Re: coupled cluster methods question
On Mar 13 5:37pm, "DOUGLAS A. SMITH" wrote:
> 1. Is it the best method currently available for open shell molecules?
On Mar 14 3:11am, Ole Swang wrote:
> This seems a rather vague concept to me... [And goes on to make some
> good points about zeroth-order descriptions, multireference character,
> etc.]
This is indeed very vague -- and very application dependent, so I
won't try to answer it. I will mention, however, that CC methods,
especially those including triple excitations, can handle many systems
that might otherwise be considered multireference quite well.
On Mar 14 3:11am, Ole Swang wrote:
> (unfortunately, as far as I know nobody has implemented a general
> multireference coupled cluster algorithm as yet).
Multi-reference CC methods are presently a very active research in a
number of groups. Work is progressing rapidly.
On Mar 13 5:37pm, "DOUGLAS A. SMITH" wrote:
> 2. How will CCD or CCSD(T) perform on closed shell molecules? I have
> several systems to compare, and I would like to use a single
> consistant method for my calculations.
In general CC methods well perform quite well for closed shells, and
if the study is done carefully, there is no reason why CC results for
closed- and open- shell systems can't be compared.
For a more detailed comparison, it is necessary to have a better idea
of what you intend to study. There are numerous studies in the
literature of the quality of CC methods. Your best bet is probably to
look up some of these involving species related to your own interests.
Note: CCD is rarely used these days -- CCSD is much better at minimal
additional cost, particularly if you intend to treat open shells.
On Mar 13 5:37pm, "DOUGLAS A. SMITH" wrote:
> 3. Is there a minimum basis set that should be used for CCD or
> CCSD(T)? Is there a practical limit to the size of a molecule
> which can be looked at using these methods?
On Mar 14 3:11am, Ole Swang wrote:
> Generally, a double-zeta valence plus polarization basis set (like,
> say 4-31g*) should be thought of as a minimal basis set for
> calculations involving correlation (something like STO-3g for
Hartree-Fock).
The collected experience in our group is also that you need at least a
polarized double-zeta quality basis. I would caution, however, that
4-31G* _doesn't_ qualify. From work we've done, even 6-31G* is marginal
for any system that is not exceedingly well-behaved. Basically, the
restriction of exponents use in the construction of the Pople-type
basis sets means that even something with enough funcitons to be
double-zeta really doesn't have the necessary flexibility. You're
much better off with (for example) a nice standard Dunning DZP basis.
For larger-than-DZP bases, and assuming you have a package that can
handle it, generally contracted basis sets, such as Almlof & Taylor's
ANO bases or Dunnings "correlation consistent" bases are a significant
improvement over the more traditional segmented contraction sets.
On Mar 14 3:11am, Ole Swang wrote:
> There is, of course, a practical limit for the size of the systems which
can
> be studied using CC, or any other ab initio method. Where the limit goes
> is depedent of the available resources.
This is true for _any_ method, not just ab initio ones -- there is
always a tradeoff, and there is not enough informaton provided to give
meaningful answers.
In our group, we generally study molecules that _aren't_ well-bahaved
and require a high-level correlated treatment. We're accustomed to
doing CC with triples corrections of various sorts for almost
everything we work on. Given enough CPU and disk resources, it is
possible to routinely calculate triples corrections for several
hundred basis functions. There is no reason in principle why one
can't go beyond that, but since triples costs scale roughly as n**7 (n basis
functions) or higher, it can become quite expensive, and at some point
you have to ask yourself if its really worth it (larger molecules are
often _better_ behaved -- and therefore less in need of triples
corrections -- than smaller ones).
On Mar 14 3:11am, Ole Swang wrote:
> in most cases a CC treatment is a factor 3 or 6 or something more
> expensive than a CI treatment. (there shouldn't be an exponential
> differennce betweeen the two).
This is a fairly common misconception. For a CC calculation and a CI
calculation with the same excitation operators included and the same
underlying formalism, the cost is essentially the same. A difference
you will see is that most CI codes are completely spin adapted, while
most CC codes are written in a spin-orbital framework.
The differences between a CC and CI implementation, then, depend on
whether the spin-orbital CC has been implemented to also take
advantage of the spin adaption when possible (for a closed shell
reference). For an open-shell molecule, there are effectively several
CC equations being solved simultaneously, to it's going to take longer
than a spin-adapted CI code.
On the other hand, CC methods are size-extensive (important for
treating larger systems), and recover more of the correlation energy
for a fixed amount of computational resources than CI.
--
David Bernholdt bernhold "-at-" qtp.ufl.edu
Quantum Theory Project bernhold "-at-" ufpine.bitnet
University of Florida
Gainesville, FL 32611 904/392 6365