7.11. SCF Stability Analysis¶
The SCF stability analysis evaluates the electronic Hessian (with respect to orbital rotations) at the point indicated by the SCF solution to determine the lowest eigenvalues of the Hessian. If one or more negative eigenvalues are found, the SCF solution corresponds to a saddle point and not a true local minimum in the space considered in the analysis. A typical case are stretched bonds of diatomics, where the symmetry of the initial guess leads to a restricted solution instead of the often preferred unrestricted one. Several spaces are theoretically possible[780]. However, ORCA limits itself to the analysis RHF/RKS in the space of UHF/UKS or UHF/UKS in the space of UHF/UKS. As such, it is on the available for the SCF parts of DFT and HF.[80] We mention passing, that a stability analysis is also available for the CASSCF type wave function and is described elswhere in more detail (Section Detecting CASSCF Instabilities). In the following, HF is used to indicate both HF and KS. Consider the following input (unless indicated otherwise, default values are shown):
! BHLYP def2-SVP NORI
%scf
guess hcore # for illustrative purposes only
HFTyp UHF # default based on spin multiplicity
STABPerform true # default false
STABRestartUHFifUnstable true # restart the UHF-SCF if unstable
STABNRoots 3 # number of eigenpairs sought
STABMaxDim 3 # Davidson expansion space = MaxDim * NRoots
STABMaxIter 100 # maximum number of Davidson iterations
STABNGuess 4096 # size of initial guess matrix: 4096 x 4096
STABDTol 0.0001 # convergence criterion from iteration to iteration
STABRTol 0.0001 # convergence criterion max residual norm
STABlambda +0.5 # mixing parameter
STABORBWIN -1, -1, -1, -1, -1, -1, -1, -1 # defines the donor / acceptor spaces
# 4 parameters for RHF
# 8 paramters for UHF (4 alpha, 4 beta)
# orbital window, -1 refers to automatic determination
STABEWIN -5.0, 5.0 # lower and upper cutoff in Eh for automatic freezing
#---------------------------------------------------------------------------------
# alternative specification using a sub-block:
stab
NRoots 3
MaxDim 3 # etc.
end
end
* xyz 0 1
h 0.0 0.0 0.0
h 0.0 0.0 1.4
*
The determination of the electronic Hessian is structurally comparable
to the TDHF/CIS/TDDFT procedure. Thus, many options are very similar and
the user is encouraged to read the section on TDDFT (Section
Excited States via RPA, CIS, TD-DFT and SF-TDA) to clarify some of the options given
here. Since one is usually only interested in the qualitative
determination “stable or not?”, three roots should be sufficient to find
the lowest eigenvalue. By the same philosophy, StabMaxDim
,
StabMaxIter
, StabNGuess
and the convergence criteria were chosen.
The parameter StabLambda
refers to the \(\lambda\) of equation 37 of
reference [780], which determines the mixing of
the original SCF solution and the new orbitals to yield a new guess.
Choosing this value is not trivial, since positive and negative values
can lead to different new solutions (at least in principle). The
convergence of the ensuing SCF depends on it, as well, since all SCF
procedures require a sufficiently good guess to converge in a decent
number of iterations (or even at all).
The orbital window and the energy window can be specified. Note that the
StabEWIN
will be overridden by the appropriate StabORBWIN
values.
The automatic determination is also influenced by the
%method FrozenCore
settings. Tests have shown that significant
curtailing of the actual orbital window can drastically influence the
results to the point of qualitative failure.
Current limitations on the method are:
Only single-point-like calculations are supported. For geometry optimizations etc., one must use the guess MORead feature Choice of Initial Guess and Restart of SCF Calculations to employ the guess obtained here. Likewise, one must extract a geometry and run a separate calculation if one is interested in the SCF stability.
As for TDDFT, NORI, RIJONX, and RIJCOSX are supported. RI-JK is not supported.
Other, more advanced features like finite-temperature calculations and relativistic calculations (beside ECPs) are not possible at this time.
Overall, the user is cautioned against using the stability analysis blindly without critically evaluating the result in terms of energy difference and by investigating the orbitals (by the printout or by plotting). Its usefulness cannot be denied, but it is certainly not black-box.
An SCF stability analysis with default settings can be requested via STABILITY, SCFSTABILITY, SCFSTAB or STAB on the simple input line.