[gpaw-users] NEB convergence

[Lasse Vilhelmsen]

I have often run into the same problem with the CI-NEB method being 
significantly harder to converge.

I have found that it often helps breaking the path into smaller sections 
instead of optimizing the entire path in one go. So if  you know that 
image i is the transition state image, you can try running the CI-NEB 
with image i-2 and i+2 as the end points.

I can see that you are using the FIRE optimizer which I usually also 
find the one working the best for CI-NEB calculations, but you could try 
using another as well, i.e. BFGS.

Best Regards
Lasse Vilhelmsen

On 5/8/2012 13:35, Marcin Dulak wrote:
> On 05/08/12 19:05, Hieu Doan wrote:
>> Hi,
>>
>> I have recently had convergence issues with NEB calculations whenever a climbing image is turned on. Images are usually relaxed with fmax=0.5 and climbing=false before setting the final criteria to fmax = 0.05 and climbing=true.  Here are several things I have tried but did not help:
>> - Change the eigensolver to davidson
> eigensolver should not help for NEB (optimizer convergence),
> but, intuitively, what may help is converging SCF better
> https://wiki.fysik.dtu.dk/gpaw/documentation/manual.html#manual-convergence
>> - Lower dtmax from 0.3 to 0.1
>> - Further relax images to fmax=0.2 before climbing image
>>
>> These calculations often reach the walltime limit while the force is still fluctuating nowhere around 0.05. I also notice that the climbing TS image either beaks and makes no sense.  My script is posted below. Any suggestion would be greatly appreciated!
> how much properties depend on optimizer force stopping threshold?
> What's the expected precision of NEB?
> Maybe 0.05 is too strict to achieve?
>
> Best regards,
>
> Marcin
>> ----------------------------------
>> from ase import *
>> from gpaw import GPAW, Mixer, FermiDirac
>> from ase.visualize import view
>> from ase.neb import NEB
>> from ase.optimize import BFGS, FIRE
>> from ase.parallel import rank, size
>> from ase.io.trajectory import PickleTrajectory
>> from ase.io import read, write
>> import numpy as np
>> import os
>>
>> token = 'image'
>> nimages = 7      # number of images including initial and final
>> parallel = True
>> climbing = True
>> maxforce = 0.05
>>
>> # Read initial and final states:
>> initial = read(token+'_0.traj')
>> final = read(token+'_'+str(nimages-1)+'.traj')
>>
>> # Make a band consisting of nimages images:
>> interpolate = False
>>
>> n = size //(nimages-2)    # number of cpu's per image
>> j = 1 + rank//n           # intermediate image number, start from 1
>> assert (nimages-2) * n == size
>> images = [initial]
>>
>> for i in range(1,nimages-1):
>>      filename = token+'_'+str(i)
>>      try:
>>          atoms = read(filename+'.traj.bak')
>>          print "Updated from traj file ",filename+'.traj.bak'
>>      except:
>>          atoms = initial.copy()
>>          interpolate = True
>>
>>      ranks = np.arange((i-1)*n, i*n)
>>
>>      if rank in ranks:
>>
>>          calc = GPAW(h=0.18,
>>                      xc='RPBE',
>>                      kpts=(4,4,1),
>>                      occupations=FermiDirac(0.1),
>>                      eigensolver='dav',
>>                      mixer=Mixer(beta=0.05, nmaxold=5, weight=50),
>>                      maxiter=500,
>>                      spinpol=False,
>>                      txt=token+'_%d.txt' % j,
>>                      communicator=ranks)
>>          atoms.set_calculator(calc)
>>
>>      images += [atoms]
>>
>> images += [final]
>>
>> neb = NEB(images,parallel=parallel,climb=climbing)
>>
>> if interpolate:
>>      neb.interpolate()
>>
>> # Optimize:
>> qn = FIRE(neb,logfile='qn-'+token+'.log',dtmax=0.1)
>> traj = PickleTrajectory(token+'_%d.traj' % j, 'w', images[j],
>>                          master=(rank % n == 0))
>> qn.attach(traj)
>> qn.run(fmax=maxforce)
>> ----------------------------------------
>>
>> Hieu A. Doan
>> Graduate Research Assistant
>> Department of Chemical and Biomolecular Engineering
>> University of Houston
>> _______________________________________________
>> gpaw-users mailing list
>> gpaw-users at listserv.fysik.dtu.dk
>> https://listserv.fysik.dtu.dk/mailman/listinfo/gpaw-users
>>
>