Deborah L.
Crittenden
*
School of Physical and Chemical Sciences, University of Canterbury, Christchurch, 8140, New Zealand. E-mail: deborah.crittenden@canterbury.ac.nz
First published on 28th October 2022
Using an extensive set of benchmark geometric data and absolute NMR chemical shieldings for gas phase molecules compiled from the literature, we introduce and test a novel shift-and-scale correction scheme for bringing computed values into closer agreement with experiment. Our approach requires only one additional reference calculation relative to existing methods for computing chemical shifts, and all reference species are small molecules whose absolute nuclear shieldings can be easily obtained, both computationally and experimentally. We demonstrate that our approach is capable of correcting for errors due to choice of electronic structure method and basis set, vibrational averaging effects, and scalar relativistic effects, but cannot account for the influence that heavy atoms have on the chemical shieldings of neighbouring light atoms, via spin–orbit coupling. We particularly recommend using our approach in conjunction with nuclear shieldings computed at DFT optimized geometries, employing hybrid functionals with moderate-to-high quality atomic orbital basis sets, e.g. pc-2.
Given the utility and widespread use of NMR spectroscopy in determining the identity of small molecules, it is surprising that spectral assignment remains by and large a manual process. This can be traced back to the fact that simple heuristic models are very effective,1 and predicting NMR parameters to high enough accuracy for qualitative assignment of experimental spectra is computationally challenging.2,3 Spin–spin coupling constants can predicted to sufficient accuracy (<10% relative error) using computationally affordable density functional theory models with modestly sized atomic orbital basis sets,4,5 but chemical shieldings are more challenging.2,6
For example, previous benchmarking studies6 have shown that in order to obtain absolute chemical shifts for gas phase molecules to within 5 ppm (mean absolute deviation) or ∼15% (mean relative deviation) of experimental values (across 28 molecules with measured shieldings primarily available for 1H, 13C, 17O, 19F nuclei), it is necessary to employ highly correlated wavefunction models – CCSD(T) – in conjunction with large and flexible segmented-valence atomic orbital basis sets such as Dunning's correlation-consistent polarizable core–valence basis sets of at least triple zeta quality, i.e. cc-pCVXZ, where X = T or Q. It is also necessary to account for vibrational-averaging effects, which would otherwise contribute additional errors of ∼5 ppm or 25%. By way of comparison, DFT models predictions typically incurred average absolute errors around 20 ppm and average absolute relative errors of ∼50%.
Similarly, relativistic effects cannot be ignored,7 even for hydrogen atoms, whose chemical shifts are influenced by neighbouring heavy atoms through spin–orbit coupling.8 For example, relativistic corrections to 1H NMR chemical shieldings range from 0.15 ppm for the hydrogen nucleus in hydrogen fluoride to 18.40 ppm for the hydrogen nucleus in hydrogen iodide.9 For heavy atoms, scalar relativistic effects dominate and are even larger in magnitude.9 For halogen nuclei within halogen hydrides, total relativistic corrections range from 8.9 ppm (HF) to 2060.8 ppm (HI).9
Of course, predicting absolute chemical shieldings is a much harder task than predicting relative chemical shifts, which are generally the more experimentally relevant quantity. In this work, we take this idea one step further and posit that by uniformly shifting and scaling computational absolute shielding values obtained using computationally inexpensive electronic structure methods to exactly reproduce experimental values for two carefully chosen reference molecules, it may be possible to account not only for methodological errors that are constant across all molecules but also methodological errors that vary systematically between molecules.
Although this sounds simple in theory, in practice it is complicated by the vast array of computationally-affordable electronic structure models that could be tried (the density functional theory “zoo”10) and multiple potential sources of error:
• Electronic: incomplete, approximate or absent treatment of electron correlation, basis set incompleteness.
• Geometric: resultant errors in equilibrium geometries, vibrational-averaging effects.
• Relativistic: spin–orbit and scalar relativistic effects that cannot be fully accounted for outside a full four-component relativistic Dirac equation framework.
Therefore, it is necessary to be as systematic and selective as possible. “Computational error bars” for DFT predictions may be established using Hartree–Fock theory and Hartree–Fock–Slater density functional theory.11 Despite both lacking any attempt at modelling electron correlation, these simple methods nonetheless exhibit systematic electron under- and over-localisation errors, respectively, and have proven effective as “bounding” functionals for molecular property estimation.11
An alternative strategy is to select density functionals from the literature that have been shown to be particularly accurate for predicting NMR chemical shieldings, such as the modified B3LYP variant with 5% HF exchange recommended by Handy et al.12,13 Alternatively, Keal and Tozer have developed density functionals designed specifically for predicting NMR shifts and shieldings,14,15 of which the KT3 functional has also been optimized for other molecular properties including atomization energies, ionization potentials, electron affinities, proton affinities, geometric parameters, vibrational frequencies and electronic polarizabilities.15
When it comes to selecting basis sets for NMR shielding calculations, it is necessary to choose high quality basis sets that can accurately describe electron densities close to nuclear centres, i.e. “conventional” highly-contracted basis sets optimized for computing valence properties generally do not suffice.16 Polarization consistent (pc-n) and segmented polarization consistent (pcS-n) basis sets are particularly appropriate, because they converge rapidly with respect to n.17,18 Even at n = 2, basis set incompleteness errors are substantially lower than those associated with the choice of electronic structure method.17,18
Geometric errors may arise from the choice of electronic structure model and neglecting to account for vibrational-averaging effects. Fortunately, these sources of error can be accounted for separately. As proposed by Vuckovic,19 a simple way of quantifying geometric model-related errors is to compare property values computed at an “exact” reference equilibrium geometry with those computed at the optimized geometry at a given level of theory. Accounting for vibrational averaging effects at each level of theory is harder and much more computationally intensive, because it requires higher order derivatives of the energy and nuclear shieldings with respect to atomic displacements.20 A more feasible approach is to derive anharmonic corrections at a single level of theory and then use these to “correct” the vibrationally-averaged experimental values to obtain semi-experimental equilibrium shieldings, which can then be compared directly with computed values.6 This approach can be justified on the basis that anharmonic corrections are small in magnitude relative to overall shieldings,6 so methodological errors are expected to be negligible, except perhaps in cases where the chosen level of theory yields qualitatively incorrect potential energy curves. For example, Hartree–Fock theory provides a qualitatively incorrect description of bonding within F2.3
Relativistic effects are the most challenging to account for, because full 4-component relativistic Dirac calculations are required for quantitative prediction of absolute NMR chemical shieldings.21,22 These are more computationally intensive than scalar and 2-component relativistic models,23–25 and are not as widely implemented in molecular quantum chemistry programme packages.
Molecule | Value | Ref. | Molecule | Value | Ref. |
---|---|---|---|---|---|
HBr | 35.037 | 34 | CH3NH2 | 28.305 | 34 |
HCl | 31.124 | 34 | CH3Cl | 27.932 | 34 |
NH3 | 30.727 | 34 | HCN | 27.78 | 35 |
CH3OH | 30.671 | 34 | SiH4 | 27.625 | 34 |
CH4 | 30.633 | 34 | CH3OH | 27.350 | 34 |
H2S | 30.572 | 34 | CH3F | 26.635 | 34 |
CH3NH2 | 30.367 | 34 | H2 | 26.293 | 34 |
H2O | 30.102 | 34 | CH2CCH2 | 26.279 | 34 |
CH3SH | 29.888 | 34 | LiH | 25.7 | 36 |
CH3CH3 | 29.887 | 34 | CH2Cl2 | 25.670 | 34 |
PH3 | 29.761 | 34 | CH2CH2 | 25.463 | 34 |
HCCH | 29.317 | 34 | CH2F2 | 25.291 | 34 |
CH3CN | 29.149 | 34 | CHF3 | 24.543 | 34 |
CH3CO | 28.857 | 34 | CHCl3 | 23.659 | 34 |
CH3SH | 28.831 | 34 | C6H6 | 23.535 | 34 |
HF | 28.51 | 35 | CH3COH | 21.068 | 34 |
CH3Br | 28.328 | 34 | HOF | 18.511 | 37 |
Molecule | Value | Ref. | Molecule | Value | Ref. |
---|---|---|---|---|---|
CH4 | 195.07 | 38 | CH3CN | 74.03 | 38 |
CH3CN | 185.05 | 38 | CHF3 | 68.74 | 38 |
CH3Br | 182.41 | 38 | CF3CF3 | 68.49 | 38 |
CH3SH | 182.34 | 38 | CFCl3 | 67.21 | 38 |
CH3CH3 | 180.84 | 38 | CF4 | 64.48 | 38 |
CH3Cl | 163.84 | 38 | CH2CH2 | 64.43 | 38 |
CH3NH2 | 158.32 | 38 | CO2 | 58.71 | 38 |
CH3COH | 156.81 | 38 | C6H6 | 57.17 | 38 |
CH3OH | 136.51 | 38 | OCS | 30.97 | 39 |
CH3F | 116.75 | 38 | CO | 0.80 | 38 |
HCCH | 116.65 | 38 | H2CO | −0.5 | 40 |
CH2CCH2 | 115.29 | 38 | CH3COH | −6.66 | 38 |
CCl4 | 88.99 | 38 | CS2 | −8.11 | 38 |
HCN | 82.1 | 41 | CH2CCH2 | −29.35 | 38 |
CH2F2 | 77.73 | 38 |
Benchmark equilibrium geometries and geometric parameters were compiled from both the spectroscopic26–30 and quantum chemical31–33 literature. Computed geometries are of at least CCSD(T)/cc-pVQZ quality. Cartesian coordinates for all benchmark equilibrium geometries used in this work are reported in the ESI.†
Category | Options |
---|---|
a A simple hybrid functional comprising 50% exact (HF) exchange and 50% Slater–Dirac exchange. b A modified variant of the B3LYP functional comprising 5% HF exchange and 95% Slater–Dirac exchange (cf. 20% HF and 80% Slater–Dirac exchange in the original). | |
Method | HF, HnHa, B3LYP, B3LYP0.05b, HFS, KT3 |
Basis set | pc-2, pcS-2 |
For the polyatomic molecules in our data set, vibrationally-averaged geometries54 and chemical shieldings20 are computed at HF/pc-2, following the procedure outlined in Ruud et al.,55 Lutnaes et al.56 and Teale et al.6 In brief, the geometry is first optimized and subject to harmonic vibrational analyis. Semi-cubic force constants required to define zero-point vibrational displacements along each normal mode are computed by numerical differentiation, using a step size of 0.0075 a.u. The vibrationally-averaged geometries thus obtained are used as an expansion point for computing vibrationally-averaged shielding constants. The required cubic and quartic derivatives of the energy and second derivatives of the nuclear shielding tensor with respect to displacement along normal mode coordinates are computed numerically, using a step size of 0.05 a.u. We note that these step sizes may not always be optimal in all cases, but will consider this as a potential source of error rather than attempting to optimize these values on a case-by-case basis which would be both computationally prohibitive and may introduce further uncontrolled errors.
For diatomics, accurate and precise equilibrium and vibrationally-averaged bond lengths can be determined directly from spectroscopic data. Therefore, vibrationally-averaged shieldings are computed using these coordinates rather than their HF-optimized equivalents.
Finally, anharmonic corrections are computed as:
δZPVE = σHF/pc-20 − σHF/pc-2e | (1) |
σcorr = mσcalc + b | (2) |
The most statistically robust way of obtaining the required m and b values, that guarantees residual root-mean-squared errors will be minimised across the training data set, is linear regression analysis. However, this is not particularly practical as a general approach, because the parameters thus obtained are not transferable, and so a full calibration procedure must be performed at each intended level of theory.
However, all is not lost. The required scaling parameters may instead be estimated by comparing computed and experimental shieldings using two carefully chosen reference molecules:
(3) |
(4) |
Key considerations in choosing appropriate reference molecules include:
• They should be small.
• They should have significantly different shieldings.
• They should lie on or near the line of best fit one would obtain via a full regression analysis, i.e. be representative of the data set as a whole.
The molecules selected according to these criteria are listed in Table 5.
Nucleus | 1H | 13C | 19F |
---|---|---|---|
Molecule 1 | CH4 | CH4 | HF |
Molecule 2 | H2 | CO | CF4 |
Fig. 1 Computed vs. semi-experimental equilibrium shieldings for 1H (top), 13C (middle) and 19F (bottom) nuclei. |
Model | 1H | 13C | 19F | ||||||
---|---|---|---|---|---|---|---|---|---|
Slope | Int. | R 2 | Slope | Int. | R 2 | Slope | Int. | R 2 | |
HF/pc-2 | 0.9301 | 2.352 | 0.9715 | 1.110 | −18.48 | 0.9708 | 0.987 | 18.0 | 0.9917 |
HF/pcS-2 | 0.8890 | 3.085 | 0.9479 | 1.108 | −17.55 | 0.9712 | 0.989 | 21.9 | 0.9923 |
HnH/pc-2 | 0.9837 | 0.446 | 0.9849 | 1.099 | −22.25 | 0.9900 | 1.057 | −29.5 | 0.9886 |
HnH/pcS-2 | 0.9817 | 0.435 | 0.9853 | 1.110 | −24.74 | 0.9902 | 1.067 | −29.1 | 0.9894 |
B3LYP/pc-2 | 1.0019 | 0.140 | 0.9866 | 1.057 | −23.44 | 0.9848 | 1.075 | −50.2 | 0.9917 |
B3LYP/pcS-2 | 1.0002 | 0.124 | 0.9872 | 1.072 | −27.47 | 0.9864 | 1.088 | −51.5 | 0.9923 |
B3LYP0.05/pc-2 | 1.0282 | −0.751 | 0.9832 | 1.048 | −23.91 | 0.9791 | 1.097 | −66.9 | 0.9888 |
B3LYP0.05/pcS-2 | 1.0264 | −0.762 | 0.9842 | 1.066 | −28.95 | 0.9808 | 1.112 | −69.3 | 0.9896 |
HFS/pc-2 | 1.0666 | −2.376 | 0.9765 | 1.073 | −24.31 | 0.9793 | 1.128 | −82.5 | 0.9775 |
HFS/pcS-2 | 1.0643 | −2.371 | 0.9781 | 1.094 | −30.12 | 0.9796 | 1.146 | −85.9 | 0.9789 |
KT3/pc-2 | 0.9914 | 0.482 | 0.9767 | 0.958 | −3.86 | 0.9747 | 1.023 | −42.9 | 0.9869 |
KT3/pcS-2 | 0.9941 | 0.349 | 0.9782 | 0.957 | −3.33 | 0.9782 | 1.024 | −39.2 | 0.9878 |
Fig. 1 shows that the KT3 functional displays qualitatively different behaviour to the hybrid functionals – it is much more accurate for predicting 13C shielding constants, although it displays similar accuracy for 1H and 19F shieldings. Analysis of the regression parameters in Table 6 illustrates these differences even more clearly. Hybrid functionals display systematic model-dependent trends in both the slope and intercept values according to the proportion of HF exchange in the underlying functional; regression parameters for the KT3 functional do not fit these trends. KT3 functionals tend to yield more accurate predictions of shielding constants (slopes and intercepts closer to 1 and 0, respectively) but display more scatter about their lines of best fit (lower R2 values).
These observations likely reflect differences in parameterisation strategy; hybrid functionals are parameterised to accurately model the energetics of valence electrons57 whereas the KT functionals are designed to accurately reproduce exchange-correlation potentials, especially for electrons within 0.4–1.0 a.u. of each nuclear centre.14 Because valence electron behaviour varies widely across our data set whereas smaller fluctuations are expected in core regions, it stands to reason that hybrid functionals display larger but more systematic errors, whereas the KT3 functional exhibits smaller but less systematic errors. This suggests that hybrid functionals may be more appropriate for our proposed shift-and-scale correction procedure. In particular, it is best to choose “well-balanced” hybrid functionals in which electron under- and over-localisation errors approximately cancel and that are parameterised to model electron behaviour well in a wide range of chemical environments, which will show up as consistently high R2 values across all nuclear centres in this context.
However, from a practical standpoint, it does not make sense to correlate DFT shieldings computed at near-exact equilibrium geometries with semi-experimental equilibrium shieldings, because both of these quantities are hard to obtain. Benchmark equilibrium geometries can only be obtained by painstaking analysis of spectroscopic data or through high level ab initio geometry optimizations. Anharmonic corrections are computationally intensive because they require calculation of third derivatives of the energy with respect to atomic displacements and second derivatives of the nuclear shielding tensor. Therefore, it is prudent to investigate whether errors introduced by ignoring anharmonicity and/or optimizing geometries using density functional theory are systematic, and could therefore be compensated for using our proposed shift-and-scale procedure.
Fig. 2 Computed (B3LYP/pc-2) vs. experimentally-derived shieldings for 1H (top), 13C (middle) and 19F (bottom) nuclei. |
For 13C nuclear shieldings, neither geometry optimization nor vibrational averaging effects significantly affect the correlation between computed and observed values. Therefore, the most sensible quantities to correlate are shieldings computed at the DFT-optimized geometry and vibrationally-averaged shielding constants obtained directly from experiment.
For 19F nuclear shieldings, the picture becomes a bit more complicated. The most obvious outliers are due to limitations of mean-field models for describing bonding in F2, which shows up in longer-than-expected optimized bond lengths, and therefore shieldings that are lower in magnitude than expected. In fact, perhaps the most surprising thing about this data is how well the B3LYP data correlates to experiment when the molecule is constrained to its equilibrium bond length, given the known deficiencies of the electronic structure model. Because this outlier is so extreme and also because F2 has a very different shielding constant to other fluorine nuclei in our data set, this disproportionately affects the R2 value. However, otherwise we observe a mix of smaller sources of scatter due to both vibrational-averaging corrections and geometry optimization.
For 1H shieldings, a very different pattern emerges. In this case, geometry optimization does not seem to have much of an effect, but the data clusters strongly based upon whether zero-point vibrational averaging corrections have been applied or not. This is perhaps unsurprising, as zero-point vibrational effects are expected to be strongest for lighter, peripheral atoms. What is perhaps surprising is that 1H nuclear shieldings computed at DFT-optimized geometries correlate just as well with vibrationally-averaged shieldings obtained directly from experiment as computational values obtained at reference equilibrium geometries correlate to semi-experimental equilibrium data.
Therefore, we conclude that the most computationally-effective strategy will be to apply our two-point correction model to NMR shieldings computed at DFT-optimized geometries, using experimental reference data that has not been corrected for vibrational averaging effects. This approach will obviously incur significant errors if the DFT model of choice is not appropriate for a given molecule, but such errors will remain isolated provided that such molecules are not used in the calibration process.
Model | 1H | 13C | 19F | ||||||
---|---|---|---|---|---|---|---|---|---|
MAD (ppm) | STD (ppm) | MARD (%) | MAD (ppm) | STD (ppm) | MARD (%) | MAD (ppm) | STD (ppm) | MARD (%) | |
HnH/pc-2 | 0.539 | 0.472 | 1.6 | 8.26 | 7.71 | 5.4 | 60.4 | 39.0 | 19.2 |
HnH/pc-2 (corr) | 0.321 | 0.277 | 0.8 | 3.22 | 4.80 | 2.7 | 10.8 | 20.6 | 0.9 |
B3LYP/pc-2 | 0.797 | 0.286 | 2.7 | 12.85 | 5.34 | 14.7 | 29.0 | 18.8 | 9.3 |
B3LYP/pc-2 (corr) | 0.238 | 0.195 | 0.7 | 2.75 | 3.36 | 2.5 | 7.5 | 9.3 | 1.8 |
B3LYP0.05/pc-2 | 0.450 | 0.332 | 1.4 | 15.64 | 6.42 | 16.9 | 47.0 | 28.6 | 14.2 |
B3LYP0.05/pc-2 (corr) | 0.164 | 0.236 | 0.3 | 4.40 | 5.09 | 3.8 | 10.7 | 10.6 | 3.1 |
KT3/pc-2 | 1.062 | 0.412 | 3.7 | 4.60 | 5.69 | 4.6 | 34.5 | 16.4 | 10.3 |
KT3/pc-2 (corr) | 0.235 | 0.358 | 0.6 | 4.60 | 5.60 | 3.9 | 12.8 | 13.6 | 3.9 |
Fig. 3 Scaled-and-shifted B3LYP/pc-2 nuclear shieldings obtained using our two-point correction scheme. The dotted diagonal line represents exact equivalence between computed and experimental values. |
Model | MAD (ppm) | MARD (%) |
---|---|---|
HnH/pc-2 (corr) | 3.8 | 1.4 |
B3LYP/pc-2 (corr) | 3.1 | 1.0 |
B3LYP0.05/pc-2 (corr) | 4.4 | 1.3 |
KT3/pc-2 (corr) | 4.8 | 1.6 |
The main take-home message is that our two-point correction procedure is extremely effective at controlling for most sources of error in equilibrium DFT predictions of vibrationally-averaged NMR shielding constants, based upon DFT-optimized geometries. This enables us to identify and characterise outliers, which fall into two categories; those that arise due to neglect of relativistic effects and those that arise due to grossly inadequate modelling of electron correlation.
Hydrogen nuclei adjacent to third- and fourth-row atoms within HBr, HCl, H2S, CH3SH and PH3 exhibit significant spin–orbit coupling that is inherently captured experimentally but absent from non-relativistic quantum chemical models and not accounted for by the present formulation of our two-point correction procedure. Nonetheless, the residual errors appear quite systematic, and in line with expectations from relativistic calculations of spin–orbit contributions to nuclear shielding constants of hydrogen halides in the literature.9 Similarly, carbon nuclei adjacent to multiple chlorine atoms or a single, heavier bromine atom also display spin–orbit heavy-atom–light-atom effects consistent with trends reported in the literature.8,58
Molecular fluorine, F2 is a canonical example of a molecule in which electron correlation effects are extremely important and must be accounted for using correlated wavefunction theories to obtain get its potential energy curve even qualitatively correct.59 Similarly, electron correlation is known to be particularly important for molecules with O–F bonds, such as HOF.60 Therefore it is appropriate that DFT-based models yield inaccurate predictions in these cases.
Digging a bit deeper into the data presented in Table 7 reveals that the effectiveness of our correction procedure depends on both the level of theory and the identity of the nuclear centre. The KT3 functional displays a different error profile to the hybrid functionals; it yields more accurate predictions of 13C shieldings that cannot be improved using our correction scheme, but less accurate predictions of 1H shieldings that show the highest fold improvement in accuracy upon correction. Overall, this supports our earlier assertion that KT functionals appear to trade off the ability to capture subtle changes in valence electron behaviour for an improved description of the near-core region.
Errors in shielding constants computed using hybrid functionals tend to be larger but more systematic, so our correction procedure improves prediction accuracy across the board. For hybrid functionals, mean absolute deviations in 1H NMR shieldings are reduced by a factor of ∼2–3, 13C by ∼2.5–4.5 and 19F by ∼4–6. In general, the less accurate the initial predictions, the more they are improved upon correction.
Finally, we note that our correction model holds up well against other benchmark computational results in the literature. To facilitate comparison, we have computed aggregate mean absolute deviations and median absolute relative deviations for all corrected shielding values obtained using hybrid functionals, excluding outliers. These values are presented in Table 8. Based on mean absolute deviations, our results are comparable in accuracy to the CCSD(T)/aug-cc-pCV nZ (n = T, Q) results of Teale et al.,6 but are obtained at a fraction of the computational cost.
We have demonstrated that this approach yields predictions of approximately CCSD(T)/aug-cc-pVT/QZ accuracy for a fraction of the computational cost. Further, it enables outliers that arise due to non-systematic or non-systemic effects such as spin–orbit coupling and electron correlation model failure to be clearly identified and characterised. Spin–orbit heavy-atom–light-atom effects are particularly problematic, and will need to be accounted for to ensure high accuracy predictions for all molecules, including those that contain heavy atoms adjacent to much lighter nuclei. In keeping with the spirit of this work, we propose that this too be achieved using a simple empirical correction approach, although this will require careful calibration against benchmark computational data obtained by solving the full 4-component Dirac equation.25
Similar correlations between computed nuclear shieldings and experimental chemical shifts have been observed for molecules in solution,61 and so we expect our two-point correction procedure to generalise well – noting that the “shift” aspect of our procedure will naturally account for referencing when applied to predict relative experimental shifts from absolute computed shieldings. In other words, it is equally appropriate to use relative chemical shifts for reference points in our procedure as it is to use absolute shielding constants – provided the chemical shifts are all measured under the same experimental conditions. The question of which reference molecules are most appropriate – and the importance of accounting for solvation effects – will be addressed in future work.
Footnote |
† Electronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d2cp03992c |
This journal is © the Owner Societies 2022 |