UT theoretical chemistry code forum

Professor Henkelman,

I've been running several NEB calculations on CO-saturated graphene-supported and free Pt13 nanoclusters to calculate the barrier for CO oxidation. I run 20 steps without climbing image and then I turn CI on and continue running iteratively. The problem is that after several hundreds steps and interchanging between the FIRE and QM optimizers, using 0.03 eV/A convergence criterion, relatively low K-Point mesh and ECUT, my calculations converge very slowly…

Any suggestions on how to speed up convergence?

Thank you,

Ioanna

In principle, you can switch to a second-order optimizer such as L-BFGS (IOPT=1), but this will only converge rapidly if you are near the minimum energy path. In my experience, problems with convergence are often related to other problems. If you post an example problem calculation, I can take a look at it.

Thanks,

I will email you one of my runs as an example.

Ok, or post it here or use dropbox or a web link to avoid email size limitations.

I believe I shared a folder with you using Dropbox.

Thank you in advance for looking at it!

That band looks reasonable. The forces have dropped from about 1 eV/Ang to 0.5 eV/Ang in the calculation that you sent me, over the course of 20 iterations. You could try IOPT=1 or 2 to speed things up.

Perhaps the most speed up you will get is by doing a gamma point calculation. I expect that will not change the geometry of the saddle point in any way. Once you have a converged band, you can always reconverge it at 3x3x1 to check. This change alone, will make your calculations 10x faster (/5 kpoints /2 for gamma point)!

Then, I would suggest relaxing more atoms, either in an NEB calculation, or using a dimer approach to reconverge the saddle. I like how you have restricted movement to find a saddle, but then you should relax things.

Thank you for your suggestions, Prof. Henkelman.

My plan is to unfix more atoms - actually all the Pt atoms and CO molecules, once I find a converged band,
however, how could I take advantage of the converged band in this case?

Since I will have to relax the initial and final states and interpolate again between the new initial and final images,
unfixing the atoms, wouldn't it be the same as running a new NEB calculation from scratch?

Thank you for your help,

Best,
Ioanna

Certainly not. It is much faster to start from a band that is close and reconverge it. That is true both for increase the k-point mesh, and relaxing atoms. The hardest part of relaxing a band is getting close to the minimum energy path. When you are close, the quadratic optimizers (IOPT=1,2), or a dimer calculation, should rapidly converge from a rough saddle to the nearby more-accurate saddle.

Thank you,

I very much appreciate your help!

Best,
Ioanna