VASP5.2.11 relaxation of defect structure cannot finish

Message

chorawut · #1 Post by **chorawut** » Sun Jul 24, 2011 7:11 am

I'm try to relax the defect system by using both VASP4.6 and VASP5.2. I set the criteria of our relaxation by using EDIFFG=-0.01 in INCAR. For VASP4.6, I can relax our defect structure within a few ionic relaxation steps (~ 30 -50). But, for VASP5.2, I cannot relax our defect structure. The problem is very similar to the previous one in this Forums. Even if I copy CONTCAR to POSCAR and rerun it, I still cannot get the optimized defect structure with EDIFFG=0.01 criteria. Anyway, the total ground state energy is already the same but the force is still higher than criteria. Moreover, I see the forces in OUTCAR file and I found that there is something strange about the relaxation algorithm in VASP5.2.11. The forces seem do not get closer to the criteria value. I have to note why I used EDIFFG for our criteria. This is because I want to get the defect structure without any strain. If I used EDIFF=0.0001, I may still get the defect structure with some strain in the crystal.
For VASP5.2.11, the problem is something like the following,
47 F= -.91218170E+03 E0= -.91217749E+03 d E =0.128814E+03 mag= 0.0000
48 F= -.91218172E+03 E0= -.91217750E+03 d E =0.128814E+03 mag= 0.0000
49 F= -.91218175E+03 E0= -.91217752E+03 d E =0.128814E+03 mag= 0.0000
50 F= -.91218177E+03 E0= -.91217771E+03 d E =0.128814E+03 mag= 0.0000
51 F= -.91218177E+03 E0= -.91217755E+03 d E =0.128814E+03 mag= 0.0000
52 F= -.91218177E+03 E0= -.91217766E+03 d E =0.128814E+03 mag= 0.0000
53 F= -.91218177E+03 E0= -.91217756E+03 d E =0.128814E+03 mag= 0.0000
54 F= -.91218177E+03 E0= -.91217760E+03 d E =0.128814E+03 mag= 0.0000
55 F= -.91218177E+03 E0= -.91217756E+03 d E =0.128814E+03 mag= 0.0000
56 F= -.91218177E+03 E0= -.91217757E+03 d E =0.128814E+03 mag= 0.0000
57 F= -.91218177E+03 E0= -.91217757E+03 d E =0.128814E+03 mag= 0.0000
58 F= -.91218177E+03 E0= -.91217759E+03 d E =0.128814E+03 mag= 0.0000
59 F= -.91218177E+03 E0= -.91217756E+03 d E =0.128814E+03 mag= 0.0000
60 F= -.91218177E+03 E0= -.91217758E+03 d E =0.128814E+03 mag= 0.0000
61 F= -.91218177E+03 E0= -.91217759E+03 d E =0.128814E+03 mag= 0.0000
62 F= -.91218177E+03 E0= -.91217757E+03 d E =0.128814E+03 mag= 0.0000
63 F= -.91218177E+03 E0= -.91217757E+03 d E =0.128814E+03 mag= 0.0000

I'm very curious whether the relaxation algorithm using in VASP4.6 and 5.2 is different. If yes, Can I go back to use the previous algorithm as used in VASP4.6 ? or Could anyone help me about this problem ?

My computer is opteron with x86-64 and I used IFORT-11.1.072 and lammpi-7.0.4 and ATLAS lib. for compilation.

superyoyo · #2 Post by **superyoyo** » Mon Jul 25, 2011 6:07 am

Hi Chorawut,

I am also facing this problem all the time. I have posted this issue:

http://cms.mpi.univie.ac.at/vasp-forum/ ... 875.0#9802

However, there is no solution yet. I hope the VASP admin might figure and help us out, regarding this problem.

Cheers,
<span class='smallblacktext'>[ Edited Fri Jul 29 2011, 09:39AM ]</span>

chorawut · #3 Post by **chorawut** » Tue Jul 26, 2011 11:50 am

Yep...I found that you also have the same problem. I hope that if the algorithm used in VASP4.6 or 5.2.2 did not give any problem on force, I hope that VASP admin can change the algorithm used in VASP5.2.11 to old one.

chorawut · #4 Post by **chorawut** » Mon Aug 15, 2011 2:25 pm

Hello Admin
Are you still fixing this problem ?

chorawut · #5 Post by **chorawut** » Tue Aug 16, 2011 12:49 pm

Ok....now I can solve the problem
For VASP5, I have to do two steps:
(1) IBRION = 2, EDIFFG=-0.2 (low criteria)
(2) IBRION = 1, EDIFFG=-0.01 (high criteria)
Then, the defect structure can be relaxed and reaches the required accuracy....

For VASP4, I don't have to this.

<span class='smallblacktext'>[ Edited Tue Aug 16 2011, 12:53PM ]</span>

seniyajw · #6 Post by **seniyajw** » Tue Aug 23, 2011 9:57 am

I only ask because when I started using VASP, let iSpin = 2, instead of changing iSpin = 1, and maintained the ZBRENT error. However, once turned off the spin polarization, I came to convergence quite rapidly.

chorawut · #7 Post by **chorawut** » Tue Aug 23, 2011 7:10 pm

I think this problem might depend on the running machine because I feel that not everyone has this problem. My computer is AMD Opteron (x86_64). I never found this problem in VASP4.6 but I always found this problem in VASP5. I think I waste a lot of time in relaxation in VASP5. How can I solve this problem ?

#8 Post by **admin** » Tue Aug 30, 2011 5:47 pm

sorry, we CANNOT reproduce this behaviour in our tests (vasp.5.2.2, vasp.5.2.12, vasp.4.6.28).
please check the following:
1) use the makefiles we deliver together with the respective code version
2) use the same compiler as we do (Intel 11.1)
3) compile the parallel executables WITHOUT the SCALAPACK (this may cause problems)
4) check if the behaviour is the same for serial and parallel versions (to avoid errors due to MPI_over-optimization)
5) reduce the optimization level of the compilation. if this gives comparable
results, increase it again step by step

#9 Post by **admin** » Fri Sep 16, 2011 5:05 pm

there has been a bug in nonlr.F which did not show up in all our test cases but seems to cause numerical instabilities to calculated forces on some atoms in rare cases. The bug has been fixed, the fixed version will be included in the upcoming release of vasp.5.2.12