UT theoretical chemistry code forum

Hello,

I have some NEB calculations involving floppy molecules on a surface producing strange and circuitous trajectories. I was glancing through JCP 128 134106 "Optimization methods for finding minimum energy paths" where you discuss the different techniques that are available in VTSTTools, when I saw the section regarding DNEB. Based on this paper, it would appear that DNEB is useful as a tool to anneal the initial images into a reasonable trajectory guess, but that the double nudging causes poor convergence behavior. I saw no mention about the "modified" double nudging that is enabled when the LDNEB flag is set to .TRUE. What is the proper way to use double nudging? Do I want to run a trajectory with LDNEB enabled and a very loose convergence criterion (EDIFFG = -0.5) followed by regular NEB? Or does LDNEB enable what is referred to in the paper as "swDNEB", where the double nudging is smoothly switched off during convergence?

Do you have any other tips for difficult-to-converge NEB calculations? I've been running these same calculations repeatedly changing the optimizer (GL-BFGS, L-BFGS, FIRE) and enabling or disabling climbing image, and each time I get a different absurd trajectory which involves bonds breaking and being formed between nearly every image. Should I try changing SPRING? If so, what is a reasonable value to use for preventing images from diverging too wildly?

Thanks for your help,
Eric Hermes

Hi Eric,

I think that two issues could be separated here.

1) The DNEB method was developed by the Wales group to deal with long paths which typically have several intermediate minima. For those paths, the double-nudging keeps the path straighter than the regular NEB. Since it is critical to maintain a smooth path, this is helpful during the initial stages of convergence. One problem with the original DNEB is that it does not allow for tight convergence. This is not a problem for the Wales group because they typically switch to a single ended mode follow method once an image is in the region of a saddle. In the comparison that you found in JCP, we incorporated a switching function that would smoothly turn of the double nudging force and allow for convergence. This is what is incorporated into vasp.

2) It is my experience that issues related to NEB convergence are not typically related to specific optimizers or details of the methods. Often, they are related to more qualitative problems in regards to how the path is set up, or the parameters in the INCAR file. If you post an example failed calculation (without CHG*, WAV*) I can see if there is an obvious problem.

As far as debugging problems, I would always look at the initial path. Check to see if any atoms run into each other. After one iteration, check to make sure that the forces are reasonable (<20 eV/Ang). Then, you can try any first-order optimizer (e.g. IOPT=3 or IBRION=3) and use a small time step of about 0.1 and check for systematic (and slow) convergence. If that works, you can try more aggressive optimizers to improve the rate of convergence. If you see evidence of intermediate minima, converge them and set up new NEB calculations going between neighboring states.

Professor Henkelman,

Thank you for responding to my post. Here is an example of one of my pathological NEB calculations. Note that the tarball is 94MB, and will expand to about 900MB when uncompressed. There are no WAVECAR, CHGCAR, or CHG files, but the OUTCAR files are about 100MB for each image. Also note that the product OUTCAR in 09 does not exactly correspond to the final image used in the NEB.

https://www.dropbox.com/s/q1mpf99fgr3b4 ... r.bz2?dl=0

Edit: Another thing that may be worth mentioning is that I am using the Grimme D3 correction (ivdw=11), but I have patched VASP to calculate the three-body Axilrod-Teller-Muto contribution to the energy and forces. This term is defined in the original DFT-D3 paper, but is unimplemented in most periodic implementations of DFT-D3.

Thank you for your assistance,
Eric Hermes

It looks to me as if your reaction mechanism will be significantly more complicated than you assumed. Specifically, there is a very long hydrogen transfer demanded by the choice of initial and final states, and the identity of the hydrogen atoms. Since there is likely a different mechanism, or a set of mechanisms for this reaction, I suggest finding other minima and trying to break up your band into elementary steps. So you can, for example, minimize the intermediate minima found in the band that you posted. See what the first reaction step is, and then decide if that is appropriate and use an NEB calculation to identify that barrier. Then proceed along the path. I expect, as you do this, that you may relabel some H atoms to get a lower barrier mechanism.

I do not recommend trying to use any fancier methods such as DNEB for this. That works nicely with empirical potentials and long paths, but these DFT calculations are typically too expensive for that approach. It makes more sense to provide some guidance to the algorithm by looking at intermediates and finding the appropriate mechanism.