Haoyu S.
Yu
a,
Wenjing
Zhang
ab,
Pragya
Verma
a,
Xiao
He
ac and
Donald G.
Truhlar
*a
aDepartment of Chemistry, Chemical Theory Center, Inorganometallic Catalyst Design Center, and Supercomputing Institute, University of Minnesota, Minneapolis, Minnesota 55455-0431, USA. E-mail: truhlar@umn.edu
bThe College of Chemistry and Molecular Engineering, Zhengzhou University, Zhengzhou, Henan 450001, China
cState Key Laboratory of Precision Spectroscopy, East China Normal University, Shanghai, China
First published on 9th April 2015
The goal of this work is to develop a gradient approximation to the exchange–correlation functional of Kohn–Sham density functional theory for treating molecular problems with a special emphasis on the prediction of quantities important for homogeneous catalysis and other molecular energetics. Our training and validation of exchange–correlation functionals is organized in terms of databases and subdatabases. The key properties required for homogeneous catalysis are main group bond energies (database MGBE137), transition metal bond energies (database TMBE32), reaction barrier heights (database BH76), and molecular structures (database MS10). We also consider 26 other databases, most of which are subdatabases of a newly extended broad database called Database 2015, which is presented in the present article and in its ESI. Based on the mathematical form of a nonseparable gradient approximation (NGA), as first employed in the N12 functional, we design a new functional by using Database 2015 and by adding smoothness constraints to the optimization of the functional. The resulting functional is called the gradient approximation for molecules, or GAM. The GAM functional gives better results for MGBE137, TMBE32, and BH76 than any available generalized gradient approximation (GGA) or than N12. The GAM functional also gives reasonable results for MS10 with an MUE of 0.018 Å. The GAM functional provides good results both within the training sets and outside the training sets. The convergence tests and the smooth curves of exchange–correlation enhancement factor as a function of the reduced density gradient show that the GAM functional is a smooth functional that should not lead to extra expense or instability in optimizations. NGAs, like GGAs, have the advantage over meta-GGAs and hybrid GGAs of respectively smaller grid-size requirements for integrations and lower costs for extended systems. These computational advantages combined with the relatively high accuracy for all the key properties needed for molecular catalysis make the GAM functional very promising for future applications.
Even though the meta and nonlocal functionals can give more accurate results than GGAs and LSDAs, GAs are still of great interest for four reasons. First, GAs are widely implemented in many programs because of their ease of coding. Second, GAs often have better self-consistent field (SCF) convergence and smaller grid requirements than meta functionals. Third, calculations employing GAs are less expensive than calculations involving nonlocal functionals, with the difference being more pronounced for extended and large systems and when geometries are optimized.
The fourth reason for special interest in GAs is that local functionals often have better performance than hybrid functionals, on average, for systems with high multi-reference character. Multi-reference character is the extent to which a wave function is inherently multi-configurational so that a single Slater determinant does not provide a good starting point (reference function) for approximating the complete wave function. Although KS theory does not calculate the wave function of the interacting system, it does use a Slater determinant to represent the density, and calculating the exchange from the Slater determinant, as in Hartree–Fock exchange, can introduce static correlation error, a result of which is that it is often more challenging to obtain good approximations for multi-reference systems when Hartree–Fock exchange is included. (The unknown exact exchange–correlation energy functional includes nonlocal effects and does not have static correlation error, but the problem just mentioned is not completely solved by currently available functionals.) Multi-reference systems are sometimes called strongly correlated systems. Many open-shell systems and transition-metal systems have multi-reference character, and hence the ability to treat multi-reference systems is critical to the ability to treat many catalytic reaction mechanisms. Systems without high multi-reference character are called single-reference systems.
Most GAs have a form that separately approximates exchange and correlation, as first introduced by Langreth and Mehl2 and usually called a generalized gradient approximation3 (GGA); however, it has been shown that a nonseparable gradient approximation4 (which has more flexibility at the cost of satisfying less exact constraints) is capable of performing well for a broader set of properties. The original NGA, called N12, was designed to give good predictions both of solid-state lattice constants and of cohesive energies and molecular atomization energies; it also gives good predictions of molecular bond lengths.4 Here we show that we can get improved performance for barrier heights (which are important for studies of both uncatalyzed and catalyzed reactions) by relaxing the accuracy for lattice constants, which are not needed for molecular (as opposed to solid-state) processes. By diminishing the emphasis on obtaining good lattice constants we can obtain an exchange–correlation functional that may be more useful for treating many large and complex homogeneous and enzymatic catalysts that do not require the calculations on solid-state material.
A second goal of this work is to obtain improved results for compounds containing metal atoms, including transition metal compounds with high multi-reference character, by incorporating a greater amount of representative data for metal–ligand bond energies in the training set of a density functional. A third goal of the present work is to obtain a very smooth exchange–correlation functional by enforcing an unsmoothness penalty as part of the optimization process.
Combining these three goals, we have designed a new exchange–correlation functional called gradient approximation for molecules, or GAM, and this new functional is presented here. The GAM functional is an NGA, and so it depends only on spin densities and spin density gradients. The parameters of the GAM functional are optimized against a broad set of molecular and solid-state data in a new database called Database 2015, which is also presented here. We will show that the resulting GAM functional yields good results for main group bond energies, chemical reaction barrier heights, transition-metal bond energies, weak interaction energies between noble gas atoms, and bond lengths of diatomic molecules.
Section 2 describes the computational details. Section 3 describes Database 2015, for which complete information is given in the ESI.† Section 4 describes previously available functionals to be used for comparison. Section 5 describes the design and optimization of the GAM functional. Section 6 gives results; Section 7 provides discussion; and Section 8 summarizes the main conclusions.
Besides testing the new functional on the training subset of Database 2015, we made several tests outside the training set. First we tested the new functional against subdatabases for semiconductor band gaps (SBG31) and solid-state cohesive energies (SSCE8), which are in Database 2015 but outside the training set. We also tested our functional against other data that is not in the training set. This data includes a recently published database WCCR for transition metal coordination reactions56 (renamed here as WCCR10 for consistency with our general naming scheme), the enthalpies of binding of O2 and N2 to the metal–organic framework Fe2(dobdc), the binding of C2H4 to Pd(PH3)2, Ag2 dimer and FeC bond dissociation energies, transition metal dimer bond distances, and the Ar2 potential energy curve.
For the WCCR10 database we use the same basis set (def2-QZVPP) and geometries as used in the original paper; these geometries, which were optimized by functional BP86,33,34 are provided in the ESI† of the WCCR paper.56
For calculating the binding enthalpies of O2 and N2 bound to Fe2(dobdc), we used an 88-atom cluster model of the experimental structure of Fe2(dobdc) containing three iron centers. The details of this cluster and rationale for its design are described in our earlier work.8 This cluster has three iron atoms, and here we studied binding at the central iron, which best represents the immediate environment around iron in the actual MOF. During optimization, the cluster of the MOF was frozen and the guest molecules (O2 or N2) were allowed to relax. The binding enthalpies were calculated using the formula given in eqn (1) of ref. 8.
The binding energy of the Pd(PH3)2C2H4 complex were computed using four basis sets. In all four basis sets, Pd atom has 18 active electrons and 28 core electrons that are replaced by an effective core potential. Basis set BS1 denotes the Stuttgart–Dresden–Dunning (SDD) basis set for Pd9 and the cc-pVTZ basis set for P,10 C, and H.11 Basis set BS2 denotes the def2-TZVP basis set for Pd12 and the cc-pVTZ basis set for P, C, and H. Basis set BS3 denotes the def2-TZVP basis set for Pd, the cc-pV(T+d)Z basis set for P,13,14 and the cc-pVTZ basis set for C and H. Basis set BS4 denotes the def2-TZVP basis set for Pd, the maug-cc-pV(T+d)Z basis set for P,15 the maug-cc-pVTZ basis set for C,15,16 and the cc-pVTZ basis set for H.
One basis set was used for Ag dimer, namely jun-cc-pVTZ-PP,17–19 one basis set was used for homonuclear transition metal bond distance, namely LanL2DZ,20–23 and two basis sets were used for Ar dimer, namely the aug-cc-pVQZ10,24 and aug-cc-pV6Z25 basis sets.
We divide the previous bond energy databases according to two types of classification: (i) whether the molecule contains only main-group nonmetal atoms or it also contains main-group-metal atoms or transition-metal atoms; (ii) whether the molecule has singe-reference character, i.e., can be well described by a single configuration wave function, or multi-reference character, i.e., cannot be so described. Then we added additional data to the underpopulated classes. Accordingly we have six new subdatabases for bond energies. Theses subdatabases are as follows (their shorthand names are in parentheses, where the final number in the shorthand name of a subdatabase is the number of data):
• single-reference main-group-metal bond energies (SR-MGM-BE9),
• single-reference main-group-nonmetal bond energies (SR-MGN-BE107),
• single-reference transition-metal bond energies (SR-TM-BE17),
• multi-reference main-group-metal bond energies (MR-MGM-BE4),
• multi-reference main-group nonmetal bond energies (MR-MGN-BE17),
• multi-reference transition-metal bond energies (MR-TM-BE15).
A new subdatabase called NGDWI21 has been added for noble-gas-dimer weak interactions. It comprises both homodimers and heterodimers.
We have added three new subdatabases for atomic excitation energies, namely
• 3d transition metal atomic excitation energies (3dAEE7),
• 4d transition metal atomic excitation energies (4dAEE5),
• p-block excitation energies (pEE5).
Two new subdatabases for p-block isomerization energies are added:
• 2p isomerization energies (2pIsoE4),
• 4p isomerization energies (4pIsoE4).
A new subdatabase for molecular geometries has been added; it is called diatomic geometries for heavy atoms (DGH4).
The above points summarize the main changes made to our previous database,26 called Database 2.0. A complete list of the subdatabases included in Database 2015 is given in Table 1, which also shows the number of data in each category (the inverse weight column of this table will be explained in Section 5). The database is divided into primary subdatabases, and some of the primary subdatabases are further divided into secondary subdatabases. Complete details of the new database and its layers of subdatabases, including geometries and references for the included data and also the basis sets we use for calculations on the various subdatabases, are given in the ESI.†
n | Primary subset | Secondary | Description |
I
n
![]() |
Ref. |
---|---|---|---|---|---|
a Databases 1–27 were used with various inverse weights in training, and databases 1–29 constitute Database 2015. Database 30 is from T. Weymuth et al. (ref. 56), and – like databases 28 and 29 – it was used only for testing. b In the name of a database or subdatabase, the number at the end of the name or before the solidus is the number of data. For example, ME417, SR-MGM-BE9, IsoL6/11, and DGH4 contain respectively 417, 9, 6, and 4 data. c Inverse weights with units of kcal mol−1 per bond for databases 1–7, kcal mol−1 for databases 8–24, and Å for databases 25–27. d TM denotes transition metal. e NA denotes not applicable. | |||||
ME417 | |||||
1 | SR-MGM-BE9 | Single-reference main-group metal bond energies | 2.00 | ||
SRM2 | Single-reference main-group bond energies | 26 | |||
SRMGD5 | Single-reference main-group diatomic bond energies | 26 and 65 | |||
3dSRBE2 | 3d single-reference metal–ligand bond energies | 66 | |||
2 | SR-MGN-BE107 | Single-reference main-group nonmetal bond energies | 0.20 | 26 | |
3 | SR-TM-BE17 | Single-reference TMd bond energies | 3.15 | ||
3dSRBE4 | 3d single-reference metal–ligand bond energies | 66 | |||
SRMBE10 | Single-reference metal bond energies | 26 | |||
PdBE2 | Palladium complex bond energies | 67 | |||
FeCl | FeCl bond energy | 68 | |||
4 | MR-MGM-BE4 | Multi-reference main-group metal bond energies | 4.95 | 65 | |
5 | MR-MGN-BE17 | Multi-reference main-group nonmetal bond energies | 1.25 | 26 | |
6 | MR-TM-BE13 | Multi-reference TM bond energies | 0.76 | ||
3dMRBE6 | 3d multi-reference metal–ligand bond energies | 66 | |||
MRBE3 | Multi-reference bond energies | 26 | |||
Remaining | Bond energies of remaining molecules: CuH, VO, CuCl, NiCl | 68 | |||
7 | MR-TMD-BE2 | Multi-reference TM dimer bond energies (Cr2 and V2) | 10.00 | 26 | |
8 | IP23 | Ionization potentials | 5.45 | 26 and 69 | |
9 | NCCE30 | Noncovalent complexation energies | 0.10 | 26 and 70–73 | |
10 | NGDWI21 | Noble gas dimer weak interactions | 0.01 | 26 and 74 | |
11 | 3dAEE7 | 3d TM atomic excitation energies | 0.40 | 69 and 75 | |
12 | 4dAEE5 | 4d TM atomic excitation energies | 6.90 | 76 | |
13 | pEE5 | p-block excitation energies | 1.74 | 77 | |
14 | 4pIsoE4 | 4p isomerization energies | 8.00 | 78 | |
15 | 2pIsoE4 | 2p isomerization energies | 7.81 | 78 | |
16 | IsoL6/11 | Isomerization energies of large molecules | 2.00 | 26 | |
17 | EA13/03 | Electron affinities | 2.96 | 26 | |
18 | PA8 | Proton affinities | 2.23 | 26 | |
19 | πTC13 | Thermochemistry of π systems | 5.75 | 26 | |
20 | HTBH38/08 | Hydrogen transfer barrier heights | 0.25 | 26 | |
21 | NHTBH38/08 | Non-hydrogen transfer barrier heights | 0.80 | 26 | |
22 | AE17 | Atomic energies | 10.22 | 26 | |
23 | HC7/11 | Hydrocarbon chemistry | 6.48 | 26 | |
24 | DC9/12 | Difficult cases | 10.00 | 26 | |
MS10 | |||||
25 | DGL6 | Diatomic geometries of light-atom molecules | 0.01 | 26 | |
26 | DGH4 | Diatomic geometries of heavy-atom molecules: ZnS, HBr, NaBr | 0.01 | 79 | |
Diatomic geometry of Ag2 | 0.0013 | 80 | |||
SS17 | |||||
27 | LC17 | Lattice constants | 0.013 | 26 | |
SSE39 | |||||
28 | SBG31 | Semiconductor band gaps | NAe | 26 | |
29 | SSCE8 | Solid-state cohesive energies | NA | 26 | |
WCCR10 | |||||
30 | WCCR10a | Ligand dissociation energies of large cationic TM complexes | NA | 56 |
Category | X | Type | Year | Method | Ref. |
---|---|---|---|---|---|
a X is the percentage of nonlocal Hartree–Fock exchange. When a range is given, the first value is for small interelectronic distances, and the second value is for large interelectronic distances. Details of the functional form that joins these regions of interelectronic separation are given in the references. b GVWN5 denotes the Gáspár approximation for exchange and the VWN5 fit to the correlation energy; this is an example of the local spin density approximation (LSDA), and it has the keyword SVWN5 in the Gaussian 09 program. Note that Kohn–Sham exchange is the same as Gáspár exchange, but Slater exchange (not tested here) is greater by a factor of 1.5. c PW91 formally satisfies the gradient expansion for exchange to second order but only at such small values of the gradient that for practical purposes it should be grouped with functionals that do not satisfy the gradient expansion to second order. | |||||
Local | 0 | LSDA | 1980 | GKSVWN5b | 27–29 |
0 | GGA – correct to 2nd order in exchange | 2008 | SOGGA | 30 | |
0 | 2008 | PBEsol | 31 | ||
0 | 2011 | SOGGA11 | 41 | ||
0 | GGA – other | 1988 | BP86 | 33 and 34 | |
0 | 1988 | BLYP | 34 and 36 | ||
0 | 1991 | PW91c | 35 | ||
0 | 1991 | BPW91 | 34 and 35 | ||
0 | 1996 | PBE | 32 | ||
0 | 1997 | mPWPW | 37 | ||
0 | 1997 | revPBE | 38 | ||
0 | 1999 | RPBE | 39 | ||
0 | 2000 | HCTH407 | 40 | ||
0 | 2001 | OLYP | 36 and 42 | ||
0 | 2009 | OreLYP | 36, 42 and 43 | ||
0 | NGA | 2012 | N12 | 26 and 74 | |
0 | 2015 | GAM | Present | ||
0 | Meta-GGA | 2003 | TPSS | 44 | |
0 | 2006 | M06-L | 46 | ||
0 | 2009 | revTPSS | 45 | ||
0 | 2011 | M11-L | 47 | ||
Nonlocal | 20 | Global hybrid GGA | 1994 | B3LYP | 42 and 48 |
0–25 | Range-separated hybrid GGA | 2009 | HSE06 | 50 and 51 |
Exc = ENSGAnxc + Ec | (1) |
![]() | (2) |
![]() | (3) |
ΓNSGAnxcσ = εUEGxσ(ρσ)Fx(ρσ, xσ) | (4) |
![]() | (5) |
uxσ = γxσxσ2/(1 + γxσxσ2) | (6) |
vxσ = ωxσρ1/3σ(1 + ωxσρ1/3σ) | (7) |
![]() | (8) |
![]() | (9) |
A GGA exchange functional can be written like eqn (4) but where the enhancement factor Fx depends only on the reduced spin density gradient xσ. For an NGA we allow the enhancement factor to depend also on the spin density ρσ.
![]() | (10) |
![]() | (11) |
![]() | (12) |
![]() | (13) |
We optimize our functional against 27 primary databases, including 24 molecular energy databases, two molecular structure databases, and one solid-state structure database. We optimize the GAM functional self-consistently by minimizing the following unfitness function:
![]() | (14) |
![]() | (15) |
![]() | (16) |
![]() | (17) |
In order to design a functional with good across-the-board performance, we include various molecular and solid-state properties in our training set; these properties include main-group bond energies, transition metal bond energies, transition metal atomic excitation energies, barrier heights, ionization potentials, proton affinities, electron affinities, lattice constants, etc. In Table 1, the inverse weight of each primary database is given. The smaller the inverse weight is, the more emphasis we put on that primary database. The inverse weights were chosen as follows: first we calculated the mean unsigned errors (MUEs) of 80 exchange–correlation functionals (previously published functionals developed in many different groups) for all the molecular subdatabases in Database 2015; this shows how well previous exchange–correlation functionals typically perform for each kind of data. The average of these MUEs for a given subdatabase were used as our initial inverse weights. Then we modified the inverse weights iteratively to improve performance on the various subdatabases where we wished to reduce the error. Our goal in this process was to obtain good across the board performance for as many subdatabases as possible, not to simply reduce the overall mean unsigned error.
Whereas the N124 functional involved 20 optimized linear coefficients and the constraint that it reduced to PBEsol at low density, the new GAM functional involves optimizing 26 linear coefficients in eqn (5), (10), and (11) with no constraints. We use the same values as N12 for the nonlinear parameters ωxσ, γxσ, γcαβ, and γcσσ. A key element in the optimization is the choice of weights. We do not choose them to minimize the overall error but rather to try to get small errors across the board, i.e., relatively small errors for each of the subdatabases, to the greatest extent possible. The final choice of weights was determined after considerable trial and error and is a subjective decision that cannot be justified by any numerical argument.
Table 3 lists the values for the optimized and inherited parameters of the GAM functional.
Exchange | Correlation | ||
---|---|---|---|
Optimized parameters | |||
a 00 | 1.32730 | b 0 | 0.860548 |
a 01 | 0.886102 | b 1 | −2.94135 |
a 02 | −5.73833 | b 2 | 15.4176 |
a 03 | 8.60197 | b 3 | −5.99825 |
a 10 | −0.786018 | b 4 | −23.4119 |
a 11 | −4.78787 | c 0 | 0.231765 |
a 12 | 3.90989 | c 1 | 0.575592 |
a 13 | −2.11611 | c 2 | −3.43391 |
a 20 | 0.802575 | c 3 | −5.77281 |
a 21 | 14.4363 | c 4 | 9.52448 |
a 22 | 8.42735 | ||
a 23 | −6.21552 | ||
a 30 | −0.142331 | ||
a 31 | −13.4598 | ||
a 32 | 1.52355 | ||
a 33 | −10.0530 | ||
Inherited parameters | |||
ω xσ | 2.5 | γ cαβ | 0.006 |
γ xσ | 0.004 | γ cσσ | 0.2 |
Type | LSDA | GGA | GGA | GGA | GGA | GGA | GGA | GGA | GGA | GGA | GGA | GGA | GGA | GGA | GGA | NGA | NGA |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Functional | GKSVWN5 | SOGGA | PBEsol | PBE | BP86 | PW91 | BLYP | mPWPW | revPBE | BPW91 | RPBE | HCTH407 | SOGGA11 | OLYP | OreLYP | N12 | GAM |
a The MGBE137 database consists of SR-MGM-BE9, SR-MGN-BE107, MR-MGM-BE4, and MR-MGN-BE17. b The TMBE32 database consists of SR-TM-BE17, MR-TM-BE13, and MR-TMD-BE2. c The BH76 database consists of HTBH38/08 and NHTBH38/08. d The ME417 database consists all the 24 subdatabases above and the ME400xAE consists all the subdatabases except AE17. | |||||||||||||||||
SR-MGM-BE9 | 11.64 | 4.43 | 4.47 | 2.72 | 3.10 | 2.57 | 5.07 | 2.87 | 4.26 | 3.20 | 4.57 | 3.52 | 8.79 | 4.67 | 4.06 | 5.92 | 2.00 |
SR-MGN-BE107 | 16.21 | 7.27 | 7.28 | 3.40 | 4.06 | 3.51 | 2.78 | 2.80 | 2.99 | 2.49 | 3.35 | 2.55 | 2.77 | 2.32 | 2.56 | 2.38 | 2.27 |
SR-TM-BE17 | 20.89 | 11.59 | 11.33 | 7.20 | 7.39 | 8.76 | 6.52 | 6.73 | 6.22 | 7.34 | 6.24 | 8.36 | 11.44 | 9.32 | 7.15 | 8.31 | 6.31 |
MR-MGM-BE4 | 24.56 | 14.48 | 15.81 | 9.31 | 9.49 | 10.26 | 8.75 | 9.02 | 6.24 | 8.03 | 6.43 | 10.11 | 7.44 | 8.39 | 8.35 | 9.10 | 7.76 |
MR-MGN-BE17 | 36.89 | 21.29 | 23.16 | 14.80 | 13.87 | 14.80 | 6.67 | 12.45 | 5.94 | 10.74 | 5.51 | 5.24 | 8.57 | 5.15 | 4.25 | 6.93 | 4.22 |
MR-TM-BE13 | 34.07 | 22.03 | 21.24 | 12.73 | 12.11 | 13.25 | 10.64 | 11.67 | 8.55 | 10.81 | 7.73 | 19.70 | 18.79 | 5.77 | 5.10 | 12.54 | 4.94 |
IsoL6/11 | 2.05 | 1.89 | 1.55 | 1.98 | 2.28 | 1.92 | 3.73 | 2.16 | 2.82 | 2.38 | 2.99 | 3.02 | 1.73 | 3.44 | 3.39 | 1.73 | 1.96 |
IP23 | 9.59 | 4.84 | 5.82 | 6.19 | 8.44 | 7.29 | 6.52 | 6.85 | 5.00 | 6.30 | 4.92 | 6.81 | 5.92 | 3.12 | 3.03 | 4.36 | 4.53 |
EA13/03 | 5.70 | 2.70 | 2.16 | 2.27 | 4.21 | 2.60 | 2.68 | 2.31 | 2.40 | 2.26 | 2.37 | 3.70 | 5.23 | 3.60 | 2.32 | 4.12 | 4.49 |
PA8 | 5.07 | 2.33 | 2.10 | 1.34 | 1.41 | 1.30 | 1.58 | 1.52 | 2.00 | 1.88 | 1.98 | 2.84 | 2.11 | 2.40 | 1.70 | 1.35 | 3.84 |
πTC13 | 4.80 | 4.06 | 4.20 | 5.59 | 5.85 | 5.73 | 6.07 | 6.41 | 7.15 | 7.08 | 7.20 | 8.23 | 7.41 | 8.26 | 7.27 | 8.61 | 8.59 |
HTBH38/08 | 17.56 | 12.88 | 12.69 | 9.31 | 9.16 | 9.60 | 7.52 | 8.43 | 6.58 | 7.38 | 6.43 | 5.48 | 6.57 | 5.63 | 6.28 | 6.94 | 5.35 |
NHTBH38/08 | 12.42 | 9.68 | 9.86 | 8.42 | 8.72 | 8.80 | 8.53 | 8.03 | 6.82 | 7.26 | 6.82 | 6.29 | 4.32 | 5.25 | 5.57 | 6.86 | 5.15 |
NCCE30 | 3.61 | 2.12 | 2.07 | 1.46 | 1.53 | 1.60 | 1.64 | 1.42 | 1.71 | 1.74 | 1.61 | 1.32 | 1.48 | 2.52 | 2.68 | 1.38 | 1.29 |
AE17 | 421.13 | 283.06 | 245.90 | 47.24 | 16.92 | 4.63 | 8.68 | 12.55 | 10.88 | 11.95 | 9.39 | 16.80 | 10.06 | 10.13 | 2.37 | 14.21 | 10.18 |
HC7/11 | 21.45 | 17.88 | 13.31 | 3.97 | 9.95 | 4.55 | 27.39 | 8.08 | 13.65 | 10.77 | 14.96 | 14.97 | 6.26 | 17.01 | 16.34 | 4.27 | 6.24 |
3dAEE7 | 11.86 | 10.87 | 10.77 | 9.80 | 10.36 | 10.47 | 10.27 | 10.63 | 10.05 | 10.84 | 9.78 | 12.00 | 12.50 | 11.56 | 10.98 | 18.51 | 9.82 |
4dAEE5 | 14.10 | 4.77 | 8.48 | 4.70 | 5.07 | 4.73 | 5.73 | 4.89 | 4.49 | 5.03 | 4.27 | 7.75 | 7.60 | 5.94 | 6.42 | 10.24 | 5.23 |
pEE5 | 4.36 | 6.30 | 5.15 | 3.96 | 3.46 | 4.14 | 5.10 | 5.22 | 4.37 | 6.33 | 3.51 | 4.27 | 5.01 | 2.09 | 3.25 | 14.86 | 2.99 |
DC9/12 | 17.35 | 14.61 | 13.34 | 14.99 | 15.11 | 13.94 | 17.88 | 14.76 | 20.35 | 16.21 | 21.48 | 19.74 | 16.65 | 21.71 | 22.57 | 10.20 | 23.07 |
2pIsoE4 | 2.05 | 1.44 | 1.71 | 2.73 | 3.21 | 2.87 | 5.45 | 3.20 | 3.59 | 3.43 | 3.70 | 4.59 | 1.72 | 3.95 | 3.72 | 3.41 | 5.02 |
4pIsoE4 | 3.05 | 2.29 | 2.28 | 2.43 | 2.87 | 2.58 | 4.00 | 2.50 | 2.16 | 2.41 | 2.16 | 3.29 | 3.27 | 2.15 | 2.22 | 1.73 | 3.57 |
NGDWI21 | 0.212 | 0.082 | 0.081 | 0.102 | 0.528 | 0.165 | 0.385 | 0.220 | 0.282 | 0.587 | 0.179 | 0.246 | 0.650 | 0.323 | 0.389 | 0.387 | 0.019 |
MR-TMD-BE2 | 51.28 | 33.08 | 30.87 | 28.10 | 24.40 | 27.97 | 42.70 | 29.43 | 28.40 | 30.96 | 26.79 | 20.09 | 35.20 | 25.18 | 12.74 | 27.97 | 10.67 |
MGBE137a | 18.72 | 9.04 | 9.31 | 4.94 | 5.38 | 5.05 | 3.58 | 4.18 | 3.54 | 3.73 | 3.79 | 3.17 | 4.02 | 3.00 | 3.04 | 3.37 | 2.65 |
TMBE32![]() |
28.14 | 17.18 | 16.58 | 10.75 | 10.37 | 11.79 | 10.45 | 10.15 | 8.55 | 10.22 | 8.13 | 13.70 | 15.91 | 8.87 | 6.66 | 11.26 | 6.03 |
BH76c | 14.99 | 11.28 | 11.27 | 8.87 | 8.94 | 9.20 | 8.02 | 8.23 | 6.70 | 7.32 | 6.62 | 5.88 | 5.44 | 5.44 | 5.92 | 6.90 | 5.25 |
ME417d | 30.67 | 19.55 | 18.04 | 7.45 | 6.68 | 5.98 | 5.89 | 5.80 | 5.27 | 5.60 | 5.26 | 5.90 | 5.74 | 5.01 | 4.56 | 5.57 | 4.51 |
ME400xAEd | 14.07 | 8.36 | 8.36 | 5.76 | 6.25 | 6.03 | 5.77 | 5.51 | 5.03 | 5.33 | 5.09 | 5.44 | 5.56 | 4.79 | 4.66 | 5.20 | 4.27 |
Type | NGA | Meta | Meta | Meta | Meta | Hybrid | Hybrid |
---|---|---|---|---|---|---|---|
Functional | GAM | TPSS | revTPSS | M06-L | M11-L | B3LYP | HSE06 |
a The MGBE137, TMBE32, BH76, ME417, and ME400xAE notations are explained in footnotes to Table 4. | |||||||
SR-MGM-BE9 | 2.00 | 2.55 | 2.91 | 3.40 | 7.24 | 4.58 | 3.47 |
SR-MGN-BE107 | 2.27 | 2.43 | 2.24 | 2.03 | 1.76 | 2.45 | 2.08 |
SR-TM-BE17 | 6.31 | 6.11 | 6.13 | 6.24 | 5.73 | 5.48 | 4.96 |
MR-MGM-BE4 | 7.76 | 6.69 | 5.98 | 6.15 | 13.50 | 7.76 | 8.52 |
MR-MGN-BE17 | 4.22 | 4.25 | 4.62 | 3.11 | 4.02 | 5.09 | 5.30 |
MR-TM-BE13 | 4.94 | 8.87 | 6.81 | 4.40 | 4.44 | 5.33 | 4.87 |
IsoL6/11 | 1.96 | 3.66 | 3.96 | 2.76 | 1.57 | 2.61 | 1.25 |
IP23 | 4.53 | 4.29 | 4.07 | 3.91 | 4.77 | 5.51 | 4.06 |
EA13/03 | 4.49 | 2.35 | 2.59 | 3.83 | 5.54 | 2.33 | 2.77 |
PA8 | 3.84 | 2.66 | 2.79 | 1.88 | 2.17 | 1.02 | 1.10 |
πTC13 | 8.59 | 8.12 | 7.85 | 6.69 | 5.14 | 6.03 | 6.20 |
HTBH38/08 | 5.35 | 7.71 | 6.96 | 4.15 | 1.44 | 4.23 | 4.23 |
NHTBH38/08 | 5.15 | 8.91 | 9.07 | 3.81 | 2.86 | 4.55 | 3.73 |
NCCE30 | 1.29 | 1.34 | 1.33 | 0.90 | 0.81 | 1.09 | 0.95 |
AE17 | 10.18 | 18.04 | 23.81 | 7.04 | 21.81 | 18.29 | 32.82 |
HC7/11 | 6.24 | 10.48 | 6.42 | 3.35 | 2.42 | 16.80 | 7.34 |
3dAEE7 | 9.82 | 10.78 | 10.47 | 7.84 | 14.03 | 8.47 | 10.62 |
4dAEE5 | 5.23 | 5.19 | 5.11 | 6.58 | 11.04 | 5.67 | 5.07 |
pEE5 | 2.99 | 2.25 | 2.31 | 7.50 | 10.39 | 2.87 | 5.70 |
DC9/12 | 23.07 | 14.20 | 14.94 | 10.67 | 5.90 | 12.02 | 9.08 |
2pIsoE4 | 5.02 | 3.54 | 2.53 | 3.16 | 3.32 | 4.69 | 2.44 |
4pIsoE4 | 3.57 | 2.60 | 3.27 | 2.88 | 5.03 | 4.24 | 2.64 |
NGDWI21 | 0.019 | 0.171 | 0.174 | 0.125 | 0.568 | 0.276 | 0.102 |
MR-TMD-BE2 | 10.67 | 26.21 | 26.59 | 7.22 | 22.18 | 31.21 | 45.13 |
MGBE137a | 2.65 | 2.79 | 2.69 | 2.37 | 2.74 | 3.07 | 2.76 |
TMBE32 | 6.03 | 8.49 | 7.68 | 5.55 | 6.24 | 7.03 | 7.43 |
BH76 | 5.25 | 8.31 | 8.01 | 3.98 | 2.15 | 4.39 | 3.98 |
ME417 | 4.51 | 5.40 | 5.42 | 3.55 | 4.15 | 4.68 | 4.83 |
ME400xAE | 4.27 | 4.86 | 4.64 | 3.41 | 3.40 | 4.10 | 3.64 |
Functional | Type | DGL6 | DGH4 | MS10a |
---|---|---|---|---|
a The MS10 database consists of DGL6 and DGH4 subdatabases. The functionals are listed in the same order as in Tables 4 and 5. | ||||
GKSVWN5 | LSDA | 0.011 | 0.031 | 0.019 |
SOGGA | GGA | 0.009 | 0.013 | 0.010 |
PBEsol | GGA | 0.010 | 0.007 | 0.009 |
PBE | GGA | 0.013 | 0.020 | 0.016 |
BP86 | GGA | 0.015 | 0.021 | 0.018 |
PW91 | GGA | 0.012 | 0.019 | 0.015 |
BLYP | GGA | 0.019 | 0.037 | 0.026 |
mPWPW | GGA | 0.012 | 0.021 | 0.016 |
revPBE | GGA | 0.015 | 0.034 | 0.023 |
BPW91 | GGA | 0.013 | 0.022 | 0.017 |
RPBE | GGA | 0.016 | 0.038 | 0.025 |
HCTH407 | GGA | 0.004 | 0.033 | 0.015 |
SOGGA11 | GGA | 0.008 | 0.053 | 0.026 |
OLYP | GGA | 0.009 | 0.036 | 0.020 |
OreLYP | GGA | 0.011 | 0.034 | 0.020 |
N12 | NGA | 0.008 | 0.007 | 0.008 |
GAM | NGA | 0.007 | 0.034 | 0.018 |
TPSS | Meta | 0.010 | 0.015 | 0.012 |
revTPSS | Meta | 0.011 | 0.009 | 0.010 |
M06-L | Meta | 0.006 | 0.018 | 0.011 |
M11-L | Meta | 0.012 | 0.033 | 0.021 |
B3LYP | Hybrid | 0.009 | 0.027 | 0.016 |
HSE06 | Hybrid | 0.003 | 0.015 | 0.008 |
Table 7 gives the performance for solid-state databases, but since B3LYP calculations with periodic boundary conditions are very expensive, we only compare 21 density functionals for the solid-state lattice constant and energetic data of Table 7. Table 8 compares the performance of GAM to that of eight density functionals for the WCCR10 database of Weymuth et al.56Table 9 is a test for the binding of dioxygen and dinitrogen to Fe2(dobdc), which is also called Fe-MOF-74, where we compare to experiments of Bloch et al.57Table 10 presents results for the binding of ethylene to Pd(PH3)2, where we compare the results of GAM to the best estimate computed using BCCD(T)58 in our earlier work.67Table 11 presents results for the bond distance of homonuclear transition metal dimers, where we compare the results of GAM and N12 with 5 functionals in a recent paper.80
Functionala | Type | LC17 | NNDb | SBG31 | SSCE8 |
---|---|---|---|---|---|
a The functionals are listed in the same order as in Tables 4 and 5. b The values in this column are obtained by dividing the previous column by 2.15 (a standard factor determined in previous work – see text) so that the results may be compared more physically to errors in molecular bond lengths. | |||||
GKSVWN5 | LSDA | 0.069 | 0.032 | 1.14 | 0.70 |
SOGGA | GGA | 0.022 | 0.010 | 1.14 | 0.31 |
PBEsol | GGA | 0.023 | 0.011 | 1.14 | 0.27 |
PBE | GGA | 0.068 | 0.031 | 0.98 | 0.11 |
BP86 | GGA | 0.073 | 0.034 | 1.12 | 0.12 |
PW91 | GGA | 0.065 | 0.030 | 1.11 | 0.50 |
BLYP | GGA | 0.111 | 0.052 | 1.14 | 0.37 |
mPWPW | GGA | 0.075 | 0.035 | 1.11 | 0.10 |
revPBE | GGA | 0.110 | 0.051 | 1.08 | 1.12 |
BPW91 | GGA | 0.083 | 0.038 | 1.10 | 0.20 |
RPBE | GGA | 0.119 | 0.055 | 1.07 | 0.61 |
HCTH407 | GGA | 0.120 | 0.056 | 0.89 | 0.30 |
SOGGA11 | GGA | 0.125 | 0.058 | 0.89 | 0.07 |
OLYP | GGA | 0.118 | 0.055 | 0.90 | 0.36 |
OreLYP | GGA | 0.113 | 0.053 | 0.92 | 0.20 |
N12 | NGA | 0.027 | 0.012 | 0.99 | 0.13 |
GAM | NGA | 0.092 | 0.046 | 0.99 | 0.13 |
TPSS | Meta | 0.055 | 0.025 | 0.85 | 0.22 |
revTPSS | Meta | 0.039 | 0.018 | 1.00 | 0.13 |
M06-L | Meta | 0.080 | 0.037 | 0.73 | 0.17 |
M11-L | Meta | 0.073 | 0.034 | 0.54 | 0.24 |
HSE06 | Hybrid | 0.041 | 0.019 | 0.26 | 0.11 |
M S(Fe, X2)c | ΔHb | ||
---|---|---|---|
GAM | Expt.e | ||
a The basis set is def2-TZVP. b The binding enthalpy (a positive value indicates exothermic binding). c This column has the MS values for the central Fe and the guest molecule in the initial iteration of self-consistent field calculations. The two peripheral Fe centers where no guest is bound were taken to have MS = 2 for all the calculations. d This column is calculated by eqn (1) of ref. 8. e The most recent experimental value is shown, as discussed in the text. f NA denotes not applicable. | |||
Fe–N2 | 2, 0 | 3.9 | 5.5 |
Fe–O2 | 2, 1 | 10.8 | 9.8 |
2, 0 | 7.8 | NAf | |
2, −1 | 5.0 | NA |
Basis seta | GAM | Best estimateb |
---|---|---|
a The various basis sets used are: BS1 = SDD (Pd), cc-pVTZ (P, C, H); BS2 = def2-TZVP (Pd), cc-pVTZ (P, C, H); BS3 = def2-TZVP (Pd), cc-pV(T+d)Z (P), cc-pVTZ (C, H); BS4 = def2-TZVP (Pd), maug-cc-pV(T+d)Z (P), maug-cc-pVTZ (C), cc-pVTZ (H). b The best estimate was calculated in an earlier work using BCCD(T) and is described in ref. 67. | ||
BS1 | 11.0 | 17.6 |
BS2 | 11.1 | |
BS3 | 11.1 | |
BS4 | 11.1 |
Cu2 | Au2 | Ni2 | Pd2 | Pt2 | Ir2 | Os2 | MUE | |
---|---|---|---|---|---|---|---|---|
a The experimental bond length is taken from ref. 80. | ||||||||
LSDA | 2.215 | 2.495 | 2.118 | 2.373 | 2.353 | 2.271 | 2.354 | 0.038 |
PBE | 2.278 | 2.552 | 2.135 | 2.397 | 2.391 | 2.302 | 2.384 | 0.062 |
B3LYP | 2.292 | 2.577 | 2.099 | 2.411 | 2.392 | 2.301 | 2.387 | 0.071 |
B3PW91 | 2.288 | 2.552 | 2.095 | 2.367 | 2.375 | 2.287 | 2.373 | 0.068 |
mPWPW | 2.293 | 2.549 | 2.088 | 2.359 | 2.369 | 2.282 | 2.369 | 0.068 |
N12 | 2.224 | 2.543 | 2.110 | 2.501 | 2.366 | 2.262 | 2.282 | 0.026 |
GAM | 2.306 | 2.543 | 2.189 | 2.536 | 2.408 | 2.283 | 2.292 | 0.050 |
Exp.a | 2.219 | 2.472 | 2.155 | 2.480 | 2.333 | 2.270 | 2.280 | 0.000 |
Table 4 shows that among LSDA, all the GGAs, and the previous NGA, the new functional GAM gives the smallest overall mean unsigned error for the entire molecular energy database ME417; the mean unsigned error is only 4.51 kcal mol−1. We also show the overall error of ME400xAE, which is the average error for the molecular energy database when we exclude absolute atomic energies, and in this case too, the GAM functional gives the smallest error among GGAs, LSDA, and NGA. We emphasize that we could reduce these total errors more, if that were our goal, but that is not our goal. Our goal is rather to obtain good performance across a broad range of databases. In order to have a functional that is especially good for studying molecular catalysis, the functional should be good for main-group bond energy (MGBE137), which includes the SR-MGM-BE9, SR-MGN-BE107, MR-MGM-BE4, and MR-MGN-BE17 subdatabases, for transition metal bond energy (TMBE32), which includes the SR-TM-BE17, MR-TM-BE13, and MR-TMD-BE2 subdatabases, for barrier heights (BH76), which includes the HTBH38/08 and NHTBH38/08 subdatabases, and for molecular structure (MS10), which includes DGL6 and DGH4 subdatabases. In Tables 4 and 5 we calculate the average error for each of these four categories by averaging the errors from each subdatabase. Among LSDA and all GGAs and NGAs, the GAM functional ranked the best for the MGBE137, TMBE32, and BH76 subdatabases. If we consider all the functionals in Tables 4 and 5, the GAM functional ranks the second best for TMBE32 subdatabase, for which M06-L is the best with an error 0.48 kcal mol−1 smaller than the GAM; the GAM functional ranks the second best for the MGBE137 subdatabase, for which M06-L is the best with an error 0.28 kcal mol−1 smaller than the GAM; and the GAM functional ranks the fifth best for BH76 subdatabase, for which M11-L is the best followed by M06-L, B3LYP, and HSE06. We note that M06-L is a meta functional, and therefore it should be better than a simpler gradient approximation, but we gave several reasons for optimizing a gradient approximation in the introduction.
In addition to the databases mentioned above, the GAM functional also provides good results for 3d transition metal atomic excitation energies, which are very hard for most available density functionals, but which we have recently shown60 can be very important for understanding metal–metal bonding. The GAM functional ranks the fifth best for the 3dAEE7 subdatabase, behind M06-L, B3LYP, PBE, and RPBE.
Next we consider noble-gas weak interactions. From Tables 4 and 5 we can see that all the functionals tested except GAM give a mean unsigned error larger than 0.081 kcal mol−1 for the NGDWI21 subdatabase, for which GAM only gives 0.019 kcal mol−1. The average value of all the noble gas weak interaction energies in our database is 0.160 kcal mol−1, which means that most functionals give an average error that is larger than 50% of the average of the reference values. The GAM functional gives the best results for NCCE30 subdatabase as compared to all tested GGAs and N12.
The GAM functional also provides the second best results for MR-TMD-BE2 (Cr2 and V2, which are known to be very hard cases for density functional theory) among all functionals tested.
Table 6 shows that the relative performance of GAM for molecular structures is not quite as good as for energies. The GAM functional ranks the 13th for MS10 subdatabase with an MUE of 0.018 Å, which is 0.002 Å larger than the average MUE of all functionals tested in Table 6. However, a more fair comparison in this case is to compare to the 12 GGAs excluding PBEsol and SOGGA (we exclude PBEsol and SOGGA at this point since their design is understood to make them better for structures than for energies, and so we do not consider them to be general-purpose functionals). As compared to the remaining group of 12 GGAs, only HCTH407 does better for DGL6 and only four of the 12 do better for MS10. The less than stellar performance of GAM on MS10 is primarily due to a large overestimation of the bond length of Ag dimer; this bond length behaves differently than other bond lengths in MS10, and success for this bond length is highly correlated to performance on lattice constants, which we downplayed. This downplay is evidenced in Table 7, which shows that the GAM functional does not give good results for the solid-state lattice constant database with a mean unsigned error of 0.046 Å for the quantities we nominally call nearest neighbor distances (NND – see Section 6.3); this error is 0.010 Å larger than the average mean unsigned error for NND. As discussed in the introduction, this results from a strategic decision to emphasize molecular energies over lattice constants in the creation of GAM. The “M” (for “molecules”) at the end of GAM is primarily to indicate our awareness that we still do not have a universally good functional, which is so far unattainable by any functional containing only density and density gradient ingredients. Nevertheless, despite not being universal, the performance of the new functional developed here is very good if we consider molecules rather than solid-state lattice constants.
Next we turn to data not used for training.
Table 7 shows that the GAM functional also shows reasonably good results for the solid-state energies databases. Among the 17 LSDA, GGAs, and NGAs, the GAM functional ranks the sixth best for the SBG31 database and fifth best for the SSCE8 database. These databases were not used for training.
In Table 8, the GAM functional ranks the second best among all functionals being tested, where the functionals tested are those chosen by the previous56 authors. The WCCR10 database includes ten transition metal coordination reactions. The molecules involved in these reactions are very large and very different from the training sets in Database 2015. The performance against these large molecules is slightly worse than that for the transition metal molecules in our training set, but within a reasonable range.
Table 9 presents the results for the performance of GAM on MOFs. We find that GAM gives good results when compared to experiments for the separation of O2 and N2 on Fe2(dobdc), with a magnitude of the deviation from experimental adsorption enthalpies of 1.0 kcal mol−1 for O2 and 1.6 kcal mol−1 for N2. It should be noted here that our training set has no data on MOFs or any other type of nanoporous materials. This average deviation from the latest experimental values is under 3 kcal mol−1 and is within experimental error. This indicates that the GAM functional shows good agreement with experimental data that are not used for training.
In Table 10, results for the binding of C2H4 to Pd(PH3)2 are presented. This datum is outside the training set. This is a difficult case for functionals; for example, BLYP gives a binding energy of 10.2 kcal mol−1 as compared to the best estimate of 17.6 kcal mol−1. Table 10 shows good stability with respect to changes in the basis set and that the GAM functional deviates from the best estimate by 6.5 kcal mol−1 with the largest basis set used. This is comparable to the 6.3 kcal mol−1 mean unsigned error for single-reference transition metal bond energies of molecules in the training set, and therefore it is an example where we obtain comparable performance inside and outside of the training set.
A very recent paper, which we considered only after our training set weights were final, reported bond distance for eight transition metal dimers, only one of which (Ag2) is in our training set. We therefore use the bond distances of the seven others as a test against data quite different from that used for training. These seven dimers, Cu2, Au2, Ni2, Pd2, Pt2, Ir2, and Os2, include two 3d metals, one 4d metals, and four 5d metals. (No 5d data were used for training.) The GAM functional gives the third best results among all the functionals tested in Table 11, with an MUE of 0.05 Å; the only functionals that do better are LSDA, which is much better for bond lengths than for molecular energies, and N12, the only previous NGA. This is very encouraging performance well outside the training set.
We also tested our new functional against the experimental bond dissociation energies of Ag2 and FeC, which are 38.0 kcal mol−1 and 88.32 kcal mol−1 respectively.61,62 The GAM functional predicts these bond dissociation energies to be 39.21 kcal mol−1 and 86.49 kcal mol−1. Li et al.62 have tested the bond dissociation energy of FeC with various functionals, and in Table 12 we add our new result to their comparison. The results in Table 12 show that the GAM functional is the second best among all 18 functionals being tested, and that many of the previous functionals have large errors for this difficult case.
M11-L | SOGGA11 | τ-HCTHhyb | M06-L | BLYP | B3LYP | M05 | M06 | |
---|---|---|---|---|---|---|---|---|
FeC | −4.60 | 10.81 | −7.13 | −7.36 | 12.88 | −1.38 | 5.75 | −20.93 |
ωB97 | ωB97X | ωB97X-D | M08-SO | M08-HX | M11 | SOGGA11-X | GAM | |
---|---|---|---|---|---|---|---|---|
FeC | −38.87 | −20.01 | 21.39 | −26.68 | −35.65 | −37.03 | −67.16 | 1.83 |
Recent studies pointed out that some density functionals give unstable results for large basis sets.63Fig. 1 shows the potential energy curve of Ar2 with our new GAM functional and the aug-cc-pVQZ and aug-cc-pV6Z basis sets. Fig. 1 shows that our results are very close to the reference values64 and there is no slow convergence issue with respect to the basis sets. Moreover, the excellent agreement with the reference plot shows that the GAM functional provides good results for noble gas weak interactions. This is consistent with Tables 4 and 5 showing that the GAM functional is the best for the NGDWI21 subdatabase among all the functionals tested in the present paper.
![]() | ||
Fig. 1 Ar–Ar potential curve, the bonding energies are calculated with GAM/aug-cc-pVQZ and GAM/aug-cc-pV6Z level of theory. The reference is from the Tang–Toennies model. |
Exc = ∫dr![]() | (18) |
![]() | (19) |
![]() | (20) |
A key design element of the NGA functional form is that, unlike GGAs, we do not attempt to separately fit exchange and correlation. Therefore, unlike a GGA, we do not have a pure-exchange enhancement factor that depends only on s. However, Fig. 2 and 3 show that after we add correlation to exchange, the extent of dependence on ρ for closed-shell systems is not qualitatively different in GAM and in the GGAs.
(1) The GAM functional gives the smallest mean unsigned error for main group bond energies (MGBE137), transition metal bond energies (TMBE32), and reaction barrier heights (BH76).
(2) The GAM functional gives the smallest mean unsigned error of 0.019 kcal mol−1 for the noble gas dimer weak interaction energies (NGDWI21), with all the other functionals tested here giving a mean unsigned error larger than 0.081 kcal mol−1, which is about 50% of the reference value.
(3) GAM is best of any LSDA, GGA, or NGA for both the overall mean unsigned error for molecular energies, either including total atomic energies (ME417) or excluding them (ME400xAE). OreLYP (which has not previously been widely tested) and OLYP are the second and the third best.
The GAM functional gives an MUE of 0.018 Å for the molecular structure subdatabase (MS10), which is reasonable, although not outstanding.
Besides the training sets tested in the paper, we also tested the performance of the GAM functional against band gaps (SBG31), solid-state cohesive energies (SSCE8), transition metal coordination reactions (WCCR10), the bond energies of Ag2 and FeC, adsorption enthalpies of gases on MOFs, the binding of C2H4 to Pd(PH3)2, and the bond distances of homonuclear transition metal dimers (HTMD7). The last-named test includes four 5d transition metals, although no 5d transition metal data was used for training. The GAM functional does acceptably well in these tests. We conclude that the GAM functional we designed is transferable to molecular problems outside our training sets.
The linear coefficients optimized for GAM are in a narrow range of magnitude so there is no excessive cancellation between terms. The self-consistent-field convergence of the GAM functional has been tested against more than one thousand data; only one of them shows some convergence problems. The enhancement factor plot of the GAM functional is reasonably smooth.
With all these advantages over the GGAs and the previous NGA, with the advantage of an NGA requiring smaller grids than meta-GGAs or meta-NGAs, and with the advantage of an NGA requiring considerably less computation time for extended systems than hybrid functionals, we expect the GAM functional to be very useful for molecular catalysis and a wide variety of other applications to large and complex molecular systems.
Footnote |
† Electronic supplementary information (ESI) available: A complete description of Database 2015 and the geometries and basis sets used with it. See DOI: 10.1039/c5cp01425e |
This journal is © the Owner Societies 2015 |