[CPMD-list] stack entry size mismatch

[Salomon Billeter]

Hi,

on our machines, your calculation using P-RFO converges within only 15
cycles to a reasonable transition state (one negative eigenvalue,
reasonable geometry, convergence criteria satisfied). However, also both
our initial and final energies differ by almost 0.1 atomic units from the
values in your output.

We probably best upload a summary to the list when we find the source of
the problems.

However, some comments of possible general interest here:
- Under normal circumstances, the number of steps a good local geometry
optimizer needs to get to a stationary point should be in the order of
magnitude of the number of degrees of freedom. A much larger number of
steps should not be necessary.
- Useful diagnostics for L-BFGS and P-RFO can be switched on using PRINT
LSCAL ON. In the case of P-RFO, this includes among others the eigenvalue
structure and the trust radius and mode overlap histories.
- If you are not using the microiterative (linear scaling) scheme for
transition state search, you can remove the components of translational
(and, depending on your system input, rotational) motion of the system from
the Hessian. These components can e.g. be caused by the XC potential
evaluated on a finite real-space mesh and anharmonicities entering through
the finite-difference scheme. The keyword is PRFO PRJHES.
- It's a good idea to allow some extra space for the changing shape when
optimizing the geometry of an isolated system. Your box is somewhat tight,
even for Hockney, and the is cutoff somewhat low, although this does not
explain the error.

Best regards,
Salomon