Convergence Issues

Try to use default parameters for the calculator. Simple and often useful.

Here you find a list of suggestions that should be considered when encountering convergence problems:

  1. Make sure the geometry and spin-state is physically sound.

    Remember that ASE uses Angstrom and not Bohr or nm! For spin polarized systems, make sure you have sensible initial magnetic moments. Don’t do spin-paired calculations for molecules with an odd number of electrons. Before performing calculations of isolated atoms see Calculation of atomization energies.

  2. Use less aggressive density mixing.

    Try something like mixer=Mixer(0.02, 5, 100) or mixer=MixerSum(0.02, 5, 100), mixer=MixerDif(0.02, 5, 100) for spin-polarized calculations and remember to import the mixer classes:

    from gpaw import Mixer, MixerSum, MixerDif

    For some systems (for example transition metal atoms) it is helpful to reduce the number of history steps in the mixer to 1 (instead of 5).

  3. Solve the eigenvalue problem more accurately at each scf-step.

    Import the Davidson eigensolver:

    from gpaw import Davidson

    and increase the number iterations per scf-step eigensolver=Davidson(3).

    CG eigensolver tends converge fastest the unoccupied bands eigensolver='cg'.

  4. Use a smoother distribution function for the occupation numbers.

    Remember that for systems without periodic boundary conditions (molecules) the Fermi temperature is set to zero by default. You might want to specify a finite Fermi temperature as described here and check the convergence of the results with respect to the temperature!

  5. Try adding more empty states.

    If you are specifying the number of bands manually, try to increase the number of empty states. You might also let GPAW choose the default number, which is in general large enough.

  6. Use enough k-points.

    Try something like kpts={'density': 3.5, 'even': True} (see Brillouin-zone sampling).

  7. Don’t let your structure optimization algorithm take too large steps.

  8. Better initial guess for the wave functions.

    The initial guess for the wave functions is always calculated using the LCAO scheme, with a default single-zeta basis, i.e. one orbital for each valence electron. It is possible to use basis='szp(dzp)' to extract the single-zeta polarization basis set from the double-zeta polarization basis sets that are distributed together with the latest PAW datasets. You can also try to make a better initial guess by enlarging the Atomic basis set. Note that you first need to generate the basis file, as described in LCAO mode.

    Warning: this may in some cases worsen the convergence, and improves it usually only when the number of empty states is significantly increased.