[Wien] SECLR4 error in large supercells / high number of k points

[Peter Blaha]

I checked your files and could verify the problem.

You have to change in    SRC_lapw1/lopw.f
...
    90                   CONTINUE
                         IF (HL .LE. 1.0D-3) GOTO 10
...

the value of 1.0d-3 to 1.0d-2

Short explanation: This routine attaches to each LO a plane wave (K_n) and
checks if the basis function (Y_lm * exp(iK_n R) are not linear dependent.

The check was done with a limit of 10-3

It seems, this limit is not secure enough.

PS: A nice example of how one should use the mailing list. Kurt Lejaeghere
provided convincing arguments and information so that I decided that the problem
is not due to a user error, but might have deeper roots.

Once I had the example, the solution was not too difficult (in this case).

Thanks for the good "preparation" of the problem.

> I have looked into the matter again and have tried everything I could 
> think of. Now I am however able to describe the problem more in detail 
> and more systematically. Tests have shown that the SECLR4 error occurs 
> in all supercells containing bcc Fe and impurities, as long as the 
> number of k points is high enough. Even if Fe itself is taken as an 
> impurity, this is the case (the supercell is then pure iron). When the 
> same size of supercell is taken, the problem always occurs at the same k 
> point, independent of the value of RKMax or the chemical nature of the 
> impurity. Most of the times this is the only k point where the error 
> takes place (a parallel computation only fails for the subprocess 
> containing that specific k point). Below an example is given.
> 
> For a pure Fe supercell containing 128 atoms (4x4x4 supercell, P 
> sublattice), RKMax 7.0 and 10 k points in the IBZ (200 in the FBZ) runs 
> without a glitch. When the converged case is restored in a new and empty 
> directory and the number of k points is increased up to 56 (2000 in the 
> FBZ), the 2nd k point gives problems (and only that one):
> 
>          2         1         1         3        24 24.0
> 
> The first k point is computed correctly, giving:
> 
>      K=   0.04167   0.04167   0.04167            1
>       MATRIX SIZE11270  WEIGHT= 8.00  PGR:
>      EIGENVALUES ARE:
>      -5.8279952   -5.8264711   -5.8264711   -5.8264462   -5.8262132
> (...)
>       1.9860526    1.9860593    1.9861886    1.9862001    1.9862001
>       1.9862628    1.9897641    1.9897641    1.9897724    1.9897724
>       1.9897768    1.9899170    1.9901205    1.9901205    1.9901247
>             0 EIGENVALUES BELOW THE ENERGY  -10.00000
>        ********************************************************
> 
> Afterwards the information of the 2nd k point is printed, resulting in:
> 
>  Matrix size        11235
>          allocate HS       963.0 MB
>          allocate Z        342.9 MB
>      allocate spanel        11.0 MB          dimensions 11235   128
>      allocate hpanel        11.0 MB          dimensions 11235   128
> (...)
>        allocate YL          65.8 MB          dimensions    15 11235    24
>        allocate phsc         0.2 MB          dimensions 11235
> Time for los      (hamilt) :         50.8
> Time for alm         (hns) :          2.9
> Time for vector      (hns) :         13.5
> Time for vector2     (hns) :         11.9
> Time for VxV         (hns) :        265.3
> Wall Time for VxV    (hns) :          2.1
>  Cholesky INFO =        10819
> 
> the last line being the SECLR4 error. Because this supercell only 
> contains identical Fe atoms (at 14 artificially inequivalent positions) 
> the number of possible changes in the case.in1 file is limited. Whatever 
> I try, be it -in1ef or -in1_orig, with or without an additional d-LO, 
> raising or lowering the linearization energies by 0,1 Ry ..., nothing 
> works and the SECLR4 error persists. With each test I started from a 
> restore based on the case with 10 k points, using a new and empty 
> directory.
> 
> I sincerely hope anyone can help me with this problem, that keeps 
> hunting me in every supercell I need.
> 
> With kind regards
> 
> Kurt Lejaeghere
> Ghent University, Belgium
> 
> _______________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.at
> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien

-- 

                                       P.Blaha
--------------------------------------------------------------------------
Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
Phone: +43-1-58801-15671             FAX: +43-1-58801-15698
Email: blaha at theochem.tuwien.ac.at    WWW: http://info.tuwien.ac.at/theochem/
--------------------------------------------------------------------------