Wavefunction Convergence

Some general comments on wavefunction optimizations:


Any optimization that takes more than 100 steps should be considered ``slow''.

Optimizations using ODIIS that have repeated resets (more than a few) will probably never converge. The default settings are optimal for well behaving systems, however changing this to

  ODIIS NO_RESET=20
   5

will frequently result in more robust behavior and reduce the memory usage at the same time. Please note that convergence for LSD is usually slower and more difficult than for unpolarized cases.

If the ODIIS optimizer gets stuck (more than a few resets) you can stop the run and restart using

  PCG MINIMIZE
  TIMESTEP
   20

The conjugate gradient minimizer with line search is much more robust. For LSD and larger systems it should be used from the start.

A typical behavior will be that after the restart the energy goes down and the gradient increases. This means that we are in a region where there are negative curvatures. In such regions the DIIS minimizer moves in the wrong direction. After some iterations we will be back to normal behavior, energy and gradient get smaller. At this point it may be save to switch back to ODIIS.

Sometimes, it can also be helpful to not only wait longer for a DIIS reset but also to diagonalize after repeated resets to get out of this region. This can be accomplished using

  ODIIS NO_RESET=20
    5
  LANCZOS DIAGONALIZATION RESET=2

Finally, starting a Car-Parrinello MD from a random wavefunction with all atom positions fixed, and using DAMPING ELECTRONS or ANNEALING ELECTRONS is an alternative to get to a converged wavefunction. Due to the unfavorable convergence behavior close to the minimum, you may want to switch to a different optimizer as soon as the fictitious kinetic energy of the electrons drops significantly below the typical range for a (normal) CP-MD run.

Wavefunction optimizations for geometries that are far from equilibrium are often difficult. If you are not really interested in this geometry (e.g. at the beginning of a geometry optimization or this is just the start of a MD) you can relax the wavefunction convergence criterion to $10^{-3}$ or $10^{-4}$ and do some geometry or MD steps. After that wavefunction optimization should be easier.

Some general remarks on comparing the final energies:


Converge the wavefunction very well, i.e. set CONVERGENCE ORBITALS to $10^{-6}$ or tighter.


Make sure that all parameters are the same:
- same geometry,
- same functional,
- same number of grid points (this may differ if you use different FFT libraries)
- same number of spline points for the pseudopotential

(IMPORTANT: the default for SPLINE POINTS has changed between different CPMD versions, $500 \to 3000 \to 5000$ ). A very good test is to start always from the same RESTART http and only do one single step. This way ALL energies have to be exactly the same and no problems with different convergence rates occur.