Original file was Chapter5.tex - RUA:...
-
Upload
truongtram -
Category
Documents
-
view
219 -
download
0
Transcript of Original file was Chapter5.tex - RUA:...
This is a previous version of the article published in Theoretical Chemistry Accounts. 2015, 134: 1602. doi:10.1007/s00214-014-1602-6
Non parametrized functionals with empirical
dispersion corrections: a happy match?
Diane Bousquet1, Eric Brémond1, Juan C. Sancho-García2, Ilaria Ciofini1 and Carlo Adamo1,3*
1Laboratoire d’Electrochimie, Chimie des Interfaces et Modélisation pour l’Energie, CNRS
UMR-7575, Chimie ParisTech, 11 rue P. et M. Curie, F-75231 Paris Cedex 05, France.2Departamento de Química-Física, Universidad de Alicante, E-03080 Alicante, Spain 3Institut Universitaire de France, 103 Boulevard Saint Michel, F-75005 Paris, France.
Abstract
The performances of two parametrized functionals (namely B3LYP and B2PYLP) have been
compared to those of two non parametrized functionals (PBE0 and PBE0-DH) on a relatively
large benchmark set when three different types of dispersion corrections are applied (namely
the D2, D3 and D3(BJ) models). Globally the MAD computed using non parametrized
functionals decreases when adding dispersion terms although the accuracy not necessarily
increases with the complexity of the model of dispersion correction used. In particular, the
D2 correction is found to improve the performances of both PBE0 and PBE0-DH, while no
systematic improvement is observed going from D2 to D3 or D3(BJ) corrections. Indeed
when including dispersion, the number of sets for which PBE0-DH is the best performing
functional decreases at the benefit of B2PLYP.
Overall our results clearly show that inclusion of dispersion corrections is more beneficial to
parametrized DH functionals than for non parametrized ones. The same conclusions globally
hold for the corresponding global hybrids, showing that the marriage between non
parametrized functionals and empirical corrections may be a difficult deal.
*corresponding author: [email protected]
1. Introduction
In the field of Density Functional Theory (DFT), the development of more accurate
approximate forms of exchange correlation functionals for the description of a wider variety
of properties, ranging from thermochemistry to photophysics, still remains a major research
topic. In the last few years, many functionals belonging to the higher rung of the Perdew
ladder1 have indeed been developed. Among these, global or range-separated hybrid
functionals, which are characterized by a dependence on occupied Kohn-Sham (KS) orbitals
due to the inclusion of a fraction of Hartree-Fock (HF) like exchange, such as for instance
PBE02, B3LYP3, CAM-B3LYP4 or LC-wPBE5-7, normally present a good performances to
computational time ratio, and for this reason they have been intensively applied to model the
reactivity and properties of many chemical systems during the last years8-11. More recently,
following an approach firstly proposed by Truhlar12, different double-hybrid (DH) functionals
have been developed including also a dependency on the virtual KS orbitals. This latter
dependence is related to the presence of a second-order perturbative Moller-Plesset (MP2)
contribution to the correlation energy13,14. Since that pioneering work, a significant number of
DHs have been developed, the most famous family of parametrized DHs being the one
introduced by Grimme and including functionals such as B2PLYP15,16 and mPW2-PLYP17.
Following the same parametrization philosophy, other double-hybrids have been developed
by Martin, such as B2GP-PLYP18, B2T-PLYP and B2K-PLYP19, demonstrating to be
particularly accurate on an impressing number of properties. Furthermore, the inclusion of a
spin component or spin-opposite scaling of the MP2 contribution was recently proven to
ameliorate the performances such in the case of DSD-BLYP20, DSD-PBEB8621 or the latest
Grimme’s functional (PWPB9522).
In this overall context a DH, the so-called PBE0-DH23-25, was recently developed by Adamo
and coworkers, following the same parameter free approach used for the development of the
parent and well performing global hybrid PBE026-32. This functional actually includes the
Perdew-Burke-Ernzerhof (PBE) exchange-correlation functional in a double hybrid
formulation without the inclusion of any empirical parameters since both the fraction of HF-
like exchange and MP2 correlation contribution were fixed making use of theoretical
constrains and physical considerations33 without resorting on fitting over different
experimental sets.
In a recent paper34, the performances of this double hybrid were tested on the extended
benchmark set GMTKN30, developed by Goerigk and Grimme35,36 and compared to those of
B2PLYP and of their parent Global Hybrids (GHs), PBE0 and B3LYP, in order to compare
2
two DHs functionals possessing the same level of complexity and representatives of the
parametrized (B2PYLP) and non parametrized (PBE0-DH) development philosophies. In
particular, many recent papers of Grimme and collaborators clearly showed that explicit
inclusion of dispersion corrections strongly ameliorate the overall performances of
functionals not only for properties directly related to dispersion, such as for instance
structures of systems of biological relevance, but also for properties that could appear, at first
sight, hardly affected by London interactions37-39.
The development of corrections able to correctly deal with London forces, has indeed become
a flourishing field in DFT. Clearly, the inclusion of a fraction of MP2 correlation in global
hybrids represents a good starting point to correctly estimate non covalent interactions, such
as van der Waals ones. Indeed, a systematic improvement of functionals’ performances going
from global to corresponding double hybrid functionals was observed for benchmark sets
dominated by dispersion and non covalent interactions36. The better performances of DH with
respect to GH is indeed related to the different long range behavior of the correlation energy
terms in the two families of functionals: an exponential decrease exponentially with the
interatomic distance in the case of GH versus the presence of a long range term in DH, due
the inclusion of an MP2 correlation fraction. However, according to Grimme works22,40, the
contribution of the MP2 correlation term is not sufficient to compensate the DFT repulsive
part, so that even at DH level, dispersion forces are not correctly described making the
addition of an empirical correction term necessary for the correct description dispersion
interactions41.
To this end, different empirical corrections, containing parameters fitted on relevant
benchmark sets, were developed and tested in conjunction of a wide variety of parametrized
DHs showing excellent results. Nonetheless, up to now none of these corrections were tested
when combined to a non-parametrized DH such as PBE0-DH. In order to test if the
systematic and sizable improvement of DH performances by inclusion of empirical
dispersion corrections observed for parametrized DHs also holds in the case of PBE0-DH, the
inclusion of three different empirical dispersion corrections, namely the so-called –D238, D339
and D3(BJ)42-45 ones, in the PBE0-DH functional form will be considered in this work. The
performances of the so-obtained dispersion corrected DH will be compared to those of an
empirical DH of analogous complexity (B2PLYP), also dispersion corrected, in order to
basically show if non-empirical DH and empirical dispersion corrections really make a happy
match.
3
The paper is structured as follows: after a brief recall of the dispersion corrections that will be
applied and a description of the fitted parameters that will be used (Section 2), the
computational details will be given in Section 3. Next, the performances of dispersion
corrected PBE0-DH and B2PYLP will be compared over the GMTKN30 benchmark set
(Section 4) and some general conclusions will be drawn in Section 5.
2. Dispersions interactions: the models and the parameters
In order to deal with inter- and intra-molecular dispersion contributions in the DFT
framework, a set of corrections to Kohn-Sham energies where recently introduced by
Grimme, namely the DFT-D137, DFT-D238 and, more recently, the DFT-D339 models, these
can be viewed as a kind of hierarchy of increasing complexity arising from the approximate
yet accurate modeling of pair-wise long-range interactions between weakly interacting
densities or fragments.
Basically, the total energy of a system can be written as the sum of the Kohn-Sham energy
and of an energy correction for dispersion interactions as follows:
EDFT−Dx=EKS−DFT+ EdispDx
(
1
)
where EdispDx is an atom pair wise correction term expressed as:
EdispDx =∑
k=2
x
Edisp(k)
(
2
)
and where Edisp(k) (for k>1) represents respectively the two body (k=2) and three-body (k=3)
dispersion correction terms. Truncated expression of (2) to the first order correction (x=2)
leads to the so-called DFT-D2 correction, while the DFT-D3 model is obtained truncating for
x=3. The leading term is indeed the two-body one, expressed as:
Edisp(2) =−∑
n=3
Norder
s2 n ∑i=1
N n−1
∑j=i+1
Nn C2 nij
r ij2 n f dmp ,2n ( Rij )
(
3
)
where Nn represent the number of atoms in the system, n represent the order (here limited to
3) Rij'=|Ri
'−R j'|is the internuclear distance between atomsi and j and C2n
ij is the dispersion
coefficient for a pair of atoms ij, calculated as the geometrical average of the dispersion
coefficient of the atoms pair:
4
C2nij =√C2 n
i C2 nj
(
4
)
In equation (3), f dmp ,2n is a damping function introduced in order to avoid as much as possible
the double counting of electronic correlation at intermediate distances. f dmp actually defines
the range of in which the correction plays a role, and, in the case of the DFT-D2 model is
expressed as:
f dmp ,2 n ( rij)=1
1+exp[−d ( r ij
r iVdW +r j
VdW −1)](
5
)
Starting from the DFT-D2 model, the DFT-D3 one was developed allowing to obtain a better
accuracy for the description of all systems, including particularly metallic and heavy atoms
containing systems39, while reducing concomitantly the empiricism contained in the model. In
this latter model, the dispersion coefficients specific to each atoms pair and the cutoff radii
are explicitly calculated, taking into account the chemical environment too. On the other hand
the three body terms have been proven to yield negligible effects or even a deterioration of
the results when included for small sized systems, most probably related to an overestimation
of the three body terms in the strong density regions39. Therefore these terms are often
neglected, although they may play a role, as in the case of the Grimme S12L subset46 for
which three-body corrections are proved to play an important role. In this line it is worth
mentioning the recent work of Tkatchenko and collaborators dealing with the effect of many
body corrections47.
Overall, the dispersion correction term at DFT-D3 level (Edisp(3) ) is thus very similar to DFT-D2
one, except that all the terms in n should be taken into account. The scaling factors s2n (eq. 3)
are adjusted in the case of n<3, to allow a correct asymptotic behavior, which is only the case
when C6ij is exact (s6=1). For functionals that already take into account a part of the long
range correlation, a correct asymptotic behavior will only be reproduced when using s6 ≠1,
which will be the case for the studied double hybrids PBE0-DH and B2PLYP. Indeed, after
intensive testing, Grimme found out that the inclusion of terms higher in n−¿order than 8 can
generate instability while not upgrading the performances significantly39. Thus, as in the case
of DFT-D2, only s6 and s8 terms will be included. On the other hand, the DFT-D3 zero
damping function used is slightly different from the one of the DFT-D2 model, as proposed
by Chai and Head Gordon48:
5
f dmp ,2n ( rij)=1
1+d ( rij
sr ,2n+R0ij )
−α 2 n
(
6
)
where sr ,n is the scaling factor dependent on the cutoff radii (R0ij). Grimme38 proposed to fix
sr ,8 to a value of unity, needing thus to only optimize the sr ,6 value. Regarding the α n
parameters, they were adjusted manually to fit some physical constrains39, being α 6=¿ 14 and
α 8=α6+2=¿ 16.
At last, the DFT-D3 model was also tested using the damping function presented by Becke
and Johnson, leading to the so-called DFT-D3(BJ). In this variant of the DFT-D3 dispersion
model, the dispersion energy is expressed as:
EdispDFT− D3(BJ )=− ∑
n=6,8 , …sn ∑
i=1
Nn−1
∑j=i+1
N n C 2nij
rij2 n+ f BJ ,2n(Rij
0)2 n
(
7
)
with the corresponding damping function defined as f BJ , 2n(R¿¿ij0)=a1 Rij0+a2¿. Because of
the form of the dumping funtion, there is no need to define cutoff radii. The parameters a1
and a2 are fitted parameters, and Rij0 is defined as:
Rij0=√C8
ij
C6ij
(
8
)
6
In order to use the three above mentioned models in conjunction with the PBE0-DH
functional, some parameters were needed to be defined. All coefficients used in this work are
listed in Table 1, along with the corresponding values obtained from literature when
available35,40,49-51. In the case of the DFT-D2 model, being d fixed to 20, the only parameter
needed, that is the s6coefficient51, was obtained via a fitting procedure on the S2249 subset.
For the DFT-D3 model, for which the d value is fixed to 6, three parameters need to be fitted,
and the procedure described by Grimme39 was followed. This latter consists in fixing the s6
parameter to one in order to ensure a correct asymptotic behavior for global hybrid
functionals (ie PBE0 and B3LYP). The determination of s6 for double hybrids is done based
on a system of three rare gas dimers (Ne2, Ar2 et Kr2). By applying this method to PBE0-DH,
a value of 0.84 is obtained for s6 . Then, the other parameters are determined by minimizing
the Mean Absolute Deviation (MAD) with a non linear least square method on the S130 set.
This set is composed of the following subsets: S2239,49,50 S22+, PCONF52, SCONF53,
ACONF54, CYCONF55, ADIM656, RG657-60. Note that S22+ contains the same complexes as
the S22 set, only considering systems at larger internuclear distances. The total MAD is then
calculated, using the MAD of RG6 weighted by 20. The mapping of the MAD evolution and
its dependence on the s8 and sr ,6 parameters, can be found in Figure 1. This procedure allows
us to find s8 and sr ,6 values of 0.784 and 1.394, respectively as reported in Table 1. Regarding
the optimization of the DFT-D3(BJ) parameters, the procedure followed is strictly the same
applied for the DFT-D3 model, and allow obtaining s8 = 0.095, a1 = 0.0 and a2 = 6.102.
Inspection of the parameters obtained for the different GHs (B3LYP and PBE0) and DHs
(B2PYLP and PBE0-DH) allows to drawn some general comments. First, in the case of the
DFT-D2 model, the s6 values are substantially different for B3LYP and PBE0; the latter
being significantly smaller while PBE0-DH and B2PYLP shows significantly similar
parameters, the latter being only slightly larger. On the other hand, in the case of DFT-D3
models, PBE0-DH always possesses a larger s6 and a smaller s8. Interestingly, for PBE0-DH
in the case of D3(BJ) model for the same s6 value two sets of parameters providing very
similar MAEs on the fitting subset (i.e. the S130 one) have been found. The one (hereafter
referred to as PBE0-DH-D3(BJ)) providing the smallest MAE (s6= 0.840; a1=0.0 ; s8= 0.095
and a2= 6.102, Table 1) will be used throughout in the paper, and the results obtained using
the second set parameters, referred to as PBE0-DH-D3(BJ)2 (s6= 0.840; a1=0.345 ;s8= 0.747
and a2= 5.276, Table 1) and providing a MAE of only 0.003 kcal/mol larger on the S130, will
be discussed in section 4.
7
Interestingly, these two sets of parameters closely resemble to the D3(BJ) parameters
obtained for DSD-BLYP and PWPB95 functionals (both having a null a1 value) and for the
B2PLYP one, clearly pointing out how the search of global minima over the MAE surface
can be quite tricky and that parameter of the same range can be obtained independently of the
functional used.
3. Computational details
The PBE0-DH functional has been implemented in a local version of the Gaussian program61,
so that all the standard features, including analytical derivatives for geometry optimization,
harmonic frequencies and properties, are also available. In analogy with previous work (see
for instance references 2, 36, 62, 63), the performances of the PBE0, B3LYP, B2PLYP and
PBE0-DH functionals have been tested on the GMTKN30 database using both Gaussian and
Orca64 Version 2.9.1-RELEASE program packages.
To easily compare the obtained results with the values presented by Grimme, we used in
particular the large Ahlrichs’ type quadruple-ζ basis set def2-QZVP,65 which is expected to
provide converged results even for DHs. Indeed, the use of a quadruple-ζ base is sufficient to
reach the convergence limit,51,66 even if DH approaches could have a significantly larger
dependency on basis set then corresponding non-DH forms, due to the presence of the MP2
component. Regarding the calculations of electron affinities, and of binding energies of water
(subsets G21EA67 and WATER2768), diffuse s and p functions were taken from the Dunning
aug-cc-pVQZ,69 obtaining the aug-def2-QZVP basis, which was recommended by Grimme
for those particular subsets.
The PCONF52, WATER2768, IDISP70, ISOL2271, ADIM656, BHPERI72 and BSR3673-77 were
computed making use of the RI-MP278-80 approach as implemented in Orca. Available data
for B3LYP, B2PYLP, PBE0, B3LYP-D3, B2PYLP-D3 and PBE0 were retrived from the
Grimme website81.
8
4. Results and discussion
Being mostly interested in the effect of dispersion correction on non parametrized
functionals, firstly the global behavior of PBE0-Dx and PBE0-DH-Dx functionals will be
compared to that computed in absence of dispersion correction on the whole GMTKN3035
set. Next, the performances of the different functionals will be analyzed over different groups
of subsets, defined in a previous paper36 and based on the dominant properties they are aimed
to benchmark, in order to see which type of interaction are mostly affected by the inclusion of
dispersion. In particular, the GMTKN30 datasets will be classed in subset probing
thermochemistry (decomposition and atomization energies), adiabatic processes, reaction
energies and barriers heights, conformers and isomers relative energies. On the other hand the
DC982-89 subset, which includes nine reactions known to be difficult to be treated at DFT
level, and SIE1133 subset, dealing with test cases for self interaction error, will be discussed
separately. The MADs computed on each GMTKN30 subset are reported in Supporting
Information.
4.1 Dispersion corrected global (PBE0) and double hybrid (PBE0-DH) benchmarked over
the GMTKN30 set
The overall performances of PBE0 and PBE0-DH over the GMTKN30 set with and without
inclusion of dispersion corrections are summarized in Figure 2 by their associated Mean
Absolute Deviation (MAD), together with the results obtained at B3LYP and B2PLYP level.
Globally speaking, the MAD computed with both PBE0 and PBE0-DH decreases when
adding dispersion terms, although not necessarily after increasing the complexity of the
correction used. In particular, if the D2 correction sizably ameliorates the performances of
both PBE0 and PBE0-DH, the systematic improvement observed going from -D2 to -D3 (or -
D3(BJ)) corrections in the case of B3LYP is here not observed, in analogy to what is
computed in the case of B2PLYP.
On the other hand, both in the case of the non-parametrized and parametrized families of
hybrids, the effects of dispersion correction terms are quantitatively larger on global hybrids
with respect to double hybrids so that the computed MAD difference between DH and
corresponding GH decreases by addition of dispersion. For instance, the difference of MADs
between PBE0 and its corresponding double hybrid (PBE0-DH) is smaller when adding
dispersion corrections, and even if PBE0-DH still performs better than PBE0, the difference
in MAD decreases from 0.29 kcal/mol when dispersion is not included to 0.08, 0.19 or 0.15
kcal/mol when dispersion is included at the D2, D3 and D3(BJ) level, respectively.
9
Globally speaking, this is also what is observed in the case of the parameterized family of
functionals (BLYP and B2PLYP) although the overall MAD difference between GH and DH
is always larger than 1 kcal/mol. Furthermore, while the largest decrease in MAD is
computed when applying the D3 type corrections in the case of B3LYP, for PBE0 the largest
effect is found when using the D2 corrections. Indeed, both for PBE0-DH and PBE0, the
more complex D3 corrections seem to add a negligible accuracy.
The difference in performances of the two families of functionals (i.e. the PBE-based and
BLYP-based) could be related to the dissimilar behavior of the exchange component in the
limit of large reduced density gradient. Indeed, it has been shown that the Becke's exchange
functional (B88)90 is repulsive in dispersion-bound complexes so that simple van der Waals
systems are precited to be unbounded91,92. This is not the case for the PBE exchange, which is
more actractive93. It is therefore likely that BLYP based functionals benefit more from the
always attractive, DFT-D type corrections. More insigths on this point could be obtained by
analyzing the performances of “cross” functionals, such as BPBE-DH or PBE-LYP-DH.
However, these unusual exchange and correlation functioals’ combinations are of scarse
theoretical interest, since the PBE contributions were defined to match each-other94.
Furthermore, a correct definition of these functionals requires the full optimizations of the
coefficients ruling the different contributions to the functionals, so that the differences
between the two orginal DH functionals (PBE0-DH and B2PLYP) and their derivatives could
be not clearly ascribed to a single effect, making the obtained results difficult to interpret. In
this intricate context, we believe that this point does not deserve further explorations and
prefer to rest on the above mentioned qualitative interpretation of the exchange behavior.
Analyzing in more details the results obtained over the GMTKN30 database, with and
without dispersion corrections for PBE0 and PBE0-DH, one can notice that, in absence of
dispersion the PBE0-DH functional is better than its parent PBE0 one for 20 over 31 subsets.
Adding the dispersion corrections computed using either the -D2 or the -D3 models, PBE0-
DH is still better than PBE0 for 20 subsets only. Indeed, even if the number of subsets where
PBE0-DH outperforms PBE0 does not change, this is only fortunate, because the concerned
sets are not the same. For example, the O3ADD694 and ALK639 subsets, for which PBE0
provides better results in absence of correction, are better described by PBE0-DH when
adding dispersion terms. Analogously, the IDISP70 subset presenting lower MADs for PBE0-
DH is actually better described at PBE0 level when adding the -D2 correction. Generally, the
MADs over the different subsets vary in consistent way for the two functionals when adding -
10
D2 or -D3 dispersion corrections with the only exception of the O3ADD6, AL2X95,
ISOL2271, DC9, ALK6, and IDISP subsets.
On the other hand, regarding the results obtained when adding dispersion corrections at DFT-
D3(BJ) level, a significant deterioration of the performances of PBE0-DH functional with
respect to the parent PBE0 one is found. Indeed, PBE0-DH-D3(BJ) outperforms PBE0-
D3(BJ) only for 15 subsets over 31.
For PBE0 and PBE0-DH, and for all subsets concerning thermochemistry, a deterioration of
the performances can be observed when adding dispersion terms, except for the BSR3672-77
subset, which shows a reduction of the MAD of 6.48 kcal/mol (for PBE0-D2) and 4.55
kcal/mol (for PBE0-DH-D2). For adiabatic processes, the inclusion of dispersion does not
play an important role while the change in accuracy when including dispersion corrections
for subsets probing reactions energies is quite inhomogeneous.
Indeed, for some subsets, such as BH76, BH76RC, the inclusion of dispersion does not
change the overall behavior of PBE0 and PBE0-DH while for other subsets, O3ADD694,
G2RC96, NBPRC97,98 , DARC95, ALK639 et AL2X95, results are sizably worst or better, such as
for the BHPERI72 or RSE4399,100 subset. On the other hand, and as it was expected, the
inclusion of dispersion corrections allows for a significant amelioration of the results and a
systematic decrease of the MADs for all subsets probing dispersion and relative energies of
conformers.
4.2 Decomposition and atomization energies
Figure 3 reports the MAD computed for PBE0, B3LYP, PBE0-DH, B2PYLP (both in
presence and in absence of different dispersion corrections) for all four subsets concerning
thermochemistry, that are: the MB08-165101, containing the decomposition energies of
artificial molecules, the W4-0818, including atomization energies of 99 small artificial
molecules, the W4-08woMR, obtained by removing multireference cases of the previous
subset, and the BSR3673-77 which includes bond separations reactions. Those four sets present
large MAD but, since the reference data are large, the MADs obtained are, on a relative scale,
not so large. Interestingly, a reduction of MADs by inclusion of dispersion terms is observed.
For B2PLYP and PBE0, the smallest MADs are found using the -D2 correction, with a
difference of 1.44 kcal/mol and 1.47 kcal/mol, respectively, with respect to non corrected
values. For B3LYP functional, a systematic decrease of the MAD is observed following the
dispersion correction model used, with a reduction of the MAD with respect to the non-
dispersion corrected results going from 0.23 kcal/mol (DFT-D2), 1.82 kcal/mol (DFT-D3)
11
and 2.84 kcal/mol for the DFT-D3(BJ) model. Basically, it seems that B3LYP benefits most
of corrections of the D3 form. The same trend is indeed found for PBE0-DH although the
overall amelioration of performances is quantitatively smaller (MADs difference between
PBE0-DH results and PBE0-DH-D3(BJ) of only 1.17 kcal/mol).
Overall, the results obtained show that the parameterized family of functionals, already
providing the best results in absence of dispersion corrections in the case of B2PYLP, is the
one which benefits the most (in the case of the B3LYP functional) of the inclusion of
dispersion corrections especially of the -D3 type. Indeed, besides global MADs computed for
all the different subsets, it is interesting to analyze the effects of the inclusion of dispersion
for each subset, since, as previously mentioned, the behavior is rather inhomogeneous.
Indeed for the MB08-165, a significant decrease in errors is computed for B2PLYP with
MADs going from 5.12 kcal/mol (B2PLYP) to 3.67 kcal/mol (B2PLYP-D3(BJ)). For
B3LYP, however, a reduction of MAD is only obtained when using the -D3 or -D3(BJ)
corrections, with differences of 1.8 kcal/mol and 2.64 kcal/mol, respectively, with respect to
the B3LYP MADs. For the PBE0 family, inclusion of dispersion only leads to an increase of
the errors: the most important deviations are observed for the DFT-D2 model, with an
increase in MAD with respect to the non dispersion-corrected values of 0.66 (PBE0-D2) and
0.54 kcal/mol for PBE0-DH-D2.
For the W4-08 subset on the other hand, only small variations are observed among inclusion
of dispersion independently of the functional used. The most important improvements are
observed using -D3(BJ) dispersion model, with reduction of the MADs of 0.87 kcal/mol
(B3LYP), 0.59 kcal/mol (B2PLYP), 0.07 kcal/mol (PBE0) and 0.45 kcal/mol (PBE0-DH). A
similar behavior is found for the W4-08 woMR subsets, with the best improvement on MAD
obtained also applying the -D3(BJ) model but with very small overall effects ranging from
0.01 kcal/mol (PBE0) to 0.46 kcal/mol ( B3LYP).
On the other hand, the BSR36 subset, probing bond separation reactions, seems to be the
most affected by the inclusion of dispersion corrections. The minimal MADs are computed
making use of the -D2 model, with extremely low MAD values of 0.76 kcal/mol (B2PLYP),
1.60 kcal/mol (PBE0) and 1.14 kcal/mol (PBE0-DH). The most important reduction in MAD
observed after introduction of dispersion terms is of 6.48 kcal/mol (PBE0-D2), whereas a
decrease of 4.55 and 4.52 kcal/mol is computed for PBE0-DH-D2 and B2PLYP-D2
respectively. As noticed above, the B3LYP functional presents a behavior slightly different
from the other functionals with an increase in accuracy obtained only starting from -D3 type
dispersion models, the best results being obtained at DFT-D3(BJ) level with an associated
12
MAD of 3.84 kcal/mol, representing thus a decrease of 7.41 kcal/mol with respect to the
value computed without any dispersion corrections.
4.3 Adiabatic processes
Three different subsets, namely the G21IP, G21EA67 and the PA102-104 ones, probing adiabatic
ionization potentials, electron and proton affinities, are used to estimate the performances of
the functionals in the prediction of adiabatic processes. When looking at Figure 4, which
reports the MADs for these subsets, we can observe that the performances of all the
considered functionals are practically unaffected by the inclusion of dispersion. Indeed,
adding dispersion correction implies a systematic but almost negligible increase of the MAD
for B3LYP, B2PLYP and PBE0-DH, whereas the best value for PBE0 is obtained for DFT-
D3 model but with a reduction of MAD with respect to the pure PBE0 functional of only 0.02
kcal/mol.
Analyzing the subsets, the G21IP one, including 36 adiabatic ionization potentials of atoms
and small molecules, non parameterized functionals are clearly not influenced by dispersion
terms with difference in MAD with respect to pure functionals smaller than 0.03 kcal/mol.
For the G21EA subset, including 25 adiabatic electron affinities, the behavior is similar for
the two double hybrids, presenting a maximum enhancement of 0.03 kcal/mol. The
dependence on dispersion for hybrids is slightly larger, since the MAD computed with
B3LYP varies from 1.81 (for DFT and DFT-D3) to 2.44 kcal/mol for DFT-D2 and DFT-
D3(BJ).
Analogously, for the PA set, which includes 12 adiabatic proton affinities of four conjugated
polyenes104 and eight small molecules102-103, B2PLYP and B3LYP results are sizably better
than PBE0-DH and PBE0 ones, and dispersion corrections do not change significantly the
functional behavior, actually very slightly deteriorating the MADs.
Globally, a dispersion correction for non parameterized functionals seems to be useless for
the description of adiabatic processes.
4.4 Barriers heights and reaction energies
The Figure 5 shows the MADs computed for all subsets discussed in this section. Generally
speaking, if the parametrized functionals benefit from the inclusion of dispersion terms, with
a decrease of the MADs from 5.75 kcal/mol (B3LYP) to 3.16 kcal/mol (B3LYP-D3(BJ)) for
example, this global trend is not observed for the PBE0 family. For this family indeed, an
increase of the MADs is systematically found upon inclusion of dispersion corrections with a
13
maximal deterioration of performances of 0.71 kcal/mol for PBE0-D2 and 1.47 kcal/mol for
PBE0-DH-D3(BJ).
Considering the performances obtained for barrier heights on two subsets, the BHPERI54 and
the BH76105,106, as for adiabatic processes, no amelioration in the MADs regarding the BH76
subset can be found when adding dispersion corrections. Better results are indeed obtained
for the non dispersion corrected functionals, and for example, the MAD of PBE0 increases
from 4.12 to 4.56 kcal/mol when adding the -D2 correction. The same behavior is observed
for all functionals, the largest increase being obtained for B3LYP-D3(BJ) with a
augmentation of 0.76 kcal/mol.
Globally, the dispersion correction allows a nuanced amelioration of the results. Indeed, if
PBE0-DH-D3 gives smaller MADs than PBE0-DH-D2 (improvement of 0.31 kcal/mol),
functional which improves the results shown by PBE0-DH of 0.42 kcal/mol, this tendency is
not followed by the others functionals. The two global hybrids are not getting better when
adding dispersion using the DFT-D2 model, and accurate performances are only found
starting from DFT-D3 correction models, with a MAD of 2.79 kcal/mol for B3LYP.
B2PLYP-D2, on the contrary shows smaller values of MADs than B2PLYP-D3. Overall,
double hybrids are better than their hybrids parents for barrier heights, but on the BHPERI
set, the differences become smaller and smaller when adding the dispersion, to finally obtain
similar MAD for all functionals : from 1.29 kcal/mol (B2PLYP-D3(BJ) to 1.94 for PBE0-
D3(BJ), for example.
For the corresponding reaction energies (included in the BH76RC) the computed MADs are
overall smaller, the largest MAD being 2.59 kcal/mol at PBE0-D3(BJ) level. Including
dispersion, the PBE0 family shows a slight increase of the MADs, with difference in MAD
with respect to non-dispersion corrected functional between 0.03 kcal/mol (PBE0-D3) and
0.08 kcal/mol (PBE0-DH-D2), whereas B2PLYP results are practically insensitive and
BL3YP ones are slightly enhanced by dispersion corrections.
Also in the case of the O3ADD694 set, including the reactions of addition of ozone on ethane
or ethyne, the computed MADs are found to increase upon inclusion of dispersion correction
terms. One interesting example is given by the behavior of the B3LYP functional, with the
best MAD of 1.97 kcal/mol, whose accuracy is worsen to 3.12 kcal/mol when using the
D3(BJ) correction. Indeed, the B3LYP family, with or without correction, is still
outperforming the PBE0 one on the O3ADD6 subset.
For the G2RC subset, the behavior of all functionals is almost identical in presence or
absence of dispersion interactions, while large MADs are computed when considering 14
14
Diels-Alder reactions, contained in the DARC95 set. The PBE0 functional presents in this case
the best behavior with the smallest MAD (3.13 kcal/mol) computed at PBE0-D3 level, while
PBE0-D2 and PBE0-D3(BJ) model give MADs larger than 4.0 kcal/mol. This behavior is not
followed by the PBE0-DH functional since in this case the inclusion of dispersion terms is
systematically deteriorating the MADs with an increase in MADs of 2.27 and 3.84 at PBE0-
DH-D3 and PBE0-DH-D3(BJ) level, respectively. B2PLYP results, on the other hand, are
enhanced by the inclusion of dispersion terms, the best performances being found at
B2PLYP-D2 level (4.23 kcal/mol). B3LYP benefits the most of the inclusion of dispersion
and it significantly increases its performances as the dispersion model included becomes
more complex, to reach a MAD of 7.4 kcal/mol for B3LYP-D3(BJ) which represents a half
of the MAD computed at pure B3LYP level (15.41 kcal/mol).
The performances of the functionals are not significantly modified by inclusion of dispersion
in the case of the RSE3499,100 subset, and the best results are still obtained for at double
hybrids level with MAD of 0.91 kcal/mol for B2PLYP and 0.86 kcal/mol for PBE0-DH
including dispersion at -D2 level. At this level, the MAD of PBE0 is also reduced of 0.37
kcal/mol. Regarding H2 activation, (NBPRC subset97,98), the B3LYP family benefits most
from the dispersion, with MADs of 1.86 kcal/mol for B3LYP-D3(BJ) and 1.20 kcal/mol for
B2PLYP-D3(BJ) whereas for the PBE0 family, results are getting worst by inclusion of
dispersion. The same trends are also observed for the ALK6 subset: the PBE0 family shows
an increase of MAD of 2.39 kcal/mol for D2 and 1.76 kcal/mol for D3 corrections using the
PBE0-DH functional. On the other hand, using the DFT-D3 model allows a decrease of 2.34
kcal/mol for the B2PLYP functional, which was not the case of the inclusion of the D2 model
leading, for B2PYLP, an increase of 2.79 kcal/mol.
Finally, also in the case of the AL2X95 subset, containing the binding energies of seven
dimers of AlX3 (X=H, CH3, F, Cl and Br), only the B3LPY family benefits from the inclusion
of dispersion corrections although overall the PBE0-DH functional without inclusion of
dispersion is still providing the smallest MAD.
4.5 Conformers and isomers
In the case of isomer relative energies, probed by the ISOL2271 and ISO34107 sets, the
addition of dispersion correction terms seems beneficial especially in the case of the B3LYP
family. Nonetheless, even the inclusion of dispersion does not significantly enhance the
results of the PBE0 family of functional, these latter are still overall better than corresponding
parametrized ones.
15
More in details, the ISOL22 subsets presents more important MADs than the ISO34 subset.
B2PLYP, PBE0 and PBE0-DH are more sensitive to dispersion computed using the DFT-D2
model, whereas B3LYP shows better performances with the DFT-D3 and D3(BJ) models. As
in the general case for isomers, the PBE0 family performs better than the B3LYP one in the
case of this set. The tendency is reversed for ISO34 subset, since B2PLYP is the best
performing functional (MAD of 1.08 kcal/mol for B2PLYP-D2), the PBE0 family also
presenting also better results for the DFT-D2 model, and all functionals being actually pretty
accurate when using dispersion corrections allowing to obtain MAD always smaller than 1.74
kcal/mol.
The PCONF,52 ACONF,54 SCONF53 and CYCONF55 sets all include relative energies of
conformers. The Figure 6 shows the change in the mean absolute deviation as function of the
dispersion correction used. Conformers energies are supposed to be dominated by dispersion
interactions, and this assumption can be verified by inspection of the results obtained using
the -D2 model: the MADs decrease with respect to the non corrected values of 0.66 (PBE0),
0.51 (PBE0-DH) and 0.56 (B2PLYP) practically reducing to half the errors. Furthermore,
more complex dispersion models give more accurate results, although the relative decrease in
MAD becomes less important, and so that for all functional but B3LYP the DFT-D2 model
seems to provide sufficiently good MADs. Only in the case of B3LYP, indeed the inclusion
of the –D3(BJ) correction seem to provide a sizable improvement (of 1.27 kcal/mol).
Globally, for all these subsets, double hybrids always outperform the corresponding hybrids
and the PCONF, showing the largest MADs, it is also the one that benefits the most of the
inclusion of dispersion. Again for this subset important improvement of the performances of
all functionals are observed by inclusion of dispersion at D2 level, with MAD decreasing of
1.55 (B2PLYP), 2.09 (PBE0) and 1.47 kcal/mol (PBE0-DH) with respect to the non-
corrected values.
Also in the case of the other three subsets (ACONF, SCONF, and CYCONF), dispersion
terms globally improve the MADs, leading to extremely accurate results: for example the
largest MAD for ACONF at D3(BJ) level is only 0.07 kcal/mol obtained using B2PLYP.
4.6 Dispersion and non covalent interactions
The performances of the functionals to describe dispersion and non covalent interactions
were tested using the six following sets: ADIM656, S2239,49,50, HEAVY2839 and RG657-61,
WATER2768 and IDISP70. The corresponding global MADs obtained are collected in Figure
7. Globally, double hybrids functionals present better performances than corresponding
16
hybrids even after inclusion of dispersion terms. These latter particularly enhance the
accuracy of the B3LYP family especially when introduced at -D3 and -D3(BJ) level for
B3LYP functional. Large effects are indeed computed already at -D2 level for PBE0, PBE0-
DH and B2PLYP: the change in MAD going from pure to -D2 corrected functional is indeed
larger than the one observed going from DFT-D2 to -D3 models. For example, the MAD of
PBE0-DH decreases from 2.39 to 1.55 kcal/mol when using the -D2 correction, while a
further decrease of 0.11 kcal/mol is computed going from the -D2 to the -D3(BJ) correction
models.
Considering the different subsets, the S22 and ADIM6 ones show exactly the same behavior
with best results obtained at DFT-D3(BJ) level, although very close, except for B3LYP
functionals, to errors obtained at DFT-D2 level. For these two sets, the MADs are indeed
very low, in the 0.11 kcal/mol (PBE0, ADIM6) to 0.48 kcal/mol (PBE0, S22) range.
Dispersion corrections over these two subsets also tend to level off the difference between
double and corresponding global hybrids accuracy. The same global trends are also valid for
the RG6 and HEAVY28, which indeed show very low MADs.
On the other hand, it is difficult to define global trends for the WATER27 subset. Indeed, for
the WATER27 subset, larger MADs are computed when including dispersion correction,
with the notable exception of the B3LYP functional which presents some improvement in
accuracy for all dispersion models tested. A slight enhancement of the accuracy is also found
for B2PLYP-D3 (0.29 kcal/mol).
Finally, concerning the IDISP subset, the behavior is less erratic, and the same patterns
described for the ADIM and S22 subsets can be found : an improvement of the results for
DFT-D2 models, with an impressive decrease in MAD, especially for global hybrids, with
respect to the non corrected functionals of 6.88 (B2PLYP), 9.32 (PBE0) and 4.0 kcal/mol
(PBE0-DH). The smallest MADs, of 1.27 kcal/mol, are indeed obtained at PBE0-D2 level of
theory.
4.7 Self interaction error
The performances of the four functionals over the SIE1134 subset, probing issues related to
self-interaction error, are not enhanced by the inclusion of dispersion corrections. Indeed, a
systematic increase of the MADs as a function of the inclusion of dispersion terms can be
observed. The variations in MADs are comprised between 0.02 kcal/mol (when going from
PBE0 to PBE0-D3) and 0.86 kcal/mol (when going from B3LYP to B3LYP-D3). However,
17
also when considering dispersion corrections double hybrids outperform their parent global
hybrids.
4.8 Nine difficult cases for DFT: the DC9 subset
The DC982-89 set contains nine reactions, not sharing any common features, considered as
particularly difficult cases for DFT approaches.
Globally, an improvement of the accuracy of the functionals due to the inclusion of
dispersion corrections can only be found for four reactions out of nine. Analyzing each
reaction in more details, it can be noticed that the reaction energy associated to the 2-
pyridone -> 2-hydropyridine reaction82, of only -1.0 kcal/mol, is actually very difficult to be
correctly reproduced. Only PBE0-DH gives a qualitatively and quantitatively correct result
predicting a reaction energy of -0.70 kcal/mol. Introduction of dispersion corrections globally
deteriorates the results and reaction energies of -0.48 kcal/mol (PBE0-DH-D2) and -0.55
kcal/mol (PBE0-DH-D3) are found. No change upon addition of dispersion is observed for
four reactions: the energy difference between hepta-1,2,3,5,6 hexaene and hepta-1,3,5 tryine
isomers80, the decomposition89 of S8 to diatomic S2, isomerization85 of (CH)12 and
cycloaddition between ethene and diazomethane87. A decrease of the MADs is observed for
the dimerization of tetramethyl-ethene to octamethyl-cyclobutane, especially in the case of
B2PLYP and PBE0 : 2.67 kcal/mol for PBE0-DH-D3, 1.51 kcal/mol for PBE0-D3, and 5.85
kcal/mol for B2PLYP-D3. Regarding the isomerization of C20 cage to its bowl isomer83, the
relative stability (of -13.3 kcal/mol) of the isomers is affected by large errors, independently
of the functional and of the dispersion corrections used. Only B2PLYP are improved by a –
D2 dispersion term allowing a reduction of the MAD of 5.03 kcal/mol. The same behavior is
found for the isomerization reaction of carbo-[3]-oxacarbon86, where the B3LYP family
provides relatively small errors and the dispersion further reduces the MAD computed with
B2PLYP: B2PLYP-D2 (1.63 kcal/mol) and B2PLYP-D3 (1.92 kcal/mol). The same holds for
the decomposition of the Be4 cluster into beryllium atoms88 with MADs of 6.87 (B3LYP-D3)
and 3.7 kcal/mol (B2PLYP-D3).
5.Effect of the parameterization of D3(BJ) model: the PBE0-DH case
As previously discussed in section 2, two possible set of D3(BJ) parameters can be defined
providing a very similar MAD on the S130 set for the PBE0-DH functional. Since the
difference in MAD obtained using the D3(BJ) parameterization proving the minimal MAD
and the D3(BJ)2 is very tiny (only 0.003 kcal/mol) and since the latter parameters are much
18
closer to those of Grimme for B2PLYP (refer to Table 1), we also decided to investigate the
performances of the PBE0-DH-D3(BJ)(2) dispersion model on all different GMTKN30
subsets. Detailed results are reported in Supporting Information.
Overall, the use of the PBE0-DH-D3(BJ)(2) model allows to decrease the global MAD on the
GMTKN30 database from 3.54 (computed using PBE0-DH-D3(BJ) model) to 3.49 kcal/mol
(as reported in Figure 2), that is a slight amelioration of the performances.
Most of the subsets are not affected by the change in parametrization going from the D3(BJ)
to the D3(BJ)(2) model, except for ALK6, DARC, DC9 and AL2X ones.
Indeed, considering the sets probing thermochemistry or conformers and isomers stability, a
slight increase is even found (from 5.74 to 5.60 kcal/mol for the W4-08 subset, for instance)
while the change in parametrization practically do not affect the accuracy in the prediction of
the adiabatic processes. On the other hand, subsets proving non covalent and dispersion
interactions present an overall slight improvement of the MAD for five sets out of six.
Regarding the four sets presenting largest amelioration of performances with respect to the
PBE0-DH-D3(BJ) parametrization, a decrease of 0.85, 1.18, 1.68 and 4.44 kcal/mol in MAD
is computed for the DARC, DC9, AL2X and ALK6 subsets, respectively, the latter subset
thus being the one mostly affected by the change in parametrization.
6. Conclusions
The overall performances of the functionals studied upon addition of dispersion corrections
are summarized in Figure 2 and 9.
As discussed above, globally the MAD computed using non parametrized functionals
decreases when adding dispersion terms although not necessarily increasing with the
complexity of the model of correction used. In particular, the D2 correction is found to
improve the performances of both PBE0 and PBE0-DH, but the systematic improvement
observed going from D2 to D3 (or D3(BJ)) corrections in the case of B3LYP is not observed.
On the other hand, both in the case of the non-parametrized and parametrized families of
hybrids, the effects of dispersion corrections terms are quantitatively larger on global hybrids
with respect to double hybrids so that the computed MAD difference between DH and
corresponding GH decreases by addition of dispersion terms. The Figure 9 shows how double
hybrids perform better for 28 sets out of the 31 studied: 13 times for PBE0-DH and 15 in the
case of B2PLYP. Adding dispersion does not change this ratio, since 27 sets out of 31 present
better MADs for double hybrids corrected with D2, and 29 out of 31 with D3 dispersion
correction terms.
19
Indeed when including dispersion, the number of sets for which PBE0-DH is the best
performing functional decreases at the benefit of B2PLYP: from 15 sets to 22 sets for
B2PLYP-D3. This data clearly shows that inclusion of dispersion is clearly more beneficial in
the case of the parametrized DH functionals than for non parametrized ones, independently of
the BJ parametrization used. The same conclusion globally holds for the corresponding
global hybrids, showing that the marriage between non parametrized functionals and
empirical corrections may be a difficult deal.
Supporting Information Available: Computed MADs over the different subsets.
Acknowledgments
This work was funded by the ANR agency under the project DinfDFT ANR 2010 BLANC n.
0425.
20
References
[1] J. P. Perdew, A. Ruzsinszky, L. A. Constantin, J. Sun, G. I. Csonka, J. Chem. Theory
Comput. 5, 902 (2009).
[2] C. Adamo, V. Barone, J. Chem. Phys. 110, 6158 (2009).
[3] A. D. Becke, J. Chem. Phys. 98, 1372 (1993).
[4] T. Yanai, D. Tew, N. Handy, Chem. Phys. Lett. 393, 51-57 (2004).
[5] O. A. Vydrov, G. E. Scuseria, J. Chem. Phys. 125, 234109 (2006).
[6] O. A. Vydrov, J. Heyd, A. Krukau, G. E. Scuseria, J. Chem. Phys. 125, 074106 (2006).
[7] O. A. Vydrov, G. E. Scuseria, J. P. Perdew, J. Chem. Phys. 126, 154109 (2007).
[8] P. Perdew, A. Ruzsinszky, L. A. Constantin, J. Sun, and G. I. Csonka, J. Chem. Theory
Comput. 5, 902 (2009).
[9] I. Y. Zhang, J. M. Wu, and X. Xu, Chem. Commun. 46, 3057 (2010).
[10] X. Xu and I. Y. Zhang, Int. Rev. Phys. Chem. 30, 115 (2011).
[11] J. C. Sancho-Garcia, C. Adamo, Phys. Chem. Chem. Phys. 15, 14581-14594 (2013)
[12] Y. Zhao, B. J. Lynch, D. G. Truhlar, J. Phys. Chem. A 128, 4786 (2004).
[13] M. Erzhernof, Chem Phys Lett. 263, 499-506 (1996).
[14] Y. Zhang, X. Xu, and W. A. Goddard III, Proc. Natl. Acad. Sci. U.S.A. 106, 4963
(2009).
[15] S. Grimme, J. Chem. Phys. 124, 034108 (2006).
[16] T. Schwabe, S. Grimme, Phys. Chem. Chem. Phys. 9, 3397 (2007).
[17] T. Schwabe, S. Grimme, Phys. Chem. Chem. Phys. 8, 4398 (2006).
[18] A. Karton, A. Tarnopolsky, J. F. Lamere, G. C. Schatz, J. M. L. Martin, J. Phys. Chem.
A 112, 12868-12886 (2008).
[19] A. Tarnopolsky, A. Karton, R. Sertchook, D. Vuzman, J. M. L. Martin, J. Phys. Chem. A
112, 3-8 (2008).
[20] S. Kozuch, D. Gruzman, J. M. L. Martin, J. Phys. Chem. C 114, 20801 (2010).
[21] S. Kozuch, J. M. L. Martin, Phys. Chem. Chem. Phys. 13, 20104 (2011).
[22] L. Goerigk, S. Grimme, Phys. Chem. Chem. Phys. 13, 6670–6688 (2011).
[23] K. Sharkas, J. Toulouse, A. Savin, J. Chem. Phys. 134, 064113 (2011).
[24] E. Brémond, C. Adamo, J. Chem. Phys. 35, 24106 (2011).
[25] J. Toulouse, K. Sharkas, E. Brémond, C. Adamo, J. Chem. Phys. 135, 101102 (2011).
[26] M. Ernzerhof, G. E. Scuseria, J. Chem. Phys. 110, 5029 (1999).
[27] I. Ciofini, C. Adamo, V. Barone, J. Chem. Phys. 121, 6710 (2004).
21
[28] C. Adamo, M. Cossi, G. Scalmani, V. Barone, Chem. Phys. Lett., 1999, 307, 265.
[29] X. Wu, A. Selloni, R. Car, Phys. Rev. B, 2009, 79, 085102.
[30] C. Zhang, D. Donadio, F. Gygi, G. Galli, J. Chem. Theory. Comput., 2011, 7, 1443.
[31] J. L. F. Da Silva, M. V. Ganduglia-Pirovano, J. Sauer, V. Bayer, G. Kresse, Phys. Rev.
B, 2007, 75, 045121.
[32] M. P. Waller, H. Braun, N. Hojdis, M. Bühl, J. Chem. Theory. Comput., 2007, 3, 2234.
[33] J. Harris, Phys. Rev. A, 1984, 29, 1648.
[34] D. Bousquet, E. Brémond, J. C. Sancho-Garcia, I. Ciofini, C. Adamo, J. Chem. Theory
Comput., 2013, 9, 3444-3457.
[35] L. Goerigk, S. Grimme, J. Chem. Theory Comput., 2010, 6, 107-126.
[36] L. Goerigk, S. Grimme, J. Chem. Theory Comput., 2011, 7, 291-309.
[37] S. Grimme, J. Comp. Chem., 2004, 25, 1463–1473.
[38] S. Grimme, J. Comp. Chem., 2006, 27, 1787–1799.
[39] S. Grimme, J. Antony, S. Ehrlich, H. Krieg, J. Chem. Phys., 2010, 132, 154104.
[40] T. Schwabe, S. Grimme, Accounts of Chemical Research, 2008, 41, 569–579.
[41] J. Aragó, E. Ortí, J. C. Sancho-García, J. Chem. Theory. Comput., 2013, 9, 3347-3443.
[42] S. Grimme, S. Ehrlich, L. J. Goerigk, Comput. Chem., 2011, 32, 1456.
[43] A. D. Becke, E. R. Johnson, J Chem Phys., 2005, 122, 154101.
[44] E. R. Johnson, A. D. Becke, J Chem Phys., 2005, 123, 024101.
[45] E. R. Johnson, A. D. Becke, J Chem Phys., 2006, 124, 174104.
[46] T. Risthauss, S. Grimm, J. Chem. Theory Comput., 2013, 9, 1580.
[47] A. Tkatchenko, Jr. R. A. DiStatio, R. Car, M. Scheffler, Phys. Rev. Lett., 2012, 108,
236402.
[48] J. Chai, M. Head-Gordon, Phys. Chem. Chem. Phys., 2008, 10, 6615–6620.
[49] P. Jurecka, J. Sponer, J. Cerný, P. Hobza, Phys. Chem. Chem. Phys., 2006, 8, 1985–
1993.
[50] T. Takatani, E. G. Hohenstein, M. Malagoli, M. S. Marshall, C. D. Sherrill, J. Chem.
Phys., 2010, 132, 144104.
[51] J. C. Sancho-Garcìa, Chem. Phys. Lett., 2012, 136, 535.
[52] D. Řeha, H. Valdés, J. Vondrášek, P. Hobza, A. Abu-Riziq, B. Crews, M. S. de Vries,
Chem. – A Eur. J., 2005, 11, 6803–6817.
[53] G. I. Csonka, A. D. French, G. P. Johnson, C. A. Stortz, J. Chem. Theory Comput.,
2009, 5, 679–692.
[54] D. Gruzman, A. Karton, J. M. L. Martin, J. Phys. Chem. A, 2009, 113, 11974–11983.
22
[55] J. J. Wilke, M. C. Lind, H. F. Schaefer III, A. G. Csaszar, W. D. Allen, J. Chem. Theory
Comput., 2009, 5, 1511– 1523.
[56] S. Tsuzuki, K. Honda, T. Uchimaru, M. Mikami, J. Chem. Phys., 2006, 124, 114304.
[57] E. Goll, H. J. Werner, H. Stoll, Phys. Chem. Chem. Phys., 2005, 7, 3917–3923.
[58] J. F. Ogilvie, F. J. H. Wang, J. Mol. Struct., 1992, 273, 277–290.
[59] J. F. Ogilvie, F. J. H. Wang, J. Mol. Struct., 1993, 291, 313–322.
[60] N. Runeberg, P. Pyykö, Int. J. Quantum Chem., 1998, 66, 131–140.
[61] Gaussian DV, Revision A.1, M. J. Frisch, G. W. Trucks, H. B. Schlegel, G. E. Scuseria,
M. A. Robb, J. R. Cheeseman, G. Scalmani, V. Barone, B. Mennucci, G. A. Petersson, H.
Nakatsuji, M. Caricato, X. Li, H. P. Hratchian, A. F. Izmaylov, J. Bloino, G. Zheng, J. L.
Sonnenberg, M. Hada, M. Ehara, K. Toyota, R. Fukuda, J. Hasegawa, M. Ishida, T.
Nakajima, Y. Honda, O. Kitao, H. Nakai, T. Vreven, J. A. Montgomery, Jr., J. E. Peralta, F.
Ogliaro, M. Bearpark, J. J. Heyd, E. Brothers, K. N. Kudin, V. N. Staroverov, R. Kobayashi,
J. Normand, K. Raghavachari, A. Rendell, J. C. Burant, S. S. Iyengar, J. Tomasi, M. Cossi,
N. Rega, J. M. Millam, M. Klene, J. E. Knox, J. B. Cross, V. Bakken, C. Adamo, J. Jaramillo,
R. Gomperts, R. E. Stratmann, O. Yazyev, A. J. Austin, R. 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. Daniels, Ö. Farkas, J. B. Foresman, J. V. Ortiz, J.
Cioslowski, and D. J. Fox, 2009, Gaussian Inc., Wallingford CT.
[62] J. Toulouse, A. Savin, C. Adamo, J. Chem. Phys., 2002, 117, 10465.
[63] V. Tognetti, P. Cortona, C. Adamo, J. Chem. Phys., 2008, 128, 34101.
[64] F. Neese, Wiley Interdisciplinary Reviews: Computational Molecular Science, 2012, 2, 73–78.[65] T. Weigend, R. Ahlrichs, Phys. Chem. Chem. Phys., 2005, 7, 3297-3305.
[66] A. Karton, J. M. L. Martin, J. Chem. Phys., 2011, 135, 144119.
[67] L. A. Curtiss, K. Raghavachari, G. W. Trucks, J. A. Pople, J. Chem. Phys., 1991, 94,
7221–7230.
[68] V. S. Bryantsev, M. S. Diallo, A. C. T. van Duin, W. A. Goddard, J. Chem. Theory
Comput., 2009, 5, 1016–1026.
[69] R. A. Kendall, T. H. Dunning, R. J. Harrison, J. Chem. Phys., 1992, 96, 6796-6806.
[70] S. Grimme, Angew. Chemie Int. Ed., 2006, 45, 4460–4464.
[71] R. Huenerbein, B. Schirmer, J. Moellmann, S. Grimme, Phys. Chem. Chem. Phys., 2010,
12, 6940–6948.
[72] V. Guner, K. S. Khuong, A. G. Leach, P. S. Lee, M. D. Bartberger, K. N. Houk, J. Phys.
23
Chem. A, 2003, 107, 11445–11459.
[73] S. N. Steinmann, G. Csonka, C. Carminboeuf, J. Chem. Theory Comput., 2009, 5, 2950–
2958.
[74] W. J. Hehre, R. Ditchfield, L. Radom, J. A. Pople, J. Am. Chem. Soc., 1970, 92, 4796–
4801.
[75] L. Radom, W. J. Hehre, J. A. Pople, J. Am. Chem. Soc.,1971, 93, 289–300.
[76] M. D. Wodrich, D. F. Jana, von Rague, P. Schleyer, C. Corminboeuf, J. Phys. Chem. A,
2008, 112, 11495–11500.
[77] H. Krieg, S. Grimme, Mol. Phys. 2010, 108, 2655–2666.
[78] D. Feller, E. D. Glendening, D. E. Woon, M. W. Feyereisen, J. Chem. Phys., 1995, 103,
3526.
[79] M. W. Feyereisen, D. Feller, D. A. Dixon, J. Phys. Chem.,1996, 100, 2993.
[80] D. E. Bernholdt, R. J. Harrison, Chem. Phys. Lett., 1996, 250, 477.
[81] http://www.thch.uni-bonn.de/tc/
[82] M. Piacenza, S. Grimme, J. Comput. Chem., 2004, 25, 83– 99.
[83] S. Grimme, C. Mück-Lichtenfeld, Chem. Phys. Chem., 2002, 2, 207–209.
[84] H. L. Woodcock, H. F. Schaefer, P. R. Schreiner, J. Phys. Chem. A, 2002, 106, 11923–
11931.
[85] P. R. Schreiner, A. A. Fokin, R. A. Pascal, A. de Meijere, Org. Lett., 2006, 8, 3635–
3638.
[86] C. Lepetit, H. Chermette, M. Gicquel, J. L. Heully, R. Chauvin, J. Phys. Chem. A, 2007,
111, 136–149.
[87] S. Grimme, C. Mück-Lichtenfeld, E. U. Würthwein, A. W. Ehlers, T. P. M. Goumans,
K. Lammertsma, J. Phys. Chem. A, 2006, 110, 2583–2586.
[88] J. S. Lee, J. Phys. Chem. A, 2005, 109, 11927–11932.
[89] S. Grimme, J. Chem. Phys., 2006, 124, 034108.
[90] A. D. Becke, Phys. Rev. A 1988, 38, 3098-3100.
[91] D. J. Lacks, R. G. Gordon, Phys. Rev. A 1993, 47, 4681.
[92] C. Adamo, V. Barone J. Chem. Phys., 1998, 108, 664-675
[93] F. Kannemann, A. D. Becke, J. Chem. Theory Comput., 2009, 5, 719-727
[94] Y. Zhao, O. Tishchenko, J. R. Gour, W. Li, J. J. Lutz, P. Piecuch, D. Truhlar, J. Phys.
Chem. A, 2009, 113, 5786– 5799.
[95] E. R. Johnson, P. Mori-Sànchez, A. J. Cohen, W. J. Yang, Chem. Phys., 2008, 129,
204112.
24
[96] L. A. Curtiss, K. Raghavachari, P. C. Redfern, J. A. Pople, J. Chem. Phys., 1997, 106,
1063–1079.
[97] A. Halkier, T. Helgaker, P. Jørgensen, W. Klopper, H. Koch, J. Olsen, A. K. Wilson,
Chem. Phys. Lett., 1998, 286, 243–252.
[98] T. M. J. Gilbert, J. Phys. Chem. A., 2004, 108, 2550–2554.
[99] F. Neese, T. Schwabe, S. Kossmann, B. Schirmer, S. Grimme, J. Chem. Theory
Comput., 2009, 5, 3060-3073.
[100] H. Zipse, Top. Curr. Chem., 2006, 263, 163–189.
[101] M. Korth, S. Grimme, J. Chem. Theory Comput., 2009, 5, 993-1003.
[102] S. Parthiban, J. M. L. Martin, J. Chem. Phys., 2001, 114, 6014–6029.
[103] Y. Zhao, D. G. Truhlar, J. Phys. Chem. A, 2006, 110, 10478–10486.
[104] J. M. L. Martin, G. de Oliveira, J. Chem. Phys., 1999, 111, 1843–1856.
[105] Y. Zhao, B. J. Lynch, D. G. Truhlar, J. Phys. Chem. A, 2004, 108, 2715–2719.
[106] Y. Zhao, N. Gonzalez-Garcia, D. G. Truhlar, J. Phys. Chem. A, 2005, 109, 2012.
[107] S. Grimme, M. Steinmetz, M. Korth, J. Org. Chem., 2007, 72, 2118–2126.
25
Table 1: Empirical coefficients of the different models of dispersion for each functional considered in this paper.
DFT-D2 DFT-D3 DFT-D3(BJ)s6 s6 sr ,6 s8 s6 a1 s8 a2
B3LYP 1.05a 1.000 c 1.261 c 1.703 c 1.000 d 0.398 d 1.989 d 4.421 d
B2PLYP 0.55 a 0.640 c 1.427 c 1.022 c 0.640 d 0.307 d 0.915 d 5.057 d
PBE0 0.67 a 1.000 c 1.287 c 0.928 c 1.000 d 0.415 d 1.218 d 4.859 d
PBE0-DH 0.47 b 0.840 1.394 0.748 0.840 0.000 0.095 6.1020.840 e 0.345 e 0.747 e 5.276e
a Reference [49,50], b Reference [51], c Reference [36] ], d Reference [40]. e The model developed using this set of parameters will be referred to as PBE0-DH-D3(BJ)(2).
26
Figure captions
Figure 1: MAD (in kcal/mol) computed with the PBE0-DH functional for the S130 set, when
s8 and sr ,6 are varying, at fixed s6=¿0.84.
Figure 2: MAD (in kcal/mol) computed with the PBE0, PBE0-DH, B3LYP and B2PLYP
functionals over the GMTKN30 set including the dispersion corrections at different level.
Figure 3: MAD (in kcal/mol) computed with the PBE0, PBE0-DH, B3LYP and B2PLYP
functionals for sets concerning decomposition and atomization energies, including dispersion
corrections with different models.
Figure 4: MAD (in kcal/mol) computed with the PBE0, PBE0-DH, B3LYP and B2PLYP
functionals for sets concerning adiabatic processes, including dispersion corrections with
different models.
Figure 5: MAD (in kcal/mol) computed with the PBE0, PBE0-DH, B3LYP and B2PLYP
functionals for sets regarding barriers heights and reactions energies, including dispersion
corrections with different models.
Figure 6: MAD (in kcal/mol) computed with the PBE0, PBE0-DH, B3LYP and B2PLYP
functionals for sets regarding isomers (left) and conformers (right) stability, including
dispersion corrections with different models.
Figure 7: MAD (in kcal/mol) computed with the PBE0, PBE0-DH, B3LYP and B2PLYP
functionals for sets regarding dispersion and non covalent interactions, including dispersion
corrections with different models.
Figure 8: MAD (in kcal/mol) computed with the PBE0, PBE0-DH, B3LYP and B2PLYP
functionals for the sets DC9 (left) and SIE11(right), including dispersion corrections with
different models.
Figure 9: Count of best (up) and worst (low) MAD obtained for PBE0, PBE0-DH, B3LYP
and B2PLYP functionals, including dispersion at different level, over all subsets of the
GMTKN30 database.
27
SIE11
BH76
BSR36ACONF
HEAVY28
MB08-165
RSE34
DARC
BHPERIISO34
G2RC
DC9
G21EAALK6
WATER 27O3ADD6
AL2X
ISOL22
S22
IDISP
HEAVY 28
RG6
Table of contents
The performances of the B3LYP and B2PYLP functionals have been compared to those of PBE0 and PBE0-DH when three different types of dispersion corrections are applied (D2, D3 and D3(BJ)).
37
CYCONF