[CPMD-list] slow geometry optimization - would underrelaxation help?

[Salomon Billeter]

Dear Wolfgang,

instead deliberately using too small a step, you can require your trial
move to satisfy the two Wolfe conditions (sufficient decrease and curvature
conditions). The L-BFGS geometry optimizer in CPMD (keyword LBFGS in the
&CPMD section) has a built-in dynamic trust radius method for predicting
maximum trial move lengths based on these conditions (and for rejecting
steps in case they fail to reduce the energy). In addition, I'd recommend
using adaptive wavefunction convergence criteria (CONVERGENCE ADAPT and
CONVERGENCE INITIAL) for saving wavefunction cycles in earlier stages of
the optimization.

Best regards,
Salomon