[CPMD-list] floating point exception & unconverged geometry

[Salomon Billeter]

Dear Masroor,

due to its scaling properties with system size, the conjugate gradient
geometry optimizer it is mainly used for large systems. A drawback is its
slower convergence than many other algorithms. For your system, I would try
another optimizer.

The next version of CPMD will have a new, linear scaling geometry optimizer
based on the L-BFGS algorithm which is also suitable for smaller systems.
It has a trust radius algorithm based on the Wolfe conditions ensuring that
the trial moves are neither too small nor too big to reach convergence
quickly.

In the meantime, you can use one of the other geometry optimizers in CPMD,
e.g. GDIIS with BFGS update. It can give you very fast convergence, but the
trust region it is using does not always ensure convergence (experimenting
with the time step can help). If you see the total energy rising, it's a
good idea to stop the optimization and to restart from the saved
wavefunction and coordinates.

Best regards,
Salomon