[QE-users] sternheimer_kernel: root not converged in phonon calculation

[RABONE Jeremy]

I tell a lie - calculations don't always converge at single k-points. In fact the bands calculation for the system I'm studying fails more often than not at any given k-point (it just worked 10 times in a row on Friday - could be the crunchy effect). After more digging, part of the problem seems to be that the random wavefunctions may or may not work at any given k-point, but since there is no seeding of the randum number generator, the exact point at which calculations fail is the same for any given set of conditions but changes as the conditions change. Essentially this makes it impossible to use bands calculation over lines for this system. It is possible to calculate the bands k-point at a time with some small changes to the code (seeding the random number generator so that different random wavefunctions are used for the same calculation, not stopping the band calculation if there are too many unconverged bands but trying again with a new set of random wavefunctions) and repeating any failed calculations until they work. The most awkward problem is that when the Hamiltonian is not positive definite, the LAXlib calls MPIabort so rather more modification would be required to retry band calculations with new wavefunctions.


Kind regards,

Jeremy


________________________________
De : users <users-bounces at lists.quantum-espresso.org> de la part de RABONE Jeremy <jeremy.rabone at get.omp.eu>
Envoyé : vendredi 22 mars 2024 16:45
À : users at lists.quantum-espresso.org
Objet : Re: [QE-users] sternheimer_kernel: root not converged in phonon calculation


Hello again,


  After some further tests (and attempts to calculated electronic band structure), there seems to be something spectacularly broken in calculations where series of q-points (phonon) or k-points (bands) are used. The points where failures occur all work fine if they are calculated as single points, but calculations along a line or over the Brillouin zone always throw one or other of the errors to do with diagonalisation (too many bands unconverged or loss of positive definiteness or root not converged) at some point. The failures all happen at the same point in any given calculation but what is weird is that one can, for example, do a band calculation with 7 k-points with 10 points along each line and it will fail at point 23 but doing the same calculation with just the line between the first 2 k-points it fails on the 4th point...

Is it the case that the wavefunctions from the previous calculation are used as starting point for the next one ? If so is there a way to stop this and always restart from random wavefunctions ?

Kind regards,

Jeremy


________________________________
De : users <users-bounces at lists.quantum-espresso.org> de la part de RABONE Jeremy <jeremy.rabone at get.omp.eu>
Envoyé : mardi 19 mars 2024 15:25
À : users at lists.quantum-espresso.org
Objet : [QE-users] sternheimer_kernel: root not converged in phonon calculation


Hello,

  I seem to be having a similar problem to one previously reported in 2022 where phonon calculation fails to converge with huge errors for a particular q point having worked fine for the gamma point. The exchange seemed to conclude that the math libraries (mkl) of a particular machine were at fault, and I'm wondering if this is also the case with the cluster I'm using. Has there been any conclusive identification of mkl libraries causing problems (I'm using intel 17.0.1.132) ?

Kind regards,

Jeremy

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20240326/04a772c6/attachment.html>