Difficulty converging all electron magnetic system

[Craig Schwartz]

Hello,

         I am trying, and struggling, to converge an all electron calculation on Dysprosium ferroborate. An example input and basis set are attached.  The system is challenging in that its magnetic and we need all the electrons to do certain x-ray calculations at the heavy element’s low-lying edges (so we can’t just use the MOLOPT and GTH PBE pseudopotentials for instance).  The basis sets are all taken directly from basis set exchange (cp2k format) and the ‘ALL’ potential for Dy is roughly a copy of I from the standard cp2k ‘POTENTIAL’ file.

Dy  ALLELECTRON ALL

   12   24   20   10

     0.50000000    0

 

 I’ve been using a 2026.1 release from the git (pulled April 29, 2026) which includes Gaussian smearing, but I’ve also tried Fermi-Dirac smearing as well as the 2025.2 development version.  The system is a bit too big for hybrid functionals, and things like PBE have not worked.

       I’ve messed around with Boyden, Pulay, Kerker and direct mixing and nothing has worked.  The best convergence I’ve seen has been with DIRECT_P_MIXING and  a small ALPHA 0.1, although even this does not converge. Broyden, Kerker, and Pulay mixing attempts with different values of ALPHA and BETA terminate with fluctuations on the order of E+02.

  1990 P_Mix/Diag. 0.10E+00   80.9    21.05789673    -50737.0152956767  2.02E-01

  1991 P_Mix/Diag. 0.10E+00   80.9    12.46218230    -50736.8642879388  1.51E-01

  1992 P_Mix/Diag. 0.10E+00   80.8    11.07681641    -50737.0034034056 -1.39E-01

  1993 P_Mix/Diag. 0.10E+00   80.8    12.96840862    -50736.7868106099  2.17E-01

  1994 P_Mix/Diag. 0.10E+00   80.7    11.69891226    -50736.9641743435 -1.77E-01

  1995 P_Mix/Diag. 0.10E+00   80.8    15.49182193    -50737.0700617575 -1.06E-01

  1996 P_Mix/Diag. 0.10E+00   80.9    20.70203774    -50737.1124326006 -4.24E-02

  1997 P_Mix/Diag. 0.10E+00   80.8    14.68011752    -50737.0046223170  1.08E-01

  1998 P_Mix/Diag. 0.10E+00   80.7    11.80777138    -50737.1114555977 -1.07E-01

  1999 P_Mix/Diag. 0.10E+00   80.7    14.55452265    -50736.6616467933  4.50E-01

  2000 P_Mix/Diag. 0.10E+00   80.9    10.58192107    -50736.8174171267 -1.56E-01

 

  Leaving inner SCF loop after reaching  2000 steps.

 

Experimentally this structure has a net magnetization of 7 bohr magnetor per formula unit when in field, so I’ve been using both an unspecified multiplicity low-spin initial guess and a high spin initial guess that is close to per 7 bohr mangeton/formula unit (i.e. the Dy f electrons align with the field).

      Does anyone have any suggestions on how to converge this problem?  It’s annoying because of the combination of all electron, magnetization, and rare-earth f-electrons that aren’t in the pseudopotential.  This is the standard R-centered cell for this structure, but I’ve also tried the primitive cell with a denser k-grid and have had similar issues.

Thanks for any help,

Craig