Follow-up NEB questions: Job convergence and optimization

[whdxsn]

Hi, graeme and dano . Thanks for your help for my previous questions. Now I have some additional questions as the calculations go on. Please bear me with since I put all relevant details i think useful below, so it is a bit lengthy.

The situations are the following :

My 1st run of calculation A stopped due to exceeding over walltime . I restarted it following the previous posts and do the 2nd run. Below are the energy output of the 1st run ,

457 N E dE d eps ncg rms rms(c)
458 DAV: 1 -0.297683758207E+03 -0.29768E+03 -0.20963E-01 13440 0.132E+00 0.743E-02
459 DAV: 2 -0.297684060285E+03 -0.30208E-03 -0.44749E-03 13808 0.185E-01 0.469E-02
460 DAV: 3 -0.297684045316E+03 0.14969E-04 -0.11112E-04 15976 0.313E-02 0.282E-02
461 DAV: 4 -0.297684085767E+03 -0.40452E-04 -0.41170E-05 12000 0.226E-02 0.515E-03
462 DAV: 5 -0.297684093122E+03 -0.73551E-05 -0.11262E-05 11328 0.113E-02
463 37 F= -.29768409E+03 E0= -.29768409E+03 d E =-.448890E-02 mag= 0.0000
464 curvature: -2.02 expect dE=-0.165E-01 dE for cont linesearch -0.398E-07
465 trial: gam= 0.84770 g(F)= 0.820E-02 g(S)= 0.000E+00 ort = 0.339E-04 (trialstep = 0.448E+00)
466 search vector abs. value= 0.502E-01
467 bond charge predicted
468 N E dE d eps ncg rms rms(c)
469 DAV: 1 -0.297686643549E+03 -0.29769E+03 -0.63576E-01 13440 0.227E+00 0.128E-01
470 DAV: 2 -0.297687646884E+03 -0.10033E-02 -0.14089E-02 13728 0.326E-01 0.818E-02
471 DAV: 3 -0.297687591709E+03 0.55175E-04 -0.32625E-04 17008 0.540E-02 0.505E-02
472 DAV: 4 -0.297687680657E+03 -0.88948E-04 -0.15012E-04 13440 0.420E-02 0.101E-02
473 DAV: 5 -0.297687689918E+03 -0.92610E-05 -0.38064E-05 13696 0.206E-02

For the 2nd run, I also monitored the force and energy while the job is running,
Below are the outputs:

0 0.07161400 -297.68252000 0.00000000
1 0.05848100 -297.52924200 0.15327800
2 0.05592600 -297.49658200 0.18593800
3 0.05831800 -297.60763100 0.07488900

0 0.06929400 -297.68769000 0.00000000
1 0.08378800 -297.53136500 0.15632500
2 0.07602000 -297.49675300 0.19093700
3 0.07906600 -297.60748100 0.08020900

0 0.06929400 -297.68769000 0.00000000
1 0.08378800 -297.53136500 0.15632500
2 0.07602000 -297.49675300 0.19093700
3 0.07906600 -297.60748100 0.08020900

0 0.03924600 -297.68912400 0.00000000
1 0.05203100 -297.53183000 0.15729400
2 0.03996500 -297.49689400 0.19223000
3 0.03230400 -297.60760400 0.08152000
One thing I notice is that the energy and force of images do not change too much from one check to the other. I also get the following information after grepping the current output :

./01/OUTCAR:8884:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./01/OUTCAR:16219:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./01/OUTCAR:23564:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./01/OUTCAR:30901:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./01/OUTCAR:38246:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./01/OUTCAR:45583:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./02/OUTCAR:8884:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./02/OUTCAR:16219:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./02/OUTCAR:23564:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./02/OUTCAR:30901:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./02/OUTCAR:38246:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./02/OUTCAR:45583:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./03/OUTCAR:8933:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./03/OUTCAR:16316:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./03/OUTCAR:23662:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./03/OUTCAR:31000:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./03/OUTCAR:38346:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./04/OUTCAR:8884:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./04/OUTCAR:16219:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./04/OUTCAR:23564:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./04/OUTCAR:30901:------------------------ aborting loop because EDIFF is reached ----------------------------------------
./04/OUTCAR:38246:------------------------ aborting loop because EDIFF is reached ----------------------------------------

./03/OUTCAR:38541: reached required accuracy - stopping structural energy minimisation
./04/OUTCAR:38440: reached required accuracy - stopping structural energy minimisation

It looks like image 03 and 04 have reached required accuracy and stopped structural minimization. However, image 01 and 02 do not stop. However, for all image directories, the last write to the outcar file was several hours before(which i do not expect), even though the job is shown running in the queue.

Here are my questions:
1. Based on the above information, especially the observation that the energy and force of images do not change too much from one check to the other, Do I encounter the abnormal NEB energy and Convergence problem or my calculation has reached required accuracy? From previous posts, graeme suggested forces on images greater than 10 eV/Ang are very high and IOPT flag could be used to fix it. The forces on the images for my case are on order of 0.1, so do I need to do something to further minimize the force ? Currently my INCAR is the following :

1 NELMIN=5
2 EDIFF=1E-5 # default is 0.0001, default EDIFFG is 10*EDIFF
3 NSW=1000
4 IBRION=2
5 ISIF=1 #no cell shape and no cell volumn, no tensor calculation, just force
6 PREC=ACU
7 ISMEAR=0
8 SIGMA=0.002
9 ISPIN=2
10 IMAGES=4
11 SPRING = -5

the IBRION=2 in my case.

My another run 1st for calculation B does not have "reached required accuracy - stopping structural energy minimisation" grepped in any image directory yet. The force and energy information is :

0 0.61028700 -297.66883000 0.00000000
1 0.95960300 -297.50246000 0.16637000
2 1.15943300 -297.39468500 0.27414500
3 0.98762200 -297.36664800 0.30218200
4 0.59985600 -297.36459700 0.30423300

0 0.09215400 -297.69497300 0.00000000
1 0.12711300 -297.54091900 0.15405400
2 0.16876400 -297.45303300 0.24194000
3 0.11106100 -297.41097500 0.28399800
4 0.09488100 -297.38129600 0.31367700

0 0.08990400 -297.69912300 0.00000000
1 0.10193800 -297.54432800 0.15479500
2 0.11274900 -297.45816900 0.24095400
3 0.09798200 -297.41411300 0.28501000
4 0.09531400 -297.38247700 0.31664600

I am thinking that it is still not converged yet until I see either "reached required accuracy " message or force and energy do not change much . Right?


Thanks for your attention and replies.

[dano]

We typically decide that the energy is converged on a force criteria because the NEB is really defined for via forces. And in this example your forces are still fairly high
I would typically run my INCAR something like this

IMAGES=4
ICHAIN=0
LCLIMB=.TRUE. # you are going to want this turned on to quantitatively identify the saddle

IBRION=3 # quick-min is more stable the CG to to the lack of an object function
POTIM=0.1
EDIFF=1E-5
EDIFFG=-0.001 # this sets convergence when the forces on all degrees of freedom is less than 0.001 eV (0.01 is also very reasonable to speed things up)
ISIF=0
Other VASP flags

[whdxsn]

Hi, dano
Thanks for your reply. I am using the following in the INCAR :

1 NELMIN=5
2 EDIFF=1E-5 # default is 0.0001, default EDIFFG is 10*EDIFF
3 NSW=1000
4 ISIF=1 #no cell shape and no cell volumn, no tensor calculation, just force
5 PREC=ACU
6 ISMEAR=0
7 SIGMA=0.002
8 ISPIN=2
9 IMAGES=5
10 SPRING = -5
11 IBRION=3
12 POTIM=0.1
13 EDIFFG=-0.01

The jobs i have been running are : diffusion from A->B , A->C and B->C
Point A has the lowest energy, while point B is 0.27eV above A and point C is 0.40eV above A.
The job were running for 60-75 hours on 5 nodes , 4cpu per nodes cluster.
I have monitored the force and energy vs ionic steps. see attachment
By examining them, I found that :
1. force goes down rather continuously while the energy begins to bottom and goes up( within range of 0.02eV) see vaspoutput3-5 of A->B and vaspout5 in A-C
2. forces goes down more ruggedly while the energy decrease. See A-C and B-C. I also found A-C and B-C has fewer ionic steps even though they run longer time.
3. By examining the last nebef output: I also notice that initial images like 0 and 1 has relative smaller forces than the rest. Some forces are so large even though they have run a lot of ionic steps.
like A-B.
A-B
0 0.01986100 -297.75357000 0.00000000
1 0.03546700 -297.68351500 0.07005500
2 0.02602000 -297.58788600 0.16568400
3 0.02833900 -297.52530900 0.22826100
4 0.04229900 -297.49960100 0.25396900
A-C
0 0.08147600 -297.69055800 0.00000000
1 0.08463300 -297.53996600 0.15059200
2 0.12700400 -297.45160600 0.23895200
3 0.10997000 -297.40942600 0.28113200
4 0.10378500 -297.38011700 0.31044100
B-C
0 0.01789200 -297.48624600 0.00000000
1 0.02653300 -297.45991700 0.02632900
2 0.03338400 -297.42918400 0.05706200
3 0.06238500 -297.40021700 0.08602900
4 0.07984700 -297.37617300 0.11007300

Questions:
1. Could the small energy differences between A B and C make it harder for the force to go down
2. Are the attached force/energy vs ionic steps make sense to you
3. what could i do further to speed up the force minimization.

thanks