High saddle point Energy caused slow NEB convergence？

[ross]

Dear everyone：
Recently I was confused a lot by my NEB performance. Below is one of my IMAGE force.
FORCES: max atom, RMS 0.081813 0.041618
FORCES: max atom, RMS 0.081572 0.041002
FORCES: max atom, RMS 0.081090 0.040789
FORCES: max atom, RMS 0.082272 0.041200
FORCES: max atom, RMS 0.081963 0.040259
FORCES: max atom, RMS 0.082376 0.041074
FORCES: max atom, RMS 0.082210 0.040819
FORCES: max atom, RMS 0.082831 0.040742
FORCES: max atom, RMS 0.083369 0.040679
FORCES: max atom, RMS 0.083434 0.040673
FORCES: max atom, RMS 0.084609 0.040759
FORCES: max atom, RMS 0.084396 0.040847
FORCES: max atom, RMS 0.085675 0.040965

FORCES can not converge to the setted criteria 0.05. Also I attach part of my INCAR related to NEB.
Ts search
SPRING = -5
LCLIMB = .TRUE.
IMAGES = 8
IBRION = 3
#SMASS = 2
POTIM = 0.005
#LSCALAPAK = .FALSE.
LPLANE = .TRUE.
NPAR = 8
LSCALU = .FALSE.
NSIM = 4
#IOPT = 2
I also tried "IBRION=3 POTIME=0.01" it nearlly has the same result. And "IBRION=3 POTIM=0 IOPT =1 or 2" which is expected to converge the result when FORCES are samll, howerever makes the FORCES larger and larged. Why was that? Because the high migration Energy as large as 1.5ev, or I have to use more IMAGES？I hope someone may help me with it.

By the way , in the paper "J. Phys.: Condens. Matter 20 (2008) 064211 (9pp)",the author said that they used the metadynamic implemented in the code VASP to calculate the free energy reaction barrier which take the temperature into account. I was confused since I can not find anything about metadynamic in VASP's manual. How can the author make that？
Also I wonder that why NEB code developer(Professor Henkelman's group) write some TS search code that could be used with ab initio MD in VASP, since as far as I know the NEB has something in common with the Umbrella Sampling method which is used in some MD TS search.
Best Wishes
Ross

[graeme]

Using IBRION=3 and POTIM=0.005 is a very conservative setting. Things will change only very slowly. Using a POTIM value of 0.1 would be more typical, or the faster second-order algorithms that you mentioned.

The force on each image will not always drop in an NEB calculation. The images are relaxing collectively so that some can be dragged to a higher force region during convergence.

There is no evidence that more images are needed. If you have our scripts installed, try running nebresults.pl to get a summary of the force and energy at each image. My guess is that you are not converged and looking at an artificially large barrier.

Metadynamics has not been implemented in the standard version of vasp.

I can't decipher the last sentence. The NEB finds energy barriers. It is not a sampling method. The only connection to dynamics that I can think of is that you can used a damped dynamics algorithm to optimize it.

[ross]

Thanks for your reply, professor. In fact first I use "IBRION=3 and POTIM=0.01", the NEB nearly converged to my criterion, but it was just close, can not reached even running for many steps. Result(nebbef.dat) by run nebresult.pl is as follows:
0 0.00000000 -540.60921800 0.00000000
1 0.02128900 -540.55727500 0.05194300
2 0.04330500 -540.43880600 0.17041200
3 0.09133800 -540.13776600 0.47145200
4 0.08303200 -539.50620500 1.10301300
5 0.04825200 -539.16487200 1.44434600
6 0.05730100 -539.70215000 0.90706800
7 0.02887700 -540.33266800 0.27655000
8 0.01642000 -540.58323700 0.02598100
9 0.00000000 -540.66757900 -0.05836100
the second column should be the forces. Then I switched to IOPT =1 ,the forces vibrate to a large value. So I want to know weather it is related to the large barrier.

As the last sentence of my former topic, I refer to that" the biased potential" used in umbrellar sampling just like the "spring forces" in the NEB.
Best Wishes
ross

[graeme]

That looks actually fairly well converged. Using a larger POTIM with IBRION=3 should speed up the final convergence. Also, the IOPT=1 should be suitable, perhaps with a lower INVCURV value such as 0.002. If the IOPT=1 optimizer is not working, we would be interested to look into it and see what's going wrong.

[ross]

graeme wrote :
> That looks actually fairly well converged. Using a larger POTIM with
> IBRION=3 should speed up the final convergence. Also, the IOPT=1 should be
> suitable, perhaps with a lower INVCURV value such as 0.002. If the IOPT=1
> optimizer is not working, we would be interested to look into it and see
> what's going wrong.
Dear graeme:
In the past few days, I have tried the IOPT=1, hoping for the final convergence, however, it semms impossible. Here are some results of the FORCES in several IMAGES:
IMAGE 02 :
FORCES: max atom, RMS 0.034319 0.019566
FORCES: max atom, RMS 0.058612 0.031156
FORCES: max atom, RMS 0.133011 0.075076
FORCES: max atom, RMS 0.096282 0.056843
FORCES: max atom, RMS 0.089895 0.047375
FORCES: max atom, RMS 0.083439 0.034487
FORCES: max atom, RMS 0.120520 0.042008
IMAGE 03:
FORCES: max atom, RMS 0.074223 0.037742
FORCES: max atom, RMS 0.150202 0.055794
FORCES: max atom, RMS 0.303425 0.086255
FORCES: max atom, RMS 0.201707 0.070798
FORCES: max atom, RMS 0.181551 0.062243
FORCES: max atom, RMS 0.229307 0.059238
FORCES: max atom, RMS 0.346494 0.072348
IMAGE 05:
FORCES: max atom, RMS 0.061746 0.028559
FORCES: max atom, RMS 0.099103 0.035366
FORCES: max atom, RMS 0.123791 0.048295
FORCES: max atom, RMS 0.120389 0.042994
FORCES: max atom, RMS 0.118182 0.040148
FORCES: max atom, RMS 0.115461 0.036341
FORCES: max atom, RMS 0.117877 0.036425
In fact, FORCES in all IMAGES are larger and larger. I am afraid I have to say that for my system, the FORCES can only converge to around 0.08ev/A. By the way, can the criterion of 0.08ev/A be used to estimate the migration energy with a quite small error.
Best Wishes
ross

[graeme]

If you mail a .tar.gz file of this calculation, I would like to see what's going wrong.