[CPMD-list] imaginary frequency for H2O and (H2O)2

[Salomon Billeter]

Hi Lan-Feng,

>   For the sym=0 run, CENTER MOLECULE, ISOLATED MOLECULE, CONVERGENCE
ORBITALS 1.D-8, CONVERGENCE GEOMETRY 1.D-4, > 10 A cubic cell, 100 Ry, PBE,
GC-CUTOFF 5.D-8 (raised from default 1.D-8):
>
>    ATOM          COORDINATES            GRADIENTS (-FORCES)
>    1  O  6.7858  9.4486  9.5174   1.169E-04  2.731E-08 -1.883E-05
>    2  O 12.2342  9.4486  9.3849   7.327E-05 -2.303E-08  1.437E-04
>    3  H  6.2825 10.9020  8.5169   1.717E-04  6.366E-07 -4.552E-05
>    4  H  6.2825  7.9953  8.5169   1.717E-04 -6.612E-07 -4.554E-05
>    5  H 10.3927  9.4486  9.5889   4.167E-05  3.440E-08 -1.624E-04
>    6  H 12.8841  9.4486 11.0973  -2.178E-04 -2.295E-09  2.536E-04
>
> GNMAX=  2.536102E-04 [9.64E-07], GNORM=  1.161934E-04, DETOT=     3.409
E-10.
>
>   I don't fully understand why it converged here, since both the max and
the average forces are still > 1.D-4,
> but anyway it's a bit better than using default GC-CUTOFF, where
convergence was achieved only for threshold
> 4.D-4.

even 10^-4 atomic units for the convergence of the forces on the ions might
be too loose for a subsequent vibrational analysis. However, depending on
the geometry optimizer you are using, the fact that your optimization stops
even earlier than that may hint that are near a saddle point (what the
vibrational analysis supports). Which geometry optimizer are you using? I'd
try using LBFGS (don't forget to say PRINT LSCAL ON to get information
about the progress, especially in a hard case).

>   Then it failed to compute frequencies by linear response:
>
>  PROGRAM STOPS IN SUBROUTINE SDLINRES| HOCKNEY PS not impl. [PROC=   2]
>
>   Sounds like the Hockney Poisson solver (default for sym=0) is not yet
available for linear response. Could the > other 2 solvers (Tuckerman and
Mortensen) be ok?

That's because the second variational derivative of the electrostatic
energy is not implemented for Hockney. You can use either Tuckerman or
Mortensen, but your box (10 A for this system) most likely is not large
enough for them, possibly even for Hockney. Before starting to use them,
you might want to scan this list for the several postings regarding the box
requirements of these Poisson solvers or see the section on this in the
manual. Which Poisson solver did you use for the optimization of the
geometry?

Best regards,
Salomon