CSIRO PUBLISHING Rapid Communication

www.publish.csiro.au/journals/ajc Aust. J. Chem. 2004, 57, 365–370

EDF2: A Density Functional for Predicting Molecular

Vibrational Frequencies

Ching Yeh Lin,A Michael W. George,A and Peter M. W. GillA,B

A School of Chemistry, University of Nottingham, Nottingham NG7 2RD, UK.
B Author to whom correspondence should be addressed (e-mail: peter.gill@nott.ac.uk).

The majority of calculations of molecular vibrational spectra are based on the harmonic approximation but are com-
pared (usually after empirical scaling) with experimental anharmonic frequencies. Any agreement that is observed
in such cases must be attributable to fortuitous cancellation of errors and it would certainly be preferable to develop
a more rigorous computational approach. In this paper, we introduce a new density functional model (EDF2) that
is explicitly designed to yield accurate harmonic frequencies, and we present numerical results for a wide variety
of molecules whose experimental harmonic frequencies are known. The EDF2 model is found to be significantly
more accurate than other DFT models and competitive with the computationally expensive CCSD(T) method.

Manuscript received: 2 October 2003.

Final version: 3 February 2004.

Spectroscopy has had an enormous influence on the under- measurement of isotopic variants but, of course, there is
standing of all aspects of chemistry. In particular, by provid- a catch. Although quantum mechanics provides a formally
ing direct information about the curvature of the potential exact theoretical basis for the whole of chemistry, all of the
energy surface (PES) near the minimum in which a molecule computationally tractable quantum chemical models involve
resides, vibrational spectroscopy provides a powerful probe approximations that introduce errors into the resulting predic-
of the structural and bonding characteristics of solids, liquids, tions. Whether such errors are acceptably small depends on
and gases. the model employed and the nature of the chemical question
The infrared or Raman spectrum of a medium-size being asked. The most accurate models usually agree very
molecule contains a wealth of information but, regrettably, well with experiment but, unfortunately, consume impracti-
this treasure trove is normally regarded as a liability, not an cally large amounts of computer time. Less accurate, less
asset. Normally, analysis focuses on a small subset of famil- time-consuming models, such as the various flavours of
iar bands (e.g. carbonyl stretches) and largely neglects the density functional theory (DFT), mirror reality rather less
rest of the information. Why is such a profligate approach faithfully but have nonetheless found widespread use in all
the established norm? Simply because it is difficult to branches of chemistry.
assign the majority of the peaks in a vibrational spectrum The ideal theoretical procedure is to solve the vibrational
unless one performs a sophisticated normal-mode analy- Schrödinger equation directly in a fully coupled, anharmonic
sis requiring the synthesis of several isotopomers. Sadly, basis.[1] This can be done but the approach is limited by its
the time required for such synthetic excursions is usually complexity to small systems. Alternatively, one can first cal-
prohibitive. culate normal-mode (harmonic) frequencies and then apply
There is, however, another way forward. As the speed the necessary corrections for anharmonicity but, whereas the
of desktop computers continues to rise exponentially, it is first of these steps is now routine, current approaches to
becoming increasingly practical to apply quantum mechan- the second step remain computationally demanding.[2] Much
ics to make first-principles predictions of chemical behaviour, more commonly, harmonic calculations are performed and
and it is now possible to calculate the vibrational spectra of the resulting frequencies scaled by empirical scale factors[3]
twenty-atom molecules in a few hours using a standard PC which account (it is hoped) for the neglect of anharmonic-
and an efficient quantum chemistry package. Armed with ity. However, when this works, it obviously gives the right
this combination of hardware and software, it is straightfor- answers for the wrong reasons. Moreover, calculated har-
ward to predict the infrared spectrum of a newly synthesized monic frequencies are often found to deviate significantly
molecule and gain significant insights by comparing the result from experimental harmonic frequencies, in cases where the
to an experimental measurement. latter are known, implying that the theoretical method is
Performing theoretical calculations of vibrational yielding a PES with the wrong curvature (namely, second
spectra is much more cost-effective than the synthesis and derivatives) at the bottom of the well.

© CSIRO 2004 10.1071/CH03263 0004-9425/04/040365

366 C. Y. Lin, M. W. George and P. M. W. Gill

This led us to pursue a two-stage approach. First, build (4) EDF1,[26] an empirical functional that includes the mod-
a DFT model that yields PESs with correct curvatures and, ified Becke exchange functional EX EDF1 and a modified

thus, correct harmonic frequencies. Second, develop compu- Lee–Yang–Parr correlation functional ECEDF1 ,
tationally efficient schemes for the inclusion of anharmonic
and three widely used wavefunction methods:
corrections. The present manuscript describes our progress
on the first stage. (1) HF,[27,28] which uses the Fock exchange functional EX F30 ;

We began by assembling a large set of experimental har- (2) MP2, [29] which uses second-order Møller–Plesset per-
monic frequency data. These are available for a wide variety turbation theory;
of diatomics and a smaller number of larger systems. The (3) CCSD(T),[30] which uses coupled-cluster theory with
diatomic data come primarily from Huber and Herzberg[4] single, double, and perturbative triple substitutions.
but the polyatomic data that we have used are drawn from a
Because coupled-cluster calculations are computationally
variety of sources.[5–19] The complete set of 315 molecules
expensive for large systems, we split the full dataset into 296
is listed in Table 1, and they possess 612 distinct normal
modes. As Fig. 1 indicates, the dataset covers approximately
two-thirds of the elements. H
We then calculated the harmonic vibrational frequencies Li Be B C N O F
of each molecule at a variety of levels of theory. These Na Mg Al Si P S Cl
included four popular DFT methods: K Ca Sc Ti V Cr Mn Fe Ni Cu Zn Ga Ge As Se Br

(1) LSDA,[20,21]a combination of the Dirac exchange func- Rb Sr Y Zr Nb Rh Ag Cd In Sn Sb Te I Xe

tional EX
D30 with the Vosko–Wilk–Nusair correlation Cs Ba La Au Hg Tl Pb Bi
functional ECVWN ; Ra
(2) BLYP,[22,23] a combination of the Becke exchange Ho
functional EX
B88 with the Lee–Yang–Parr correlation
Th
functional ECLYP ;
(3) B3LYP,[24,25] the three-parameter hybrid functional of Fig. 1. Elements in dataset (dark symbols) and the basis sets used
Stephens et al.; (shading; cc-pVTZ light grey, CRENBL dark grey, SRSC white).

Table 1. The complete set of molecules used in this study

DiatomicsA
Triatomics
Ag2 Au2 BeH Cl2 FO InF MnF PbF ScF SrH CO2 E HCNB
AgAl AuAl BeI ClF GaBr InH MnH PbH ScO SrI CS2 D HNCF
AgBi AuBa BeO ClO GaCl InI MnI PbI ScS SrO H2 OB HNOF
AgBr AuBe BeS CN GaF InO MnO PbO SeBr Te2 H2 SB SO2 F
AgCl AuCa BF CO GaH IO MnS PbS SeO TeBr
AgCu AuCl BH CrH GaI K2 N2 PbTe SeS TeI Tetraatomics
AgF AuCu BiBr CrO GaO KBr Na2 PCl SH TeS H2 COB NH3 F
AgGa AuGa BiCl CrS GeBr KCl NaBr PF Si2 TeSe HCCF E PH3 F
AgH AuGe BiH CS GeCl KF NaCl PH SiCl ThO HCCHC SO3 F
AgI AuH BiI Cs2 GeH KH NaF PI SiF TiO
AgIn AuMg BN CsBr GeI KI NaH PN SiH TlBr Pentatomics
AgTe AuSi BO CSe GeO LaO NaI PO SiI TlCl CH2 Cl2 B SiH4 F
Al2 AuSn Br2 CsF GeS LaS NaK PS SiN TlF CH3 ClB CH3 BrB,Q
AlBr AuSr BrCl CsH GeSe Li2 NaRb RaCl SiO TlH CH3 FF CH3 IB,Q
AlCl AuTe BrF CsI GeTe LiBr NbO Rb2 SiS VO CH4 F HCOOHN,Q
AlF B2 BS Cu2 H2 LiCl NBr RbBr SiSe XeCl
AlH BaBr C2 CuBr HBr LiF NCl RbCl SiTe YCl Larger moleculesQ
AlI BaCl CaBr CuCl HCl LiH NF RbH SnBr YF CH2 CH2 B SF6 N
AlO BaF CaCl CuF HF LiI NH RbI SnCl YO CH3 CH3 B C3 H8 B
AlS BaH CaF CuGa HgCl LiNa NiH RhC SnF ZnBr C6 H6 G C6 D6 H
AlSe BaI CaH CuH HgI MgBr NO S2 SnI ZnCl 13 C H I 13 C D I
6 6 6 6
As2 BaO CaI CuI HI MgCl NS Sb2 SnO ZnF n-butaneO 3-hexyneP
AsCl BaS CaS CuO HoF MgF NSe SbBi SnS ZnH n-pentaneP
AsF BBr CCl CuSe I2 MgH O2 SbBr SnTe ZnI
AsN BCl CdCl CuTe IBr MgI OH SbCl SO ZrO Metal carbonylsQ
AsO BeBr CdI F2 ICl MgO P2 SbF SrBr Ni(CO)2 J Fe(CO)5 K
AsP BeCl CF FeCl IF MgS PbBr SbP SrCl Ni(CO)3 J Cr(CO)6 L
AsS BeF CH FeF InCl MnBr PbCl ScCl SrF Ni(CO)4 J
A Ref. [4]. B Ref. [5]. C Ref. [6]. D Ref. [7]. E Ref. [8]. F Ref. [9]. G Ref. [10]. H Ref. [11]. I Ref. [12]. J Ref. [13]. K Ref. [14].
L Ref. [15]. M Ref. [16]. N Ref. [17]. O Ref. [18]. P Ref. [19]. Q CCSD(T) calculations not performed.
A Density Functional for Vibrational Frequencies 367

‘small’ and 19 ‘large’ molecules (see footnote Q of Table 1) Table 2. Errors [cm−1 ] in calculated harmonic frequencies for
and did not attempt CCSD(T) calculations on any of the some small molecules
‘large’ molecules. Molecule Expt BLYP EDF1 B3LYP EDF2 HF MP2 CCSD(T)
All of the DFT, HF, and MP2 geometry optimization H2 4401 −54 −26 19 2 186 125 8
and frequency calculations were performed using Q-Chem HF 4138 −206 −120 −48 −33 344 57 39
−69 −18 −2
2.0.[31] Coupled-cluster calculations were carried out using CO2 2397
1354 −48 −25
19
18
42
26
167 42
157 −11 −8
Molpro.[32] The Dunning cc-pVTZ basis set[33] was used for 673 −40 −30 −2 −1 100 −12 −13
light atoms, and CRENBL[34] and SRSC[35] pseudopotential H2 O 3943 −188 −107 −42 −24 284 50 2
3832 −176 −98 −31 −15 295 40 8
basis sets for heavy atoms (Fig. 1). In the DFT calcula- 1649 −38 −25 −10 −18 104 2 19
tions, we employed an Euler–Maclaurin–Lebedev quadrature SO2 1381 −135 −91 −30 −20 195 −45 −28
scheme[36–38] to integrate the exchange–correlation func- 1167 −98 −60 −2 6 212 −47 −19
526 −45 −33 −9 −8 82 −22 −13
tional and used 100 radial and 194 angular grid points on NH3 3577 −118 −53 1 20 230 99 20
each atom to ensure that low-frequency modes are treated 3506 −158 −102 −44 −30 179 31 −35
−46 −38 −13 −20 −3
satisfactorily.[39] Every frequency calculation was preceded 1691
1022 46 54 46 31
104
105
1
55 87
by an optimization—it is not physically meaningful to discuss SO3 1409 −131 −80 −34 −19 168 6 −14
vibrations about non-stationary points on a PES. 1048 −82 −42 3 13 193 6 9
539 −60 −48 −26 −24 60 −21 −21
Following the construction philosophy of the EDF1 504 −61 −45 −24 −21 79 −14 −17
functional[26] (the acronym stands for ‘Empirical Density CH2 Cl2 3182 −58 −30 17 18 182 56 24
Functional 1’), we sought to develop a new DFT functional 3123 −75 −47 −1 1 162 54 7
1464 −46 −38 −2 −7 131 31 10
by linearly combining several existing functionals. However, 1295 −53 −40 −8 −12 119 14 −2
whereas EDF1 was optimized to yield accurate thermochem- 1177 −51 −33 −7 −8 106 26 3
917 −46 −36 −17 −17 64 2 −11
istry when used with the 6–31+G* basis set, our new target 771 −109 −72 −53 −41 55 33 3
(EDF2) was optimized to give accurate harmonic frequencies 724 −70 −34 −24 −15 40 22 −4
when used with the basis sets indicated in Fig. 1. 284 −19 −12 −6 −4 20 7 −2
CH3 Cl 3166 −74 −40 1 7 149 62 15
To begin, we wrote an initial approximation to the new 3074 −73 −45 −2 2 143 60 6
functional as Equation (1) 1482 −36 −32 1 −5 120 34 12
1383 −49 −38 −6 −10 119 22 1
1038 −48 −35 −14 −15 77 10 −6
EEDF2 = a1 EHF + a2 ELSDA + a3 EBLYP 740 −74 −34 −31 −17 29 37 2
+ a4 EB3LYP + a5 EEDF1 (1) CH4 3158 −97 −48 −24 −11 88 60 −5
3026 −63 −32 3 9 122 −17 8
1567 −36 −32 −3 −9 99 33 4
and determined the five unknown ai coefficients by a simple 1357 −43 −46 −13 −21 97 −174 −14
least-squares fit of the HF, LSDA, BLYP, B3LYP, and EDF1
harmonic frequencies to the experimental ones. This strat- Table 3. RMS errors [cm−1 ] of the harmonic frequencies of
egy was termed ‘external optimization’ by Adamson et al.[26] small and large molecules
Then, since each of the five functionals is itself a sum of sim-
Molecules BLYP EDF1 B3LYP EDF2 HF MP2 CCSD(T)
pler components, we could write the exchange–correlation
part of the new functional as Equation (2) ‘Small’ 65 53 40 34 108 67 35
‘Large’ 67 54 34 33 133 55 –
EXC
EDF2
= b1 EX
F30
+ b 2 EX
D30
+ b 3 EX
B88
+ b 4 EX
EDF1 All 66 54 38 34 117 64 35

+ b5 ECVWN + b6 ECLYP + b7 ECEDF1 (2) Table 4. RMS errors [%] of the harmonic frequencies of small
and large molecules
Since external optimization leads to a linear combination of
model chemistries, not a linear combination of functionals, Molecules BLYP EDF1 B3LYP EDF2 HF MP2 CCSD(T)
the bi coefficients at this stage were not optimal. There- ‘Small’ 7.0 5.8 4.5 3.9 9.8 6.4 4.1
fore, in a second step, we performed an internal optimization ‘Large’ 4.7 4.4 2.4 2.3 9.0 3.6 –
of the coefficients using the downhill simplex minimiza- All 6.3 5.4 3.7 3.3 9.5 5.6 4.1
tion algorithm.[40] On each iteration of this (each time the
coefficients were refined), we re-optimized the geometries errors in the harmonic frequencies of ten familiar molecules,
of all of the molecules in Table 1, re-computed their har- as calculated by four DFT methods and three wavefunction
monic frequencies at these geometries, and determined the methods. The HF method is clearly the least accurate, fol-
resulting root-mean-square (RMS) error between these and lowed by BLYP and then EDF1. The most accurate methods
the experimental values. After a few months of this, the opti- are B3LYP, EDF2, and CCSD(T). Our results reveal that the
mal coefficients were found to be b1 = 0.1695, b2 = 0.2811, poor accuracy of B3LYP observed in previous studies[41,42]
b3 = 0.6227, b4 = −0.0551, b5 = 0.3029, b6 = 0.5998, and for the H2 and HF molecules is largely an artefact of the
b7 = −0.0053. neglect of anharmonicity: the harmonic frequencies from
There are several different ways in which the accuracies B3LYP agree well with the harmonic experimental results.
of different functionals can be compared and Tables 2–4 and Moreover, the new EDF2 functional is even better in this
Fig. 2 provide different perspectives on this. Table 2 shows the respect.
368 C. Y. Lin, M. W. George and P. M. W. Gill

Table 5. Errors [cm−1 ] in harmonic frequencies using EDF2 with

HF
100 various basis sets

50 Molecule Expt 6–31+G* 6–311G** cc-pVDZ cc-pVTZ cc-pVQZA

0 B2 1051 −31 −33 −35 −31 −32

MP2 BF 1402 −28 −11 −50 19 15
100 BH 2367 −5 −41 −48 −24 −21
BN 1515 61 70 47 65 68
50 BO 1886 35 50 22 43 46
C2 1855 11 36 37 40 38
0
CF 1308 −18 −34 −1 14 4
CCSD(T)
100 CH 2859 −22 −52 −87 −38 −31
CN 2069 97 98 93 100 100
50 CO 2170 43 61 41 53 54
F2 917 133 84 105 149 143
0 FO 1029 94 92 88 108 103
BLYP
100 H2 4401 34 2 −45 2 −2
HF 4138 −187 −5 −97 −33 −36
50 Li2 351 −11 −9 −10 −10 −8
LiF 910 −15 58 84 13 15
0 LiH 1406 −10 5 −20 9 7
B3LYP N2 2359 105 98 104 101 98
100
NF 1141 18 34 39 33 27
50 NH 3282 −21 −42 −102 −27 −21
NO 1904 90 98 103 87 84
0 O2 1580 56 76 85 64 70
EDF1 OH 3738 −84 −23 −100 −29 −26
100
MAE 53 48 63 48 46
50
A All g-functions were removed.
0
EDF2
100 confirms that the EDF2 functional represents a substantial
50 improvement over B3LYP across the board. For the inor-
0
ganic molecules, EDF2 is slightly more accurate than even

150
100
50 0 50 100 150 CCSD(T). Behind these, in order of decreasing accuracy, lie
v (cm
1) EDF1 and MP2 (which appear broadly comparable), BLYP,
and HF.
Fig. 2. Distributions of error ω = ωcalc − ωexp [cm−1 ] in calculated
harmonic frequencies. Does EDF2 yield equally satisfactory harmonic frequen-
cies when combined with other basis sets? To test this,
we calculated the EDF2 frequencies of a selection of first-
The histograms in Fig. 2 show the distribution of errors row diatomics using the 6–31+G*, 6–311G**, cc-pVDZ,
for the seven methods and normal modes of the full set of cc-pVTZ, and cc-pVQZ basis sets and these are shown in
molecules. (As noted above, CCSD(T) was applied only to Table 5. We see that, though individual frequencies vary con-
the small molecules.) It is well known that HF usually over- siderably as the basis is changed, the mean absolute errors
estimates harmonic frequencies, but the tendency of BLYP (MAE) are surprisingly constant. The double-zeta Dunning
and EDF1 to underestimate is less well documented. Bearing basis (cc-pVDZ) gives the largest mean error (63 cm−1 )
in mind that anharmonic corrections are often negative, this but the errors of the four other bases all lie between 46
helps to explain the curious observation that unscaled BLYP and 53 cm−1 . The most difficult molecule, across the board,
and EDF1 harmonic frequencies often agree well with exper- is F2 . Judging by this small set of molecules, the EDF2/
imental anharmonic frequencies.[3,43,44] The most accurate 6–31+G* model seems to offer a promising route to fairly
predictions come from B3LYP, CCSD(T), and EDF2, for accurate harmonic vibrational frequencies at a modest com-
which errors greater than 100 cm−1 are rare. putational cost.
Table 3 shows the RMS errors of the harmonic frequen- The EDF2 functional was explicitly constructed to yield
cies predicted by the seven methods. Since CCSD(T) was not PES curvatures, and it is obviously important to enquire
applied to the ‘large’ molecules, the results for the ‘small’ whether our demand for accuracy in the associated second
molecules, the ‘large’ molecules, and the full dataset are derivatives has entailed serious sacrifices in the quality of the
presented separately. For the large molecules, which are pre- first derivatives (which affect structure) and/or the energy
dominantly organic, the new EDF2 functional is comparable itself (which affects thermochemistry). We have addressed
to B3LYP. For the small molecules, most of which are inor- both of these questions.
ganic, EDF2 is superior to B3LYP and, remarkably, appears To assess the accuracy of EDF2 for structural prediction,
comparable to the much more expensive CCSD(T) method. we have examined the diatomics in our set for which (a) the
Because the harmonic frequencies span two orders of cc-pVTZ basis set is available and (b) the bond length is
magnitude (from 26 to 4401 cm−1 ), it is more appropriate known experimentally. The results are shown numerically in
to discuss relative, rather than absolute, errors. Table 4 gives Table 6 and graphically in Fig. 3 and reveal that, for bond
the RMS percentage errors of the predicted frequencies and length predictions, EDF2 is more accurate than any of the
A Density Functional for Vibrational Frequencies 369

Table 6. Errors [pm] in calculated bond lengths using the cc-pVTZ 15 HF

basis set
10
Molecule Expt BLYP EDF1 B3LYP EDF2 HF MP2 CCSD(T)
5
AlF 165.4 4.6 4.1 2.8 2.2 0.0 2.4 2.1 0
AlH 164.8 3.1 2.9 1.8 1.6 0.5 0.1 0.7 15 MP2
AlO 161.8 3.0 2.1 1.7 0.9 7.5 1.0 2.0 10
BeCl 179.7 1.9 1.5 1.0 0.4 1.7 0.2 0.3
BeF 136.1 1.7 1.6 0.6 0.1 −0.5 0.9 0.6 5
BeH 134.3 0.5 1.0 0.0 −0.1 0.1 −1.0 −0.4 0
BeO 133.1 0.9 0.6 −0.8 −1.1 −3.7 1.8 0.9 15 CCSD(T)
BeS 174.2 1.5 0.9 0.0 −0.5 −1.3 0.7 0.8 10
B2 159.0 2.8 2.7 2.2 1.6 4.8 −0.2 −0.3
BCl 171.6 2.4 2.0 0.9 0.2 0.7 −0.7 0.5 5
BF 126.3 1.4 1.3 0.2 −0.3 −1.4 0.0 0.3 0
BH 123.2 0.9 1.4 0.1 0.1 −1.0 −1.3 −0.6 15 BLYP
BN 128.1 5.4 5.2 3.8 3.4 0.9 3.0 4.8 10
BO 120.5 1.1 0.9 −0.1 −0.5 −2.2 0.3 0.4
C2 124.3 1.3 1.0 0.5 0.1 11.3 14.3 0.3 5
CaF 196.7 −1.1 −2.4 −1.5 −2.8 2.2 0.1 2.0 0
CaH 200.3 −2.3 −3.6 −2.2 −3.3 0.2 0.0 2.3 15 B3LYP
CCl 164.5 4.1 2.5 2.2 1.3 2.2 −0.3 1.4 10
CF 127.2 2.2 1.5 0.4 −0.2 −1.8 −0.3 0.2
CH 112.0 1.4 1.4 0.4 0.3 −1.4 −1.5 −0.6 5
Cl2 198.8 6.6 3.1 3.5 2.4 −0.4 0.5 2.5 0
ClF 162.8 5.2 2.8 1.9 1.0 −3.8 0.6 1.4 15 EDF1
ClO 157.0 4.9 2.5 2.4 1.3 2.5 −0.2 2.3 10
CN 117.2 0.3 0.1 −0.9 −1.2 −2.1 −4.7 6.3
CO 112.8 1.0 0.7 −0.2 −0.5 −2.4 0.7 0.4 5
CS 153.5 1.9 1.3 0.2 −0.2 −1.9 0.4 1.0 0
CSe 167.6 2.4 1.4 0.4 −0.2 −2.1 0.0 0.7 15 EDF2
F2 141.2 2.0 0.0 −1.5 −2.3 −8.3 −1.6 0.2 10
FO 132.6 5.4 3.4 2.4 1.6 −1.3 0.3 2.7
GaCl 220.2 6.3 4.2 4.0 2.6 3.1 0.0 −1.4 5
GaF 177.4 4.4 3.3 2.2 1.3 −0.8 0.0 −0.3 0
GaH 166.3 4.0 3.2 2.3 1.8 1.0 0.0 −1.6
10
5 0 5 10 15 20
GaO 174.4 −1.7 −3.4 −2.4 −3.8 4.3 0.0 −3.6 r (pm)
GeH 158.8 1.9 1.9 1.9 1.3 1.9 1.9 −0.7
GeO 162.5 2.8 1.7 0.5 0.0 −3.4 0.0 0.7 Fig. 3. Distributions of error r = rcalc − rexp [pm] in calculated bond
GeS 201.2 4.2 2.5 1.8 1.0 −1.3 0.0 1.0 lengths.
H2 74.1 0.5 0.4 0.1 0.3 −0.7 −0.4 0.1
HBr 141.4 2.2 1.3 1.1 0.8 −0.6 0.0 −0.5
HCl 127.5 1.9 1.1 0.9 0.7 −0.7 −0.3 0.1
other methods tested, eclipsing both B3LYP and CCSD(T).
HF 91.7 1.6 0.9 0.5 0.4 −1.9 0.0 −0.1 Its worst failure (6.9 pm) occurs for the weakly bound LiNa
Li2 267.3 3.3 6.4 2.5 2.5 11.0 4.5 −0.6 dimer but its mean absolute error for the 69 molecules tested
LiCl 202.1 1.2 1.0 0.3 −0.4 1.7 0.8 0.7
is only 1.1 pm.
LiF 156.4 1.7 1.8 0.6 0.1 0.0 1.6 1.2
LiH 159.6 0.1 0.6 −0.5 −0.6 1.2 −0.4 0.0 To assess the accuracy of EDF2 for thermochemical pre-
LiNa 281 8.4 13.5 7.4 6.9 18.8 11.0 8.6 diction, we have applied it to the G2 dataset of Pople et al.[45]
MgF 175.0 3.8 3.4 2.0 1.4 −0.2 1.7 1.3 which includes 56 atomization energies, 40 ionization ener-
MgH 173.0 2.8 2.6 1.4 1.0 0.7 −0.7 0.3
MgO 174.9 1.1 0.4 −0.6 −1.3 −2.1 −1.3 0.5 gies, 25 electron affinities, and 8 proton affinities. RMS errors
N2 109.8 0.6 0.3 −0.6 −0.9 −3.1 1.2 0.3 for the four DFT methods are compared with HF, MP2, and
NaF 192.6 3.0 3.3 1.3 0.5 0.4 2.3 1.8 CCSD(T) in Tables 7 and 8.
NaH 188.7 0.6 1.8 −0.2 −0.7 3.2 1.9 2.8
NF 131.7 2.6 1.1 0.1 −0.6 −2.5 −0.9 −0.1
Table 7 shows thermochemical results obtained using the
NH 103.6 1.6 1.2 0.5 0.4 −1.5 −0.8 0.0 6–31+G* basis set, the set for which the original EDF1 func-
NO 115.1 1.2 0.6 −0.5 −0.8 −3.5 −1.5 0.2 tional was designed to be optimal. It is not surprising to
NS 149.4 2.7 1.8 0.7 0.2 4.2 −7.4 1.4 find that, when this modest basis is used, EDF1 is the most
O2 120.8 2.4 1.1 −0.2 −0.7 −4.9 1.3 0.1
OH 97.0 1.6 1.0 0.5 0.4 −1.9 −0.4 0.0 accurate of the functionals tested. It is surprising, however,
P2 189.3 2.5 1.3 0.2 −0.3 −3.6 2.8 1.9 to discover that CCSD(T) is no better than MP2 (because of
PF 159.0 4.5 3.3 2.2 1.5 −1.3 1.2 1.3 the small basis set) and EDF2 is comparable to B3LYP.
PH 142.2 2.0 1.5 0.9 0.7 −0.8 −0.5 0.3
PN 149.1 1.5 1.0 −0.3 −0.6 −3.8 3.3 1.2
Table 8 shows thermochemical results obtained using
PO 147.6 3.1 2.4 1.1 0.7 −2.9 1.9 1.6 the aug-cc-pVTZ basis set, which is much larger than 6–
SeO 164.8 3.0 1.3 0.0 −0.6 −5.5 0.0 1.2 31+G*. Overall, we find that EDF2 is the most accurate of
SH 134.1 1.9 1.3 0.9 0.7 −0.8 −0.5 0.1
the four DFT methods tested, bettering even the venerable
SiF 160.1 4.7 3.9 2.7 2.1 −0.5 1.9 1.7
SiH 152.0 2.5 2.2 1.3 1.1 −0.2 1.7 0.6 B3LYP method in all tests except proton affinities. Of course,
SiN 157.2 1.7 1.0 0.1 −0.3 1.4 −4.3 8.3 none of the DFT methods can approach the thermochemical
SiO 151.0 2.8 2.2 0.9 0.5 −2.5 2.2 1.4 accuracy of CCSD(T) with a large basis set, a combina-
SO 148.1 4.4 3.2 1.9 1.4 −3.0 1.7 1.7
tion that Dunning has called ‘the gold standard of quantum
MAE 2.6 2.1 1.3 1.1 2.6 1.5 1.3
chemistry’.
370 C. Y. Lin, M. W. George and P. M. W. Gill

[11] N. C. Handy, P. E. Maslen, R. D. Amos, J. S. Andrews,

Table 7. RMS errors [kcal mol−1 ] of the thermochemical properties C. W. Murray, G. J. Laming, Chem. Phys. Lett. 1992, 197, 506.
of the molecules in the G2 dataset using the 6–31+G∗ basis set doi:10.1016/0009-2614(92)85808-N
[12] L. Goodman, A. G. Ozkabak, S. N. Thakur, J. Phys. Chem.
BLYP EDF1 B3LYP EDF2 HF MP2 CCSD(T) 1991, 95, 9044.
[13] L. Manceron, M. E. Alikhani, Chem. Phys. 1999, 244, 215.
Atomization energies 5.75 4.41 8.10 6.17 91.25 27.60 28.89
doi:10.1016/S0301-0104(99)00110-X
Ionization potentials 5.14 4.34 5.05 4.77 24.65 9.60 9.71
Electron affinities 4.39 3.79 4.40 4.38 29.64 11.47 11.44 [14] R. S. Weiner, Ph.D. Thesis 1970 (Purdue University: West
Proton affinities 5.23 3.77 5.51 6.38 7.20 6.02 5.16 Lafayette, IN).
[15] L. Jones, R. S. McDowell, M. Goldblatt, Inorg. Chem. 1969,
8, 2349.
[16] M. Freytes, D. Hurtmans, S. Kassi, J. Liévin, J. Vander Auwera,
Table 8. RMS errors [kcal mol−1 ] of the thermochemical properties
A. Campargue, M. Herman, Chem. Phys. 2002, 283, 47.
of the molecules in the G2 dataset using the aug-cc-pVTZ basis setA
doi:10.1016/S0301-0104(02)00507-4
BLYP EDF1 B3LYP EDF2 HF MP2 CCSD(T) [17] R. S. McDowell, J. P. Aldridge, R. F. Holland, J. Phys. Chem.
1976, 80, 1203.
Atomization energies 7.67 7.50 6.12 5.94 86.34 6.33 3.84
[18] R. M. Levy, O. d. l. L. Rojas, R. A. Friesner, J. Phys. Chem.
Ionization potentials 5.14 4.47 5.28 4.96 25.41 5.07 2.35
1984, 88, 4233.
Electron affinities 3.81 2.80 3.80 2.97 29.70 4.98 2.21
Proton affinities 2.50 1.63 1.91 2.03 3.96 1.25 1.20 [19] B. R. Henry, D. M. Turnbull, D. P. Schofield, H. G. Kjaergaard,
J. Phys. Chem. A 2003, 107, 3236. doi:10.1021/JP021650N
A Except for Li, Be, Na, and Mg for which cc-pVTZ was used. [20] P. A. M. Dirac, Proc. Camb. Phil. Soc. 1930, 26, 376.
[21] S. J. Vosko, L. Wilk, M. Nusair, Can. J. Phys. 1980, 58, 1200.
[22] A. D. Becke, Phys. Rev. A. 1988, 38, 3098. doi:10.1103/
The EDF2 functional was constructed to yield accurate PHYSREVA.38.3098
harmonic vibrational frequencies but it appears also to afford [23] C. Lee, W. Yang, R. G. Parr, Phys. Rev. B 1988, 37, 785.
accurate structures and thermochemistry. When used with a doi:10.1103/PHYSREVB.37.785
large basis set it yields harmonic frequencies that are compa- [24] A. D. Becke, J. Chem. Phys. 1993, 98, 5648. doi:10.1063/
rable to those of CCSD(T) and thermochemical predictions 1.464913
[25] P. J. Stephens, F. J. Devlin, C. F. Chabalowski, M. J. Frisch,
that are generally superior to those of B3LYP, at least for the J. Phys. Chem. 1994, 98, 11623.
G2 molecular dataset. We are encouraged by these results [26] R. D. Adamson, P. M. W. Gill, J. A. Pople, Chem. Phys. Lett.
and look forward to applying EDF2 to a range of chemical 1998, 284, 6. doi:10.1016/S0009-2614(97)01282-7
problems. [27] D. R. Hartree, Proc. Camb. Phil. Soc. 1928, 24, 89.
[28] V. Fock, Z. Physik 1930, 61, 126.
[29] C. Møller, M. S. Plesset, Phys. Rev. 1934, 46, 618. doi:10.1103/
Acknowledgements PHYSREV.46.618
We thank the referees for several very helpful comments. [30] J. Cizek, J. Paldus, Phys. Scripta 1980, 21, 251.
[31] J. Kong, C. A. White, A. I. Krylov, D. Sherrill, R. D. Adamson,
C.Y.L. gratefully acknowledges an ORS Award from the T. R. Furlani, M. S. Lee, A. M. Lee, et al., J. Comput. Chem. 2000,
DfES. M.W.G. thanks the Royal Society of Chemistry for the 21, 1532. doi:10.1002/1096-987X(200012)21:16<1532::AID-
support of the Sir Edward Frankland Fellowship. P.M.W.G. JCC10>3.0.CO;2-W
thanks the organizers for the opportunity to present this [32] H.-J. Werner, P. J. Knowles, Users’ Manual for MOLPRO, ver.
work at the 19th RACI Organic Conference in Lorne, Aus- 2000.1 2000 (University of Birmingham: Birmingham).
[33] D. E. Woon, T. H. Dunning, J. Chem. Phys. 1993, 98, 1358.
tralia. This research was supported by a JREI computer doi:10.1063/1.464303
hardware grant (GR/R62052) from the Engineering and [34] M. M. Hurley, L. Fernandez Pacios, P. A. Christiansen,
Physical Sciences Research Council. R. B. Ross, W. C. Ermler, J. Chem. Phys. 1986, 84, 6840.
doi:10.1063/1.450689
References [35] M. Dolg, U. Wedig, H. Stoll, H. Preuss, J. Chem. Phys. 1987,
86, 866. doi:10.1063/1.452288
[1] S. Carter, J. M. Bowman, J. Chem. Phys. 1998, 108, 4397. [36] V. I. Lebedev, Sibirsk. Mat. Zh. 1977, 18, 132.
doi:10.1063/1.475852 [37] C. W. Murray, N. C. Handy, G. J. Laming, Mol. Phys. 1993,
[2] J. M. Bowman, Acc. Chem. Res. 1986, 19, 202. 78, 997.
[3] A. P. Scott, L. Radom, J. Phys. Chem. 1996, 100, 16502. [38] P. M. W. Gill, B. G. Johnson, J. A. Pople, Chem. Phys. Lett.
doi:10.1021/JP960976R 1993, 209, 506. doi:10.1016/0009-2614(93)80125-9
[4] K. P. Huber, G. Herzberg, Molecular Spectra and Molecular [39] J. M. L. Martin, C. W. Bauschlicher, A. Ricca, Comput.
Structure 1979 (Van Nostrand Reinhold: New York, NY). Phys. Commun. 2001, 133, 189. doi:10.1016/S0010-4655(00)
[5] K. J. Miller, F. S. Ganda-Kesuma, J. Mol. Spec. 1991, 145, 429. 00174-0
doi:10.1016/0022-2852(91)90130-3 [40] J. A. Nelder, R. Mead, Comput. J. 1965, 7, 308.
[6] B. G. Johnson, P. M. W. Gill, J. A. Pople, J. Chem. Phys. 1993, [41] M. W. Wong, Chem. Phys. Lett. 1996, 256, 391. doi:10.1016/
98, 5612. doi:10.1063/1.464906 0009-2614(96)00483-6
[7] N. J. Harris, J. Phys. Chem. 1995, 99, 14689. [42] I. Bytheway, M. W. Wong, Chem. Phys. Lett. 1998, 282, 219.
[8] J. R. Thomas, B. J. DeLeeuw, G. Vacek, T. D. Crawford, doi:10.1016/S0009-2614(97)01281-5
Y. Yamaguchi, H. F. Schaefer III, J. Chem. Phys. 1993, 99, 403. [43] B. G. Johnson, P. M. W. Gill, J. A. Pople, J. Chem. Phys. 1993,
doi:10.1063/1.465764 98, 5612. doi:10.1063/1.464906
[9] W. J. Hehre, L. Radom, P. von R. Schleyer, J. A. Pople, Ab [44] T. M. Watson, J. D. Hirst, J. Phys. Chem. A 2002, 106, 7858.
Initio Molecular Orbital Theory 1986 (Wiley: New York, NY). doi:10.1021/JP025551L
[10] A. Miani, E. Cane, P. Palmieri, A. Trombetti, N. C. Handy, [45] L. A. Curtiss, K. Raghavachari, G. W. Trucks, J. A. Pople,
J. Chem. Phys. 2000, 112, 248. doi:10.1063/1.480577 J. Chem. Phys. 1991, 94, 7221. doi:10.1063/1.460205