Alastair J. A.
Price
a,
Alberto
Otero-de-la-Roza
*b and
Erin R.
Johnson
*a
aDepartment of Chemistry, Dalhousie University, 6274 Coburg Rd, Halifax, B3H 4R2, Nova Scotia, Canada. E-mail: erin.johnson@dal.ca
bDepartamento de Química Física y Analítica and MALTA-Consolider Team, Facultad de Química, Universidad de Oviedo, Oviedo 33006, Spain. E-mail: aoterodelaroza@gmail.com
First published on 15th December 2022
Molecular crystals are important for many applications, including energetic materials, organic semiconductors, and the development and commercialization of pharmaceuticals. The exchange-hole dipole moment (XDM) dispersion model has shown good performance in the calculation of relative and absolute lattice energies of molecular crystals, although it has traditionally been applied in combination with plane-wave/pseudopotential approaches. This has limited XDM to use with semilocal functional approximations, which suffer from delocalization error and poor quality conformational energies, and to systems with a few hundreds of atoms at most due to unfavorable scaling. In this work, we combine XDM with numerical atomic orbitals, which enable the efficient use of XDM-corrected hybrid functionals for molecular crystals. We test the new XDM-corrected functionals for their ability to predict the lattice energies of molecular crystals for the X23 set and 13 ice phases, the latter being a particularly stringent test. A composite approach using a XDM-corrected, 25% hybrid functional based on B86bPBE achieves a mean absolute error of 0.48 kcal mol−1 per molecule for the X23 set and 0.19 kcal mol−1 for the total lattice energies of the ice phases, compared to recent diffusion Monte-Carlo data. These results make the new XDM-corrected hybrids not only far more computationally efficient than previous XDM implementations, but also the most accurate density-functional methods for molecular crystal lattice energies to date.
Having a method that is able to rank molecular crystal structures accurately is essential for crystal structure prediction (CSP) – the prediction of the crystal structure of a compound from its molecular diagram only.6,16,17 A reliable CSP protocol would be extremely useful in the disciplines listed above, as it would allow circumventing experimental solid-form screening processes, which are expensive and time-consuming,18–22 and would provide a detailed energy–structure–function map for the chosen molecule and property of interest.11,12 To gauge progress in the field, the Cambridge Crystallographic Data Centre (CCDC) periodically runs CSP blind test competitions in which participant groups try to predict the observed crystal structures of a few molecular compounds.23–28 The 5th blind test, held in 2011, showed that final ranking of the candidate structures using dispersion-corrected DFT is quite effective, and far superior to force fields in most cases,27,29–32 and this conclusion was further supported by the 6th blind test.28 Although other techniques such as fragment-based methods,2,33–37 wavefunction theory,38,39 and machine-learning methods37,40 have been used, DFT is arguably the current workhorse for modeling molecular materials.17,41–54
Dispersion-corrected functionals based on the exchange-hole dipole moment (XDM) model,55–58 in particular the semilocal functional B86bPBE-XDM,59,60 have shown excellent performance for description of molecular crystals44,45,57,61 and non-covalent interactions in general.62,63 In its current plane-wave/pseudopotentials implementation, while still effective for CSP, B86bPBE-XDM is affected by outstanding drawbacks shared by all semilocal functionals. First, the use of a plane-wave basis set makes the computational requirements scale significantly with system size, such that calculations involving unit cells with hundreds to thousands of atoms are on the verge of being infeasible. Second, GGA functionals spuriously over-stabilize systems affected by delocalization error,64–66 which negatively impacts the modeling of molecular salts, acid–base co-crystals, hydrogen bonding, and halogen bonding, to list only a few examples.44,67–69 Lastly, GGA functionals give a poor description of conformational energies, which are important when comparing crystal polymorphs composed of flexible molecules.45,70–72 Notably, several studies have demonstrated the poor performance of B86bPBE-XDM for relative lattice energies in cases where delocalization error is prevalent,44,45,70–72 emphasizing the need for hybrid DFT.
In this work, we address these shortcomings by combining XDM functionals with the numerical atomic orbital (NAO) basis sets in the Fritz Haber Institute ab initio materials simulations (FHI-aims) package.73–76 FHI-aims offers near linear scaling with system size for self-consistent DFT calculations73,74 and enables relatively inexpensive use of hybrid functionals,75 compared to plane-wave approaches. This is important because hybrid functionals can be used to mitigate delocalization error66,68,77–81 and are generally more accurate than GGAs for conformational energies.72 One drawback of NAOs is the possible appearance of basis-set incompleteness error (BSIE), which is known to have a deleterious effect on the description of non-covalent interactions,82–84 although we show that BSIE can be effectively mitigated by parametrization of the dispersion damping function. Dispersion-corrected DFT methods with NAOs have been applied to molecular crystals in combination with the Tkachenko–Scheffler (TS)85 and many-body dispersion (MBD)86,87 family of corrections.49–53
To assess the new XDM-corrected hybrid functionals, we focus on molecular crystal lattice energies, as they are the key property for CSP ranking88 and one of the most demanding tests for computational methods regarding non-covalent interactions.2 The lattice energy of a molecular crystal is the energy required to separate the crystal at its equilibrium geometry into its component molecules. This is an essential parameter when assessing the accuracy of computational methods for modeling molecular crystals because the lattice energy is determined by a delicate balance between intermolecular and intramolecular interactions.2 The accurate calculation of lattice energies is also a stricter performance test for computational methods than energy differences between molecular crystal pairs because the benefits from error cancellation are minimized, while longer-range interactions and many-body effects become far more important. Here, we consider the lattice energies of the X23 set35,38,39,57,89,90 and of 13 ice phases, for which diffusion Monte Carlo (DMC) data has been generated.91 The latter is a particularly stringent test because determining accurate lattice energies for ice relies on a fine balance of dispersion, electrostatic, and many-body induction effects. At present, there is no functional that gives a good description of the absolute and relative energies of all ice phases,91 and therefore the reliable treatment of water and ice with DFT methods remains an unsolved problem.91–93
Herein, we show that the NAO implementation of XDM-corrected functionals provides excellent performance for the description of molecular dimers, ice, and molecular crystal lattice energies in general, with high computational efficiency. In particular, a composite method combining B86bPBE-XDM and its sequent 25% hybrid functional achieves mean absolute errors (MAEs) for the X23 and ice lattice energies of 0.48 kcal mol−1 and 0.19 kcal mol−1, respectively. For the X23, the reported MAE is roughly half the previous best value, making the new XDM methods the most accurate DFT approaches for modeling of molecular materials currently available.
(1) |
E = Ebase + EXDM. | (2) |
In eqn (1), i and j run over atoms, Rij are the interatomic distances, Cn,ij are the dispersion coefficients, and the RvdW,ij are damping lengths calculated as
RvdW,ij = a1Rc,ij + a2, | (3) |
(4) |
The a1 and a2 parameters are the damping function coefficients, which are determined for every functional and basis set combination by minimizing the root-mean-square percent error in binding energies for 49 small molecular dimers, relative to high-level reference data (the Kannemann–Becke set,77,94,95 KB49). The damping function is therefore used to match the XDM dispersion contribution to the particular exchange-repulsion behavior of the chosen functional, as well as to mitigate any (moderate) BSIE from an incomplete basis set. BSIE generally causes some overbinding of intermolecular dimers provided reasonable basis sets of at least double-ζ quality are used. Therefore, if the dispersion energy is damped slightly more strongly for an incomplete basis set, the overall binding energies can provide a good approximation to basis-set limit results. Importantly, once the a1 and a2 parameters are determined, they remain the same for every system to which the functional and basis set are applied, molecular or periodic, and are never re-parametrized for specific cases.
The dispersion coefficients in eqn (1) (Cn,ij) are calculated non-empirically from the self-consistent electron density, its derivatives, and the kinetic energy density. It has been shown that the dependence of these coefficients on the chemical environment (the electronic many-body dispersion effects) is essential to the accuracy of the XDM method.96 Calculation of three-body and higher-order dispersion coefficients, of which the Axilrod–Teller–Muto (C9) is the leading term, is possible in XDM,97 but we have found that including this term has either little impact or degrades the accuracy of XDM-corrected functionals.96
The performance of an XDM-corrected method depends critically on the base functional with which it is paired. In this article, we consider two generalized-gradient-approximation (GGA) functionals: PBE,60 due to its popularity in the solid-state community, and B86bPBE,59,60 which is our GGA functional of choice when non-covalent interactions are dominant, thanks to its ability to accurately describe non-bonded repulsion.57,62,95,98,99 In addition, we consider multiple hybrid density functionals with exchange-correlation (XC) energies of the form
EXC = (1 − aX)EPBE/B86bX + aXESDX + EPBEC. | (5) |
The memory requirements of hybrid functional calculations with the tight basis set exceeded our current computational resources, so we approximated the hybrid/tight result using a correction calculated by evaluating the energy difference between tight and light bases at the GGA level:
Etighthybrid ≈ Elighthybrid + (EtightGGA − ElightGGA). | (6) |
This type of basis-set correction is analogous to using the difference between large- and small-basis MP2 energies to correct small-basis CCSD(T) energies, as in common practice in wavefunction theory calculations.103 In addition to the XDM-corrected functionals mentioned above, we also considered the Tkachenko–Scheffler (TS)85 and many-body dispersion (MBD)86,87 methods for comparison, since they are already implemented in FHIaims and are routinely used for molecular crystals and CSP.49–53 In the case of MBD, we used MBD@rsSCS87 as recommended by the FHI-aims documentation. In the rest of the article, MBD@rsSCS is referred to simply as MBD.
All calculations for the KB49,77,94 S22×5,104 and S66×8105,106 benchmarks of gas-phase dimer binding energies, as well as the 3B-69107 set of three-body interaction energies in molecular trimers, were carried out as single-point energy evaluations at the literature geometries. This is standard for these benchmark sets and is done to facilitate direct comparison with the CCSD(T) reference data. Since the S22×5, S66×8, and 3B-69 all contain dimer and trimer geometries far from equilibrium, geometry optimization would be meaningless for these systems. Conversely, full geometry optimizations were performed with each functional on both the molecular crystals and isolated molecules forming the X23 set89,90,108 of lattice energies. The geometries of the 13 ice polymorphs forming the ICE13 set91 were also fully optimized, although the geometry of the isolated water molecule was kept fixed, as described in ref. 91, for consistency. For the crystals, reciprocal-space k-point grids were selected with the number of points, n1 × n2 × n3, given by
ni = int[max(1,Rk|bi| + 0.5)], | (7) |
Functional | a X | a 1 | a 2 (Å) | MAE | MAPE |
---|---|---|---|---|---|
a The mean absolute errors (MAE, in kcal mol−1) and mean absolute percent errors (MAPE) for the KB49 fit set are also shown. The best overall results for each basis set are indicated in bold. b This value is the range-separation parameter (ω) instead of the exact-exchange fraction. c The optimal a1 value was negative, so it was set to zero during the parametrization. | |||||
Light basis set | |||||
PBE | 0.00 | 0.5312 | 2.3270 | 0.67 | 19.0 |
B86bPBE | 0.00 | 0.8219 | 1.2069 | 0.54 | 14.9 |
HSE06 | 0.11b | 0.3268 | 3.0431 | 0.52 | 13.6 |
PBE0 | 0.25 | 0.3302 | 3.0042 | 0.46 | 12.5 |
PBE-50c | 0.50 | 0.0000 | 4.1971 | 0.38 | 9.8 |
B86bPBE-25 | 0.25 | 0.5235 | 2.1995 | 0.35 | 9.7 |
B86bPBE-50 | 0.50 | 0.0831 | 3.7362 | 0.30 | 8.5 |
Tight basis set | |||||
PBE | 0.00 | 0.6438 | 1.8533 | 0.50 | 14.1 |
B86bPBE | 0.00 | 0.8976 | 0.8518 | 0.38 | 11.0 |
HSE06 | 0.11b | 0.5020 | 2.3000 | 0.46 | 11.1 |
PBE0 | 0.25 | 0.5053 | 2.2527 | 0.41 | 10.2 |
PBE-50 | 0.50 | 0.3983 | 2.5986 | 0.42 | 9.6 |
B86bPBE-25 | 0.25 | 0.6546 | 1.6097 | 0.32 | 8.4 |
B86bPBE-50 | 0.50 | 0.4887 | 2.1855 | 0.36 | 8.5 |
The errors shown in Table 1 are comparable to those obtained with our previous plane-wave (Quantum ESPRESSO57,109) and Gaussian basis-set (using Gaussian110 or psi4111 with the postg program112) results contained in the current XDM parametrization database.113 For example, the MAPE for B86bPBE/tight in Table 1 (11.0%) is very close to the MAPE obtained for the same functional using the projector augmented wave (PAW) method114 (11.8%), plane waves plus norm-conserving pseudopotentials (12.4%), and the aug-cc-pVTZ Gaussian basis set (11.4%). The MAPEs obtained with other functionals, such as PBE, PBE0, or HSE06, also deviate from those in the parametrization database by around 1% at most. This is a strong indication that our FHIaims XDM implementation is working correctly.
Focusing on the results for the tight basis set, Table 1 shows that hybrid functionals outperform GGAs, and that B86b-based functionals consistently give lower errors than the analogous PBE-based functionals. This is also in agreement with our previous works.57,77 The lowest errors among the functionals studied are obtained for the B86bPBE-25 hybrid, with a MAE of 0.32 kcal mol−1 and a MAPE of 8.4%.
Because the tight basis set is too expensive for routine geometry optimizations, we resort to using the smaller, light basis set and relying on the XDM damping function to partially alleviate any BSIE.77,82Table 1 also shows the average errors for the light (double-ζ) basis set. While lower errors are obtained with the tight (triple-ζ) basis set, the good performance of the light basis set indicates a reasonably low impact on the accuracy caused by BSIE. This is in stark contrast to our previous results using the double-ζ basis set in the SIESTA NAO code,115,116 where the MAPE was in the 20% to 30% range and could not mitigated by using counterpoise corrections. The differences between FHI-aims and SIESTA are likely due to the different strategies employed by their developers for NAO construction (see ref. 73 and 115 for details). The small magnitude of the BSIE with FHI-aims can be confirmed by comparing the dispersion-uncorrected binding energies calculated with the light and tight basis sets using the same functional. For example, the mean absolute difference between the light and tight binding energies obtained using B86bPBE (without XDM) is 0.32 kcal mol−1, with individual errors not exceeding 0.75 kcal mol−1.
The S22×5 and S66×8 error statistics for the various combinations of functional, basis set, and dispersion correction are shown in Table 2. As for the KB49 set, the average errors are lower for the tight basis set, and hybrid functionals slightly outperform GGAs regardless of the dispersion correction employed. The XDM values in the table are also similar to those reported for the same benchmarks using the aug-cc-pVTZ Gaussian basis set.58 While all basis set, functional, and dispersion method combinations perform generally well, B86bPBE-25-XDM consistently gives the lowest errors by a small margin, with MAEs in the range of 0.2–0.4 kcal mol−1.
Functional | Dispersion | S22×5 | S66×8 | 3B-69 | |||
---|---|---|---|---|---|---|---|
Correction | Light | Tight | Light | Tight | Light | Tight | |
a The best overall results in each column are indicated in bold. | |||||||
PBE | TS | 0.57 | 0.39 | 0.60 | 0.38 | 0.078 | 0.080 |
HSE06 | TS | 0.63 | 0.45 | 0.64 | 0.38 | 0.046 | 0.042 |
PBE0 | TS | 0.58 | 0.42 | 0.59 | 0.33 | 0.044 | 0.039 |
PBE | MBD | 0.55 | 0.44 | 0.44 | 0.28 | 0.113 | 0.113 |
HSE06 | MBD | 0.53 | 0.48 | 0.45 | 0.29 | 0.069 | 0.066 |
PBE0 | MBD | 0.50 | 0.46 | 0.40 | 0.26 | 0.060 | 0.055 |
PBE | XDM | 0.58 | 0.44 | 0.45 | 0.29 | 0.101 | 0.099 |
B86bPBE | XDM | 0.46 | 0.34 | 0.35 | 0.20 | 0.050 | 0.052 |
HSE06 | XDM | 0.52 | 0.45 | 0.41 | 0.28 | 0.054 | 0.055 |
PBE0 | XDM | 0.49 | 0.43 | 0.38 | 0.25 | 0.044 | 0.045 |
PBE-50 | XDM | 0.47 | 0.47 | 0.37 | 0.28 | 0.047 | 0.030 |
B86bPBE-25 | XDM | 0.39 | 0.35 | 0.30 | 0.19 | 0.037 | 0.040 |
B86bPBE-50 | XDM | 0.40 | 0.41 | 0.32 | 0.24 | 0.055 | 0.051 |
Beyond-pairwise intermolecular interactions are also important in molecular crystals, since they represent a small but significant fraction of the total lattice energy.2,117 For this reason, we consider as an additional benchmark the 3B-69 set of molecular trimers.107 In this case, the reference data corresponds to the difference between the trimer binding energy and the pairwise sum of the constituent dimer binding energies. This is a good measure of whether the considered methods can describe non-additive many-body intermolecular interactions96 and, as such, highlight their performance in the treatment of beyond-pairwise effects. Table 2 shows that BSIE has less impact on the three-body energies than it does for the pairwise binding energies, with the light and tight MAEs being approximately the same.
MBD might be expected to be the most accurate dispersion correction for the 3B-69 benchmark due to the many-body nature of the interactions. However, we observe that all three dispersion methods provide roughly comparable performance, with XDM being slightly superior to MBD for the same functional and basis set combination. Instead, it is the choice of base functional that is the determining factor, with PBE consistently giving the largest errors, while use of either B86b or exact exchange improves performance in the treatment of three-body interactions. This confirms our previous observation that the choice of base functional is critical for accurate treatment of beyond-pairwise non-covalent interactions.118 Overall, XDM paired with either the B86bPBE-25 or PBE-50 hybrid functionals (depending on basis set) gives the lowest MAE. The fact that XDM (which does not incorporate a three-body dispersion contribution) outperforms MBD (which does) for the description of three-body intermolecular interactions suggests that electronic many-body effects are much more important than the atomic many-body dispersion effects encapsulated by the Axilrod–Teller–Muto term.96
Functional | Dispersion | Light | Tight |
---|---|---|---|
a Literature value obtained from ref. 87. b The hybrid energies with the light settings are corrected using the difference between light and tight results at the GGA level (viaeqn (6)). | |||
Full geometry relaxation | |||
PBE | TS | 4.17 | 3.14 |
HSE06 | TS | 4.57 | — |
PBE0 | TS | 4.44 | 2.39a |
PBE | MBD | 1.61 | 0.94 |
HSE06 | MBD | 2.12 | — |
PBE0 | MBD | 1.98 | 0.84a |
PBE | XDM | 1.14 | 1.04 |
HSE06 | XDM | 1.20 | — |
PBE0 | XDM | 1.14 | — |
PBE50 | XDM | 1.25 | — |
B86bPBE | XDM | 0.83 | 0.72 |
B86bPBE-25 | XDM | 0.81 | — |
B86bPBE-50 | XDM | 1.06 | — |
Single points at GGA/light geometries | |||
PBE0//PBE | MBD | 1.97 | 1.07b |
PBE0//PBE | XDM | 1.01 | 0.96b |
PBE-50//PBE | XDM | 1.00 | 0.87b |
B86bPBE-25//B86bPBE | XDM | 0.66 | 0.48 |
B86bPBE-50//B86bPBE | XDM | 0.70 | 0.53b |
The table is separated into two sections, with the upper part showing results obtained with full geometry optimization of the molecular crystals at each listed level of theory. As noted in the computational methods section, we were only able to perform calculations using the tight basis set for GGA functionals due to the high memory requirements for hybrids. However, literature results87 for PBE0-TS and PBE0-MBD with tight settings (which used the earlier X23 reference data89) are provided as these combinations give the lowest MAEs obtained with each of these dispersion corrections. We note that updating the reference data causes the MAEs to change by at most 0.25 kcal mol−1, although often the deviation is lower.
As observed previously,87,108 TS massively overbinds these molecular crystals. With TS and MBD, there is a significant difference between the light and tight results, which occurs because the damping parameters within these dispersion corrections are not optimized for each basis set independently. As a result, XDM significantly outperforms MBD with the light basis set, although the two methods give comparable MAEs with the tight basis. Also, the PBE-XDM and B86bPBE-XDM MAEs with light are in excellent agreement with previous results obtained using the Quantum ESPRESSO plane-wave code.62
B86bPBE-XDM with the tight basis set yields the lowest MAE (0.72 kcal mol−1) yet obtained for the X23 set with any dispersion-corrected GGA, although this is largely due to the improvement in the reference data (the MAE compared to the values in ref. 89 is 0.90 kcal mol−1). For comparison, the MAE for PBE0-MBD in Table 3 is 0.84 kcal mol−1 and the lowest MAE reported by Thomas et al.122 for the X23 is 0.81 kcal mol−1, obtained with the TPSS-D3 dispersion-corrected meta-GGA. (The D3 dispersion correction by Grimme et al. is not available in FHIaims, so a direct comparison was not possible.) The best GGA results given by Thomas et al.122 are 0.93 kcal mol−1 (B86bPBE-XDM with plane-waves, a slightly outdated XDM implementation, and only for the C21) followed by PBE-D3 at 0.98 kcal mol−1. Regarding the dispersion-corrected hybrid functionals, the best results are obtained with B86bPBE-25/light (0.81 kcal mol−1 with either the Dolgonos et al.90 or the older Reilly et al.89 reference data) followed by PBE0-MBD/tight (0.84 kcal mol−1), with the former being considerably more efficient. For comparison, Thomas et al.122 report MAEs of 0.93 kcal mol−1 for PBE0-MBD at the PBE-TS optimized geometries and 1.03 kcal mol−1 for PBE0-D3.
Computational efficiency is an important consideration in CSP, where hundreds to thousands of candidate crystal structures must be ranked with DFT for a given compound. Composite approaches, in which a relatively low level of theory is used for geometry optimization, followed by single-point energy evaluation at a higher level of theory, are an excellent strategy to reduce the computational cost without losing accuracy.46,123 In this work, we consider composite approaches that use dispersion-corrected GGA functionals (PBE-MBD, PBE-XDM, or B86bPBE-XDM) and the light basis set for geometry optimization. Single-point energies are then evaluated with the corresponding 25% or 50% hybrid functionals and the light basis set and, in some cases, also with the same GGA and the tight basis set. This allows us to obtain energies (viaeqn (6)) with an accuracy comparable to what would be expected from full hybrid/tight calculations, but with a drastically reduced computational cost. MAEs in the X23 lattice energies obtained using this type of composite approach are shown in the lower portion of Table 3. The notation in the table is high-level (hybrid)//low-level (GGA).
The MAEs obtained with the composite approach using B86bPBE-25-XDM and B86bPBE-50-XDM are the lowest errors yet obtained for the X23 set with any DFT method. The composite B86bPBE-25-XDM//B86bPBE-XDM approach with basis-set correction gives an MAE of only 0.48 kcal mol−1, well below the usual target of 1 kcal mol−1 deemed to be chemical accuracy and almost exactly on the 2 kJ mol−1 mark commonly cited as the average energy difference between polymorphs.124 It is also reassuring that the average error for molecular crystals is similar to that for dimers formed from molecules with similar sizes (Table 2), for which the MAEs were in the 0.2 kcal mol−1 to 0.4 kcal mol−1 range.
While good performance for absolute lattice energies is highly desirable, it does not necessarily ensure reliable polymorph ranking, which is dependent on accurate lattice-energy differences (as well as thermal and kinetic factors). The performance of the proposed methods for relative lattice energies will be examined in detail elsewhere. Nonetheless, improvements in absolute lattice energies do tend to result in more accurate relative lattice energies, as seen for the two oxalic acid polymorphs (α and β forms) appearing in the X23 set. The choice of dispersion correction has only a minor effect on this energy difference, although the polymorph ordering is highly dependent on exact-exchange mixing in the base functional. As shown in Fig. 1, the GGAs predict the incorrect energy ordering and relatively large fractions of exact exchange (near 50%) are needed to recover the reference lattice-energy difference. This suggests that delocalization error is a factor in determining the most stable oxalic acid polymorph, with the β form likely favored by GGAs due to its dimeric hydrogen-bonding.
A strict test of density functionals and dispersion corrections for water is calculation of the absolute lattice energies of the various (ordered) ice phases. Different ice phases vary in molecular arrangements and in the extent of electron delocalization, which has been shown to correlate with the absolute lattice energy.125 In a recent work, Della Pia et al.91 reported absolute lattice energies of 13 ordered ice phases calculated using Diffusion Monte Carlo (DMC), and subsequently benchmarked a number of functionals using a plane-wave approach. Relative to the X23, this set has the advantage that no vibrational or nuclear quantum effects need to be removed before comparison to DFT results. We now use this ICE13 set, which is a superset of the previous ICE10 set proposed by Brandenburg et al.,92 to evaluate the performance of our XDM-corrected methods.
Table 4 shows the MAEs calculated for the absolute and relative lattice energies of the ICE13 set with respect to the DMC data. The MAE of the relative energies is calculated by considering all 78 pairs of crystals in the ICE13 set, to avoid singling out any particular ice phase. As in the case of the X23, the MAEs with the tight basis set are lower (in most cases) than with light, both for the absolute and for the relative lattice energies. XDM outperforms TS and MBD for absolute lattice energies by around 0.4–0.5 kcal mol−1, but gives higher errors by a few tenths of a kcal mol−1 for the relative lattice energies. The average errors from the GGA functionals are quite high, in the vicinity of 2 kcal mol−1 for the absolute lattice energies and ca. 0.5 kcal mol−1 or more for the relative lattice energies. Hybrids give improved results, providing another indication that the cooperative hydrogen bonding networks in ice exhibit considerable delocalization error.
Functional | Dispersion | Absolute | Relative | ||
---|---|---|---|---|---|
Correction | Light | Tight | Light | Tight | |
a The hybrid energies with the light settings are corrected using the difference between light and tight results at the GGA level (viaeqn (6)). | |||||
Full geometry relaxation | |||||
PBE | TS | 3.69 | 2.18 | 0.56 | 0.51 |
HSE06 | TS | 2.68 | — | 0.40 | — |
PBE0 | TS | 2.38 | — | 0.37 | — |
PBE | MBD | 3.70 | 2.19 | 0.66 | 0.60 |
HSE06 | MBD | 2.68 | — | 0.40 | — |
PBE0 | MBD | 2.38 | — | 0.37 | — |
PBE | XDM | 2.79 | 1.71 | 0.91 | 0.73 |
HSE06 | XDM | 1.60 | — | 0.72 | — |
PBE0 | XDM | 1.35 | — | 0.65 | — |
PBE50 | XDM | 0.79 | — | 0.46 | — |
B86bPBE | XDM | 2.69 | 1.78 | 0.67 | 0.45 |
B86bPBE-25 | XDM | 1.16 | — | 0.53 | — |
B86bPBE-50 | XDM | 0.55 | — | 0.42 | — |
Single points at GGA/light geometries | |||||
PBE0//PBE | MBD | 2.13 | 0.61a | 0.34 | 0.29a |
PBE0//PBE | XDM | 1.11 | 0.30a | 0.61 | 0.43a |
PBE-50//PBE | XDM | 0.25 | 1.16a | 0.35 | 0.21a |
B86bPBE-25//B86bPBE | XDM | 0.93 | 0.19 | 0.49 | 0.28a |
B86bPBE-50//B86bPBE | XDM | 0.32 | 1.20a | 0.31 | 0.19 |
While we could not run the hybrid calculations with the tight basis set, the light results indicate that 25% hybrid functionals reduce the MAE, and 50% hybrids reduce it even further. However, the statistics for the composite methods in Table 4 allow us to understand the effects of BSIE and the incorporation of exact exchange separately, and reveal that this may be ascribed to error cancellation. While the results with the 25% hybrids improve when the basis-set correction of eqn (6) is added, the 50% hybrid functionals perform better with the light basis set and no BSIE correction. This suggests that, when half-and-half functionals are used for water, there is error cancellation between delocalization error and BSIE.
As for the X23, the best-performing method for ice is found to be the composite approach using B86bPBE-25-XDM with the additive BSIE correction, which yields MAEs of 0.19 kcal mol−1 and 0.28 kcal mol−1 for the absolute and relative lattice energies, respectively. This absolute lattice-energy error is lower than the MAEs of all functionals studied by Della Pia et al.,91 and the relative-energy error is also among the best. For comparison, the best-performing functional reported91 for absolute lattice energies is the revPBE-D3 GGA, with a MAE of 0.22 kcal mol−1, and the best functionals in each of the other classes are rSCAN (meta-GGA, 0.23 kcal mol−1), vdw-DF2 (non-local, 0.32 kcal mol−1), and revPBE0-D3 (hybrid, 0.39 kcal mol−1). Naturally, all these functionals are well within the “good functional” category established by the authors (MAEs <0.96 kcal mol−1 and <0.48 kcal mol−1 for absolute and relative lattice energies, respectively).
To test the accuracy of the new XDM-corrected hybrid functionals, we assessed their performance for binding energies of molecular gas-phase dimers and trimers, as well as lattice energies of small molecular crystals (the X23 set) and 13 phases of ice (the ICE13 set). The results were compared to the Tkachenko–Scheffler (TS) and state-of-the-art many-body dispersion (MBD@rsSCS) methods. For molecular dimers, XDM-corrected functionals achieve a mean average error (MAE) of between 0.2 and 0.4 kcal mol−1, slightly outperforming TS and MBD. More importantly, XDM-corrected functionals also show excellent performance for three-body interaction energies (the 3B-69 set), suggesting that electronic many-body effects are much more important than atomic many-body dispersion effects, which are not included in the canonical XDM methods.
The XDM-corrected methods also yield very low average errors for the X23 set of lattice energies, particularly if hybrid methods or relatively large (“tight”) basis sets are used. The most intriguing result is the spectacular performance of composite methods, in which a GGA geometry optimization (e.g. B86bPBE-XDM) is followed by a single-point energy calculation to incorporate the benefits of using a hybrid functional (e.g. B86bPBE-25-XDM), and perhaps an additional single-point correction to treat basis-set incompleteness error (the difference between tight and light energies at the GGA level). The best-performing composite method (B86bPBE-25-XDM with basis-set correction, at the B86bPBE-XDM equilibrium geometries) achieves a MAE of only 0.48 kcal mol−1 for the X23 set, roughly half the error of other similar DFT methods. Moreover, this composite approach can be routinely applied to molecular crystals containing as many as 1000 atoms within the unit cell.
The excellent performance of the basis-corrected B86bPBE-25-XDM//B86bPBE-XDM composite method extends to the calculation of the absolute lattice energies of 13 ice phases, for which it achieves an MAE of only 0.19 kcal mol−1, outperforming all DFT functionals reported to date. The calculation of absolute lattice energies of ice is particularly difficult due to the presence of delocalization error and the delicate balance between electrostatics, dispersion, and induction. It is an key point that one single methodology works well for molecular dimers and trimers, and achieves the lowest MAE for both the X23 and ICE13 lattice energies. Obtaining good across-the-board performance in all tests examined is of paramount importance when modeling complex materials that feature several disparate types of non-covalent interactions. This makes us confident that the proposed XDM-corrected methods will serve nicely for accurate energy ranking in crystal structure prediction.
Footnote |
† Electronic supplementary information (ESI) available: Tables of computed lattice energies and example FHI-aims control.in files. See DOI: https://doi.org/10.1039/d2sc05997e |
This journal is © The Royal Society of Chemistry 2023 |