Large Maximum Im/Re Ratio

[GCatarina]

Wannier90 for benzene is giving me sensible WFs (C:pz) and Hamiltonian, but I keep getting large "Im/Re ratios" (~100-1000), which from what I read suggest that I am getting stuck in a local minimum.
Input and output files below.

I am using QE (tried both v7.2 and v7.5), and I already tried both the stable v3.1 and the developer version of wannier90, with similar results.

For completeness, I also tried this example for methane and I still obtain large Im/Re ratios.
I will now explore the random noise functionality (conv_noise_*), which I never tried.
Any other suggestion?

wan.win.txt
wan.wout.txt

[AndresOrtegaGuerrero]

@giovannipizzi Do you know what could be the reason of this? , We also tested the library mode with the PySCF interface and we get zero for this ratio, we wonder the source of the difference

[GCatarina]

Update: exact same calculation but without the 'gamma_only' option (both in pw.x and wannier90) gives similar outputs (same spreads, WFs and Hamiltonian also look the same) but now the Im/Re ratio is very close to zero

[jryates]

I understand this issue. It is a PWSCF problem - not really a Wannier90 issue.

The centres and spreads you have calculated with the Gamma_only methods are indeed correct. However, there is a bug which crept into pw2wannier90.f90 when it was partially re-written a few years ago. The routine which outputs the UNK files does not take account of the way PWSCF stores wave functions in the Gamma branch. (it did in the earlier versions of pw2wannier90). You can see the problem if you write a formatted UNK file - as the wave functions are not real (as they should be in a Gamma-point formalism).

I fixed this issue locally - and the WF are now entirely real. Of course, in the Gamma point branch they can't be anything other than real - as we start with real wave functions, and construct the WF as real combinations of them. (the comment about false minima being complex, only applies to the multi k-point case).

I am not a PWSCF developer - and the last time I seriously did any coding with it was 20 years ago. So I'm not up to speed with how it handles parallelisation etc. So the code below is unlikely to work in parallel - and no one should rely on it. It worked for me in serial with 7.3. I just include it in the hope someone more expert can use it as a starting point for a robust fix.

  !
  ! Transform wavefunction to real space
  !
   evc_r(:, :) = (0.d0, 0.d0)
   DO ipol = 1, npol
      evc_r(dffts%nl(igk_k(1:npw,ik)), ipol) = evc(1+npwx*(ipol-1):npw+npwx*(ipol-1), ibnd)
      if(gamma_only) evc_r(dffts%nlm(igk_k(1:npw,ik)),1) = conjg(evc(1:npw,ibnd)) !JRY HACK
      CALL invfft('Wave', evc_r(:, ipol), dffts)
   ENDDO
   !

As a final comment for the other developers. Wannier90 plotting routines do not take account of the gamma branch. They work perfect well as is. However, they could be sped up by (a) checking the UNK are real (b) then using real arrays to construct the WF.

[jaemolihm]

Dear @GCatarina and @jryates ,

Thank you for the fix, and sorry for missing this case during the rewriting. I think the fix also works in the parallel run, it essentially restores psic(dffts%nlm(igk_k(1:npw,ik))) = conjg(evc(1:npw, ibnd)) which was there before the rewrite. I checked with the methane example that the fix works in the parallel case.

If you are fine, I can proceed with opening a merge request and mentioning this issue.

[jaemolihm]

I created a MR : https://gitlab.com/QEF/q-e/-/merge_requests/2779

[JeromeCCP9]

@jryates am I right that this is open because we would like to print a diagnostic message for the im/re ratio? Is something else needed:

However, they could be sped up by (a) checking the UNK are real (b) then using real arrays to construct the WF.

shall we do this?

[jryates]

@JeromeCCP9 I left this issue open until QE had accepted the PR. They merged it 3 months ago. So I think we can close this issue as resolved. Perhaps we could open a new issue with the gamma point optimisations I suggested.