[CPMD-list] Geometry optimization!

[Salomon Billeter]

Dear Aijun Du,

transitory energy increases are a normal behavior of the L-BFGS
implementation in CPMD. In order to avoid the extra effort associated with
a proper line search algorithm and the robustness and performance  problems
of a fixed trust radius, this optimizer tries to predict the optimum step
length along the new search direction based on the quality of the previous
step in terms of the Wolfe criteria.

Sometimes, such a step can be too large, and the energy can increase. This
is no problem since the optimizer rejects that step and tries a smaller
displacement starting from the last good geometry. Even the wavefunction is
restored.

However, if the fraction of rejected steps is large, there is a problem.
L-BFGS reacts quite sensitively to an inaccurate gradient which can be
caused by poor convergence of the wavefunction, etc. In order to see what
is going on, I would recommend using
PRINT LSCAL ON
which turns on the diagnostics of the linear scaling optimizers.

You might also want to use
CONVERGENCE ADAPT
  0.02
to save unnecessary wavefunction optimization cycles for the early part of
the optimization, and to use
CONVERGENCE ENERGY
  0.05
to enforce more consistent energy changes for the late part of the
optimization (you can change these parameters, of course).

Best regards,
Salomon Billeter