Chemistry Reference
In-Depth Information
identity (RI-MP2), the local approximation (LMP2), and both approximations
simultaneously (RI-LMP2). All comparisons are made using the reasonably
large aug-cc-pVTZ basis, and results were obtained with the MOLPRO pro-
gram.
89
Considering first the RI approximation, we see that it leads to errors
smaller than 0.01 kcal mol
1
for binding energies, while decreasing computa-
tional times from 14-17 h per point to 3-4 h per point, a substantial savings.
This reduced computational cost is even more impressive given that the canoni-
cal computation used point-group symmetry, whereas for technical reasons the
RI and local approximations in MOLPRO do not.
Considering next the local approximation, the LMP2 errors are some-
what larger but are still modest (the largest error is 0.155 kcal mol
1
for the
parallel-displaced configuration). Disappointingly, the computational cost for
LMP2 is actually greater than that of conventional MP2 for this test case. This
is probably because the conventional computation has the advantage of point-
group symmetry while the LMP2 computation does not, and the benzene
dimer is too small a system to have reached the crossover point where
LMP2 becomes less computationally expensive. Adding the RI approximation
to the LMP2 approximation (the RI-LMP2 column) decreases the computa-
tional time again, but the computational cost remains greater than that of
RI-MP2. The errors for RI-LMP2 are essentially the same as for LMP2.
It should be noted that the RI and local approximations are not comple-
tely ''black box.'' For the RI approximation, one must choose an auxiliary basis
set. In Table 4 we have used the coulomb/exchange-fitting (JK-fitting) auxiliary
basis
71
for the Hartree-Fock part of the computation, and the MP2-fitting aux-
iliary basis
72
for the computation of the correlation energy. It is possible that
further investigation might yield even more efficient auxiliary basis sets.
In the local correlation methods, the program must determine the
orbital
domains
that specify which pairs of orbitals should be correlated. This can be a
little problematic for studies of weakly interactingmolecules because as one scans
the potential energy surface, twomolecules might suddenly become close enough
that the orbital domains change, leading to a discontinuity in the potential energy
surface.
90
The simplest solution to this may be to define the orbital domains when
the molecules are a large distance away, and fix the domains in this form for the
entire study, as was done in the LMP2 study of Takatani and Sherrill.
88
Additionally, we were surprised to discover that the local approximation
in LMP2 can show significant errors when paired with the aug-cc-pVDZ basis
for several example noncovalent interactions.
88
This error was reduced by using
larger basis sets or if one removed diffuse functions on hydrogen. More systema-
tic investigations may shed light on which basis sets are best to use for LMP2
computations.
Spin-Component-Scaled MP2
Grimme has argued that the MP2 method exhibits a bias toward same-
spin excitations because Hartree-Fock-based methods already include some
