Use of different diagonalization methods #6
Replies: 1 comment 2 replies
-
Hello Julio, Are you using CG ins pw.x? Because there is not specific reason for the NSCF convergence in d3q to be harder to converge than the scf one in pw.x. Maybe if you have a large number of electrons, the threshold becomes too strict. Can you change at the line 202 of run_nscf_d3.f90 It is an old default that "should be good for all cases" but has probably never been tested. Let me know if this solves your problem, this way I can fix it in future releases. cheers |
Beta Was this translation helpful? Give feedback.
-
Dear contributors,
First of all thanks for releasing the d3q code, I'm hopeful it will be very helpful in my research.
I'm starting this discussion here because I've been having some issues with the code where too many bands are not being converged. I tried solving the problem by running nscf runs with different parameters and the only solution that gave me consistent results was changing the diagonalization method from 'david' to 'cg'.
I tried changing it, by changing the source code variables from:
isolve = 0
david = 4
to:
isolve=1
david=1
(as indicated in the Quantum Espresso source code setup_nscf.f90)
in the file run_nscf_d3.f90 and recompiling it, but the code used the Davidson method anyway.
Is there something that I'm missing or some guidance that could help me get around it?
kind regards,
Julio
Beta Was this translation helpful? Give feedback.
All reactions