self-interaction error in density functional theory: a mean-field correction for molecules and large...
TRANSCRIPT
Chemical Physics 309 (2005) 67–76
www.elsevier.com/locate/chemphys
Self-interaction error in density functional theory: amean-field correction for molecules and large systems
Ilaria Ciofini a, Carlo Adamo a, Henry Chermette b,*
a Ecole Nationale Sup�erieure de Chimie de Paris, Laboratoire d’Electrochimie et Chimie Analytique, UMR CNRS-ENSCP no. 7575,
11 rue P. et M. Curie, F-75231 Paris Cedex 05, Franceb Laboratoire de Chimie Physique Th�eorique, Universit�e Claude Bernard, Bat. 210, Lyon I and CNRS UMR 5182, 43 Boulevard du 11 Novembre 1918,
F-69622 Villeurbanne Cedex, France
Received 26 April 2004; accepted 5 May 2004
Available online 28 July 2004
Abstract
Corrections to the self-interaction error which is rooted in all standard exchange-correlation functionals in the density functional
theory (DFT) have become the object of an increasing interest. After an introduction reminding the origin of the self-interaction
error in the DFT formalism, and a brief review of the self-interaction free approximations, we present a simple, yet effective, self-
consistent method to correct this error. The model is based on an average density self-interaction correction (ADSIC), where both
exchange and Coulomb contributions are screened by a fraction of the electron density. The ansatz on which the method is built
makes it particularly appealing, due to its simplicity and its favorable scaling with the size of the system. We have tested the ADSIC
approach on one of the classical pathological problem for density functional theory: the direct estimation of the ionization potential
from orbital eigenvalues. A large set of different chemical systems, ranging from simple atoms to large fullerenes, has been con-
sidered as test cases. Our results show that the ADSIC approach provides good numerical values for all the molecular systems, the
agreement with the experimental values increasing, due to its average ansatz, with the size (conjugation) of the systems.
� 2004 Elsevier B.V. All rights reserved.
1. Introduction
Density functional theory (DFT), combining good
performances and low computational costs, has become
an invaluable tool for chemists and physicists in un-
derstanding the electronic structure of atoms, molecules
or solids and related properties [1]. In the framework ofthe Kohn–Sham (KS) approach to DFT, the quality of
the results is strictly related to the functional used to
evaluate the exchange and correlation energy, the only
contribution that needs to be approximated in the ex-
pression of the total KS energy [2]. The research for
improved approximations to this contribution has
therefore become one of the main streams in theoretical
DFT development (see for instance [3] and [4]). In such aquest for higher accuracy, some failures of the different
* Corresponding author. +33-4-7244-8427; fax: +33-4-7244-8004.
E-mail address: [email protected] (H. Chermette).
0301-0104/$ - see front matter � 2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.chemphys.2004.05.034
models have been considered as ‘‘pathological’’, that is
intrinsic to the DFT approach itself. Among others,
activation energies for SN2 and proton transfer reac-
tions, dissociation energies of two center-three electron
systems, ionization potentials and charge transfer sys-
tems, can be considered as representative examples [5–
8]. In many cases, these faults only depend on theapproximate nature of the used functionals, which leads
to the so-called self interaction (SI) error. This spurious
effect arises from the interaction of an electron with it-
self, and it is related to Coulomb energy of the Kohn–
Sham (KS) Hamiltonian which is not, in contrast to the
Hartree–Fock approach, totally cancelled by the ex-
change contribution [2,9]. One of the consequences is
that the eigenvalue of the highest occupied KS orbitaldoes not correspond to the ionization potential (IP), as
it should. The SI problem has been clearly identified in
the early DFT approaches and some solutions were
proposed by Fermi and Amaldi [10] or by Slater [11].
68 I. Ciofini et al. / Chemical Physics 309 (2005) 67–76
More complex is, instead, the handling of SI in the
framework of the KS approach. Some years ago Perdew
and Zunger [12] have introduced a self-interaction cor-
rection (SIC), which is yet the only approach giving a
correction for both exchange and correlation contribu-tions. In fact, while some of the most recent correlation
functionals are SI-free by construction, the handling of
the exchange part is still troublesome [13–15].
The more effective recipe for the elimination of the SI
error is the evaluation of an orbital dependent exchange-
correlation potential, which is, in practical implemen-
tations, computationally excessively demanding, so that
its use is so far restricted to small model systems orsimplified Hamiltonians [16,17]. For this reason, various
attempts have been made to develop simplified SIC
scheme, based mainly on mean field approximations,
since they couple a reduced computational effort with an
almost complete removing of the SI error. When they do
not involve SIC pseudopotentials [18–22], the most
successful approaches are those based on the optimized
effective potential (OEP) [23,24] and in particular thatdeveloped by Krieger, Li and Iafrate (KLI), which in-
volves an integral equation for the averaged SIC field
[25]. Even if such approach has been successfully applied
in a number of different cases, including molecular or
solid state problems [26–28], still it requires a significant
amount of computer resources in order to evaluate the
orbital-dependent Coulomb part [29].
We have recently implemented in a molecular code asimplified, yet effective, approach based on an average-
density SIC (ADSIC) approximation, which takes into
account the screening of both exchange-correlation and
Coulomb contribution through a subtraction from the
total density of a fraction proportional to 1=N , N being
the total number of electron [30]. This simple, self con-
sistent approach, first proposed by Legrand et al. [31]
and applied to a jellium model is technically not ex-pensive and, at the same time, still retains a number of
theoretical features, like the correct behavior for the
asymptotic potential and a variational formulation [31].
The first applications to molecular systems are quite
promising, but some potentialities of ADSIC are still
unexplored [30].
The aimof this paper is to investigate some of the limits
of such an approach in the field ofmolecular applications.To this end, the evaluation of the vertical ionization po-
tentials (IPs) for selected test cases, ranging from atoms to
large conjugate systems, has been chosen as a difficult
playground. Although the first IPs could be directly cal-
culated using the DFT extension of the Koopmans the-
orem [25,32], or more exactly of the Janak theorem [33],
the negative of the KS HOMO energy is too small with
respect to the experimental values [34], even if the shift,usually of several electronvolts, is rather constant [35–39].
It has been demonstrated that with the self interaction
correction included, the HOMO energies are much closer
to the first IPs, and that this effect is directly related to the
SI error [40]. In this respect, IPs belong to key properties
to validate any new SI corrections.
2. Theoretical background
In the Kohn–Sham (KS) approach to DFT the total
exact energy can be written as [2]:
EKS qa; qb� �
¼ Ts uri
� �� �þ J q½ � þ
ZqðrÞvðrÞdr
þ Exc qa; qb� �
; ð1Þ
where fuirgðr ¼ a or bÞ are the spin orbitals and qr are
the total density of spin r. The first term in Eq. (1) is the
kinetic energy of a system of non-interacting particle,
the second is the Coulomb interaction and the third isthe interaction energy between the electron density qðrÞand the external potential mðrÞ. These terms are all
known exactly, which is not the case for the last con-
tribution, the Exc term, containing all the remaining
contribution to both the kinetic energy and the electron-
electron interaction, and usually is approximated by
some functional form of the density, Eapproxxc . Even if the
choice of the functional form is completely arbitrary,some physical constraints which have to be respected by
the exact exchange-correlation functionals (potential)
are well known (see [41] for an interesting summary on
this point). In particular, the exact potential should
manifest a derivative discontinuity in the bulk region, as
a constant shift [42]:
vþxcðrÞ � v�xcðrÞ ¼oExc
oN
� �þ
v
�� oExc
oN
� ��
v
�
� eLUMOð � eHOMOÞ; ð2Þ
where vþxcðrÞ and v�xcðrÞ are the right and left derivatives
of the vxcðrÞ potential with respect to the number of
electrons, evaluated at N � d, d being a positive infini-
tesimal. At the same time vxcðrÞ should fall off as 1=r inthe asymptotic region [43]:
vxc ! � 1
r; r ! 1 ð3Þ
Because LDA and GGA XC functionals are not self-
interaction free, and they do not cancel the SI contained
in the Coulomb potential, they do not exhibit the correct
asymptotic decay [44]. Moreover, because the correla-
tion contribution of the XC energy and potential are
short ranged, describing only the so-called dynamic
correlation, only the exchange potential controls theasymptotic decay of the potential. Several attempts to
improve exchange functionals have borne on the one
hand on the exchange potential itself, in order to obtain
model potential with the correct asymptotic behavior,
[25,44–48], and on the other hand on self-interaction
corrections to the XC energy [12,16,17,31,45,49–55].
I. Ciofini et al. / Chemical Physics 309 (2005) 67–76 69
In a recent work, Della Sala and G€orling [56] have
demonstrated that in molecular systems, the asymptotic
region of the Kohn–Sham exchange potential is even
more complicated (than so far assumed), because the 1=rdecay has to be replaced by a C � 1=r, C being a con-stant, in directions belonging to nodal surfaces of the
HOMO of the molecule (this is also shown in [29]). The
C constant is directly related to the energy difference
between the HOMO and the highest occupied MO be-
low the HOMO possessing no nodal surface in the same
region. They have introduced a localized Hartree–Fock
method (LHF) allowing to approximately determine a
KS exchange potential by localizing the (non-local)Hartree–Fock potential [44]. Furthermore, they have
shown that the KLI procedure can be considered as an
approximation of the LHF method. These methods lead
to qualitatively correct SI free KS orbitals, whose ei-
genvalues can be compared to ionization energies or
electronic affinities, and be used for property calcula-
tions, e.g., TDDFT calculations of optical spectra.
The self interaction correction method of Perdew andZunger (PZ) effectively yields a discontinuity for the
highest occupied state when N is increased though in-
tegral occupancy [12]. This approach is de facto the
starting point for any SIC:
ESICxc qa; qb
� �¼ Eapprox
xc qa; qb� �
�Xr¼a;b
�XNr
i¼1
Eapproxxc qr
i ; 0� �
þ J qri
� �; ð4Þ
where the second sum runs on all the occupied spin
orbitals of density qri . The self-interaction error is thus
explicitly subtracted for each orbital from the standard
KS exchange-correlation energy. The correspondinglocal spin–orbital potential is then:
vSICxc;i qa; qb� �
¼ dEapproxxc
dqri
�Z
qri ðr0Þr� r0j j dr
0 � dExc½qri ; 0�
dqri
: ð5Þ
Since the exchange-correlation energy and potential are
dependent on each spin–orbital, this correction proce-
dure is computationally expensive. Moreover, SIC are
not invariant with respect to unitary transformations of
the orbitals. This leads to the necessity to employ orbital
localization techniques [27,57]. Other SIC techniques
may be found in [58–61].
A more convenient way is to average the resultingpotential over the different spin–orbitals in order to have
an orbital-independent potential. This is the philosophy
of the Krieger–Li–Iafrate approach [25]. Here the po-
tential, mKLIxc , is an explicit functional of the orbitals:
vKLIxc;i qa; qb
� �¼
Xr
Xi
qri
qrvSIC;rxc;i
��Z
qri vKLI;r
xc;i ðrÞ�
� vSIC;rxc;i ðrÞ�dr�; ð6Þ
where the vSIC;rxc;i is expressed by Eq. (5). The correction
thus obtained is, in practical applications, equivalent to
the PZ approximation, but has the advantage of coming
from a single local potential. In particular, atomic ex-
change correlation potentials, with the correct asymp-totic behavior have been obtained, giving very accurate
ionization potentials [62]. However, the SIC potential
needs to be determined through an iterative scheme, as it
appears in both sides of Eq. (6). The KLI scheme can be
further simplified by neglecting the orbital-dependent
part. The result is the so-called Slater potential [11,25]:
vSlaterxc qa; qb� �
¼Xr
Xi
qri
qrvSIC;rxc;i ðrÞ: ð7Þ
Of course, the Coulomb part, J ½qri �, still requires
some computational efforts.
Further simplification of this model is its adaptation
to the free electron gas [11], leading to the Dirac–Slater
exchange or the Xa exchange-correlation potentials [63],
which are not SI free, as it is known.An even more crude simplification of Eqs. (4) and (5)
can be obtained by subtracting a fraction 1=N from the
total density, so that qðN � 1Þ=N can be considered as
the density of the (N � 1) electrons seen by the spectator
electron. This approach rests on the original idea of
Fermi and Amaldi [10], but the inclusion of the ex-
change-correlation term makes it more suitable for
dealing with current approximate functionals. The en-ergy correction in such a case, takes a really simple form
[31]:
EADSICxc qa; qb
� �¼ Eapprox
xc qa; qb� �
� NJqN
h i
�Xr¼a;b
NrEapproxxc
qr
Nr; 0
� �; ð8Þ
whereas the potential is
vADSICxc qr½ � ¼ dEapprox
xc
dqr�Z
qrðr0Þ=Nr� r0j j dr0 � dExc½qr=N ; 0�
dqr:
ð9ÞThis simple model retains a number of properties,
and, in particular the potential (9) has the correct 1=rasymptotic behavior, can be variationally related to an
energy potential and, of course, give zero for one-elec-
tron potential [31]. In the same vein, Lundin and Eri-
ksson [58] corrected only the density of the localized
electrons of their system (f electrons).
3. Computational details
We have implemented the ADSIC approach in one of
the development versions of the Gaussian 03 code [64] for
the generalized gradient approximation (GGA) for the
exchange based on the Becke 88 functional [65]. Thiscontribution has been next coupled with the GGA
70 I. Ciofini et al. / Chemical Physics 309 (2005) 67–76
correlation of Lee, Yang and Parr [66], a functional by
construction SI-free, so to have the BLYP approach.
All the molecules have been optimized using the 6-
311G(d,p) basis set and energy evaluation have been done
with the same basis, eventually augmented with a diffusefunction on heavier atoms (6-311+G(d,p) basis) [67,68].
Some tests have been also carried out with the extended 6-
311++G(3df,3pd) basis set, while the smaller 6-31G basis
has been used for fullerenes. In this case the geometries of
the most stable isomers obtained with a 6-31G basis were
used. All fullerenes structures were optimized at the PM3
level. ForC78, all isomers respecting the isolated pentagon
rule have been computed to check the variation of theproperties with the isomer symmetry. The reliability of
such structures was verified by carrying out geometry
optimization on themost stable isomer at the BLYP level.
Since no significant difference was found, we expected no
difference in the DFT properties.
4. Results and discussion
As first test, we have computed the IPs, from the
highest orbital eigenvalue, for some atomic systems,
even if we do not expect that the ADSIC approach
performs particularly well on such systems, due to its
average hypothesis. For this reason, only 10 spherical
atoms, all belonging to the first and second group of the
periodic table (configuration [core] n1 or [core] n2) havebeen chosen. The results are reported in Fig. 1 as
function of the number of electrons, together with the
experimental values [69] and the corresponding uncor-
rected energies (labeled noSIC on the plot). The parallel
behavior of the computed values for the two sets, with
and without SIC, is evident from the plots of Fig. 1.
From one hand, the IPs calculated as the �eHOMO
without the SI correction exhibit a large error with re-
3 4 11 12 19 20 37 382
4
6
8
10
12 noSIC ADSIC exp
IP(e
V)
n. electrons
Fig. 1. Experimental (from [68]) and theoretical ionization potential
for some selected atoms, ranging between Li (no. electrons¼ 3) and Sr
(no. electrons¼ 38). The values have been computed as �eHOMO, using
the 6-311+G(d,p) basis set.
spect to the experimental data, ranging from )1.8 (Rb)
and )3.9 (Be) eV. On the other hand, the ADSIC cor-
rection shifts the orbital energies to lower values, so that
the corresponding IPs are overestimated, between +3.6
(Na) and +2.3 (Sr). Nevertheless, we have to notice thatwhile the eigenvalues for the highest occupied atomic
orbital are quite sensitive to the basis set for standard
functional, this is not the case for the ADSIC approach.
For instance, the differences found for the considered
atoms in going from the 6-311+G(d,p) to the large 6-
311+G(2df,2p) basis is always less than 0.0003 eV.
As next test, we have chosen an ensemble of five
molecules, containing 14 and 15 main group elements.These molecules have been chosen since they contain a
double bond, N@X (X¼ Si, C), bonded to hydrogen or
methyl groups, and so the electronic properties are ex-
pected changing with the X atoms or the substitutes in a
regular way. Furthermore, one of such molecules,
namely trimethylsilanimine, is characterized by two IPs
corresponding to the ejection of an electron from the
non-bonding orbital (nY ) or from the bonding pX¼Y
orbital. The two corresponding IPs have the peculiarity
of being close so that they are difficulty assigned without
theoretical help [39].
All the calculations have been carried out at the, re-
spectively, BLYP optimized geometries, obtained with-
out the ADSIC correction, and the results are reported
in Table 1. In order to have a better picture of the
performances of the ADSIC approach, the IPs havebeen calculated either using the orbital energies and the
so-called DSCF method, where the total energy differ-
ence between neutral and charged molecules.
The values obtained using the uncorrected �eHOMO
(and �eHOMO�1) energies (no SIC approach) have a
mean average error (MAE) of 3.6 eV. As already found
for the atoms, these IPs are significantly underestimated.
In contrast a much smaller error is found for the DSCFmethod, with the MAE around 0.7 eV, as expected on
the basis of previous calculations [39]. Close deviations
have been found for the ADSIC approach, the MAE
being 0.9 eV.
In one case (silaethene) we have checked the basis set
dependency of the ADSIC orbital energies by using the
large 6-311++G(3df,3pd) set, and we did not find any
significant difference (DeHOMO < 0:02 eV).More in general, it appears from our results that the
ADSIC approach provides, for such systems, IPs with
the same accuracy than the DSCF method: when the
latter fails, the ADSIC method seems to be neither
better, not worse than DSCF with respect to the exper-
iments. In the case of the two states of the trimethylsi-
lanimine (2A0 and 2A00), they are found significantly too
separated than experiment suggests with both methods.This is probably due to a multideterminant nature of (or
at least of one of) these states, which is poorly described
by standard GGAs and, consequently, by the ADSIC
Table 1
Ionization potential (eV) for some unsaturated molecules, computed either using the DSCF approach (first column) or the HOMO eigenvalues
(second and third columns)
Molecule DSCF �eHOMO Exp.a
NoSIC ADSIC
Ethylene 10.5 6.4 12.1 10.5
Isobutene 8.9 5.1 9.5 9.5
Silaethene 8.7 6.2 10.5 9.0
Dimethylsilaethene 7.8 4.7 8.6 8.3
Trimethylsilanimine (2A0) 6.9 4.2 7.0 7.9
Trimethylsilanimine (2A00) 9.7 7.0 10.0 8.3
All the calculations have been carried out with the 6-311G(d,p) basis set.aAll experimental values are from [32].
I. Ciofini et al. / Chemical Physics 309 (2005) 67–76 71
approach (for multideterminantal DFT approaches see
[70]).
The results obtained for the atoms and the chosen
molecular systems are consistent with the average ap-
proximations: more electrons are involved and more
delocalized the system is, more the results are accurate.
So adding methyl groups to unsaturated carbons or
changing these latter with heavier atoms (Si) alwaysleads to improved results. At the same time, the ADSIC
seems to provide correct eigenvalues also for orbitals
lower in energy than the HOMO.
These considerations induced us to investigate some
aromatic systems for which, at least in principle, the
ADSIC model should work at its best. To this end, we
have chosen two different sets. The first one is composed
by 7 aromatic systems, relatively large (5 and 6 memberrings) and containing not only carbon atoms (benzene
and fulvene), but also nitrogen (pyridine, pyrimidine
and pyrazine) and sulphur (furane and thiophene). For
these systems several IPs have experimentally observed
and theoretically computed, thus providing a further
verification for our approach [38,40]. The second set is
composed by three different fullerenes, namely C60, C70
and C78 (see Fig. 2), whose IPs have been recently ex-perimentally determined [71]. While the first two fulle-
renes are symmetric, the last one, C78, is not, so that a
different behavior of the ADSIC approach is expected.
The results obtained for the first set of aromatic
molecules are collected in Table 2. Without entering in a
tedious analysis of all the ionization potentials, we note
that the ADSIC approaches provides Mean Average
Errors (MAE) which are about one tenth of that ob-tained with the uncorrected BLYP functional (0.6 vs 4.0
eV, in average) and close to the results provides by the
DSCF approach (MAE¼ 0.5 eV, see [35]). Furthermore,
the data of Table 2 show that the ADSIC approach
treats on the same foot both r and p valence orbitals,
thus allowing accurate calculations also for high energy
transitions. It also interesting to compare our results for
pyridine, furane and thiophene, with those recently ob-tained by Chong and co-workers, using a statistical
average orbital potential (SAOP), starting from post-
HF wavefunctions [41]. For these three molecules the
SAOP deviation is 0.5 eV, which is also our results at the
BLYP/ADSIC level. The overall trends are summarized
in Fig. 3 where the BLYP/ADSIC results are reported
versus the experimental values for four molecules (py-
rimidine, furane, benzene and fulvene). From these
plots, a linear correlation between the two data sets is
quite apparent, the correlation factor (R) being alwaysgreater that 0.99.
Finally we did compute both the ionization potentials
and the electron affinities of three fullerenes in what we
suppose to be their most stable isomers. The computed
values are reported in Table 3 in comparison with the
available experimental data [62,72,73]. It is worthwhile
to note that, in some cases (e.g., C78), the experimental
values refer to a mixture of isomeric forms for a givenCn. Nevertheless the values for a given isomer are nor-
mally estimated within 0.1eV from the average mixture
value [74].
The computed SIC IP of C60 and C70 compare well to
the experimental ones (deviation of 0.5 eV), whereas the
less symmetric C78 deviates more strongly from the ex-
periment (0.7–1.0 eV). This cannot be related to the fact
that the experimental data refer to a Boltzman distri-bution of different isomers, since our calculations con-
sidering all stable isomers show a limited spread (Table
3). For what concerns the EA, the values computed
from the LUMO eigenvalue of the neutral systems
(�eLUMO in Table 3) strongly overestimate the EA both
when neglecting and including the SIC. Nevertheless
when electronic relaxation is allowed, that is when
computing the EAs as SOMO eigenvalue of the C�n
anionic systems (�eSOMO in Table 3), the ADSIC results
become in a 0.5 eV agreement with the experimental
data whereas the non-corrected values do strongly de-
viates (1.5–2.0 eV). Furthermore, our results are closed
to those obtained with parameterized semi-empirical
approaches [71] or with DFT/DSCF method [75]. In this
last case, the IP for the C60 computed by Martin and co-
workers using the B3LYP/6-31G(d) approach and theDSCF method is 7.14 eV, in excellent agreement with
our ADSIC value (7.15 eV) [75].
Fig. 2. Sketches of the considered molecules: (a) small organic compounds; (b) fullerenes: C60 (left), C70 (center), C78 (right).
72 I. Ciofini et al. / Chemical Physics 309 (2005) 67–76
In summary the ADSIC approach provides results as
accurate as the standard DSCF approach for large,
conjugate systems, using the corrected KS orbital ei-
genvalues. This result is even more interesting, since the
latter approach requires as many energy calculations as
cationic species are searched, besides the neutral ground
state calculation, whereas the ADSIC calculation pro-
vides the whole ionization energy spectrum within one-shot calculation. In this context, the ADSIC approach is
not only accurate but twice faster for single ionization
estimation. Furthermore, the accuracy of the ADSIC
model is close to that provided by more complex and
resource-demanding DFT methods [40].
SIC methods lead to orbital patterns where on the
one hand the occupied orbitals energies are shifted
downwards with respect to LDA or GGA orbitals, be-coming close to corresponding HF eigenvalues. On the
other hand the unoccupied orbitals energies are signifi-
cantly lower than corresponding HF eigenvalues, be-
cause the latter obey to a exchange potential which acts
differently on occupied and unoccupied orbitals, leading
to artificially overestimated HOMO–LUMO gaps.
Therefore, SIC methods deliver more realistic gaps, ly-
ing between (underestimated) LDA/GGA gaps and
(overestimated) HF ones. Consequently SIC methods
should lead to significantly better properties than those
calculated at the LDA, GGA, metaGGA KS level (on
the one hand) or HF level (on the other hand). This is
visible in properties like IPs, even with approximate SIC
models such as the ADSIC used in this work, and
should stem within TDDFT calculations (work in pro-
gress). At the KLI level, Patchkovskii et al. [27] haveproven that magnetic properties are significantly im-
proved.
Whereas SIC orbitals are significantly better than
standard KS orbitals, this is not necessarily the case of
the SIC energy. Indeed, the SIC energy functionals have
lost the non-dynamic correlation part which is present in
LDA/GGA exchange functionals. Therefore energetic
properties (reaction energies, barriers, etc.) may be(barriers) or may be not (reactions) improved, according
to the difference in electronic density delocalization be-
tween reactants, products and transition states: if this
difference is large, SIC are welcome and lead to im-
proved energy differences (see for instance Patchkovskii
and Ziegler in [76]). In most cases, SIC geometry opti-
mizations are rather poor, and deviate like, but even
more strongly than HF geometries [58]. For that reason,
Table 2
Computed and experimental vertical ionisation potentials (eV) for conjugated planar molecules
Molecule Elec. state NoSICa ADSIC expa
Benzene 1e1g 5.99 9.96 9.24
3e2g 8.10 11.48 11.49
1a2u 8.66 12.73 12.3
3e1u 10.03 13.52 13.8
1b1u 10.62 14.15 14.7
2b2u 11.03 14.39 15.4
3a1g 12.59 16.17 18.85
2e2g 14.53 18.02 19.2
2e1u 18.08 21.67 22.8
2a1g 20.77 24.59 25.9
Fulvene 1a2 4.65 8.91 8.55
2b1 6.14 10.05 9.54
7b2 8.46 11.55 12.1
11a1 8.78 12.10 12.8
10a1 8.86 12.78 13.6
6b2 9.05 12.85 14
Pyridine 1a2 5.66 9.39 9.6
11a1 6.37 10.46 9.75
2b1 6.95 11.06 10.81
7b2 8.68 12.14 12.61
1b1 9.52 13.21 13.1
10a1 9.64 13.72 13.8
6b2 10.31 13.89 14.5
Pyrimidine 7b2 5.70 9.52 9.73
2b1 6.95 10.69 10.5
11a1 7.07 11.29 11.2
1a2 7.75 12.01 11.5
1b1 9.87 13.50 13.9
10a1 10.12 13.84 14.5
6b2 10.32 14.67 14.5
Pyrazine 6ag 5.63 9.36 9.4
1b1g 6.80 10.92 10.2
5b1u 7.07 11.02 11.4
1b2g 8.05 12.31 11.7
3b3g 9.38 12.92 13.3
1b3u 10.31 14.31 14
4b2u 10.62 14.65 15
4b1u 11.89 15.31 16.2
Furane 1a2 5.36 9.85 8.89
2b1 6.52 11.05 10.25
9a1 8.66 12.91 13
8a1 9.44 13.53 13.8
6b2 9.81 13.71 14.4
5b2 10.50 14.53 15.25
1b1 10.67 15.20 15.6
Thiophene 1a2 5.56 9.75 8.87
3b1 5.96 10.25 9.52
11a1 8.21 12.33 12.1
2b1 8.95 13.10 12.7
7b2 9.27 13.23 13.3
6b2 9.42 13.32 14.3
MAEb 4.0 0.5
All the calculations have been carried out using the 6-311G(d,p) basis set.aAll the BLYP and experimental values are from [31].bMean average error on all the IPs.
I. Ciofini et al. / Chemical Physics 309 (2005) 67–76 73
most papers describing SIC properties use molecular
geometries optimized with standard GGA, metaGGA
or hybrid XC functionals. This has been the case in the
present work. The poorness of geometries obtained with
model exchange potentials has early been also under-
lined in many papers [77,78].
Table 3
Computed and experimental ionization potentials (IP, eV) and electron affinities (EA, eV) for the three considered fullerenes
IP
E (kcal/mol) �eHOMO Exp.a
NoSIC ADSIC
C60Ih 0 5.37 7.18 7.57
C70D5h 0 5.3 6.97 7.36
C78C2v 0 4.77 6.33 7.26
D3h 3.2 (3.8) 4.79 6.34
C2v 7.6 (5.6) 5.01 6.57
D3 9.7 (6.3) 4.74 6.30
D3h 26.8 (17.5) 5.08 6.63
EA
�eLUMO �eSOMO Exp.b
NoSIC ADSIC NoSIC ADSIC
C60Ih 3.48 5.28 0.29 2.09 2.65
C70D5h 3.46 5.11 0.45 2.10 2.73
C78C2v 3.89 5.43 0.93 2.49 3.1
D3h 4.05 5.61 1.23 2.66
C2v 3.79 5.34 0.84 2.39
D3 3.84 5.41 0.96 2.51
D3h 3.41 4.95 0.53 2.07
All the calculations have been carried out using the 6-31G basis set.aRef. [65].bRef. [66].
8 9 10 11 12 13 14 15 168
9
10
11
12
13
14
15
16
fulvene pyrimidine benzene furane
IP c
alc
(eV
)
IP exp (eV)
Fig. 3. Correlation between experimental and computed BLYP/AD-
SIC ionisation potentials for four different aromatic molecules.
74 I. Ciofini et al. / Chemical Physics 309 (2005) 67–76
The reason of that disappointing feature is that the
removal of the SI energy removes the non-dynamical (or
left-right, or quasi-degeneracy) correlation energy in-
cluded in the exchange energy functional, whereas nosignificant change is obtained for the (short ranged)
dynamical correlation energy, described by the correla-
tion energy functional. Whereas it is now widely ad-
mitted than the long ranged non-dynamical correlation
energy is described by the exchange part of the ex-
change-correlation energy functional, as a consequence
of the structure of the (Fermi) exchange hole [79–84], it
has been argued more recently that it is the SI error
which mimics long range (left-right) non-dynamic cor-
relation [85,86].
5. Conclusions
In this paper, we have presented the validation of a
simple approach to correct the self interaction error
present in the common approximate exchange-correla-
tion functionals used in density functional theory. This
model rests on an average density self-interaction cor-
rection (ADSIC), so that the main advantages of the
method with respect to other corrections are its sim-plicity and its favorable scaling with the size of the
system. At the same time, it retains a number of theo-
retical features, such as the correct asymptotic behavior.
We have tested the ADSIC approach on the direct es-
timation of the ionization potential from orbital eigen-
values. On this difficult playground, the ADSIC
approach provides good numerical values for all the
molecular systems, while large deviations, due to itsaverage ansatz, are obtained for atomic or strongly lo-
calized molecular systems.
6. Note added in proofs
Eq. (5) and (9) of this paper were wrongly reported in
Ref. [30] (one missing term), but correctly programmed.The correct form is given here.
I. Ciofini et al. / Chemical Physics 309 (2005) 67–76 75
Acknowledgements
Authors thank Thomas Heine (University of Dres-
den) help concerning fullerene structures. I.C. and C.A.
also thank CNRS for a financial support from the ACI‘‘Jeune Equipe 2002’’ project. This work has also been
carried out within the framework of the Cost Action
D26 ‘‘Integrative Computational Chemistry’’ (action n.
D26/0013/02) and a support from the CNRS GdR
‘‘DFT’’. The CINES is acknowledged for a grant of time
(project cpt2130).
References
[1] W. Koch, M.C. Holthausen, A Chemist’s Guide to Density
Functional Theory, Wiley-VCH, Weinheim, 2000.
[2] R.G. Parr, W. Yang, Density Functional Theory of Atoms and
Molecules, Oxford University Press, New York, 1989.
[3] C. Adamo, A. di Matteo, V. Barone, Adv. Quantum Chem. 36
(1999) 4.
[4] H. Chermette, A. Lembarki, F. Rogemond, H. Razafinjanahary,
Adv. Quant. Chem. 33 (1999) 105.
[5] M.N. Glukhovtsev, R.D. Bach, A. Pross, L. Radom, Chem. Phys.
Lett. 260 (1996) 558.
[6] C. Adamo, V. Barone, J. Chem. Phys. 108 (1998) 664.
[7] H. Chermette, I. Ciofini, F. Mariotti, C. Daul, J. Chem. Phys. 115
(2001) 11068.
[8] E. Ruiz, D.R. Salahub, A. Vela, J. Phys. Chem. 100 (1996) 12265.
[9] E.J. Baerends, O.V. Gritsenko, J. Phys. Chem. A 101 (1997)
5383.
[10] E. Fermi, E. Amaldi, Accad. Ital. Rome 6 (1934) 117.
[11] J.C. Slater, Phys. Rev. 81 (1951) 385.
[12] J.P. Perdew, A. Zunger, Phys. Rev. B 23 (1981) 5048.
[13] J.P. Perdew, S. Kurth, A. Zupan, P. Blaha, Phys. Rev. Lett. 82
(1999) 2544.
[14] E. Proynov, H. Chermette, D.R. Salahub, J. Chem. Phys. 113
(2000) 10013.
[15] J.B. Krieger, J. Chen, G.J. Iafrate, A. Savin, in: A. Gonis, N.
Kioussis (Eds.), Electron Correlations and Materials Properties,
Plenum, New York, 1999.
[16] B.G. Johnson, C.A. Gonzales, P.M.W. Gill, J.A. Pople, Chem.
Phys. Lett. 221 (1994) 100.
[17] A. Svane, Phys. Rev. Lett. 72 (1994) 1248.
[18] A. Svane, W.M. Temmerman, Z. Szotek, J. Laegsgaard, H.
Winter, Int. J. Quant. Chem. 77 (2000) 799.
[19] D. Vogel, P. Kr€uger, J. Pollmann, Phys. Rev. B 54 (1996) 5495.
[20] D. Vogel, P. Kr€uger, J. Pollmann, Phys. Rev. B 58 (1998) 3865.
[21] D. Vogel, P. Kr€uger, J. Pollmann, Phys. Rev. B 54 (1996) 5495;
C. Stampfl, C.G. Van de Walle, D. Vogel, P. Kr€uger, J. Pollmann,
Phys. Rev. B 61 (2001) R7846.
[22] A. Filippetti, N.A. Spaldin, Phys. Rev. B 67 (2003) 125109.
[23] J.D. Talman, W.F. Shadwick, Phys. Rev. A 14 (1976) 36.
[24] R.T. Sharp, G.K. Horton, Phys. Rev. 90 (1953) 317.
[25] J.B. Krieger, Y. Li, G.J. Iafrate, Phys. Rev. A 45 (1992) 101.
[26] J. Garza, J.A. Nichols, D.A. Dixon, J. Chem. Phys. 112 (2000)
7880.
[27] S. Patchkovskii, J. Autschbach, T. Ziegler, J. Chem. Phys. 115
(2001) 26.
[28] S. Patchkovskii, T. Ziegler, J. Phys. Chem. A 106 (2002) 1088.
[29] S. K€ummel, J.P. Perdew, Phys. Rev. B 68 (2003) 035103.
[30] I. Ciofini, H. Chermette, C. Adamo, Chem. Phys. Lett. 380 (2003)
12.
[31] C. Legrand, E. Suraud, P.G. Reinhard, J. Phys. B 35 (2002) 1115.
[32] J.P Perdew, R.G. Parr, M. Levy, J.L. Balduz, Phys. Rev. Lett. 49
(1982) 1691.
[33] J.F. Janak, Phys. Rev. B 18 (1978) 7165.
[34] J.P. Perdew, M. Levy, Phys. Rev. B 56 (1997) 16021.
[35] R. Stowasser, R. Hoffmann, J. Am. Chem. Soc. 121 (1999) 3414.
[36] C. Zhan, J.A. Nichols, D.A. Dixon, J. Phys. Chem. A 107 (2003)
4184.
[37] F. De Proft, P. Geerlings, J. Chem. Phys. 106 (1997) 3270.
[38] W.A. Shapley, D.P. Chong, Int. J. Quantum Chem. 81 (2001) 34.
[39] S. Joant�eguy, G. Pfister-Guillouzo, H. Chermette, J. Phys. Chem.
A 103 (1999) 3505.
[40] D.P. Chong, O.V. Gritsenko, E.J. Baerends, J. Chem. Phys. 116
(2002) 1760.
[41] T. Tsuneda, T. Suzumura, K. Hirao, J. Chem. Phys. 110 (1999)
10664.
[42] J.P. Perdew, M. Levy, Phys. Rev. Lett. 51 (1983) 1884.
[43] M. Levy, J.P. Perdew, in: D.C. Langreth, H. Suhl (Eds.), Many-
body Phenomena at Surfaces, Academic, New York, 1984.
[44] F. Della Sala, A. G€orling, J. Chem. Phys. 115 (2001) 5718.
[45] D.J. Tozer, N.C. Handy, J. Chem. Phys. 109 (1998) 10180.
[46] D.J. Tozer, J. Chem. Phys. 112 (2000) 3507.
[47] M.E. Casida, D.R. Salahub, J. Chem. Phys. 113 (2000) 8918.
[48] A. Lembarki, F. Rogemond, H. Chermette, Phys. Rev. A 52
(1995) 3704.
[49] C.A. Ullrich, P.G. Reinhard, E. Suraud, Phys. Rev. A 62 (2000)
053202.
[50] X.M. Tong and, S.I. Chu, Phys. Rev. A 55 (1997) 3406.
[51] R. van Leeuwen, E.J. Baerends, Phys. Rev. A 49 (1994) 2421.
[52] O.V. Gritsenko, P.R.T. Schipper, E.J. Baerends, Chem. Phys.
Lett. 302 (1999) 199.
[53] P.R.T. Schipper, O.V. Gritsenko, S.J.A. van Gisbergen, E.J.
Baerends, J. Chem. Phys. 112 (2000) 1344.
[54] M. Gr€uning, O.V. Gritsenko, S.J.A. van Gisbergen, E.J. Baer-
ends, J. Chem. Phys. 114 (2001) 652.
[55] M.E. Casida, K.C. Casida, D.R. Salahub, Int. J. Quant. Chem..
70 (1998) 933.
[56] F. Della Sala, A. G€orling, J. Chem. Phys. 116 (2002) 5374.
[57] S. Goedecker, C.J. Umrigar, Phys. Rev. A 55 (1997) 1765.
[58] U. Lundin, O. Eriksson, Int. J. Quant. Chem. 81 (2001) 247.
[59] S.H. Vosko, L. Wilk, J. Phys. B: At. Mol. Phys. 16 (1983) 3687.
[60] P. Cortona, Phys. Rev. A 34 (1986) 769.
[61] M.F. Politis, P.A. Hervieux, J. Hanssen, M.E. Madjet, F. Martin,
Phys. Rev. A 58 (1998) 367.
[62] J. Chen, J.B. Krieger, Y. Li, G.J. Iafrate, Phys. Rev. A 54 (1996)
3939.
[63] J.C. Slater (Ed.), Quantum Theory of Molecules and Solids, The
Self-Consistent Field for Molecules and Solids, vol. 4, Mc Graw
Hill, New York, 1974.
[64] M.J. Frisch, G.W. Trucks, H.B. Schlegel, G.E. Scuseria, M.A.
Robb, J.R. Cheeseman, J.A. Montgomery Jr., T. Vreven, K.N.
Kudin, J.C. Burant, J.M. Millam, S.S. Iyengar, J. Tomasi, V.
Barone, B. Mennucci, M. Cossi, G. Scalmani, N. Rega, G.A.
Petersson, H. Nakatsuji, M. Hada, M. Ehara, K. Toyota, R.
Fukuda, J. Hasegawa, M. Ishida, T. Nakajima, Y. Honda, O.
Kitao, H. Nakai, X. Li, J.E. Knox, H.P. Hratchian, J.B. Cross, C.
Adamo, J. Jaramillo, R. Gomperts, R.E. Stratmann, R. Cammi,
C. Pomelli, J. Ochterski, P.Y. Ayala, K. Morokuma, W.L. Hase,
P. Salvador, J.J. Dannenberg, V.G. Zakrzewski, S. Dapprich,
A.D. Daniels, M.C. Strain, O. Farkas, D.K. Malick, A.D.
Rabuck, K. Raghavachari, J.B. Foresman, J.V. Ortiz, Q. Cui,
A.G. Baboul, S. Clifford, J. Cioslowski, B.B. Stefanov, G. Liu, A.
Liashenko, P. Piskorz, I. Komaromi, R.L. Martin, D.J. Fox, T.
Keith, M.A. Al-Laham, C.Y. Peng, A. Nanayakkara, M. Chal-
lacombe, P.M.W. Gill, B. Johnson, W. Chen, M.W. Wong, C.
Gonzalez, J.A. Pople, Gaussian Development Version, Revision
A.01, Gaussian Inc., Pittsburgh PA, 2002.
[65] A.D. Becke, Phys. Rev. A 38 (1988) 3098.
76 I. Ciofini et al. / Chemical Physics 309 (2005) 67–76
[66] C. Lee, W. Yang, R.G. Parr, Phys. Rev. B 37 (1988) 785.
[67] R. Krishnan, J.S. Binkley, R. Seeger, J.A. Pople, J. Chem. Phys.
72 (1980) 650.
[68] T. Clark, J. Chandrasekhar, P.v.R. Schleyer, J. Comp. Chem. 4
(1983) 294.
[69] CRC Handbook of Chemistry and Physics, 69th ed., CRC Press,
Boca Raton, 1988.
[70] R. Pollet, A. Savin, T. Leininger, H. Stoll, J. Chem. Phys. 116
(2002) 1250.
[71] O.V. Boltalina, I.N. Ioffe, L.N. Sidorov, G. Seifert, K. Vietze, J.
Am. Chem. Soc. 122 (2000) 9745.
[72] O.V. Boltalina, E.V. Dashkova, L.N. Sidorov, Chem. Phys. Lett.
256 (1996) 253.
[73] H. Razafinjanahary, F. Rogemond, H. Chermette, Int. J. Quant.
Chem. 51 (1994) 319.
[74] J. Cioslowski, K. Raghavachari, J. Chem. Phys. 98 (1993) 8734.
[75] S. D�ıaz-Tendero, M. Alcam�ı, F. Mart�ın, J. Chem. Phys. 119
(2003) 5545.
[76] S. Patchkovskii, T. Ziegler, J. Chem. Phys. 116 (2002) 7806.
[77] H. Chermette, A. Lembarki, F. Rogemond, H. Razafinjanahary,
Adv. Quant. Chem. 33 (1999) 105.
[78] R. Neumann, R.H. Nobes, N.C. Handy, Mol. Phys. 87 (1996) 1.
[79] P.R.T. Schipper, O.V. Gritsenko, E.J. Baerends, Phys. Rev. A 57
(1998) 1729.
[80] O.V. Gritsenko, B. Ensing, P.R.T. Schipper, E.J. Baerends, J.
Phys. Chem. A 104 (2000) 8558.
[81] O.V. Gritsenko, P.R.T. Schipper, E.J. Baerends, J. Chem. Phys.
107 (1997) 5007.
[82] N.C. Handy, A.J. Cohen, Mol. Phys. 99 (2001) 403.
[83] H. Chermette, I. Ciofini, F. Mariotti, C. Daul, J. Chem. Phys. 114
(2001) 1447.
[84] V. Polo, E. Kraka, D. Cremer, Theor. Chem. Acc. 107 (2002)
291.
[85] V. Polo, J. Gr€afenstein, E. Kraka, D. Cremer, Chem. Phys. Lett.
352 (2002) 469.
[86] V. Polo, E. Kraka, D. Cremer, Mol. Phys. 100 (2002) 1771.