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 Ångström 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 Getting started with GPAW.

  2. Try tuning the mixer settings, while the defaults are designed to work great in many cases, tuning the mixer can improve convergence for certain systems.

    Example:

    mixer={'beta': 0.04,  # Reducing or increasing beta
       'backend': 'msr1',  # Try other mixing algorithms
       'method': 'difference',  # Changing the spin-driver
       'nmaxold': 8,  # Reducing or increasing the history
       'weight': 100,  # Changing the damping term}

    Note that the above example show the possible changes to try, however, a combination of the above and default parameters is likely to be the optimimal choice for your system.

  3. Solve the eigenvalue problem more accurately at each scf-step (PW- and FD-mode specific). This can be easily achieved by allowing the eigensolver more iterations per step eigensolver={'niter': 5}.

  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} (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.