Page 1 of 1

Convergence problems

Posted: Fri Jan 04, 2008 3:39 pm
by buber
Dear Vasp Community:

I'm trying to do some very basic defect calculations in a class of complex oxides with the stoichiometry AB6O12, where A=U,W,Re and B=Y,Yb,Lu. In most cases, especially when A=W, I haven't had any problems. But, I have had problems with convergence in two cases that I haven't been able to make any progress on.

When A=U and B=Y, I can converge the perfect structure very nicely. But, when I look at an antisite defect, by replacing one Y with U, the calculation doesn't converge. The electronic iterations get stuck, changing by +/- the same amount. A number of colleagues have suggested I change the smearing, set ISPIN=2, set LMIXMAX=6, etc. Everything I've tried, either doesn't change the behavior or does fix the electronic iterations, but after 3 or 4 ionic iterations, the forces on the ions blow up (become unphyiscally large, like 100 eV/angstrom).

The second problem, when A=Re and B=Lu, is that I don't get any kind of convergence, even in the perfect case. I get a warning at the top

VERY BAD NEWS! internal error in subroutine IBZKPT:
Reciprocal lattice and k-lattice belong to different class of lattices. 168

But all I've done compared to the A=U, B=Y case is change the POTCAR file. If I let it go, I get similar convergence problems as before.

Any suggestions on what I might try to fix these problems would be greatly appreciated. I'm using ISYM=0 and a Monkhorst-Pack grid of 2x2x2 for these calculations, along with the GGA PAW pseudopotentials.

Thank you,
Blas

Convergence problems

Posted: Tue Jan 08, 2008 12:27 pm
by admin
1) for the electronic convergence
--) LMAXMIX=6 should be set if you have rare-earth elements in the cell, for d-elements, please set it to 4
--) it may help to decrease the mixing parameters (AMIX, BMIX, and in addition AMIX_MAX and BMIX_MAX for the spin-polarized runs to improve convergence )
--) also, it may help to increase NELMDL (the number of non-selconsistent steps at the beginning) to improve the pre-convergence of the wavefunction
--) maybe it also helps to use L(S)DA+U ( please check the electronic structure and compare to experiment, if results are available)

2) for the forces which suddenly increase:
--) please check (in OSZICAR) if each ionic step was fully converged electronically up to the requested convergence limit (EDIFF) before the forces are calculated. (i.e. whether the forces were calculated after el. convergence is reached or because the number of electronic convergence steps (NELM=60, by default) was reached without actually being converged. The latter yields unreasonable forces. (in that case, please either increase NELM and/or try one of the above mentioned to improve the electronic convergence)
--) if each of the ionic steps was converged, please check the relaxation history (XDATCAR file). Sometimes, especially the first ionic steps 'overshoot' a little. If that is the case, please choose a different relaxation algorithm (IBRION) and/or decrease the step width (POTIM)

3) the VERY BAD NEWS message is due to different lattice types obtained for the unit cell and the symmetrized k-mesh. (triclinic and simple monoclinic as it seems). maybe the axes ratios of the unit cells are not 1:1:1 ? please check.
for the symmetry analysis, the
Bravais Matrix ,
atoms' positions, and, if set,
magnetic moments and
atoms' velocities are taken into account.
If nothing has been changed except the atoms in the PP, it seems strange that this warning shows up for some calculations whereas it does not for others. In any case, if ISYM=0 is set, symmetrization is not taken into account in the calculation later on, the only point to be careful with is the axis' ratio then.