok for the NEB?
Second, I took your advice and used IPOT=3 and TIMESTEP=0.01. But the convergence was still hard and the force stayed at 2.64 from the 13th to 30th iterations. To avoid the problem of not converged electron steps, I increased the Ecut and EDIFF, which is given in the attachment. But, the force stayed at 2.63 in the 1 image scheme.
Third, I designed a 3 image scheme, and the force of the 2nd image stayed at 2.63 also. Likely, no hope to decrease further. Is the problem in the optimizer setup? If not, what's the possible problem?
Looking forward for your reply. Thanks for your patient reading.

You have IBRION=3 and POTIM=0, which is good; it tells vasp to do nothing with the ions.

But then you did not select any of our optimizers. You have the tag:

IPOT = 3

instead of

IOPT = 3

so nothing is happening in your simulations.

Hi graeme,
I'm so sorry for the silly mistake about "IPOT". As I have mentioned in the last message, because of the periodic boundary conditions, the boundary atoms A (x=0) and B (x=1) are the same atom. So, there is nothing to worry about its movement from one side of the box to the other side. Is that right?
Thanks for your patient reading.

right

Hi graeme,
I read your paper in 2008 about optimizers. In that paper, it mentioned that GL-BFGS(hess) is faster than L-BFGS, but the qualitative error exists.
(1) Is that means that GL-BFGS(hess) is not stable and for some systems a big error may happen?
(2) Is IOPT=1 stands for GL_BFGS(hess) optimizer? Or is other L-BFGS(hess) or L-BFGS(line) mentioned in your paper?
(3) If there are many local minimum positions in potential space. And the two end images are very close, around 1A. In this situation, is the first-order methods perform more efficient than the second-order methods? Bbecause the distance of the diffuion atom in reactant and product is short, should I lower the TIMESTEP and MAXMOVE at the same time?
Thanks in advance!

Hi Butterflyyh,

could you please also send me this paper?

Regards

Thomas