ionic convergence problem

[peterklaver]

Hello people,

At present I'm doing lots of interstitial relaxations. A few configurations are giving me trouble with their ionic relaxation. At first the relaxation seems to go fine. But when there are just a few atoms left with force components of 0.1x eV/A the energy starts to oscillate. I've tried both IBRION=1,2 and I've also tried using very small POTIMs. Then I thought it might be the case that the forces are not accurate enough, althought the drift is much smaller than 0.1x. Anyway, I tried setting various precision-related prameters to finer values, but that didn't help. Seems logical, as the drift was small to begin with. I also tried more k-points but to no avail. The energy convergence of one attempt is given below. The starting configuration had already been relaxed a good deal. And the first 4 steps of this relaxation also go fine, but then things go wrong:

1 F= -.10638914E+04 E0= -.10640265E+04 d E =-.106389E+04 mag= 234.4195
2 F= -.10639015E+04 E0= -.10640369E+04 d E =-.100583E-01 mag= 234.4539
3 F= -.10639055E+04 E0= -.10640406E+04 d E =-.407408E-02 mag= 234.4406
4 F= -.10639093E+04 E0= -.10640441E+04 d E =-.376387E-02 mag= 234.4244
5 F= -.10639078E+04 E0= -.10640428E+04 d E =0.147456E-02 mag= 234.4241
6 F= -.10639068E+04 E0= -.10640418E+04 d E =0.105198E-02 mag= 234.4259
7 F= -.10638890E+04 E0= -.10640250E+04 d E =0.177652E-01 mag= 234.4554
8 F= -.10638818E+04 E0= -.10640182E+04 d E =0.715494E-02 mag= 234.4691
9 F= -.10638979E+04 E0= -.10640337E+04 d E =-.160660E-01 mag= 234.4537
10 F= -.10638989E+04 E0= -.10640348E+04 d E =-.100234E-02 mag= 234.4605
11 F= -.10638998E+04 E0= -.10640355E+04 d E =-.863053E-03 mag= 234.4618
12 F= -.10638980E+04 E0= -.10640339E+04 d E =0.176092E-02 mag= 234.4625
13 F= -.10638911E+04 E0= -.10640273E+04 d E =0.694348E-02 mag= 234.4661
14 F= -.10638833E+04 E0= -.10640198E+04 d E =0.776095E-02 mag= 234.4681

I've put input files and OSZICAR/OUTCAR files of my various attempts (9 MB) on

http://titus.phy.qub.ac.uk/group/Peter/files.tar.gz

Any suggestions as to what I'm missing here?

greets,
Peter

[peterklaver]

Solved it. The calculations which were giving me trouble had configurations where the system was very close to a saddle point. Understandably a difficult task for any ionic convergence algorithm. Starting from a configuration slightly further away from the saddle point helped it to converge properly. Apparently, the 'history' built up during the relaxation can make the crucial difference.