Regenerating DFT data for diamond tests fails with QE 5.3.0

[qmc-robot]

Reported by: markdewing

In tests/solids/diamondC_1x1x1_pp (and diamondC_2x1x1_pp) when re-running the input file in dft-inputs/, QE 5.3.0 fails with the error:
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Error in routine cdiaghg (10):
S matrix not positive definite
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Full output is attached.

The psuedopotential file (C.BFD.upf) is not included in the dft-inputs directory. I used the one from pseudopotentials/BFD.

QE 5.2.1 also fails with the same error.

[qmc-robot]

Comment by: markdewing

File: try2.scf.out

[qmc-robot]

Comment by: prckent

I traced the original/correct C.BFD.upf file add added it to diamondC_1x1x1_pp and diamondC_2x1x1_pp. The total energies should agree with those in the stored scf.out files. I checked on oxygen today with QE 5.3.0.

I think that this points to a problem with the converted BFD UPF files. If Carbon is "wrong", other elements are also suspect.

[qmc-robot]

Comment by: ye-luo

The reference state is wrong 2s(2)2p(0) in that upf when generating KB projectors.

[qmc-robot]

Comment by: prckent

The "0p" wavefunction, around line 2546 is clearly wrong and contains NaNs. Most of the earlier differences are small.

[pk7 @oxygen BFD]$ grep -n -i nan *
C.BFD.upf:2554: 2.12199579097e-314 1.02335145293819053e-316 nan 1.08440373767356837e-316

Even if the intent was to generate KB projectors from the 2s(2)2p(0) configuration, some codes need a little bit of occupation in the "empty" channels to generate sane results. Possibly this includes our converter.

[qmc-robot]

Comment by: prckent

Perhaps we need to run neutral atoms in a box for all our converted potentials, or a similar test.

[qmc-robot]

Comment by: prckent

I'll assign this to Jaron since he did the conversion (squeezed before workshop), but it might need delegating.

[qmc-robot]

Comment by: jtkrogel

I see. This is not the first problem noted with the uploaded upf files.

I dredged up the files from an old (and apparently grungy) conversion
suite, without further vetting. I don¹t think we can really avoid
regenerating the projectors and doing some level of testing to satisfy
ourselves.

So, the question is: how does one economically check the viability of KB
projectors derived from a semilocal pseudopotential?

[qmc-robot]

Comment by: prckent

Perhaps we have to bite the bullet and:

• Check the neutral atom works in QE
• Compute (say) the first IP
• Compare vs result from GAMESS with same functional using the BFD gaussian fit and basis set?

Luke might be able to comment on this.

[qmc-robot]

Comment by: prckent

Maybe we could check the eigenvalue spectrum. That would allow us to do only a single QE and single GAMESS run.

[qmc-robot]

Comment by: lshulen

There are a couple of subtleties that make me think we should actually do QMC calculations for a couple of states of the isolated atom if we want to be sure. Firstly, the representation is slightly different between GAMESS and QE. For instance, numerical representation of the potential vs. gaussian ones. Also there is a question about the convergence of the basis set for the wavefunction.

My preference would be to calculate the neutral atom and something like the IP using both QE generated and GAMESS generated trial wavefunctions at the DMC level. This would directly answer the question of how these issues propagate to the calculations we really care about.

That said, probably the PBE eigenvalue spectrum between the two codes would catch a lot of things and it makes sense to use this as a screening methodology. We would just have to put in some sort of tolerance that we thought was acceptable.

[prckent]

Can anyone confirm if this is fixed? i.e. pseudopotential is now good and works with QE6.2

[ye-luo]

We have removed the whole pseudopotential folder carried by QMCPACK. https://www.pseudopotentiallibrary.org
is the right place to get validated NLPP files.