Page 1 of 1

converge slowly

Posted: Wed May 31, 2006 1:16 am
by hlzya
Hi, evergyone.

I am now calculating CH4 dissociation on Ni(100).When I use the CI-NEB method to find the MEP, the calculation is converged very slowly.

Initially, I use the NEB method without Climbing Image and I use IBRION=3 and POTIM=0.01 for several steps (e.g. 5). Then I copy CONTCAR to POSCAR, and use IBRION=1, POTIM=0.1 to converge the maxmum force in every degree of freedom to within 0.01 eV/A. Finally, the CI-NEB method is activated and IBRION=1, POTIM =0.1 are still used. It always took me several hundred steps to converge the calculation.

I wonder what is wrong with my calcultion procedure and if there is some skill using CI-NEB method.

Thanks.

Posted: Wed May 31, 2006 3:03 am
by graeme
I think it's fine to do a few steps using a small time step, but I recommend turning on the climbing image right afterwards. I don't think it makes sense to converge the neb without climbing image and then turn it on.

Also, if you are willing to try a new method, we would be very interested to see if a double nudging scheme (from Wales' group) improves the convergence using IBRION=1. You can get an updated neb.F with this modification from cvs using the command

cvs -d :pserver:anonymous@theory.cm.utexas.edu:/Groups/cvs login
cvs checkout vtsttools

and recompile vasp. Use the password 'anonymous' when prompted. You can turn on the double nudging by adding the following flags to the INCAR file:

LDNEB=.TRUE.
LDNEBNEW=.TRUE.

The second flag turns on a modified version of the double nudging which we've found to work better with bfgs (ibrion=1). If you find that this improves convergence times for your CH4/Ni(100) run, we would be interested to know about it.

Posted: Thu Jun 01, 2006 1:18 am
by hlzya
Thank you very much for your reply.

I would like to try your suggestion. But now I am using a Windows OS, and can you give me a ftp or http address where I can get the updated neb.F?

Best

Posted: Thu Jun 01, 2006 3:24 am
by graeme
Sure, I've updated the vtstcode.tar.gz file on the page:

http://theory.cm.utexas.edu/vtsttools/code/

There are also any number of free programs that allow you to access cvs servers through windows, both graphical and from the command line.

Posted: Thu Jun 01, 2006 4:33 am
by hlzya
ok. I get it.

When I use the double nudging sheme, I wonder if I must set LCLIMB = .FALSE. in the INCAR file.

Now, what is included in my INCAR file for the double nudging scheme are:

IMAGES = 6
SPRING = -5
LDNEB=.TRUE.
LDNEBNEW=.TRUE.

I wonder if these tags are enough.

Best

Posted: Thu Jun 01, 2006 1:58 pm
by graeme
You can turn on the climbing image with the double nudging. I suggest adding the tag

LCLIMB=.TRUE.

It would be interesting to start your calculation after you did a few IBRION=3 steps, and compare how long it takes to converge using the climbing image and double nudging, with your original calculation.

Posted: Mon Jun 05, 2006 12:47 am
by hlzya
Hi,

I have tried the new method. Now, the calculation still can not converge after sixty ionic steps. I have used nebef.pl and the results are:

1 0.01753400 -254.52972900 0.01544300
2 0.03279600 -254.48991500 0.05525700
3 0.06137800 -254.37285500 0.17231700
4 0.13588800 -253.65668600 0.88848600
5 0.05486500 -254.05353600 0.49163600
6 0.02485600 -254.29669500 0.24847700

I think the POTIM=0.35 is not too large according to the forces, right?

I know you have calculated the Ni(111)/CH4 system and can you share your experience with me?

Posted: Mon Jun 05, 2006 6:13 am
by graeme
A POTIM of 0.35 is larger than I usually use, but it may not be the problem.

To sort out what's going on, it would help to look at your run. If you are willing to send me a .tar.gz file of your 60 ionic step run, I'll take a look at your settings, and see if I can make any suggestions.

Posted: Mon Jun 05, 2006 7:39 am
by hlzya
I have sent my calculation files to your Email address. Thanks

Posted: Mon Jun 12, 2006 1:05 am
by hlzya
Hi,

Now the maxmum force in every degree of 04 image (the saddle point) can be converged to within 0.01eV/A. But the force on 02 image still oscillates at 0.024 ev/A. I wonder if I can determin the transition state. (I have decreased the POTIM to 0.1)

Posted: Mon Jun 12, 2006 2:52 am
by graeme
That's good to hear. It sounds like POTIM was the source of the problem. If you care most about the energy barrier, you can find the energy difference between the saddle point and your initial state. You can also use our scripts:

vfin.pl dir
cd dir
nebresults.pl

which will archive your run into the directory dir and then do a force based spline along the minimum energy path to give a better estimation of the barrier. Using the climbing image will also give you a very good estimation of the energy barrier because the highest energy image will be at the saddle point when the band converges.

If you want to make sure the entire band is converged, including this problamatic image, I suggest using even more conservative parameters. The quadratic optimizers (ibrion=1 and 2) tend to perform poorly near the minimum if the forces are not accurate enough. To improve robust convergence, you can use IBRION=3 and POTIM=0.1, or choose a smaller value of ediff (1e-6 for example), which will give you more accurate forces and better convergence using IBRION=1.

Posted: Mon Jun 12, 2006 5:40 am
by hlzya
I have also tried POTIM=0.35, but the situation is the same as that when POTIM =0.1. Therefore, I wonder if POTIM should be increased, e.g., POTOM=0.5.

In fact, I find the reason why the forces on image 02 can not be converged to 0.01 ev/A is that the initial guess may be far away from MEP. If you can see my CONTCAR in directory 00~04, you can find in image 00, 03 and 04, the two H atoms which are far away from the Ni surface are in a plane parallel to the Ni surface, while in image 01 and 02, the plane with this two H atoms involved are not parrllel to the Ni surface.

So, the problem is that how to optimize the CH4 orientation in 01 and 02 images. My present POTIM seems to be too small to change the CH4 configuration for such a large extent, right?

Best