There are three methods available for wavefunction optimization. The default
method (ODIIS) is the most powerful method but sometimes it needs
some assistance from outside. We have implemented some empirical
rules for restarting the DIIS procedure in case that the optimization
gets stuck. Nevertheless it may happen that one has to stop a run and use
at least for some steps one of the other methods implemented.
You can stop a run at the next possible breakpoint by issuing the
command
where pid is the process ID of your run.
Creating a file EXIT (e.g. with touch EXIT) has the same
effect. On the SP1 kill -30 does not work.
The default number of DIIS vectors retained is 10. For big systems 5
vectors need about the same additional memory as the PCG method.
The preconditioned conjugate gradient method
(PCG) is implemented either with a fixed steplength h that is calculated
from the timestep 
and the electron mass 
or with a three point line search algorithm.
The line search is activated by PCG MINIMIZE. This line search doubles
the cost per step with respect to the simple PCG method.
The steepest descent method also uses the steplength h.
The preconditioning in the DIIS and CG methods is controled through the
keyword HAMILTONIAN CUTOFF. The default value is 0.5 a.u. and works
reasonably well for most systems. A higher value has the effect of "less"
preconditioning, a very high value is equivalent to disabeling preconditioning.
With the PROJECT keyword you can choose from different forms of gradients.
FULL refers to
DIAGONAL to
and NONE is just 
.
Default is to use the full projection, but steepest descent also works with the
other forms. The usage of other combinations like ODIIS and PROJECTION DIAGONAL
may work in certain cases but is on your own risk.
If you want to quench the wavefunctions to the Born-Oppenheimer surface within
a molecular dynamics run, start the wavefunction optimization with
RESTART WAVEFUNCTION COORDINATES VELOCITIES
After convergence of the wavefunctions you can restart the MD run with
RESTART WAVEFUNCTION COORDINATES VELOCITIES
QUENCH ELECTRONS
There is one known situation when wavefunction optimization with ODIIS fails. Whenever the dominant direction of the gradient is in a region of negative curvature the basic assumption of ODIIS (the surface is locally approximated by a quadratic function) is not valid. A typical behaviour in these cases is, that the total energy remains constant as does the gradient. In this case one can switch to the conjugate gradient option. In a first part of the CG run the gradient will increase as will the gain in total energy. As soon as the gradient is decreasing again, it is save and also faster to switch back to the ODIIS method.