Cases where PP needs improvement include molecules with several strongly correlated electron pairs that are all localized in the same region of space, and therefore involve significant inter-pair, as well as intra-pair correlations. For some systems of this type, Coupled Cluster Valence Bond (CCVB) is an appropriate method.891, 892 CCVB is designed to qualitatively treat the breaking of covalent bonds. At the most basic theoretical level, as a molecular system dissociates into a collection of open-shell fragments, the energy should approach the sum of the ROHF energies of the fragments. CCVB is able to reproduce this for a wide class of problems, while maintaining proper spin symmetry. Along with this, CCVB’s main strength, come many of the spatial symmetry breaking issues common to the GVB-CC methods.
Like the other methods discussed in this section, the leading contribution to the CCVB wave function is the perfect pairing wave function, which is shown in Eq. (6.44). One important difference is that CCVB uses the PP wave function as a reference in the same way that other GVBMAN methods use a reference determinant.
The PP wave function is a product of simple, strongly orthogonal singlet geminals. Ignoring normalization, two equivalent ways of displaying these geminals are
| (6.45) |
where on the left and right we have the spatial part (involving and orbitals) and the spin coupling, respectively. The VB-form orbitals are non-orthogonal within a pair and are generally AO-like. The VB form is used in CCVB and the NO form is used in the other GVBMAN methods. It turns out that occupied UHF orbitals can also be rotated (without affecting the energy) into the VB form (here the spin part would be just ), and as such we store the CCVB orbital coefficients in the same way as is done in UHF (even though no one spin is assigned to an orbital in CCVB).
These geminals are uncorrelated in the same way that molecular orbitals are uncorrelated in a HF calculation. Hence, they are able to describe uncoupled, or independent, single-bond-breaking processes, like that found in CH 2 CH, but not coupled multiple-bond-breaking processes, such as the dissociation of N. In the latter system the three bonds may be described by three singlet geminals, but this picture must somehow translate into the coupling of two spin-quartet N atoms into an overall singlet, as found at dissociation. To achieve this sort of thing in a GVB context, it is necessary to correlate the geminals. The part of this correlation that is essential to bond breaking is obtained by replacing clusters of singlet geminals with triplet geminals, and re-coupling the triplets to an overall singlet. A triplet geminal is obtained from a singlet by simply modifying the spin component accordingly. We thus obtain the CCVB wave function:
| (6.46) |
In this expansion, the summations go over the active singlet pairs, and the indices shown in the labellings of the kets correspond to pairs that are being coupled as described just above. We see that this wave function couples clusters composed of even numbers of geminals. In addition, we see that the amplitudes for clusters containing more than 2 geminals are parameterized by the amplitudes for the 2-pair clusters. This approximation is important for computational tractability, but actually is just one in a family of CCVB methods: it is possible to include coupled clusters of odd numbers of pairs, and also to introduce independent parameters for the higher-order amplitudes. At present, only the simplest level is included in Q-Chem.
Older methods which attempt to describe substantially the same electron correlation effects as CCVB are the IP983 and RCC986 wave functions. In general CCVB should be used preferentially. It turns out that CCVB relates to the GVB-IP model. In fact, if we were to expand the CCVB wave function relative to a set of determinants, we would see that for each pair of singlet pairs, CCVB contains only one of the two pertinent GVB-IP doubles amplitudes. Hence, for CCVB the various computational requirements and timings are very similar to those for GVB-IP. The main difference between the two models lies in how the doubles amplitudes are used to parameterize the quadruples, sextuples, etc., and this is what allows CCVB to give correct energies at full bond dissociation.
A CCVB calculation is invoked by setting CORRELATION = CCVB. The number of active singlet geminals must be specified by GVB_N_PAIRS. After this, an initial guess is chosen. There are three main options for this, specified by the following keyword
CCVB_GUESS
Specifies the initial guess for CCVB calculations
TYPE:
INTEGER
DEFAULT:
NONE
OPTIONS:
1
Standard GVBMAN guess (orbital localization via GVB_LOCAL + Sano procedure).
2
Use orbitals from previous GVBMAN calculation, along with SCF_GUESS = read.
3
Convert UHF orbitals into pairing VB form.
RECOMMENDATION:
Option 1 is the most useful overall. The success of GVBMAN methods is often
dependent on localized orbitals, and this guess shoots for these. Option 2 is
useful for comparing results to other GVBMAN methods, or if other GVBMAN
methods are able to obtain a desired result more efficiently. Option 3 can be
useful for bond-breaking situations when a pertinent UHF solution has been
found. It works best for small systems, or if the unrestriction is a local
phenomenon within a larger molecule. If the unrestriction is non-local and the
system is large, this guess will often produce a solution that is not the
global minimum. Any UHF solution has a certain number of pairs that are
unrestricted, and this will be output by the program. If GVB_N_PAIRS
exceeds this number, the standard GVBMAN initial-guess procedure will be used
to obtain a guess for the excess pairs
For potential energy surfaces, restarting from a previously computed CCVB solution is recommended. This is invoked by GVB_RESTART = TRUE. Whenever this is used, or any time orbitals are being read directly into CCVB from another calculation, it is important to also set:
SCF_GUESS = READ
MP2_RESTART_NO_SCF = TRUE
SCF_ALGORITHM = DIIS
This bypasses orthogonalization schemes used elsewhere within Q-Chem that are likely to jumble the CCVB guess.
In addition to the parent CCVB method as discussed up until now, we have included two related schemes for energy optimization, whose operation is controlled by the following keyword:
CCVB_METHOD
Optionally modifies the basic CCVB method
TYPE:
INTEGER
DEFAULT:
1
OPTIONS:
1
Standard CCVB model
3
Independent electron pair approximation (IEPA) to CCVB
4
Variational PP (the CCVB reference energy)
RECOMMENDATION:
Option 1 is generally recommended. Option 4 is useful for preconditioning, and
for obtaining localized-orbital solutions, which may be used in subsequent
calculations. It is also useful for cases in which the regular GVBMAN PP code
becomes variationally unstable. Option 3 is a simple independent-amplitude
approximation to CCVB. It avoids the cubic-scaling amplitude equations of CCVB,
and also is able to reach the correct dissociation energy for any molecular
system (unlike regular CCVB which does so only for cases in which UHF can reach
a correct dissociate limit). However the IEPA approximation to CCVB is
sometimes variationally unstable, which we have yet to observe in regular
CCVB.
Example 6.32 N molecule in the intermediately dissociated region. In this case, SCF_ALGORITHM DIIS is necessary to obtain the symmetry unbroken RHF solution, which itself is necessary to obtain the proper CCVB solution. Note that many keywords general to GVBMAN are also used in CCVB.
$molecule 0 1 N 0 0 0 N 0 0 2.0 $end $rem JOBTYPE = sp UNRESTRICTED = false BASIS = 6-31g* EXCHANGE = hf CORRELATION = ccvb GVB_N_PAIRS = 3 CCVB_METHOD = 1 CCVB_GUESS = 1 GVB_LOCAL = 2 GVB_ORB_MAX_ITER = 100000 GVB_ORB_CONV = 7 GVB_RESTART = false SCF_CONVERGENCE = 10 THRESH = 14 SCF_GUESS = sad MP2_RESTART_NO_SCF = false SCF_ALGORITHM = diis MAX_SCF_CYCLES = 2000 SYMMETRY = false SYM_IGNORE = true PRINT_ORBITALS = true $end