Vous êtes sur la page 1sur 9

WWW.C-CHEM.

ORG REVIEW

A Two-Scale Approach to Electron Correlation


in Multiconfigurational Perturbation Theory
Pooria Farahani,[a] Daniel Roca-Sanjuan,[b] and Francesco Aquilante*[a,c]

We present a new approach for the calculation of dynamic approximation are contrasted by the substantial savings in
electron correlation effects in large molecular systems using both storage and computational demands compared to the
multiconfigurational second-order perturbation theory full CASPT2 calculation. Provided that static correlation effects
(CASPT2). The method is restricted to cases where partitioning are correctly taken into account for the whole system, the pro-
of the molecular system into an active site and an environ- posed scheme represent a hierarchical approach to the elec-
ment is meaningful. Only dynamic correlation effects derived tron correlation problem, where two molecular scales are
from orbitals extending over the active site are included at the treated each by means of the most suitable level of theory.
CASPT2 level of theory, whereas the correlation effects of the C 2014 Wiley Periodicals, Inc.
V
environment are retrieved at lower computational costs. For
sufficiently large systems, the small errors introduced by this DOI: 10.1002/jcc.23666

Introduction based on localized Cholesky orbitals.[18] Such gracious compu-


tational scaling is lost when performing the subsequent CD-
Recent developments in multiconfigurational quantum chemis- CASPT2 calculation, where the only possibility explored so far
try,[1] in particular within the complete active space self- to lower its computational costs has been to use the frozen
consistent field second-order perturbation theory (CASSCF/ natural orbital (FNO) approximation,[19,20] that reduces the size
CASPT2) method[2] have made it feasible to perform such cal- of a large secondary orbital space while retaining the accuracy
culations on much larger molecular systems and with more of the method.
extensive basis sets. The limit in basis set size for routine
Here, we shall describe and illustrate an approach that can
investigations has thereby been extended to the range 1000
be used to speed-up even more drastically CD-CASPT2 calcula-
2000 basis functions.[37] This has been possible through the
tions by reducing the size of the secondary, as well as the
introduction of the Cholesky decomposition (CD) representa-
inactive orbital spaces. The approach is useful in cases where
tion of the electron repulsion integrals,[3,4,8,9] that later on
the active orbitals are located in a limited region of the mole-
evolved into the so-called ab initio density fitting approxima-
cule, thus belonging to an active site, in contrast to the
tion.[1015]
remaining portion of the molecule (environment). As dynami-
Cholesky-based density fitting can today be used at most
cal correlation effects are known to be short-ranged, it seems
levels of theory[16] ranging from the integral evaluation,
reasonable to assume that relative energies can be computed
through SCF and CASSCF to multiconfigurational second-order
just by taking into account the correlation energy contribution
perturbation theory. A detailed account of the implementation
arising from the active site (henceforth indicated by A). Using
of this approximation to CASSCF wave functions can be found
a localization procedure that will be described below, it is
in the study of Aquilante et al.[4] The use of CD approximation
then possible to divide the inactive and secondary orbital
at the CASPT2 level of theory was first illustrated in an applica-
tion to the electronic structure of the low-lying electronic state
of the complex CoIII(diiminato)(NPh).[3] This calculation on a [a] P. Farahani, F. Aquilante
Department of Chemistry - Angstrom, The Theoretical Chemistry Pro-
system with 43 atoms utilized a general contracted Gaussian
gramme, Uppsala University, P. O. Box 518, SE-751 20, Uppsala, Sweden.
basis set with a total of 869 basis functions. A two-root multi- E-mail: francesco.aquilante@gmail.com
state calculation took 3.2 days wall time on a single AMD [b] D. Roca-Sanju an
Opteron 148, 2.2 GHz workstation, equipped with 1 GB of Instituto de Ciencia Molecular, Universitat de Valencia, P.O. Box 22085, ES-
memory. This should be compared with the time used for the 46071, Valencia, Spain
[c] F. Aquilante
evaluation of the Cholesky vectors, 3.5 h, and the time per
Dipartimento di Chimica G. Ciamician, Universit a di Bologna, V. F. Selmi
iteration for the CASSCF calculations, of 3.9 min (using a 10 2, 40126, Bologna, Italy
in10 active space). Since this first application, it became clear Contract grant sponsor: Swedish Natural Science Research Council (VR).;
from these numbers that the bottleneck in the use of the CD- Contract grant sponsor: Spanish MINECO (Ministero de Economa y
Competitividad); Contract grant number: CTQ2010-14892; Contract grant
based CASSCF/CASPT2 method with large basis sets and many
sponsor: Juan de la Cierva programme; Contract grant number: JCI-
electrons would be the CD-CASPT2 step. In fact, systems with 2012-13431; Contract grant sponsor: Italian Ministry of Education and
few hundred atoms can be treated efficiently at the CASSCF Research (MIUR; F.A.); Contract grant number: RBFR1248UI
level by means of the local exchange screening algorithm[17] C 2014 Wiley Periodicals, Inc.
V

Journal of Computational Chemistry 2014, 35, 16091617 1609


REVIEW WWW.C-CHEM.ORG

spaces into two partitions such that one of them has no con- space), the dimensionality of the linear equations for CASPT2
tributions from the atomic basis functions in the active site. amplitudes grows with the fourth power of the system size.
The occupied (secondary) orbitals belonging to the latter can As anticipated in the previous section, most of these compu-
be left frozen (deleted) in the CASPT2 calculation. In this pic- tational complexities can be alleviated by moving from a canoni-
ture, the nondynamical correlation effects are nonetheless cal orbital picture to one based on localized orbitals. The
taken into account without additional approximations, as the definition and uses of such localized orbital approach traces out
CASSCF calculation is performed for the entire system. Finally, the ideas central to the local correlation methods of Pulay.[34]
an estimate of the dynamical correlation energy due to the Here and in the following, greek subscripts are used to refer to
environment (henceforth, B) can still be computed. As shown AOs, whereas p, q, r,. . . are used to denote general molecular
in the next section, by neglecting the coupling terms between orbitals (MOs). For inactive and secondary MOs we use sub-
A and B in the CASPT2 equations, the set of amplitudes for B scripts i, j, k, . . . and a, b, c, . . . , respectively, whereas t, u, v, . . .
can be easily determined at the cost of second-order Mller identify active orbitals. As the active orbitals are assumed to be
Plesset perturbation theory (MP2). localized, they naturally define an active site (A) of the molecule,
A similar approach has been proven effective in reducing in contrast to the remaining set of atoms, the environment (B).
the costs of MP2 calculations.[21] More generally, the idea of For a given active orbital uu and a given atom X, we compute its
P
partitioning a large molecular system into regions to be atomic gross Mulliken population, QuX 5 < uuX juu > 5 lX m CluX
treated at different levels of theory is certainly not new. Multi- SlX m Cmu (where C is the MO coefficient matrix, S is the AO overlap
scale modeling by means of embedding methods using the matrix and lX indicates that only the AO basis functions on
charge-densities as the only descriptors for the environment atom X are included in the sum). For each atom X of the mole-
has become very popular.[2224] Although not required in the cule, the euclidean norm of QX 5fQuX ; u51; Nact g is compared
exact embedding theory, spatial separation between the against a given threshold (s) to assign X to the region A or B.
active site and the environment is a condition for robustness Once the active site has been defined, the inactive orbitals
in the application of the approximate protocols.[25] In theoreti- have to be localized. In principle, any of the standard localiza-
cal studies of large molecules for which combined quantum tion techniques can be used for this purpose,[36,37] but here we
mechanics-quantum mechanics (QM/QM) or quantum used instead the so-called Cholesky MOs, proposed by one of
mechanics-molecular mechanics (QM/MM) models can be the present authors,[18] and proven useful in combination with
applied, also the ONIOM scheme[26] is very widespread. How- screening techniques for building Fock-type matrices.[17] In con-
ever, in these cases the molecule is typically split into parts, trast to standard techniques, the Cholesky localization is noni-
with the consequence that dangling bonds need to be satu- terative as it requires only one CD of the AO density matrix
P
rated with link atoms. This artificial split of the molecule pre- computed from the inactive orbitals, DIlm 5 inact k Clk Cmk . Despite
cludes any possibility to recover the exact result of the full the somewhat less optimal locality of the resulting orbitals, the
molecular calculation. That is not the case, at least in principle, computational advantages compared to standard techniques
for analogous approaches[21,2733] based on the local correla- are sufficient to justify our choice in the present context. Each
tion paradigm of Pulay.[34] These so-called local correlation localized inactive orbital ui is, then, assigned to either A or B
methods have gone from the accurate and efficient evaluation depending on the absolute value of QiA 5 < uiA jui >, where the
of ground-state correlation energy[21,29,34,35] to computational symbol uiA is used to indicate the same ui but truncated to the
techniques for spectroscopy of large molecules.[27,30,31,33] The AO basis functions on atoms of the active site only.
present work can be considered an attempt to investigate the For the virtual orbitals, the situation is more involved, as
applicability of some of these ideas in a new context, namely standard localization techniques can suffer from convergence
that of multiconfiguration wave function theory. problems. Observing instead that the AO basis is local by defi-
nition, it is tempting to use these orbitals to span the second-
ary space. Orthogonality to the remaining subspaces is
Methodology ensured by projecting out these components, and therefore,
producing the so-called projected atomic orbitals
Definition of the active site
(PAOs).[34,38,39] These orbitals constitute a nonorthogonal and
The active space in CASSCF/CASPT2 calculations is in many linearly dependent basis for the secondary space. In local cor-
cases of interests confined to a particular region of the mole- relation methods[27,30,31,3335,3842] linear dependence is elimi-
cule, comprising very few atoms. Moreover, only a very limited nated by diagonalizing the overlap matrix and subsequently
number of orbitals belonging to these atoms play a role in removing orbitals corresponding to zero eigenvalues. Such
nondynamical correlation effects, and therefore, the number of procedure is in principle performed for each pair of occupied
active orbitals (Nact) is usually very small (limited to about 14 orbitalsone such an overlap matrix has, therefore, linear
18 units). In such situations, the bottleneck of the CASPT2 dimension proportional to the union of the small subsets of
treatment is represented by the fact that the number of exci- PAOs that are accessible by excitation from a given localized
tations in the first-order wave function (see Andersson and occupied orbital (orbital domains). For the definition of these
Roos[2]), increases rapidily with the size of the atomic orbital orbital domains, each PAO is formally assigned to the atom
(AO) basis set. In particular, due to the presence of external center of the corresponding AO basis function (e.g., as in Ham-
excitations (two orbital index belonging to the secondary sub- pel and Werner[38]). In our case, this approach is perfectly

1610 Journal of Computational Chemistry 2014, 35, 16091617 WWW.CHEMISTRYVIEWS.COM


WWW.C-CHEM.ORG REVIEW

suited to partition the secondary space between the regions A tudes t is determined through the solution of the following
and B. However, as the removal of linear dependencies is per- linear equations
formed separately for the PAOs assigned to each region, it is X
not guaranteed that the overall number of resulting orbitals ^ 0 > ;
< Um jH^ 0 2E0 jUl > tl 52 < Um jHjW (3)
(orthonormal within each region) is identical to the initial l

number of canonical secondary orbitals. This is a typical situa-


with H^ 0 and E0 5 < W0 jH^ 0 jW0 > representing the zeroth-
tion in local correlation methods, where a much larger of sec-
order Hamiltonian and zeroth-order energy, respectively.
ondary orbitals (if summed over all domains) is obtained.
Details about the solution of eq. (3) can be found in Ref. [43]
Although not of crucial importance for our method, we opted
and the references therein.
for an approach free from such a potential overparametriza-
Assuming the parametrization of eq. (2), any set of orbitals,
tion. As explained in Ref. [18], a set of nonorthogonal but line-
separately spanning the three subspaces (inactive, active, and
arly independent PAOs spanning the entire secondary space
secondary) can be used. In this work, we want to prove the
can be obtained through CD of a properly defined density-
advantages of using localized orbitals for the inactive and sec-
type matrix in AO basis:
ondary subspaces in situations where the active orbitals are
DVlm 5S21 2Dlm ; (1) confined to a small region of the molecule. We start by recall-
ing the basic assumption of single-reference local correlation
where S is the overlap matrix in AO basis and D is the projec- method, for which the first-order wave function can be
tor onto the occupied (inactive 1 active) space. These set of expressed as (see, e.g., Schutz et al.[35]:
Cholesky MOs retain most of the locality of the initial PAOs XX
W1 5 ~t ab ab
ij Uij : (4)
and their number is identical to the initial number of canonical
ij ab2ij
secondary orbitals. Moreover, as the CD procedure is per-
formed stepwise by selecting the current largest diagonal ele-
The symbol [ij] indicates the so-called pair domain, thus the
ment of the density-type matrix, each Cholesky MO can be
union of the orbital domains of ui and uj . As each orbital is
assigned to the atom carrying the corresponding AO basis
supposedly localized, only pairs of orbitals that share at least
function. In this way, there is no ambiguity in partitioning
one atom (strong pairs) need to be included in the expansion
these orbitals between the two regions A and B. After that,
the secondary orbitals in each region can be separately ortho- eq. (4) to recover about 90% or more of the correlation
normalized, therefore, keeping their local character. energy.[35] Adapting this logic to CASPT2, and assuming the
The final step before the energy evaluation itself consists in partitioning of the molecule in active site and environment
deriving a new set of canonical orbitals for each of the two described above, we can replace the doubly excited configura-
regions, such that the simplicity of the calculation in canonical tions in eq. (2) with the following approximate expression:
basis can be retained, and existing computer codes can be used
A
X B
X
with no extra programming effort. These (localized) canonical ~t pr ~t ab
W1  ^ ^ 0
qs E pq E rs W 1
^ ^ 0 :
ij E ai E bj W (5)
orbitals and their orbital energies can be obtained from the initial pqrs aibj
canonical orbital energies via diagonalization of the partial
(restricted to orbitals in either A or B) occupied (virtual) Fock Moreover, the set of linear eq. (3) splits into two uncoupled
matrix, F5UUT , where U is the orthogonal transformation matrix subsets, where the one corresponding to the orbitals in B is
for the occupied (virtual) orbital localization. In such canonical now equivalent to an MP2 model. In our (localized) canonical
basis, standard CASPT2 codes can be used without any modifica- basis, this implies that the approximate contribution to the
tion. The only thing that changes compared to the full calculation correlation energy due to B is given in a noniterative and
is the size of the inactive and virtual orbital space in input, now straightforward manner by the canonical MP2 expression. The
restricted only to canonical orbitals localized in A. For this reason, total PT2 energy can, then, be approximated as
a name such as freeze-and-delete (FD) CASPT2 seems appropri-
A1B A
ate for this approach. ECASPT2  ECASPT2 1E2B : (6)

CASPT2 formulated in localized orbital basis The complexity of the CASPT2 calculation on the whole sys-
tem is, therefore, reduced to that of the calculation of the
In CASPT2, the first-order correction to the CASSCF wave func-
(small) set of amplitudes associated with the orbitals assigned
tion W0 includes contributions from only singly and doubly
to the active site. As discussed in Ref. [3], in fact, the need to
excited configurations, thus:
store/retrieve the set of two-electron integrals in MO basis in
X X X the present implementation of the CD-CASPT2 method can
W1 5 tqp E^pq W0 1 E pq E rs W0 5
pr ^ ^
tqs tl Ul ; (2)
pq pqrs l
seriously limit its efficiency and capability for large molecules
and large AO basis sets. Conversely, the CD-MP2 energy
where in the singlet one-electron excitation operators the indi- expression in canonical orbital basis can be computed without
ces q, s are restricted to inactive or active orbitals and p, r to storing the (aijbj) integrals on disk[10,44] and it is, therefore,
active or secondary orbitals. The set of (nonredundant) ampli- much more affordable than the corresponding CD-CASPT2

Journal of Computational Chemistry 2014, 35, 16091617 1611


REVIEW WWW.C-CHEM.ORG

calculation. Furthermore, efficient linear-scaling algorithms for This implies that linear scaling of the computational costs can-
local correlation methods have been developed over the not be achieved whenever the contribution from the environ-
years[35,41] that could be used for the MP2 energy correction ment is to be included. Werner and coworkers[29] have shown
to break down the computational scaling of this part of the that by retaining a fully localized orbital picture, it is possible
calculation. We finally highlight the fact that although the to devise multiregion type correlation approaches for single-
expression for computing E2B is identical to the well-known reference wave functions. Conversely, we feel that the scaling
MP2 formula, the resulting value will in general be very differ- of the FD-CASPT2 is not the prominent issue, if analyzed in
ent from the one obtained using the closed-shell HF reference, light of the overall reduction of computational costs compared
due to the use of pseudocanonical orbitals, obtained by diago- to the CASPT2 calculation for the whole molecule.
nalization of the CASSCF Fock matrix.
Before closing this section, it is worthwhile comparing more
in detail FD-CASPT2 with local correlation methods. Although
Sample Calculations
they share a common origin, there are differences to be We have investigated the reliability of the FD-CASPT2
pointed out. For instance, in local correlation methods one can approach, with the MP2 corrections of eq. (7) for the environ-
approach the exact evaluation of the correlation energy by ment, in a typical application of the CASPT2 method, namely
including pairs of orbitals which do not have common atoms spectroscopy of isolated and solvated molecules.
in their respective domains. Depending on the relative dis- Our tests have been performed using different thresholds
tance, so-called weak and distant pairs are included in the (s) for FD-CASPT2 and comparison with the full CASPT2 results,
model[35,41,42,45] at various degrees of approximation. Due to using a development version of the MOLCAS quantum chemis-
the fact that the bottleneck in these calculations is determined try software.[46] To minimize the effect of weakly interacting
by the treatment of strong and weak pairs only,[45] the overall intruder states in some CASPT2 computations, the imaginary
scaling of the method can be maintained linear in virtue of level-shift technique (0.2 au) has been applied. To correct the
the linear scaling number of such pairs in localized orbital systematic error of the CASPT2 method affecting the processes
basis. In our case, the CASPT2 treatment is performed only for where the number of paired electrons is changed, the ioniza-
the orbital pairs belonging to the active site, thus it is equiva- tion potential electron affinity (IPEA) modified zeroth-order
lent to select some but not all strong pairs for the highest Hamiltonian has been used with a value of the IPEA parameter
level of theory. The environment is treated at MP2 level. For of 0.25 au.[47,48] The Gaussian 09[49] program has been used
infinite separation between the two regions, this approxima- for geometry optimization with the DFT and MP2 methods.
tion is exact, but in reality our method suffers from the
neglect of (nonstrong) pairs between the regions. A consistent
Deoxythymidine
way to approach the exact CASPT2 correlation energy for the
entire system seems, therefore, impossible. Conversely, similarly We start by benchmarking the vertical electronic excitation
to the subtractive QM/QM scheme[26] we can include in an energies (EVA) and the lowest vertical ionization potential (VIP)
approximate way the contribution to the correlation energy of the DNA nucleoside deoxythymidine (dT). As explained else-
due to the nonstrong pairs if we write the total PT2 energy of where,[50,51] the electron is removed from the highest-
the system as occupied MO of the nucleobase, that is, the thymine moiety,
in the ionization process. The same holds for the excitation,
A1B A
ECASPT2  ECASPT2 1E2A1B 2E2A 5ECASPT2
A
1Eenv 5EFD2CASPT2 : (7) which is localized in the thymine chromophore. Therefore, this
system is expected to be a good candidate for analyzing the
The above expression defines the energy values computed accuracy of the FD approach versus the CASPT2 method.
with the present FD-CASPT2 scheme and used throughout this According to previous theoretical studies,[50,52] the sugar moi-
publication. As the active site contains partially occupied orbi- ety does not seem to change significantly the absorption ener-
tals, we must attach a meaning to the MP2-type energies of gies, but much larger changes can be expected for the VIP
eq. (7). Similarly to the strategy used in the FNO approxima- property. In particular, a change of 0.12 eV was reported by
tion,[19] we associate to W0 a single Slater determinant in Improta and Barone for the EVA related to the lowest-lying pp*
which all orbitals with negative energy are doubly occupied. excited state.[52] Conversely, Rubio et al.[50] predicted a
The active orbitals (fractionally occupied space) with nonnega- decrease of 0.62 eV in the VIP of the dT system with respect
tive energy are excluded from the list to avoid problems with to the thymine nucleobase.[50] Next, we analyze the perform-
the onset of orbitals near degeneracy (e.g., in the study of dis- ance of the FD-CASPT2 approach in the dT system as com-
sociation processes). All the remaining orbitals with nonnega- pared with the standard CASPT2 and taking into account the
tive energy define the secondary space. This hypothetical results related to the isolated nucleobase.
single-reference is used for computing E2A1B and E2A energy The ground-state equilibrium structure of dT has been opti-
contributions from solely its doubly excited determinants. mized at the MP2/cc-pVDZ level of theory in the anti confor-
The final consideration is about the computational scaling mation, that is, the optimized structure has the nucleobase
of the FD-CASPT2 method. Due to the use of canonical orbi- oriented away from the deoxyribose as in the B-DNA form. At
tals, although the latter are confined to one of the two regions the optimized geometry, the FD-CASPT2 method with different
all possible pairs within each region are treated equivalently. thresholds (s) has been used to compute the EVA energies of

1612 Journal of Computational Chemistry 2014, 35, 16091617 WWW.CHEMISTRYVIEWS.COM


WWW.C-CHEM.ORG REVIEW

Table 2. Number of occupied and virtual orbitals for various FD and full
CASPT2 calculations.

FD(s)

Full 0.0 0.01 0.1 0.20.4 0.5


O 50 59 25 24 11
V 241 81 81 81 32
(xfull=x)2 1 3 35 38 1171
The quantity in the last row represents a suitable measure of the reduc-
tion in floating point operations.

lying pp* excited states, which are in agreement with previous


CASPT2 results from Schreiber et al. (4.885.06, 5.886.15, and
6.106.52 eV).[54] These energies are close to the results
obtained for the dT system, thus indicating that the nature of
the electronic excitation is determined solely by the chromo-
phore itself. Conversely, the sugar moiety is found to be crucial
for an accurate determination of the VIP in the nucleoside. In
this case, while the VIP of thymine is computed at 9.11 eV at
Figure 1. Molecular model of the deoxythymidine. Atoms in orange color
the CASPT2(10in8)/ANO-S-VDZP//MP2/cc-pVDZ level of theory,
are the ones assigned to the active region. a) s 5 0.010.4 and b) s 5 0.5. in agreement with previous benchmark results,[55] a value of
8.65 is obtained for the nucleoside dT at the same level of
the three lowest-lying excited states of the molecule. The theory, in agreement with the experimental data (8.58 eV[56]
double-f ANO-S-VDZP basis set[53] has been used for these cal- and 8.80 6 0.17 eV[57]. To test the applicability of the FD
culations. The active space comprises 10 electrons in 8 active approach to determine the ionization energy of dT, we have
orbitals (10in8), including the valence p and p* orbitals of the computed the VIP energy with a threshold s 5 0.01, which
nucleobase moiety. Figure 1 depicts the selected active region corresponds to an appropriate selection of the active region
(region A) using different thresholds (s). Eight atoms are including the whole p system. The result for the VIP of dT is
assigned to the active region in the FD-CASPT2 computations 8.82 eV, which is a much better estimation as compared to the
with thresholds s ranging from 0.01 to 0.4 (orange), which cor- value obtained from the thymine chromophore alone. The
respond to the atoms belonging to the p conjugated system. deviation with respect to the VIP of dT computed with the
Taking into account that the natural orbitals of the CAS(10in8) standard CASPT2 method are 0.17 eV at the FD-CASPT2 level,
space are delocalized over these atoms, the selection can be but it is 0.46 eV when determined from full CASPT2 calcula-
considered as optimal. It is confirmed by the results compiled tions on thymine, respectively. Thus, a deviation nearly three
in Table 1. Hence, a relatively small energy difference, less than times larger, and above the expected error bar of the chosen
0.2 eV, is obtained between the FD(0.010.4)-CASPT2 and the quantum chemical method.
standard CASPT2 results. Conversely, larger thresholds, such as At the same time, the FD-CASPT2 calculation is much less
s 5 0.5, imply a dramatic decrease of the active atoms (see expensive than the full CASPT2 on the entire system, and not
Fig. 1), which leads to a significant deviation of the EVA ener- significantly more expensive than the corresponding full
gies from the reference values. In general, the energy differ- CASPT2 on the chromophore alone. In Table 2, we have
ence obtained with the FD(0.010.4)-CASPT2 results can be reported a measure of the reduction in floating point opera-
considered small enough for many applications of the CASPT2 tions for FD-CASPT2 calculations compared to full CASPT2.
method in photochemistry as the deviations are within the Such measure is based on the fact that the computationally
accepted error of the CASSCF/CASPT2 method (around 0.20.3 dominant steps in CASPT2 calculation (as for MP2) are quad-
eV). ratic in the quantity x 5 O 3 V, where O and V represent the
Regarding the isolated thymine molecule, the present CASP- number of occupied (inactive plus active) and virtual orbitals,
T2(10in8)/ANO-S-VDZP/MP2/cc-pVDZ computations give rise to respectively. The square of the ratio between xs computed
the EVA values 4.98, 6.20, and 6.36 eV, for the three lowest- for the full CASPT2 and for a given FD-CASPT2 measures the
reduction in floating point operations involved in the calcula-
tion, thus indicating the corresponding expected speedup in
Table 1. FD-CASPT2 vertical excitation energies (in eV) of deoxythymidine
using different thresholds s. CPU time. Clearly, FD-CASPT2 can possibly produce much
more significant speedupsup to two orders of magnitudes
s 0.0 0.01 0.1 0.20.4 0.5 for very large systemswhile retaining sufficient accuracy in
S1 4.90 5.07 5.10 5.11 5.98 its predictions. Table 3 compiles the calculations timing in sec-
S2 6.24 6.37 6.39 6.40 7.13 onds, which confirms the significant time reduction. Hence,
S3 6.27 6.42 6.47 6.47 7.51
whereas 1 h and 49 min are needed for the conventional

Journal of Computational Chemistry 2014, 35, 16091617 1613


REVIEW WWW.C-CHEM.ORG

Table 3. Timings (in sec) of deoxythymidine using different thresholds s.

s Timing (s)
0.0 6540
0.01 965
0.1 539
0.20.5 450

CASPT2 computations of EVA, the FD(s 5 0.20.4)-CASPT2 cal-


culations only spent 7 min and 30 sec.

Acrolein in a box of water

Solvatochromic effects on electronic excitation energies are


also a potential target of the FD-CASPT2 method, especially in
those cases where an explicit treatment of the solvent mole-
cules is desirable. The electronic excitation is usually localized
in the solute, and therefore, the surrounding solvent molecules
shall play a secondary role in the accounting for correlation
effects. The present work focuses on the analysis of the per-
formance of the FD-CASPT2 approach and, therefore, no Figure 2. Molecular model of the acrolein in a water box. Atoms in orange
color are the ones assigned to the active region. a) s 5 0.10.5, b) s 5
attempt will be made to study the details of solvent effects in 0.01, c) s 5 0.001 blue.
the excited states of the chosen molecule. One molecule of
acrolein surrounded by 21 water molecules has been selected
for the computation of the vertical excitation energy (EVA) on the chosen threshold. In particular, it corresponds to 16
related to the lowest-lying 1(n ! p*) excited state of s-trans atoms for s 5 0.001, six atoms for s 5 0.01, and four atoms
acrolein. Clearly, an appropriate treatment of the solvent for s 5 0.1 (see Figure 2). The nonhydrogen atoms of acrolein,
would require a larger number of water molecules and a sta- which hold the excitations, plus some of the hydrogen atoms
tistical description of the solvation, but for the purpose of test- of the solute are the only ones in the active region with the
ing the reliability of FD-CASPT2 the chosen example is well- thresholds ranging from 0.01 to 0.5, whereas some water mol-
fitting. The structure of acrolein surrounded by water mole- ecules in the nearest surrounding of the solute molecule are
cules has been optimized at the B3LYP/6-31G* level (a box of also considered as active for s 5 0.001.
water molecules, obtained previously in a classical dynamics Table 4 reports the CASPT2 energies computed both in the
simulation with the TINKER program,[58] was used as starting gas-phase and in solution. Experimental data and previous the-
structure). Full CASPT2 calculation on the entire system would oretical values obtained with different ab initio methodologies
be extremely demanding, requiring more than 22 GB of mem- are also compiled for the sake of comparison. As can be seen,
ory. Nonetheless, we can compare FD-CASPT2 results for EVA deviations among the FD-CASPT2 energies obtained for the
with the corresponding energies from gas-phase calculations different thresholds are very small. Such agreement is a conse-
and with previous experimental and theoretical values quence of the fact that in all cases the relevant atoms are
obtained for acrolein in liquid water. Five different thresholds included as active (it did not happen for dT, as shown in
have been considered for the FD scheme in the range Figure 1). Similar values can, thus, be expected for FD-CASPT2
between 0 and 1. The active space comprises six electrons dis- computations using smaller thresholds. A good agreement is
tributed in five active orbitals. It corresponds to the valence p found with the experimental data and with previous theoreti-
and p* orbitals and the oxygen lone-pair orbital. The double-f cal results in which the solvent effect was estimated by means
ANO-S-VDZP basis set has been used throughout.[53] The num- of a more accurate treatment using the averaged solvent elec-
ber of selected atoms in the active region varies depending trostatic potential from molecular dynamics[65] method or also

Table 4. FD(s)-CASPT2 lowest vertical excitation energy (in eV) for acrolein in vacuo and in water solution.

Gas-phase In water

FD-CASPT2 (s)

CASPT2 Expt. Prev. Theor. 0.001 0.01 0.10.5 Expt. Prev. theor.
[a] [b] [c] [d] [e] [f ] [g]
S1(n !p*) 3.82 3.71 , 3.75 3.69 , 3.63 , 3.74 , 3.85 3.92 3.96 3.95 3.94 3.90[h], 3.964.06[i]
[a] Ref. [59], [b] Ref. [60], [c] CASPT2//CASSCF, [61], [d] CASPT2, [62], [e] MR-CISD, [63]
, [f ] MR-CISD-Q, [63]
, [g] Ref. [64], [h] ASEP/MD-CASPT2,[60], [i] PCM-
CASPT2, [62].
Experimental and theoretical data from the literature are also shown.

1614 Journal of Computational Chemistry 2014, 35, 16091617 WWW.CHEMISTRYVIEWS.COM


WWW.C-CHEM.ORG REVIEW

Table 5. Timings (in sec) of acrolein in a box of water using different Table 6. FD(s)-CASPT2 lowest vertical excitation energy (in eV) for
thresholds s. Coelentramide.

s Timing (s) CASPT2 FD-CASPT2 (s) Expt. Prev. theor.


0.0 0.001 0.01 0.10.5
0.001 5160
S1(p !p*) 4.14 4.15 4.20 4.21 3.63.7[a] 4.27[b]
0.01 3900
0.10.5 3800 [a] Ref. [69], [b] Ref. [68].
Experimental and theoretical data from the literature are also shown.

with the approximate polarizable continuum model (PCM)[66]


ble for the light emission in the bioluminescent phenomenon
method. To compare results in gas-phase and solvent condi-
of a number of marine organisms.[67,68] As shown in the previ-
tions, the EVA of an isolated acrolein molecule was correspond-
ous study by Chen et al.[68] the lowest-lying electronic transi-
ingly computed, using CASPT2/ANO-S-VDZP energies on top
tion in coelenteramide implies a redistribution of the electron
of B3LYP/6-31G* geometries. The blue-shift for the EVA of the
1 density basically among the conjugated phenol and pyrazine
(n ! p*) state of acrolein in water, established previously
groups. Hence, we can benchmark the FD approach of the
both experimentally[59,60,64] and theoretically,[6163] is repro-
CASPT2 method focusing in the EVA computations related to
duced in the present work semiquantitatively (see Table 4).
the S0 to S1 transition.
Hence, we are able to predict sizable shifts in vertical excita-
Among the different conformational isomers plausible in
tion energies by performing relatively inexpensive CASPT2 cal-
coelenteramide, we use in the present work the conformation
culations on the active region only. Noticeably, variations with
AI of Chen et al.[68] which was optimized at the CAM-B3LYP/6-
respect to the definition of the active region (threshold s) are
311G(d,p) level starting from the X-ray crystallographic struc-
much smaller than the physical shift sought for. With this
ture of coelenteramide in Ca21-discharged obelin (PDB code:
observation, we argue that FD-CASPT2 is indeed very robust
2F8P.pdb). At this geometry, the EVA energies of the lowest-
in accounting for the correlation effects determining such
lying pp* transition are computed at the FD-CASPT2 level with
types of physicochemical observables. Table 5 compiles the
different thresholds (s) and using the ANO-S-VDZP basis set.[53]
calculations timing in second.
The active space comprises 12 electrons in 12 active orbitals
(12in12), including the valence p and p* orbitals of the conju-
Coelenteramide gated phenol and pyrazine moieties. The selected active
region (region A) using different thresholds (s) in the range
Another example of molecule whose quantum-chemical spec-
0.0010.4 are displayed in Figure 3. In the example of coelen-
troscopy calculations could benefit from the FD-CASPT2
teramide, the s values in the FD-CASPT2 computations that
method is coelenteramide. This molecular system is responsi-
share the same region as the natural orbitals of the
CAS(12in12) space are 0.10.4. Lower s values imply the inclu-
sion as active of other atoms from the toluene and p-
hydroxyphenylacetamide groups. Computations with thresh-
olds (s) of 0.5 and larger (not shown here) has a too reduced
number of active atoms (region A) and give rise to unphysical
EVA results.
Table 6 compiles the present computed EVA values together
with previous TD-DFT results using the CAM-B3LYP functional
and the 6-311G(d,p) basis set[68] and experimental results in
DMSO and benzene solvents.[69] The FD(0.0010.4)-CASPT2
results have deviations no larger than 0.14 eV with respect to
the value obtained with the full CASPT2 method, which points
to a good performance of the FD approach as in the deoxythi-
midine and solvated acrolein systems. As can be seen in Table 6,
the CASPT2 results show a relatively good agreement with the
TD-DFT and experimental data, appearing at lower energies

Table 7. Timings (in sec) of coelenteramide using different thresholds s.

s Timing (s)
0.0 29,100
0.001 12,240
Figure 3. Molecular model of the coelenteramide molecule. Atoms in
0.01 6,540
orange color are the ones assigned to the active region. a) s 5 0.001, b) s
0.1 5,280
5 0.010.4. [Color figure can be viewed in the online issue, which is avail-
0.20.4 4,860
able at wileyonlinelibrary.com.]

Journal of Computational Chemistry 2014, 35, 16091617 1615


REVIEW WWW.C-CHEM.ORG

than the TD-DFT EVA and closer to the experimental reported into account for the whole system at the CASSCF level of
value. Regarding the timings (see Table 7), it is worth noting the theory, this methodology can be considered as a two-scale
large difference of almost 7 h between the computation times hierarchical approach to the electron correlation problem.
of the conventional CASPT2 and the FD(s 5 0.20.4)-CASPT2,
while the discrepancy in the EVA is only around 0.1 eV. Acknowledgment
rn Roos, who unfortu-
This work was initiated by late Prof. Bjo
Conclusions nately could not see the final outcome of his idea.
The applicability of CD-based CASSCF/CASPT2 to the study of
Keywords: multiconfigurational perturbation theory  correlation
large molecules is hampered by two major computational bot-
energy  excited states  deoxythymidine  acrolein  coelenteramide
tlenecks: (a) the factorial scaling of CASSCF with respect to the
size of the active space, and (b) the quartic scaling of the
number of amplitudes in the CASPT2 step. Active spaces larger
How to cite this article: P. Farahani, D. Roca-Sanjuan, Aquilante,
than 18 electrons in 18 orbitals (18in18) are at present beyond
F. J. Comput. Chem. 2014, 35, 16091617. DOI: 10.1002/
reach. To overcome this limit, the so-called RASSCF/RASPT2
jcc.23666
method,[5,7072] or the recently proposed SplitCAS/SplitGAS
approach[73,74] can be used to map the CAS eigenvalue prob-
lem to a much smaller dimensionality. In all situations where
the active space bottleneck is avoided, it remains to cope with [1] D. Roca-Sanjuan, F. Aquilante, R. Lindh, WIREs Comput. Mol. Sci. 2012,
the cost of determining the amplitudes for the CASPT2 correc- 2, 585.
[2] K. Andersson, B. O. Roos, In Modern Electron Structure Theory, Vol. 2,
tion. This is particularly problematic for large molecules and
Part I; R. Yarkony, Ed.; World Scientific Publishing: Singapore, 1995; pp.
accurate AO basis sets. 55109.
In this work, we have demonstrated the benefits arising from [3] F. Aquilante, P.-A. Malmqvist, T. B. Pedersen, A. Ghosh, B. O. Roos, J.
the use of a localized orbital picture in CD-based CASPT2 calcu- Chem. Theory Comput. 2008, 4, 694.
[4] F. Aquilante, T. B. Pedersen, B. O. Roos, A. Sanchez de Meras, H. Koch,
lations on large molecules. For situations where the active orbi- J. Chem. Phys. 2008, 129, 24113.
tals are localized within a small region of the molecule, a [5] S. M. Huber, M. Z. Ertem, F. Aquilante, L. Gagliardi, W. B. Tolman, C. J.
suitable active site can be identified as the collection of atoms Cramer, Chem. Eur. J. 2009, 15, 4886.
where the active orbitals effectively extend. Accordingly, the [6] K. Pierloot, S. Vancoillie, J. Chem. Phys. 2008, 128, 34104.
[7] G. L. Macchia, F. Aquilante, V. Veryazov, B. O. Roos, L. Gagliardi, Inorg.
inactive and secondary orbitals can be separately localized and Chem. 2008, 47, 11455.
partitioned between this active site and the remaining atoms [8] N. H. F. Beebe, J. Linderberg, Int. J. Quantum Chem. 1977, 7, 683.
(environment). The two regions are assumed to be uncoupled, [9] H. Koch, A. Sanchez de Meras, T. B. Pedersen, J. Chem. Phys. 2003, 118,
9481.
and therefore, two separate sets of canonical orbitals can be
[10] F. Aquilante, R. Lindh, T. B. Pedersen, J. Chem. Phys. 2007, 127, 114107.
deduced for the active site and the environment. In this way, [11] F. Aquilante, R. Lindh, T. B. Pedersen, J. Chem. Phys. 2008, 129, 34106.
the standard algorithm for CASPT2 can be used without the [12] F. Aquilante, T. B. Pedersen, L. Gagliardi, R. Lindh, J. Chem. Phys. 2009,
need to reformulate the method in noncanonical orbital basis. 130, 154107.
[13] J. Bostrom, F. Aquilante, T. B. Pedersen, R. Lindh, J. Chem. Theory Com-
In terms of its implementation, the method is equivalent to an
put. 2009, 5, 1545.
appropriate removal of some orbitals from the inactive and sec- [14] J. Bostrom, M. Delcey, F. Aquilante, T. B. Pedersen, L. Serrano-Andr es,
ondary space, hence the name freeze-and-delete (FD) CASPT2. R. Lindh, J. Chem. Theory Comput. 2010, 6, 747.
The performance of the FD approach has been analyzed [15] T. B. Pedersen, F. Aquilante, R. Lindh, Theor. Chem. Acc. 2009, 124, 1.
[16] F. Aquilante, L. Boman, J. Bostr om, H. Koch, T. B. Pedersen, A. S. de
through three illustrative examples of CASPT2 energy compu- Meras, R. Lindh, In Challenges and Advances in Computational Chem-
tations: the three lowest-energy vertical electronic transitions istry and Physics, Vol. 13; M. G. Papadopoulos, R. Zalesny, P. G. Mezey,
and the lowest VIP of the deoxythymidine nucleoside, the ver- J. Leszczynski, Eds.; Springer, London, 2011; pp. 301344.
tical 1(n ! p*) transition of acrolein surrounded by water mol- [17] F. Aquilante, T. B. Pedersen, R. Lindh, J. Chem. Phys. 2007, 126, 194106.
[18] F. Aquilante, T. B. Pedersen, A. Sanchez de Meras, H. Koch, J. Chem.
ecules, and the lowest-energy vertical 1(p ! p*) transition of Phys. 2006, 125, 174101.
coelenteramide. It is shown that accurate relative energies [19] F. Aquilante, T. K. Todorova, T. B. Pedersen, L. Gagliardi, B. O. Roos, J.
(within 0.2 eV) can be computed by performing FD-CASPT2 Chem. Phys. 2009, 131, 34113.
calculations in which the correlating orbitals are restricted to [20] G. La Macchia, G. Li Manni, T. K. Todorova, M. Brynda, F. Aquilante, B.
O. Roos, L. Gagliardi, Inorg. Chem. 2010, 49, 5216.
only those assigned to the active site. This allows for substan- [21] P. Surjan, D. Kohalmi, Z. Rolik, A. Szabados, Chem. Phys. Lett. 2008,
tial savings in storage demand and computational costsnot 450, 400.
uncommonly, of 12 orders of magnitude. However, it has to [22] N. Govind, Y. A. Wang, A. J. R. da Silva, E. A. Carter, Chem. Phys. Lett.
1998, 295, 129.
be pointed out that the accuracy of the FD approach is 510
[23] T. A. Wesolowski, Computational Chemistry: Reviews of Current
times lower than those observed for local time dependent Trends, Vol. 10; World Scientific, Singapore, 2006, 10, 1.
second-order coupled cluster model (TD-CC2)[75] methods over [24] T. A. Wesolowski, A. Warshel, J. Phys. Chem. 1993, 97, 8050.
a large range of different molecules and excitations. Moreover, [25] C. R. Jacob, L. Visscher, In Recent Advances in Orbital-Free Density
Functional Theory; T. A. Wesolowski, Y. A. Wang, Eds.; World Scientific,
the portion of correlation energy associated with the environ-
Singapore, 2013, 6, 297.
ment is estimated at the cost of a canonical MP2-type energy [26] S. Dapprich, I. Komaromi, K. S. Byun, K. Morokuma, M. J. Frisch, J. Mol.
calculation. Given that the static correlation effects are taken Struct. (Theochem) 1999, 461, 1.

1616 Journal of Computational Chemistry 2014, 35, 16091617 WWW.CHEMISTRYVIEWS.COM


WWW.C-CHEM.ORG REVIEW

[27] T. Korona, D. Kats, M. Sch utz, T. B. Adler, Y. Liu, H. J. Werner, Chal- [51] M. Rubio, D. Roca-Sanjuan, L. Serrano-Andr es, M. Merchan, J. Phys.
lenges and Advances in Computational Chemistry and Physics, Vol. 13; Chem. B 2009, 113, 2451.
Springer, London, 2011; pp. 345407. [52] R. Improta, V. Barone, Theor. Chim. Acta 2008, 120, 491.
[28] T. D. Crawford, R. A. King, Chem. Phys. Lett. 2002, 366, 611. [53] K. Pierloot, B. Dumez, P. O. Widmark, B. O. Roos, Theor. Chim. Acta
[29] R. A. Mata, H. J. Werner, M. Sch utz, J. Chem. Phys. 2008, 128, 144106. 1995, 90, 87.
[30] D. Kats, M. Sch utz, J. Chem. Phys. 2009, 131, 124117. [54] M. Schreiber, M. R. Silva-Junior, S. P. Sauer, W. Thiel, J. Chem. Phys.
[31] R. A. Mata, H. Stoll, J. Chem. Phys. 2011, 134, 034122. 2008, 128, 134110.
[32] T. S. Chwee, E. A. Carter, J. Chem. Theory Comput. 2011, 103, 7. [55] D. Roca-Sanjuan, M. Rubio, M. Merchan, L. Serrano-Andr es, J. Chem.
[33] G. Walz, D. Kats, D. Usvyat, T. Korona, M. Sch utz, Phys. Rev. A 2012, 86, Phys. 2006, 125, 084302.
052519. [56] C. Yu, T. J. ODonnell, P. R. LeBreton, J. Phys. Chem. 1981, 85, 3851.
[34] P. Pulay, Chem. Phys. Lett. 1983, 100, 151. [57] S. Ptasinska, P. Candori, S. Denifl, S. Yoon, V. Grill, P. Scheier, T. D. Mark,
[35] M. Sch utz, G. Hetzer, H. J. Werner, J. Chem. Phys. 1999, 111, 5691. Chem. Phys. Lett. 2005, 409, 270.
[36] J. M. Foster, S. F. Boys, Rev. Mod. Phys. 1960, 32, 300. [58] J. Ponder, F. M. Richards, J. Comput. Chem. 1987, 8, 1016.
[37] J. Pipek, P. G. Mezey, J. Chem. Phys. 1989, 90, 4916. [59] R. S. Becker, K. Inuzuka, J. King, J. Chem. Phys. 1970, 52, 5164.
[38] C. Hampel, H. J. Werner, J. Chem. Phys. 1996, 104, 6286. [60] A. D. Walsh, Trans. Faraday Soc. 1945, 41, 498505.
[39] S. Saebo, P. Pulay, Annu. Rev. Phys. Chem. 1993, 44, 213. [61] M. E. Martin, A. M. Losa, I. Fernandez-Galvan, M. A. Aguilar, J. Chem.
[40] J. W. Boughton, P. Pulay, J. Comput. Chem. 1993, 14, 736. Phys. 2004, 121, 3710.
[41] M. Sch utz, H. J. Werner, J. Chem. Phys. 2001, 114, 661. [62] F. Aquilante, V. Barone, B. O. Roos, J. Chem. Phys. 2003, 119, 12323.
[42] H. J. Werner, F. R. Manby, P. J. Knowles, J. Chem. Phys. 2003, 118, 8149. [63] S. A. do Monte, T. M uller, M. Dallos, H. Lischka, M. Diedenhofen, A.
[43] P.-A. Malmqvist, K. Pierloot, A. R. M. Shahi, C. J. Cramer, L. Gagliardi, J. Klamt, Theor. Chem. Acc. 2004, 111, 78.
Chem. Phys. 2008, 128, 204109. [64] A. E. Moskvin, O. P. Yablonskii, L. F. Bondar, Theor. Exp. Chem. 1966, 2, 469.
[44] F. Aquilante, T. B. Pedersen, Chem. Phys. Lett. 2007, 449, 354. [65] I. F. Galvan, M. L. Sanchez, M. E. Martn, F. J. O. del Valle, M. A. Aguilar,
[45] O. Masur, D. Usvyat, M. Sch utz, J. Chem. Phys. 2013, 139, 164116. Comput. Phys. Commun. 2003, 155, 244.
[46] F. Aquilante, L. D. Vico, N. Ferr e, G. Ghigo, P.-A. Malmqvist, P. Neogrady, [66] J. Tomasi, B. Mennucci, R. Cammi, Chem. Rev. 2005, 105, 2999.
T. B. Pedersen, M. Pito nak, M. Reiher, B. O. Roos, L. Serrano-Andr es, M. [67] D. Roca-Sanjuan, M. G. Delcey, I. Navizet, N. Ferr e, Y.-J. Liu, R. Lindh, J.
Urban, V. Veryazov, R. Lindh, J. Comput. Chem. 2010, 31, 224. Chem. Theory Comput. 2011, 7, 4060.
[47] F. Ruiperez, F. Aquilante, J. M. Ugalde, I. Infante, J. Chem. Theory Com- [68] S. F. Chen, I. Navizet, D. Roca-Sanjuan, R. Lindh, Y.-J. Liu, N. Ferr e, J.
put. 2011, 7, 1640. Chem. Theory Comput. 2012, 8, 2796.
[48] G. Ghigo, B. O. Roos, P.-A Malmqvist, Chem. Phys. Lett. 2004, 396, 142. [69] O. Shimomora, K. Teranishi, Luminescence 2000, 15, 51.
[49] M. J. Frisch, G. W. Trucks, H. B. Schlegel, G. E. Scuseria, M. A. Robb, J. R. [70] S. M. Huber, A. R. M. Shahi, F. Aquilante, C. J. Cramer, L. Gagliardi, J.
Cheeseman, G. Scalmani, V. Barone, B. Mennucci, G. A. Petersson, H. Chem. Theory Comput. 2009, 5, 2967.
Nakatsuji, M. Caricato, X. Li, H. P. Hratchian, A. F. Izmaylov, J. Bloino, G. [71] P.-A. Malmqvist, K. Pierloot, A. R. M. Shahi, C. J. Cramer, L. Gagliardi, J.
Zheng, J. L. Sonnenberg, M. Hada, M. Ehara, K. Toyota, R. Fukuda, J. Chem. Phys. 2008, 128, 2041092.
Hasegawa, M. Ishida, T. Nakajima, Y. Honda, O. Kitao, H. Nakai, T. Vreven, [72] A. R. M. Shahi, C. J. Cramer, L. Gagliardi, Phys. Chem. Chem. Phys.
J. A. Montgomery, J. E. Peralta, F. Ogliaro, M. Bearpark, J. J. Heyd, E. 2009, 11, 10964.
Brothers, K. N. Kudin, V. N. Staroverov, R. Kobayashi, J. Normand, K. [73] G. L. Manni, F. Aquilante, L. Gagliardi, J. Chem. Phys. 2011, 134, 34114.
Raghavachari, A. Rendell, J. C. Burant, S. S. Iyengar, J. Tomasi, M. Cossi, N. [74] G. L. Manni, D. Ma, F. Aquilante, J. Olsen, L. Gagliardi, J. Chem. Theory
Rega, J. M. Millam, M. Klene, J. E. Knox, J. B. Cross, V. Bakken, C. Adamo, Comput. 2013, 9, 3375.
J. Jaramillo, R. Gomperts, R. E. Stratmann, O. Yazyev, A. J. Austin, R. [75] K. Lederm uller, M. Schutz, J. Chem. Phys. 2014, 140, 164113.
Cammi, C. Pomelli, J. W. Ochterski, R. L. Martin, K. Morokuma, V. G.
Zakrzewski, G. A. Voth, P. Salvador, J. J. Dannenberg, S. Dapprich, A. D.
Farkas, J. B. Foresman, J. V. Ortiz, J. Cioslowski, D. J. Fox, Gaus-
Daniels, O. Received: 14 January 2014
sian, Inc., Wallingford CT, 2009. Revised: 3 June 2014
[50] M. Rubio, D. Roca-Sanjuan, M. Merchan, L. Serrano-Andr es, J. Phys. Accepted: 6 June 2014
Chem. B 2006, 110, 10234. Published online on 7 July 2014

Journal of Computational Chemistry 2014, 35, 16091617 1617

Vous aimerez peut-être aussi