NEB Energy fluctuating without trend towards convergence

[tp.joe]

Hi,

I am doing NEB calculation for a reaction of a rather complicated molecule on a fixed reaction surface. Preliminarily, I am using 3 images with LBFGS method. However, it seems like the molecule just moved back and forth so I used nebconverge.pl to generate the force. And as seen in the attachment, the energy of the first image doesn't seem to approach convergence, but fluctuating and increasing instead. Also, although the image 2 seems to slowly conveege and decrease, there are a lot of drastic spikes along the line. Image 1 seems to looks the best. I also attached INCAR, and an example of POSCAR, file just in case.

My question is, is this a typical behavior that you would expect from a complicated molecule?, will this behavior leads to convergence eventually? and what are some protocols to tackle this issues, particularly the spiking and converging larger molecules? I am currently trying to use other optimizers (FIRE), to see if it would make any difference, but not putting much hope on it.

Thank you in advance for your help,
Joe

[graeme]

Was this issue resolved?

[tp.joe]

Unfortunately, it does not. I have tried suggestions from other posts extensively, including doing damped MD with time step of 0.002, but it still fluctuates heavily with no tendency toward convergence. I have also tried using FIRE method. The results of both of these are attached.

[graeme]

If you are willing to attach the calculation or send it to me, I would like to take a look at it and see what's going on.

[tp.joe]

Yes please Prof. Attached is the detail of the calculation. Note that initially I used more than one picture for NEB, but the results were initially the same, so I reduced the number of picture to just the one with the highest energy (for quicker calculation) to see in a long run where does it lead to. Please let me know if there is any mistake and/or anything I can do to reach convergence.

Due to excessive size of the file in Damp_MD method, please download it from

https://drive.google.com/open?id=0ByHLt ... W05Z2NGa2c

instead.

Thank you Prof.

[pecher]

To me, the initial LBFGS calculations looks fine, but is just far from convergence. I've had the experience that for complicated systems (and this one is complicated!), the NEB calculation can sometimes take up to 2000 steps to converge. I also see the spikes in energy every few dozen steps and I imagine it has something to do with the Hessian being updated. What matters is the energy profile outside of those spikes and that seems to be still going down for image 1 and 3 and approximately staying constant for image 2.

Instead of doing a single-image NEB, I would recommend you to switch to the Dimer method (ICHAIN=2). You can easily generate the input files using the neb2dim.pl script. There you should use a much lower EDIFF (1E-07 or even 1E-08) and the calculation should converge to the saddle point eventually.

The single-image NEB results also look fine to me (albeit they seem to converge to different saddle points), again just far from convergence. I would recommend you to stick to LBFGS in NEB calculations. I have tested different optimizers for NEB calculations and this one was always by far the fastest. See also this very nice paper by Graeme's group on the performance of different optimizers: http://theory.cm.utexas.edu/henkelman/p ... 134106.pdf

What puzzles me is that the total energy is at -315 eV while for your initial NEB calculation they seem to be at -345 eV. Did you change some of the computational parameters? The structure doesn't look like something 30 eV above minimum to me.

In summary: Just be patient and continue OR switch to the Dimer method.


Cheers,
Josua

[pecher]

I have attached the vaspout file of one image in one of my calculations (LBFGS optimizer). I picked an image in a critical part of the potential energy curve. You see that energy and forces oscillate all the time, but the average energy goes down until it reaches convergence after approx. 1300 steps.

[tp.joe]

Thank you Joshua, I will do it longer. And yes I did change a system a bit to make it a little simpler molecule.

Also, I have tried Dimer method before using NEB image as the starting point, but the forces keep increasing with out decreasing trend at all, so I stopped. Did you have the same issue in a complicated system as well?

[graeme]

I see a couple of problems with the calculation in the FIRE tar file.

First, for a complex reaction, you will not be able to use a single image in an NEB calculation. A good definition of the reaction coordinate is required for convergence of the climbing image to the saddle.

Second, there are no frozen atoms in the 01/POSCAR file. You must set all atoms to be frozen in the same position for each image along the band.