Validation of experimental charge-density refinement strategies: when do we overfit?

A cross-validation method is supplied to judge between various strategies in multipole refinement procedures. Its application enables straightforward detection of whether the refinement of additional parameters leads to an improvement in the model or an overfitting of the given data. For all tested data sets it was possible to prove that the multipole parameters of atoms in comparable chemical environments should be constrained to be identical. In an automated approach, this method additionally delivers parameter distributions of k different refinements. These distributions can be used for further error diagnostics, e.g. to detect erroneously defined parameters or incorrectly determined reflections. Visualization tools show the variation in the parameters. These different refinements also provide rough estimates for the standard deviation of topological parameters.

1. Introduction

Although Philip Coppens (2005) wrote that `Charge densities [have] come of age', experimental charge-density studies still depend heavily on the amount and quality of the measured data. Additionally, there is no published investigation of whether or not the refinement of all possible parameters might lead to overfitting of the data, because the old benchmark of the `independent atom model' (IAM), which should not run below the 10:1 data-to-parameter ratio to keep the problem sufficiently over-determined, is hardly violated even with a demanding charge-density investigation. In a routine IAM refinement only nine parameters, three positional and six anisotropic displacement parameters, are necessary to model the structure, while in a multipole refinement (MM) via the Hansen–Coppens formalism (Hansen & Coppens, 1978) many more parameters are needed to describe the asphericity of the electron density distribution

Here, the density is divided into a core density, a spherical valence density and an aspherical valence density. The parameters κ and κ′ are used to describe the expansion and contraction of the density. This approach requires up to 27 (= 25 + 2) additional parameters per atom for multipole populations up to the hexadecapole level, and the κ and κ′ parameters. Some of the parameters are highly correlated, like the monopole population and the κ parameter of one particular atom. The valence density is mainly described by low-order data, but to derive proper thermal displacement parameters high-resolution low-temperature data are needed. Due to the development of more intense X-ray sources, more sensitive area detectors and improved cryogenic crystal cooling techniques, the data quality has improved significantly so that even an anharmonic description of the thermal motion is increasingly reported in the literature (see e.g. Dos Santos et al., 2016; Poulain et al., 2014; Domagała et al., 2014), requiring even more parameters per atom. In experimental charge-density investigations, high-resolution data up to at least 1.0 Å⁻¹ in sin(θ)/λ are necessary. Therefore, the data-to-parameter ratio is usually higher than for a routine IAM refinement. This ratio alone might suggest that overfitting is not of any concern in experimental charge-density determinations. However, this is not necessarily true because not all reflections contribute to all parameters equally, e.g. only the low-order reflections determine the valence density. Consequently, the increasing number of parameters might result in a significant drop in the R value without improving the model. A statistical method to detect that phenomenon is cross-validation. In this model-validation technique, a sample population of data is divided into complementary subsets. One subset (the `training' or `work' set) is used to derive a model, while the other is used to validate the model (the `validation' or `test' set). In macromolecular crystallography this is known as the R_free concept (Brünger, 1992, 1997). Here, the measured data are divided into a work set (often 95–90% of all data) and a test set (the remaining 5–10%). The model is refined against the work set, while the test set is never used for refinement but only to calculate an R value, the R_free value. Adding parameters will lead to a decrease in R_work, but only sensible added parameters will also decrease R_free. An increasing R_free is an unerring sign of overfitting. After determination of all sensible parameters, a final refinement can then be performed against all data.

In charge-density investigations, the following points need to be addressed:

(i) To adjudicate on overfitting, just the differences in R_free and R_work after introducing more parameters are monitored. Hence, we are not interested in their absolute values, which is different to the approach of protein crystallographers.

(ii) Determination of R_work and R_free in different resolution shells can be helpful because different regions of reciprocal space are of varying importance for different kinds of parameter.

(iii) The number of reflections in the validation set can be crucial. If the number is too small, the standard deviation of R_free is too large. The standard deviation of the free R value is given approximately by R_free/(2n)^1/2 (Tickle et al., 2000), where n is the number of reflections in the test set. As the R values, and especially the differences in the R values, are very small, this standard deviation is normally too high to provide a conclusive answer from a single R_free value. If n is too large, the completeness of the training set is too small, leading to biased models. In statistics, this is solved by k-fold cross-validation, which means that the sample is divided into k subsets. One subset is used as a test set, while the other k − 1 sets give the model. This process is repeated k times with a different subset as the test set. Hence, all data are used for both model deriving and validation. For a charge-density analysis this means that e.g. 20 test sets are produced. Every reflection is considered once in a test set and otherwise in the work set. This method was previously employed to validate the weighting of restraints (Paul et al., 2011; Zarychta et al., 2011) but should also be valid to judge e.g. local symmetry constraints, as the program XD (Volkov et al., 2006) does not offer the opportunity to introduce restraints. A final refinement is then of course employed against all available data. Out of the k different refinements the mean values 〈ΔR_free〉 can be calculated. Instead of checking 〈ΔR_free〉, a new value R_cross is defined that takes into account all test sets of all k refinements. Here, all differences between F_o and F_c for all reflections as validation reflections are used, therefore it is the R_free for all reflections as validation

A similar procedure has recently been described for macromolecules (Luebben & Gruene, 2015).

(iv) Normally the test set is chosen randomly, but caution is to be advised in non-centrosymmetric space groups or if pseudo-symmetry is present, as the Friedel pairs cannot be considered to be independent observations. The Friedel pairs or pseudo-symmetrically related reflections must be either both in the training set or both in the validation set.

(v) The validation set must be unbiased. That means it must never be used for the refinement unless the parameters are `shaken' before the cross-validation refinement starts.

For the k different models, the distribution of each refined parameter v can be checked and compared with the value v_total including its s.u., s_total, derived from a refinement against the complete data set. Model bias by omission of data can then be easily identified by outliers, e.g. values v_i that deviate by more than three times s_total from v_total. If all k refinements are independent, one should find a normal distribution with a mean value v_mean that is identical to v_total, and a standard deviation s_mean that equals s_total divided by a correction factor that can be derived from Cochran's theorem (Cochran & Wishart, 1934), e.g. 0.973 for ten test sets, 0.987 for 20 test sets and 0.995 for 50 test sets. However, as each refinement uses (k − 1)/k × n_d reflections of the total n_d data, the refinements are not independent and therefore the expected standard deviation s_mean is smaller than s_total. As a converse argument, s_mean > s_total indicates problems in the refinement.

Subsequent to a refinement with no clear indication of overfitting, the distribution of topological QTAIM (quantum theory of atoms in molecules; Bader, 1990) parameters like the density ρ, Laplacian ∇²ρ and ellipticity ∊ at the bond critical points should also be checked. Following the same arguments as above, we can estimate lower limits for the standard uncertainties of these sensitive values that are otherwise either entirely unavailable or include severe limitations (see the software manual for XD).

2. Method

A Python script was developed that runs most of the required steps automatically. As input, it needs an IAM model including hydrogen-atom positions, the XD master files (complete strategy), a parameter file and the data file. A stepwise addition of parameters, following the suggestions made in the XD manual, is highly recommended. Apart from convergence issues, a smaller step (e.g. smaller groups of parameters) supports the successful recognition of an overparameterization, along with the possibility of actually finding parameters that are likely to contribute to the overfitting of the data (see supporting information).

In this automatic process, the merged data set is first divided randomly into k (normally 20) different training and validation sets.

(i) For each training set as well as for the total data set the following steps are performed.

(a) In SHELXL (Sheldrick, 2008, 2015):

(1) Random shifts are applied to all coordinates and U_ij values;

(2) A high-order refinement of the heavy-atom positional and anisotropic displacement parameters is performed to reduce bias;

(3) A low-order refinement with fixed heavy atoms follows;

(4) The residual density peaks are automatically assigned to hydrogen atom positions.

(b) In XD (Volkov et al., 2006) the full MM refinement strategy is performed. In the first step the H-atom distances are adjusted to the neutron diffraction values (Allen & Bruno, 2010).

(c) For every refinement step, a zero-cycle structure factor calculation against the free sets (R_free) is performed.

(ii) R_cross is calculated by combining all xd.fco files of all validation sets for all steps.

(iii) Differences in 〈R_work〉 and R_cross are calculated for the individual steps and represented graphically.

(iv) In a second process, the distributions of all refined parameters of the k refinements with mean values v_mean and standard deviation s_mean are compared with the values of the refinement against all data, v_total, and the estimated standard uncertainty, s_total, calculated by XD. Individual refinement steps are selected for this comparison.

(a) To check for normal distribution, the Shapiro–Wilk test (Shapiro & Wilk, 1965) is performed for every parameter and the W and p values are given. (In this test two values, W and p, are calculated. W can be interpreted as a correlation coefficient, so has a value between 0 and 1. If the value is larger than a defined W_crit, the hypothesis of a normal distribution is accepted. p describes the probability of this particular sample distribution under the assumption of a normal distribution.)

(b) A complete list of all refined parameters is given, with v_mean, s_mean, v_total, s_total, W and p.

(c) Several tables for the diagnostics of outliers are given:

(1) A list of all parameters v_i with |v_total − v_i| > 3s_total;

(2) A list of all parameters from the Shapiro–Wilk tests with a W value smaller than 0.905 or a p value smaller than 0.05, which means that the hypothesis of a normal distribution is refused with a significance level of 0.05;

(3) A list of all parameters with |v_total − v_mean|/s_total > 0.5;

(4) A list of all parameters with s_mean > s_total;

(5) A summary of the number of outlier parameters per test set.

(d) For every parameter, a plot can be produced showing the refinement against all data in grey and the distribution of the k refinements in red (see Fig. 4).

(e) To visualize the variation in electron density of the k refinements, an error cube is mapped on a density cube. A colour-coded overlay (e.g. transparent to red) of such an error cube on a density cube calculated from the full model will highlight regions of higher uncertainties (see Fig. 6). [A density cube is calculated for all k refinements and for the complete model using the XDPROP module. The standard deviation of every grid point is calculated considering all k density cubes. A new cube containing only the deviation of each respective grid point is written. This cube is plotted as a colour-coded overlay (e.g. as a gradient from transparent to red) on specified iso-surface levels (1.0, 1.5, 2.0, 2.5) of the cube derived from the complete set of data. H atoms are automatically excluded from the calculation of the density cube because of their influence on the density combined with their unreliably determined positions that would add only a little information and clearly distract from the important parts.]

(v) Additionally, a routine can be started that runs XDPROP to search for bond critical points for all test sets at individually selectable steps of the whole refinement strategy.

(a) Again, the distribution of the k refinements is compared with the refinement against all data. From this distribution lower limits for standard deviations can be estimated.

(b) The Shapiro–Wilk values W and p are calculated.

This procedure will now be explicitly described for benchmark structure 1. Subsequently, benchmark structure 2 will be discussed much more briefly. Finally, some features of three additional structures will be presented to illustrate how this method can help to detect errors in the refinement strategy or in the data. These features emphasize, among others, the need for chemical constraints to limit pole populations of chemically equivalent atoms. The structures themselves and their topological analyses are not within the scope of the current paper and will be discussed elsewhere.

3. Experimental

3.1. Benchmark structure 1

Data for 1 (Schwendemann et al., 2011; Fig. 1) were collected on a Bruker D8 three-circle goniometer equipped with a Bruker TXS-30 Mo rotating anode with INCOATEC Helios mirror optics and an APEXII detector at 100 K. The compound crystallizes in space group with one molecule in the asymmetric unit.

Figure 1 The molecular structure of 1. Anisotropic displacement parameters are depicted at the 50% probability level. H atoms have been omitted for clarity.

The local coordinate systems were defined in such a way that the highest possible symmetry could be applied. This led to cylindrical symmetry for the H and F atoms and mirror symmetry for the ethyl and phenyl C atoms. For the para-C atoms even mm2 symmetry was adopted (for details see the supporting information). We constrained the pole population of chemically equivalent atoms to be identical (chemical constraints; for details see the supporting information). The two main questions we posed were:

(i) Can the local symmetry constraints be released without overfitting and would this add information to the model?

(ii) Can the chemical constraints be released without overfitting and would this cause differences in the parameters concerning the constrained atoms?

3.1.1. Calculation of R_cross

The Friedel pairs were not merged because the structure crystallizes in a noncentro­symmetric space group. For all validation and training sets, it was ensured that both Friedel mates were in the same set.

〈ΔR_work〉 and ΔR_cross were calculated for all data and for reflections on either side of sin(θ)/λ = 0.5 Å⁻¹ to get a feeling for the influence of low- and high-order reflections. Fig. 2 shows the results.

Figure 2 ΔR values for the initial refinement strategy. Abbreviations: M indicates monopoles, D dipoles, Q quadrupoles, O octupoles, H hexadecapoles, U Uij, K κ and KP κ′. A detailed description of the strategy is given in the supporting information.

The first step in XD is refining multipole populations. Both 〈R_work〉 and R_cross drop significantly. For both residual values, the improvement is, as expected, larger from the low-order reflections compared with the high-order ones, because the valence density is mainly described by low-order reflections. The refinement of the κ parameters has only a marginal effect on the R values but seems to improve the model slightly. Adding the monopole populations shows the expected effect: an improvement mainly from the low-order reflections. In addition, the following adjustments of displacement and positional parameters again do not indicate any overfitting. Here, a clear difference between low- and high-order reflections is not expected. The adjustment of the H-atom positional parameters in a low-order refinement followed by refinement of all other parameters against all data is a major improvement and is, as expected, mainly due to the low-order reflections. The refinement of κ′ shows only a marginal effect on the R values.

The next two steps concern the local symmetry constraints. First, the local symmetry of the F atoms is lowered from cylindrical to mirror plane symmetry. The small effect on the R values indicates neither a real model improvement nor clear overfitting. Although the changes are subtle, it seems that the high-order reflections unexpectedly contribute the most. Refining without any symmetry constraints, however, increases R_cross slightly, indicating overfitting of the data, although the total data-to-parameter ratio of 27.0 still seems to be at the safe side, while the ratio of low-order data to the sum of monopole, multipole and κ parameter is only 5.1. Overfitting is even more evident when the chemical constraints for equivalent atoms are also abandoned. Even here, the data-to-parameter ratio is still 16.7, while the ratio of low-order data to the sum of monopole, multipole and κ parameter is now only 2.5. Anyway, the residual density maps seem to be improved, because some peaks close to F atoms vanish. Although releasing constraints on the pole populations is overfitting, nevertheless the residual density still needs improvement. Hence, an anharmonic description of the thermal motion of these atoms was tested as an alternative, because we observe the typical positive and negative shashlik-like density distribution (Herbst-Irmer et al., 2013). Instead of releasing constraints, a refinement with third-order Gram–Charlier coefficients for four F atoms was performed in two steps (see Fig. 3). First, the three fluorine atoms F33, F34 and F35 of one C₆F₅ substituent and, in a second step, the para-F atom F24 of the second phenyl ring were anharmonically refined. Neither step shows any sign of overfitting and, as expected, the high-order reflections contribute the most. In the paper by Herbst-Irmer et al. (2013) we have already investigated preliminary R_free tests to validate the refinement of Gram–Charlier co­efficients but came to the conclusion that the results were not clear. Now, the k-fold cross-validation and the differences in R_cross instead of 〈R_free〉 are much more decisive.

Figure 3 ΔR values checking the anharmonic refinement of several F atoms. C denotes third-order Gram–Charlier coefficients.

It is important to note that several criteria must be fulfilled for a physically reasonable refinement of anharmonic motion (Herbst-Irmer et al., 2013):

(i) The residual density after harmonic refinement shows a positive and negative shashlik-like density distribution close to the atomic positions that vanishes after anharmonic refinement (see the supporting information).

(ii) For each anharmonically refined atom, at least one Gram–Charlier coefficient is larger than three times its s.u. (see the supporting information).

(iii) Kuhs' rule (Kuhs, 1992) should be fulfilled (see the supporting information), at least for light elements like carbon and fluorine. [Kuhs introduced a rule for estimating the minimum data resolution for meaningful refinement of anharmonic thermal parameters (Gram–Charlier coefficients) for each anisotropic atom: Q_n = 2n^1/2(2π)^−1/2(2ln2)^1/2〈u²〉^−1/2.]

(iv) The probability density function (pdf) should be reasonable. Unfortunately, the graphical representations presented by Herbst-Irmer et al. (2013) were obtained with a version of the program MoleCoolQt (Hübschle & Dittrich, 2011) that had a bug, producing wrong representations of the pdfs. Except for very strong anharmonic behaviour, normally no deviation from the harmonic ellipsoids is visible. However, the amount of negative density should be checked. For the two atoms F33 and F35 it is very small (lowest pdf values −0.87 and −0.34, respectively, and total integrated negative probability 0.000%), while for atoms F34 and F24 the lowest pdf values are −52.24 and −105.38, respectively, and the total integrated negative probabilities are 0.026 and 0.046%, respectively. These two atoms are chemically equivalent and therefore their monopole and multipole populations were constrained to be the same. Additionally, there is a high correlation between the Gram–Charlier coefficient C₂₂₂ of atom F24 and the bond-directed dipole population of 90%, indicating that this anharmonic refinement is not reasonable. Therefore only atom F34 should be refined anharmonically, while atom F24 stays harmonic. This improves the pdf for F34 considerably (lowest pdf value −1.06 and total integrated negative probability 0.000%). Now the highest correlation of 60–70% for the Gram–Charlier coefficent is to the positional parameters x, y and z, as described previously (Herbst-Irmer et al., 2013).

This example emphasizes that a small drop in R_cross is a necessary, but by no means a sufficient, condition for a physically reasonable model. After the anharmonic refinement of these three F atoms, we checked again the effect of local symmetry lowering (see the supporting information). A reduction of cylindrical to mirror plane symmetry for the F atoms still does not indicate overfitting and has no significant effect on 〈R_work〉, but there is a slightly greater effect from the high-order reflection. As the influence on the residual density is also marginal [a drop in e_gross (Meindl & Henn, 2008) from 39 to 38.8 e Å⁻³], we decided that the final refinement strategy should maintain the above-mentioned local symmetry, should keep all possible chemical constraints and should contain the anharmonic refinement of the three F atoms.

3.1.2. Distribution of the refined parameters

To prevent model bias due to omission of reflections, we checked the distribution of parameters derived by the 20 different refinements. Fig. 4 shows two examples, in part (a) a parameter with all values within v_total ± s_total and in part (b) a parameter with one outlier.

Figure 4 Distributions of two example parameters v, e.g. P20 and P43+ of atom C1. (a) A parameter with all values within vtotal ± stotal. (b) A parameter with one outlier. The distributions are plotted both as histogram bars and as fitted Gaussian functions. The value derived from the refinement against all data vtotal with the estimated standard uncertainty stotal is depicted in grey, while the distribution of the value with vmean and standard deviation smean of the refinements against the 20 different training sets is given in red.

For only nine out of 618 parameters of the final refinement did we find one to three such outliers. The entire set of 14 deviations do not belong to one particular training set (for details see the supporting information). Therefore, model bias due to omission of data can be excluded. Nine of the entire set of 14 outliers are octupole or hexadecapole populations of atoms C1 and C2. These atoms are refined without any local symmetry or chemical constraints. Therefore, we increased the adopted local symmetry in a new refinement strategy: we now adopt m symmetry for atoms C1, N1 and B1, and mm2 symmetry for all phenyl C atoms. Additionally, we started refining the multipole population only up to the octupole level and added hexadecapole populations in a later step.

Fig. 5 shows the ΔR values for this stepwise refinement. The hexadecapole populations are first refined for all atoms besides C1, C2, B1 and N1, and then for all atoms. Then the three F atoms are refined anharmonically. In the next two steps no symmetry constraints are applied to atoms C1, B1 and N1, and afterwards the phenyl atoms are relaxed from mm2 to m symmetry. None of these steps seems to overfit. In the next step, the local symmetry of the F atoms is changed from cylindrical to mirror plane symmetry. Again, no clear overfit is visible, but as before the high-order reflections seem to contribute the most. Releasing the local symmetry and chemical constraints in the last two steps clearly overfits. Therefore, we decided to stop the strategy after step 17 with the symmetry reduction of the phenyl atoms to m symmetry. For this last non-overfitting step there are only two parameters with |v_total − v_i| > 3s_total with a maximum value of 3.3s_total, and only seven parameters with s_total < s_mean. The largest (s_mean − s_total)/s_total is 0.34. From the parameters describing the valence density (monopole and multipole populations and κ values), the hypothesis of a normal distribution is confirmed with a significance of 0.05 for 235 of the 255 parameters.

Figure 5 ΔR values for an improved refinement strategy. Abbreviations: H indicates hexadecapoles and C third-order Gram–Charlier coefficients.

The idea presented in the following is based on the assumption that overfitted parameters are imprecisely determined and therefore lead to higher variations in the electron density described by them. A density cube at a relevant isosurface level will not be able to show that variation, as the model parameters do not directly reflect this uncertainty. A computational detour involving a set of density cubes calculated for all work sets allows calculation of the standard deviation of each density grid point (see §2). A colour-coded overlay (e.g. transparent to red) of such an error cube on a density cube (calculated from the full model) is able to highlight regions of higher uncertainties (Fig. 6). This error cube is related to the σ(ρ) cube, the calculation of which is implemented in XD, although it shows additional features not covered hitherto. As the H-atom positions are by far the least precisely determined parameters, their density will probably show severe features in the corresponding error cube.

Figure 6 The variation in electron density for the k refinements, displayed as an error cube mapped on a density cube calculated for the complete model. Cubes are shown for (a) the last reasonable step 17 and (b) the refinement without any constraints (step 20). Two isosurface levels are plotted, at 1 and 2 e Å^−3, and the standard deviation is capped at 0.015 e Å^−3. Hydrogen density has been omitted.

Fig. 6 compares the variation in the electron density of the k refinements in an error cube mapped on a density cube (see §2) for the last reasonable step 17 (Fig. 6a), with step 20 (Fig. 6b) refining all possible multipoles. For step 17, the atoms with the highest variations are those that are refined with no local symmetry constraints and no chemical constraints. Astonishingly, the refinement of atom B1 seems to be very stable. In step 20 all atoms show relatively high variation.

3.1.3. Distribution of properties at bond critical points

With this refinement strategy, the distribution of properties at bond critical points can now be evaluated. The calculation of errors on ρ and ∇²ρ in XD has some severe limitations, as mentioned in the manual. For the ellipticity ∊ no errors are provided. Nevertheless, an estimation of these errors would be useful. The distribution of k refinements can be used and, as previously described, the standard deviation derived from the distribution can be considered as a lower limit. The true error could be higher, because the k refinements are not independent. Table 1 shows a list of the properties at the bond critical points of the B—C and B—N bonds, comparing the final refinement with the first refinement without any local symmetry or chemical constraints.

Table 1
Properties (ρ in e Å⁻³ and ∇²ρ in e Å⁻⁵) at the bond critical points of the B—C and B—N bonds, comparing the final refinement strategy (second row) with the initial strategy without local symmetry or chemical constraints (first row)

Bond	Property	vtotal	stotal	vmean	smean	|vmean − vtotal|/smean	smean/stotal
C1—B1	ρ	1.139	0.012	1.141	0.007	0.3	0.6
		1.154	0.012	1.156	0.006	0.2	0.5
	∇^2ρ	−5.86	0.06	−6.0	0.6	0.2	11.2
		−7.16	0.06	−7.2	0.5	0.2	9.7
	∊	0.03		0.042	0.019	0.5
		0.05		0.055	0.013	0.4
C21—B1	ρ	1.023	0.011	1.024	0.010	0.1	0.9
		1.057	0.009	1.059	0.005	0.5	0.5
	∇^2ρ	−3.01	0.05	−3.1	0.6	0.1	12.4
		−5.68	0.04	−5.8	0.4	0.3	9.2
	∊	0.11		0.11	0.02	0.2
		0.09		0.094	0.009	0.5
C31—B1	ρ	1.050	0.011	1.049	0.007	0.1	0.7
		1.055	0.010	1.055	0.005	0.1	0.5
	∇^2ρ	−7.47	0.05	−7.5	0.4	0.1	7.6
		−6.20	0.04	−6.2	0.4	0.1	9.2
	∊	0.18		0.18	0.02	0.2
		0.15		0.15	0.02	0.2
N1—B1	ρ	0.779	0.011	0.790	0.005	0.6	0.5
		0.747	0.011	0.749	0.005	0.4	0.4
	∇^2ρ	−0.94	0.04	−1.1	0.3	0.4	7.1
		1.86	0.05	1.8	0.4	0.2	7.9
	∊	0.48		0.49	0.06	0.2
		0.65		0.66	0.06	0.1

The following conclusions can be drawn:

(i) For all properties, the differences between the value derived from the refinement against all data and the mean value of the 20 refinements is insignificant, |v_mean − v_total|/s_total < 0.6.

(ii) For the density ρ, the estimated standard uncertainty for the refinement against all data is slightly larger than the standard deviation of the distribution of the 20 refinements, s_mean/s_total < 1. The estimated standard uncertainty is in the range 0.009–0.012 e Å⁻³, or 1–1.5% of ρ. In an investigation using 13 different data sets of oxalic acid (Kamiński et al., 2014), the distribution has a standard deviation in the range 0.03–0.06 e Å⁻³ or 1.5–3.0%.

(iii) For the Laplacian ∇²ρ, the estimated standard uncertainty is much smaller than the standard deviation of the distribution, s_mean/s_total > 7.1. The standard deviations are between 0.3 and 0.6 e Å⁻⁵ i.e. 5–32% of v_total. In the above-mentioned investigation (Kamiński et al., 2014), the standard deviation is between 1 and 7 e Å⁻⁵, i.e. 6–28%.

(iv) For the ellipticity ∊, the standard deviation is between 0.009 and 0.02. In the work by Kamiński et al. (2014) it is between 0.01 and 0.04.

(v) Comparing the two refinements, most properties are very similar, but the Laplacian of the N—B bond changes from −0.94 to 1.86 e Å⁻⁵ and the ellipticity ∊ changes from 0.5 to 0.7. To investigate which refinement step is mainly responsible for these differences, all properties were checked for several steps. The most important point seems to be the use of chemical constraints, agreeing with the greatest effect on R_cross (for details see the supporting information).

3.1.4. Influence of the number of test sets on the model

The influence of the number of training sets on the models was also checked. We repeated the refinement using ten or 50 training sets, respectively (10% and 2% of data left out, respectively). The behaviour of R_cross is nearly identical for ten, 20 or 50 training sets (see the supporting information) but, as expected, for ten validation sets (more data left out) there are a higher number of outliers with |v_i − v_total| < 3s_total, while there are none for 50 validation sets. Accordingly, the standard deviations of the distributions shrink from ten to 50 validation sets. Here, we decided to use 20 sets, as this seems to be a sensible compromise between model bias and reasonable standard deviations.

3.1.5. Influence of Friedel mates on the model

We ensured that Friedel pairs are either both in the training set or both in the validation set because structure 1 crystallizes in a non-centrosymmetric space group. To evaluate the impact of this treatment we performed the same refinement, but this time with randomly prepared training and validation sets without any special care for the Friedel pairs (see Fig. 7). Now overfitting is not that easy to detect. If we now check the feasibility of releasing the local symmetry or chemical constraints, 〈R_work〉 responds in the familiar way but R_cross remains nearly unchanged. There is still an indication of overfitting if R_mean decreases much more than R_cross, but the picture is much less obscure with a proper treatment of the Friedel pairs. As they are not independent, neither is the model independent of a particular Friedel mate being present or not in the training set.

Figure 7 ΔR values for training sets with no special treatment of Friedel pairs

3.2. Benchmark structure 2

Data for structure 2 (Fig. 8) were collected on a Bruker D8 three-circle goniometer equipped with a Bruker TXS-30 Mo rotating anode with INCOATEC Helios mirror optics and an APEXII detector at 100 K. The compound crystallizes in space group Pbca (Krause, 2017).

Figure 8 The molecular structure of 2. Anisotropic displacement parameters are depicted at the 50% probability level. Hydrogen atoms have been omitted for clarity.

The local coordinate systems were defined so that the highest possible symmetry could be applied. This led to cylindrical symmetry for the H and S atoms, mirror symmetry for the methine C atoms and threefold symmetry for the methyl C atoms. For the anthracene C atoms mm2 symmetry was adopted. The pole populations of chemically equivalent atoms were constrained to be identical. The C-atom multipole expansion was restricted to octupoles, while for S and P atoms hexadecapoles were employed. The refinement strategy was similar to that of 1 (see the supporting information) and similar trends for the behaviour of R_cross are evident (see Fig. 9). The refinement of the multipole populations improves 〈R_work〉 and R_cross significantly. Both values show an improvement that is greater from the low-order data compared with the high-order, which is in good agreement with expectations. The subsequent introduction of monopoles displays a similar picture. Next, adjustment of the displacement parameters does not show any sign of overfitting and is affected by both low- and high-order reflections. The addition of positional parameters for non-hydrogen atoms does not need adjustment to the same extent here, indicating better starting parameters. With the density modelled, a new adjustment of the H-atom positions in step 6 leads to a drop in the low-resolution R value, which is again in good agreement with expectations.

Figure 9 ΔR values for the initial refinement strategy for structure 2. Abbreviations: M indicates monopoles, D dipoles, Q quadrupoles, O octupoles, H hexadecapoles, U Uij, K κ and KP κ′.

The refinement of the k parameters in step 8 stems pre­dominantly from the low-resolution data. The introduction of hexadecapoles to the C atoms also shows an improvement from the low-order data. The same is found for the expansion from mm2 to m (in the plane) symmetry for the anthracene C atoms. However, on releasing the local symmetry constraints, overfitting is indicated by a drop in 〈R_work〉 and an increase in R_cross, which is even more pronounced after the release of the constraints for equivalent atoms in step 14. The total data-to-parameter ratio is still 24.2, but the ratio of low-order data to the monopole or multipole populations and the κ parameter is reduced to 3.1. Ultimately, we decided to stay with the more restricted model obtained after step 12. Here, the ratio of low-order data to the monopole or multipole populations and the κ parameter is still 13.0.

3.3. Constraints due to crystallographic symmetry

Atoms on special positions need constraints for the refinement. While in IAM programs like SHELXL these constraints are generated automatically, they need to be set manually in XD.

Compound 3 (Fig. 10) crystallizes in space group Pnma with half a molecule in the asymmetric unit (Jancik et al., 2017). Atoms P1, Cl1, Cl2 and N2 are located on a crystallographic mirror plane. Cylindrical symmetry was applied to the Cl atoms, and mm2 symmetry for the P and N atoms. All atoms of each element type were constrained to share the same monopole and multipole populations. Additional parameters were then added in a stepwise manner: multipole and monopole populations, and adjustment of U, xyz, κ and κ′. No refinement step indicated any overfitting (see the supporting information). However, a typical residual density distribution, indicating anharmonic motion, was present. Hence third-order Gram–Charlier coefficients, first for the Cl atoms, then for the P atoms and finally for the N atoms, were introduced. No overfitting was visible and all performed tests for a reasonable refinement of anharmonic motion were fulfilled (Jancik et al., 2017). Then the local symmetry of the Cl atoms was released from cylindrical to mirror plane symmetry (data-to-parameter ratio 44.3, low-order to monopole or multipole and κ parameter 8.7). Subsequently, in a similar fashion, the symmetry was lowered for the N and P atoms from mm2 to m (data-to-parameter ratio still 41.2, low-order to monopole or multipole and κ parameter 6.4). The release of all chemical constraints for all atoms followed (data-to-parameter ratio 28.7, low-order to monopole or multipole and κ parameter 4.9). Finally, all symmetry constraints for all atoms on general positions were released (data-to-parameter ratio 24.7, low-order to monopole or multipole and κ parameter 3.7). While the release of the local symmetry contraints for the N and P atoms has only a small effect on the R values, the release of both chemical and local symmetry constraints for all atoms on general positions increases R_cross, clearly indicating overfitting.

Figure 10 The molecular structure of 3. Anisotropic displacement parameters are depicted at the 50% probability level.

In a first trial, the constraints for the Gram–Charlier co­efficients for the atoms on special positions were mistakenly left unset. The impact on ΔR_cross was imperceptible, but this error could easily be identified by inspection of the distribution of values of the k refinements (see Fig. 11). Of course, such an error also leads to convergence problems, but the parameter distribution is able to identify the problematic parameters quickly.

Figure 11 Distribution of the Gram-Charlier coefficient C112 of atom Cl1, which was refined by mistake but should be constrained to 0 due to the crystallographic mirror symmetry

3.4. Outlier detection

Structure 4 (Stute et al., 2012; Fig. 12) crystallizes in space group P2₁/n with one molecule in the asymmetric unit.

Figure 12 The molecular structure of 4. Anisotropic displacement parameters are depicted at the 50% probability level. Hydrogen atoms have been omitted for clarity

In this structure, the check of the parameter distribution identified all outliers with |v_total − v_i| > 3s_total (for details see the supporting information) as belonging to only three test sets. Two of these sets showed much higher R_free values than all other sets (see Fig. 13).

Figure 13 Rfree values for all 20 validations sets, blue: using all data, grey: after omission of the outlying reflections.

Careful inspection of these two validation sets showed that in both there is one very strong low-order reflection with F_o << F_c found by inspection of the DRK plots (Zhurov et al., 2008; Zavodnik et al., 1999) (see the supporting information). These two reflections are overexposed. Strong low-order reflections have a large effect on the multipole populations, so an in­accurate determination of such reflections is highly problem­atic. Since these reflections are omitted from these two training sets, the derived parameter sets from these two training sets are not outliers but proper values. An omission of these two reflections lowers the completeness in this important low-order range but improves the model indicated by a more reasonable parameter distribution (see as an example Fig. 14b).

Figure 14 Distribution of the monopole population of atom F34, (a) using all data and (b) after omitting two outlier reflections.

3.5. Chemical constraints

In all our tested structures, we observed the release of chemical constraints to result in serious overfitting. Recently, we published the charge-density investigation of a silylone (Niepötter et al., 2014). In this structure, an Si atom is co­ordinated by two identical cAAC (cyclic alkyl amino carbene) ligands. We found that the two Si—C bonds differ significantly in both length and ellipticity. Therefore, we anticipated that the release of chemical constraints is necessary for a proper modelling of this structure (for details of the R_cross procedure see the supporting information). However, unexpectedly, only the release of the local symmetry constraints of the Si atom proved necessary to describe the differences in the two Si—C bonds. In contrast, the release of the chemical constraints of the two complete carbene ligands showed severe signs of overfitting, further emphasizing chemical constraints to be important in stabilizing a multipole refinement, at least for data sets to a resolution of 0.5–0.4 Å.

4. Conclusions

The presented method of cross-validation is a valid tool in multipole refinement. Although the number of data in a high-resolution data set is high enough to achieve a global data-to-parameter ratio larger than 1:10 or even 1:20, even when all possible multipole parameters up to the hexadecapole level are refined, it has to be considered that not all reflections contribute equally to all multipole parameters. The information about the valence density is mainly gained from the relatively few low-order reflections. Therefore, the simple global data-to-parameter rule of thumb is not sufficient to decide on the advisable number of parameters. In contrast with the well established R_free in macromolecular crystallography (Brünger, 1992, 1997) or the new R_complete concept (Luebben & Gruene, 2015), this method is not interested in absolute R_cross values but in the progress of R_cross along the refinement strategy. For a reasonable refinement, not only must the normal R values decrease but also the R_cross value. Investigating several structures with this approach, we noticed the following general aspects:

(i) It helps to start with very high symmetry defined only by next neighbours.

(ii) Lowering these symmetry constraints often does not improve the model significantly. Normally it is sufficient to release these constraints for only a few atoms. Any un­expected large drop in R_cross is suspicious and helps to detect mistakes in the definition of the local coordinate system or the assumed local symmetry.

(iii) Release of all chemical constraints clearly causes overfitting in all investigated structures.

(iv) Unfortunately, XD does not offer restraints. However, using restraints instead of constraints could be a further improvement of the model (Paul et al., 2011; Zarychta et al., 2011). A protocol for oriented local atomic axes is explained in detail by Domagała & Jelsch (2008).

Additional to validation of the refinement strategy, analysis of the parameter distribution provides access to further causes of defects. Overlooking the necessary constraints required by crystallographic symmetry can easily be prevented. Inaccurately determined strong low-order reflections that bias the derived parameters can be easily identified.

The k different refinements also provide a distribution of the topological properties, leading to a rough estimate for the standard deviations of these parameters.

Supporting information

Computing details

For all compounds, data collection: APEX2 v2012.2. Cell refinement: SAINT V8.30C for 1_shelx, 1_xd; SAINT V8.37A for 2_shelx, 2_xd. Data reduction: SAINT V8.30C for 1_shelx, 1_xd; SAINT V8.37A for 2_shelx, 2_xd. Program(s) used to solve structure: SHELXT(G.M.Sheldrick,Acta Cryst.(2015)A71,3-8) for 1_shelx, 1_xd; SHELXT 2014/5 (Sheldrick, 2014) for 2_shelx, 2_xd. Program(s) used to refine structure: SHELXL2016/6 (Sheldrick, 2016) for 1_shelx, 1_xd, 2_shelx; Volkov et al., (2006) for 2_xd. For all compounds, molecular graphics: XP Version 5.1; software used to prepare material for publication: XP Version 5.1.

(1_shelx)

Crystal data

C24H18BF10N	Dx = 1.615 Mg m^−3
Mr = 521.20	Mo Kα radiation, λ = 0.71073 Å
Tetragonal, P421c	Cell parameters from 8098 reflections
a = 22.206 (3) Å	θ = 2.6–49.1°
c = 8.692 (2) Å	µ = 0.16 mm^−1
V = 4286.1 (15) Å^3	T = 100 K
Z = 8	Block, colorless
F(000) = 2112	0.10 × 0.09 × 0.09 mm

Data collection

Bruker Smart APEX II Ultra diffractometer	19346 reflections with I > 2σ(I)
Radiation source: BRUKER Rotating Anode	Rint = 0.041
ω scans	θmax = 49.2°, θmin = 1.3°
Absorption correction: multi-scan SADABS 2015/1	h = −47→47
Tmin = 0.928, Tmax = 0.971	k = −46→46
404587 measured reflections	l = −18→18
21597 independent reflections

Refinement

Refinement on F^2	Hydrogen site location: inferred from neighbouring sites
Least-squares matrix: full	H-atom parameters constrained
R[F^2 > 2σ(F^2)] = 0.030	w = 1/[σ^2(Fo^2) + (0.0489P)^2 + 0.0917P] where P = (Fo^2 + 2Fc^2)/3
wR(F^2) = 0.084	(Δ/σ)max = 0.001
S = 1.12	Δρmax = 0.51 e Å^−3
21597 reflections	Δρmin = −0.25 e Å^−3
327 parameters	Absolute structure: Flack x determined using 8376 quotients [(I+)-(I-)]/[(I+)+(I-)] (Parsons, Flack and Wagner, Acta Cryst. B69 (2013) 249-259).
0 restraints	Absolute structure parameter: −0.03 (3)

Special details

Geometry. All esds (except the esd in the dihedral angle between two l.s. planes) are estimated using the full covariance matrix. The cell esds are taken into account individually in the estimation of esds in distances, angles and torsion angles; correlations between esds in cell parameters are only used when they are defined by crystal symmetry. An approximate (isotropic) treatment of cell esds is used for estimating esds involving l.s. planes.

Fractional atomic coordinates and isotropic or equivalent isotropic displacement parameters (Ų)

	x	y	z	Uiso*/Ueq
N1	0.73830 (2)	0.51095 (2)	0.37910 (5)	0.01034 (5)
B1	0.75169 (3)	0.52441 (3)	0.56972 (7)	0.01079 (7)
C1	0.82116 (2)	0.50713 (3)	0.52843 (6)	0.01245 (7)
H1A	0.846977	0.542449	0.506464	0.015*
H1B	0.840401	0.480104	0.604519	0.015*
C2	0.79872 (2)	0.47494 (2)	0.38502 (6)	0.01141 (6)
H2	0.788807	0.432608	0.414986	0.014*
C3	0.83693 (3)	0.47216 (2)	0.24174 (6)	0.01284 (7)
C4	0.88063 (3)	0.51575 (3)	0.20902 (8)	0.01630 (8)
H4	0.886656	0.548338	0.278129	0.020*
C5	0.91536 (3)	0.51172 (4)	0.07596 (9)	0.02081 (11)
H5	0.944715	0.541686	0.054466	0.025*
C6	0.90720 (4)	0.46384 (4)	−0.02584 (8)	0.02268 (12)
H6	0.930487	0.461449	−0.117312	0.027*
C7	0.86477 (4)	0.41963 (4)	0.00732 (8)	0.02061 (11)
H7	0.859428	0.386647	−0.060950	0.025*
C8	0.83011 (3)	0.42363 (3)	0.14046 (7)	0.01619 (8)
H8	0.801502	0.393056	0.162823	0.019*
C9	0.74312 (3)	0.56601 (2)	0.28002 (6)	0.01298 (7)
H9A	0.782759	0.584948	0.299476	0.016*
H9B	0.711842	0.594982	0.313336	0.016*
C10	0.73660 (3)	0.55650 (3)	0.10712 (7)	0.01851 (9)
H10A	0.767876	0.528704	0.071275	0.028*
H10B	0.740905	0.595193	0.054031	0.028*
H10C	0.696818	0.539479	0.084903	0.028*
C11	0.68641 (3)	0.47198 (3)	0.33416 (7)	0.01371 (7)
H11A	0.684582	0.437255	0.405426	0.016*
H11B	0.693848	0.455994	0.229554	0.016*
C12	0.62563 (3)	0.50366 (3)	0.33529 (9)	0.01928 (10)
H12A	0.618016	0.520320	0.437863	0.029*
H12B	0.593895	0.474731	0.309272	0.029*
H12C	0.625845	0.536316	0.259457	0.029*
C13	0.74496 (2)	0.59298 (2)	0.63759 (6)	0.01058 (6)
C14	0.70447 (2)	0.63917 (2)	0.60470 (6)	0.01176 (6)
C15	0.70615 (3)	0.69567 (2)	0.67317 (7)	0.01316 (7)
C16	0.74840 (3)	0.70805 (3)	0.78583 (6)	0.01430 (7)
C17	0.78845 (3)	0.66311 (3)	0.82659 (6)	0.01367 (7)
C18	0.78640 (2)	0.60791 (2)	0.75232 (6)	0.01192 (6)
C19	0.71483 (2)	0.47761 (2)	0.68230 (6)	0.01201 (6)
C20	0.73789 (3)	0.42301 (3)	0.73780 (6)	0.01451 (7)
C21	0.71089 (4)	0.38950 (3)	0.85408 (7)	0.01773 (9)
C22	0.65755 (4)	0.40895 (3)	0.91887 (7)	0.01869 (10)
C23	0.63125 (3)	0.46129 (3)	0.86467 (8)	0.01780 (9)
C24	0.65998 (3)	0.49360 (3)	0.74971 (7)	0.01412 (7)
F14	0.65961 (2)	0.63126 (2)	0.50179 (5)	0.01574 (6)
F15	0.66653 (2)	0.73833 (2)	0.63231 (6)	0.01857 (7)
F16	0.74989 (3)	0.76149 (2)	0.85550 (6)	0.02063 (8)
F17	0.82815 (2)	0.67254 (2)	0.94020 (5)	0.01959 (7)
F18	0.82690 (2)	0.56711 (2)	0.80126 (5)	0.01659 (6)
F20	0.78896 (2)	0.39950 (2)	0.68118 (6)	0.02022 (8)
F21	0.73641 (3)	0.33875 (2)	0.90582 (6)	0.02569 (10)
F22	0.63184 (3)	0.37764 (3)	1.03277 (6)	0.02743 (11)
F23	0.57935 (3)	0.48065 (3)	0.92432 (8)	0.02757 (11)
F24	0.63130 (2)	0.54419 (2)	0.70519 (6)	0.01877 (7)

Atomic displacement parameters (Ų)

	U^11	U^22	U^33	U^12	U^13	U^23
N1	0.01126 (12)	0.01008 (12)	0.00968 (12)	−0.00041 (10)	−0.00024 (10)	−0.00035 (10)
B1	0.01145 (17)	0.01110 (17)	0.00984 (16)	0.00017 (14)	0.00035 (14)	−0.00064 (13)
C1	0.01123 (15)	0.01451 (17)	0.01161 (15)	0.00100 (13)	−0.00017 (12)	−0.00137 (13)
C2	0.01225 (15)	0.01161 (14)	0.01038 (14)	0.00080 (12)	0.00121 (12)	−0.00036 (12)
C3	0.01379 (16)	0.01329 (16)	0.01143 (15)	0.00195 (13)	0.00237 (13)	0.00001 (13)
C4	0.01522 (18)	0.0168 (2)	0.01693 (19)	0.00058 (15)	0.00397 (16)	0.00170 (16)
C5	0.0185 (2)	0.0231 (3)	0.0209 (2)	0.00345 (19)	0.00828 (19)	0.0057 (2)
C6	0.0240 (3)	0.0272 (3)	0.0169 (2)	0.0092 (2)	0.0091 (2)	0.0035 (2)
C7	0.0257 (3)	0.0221 (2)	0.01401 (19)	0.0077 (2)	0.00478 (19)	−0.00206 (17)
C8	0.0197 (2)	0.01564 (18)	0.01327 (17)	0.00313 (16)	0.00263 (16)	−0.00208 (15)
C9	0.01595 (18)	0.01171 (15)	0.01130 (15)	0.00062 (13)	0.00033 (13)	0.00127 (12)
C10	0.0236 (2)	0.0207 (2)	0.01126 (16)	0.00315 (19)	−0.00050 (17)	0.00285 (16)
C11	0.01416 (17)	0.01282 (16)	0.01416 (17)	−0.00193 (13)	−0.00148 (14)	−0.00216 (14)
C12	0.01339 (18)	0.0197 (2)	0.0248 (3)	−0.00107 (16)	−0.00350 (18)	−0.0019 (2)
C13	0.01096 (14)	0.01078 (14)	0.01000 (13)	−0.00056 (11)	0.00058 (11)	−0.00109 (11)
C14	0.01092 (14)	0.01186 (15)	0.01249 (15)	−0.00003 (12)	0.00081 (12)	−0.00166 (12)
C15	0.01326 (16)	0.01108 (15)	0.01513 (17)	0.00036 (13)	0.00316 (14)	−0.00174 (13)
C16	0.01786 (19)	0.01180 (16)	0.01323 (16)	−0.00334 (14)	0.00320 (14)	−0.00326 (13)
C17	0.01571 (18)	0.01447 (17)	0.01082 (15)	−0.00424 (14)	0.00012 (13)	−0.00223 (13)
C18	0.01287 (15)	0.01282 (15)	0.01008 (13)	−0.00128 (13)	−0.00011 (12)	−0.00060 (12)
C19	0.01407 (16)	0.01169 (15)	0.01026 (14)	−0.00047 (13)	0.00112 (12)	−0.00018 (12)
C20	0.0193 (2)	0.01264 (16)	0.01158 (16)	0.00084 (15)	0.00132 (14)	0.00102 (13)
C21	0.0271 (3)	0.01345 (17)	0.01267 (17)	−0.00213 (18)	0.00128 (17)	0.00189 (14)
C22	0.0270 (3)	0.0161 (2)	0.01293 (17)	−0.00758 (19)	0.00477 (18)	−0.00034 (15)
C23	0.0194 (2)	0.0168 (2)	0.0172 (2)	−0.00554 (17)	0.00663 (17)	−0.00230 (16)
C24	0.01424 (17)	0.01362 (17)	0.01451 (17)	−0.00210 (14)	0.00280 (15)	−0.00074 (14)
F14	0.01232 (12)	0.01725 (14)	0.01764 (14)	0.00206 (11)	−0.00322 (11)	−0.00331 (12)
F15	0.01686 (15)	0.01358 (13)	0.02528 (18)	0.00427 (11)	0.00281 (13)	−0.00180 (13)
F16	0.0295 (2)	0.01278 (13)	0.01956 (16)	−0.00394 (14)	0.00330 (15)	−0.00655 (12)
F17	0.02197 (17)	0.02233 (17)	0.01448 (14)	−0.00653 (14)	−0.00438 (13)	−0.00417 (13)
F18	0.01769 (14)	0.01813 (15)	0.01394 (12)	0.00282 (12)	−0.00438 (11)	−0.00059 (11)
F20	0.02397 (18)	0.01716 (15)	0.01954 (16)	0.00744 (14)	0.00495 (14)	0.00432 (13)
F21	0.0408 (3)	0.01606 (16)	0.02022 (18)	0.00214 (17)	0.00136 (19)	0.00721 (14)
F22	0.0410 (3)	0.0226 (2)	0.01865 (17)	−0.0116 (2)	0.01085 (19)	0.00293 (15)
F23	0.0237 (2)	0.0260 (2)	0.0330 (3)	−0.00421 (17)	0.0171 (2)	−0.00224 (19)
F24	0.01430 (14)	0.01618 (14)	0.02584 (18)	0.00163 (12)	0.00488 (13)	0.00142 (13)

Geometric parameters (Å, º)

N1—C11	1.4929 (7)	C14—C15	1.3892 (7)
N1—C9	1.4994 (7)	C15—F15	1.3407 (7)
N1—C2	1.5628 (7)	C15—C16	1.3838 (9)
N1—B1	1.7096 (8)	C16—F16	1.3328 (7)
B1—C1	1.6297 (8)	C16—C17	1.3829 (9)
B1—C13	1.6398 (8)	C17—F17	1.3401 (7)
B1—C19	1.6454 (8)	C17—C18	1.3860 (8)
C1—C2	1.5208 (8)	C18—F18	1.3456 (7)
C2—C3	1.5081 (8)	C19—C24	1.3976 (8)
C3—C8	1.3997 (8)	C19—C20	1.4018 (8)
C3—C4	1.3998 (9)	C20—F20	1.3419 (8)
C4—C5	1.3930 (9)	C20—C21	1.3909 (9)
C5—C6	1.3951 (13)	C21—F21	1.3391 (9)
C6—C7	1.3909 (13)	C21—C22	1.3809 (11)
C7—C8	1.3927 (9)	C22—F22	1.3377 (8)
C9—C10	1.5246 (9)	C22—C23	1.3835 (11)
C11—C12	1.5220 (9)	C23—F23	1.3349 (9)
C13—C14	1.3936 (8)	C23—C24	1.3857 (9)
C13—C18	1.3969 (7)	C24—F24	1.3480 (8)
C14—F14	1.3504 (7)
C11—N1—C9	112.18 (4)	C15—C14—C13	124.03 (5)
C11—N1—C2	112.00 (4)	F15—C15—C16	119.48 (5)
C9—N1—C2	112.01 (4)	F15—C15—C14	120.47 (5)
C11—N1—B1	119.28 (4)	C16—C15—C14	120.04 (5)
C9—N1—B1	113.68 (4)	F16—C16—C17	120.68 (6)
C2—N1—B1	84.73 (3)	F16—C16—C15	121.01 (6)
C1—B1—C13	112.59 (4)	C17—C16—C15	118.30 (5)
C1—B1—C19	116.95 (4)	F17—C17—C16	119.93 (5)
C13—B1—C19	109.10 (4)	F17—C17—C18	120.21 (6)
C1—B1—N1	84.84 (4)	C16—C17—C18	119.84 (5)
C13—B1—N1	119.67 (4)	F18—C18—C17	115.22 (5)
C19—B1—N1	112.30 (4)	F18—C18—C13	120.41 (5)
C2—C1—B1	88.91 (4)	C17—C18—C13	124.34 (5)
C3—C2—C1	120.78 (5)	C24—C19—C20	113.19 (5)
C3—C2—N1	118.47 (4)	C24—C19—B1	121.50 (5)
C1—C2—N1	93.89 (4)	C20—C19—B1	124.70 (5)
C8—C3—C4	118.65 (5)	F20—C20—C21	115.00 (5)
C8—C3—C2	119.34 (5)	F20—C20—C19	121.26 (5)
C4—C3—C2	121.97 (5)	C21—C20—C19	123.74 (6)
C5—C4—C3	120.54 (6)	F21—C21—C22	119.29 (6)
C4—C5—C6	120.23 (7)	F21—C21—C20	120.79 (7)
C7—C6—C5	119.64 (6)	C22—C21—C20	119.92 (6)
C6—C7—C8	120.09 (7)	F22—C22—C21	120.36 (7)
C7—C8—C3	120.80 (6)	F22—C22—C23	120.56 (7)
N1—C9—C10	116.51 (5)	C21—C22—C23	119.08 (6)
N1—C11—C12	114.51 (5)	F23—C23—C22	120.19 (6)
C14—C13—C18	113.37 (5)	F23—C23—C24	120.72 (7)
C14—C13—B1	131.95 (5)	C22—C23—C24	119.09 (6)
C18—C13—B1	114.66 (4)	F24—C24—C23	114.90 (5)
F14—C14—C15	114.90 (5)	F24—C24—C19	120.21 (5)
F14—C14—C13	121.07 (5)	C23—C24—C19	124.89 (6)

(1_xd)

Crystal data

C24H18BF10N	Dx = 1.615 Mg m^−3
Mr = 521.20	Mo Kα radiation, λ = 0.71073 Å
Tetragonal, P421c	Cell parameters from 8098 reflections
a = 22.206 (3) Å	θ = 2.6–49.1°
c = 8.692 (2) Å	µ = 0.16 mm^−1
V = 4286.1 (15) Å^3	T = 100 K
Z = 8	Block, colorless
F(000) = 2112	0.10 × 0.09 × 0.09 mm

Data collection

Bruker Smart APEX II Ultra diffractometer	19346 reflections with I > 2σ(I)
Radiation source: BRUKER Rotating Anode	Rint = 0.041
ω scans	θmax = 49.2°, θmin = 1.3°
Absorption correction: multi-scan SADABS 2015/1	h = −47→47
Tmin = 0.928, Tmax = 0.971	k = −46→46
404587 measured reflections	l = −18→18
21597 independent reflections

Refinement

Refinement on F^2	H-atom parameters constrained
Least-squares matrix: full	' w2 = 1/[s^2(Fo^2)]'
R[F^2 > 2σ(F^2)] = 0.018	(Δ/σ)max = 0.001
wR(F^2) = 0.016	Δρmax = 0.30 e Å^−3
S = 1.10	Δρmin = −0.23 e Å^−3
20737 reflections	Absolute structure: Flack x determined using 8376 quotients [(I+)-(I-)]/[(I+)+(I-)] (Parsons, Flack and Wagner, Acta Cryst. B69 (2013) 249-259).
618 parameters	Absolute structure parameter: −0.03 (3)
0 restraints

Fractional atomic coordinates and isotropic or equivalent isotropic displacement parameters (Ų)

	x	y	z	Uiso*/Ueq
F(22)	0.826859 (6)	0.567098 (6)	0.801390 (12)	0.021
F(23)	0.828166 (7)	0.672509 (5)	0.940220 (17)	0.024
F(24)	0.749861 (5)	0.761563 (9)	0.855557 (16)	0.025
F(25)	0.666509 (6)	0.738352 (6)	0.632345 (14)	0.023
F(26)	0.659555 (6)	0.631275 (5)	0.501802 (15)	0.020
F(32)	0.788978 (7)	0.399533 (5)	0.681213 (14)	0.025
F(33)	0.736321 (13)	0.338764 (13)	0.90565 (3)	0.030
F(34)	0.631798 (13)	0.377609 (13)	1.03259 (4)	0.032
F(35)	0.579318 (14)	0.480595 (11)	0.92421 (3)	0.032
F(36)	0.631308 (5)	0.544167 (6)	0.705155 (14)	0.023
N(1)	0.738276 (8)	0.510918 (8)	0.37901 (2)	0.015
C(1)	0.821143 (9)	0.507108 (9)	0.52838 (2)	0.017
C(2)	0.798720 (9)	0.474980 (9)	0.38496 (2)	0.016
C(3)	0.686396 (8)	0.472032 (8)	0.334067 (19)	0.019
C(4)	0.625660 (8)	0.503612 (8)	0.33512 (2)	0.024
C(5)	0.743079 (7)	0.566012 (8)	0.28001 (2)	0.018
C(6)	0.736609 (8)	0.556540 (8)	0.10727 (2)	0.023
C(11)	0.836898 (8)	0.472154 (8)	0.24180 (2)	0.018
C(12)	0.880624 (8)	0.515723 (9)	0.20898 (2)	0.021
C(13)	0.915406 (9)	0.511728 (11)	0.07601 (3)	0.026
C(14)	0.907218 (10)	0.463900 (9)	−0.02590 (3)	0.027
C(15)	0.864743 (10)	0.419543 (10)	0.00732 (2)	0.025
C(16)	0.830121 (9)	0.423620 (9)	0.14049 (2)	0.021
C(21)	0.744941 (7)	0.592987 (8)	0.637502 (18)	0.015
C(22)	0.786415 (7)	0.607899 (8)	0.752331 (19)	0.017
C(23)	0.788456 (8)	0.663106 (8)	0.826657 (19)	0.018
C(24)	0.748368 (7)	0.708093 (9)	0.785898 (19)	0.019
C(25)	0.706097 (8)	0.695748 (8)	0.67313 (2)	0.018
C(26)	0.704451 (7)	0.639175 (7)	0.604678 (19)	0.017
C(31)	0.714866 (8)	0.477614 (8)	0.682257 (19)	0.017
C(32)	0.737918 (8)	0.423083 (8)	0.737745 (19)	0.019
C(33)	0.710942 (10)	0.389436 (8)	0.85399 (2)	0.022
C(34)	0.657472 (9)	0.408874 (8)	0.91900 (2)	0.023
C(35)	0.631152 (9)	0.461276 (9)	0.86465 (2)	0.023
C(36)	0.659926 (8)	0.493656 (8)	0.74951 (2)	0.019
B(1)	0.751686 (8)	0.524439 (8)	0.56965 (2)	0.015
H(1A)	0.846149	0.547366	0.498003	0.02073 (3)
H(1B)	0.849685	0.480102	0.60257	0.02073 (3)
H(2)	0.784218	0.429646	0.413164	0.01952 (3)
H(3A)	0.686187	0.435031	0.415589	0.02230 (3)
H(3B)	0.695806	0.453308	0.221303	0.02230 (3)
H(4A)	0.623262	0.539335	0.250176	0.03604 (5)
H(4B)	0.615649	0.522104	0.447796	0.03604 (5)
H(4C)	0.591466	0.470787	0.304829	0.03604 (5)
H(5A)	0.786275	0.586107	0.307719	0.02141 (3)
H(5B)	0.707919	0.596695	0.317011	0.02141 (3)
H(6A)	0.739564	0.600865	0.055292	0.03483 (5)
H(6B)	0.692692	0.539122	0.075433	0.03483 (5)
H(6C)	0.771055	0.52782	0.057732	0.03483 (5)
H(12)	0.8883	0.552359	0.287591	0.02530 (3)
H(13)	0.949182	0.545052	0.05089	0.03067 (4)
H(14)	0.933892	0.461587	−0.12908	0.03283 (5)
H(15)	0.858457	0.382173	−0.06983	0.03043 (4)
H(16)	0.797812	0.388984	0.166593	0.02514 (3)

Atomic displacement parameters (Ų)

	U^11	U^22	U^33	U^12	U^13	U^23
F(22)	0.02235 (4)	0.02294 (4)	0.01843 (4)	0.00242 (4)	−0.00418 (4)	−0.00055 (4)
F(23)	0.02695 (5)	0.02696 (6)	0.01947 (4)	−0.00660 (4)	−0.00402 (5)	−0.00417 (4)
F(24)	0.03395 (7)	0.01791 (5)	0.02420 (5)	−0.00404 (4)	0.00344 (4)	−0.00628 (5)
F(25)	0.02163 (4)	0.01840 (4)	0.03002 (6)	0.00392 (5)	0.00284 (4)	−0.00197 (4)
F(26)	0.01724 (4)	0.02175 (5)	0.02249 (4)	0.00205 (3)	−0.00287 (4)	−0.00313 (4)
F(32)	0.02907 (6)	0.02183 (5)	0.02419 (5)	0.00712 (4)	0.00462 (5)	0.00431 (4)
F(33)	0.04507 (7)	0.02114 (5)	0.02493 (6)	0.00173 (5)	0.00136 (5)	0.00701 (4)
F(34)	0.04518 (7)	0.02751 (6)	0.02343 (5)	−0.01183 (5)	0.01003 (6)	0.00250 (6)
F(35)	0.02915 (6)	0.03037 (6)	0.03750 (7)	−0.00469 (5)	0.01673 (5)	−0.00215 (5)
F(36)	0.01893 (4)	0.02114 (5)	0.03019 (5)	0.00132 (4)	0.00473 (4)	0.00120 (4)
N(1)	0.01597 (7)	0.01511 (7)	0.01426 (6)	−0.00040 (6)	−0.00023 (6)	−0.00029 (6)
C(1)	0.01590 (7)	0.01941 (7)	0.01651 (7)	0.00083 (6)	−0.00046 (6)	−0.00132 (6)
C(2)	0.01747 (7)	0.01605 (7)	0.01527 (7)	0.00072 (6)	0.00132 (6)	−0.00040 (6)
C(3)	0.01940 (7)	0.01758 (6)	0.01876 (7)	−0.00186 (5)	−0.00144 (5)	−0.00208 (5)
C(4)	0.01812 (7)	0.02443 (7)	0.02951 (8)	−0.00113 (6)	−0.00338 (6)	−0.00204 (6)
C(5)	0.02054 (7)	0.01644 (6)	0.01654 (6)	0.00031 (5)	0.00017 (5)	0.00108 (5)
C(6)	0.02817 (8)	0.02549 (7)	0.01600 (7)	0.00309 (6)	−0.00060 (6)	0.00283 (5)
C(11)	0.01863 (6)	0.01796 (7)	0.01628 (6)	0.00169 (6)	0.00269 (6)	−0.00014 (6)
C(12)	0.02000 (7)	0.02154 (7)	0.02172 (7)	0.00010 (6)	0.00434 (6)	0.00140 (6)
C(13)	0.02321 (8)	0.02774 (9)	0.02573 (8)	0.00309 (7)	0.00857 (7)	0.00540 (8)
C(14)	0.02864 (8)	0.03155 (11)	0.02189 (8)	0.00861 (9)	0.00945 (8)	0.00347 (7)
C(15)	0.03039 (8)	0.02688 (8)	0.01881 (7)	0.00730 (8)	0.00525 (7)	−0.00238 (7)
C(16)	0.02453 (7)	0.02023 (7)	0.01809 (6)	0.00271 (6)	0.00315 (6)	−0.00235 (6)
C(21)	0.01590 (6)	0.01533 (6)	0.01500 (5)	−0.00062 (5)	0.00015 (5)	−0.00143 (5)
C(22)	0.01774 (6)	0.01759 (6)	0.01496 (5)	−0.00134 (5)	−0.00048 (5)	−0.00088 (5)
C(23)	0.02045 (6)	0.01888 (6)	0.01583 (5)	−0.00402 (5)	0.00001 (5)	−0.00232 (5)
C(24)	0.02241 (7)	0.01646 (6)	0.01827 (6)	−0.00291 (5)	0.00271 (6)	−0.00320 (5)
C(25)	0.01830 (6)	0.01554 (6)	0.02001 (6)	0.00022 (5)	0.00263 (5)	−0.00209 (5)
C(26)	0.01582 (6)	0.01626 (6)	0.01746 (6)	0.00046 (5)	0.00038 (5)	−0.00182 (5)
C(31)	0.01896 (6)	0.01633 (6)	0.01525 (6)	−0.00037 (5)	0.00146 (5)	0.00029 (5)
C(32)	0.02420 (7)	0.01717 (6)	0.01679 (6)	0.00097 (5)	0.00152 (5)	0.00163 (5)
C(33)	0.03150 (8)	0.01838 (6)	0.01745 (6)	−0.00216 (6)	0.00167 (6)	0.00218 (5)
C(34)	0.03126 (8)	0.02072 (7)	0.01818 (7)	−0.00697 (7)	0.00492 (6)	0.00005 (6)
C(35)	0.02436 (7)	0.02154 (7)	0.02184 (7)	−0.00541 (6)	0.00678 (6)	−0.00194 (6)
C(36)	0.01889 (6)	0.01826 (6)	0.01940 (6)	−0.00181 (5)	0.00324 (6)	−0.00072 (5)
B(1)	0.01598 (7)	0.01570 (7)	0.01479 (7)	0.00019 (6)	0.00048 (6)	−0.00061 (6)

Geometric parameters (Å, º)

F(22)—C(22)	1.3451 (2)	C(6)—H(6C)	1.0850 (2)
F(23)—C(23)	1.3400 (2)	C(11)—C(12)	1.4001 (3)
F(24)—C(24)	1.3332 (3)	C(11)—C(16)	1.3999 (3)
F(25)—C(25)	1.3392 (2)	C(12)—C(13)	1.3929 (3)
F(26)—C(26)	1.3507 (2)	C(12)—H(12)	1.0760 (2)
F(32)—C(32)	1.3418 (2)	C(13)—C(14)	1.3949 (3)
F(33)—C(33)	1.3362 (4)	C(13)—H(13)	1.0760 (2)
F(34)—C(34)	1.3349 (4)	C(14)—C(15)	1.3940 (3)
F(35)—C(35)	1.3330 (4)	C(14)—H(14)	1.0760 (3)
F(36)—C(36)	1.3456 (2)	C(15)—C(16)	1.3925 (3)
C(1)—C(2)	1.5201 (3)	C(15)—H(15)	1.0760 (2)
C(1)—B(1)	1.6296 (3)	C(16)—H(16)	1.0760 (2)
C(1)—H(1A)	1.0850 (2)	C(21)—C(22)	1.3979 (2)
C(1)—H(1B)	1.0850 (2)	C(21)—C(26)	1.3935 (2)
C(2)—C(11)	1.5070 (3)	C(21)—B(1)	1.6393 (2)
C(2)—H(2)	1.0850 (2)	C(22)—C(23)	1.3865 (2)
C(3)—C(4)	1.5202 (3)	C(23)—C(24)	1.3842 (3)
C(3)—H(3A)	1.0850 (2)	C(24)—C(25)	1.3846 (2)
C(3)—H(3B)	1.0850 (2)	C(25)—C(26)	1.3905 (2)
C(4)—H(4A)	1.0850 (2)	C(31)—C(32)	1.4003 (2)
C(4)—H(4B)	1.0850 (2)	C(31)—C(36)	1.3989 (2)
C(4)—H(4C)	1.0850 (2)	C(31)—B(1)	1.6455 (3)
C(5)—C(6)	1.5229 (3)	C(32)—C(33)	1.3921 (2)
C(5)—H(5A)	1.0850 (2)	C(33)—C(34)	1.3840 (3)
C(5)—H(5B)	1.0850 (2)	C(34)—C(35)	1.3852 (3)
C(6)—H(6A)	1.0850 (2)	C(35)—C(36)	1.3881 (2)
C(6)—H(6B)	1.0850 (2)
C(2)—C(1)—B(1)	88.932 (14)	C(11)—C(16)—C(15)	120.90 (2)
C(2)—C(1)—H(1A)	110.781 (16)	C(11)—C(16)—H(16)	119.299 (19)
C(2)—C(1)—H(1B)	114.793 (16)	C(15)—C(16)—H(16)	119.80 (2)
B(1)—C(1)—H(1A)	110.082 (15)	C(22)—C(21)—C(26)	113.387 (17)
B(1)—C(1)—H(1B)	123.542 (16)	C(22)—C(21)—B(1)	114.623 (15)
H(1A)—C(1)—H(1B)	107.522 (17)	C(26)—C(21)—B(1)	131.971 (15)
C(1)—C(2)—C(11)	120.823 (17)	F(22)—C(22)—C(21)	120.444 (15)
C(1)—C(2)—H(2)	110.329 (16)	F(22)—C(22)—C(23)	115.216 (15)
C(11)—C(2)—H(2)	108.354 (16)	C(21)—C(22)—C(23)	124.312 (17)
C(4)—C(3)—H(3A)	109.976 (15)	F(23)—C(23)—C(22)	120.168 (17)
C(4)—C(3)—H(3B)	110.676 (15)	F(23)—C(23)—C(24)	119.954 (15)
H(3A)—C(3)—H(3B)	107.493 (16)	C(22)—C(23)—C(24)	119.859 (17)
C(3)—C(4)—H(4A)	112.131 (16)	F(24)—C(24)—C(23)	120.699 (15)
C(3)—C(4)—H(4B)	111.211 (15)	F(24)—C(24)—C(25)	120.977 (16)
C(3)—C(4)—H(4C)	108.031 (16)	C(23)—C(24)—C(25)	118.315 (19)
H(4A)—C(4)—H(4B)	109.116 (17)	F(25)—C(25)—C(24)	119.509 (16)
H(4A)—C(4)—H(4C)	106.960 (16)	F(25)—C(25)—C(26)	120.511 (16)
H(4B)—C(4)—H(4C)	109.263 (16)	C(24)—C(25)—C(26)	119.975 (18)
C(6)—C(5)—H(5A)	111.042 (15)	F(26)—C(26)—C(21)	121.079 (15)
C(6)—C(5)—H(5B)	108.124 (14)	F(26)—C(26)—C(25)	114.836 (15)
H(5A)—C(5)—H(5B)	108.186 (15)	C(21)—C(26)—C(25)	124.084 (17)
C(5)—C(6)—H(6A)	106.233 (15)	C(32)—C(31)—C(36)	113.271 (17)
C(5)—C(6)—H(6B)	112.674 (15)	C(32)—C(31)—B(1)	124.726 (16)
C(5)—C(6)—H(6C)	113.950 (15)	C(36)—C(31)—B(1)	121.397 (15)
H(6A)—C(6)—H(6B)	105.756 (15)	F(32)—C(32)—C(31)	121.317 (15)
H(6A)—C(6)—H(6C)	108.986 (16)	F(32)—C(32)—C(33)	114.848 (16)
H(6B)—C(6)—H(6C)	108.838 (16)	C(31)—C(32)—C(33)	123.834 (18)
C(2)—C(11)—C(12)	121.979 (18)	F(33)—C(33)—C(32)	120.96 (2)
C(2)—C(11)—C(16)	119.406 (19)	F(33)—C(33)—C(34)	119.161 (19)
C(12)—C(11)—C(16)	118.58 (2)	C(32)—C(33)—C(34)	119.875 (19)
C(11)—C(12)—C(13)	120.64 (2)	F(34)—C(34)—C(33)	120.41 (2)
C(11)—C(12)—H(12)	120.195 (19)	F(34)—C(34)—C(35)	120.59 (2)
C(13)—C(12)—H(12)	119.16 (2)	C(33)—C(34)—C(35)	118.99 (2)
C(12)—C(13)—C(14)	120.21 (2)	F(35)—C(35)—C(34)	120.147 (19)
C(12)—C(13)—H(13)	120.74 (2)	F(35)—C(35)—C(36)	120.71 (2)
C(14)—C(13)—H(13)	119.05 (2)	C(34)—C(35)—C(36)	119.131 (19)
C(13)—C(14)—C(15)	119.65 (2)	F(36)—C(36)—C(31)	120.294 (16)
C(13)—C(14)—H(14)	119.60 (2)	F(36)—C(36)—C(35)	114.895 (16)
C(15)—C(14)—H(14)	120.76 (2)	C(31)—C(36)—C(35)	124.805 (18)
C(14)—C(15)—C(16)	119.99 (2)	C(1)—B(1)—C(21)	112.642 (14)
C(14)—C(15)—H(15)	120.24 (2)	C(1)—B(1)—C(31)	116.872 (14)
C(16)—C(15)—H(15)	119.77 (2)	C(21)—B(1)—C(31)	109.109 (13)

(2_shelx)

Crystal data

C20H23PS	Dx = 1.291 Mg m^−3
Mr = 326.41	Mo Kα radiation, λ = 0.71073 Å
Orthorhombic, Pbca	Cell parameters from 9781 reflections
a = 10.9323 (7) Å	θ = 2.7–52.2°
b = 14.5698 (10) Å	µ = 0.28 mm^−1
c = 21.0910 (14) Å	T = 99 K
V = 3359.4 (4) Å^3	Block, yellow
Z = 8	0.12 × 0.11 × 0.10 mm
F(000) = 1392

Data collection

Bruker Smart APEX II Ultra diffractometer	15462 reflections with I > 2σ(I)
Radiation source: BRUKER Rotating Anode	Rint = 0.025
ω scans	θmax = 52.2°, θmin = 1.9°
Absorption correction: multi-scan SADABS-2016/2	h = −24→24
Tmin = 0.885, Tmax = 0.929	k = −32→32
222190 measured reflections	l = −46→46
19329 independent reflections

Refinement

Refinement on F^2	0 restraints
Least-squares matrix: full	Hydrogen site location: inferred from neighbouring sites
R[F^2 > 2σ(F^2)] = 0.024	H-atom parameters constrained
wR(F^2) = 0.090	w = 1/[σ^2(Fo^2) + (0.0508P)^2 + 0.0744P] where P = (Fo^2 + 2Fc^2)/3
S = 1.11	(Δ/σ)max = 0.007
19329 reflections	Δρmax = 0.60 e Å^−3
203 parameters	Δρmin = −0.27 e Å^−3

Special details

Geometry. All esds (except the esd in the dihedral angle between two l.s. planes) are estimated using the full covariance matrix. The cell esds are taken into account individually in the estimation of esds in distances, angles and torsion angles; correlations between esds in cell parameters are only used when they are defined by crystal symmetry. An approximate (isotropic) treatment of cell esds is used for estimating esds involving l.s. planes.

Fractional atomic coordinates and isotropic or equivalent isotropic displacement parameters (Ų)

	x	y	z	Uiso*/Ueq
P1	0.43671 (2)	0.29822 (2)	0.56327 (2)	0.00922 (2)
S1	0.60402 (2)	0.33080 (2)	0.53581 (2)	0.01521 (2)
C1	0.40206 (3)	0.31454 (2)	0.64870 (2)	0.01034 (4)
C2	0.30308 (3)	0.26463 (2)	0.67638 (2)	0.01075 (4)
C4	0.11899 (3)	0.16839 (3)	0.66988 (2)	0.01492 (5)
H4	0.058349	0.139171	0.644711	0.018*
C3	0.20688 (3)	0.22113 (2)	0.64096 (2)	0.01242 (4)
H3	0.204017	0.229091	0.596281	0.015*
C6	0.20005 (3)	0.20333 (3)	0.77300 (2)	0.01496 (5)
H6	0.195897	0.198935	0.817888	0.018*
C5	0.11708 (3)	0.15670 (3)	0.73686 (2)	0.01648 (5)
H5	0.058972	0.117058	0.756143	0.020*
C9	0.46281 (3)	0.36539 (2)	0.75644 (2)	0.01249 (4)
C8	0.37449 (3)	0.30767 (2)	0.78264 (2)	0.01384 (4)
H8	0.369591	0.301534	0.827413	0.017*
C7	0.29318 (3)	0.25872 (2)	0.74445 (2)	0.01205 (4)
C10	0.54052 (4)	0.41836 (3)	0.79718 (2)	0.01576 (5)
H10	0.532674	0.412301	0.841845	0.019*
C11	0.62554 (4)	0.47733 (3)	0.77299 (2)	0.01712 (5)
H11	0.677982	0.510914	0.800405	0.021*
C12	0.63443 (4)	0.48772 (3)	0.70613 (2)	0.01678 (5)
H12	0.690985	0.530669	0.689118	0.020*
C13	0.56309 (3)	0.43705 (2)	0.66565 (2)	0.01450 (5)
H13	0.571264	0.446018	0.621240	0.017*
C14	0.47650 (3)	0.37091 (2)	0.68854 (2)	0.01117 (4)
C15	0.41687 (3)	0.17399 (2)	0.54644 (2)	0.01192 (4)
H15	0.331068	0.156393	0.557495	0.014*
C16	0.43759 (4)	0.15348 (3)	0.47599 (2)	0.01605 (5)
H16A	0.519861	0.173469	0.463738	0.024*
H16C	0.376730	0.186496	0.450617	0.024*
H16B	0.429444	0.087369	0.468551	0.024*
C17	0.50405 (4)	0.11832 (3)	0.58809 (2)	0.01940 (6)
H17A	0.486521	0.130758	0.632838	0.029*
H17C	0.588664	0.135782	0.578596	0.029*
H17B	0.492851	0.052749	0.579522	0.029*
C18	0.31688 (3)	0.36258 (2)	0.51945 (2)	0.01225 (4)
H18	0.243807	0.321749	0.515231	0.015*
C19	0.35876 (4)	0.38913 (3)	0.45265 (2)	0.01911 (6)
H19A	0.432669	0.426977	0.455445	0.029*
H19B	0.293831	0.423951	0.431486	0.029*
H19C	0.376578	0.333465	0.428258	0.029*
C20	0.27772 (4)	0.44846 (3)	0.55624 (2)	0.01955 (6)
H20A	0.245183	0.430371	0.597694	0.029*
H20B	0.214382	0.481141	0.532330	0.029*
H20C	0.348555	0.488766	0.562210	0.029*

Atomic displacement parameters (Ų)

	U^11	U^22	U^33	U^12	U^13	U^23
P1	0.00966 (3)	0.00933 (3)	0.00867 (3)	−0.00075 (2)	0.00098 (2)	−0.00078 (2)
S1	0.01136 (3)	0.01617 (3)	0.01811 (4)	−0.00326 (2)	0.00435 (2)	−0.00375 (2)
C1	0.01183 (9)	0.01038 (8)	0.00880 (8)	−0.00040 (7)	0.00017 (7)	−0.00053 (7)
C2	0.01143 (9)	0.01123 (9)	0.00958 (8)	0.00031 (7)	0.00080 (7)	0.00003 (7)
C4	0.01242 (10)	0.01699 (12)	0.01534 (11)	−0.00203 (8)	0.00165 (8)	−0.00001 (9)
C3	0.01142 (9)	0.01434 (10)	0.01149 (9)	−0.00088 (8)	0.00079 (7)	−0.00042 (8)
C6	0.01548 (11)	0.01714 (12)	0.01227 (10)	−0.00029 (9)	0.00285 (8)	0.00252 (8)
C5	0.01480 (11)	0.01863 (13)	0.01600 (11)	−0.00226 (9)	0.00318 (9)	0.00265 (10)
C9	0.01581 (10)	0.01208 (9)	0.00958 (8)	0.00048 (8)	−0.00150 (8)	−0.00077 (7)
C8	0.01701 (11)	0.01531 (11)	0.00920 (9)	0.00022 (9)	0.00000 (8)	0.00019 (8)
C7	0.01340 (10)	0.01293 (10)	0.00983 (8)	0.00075 (8)	0.00114 (7)	0.00080 (7)
C10	0.02043 (13)	0.01572 (11)	0.01114 (9)	−0.00081 (10)	−0.00372 (9)	−0.00173 (8)
C11	0.02093 (13)	0.01564 (12)	0.01478 (11)	−0.00247 (10)	−0.00493 (10)	−0.00240 (9)
C12	0.02077 (13)	0.01437 (11)	0.01520 (11)	−0.00526 (10)	−0.00252 (10)	−0.00138 (9)
C13	0.01903 (12)	0.01251 (10)	0.01196 (10)	−0.00426 (9)	−0.00104 (9)	−0.00057 (8)
C14	0.01384 (10)	0.01010 (8)	0.00956 (8)	−0.00033 (7)	−0.00105 (7)	−0.00057 (7)
C15	0.01332 (10)	0.01044 (9)	0.01200 (9)	−0.00067 (7)	0.00129 (8)	−0.00182 (7)
C16	0.01854 (13)	0.01643 (12)	0.01318 (10)	−0.00112 (9)	0.00189 (9)	−0.00490 (9)
C17	0.02619 (16)	0.01329 (11)	0.01871 (13)	0.00420 (11)	−0.00341 (12)	−0.00006 (10)
C18	0.01270 (9)	0.01309 (10)	0.01097 (9)	−0.00054 (8)	0.00020 (7)	0.00197 (7)
C19	0.02249 (15)	0.02290 (15)	0.01195 (10)	0.00001 (12)	0.00153 (10)	0.00518 (10)
C20	0.02356 (15)	0.01691 (13)	0.01817 (13)	0.00773 (11)	0.00076 (11)	0.00065 (10)

Geometric parameters (Å, º)

P1—C1	1.8565 (3)	C11—H11	0.9500
P1—C15	1.8572 (3)	C12—C13	1.3719 (5)
P1—C18	1.8573 (3)	C12—H12	0.9500
P1—S1	1.9765 (2)	C13—C14	1.4345 (5)
C1—C2	1.4285 (4)	C13—H13	0.9500
C1—C14	1.4292 (4)	C15—C17	1.5290 (5)
C2—C3	1.4373 (4)	C15—C16	1.5325 (5)
C2—C7	1.4423 (4)	C15—H15	1.0000
C4—C3	1.3732 (5)	C16—H16A	0.9800
C4—C5	1.4230 (5)	C16—H16C	0.9800
C4—H4	0.9500	C16—H16B	0.9800
C3—H3	0.9500	C17—H17A	0.9800
C6—C5	1.3658 (5)	C17—H17C	0.9800
C6—C7	1.4319 (5)	C17—H17B	0.9800
C6—H6	0.9500	C18—C19	1.5310 (5)
C5—H5	0.9500	C18—C20	1.5334 (5)
C9—C8	1.3946 (5)	C18—H18	1.0000
C9—C10	1.4338 (5)	C19—H19A	0.9800
C9—C14	1.4421 (4)	C19—H19B	0.9800
C8—C7	1.3955 (5)	C19—H19C	0.9800
C8—H8	0.9500	C20—H20A	0.9800
C10—C11	1.3647 (6)	C20—H20B	0.9800
C10—H10	0.9500	C20—H20C	0.9800
C11—C12	1.4215 (5)
C1—P1—C15	106.649 (14)	C12—C13—C14	121.82 (3)
C1—P1—C18	105.929 (14)	C12—C13—H13	119.1
C15—P1—C18	108.335 (15)	C14—C13—H13	119.1
C1—P1—S1	116.260 (11)	C1—C14—C13	124.34 (3)
C15—P1—S1	106.626 (11)	C1—C14—C9	119.51 (3)
C18—P1—S1	112.682 (12)	C13—C14—C9	116.11 (3)
C2—C1—C14	118.92 (3)	C17—C15—C16	111.19 (3)
C2—C1—P1	119.08 (2)	C17—C15—P1	109.54 (2)
C14—C1—P1	121.87 (2)	C16—C15—P1	110.98 (2)
C1—C2—C3	124.49 (3)	C17—C15—H15	108.3
C1—C2—C7	119.61 (3)	C16—C15—H15	108.3
C3—C2—C7	115.86 (3)	P1—C15—H15	108.3
C3—C4—C5	121.21 (3)	C15—C16—H16A	109.5
C3—C4—H4	119.4	C15—C16—H16C	109.5
C5—C4—H4	119.4	H16A—C16—H16C	109.5
C4—C3—C2	121.86 (3)	C15—C16—H16B	109.5
C4—C3—H3	119.1	H16A—C16—H16B	109.5
C2—C3—H3	119.1	H16C—C16—H16B	109.5
C5—C6—C7	121.20 (3)	C15—C17—H17A	109.5
C5—C6—H6	119.4	C15—C17—H17C	109.5
C7—C6—H6	119.4	H17A—C17—H17C	109.5
C6—C5—C4	118.99 (3)	C15—C17—H17B	109.5
C6—C5—H5	120.5	H17A—C17—H17B	109.5
C4—C5—H5	120.5	H17C—C17—H17B	109.5
C8—C9—C10	119.82 (3)	C19—C18—C20	110.06 (3)
C8—C9—C14	119.94 (3)	C19—C18—P1	112.00 (2)
C10—C9—C14	120.24 (3)	C20—C18—P1	110.92 (2)
C9—C8—C7	121.36 (3)	C19—C18—H18	107.9
C9—C8—H8	119.3	C20—C18—H18	107.9
C7—C8—H8	119.3	P1—C18—H18	107.9
C8—C7—C6	119.89 (3)	C18—C19—H19A	109.5
C8—C7—C2	119.75 (3)	C18—C19—H19B	109.5
C6—C7—C2	120.36 (3)	H19A—C19—H19B	109.5
C11—C10—C9	121.22 (3)	C18—C19—H19C	109.5
C11—C10—H10	119.4	H19A—C19—H19C	109.5
C9—C10—H10	119.4	H19B—C19—H19C	109.5
C10—C11—C12	118.98 (3)	C18—C20—H20A	109.5
C10—C11—H11	120.5	C18—C20—H20B	109.5
C12—C11—H11	120.5	H20A—C20—H20B	109.5
C13—C12—C11	121.40 (3)	C18—C20—H20C	109.5
C13—C12—H12	119.3	H20A—C20—H20C	109.5
C11—C12—H12	119.3	H20B—C20—H20C	109.5

(2_xd)

Crystal data

C20H23PS	Dx = 1.291 Mg m^−3
Mr = 326.41	Mo Kα radiation, λ = 0.71073 Å
Orthorhombic, Pbca	Cell parameters from 9781 reflections
a = 10.932 (2) Å	θ = 2.7–52.2°
b = 14.570 (2) Å	µ = 0.28 mm^−1
c = 21.091 (3) Å	T = 99 K
V = 3359.4 (9) Å^3	Block, yellow
Z = 8	0.12 × 0.11 × 0.10 mm
F(000) = 1392

Data collection

Bruker Smart APEX II Ultra diffractometer	15462 reflections with I > 2σ(I)
Radiation source: BRUKER Rotating Anode	Rint = 0.025
ω scans	θmax = 52.2°, θmin = 1.9°
Absorption correction: multi-scan SADABS-2016/2	h = −24→24
Tmin = 0.885, Tmax = 0.929	k = −32→32
222190 measured reflections	l = −46→46
19329 independent reflections

Refinement

Refinement on F^2	0 restraints
Least-squares matrix: full	Hydrogen site location: inferred from neighbouring sites
R[F^2 > 2σ(F^2)] = 0.018	H-atom parameters constrained
wR(F^2) = 0.016	' w2 = 1/[s^2(Fo^2)]'
S = 1.20	(Δ/σ)max < 0.001
18435 reflections	Δρmax = 0.22 e Å^−3
332 parameters	Δρmin = −0.20 e Å^−3

Fractional atomic coordinates and isotropic or equivalent isotropic displacement parameters (Ų)

	x	y	z	Uiso*/Ueq
S(1)	0.604030 (3)	0.330802 (3)	0.535802 (2)	0.015
P(1)	0.436726 (3)	0.298216 (2)	0.563275 (2)	0.009
C(1)	0.402036 (13)	0.314540 (9)	0.648694 (7)	0.010
C(2)	0.303081 (13)	0.264624 (9)	0.676372 (6)	0.011
C(3)	0.206907 (13)	0.221082 (9)	0.640947 (7)	0.012
C(4)	0.119010 (13)	0.168349 (10)	0.669895 (7)	0.015
C(5)	0.117016 (14)	0.156675 (11)	0.736838 (7)	0.016
C(6)	0.200110 (14)	0.203308 (10)	0.773012 (7)	0.015
C(7)	0.293180 (13)	0.258724 (9)	0.744461 (7)	0.012
C(8)	0.374461 (14)	0.307695 (10)	0.782637 (7)	0.014
C(9)	0.462788 (13)	0.365439 (10)	0.756440 (7)	0.012
C(10)	0.540525 (15)	0.418287 (10)	0.797177 (7)	0.016
C(11)	0.625718 (15)	0.477389 (11)	0.772952 (7)	0.017
C(12)	0.634569 (15)	0.487663 (10)	0.706141 (7)	0.017
C(13)	0.563095 (14)	0.437053 (10)	0.665667 (7)	0.015
C(14)	0.476514 (13)	0.370879 (9)	0.688534 (6)	0.011
C(15)	0.416881 (13)	0.173961 (10)	0.546454 (6)	0.012
C(16)	0.437598 (14)	0.153514 (10)	0.475988 (7)	0.016
C(17)	0.504184 (16)	0.118305 (10)	0.588112 (8)	0.019
C(18)	0.316860 (13)	0.362555 (10)	0.519445 (6)	0.012
C(19)	0.358729 (15)	0.389095 (11)	0.452680 (7)	0.019
C(20)	0.277762 (15)	0.448357 (11)	0.556255 (7)	0.020
H(3)	0.197651	0.232378	0.590422	0.01490 (2)
H(4)	0.049147	0.136662	0.640804	0.01786 (3)
H(5)	0.048275	0.113192	0.758399	0.01965 (3)
H(6)	0.197779	0.2016	0.824333	0.01794 (3)
H(8)	0.36641	0.302111	0.83367	0.01655 (3)
H(10)	0.52779	0.410203	0.847793	0.01894 (3)
H(11)	0.684501	0.517412	0.803676	0.02053 (3)
H(12)	0.698036	0.537317	0.686702	0.02031 (3)
H(13)	0.570857	0.44906	0.615153	0.01743 (3)
H(15)	0.322198	0.15507	0.558131	0.01430 (2)
H(16A)	0.526384	0.17641	0.460509	0.02409 (3)
H(16B)	0.431018	0.080369	0.469457	0.02409 (3)
H(16C)	0.369336	0.185059	0.446309	0.02409 (3)
H(17A)	0.487181	0.129244	0.63784	0.02905 (4)
H(17B)	0.49115	0.046477	0.578123	0.02905 (4)
H(17C)	0.597858	0.135302	0.577517	0.02905 (4)
H(18)	0.240258	0.313953	0.516063	0.01472 (2)
H(19A)	0.439974	0.430751	0.455154	0.02875 (4)
H(19B)	0.28732	0.428366	0.430283	0.02875 (4)
H(19C)	0.378086	0.330325	0.423378	0.02875 (4)
H(20A)	0.240305	0.431831	0.602084	0.02943 (4)
H(20B)	0.2097	0.483589	0.528494	0.02943 (4)
H(20C)	0.353903	0.494179	0.563184	0.02943 (4)

Atomic displacement parameters (Ų)

	U^11	U^22	U^33	U^12	U^13	U^23
S(1)	0.011368 (13)	0.016133 (15)	0.018203 (15)	−0.003237 (11)	0.004317 (11)	−0.003787 (12)
P(1)	0.009546 (12)	0.009326 (12)	0.008771 (12)	−0.000740 (10)	0.001024 (10)	−0.000802 (10)
C(1)	0.01180 (5)	0.01046 (5)	0.00875 (4)	−0.00066 (4)	0.00035 (4)	−0.00068 (4)
C(2)	0.01140 (5)	0.01136 (5)	0.00927 (4)	−0.00019 (4)	0.00088 (4)	−0.00004 (4)
C(3)	0.01121 (5)	0.01485 (5)	0.01118 (5)	−0.00124 (4)	0.00082 (4)	−0.00049 (4)
C(4)	0.01237 (5)	0.01736 (6)	0.01492 (5)	−0.00259 (5)	0.00161 (4)	0.00019 (4)
C(5)	0.01464 (6)	0.01898 (6)	0.01552 (5)	−0.00268 (5)	0.00311 (5)	0.00255 (5)
C(6)	0.01543 (5)	0.01734 (6)	0.01207 (5)	−0.00069 (5)	0.00273 (4)	0.00265 (4)
C(7)	0.01325 (5)	0.01324 (5)	0.00958 (5)	0.00037 (4)	0.00122 (4)	0.00078 (4)
C(8)	0.01682 (6)	0.01548 (6)	0.00907 (5)	−0.00036 (5)	0.00001 (4)	0.00002 (4)
C(9)	0.01593 (5)	0.01209 (5)	0.00929 (4)	−0.00004 (4)	−0.00149 (4)	−0.00077 (4)
C(10)	0.02058 (6)	0.01593 (6)	0.01085 (5)	−0.00142 (5)	−0.00376 (4)	−0.00161 (4)
C(11)	0.02105 (6)	0.01584 (6)	0.01443 (6)	−0.00319 (5)	−0.00488 (5)	−0.00224 (4)
C(12)	0.02114 (6)	0.01479 (5)	0.01484 (6)	−0.00582 (5)	−0.00260 (5)	−0.00147 (4)
C(13)	0.01926 (6)	0.01257 (5)	0.01174 (5)	−0.00476 (5)	−0.00110 (4)	−0.00057 (4)
C(14)	0.01383 (5)	0.01016 (5)	0.00940 (4)	−0.00072 (4)	−0.00108 (4)	−0.00062 (4)
C(15)	0.01327 (5)	0.01044 (5)	0.01203 (5)	−0.00066 (4)	0.00134 (4)	−0.00198 (4)
C(16)	0.01842 (6)	0.01649 (6)	0.01326 (5)	−0.00103 (5)	0.00200 (4)	−0.00511 (4)
C(17)	0.02615 (7)	0.01323 (6)	0.01872 (6)	0.00434 (5)	−0.00330 (5)	0.00001 (5)
C(18)	0.01259 (5)	0.01309 (5)	0.01112 (5)	−0.00053 (4)	0.00021 (4)	0.00206 (4)
C(19)	0.02267 (7)	0.02280 (7)	0.01203 (5)	0.00011 (5)	0.00147 (5)	0.00533 (5)
C(20)	0.02364 (7)	0.01688 (6)	0.01833 (6)	0.00791 (5)	0.00080 (5)	0.00061 (5)

Geometric parameters (Å, º)

S(1)—P(1)	1.9764 (1)	C(11)—H(11)	1.0830 (2)
P(1)—C(1)	1.8564 (2)	C(12)—C(13)	1.3722 (2)
P(1)—C(15)	1.8575 (1)	C(12)—H(12)	1.0830 (2)
P(1)—C(18)	1.8575 (2)	C(13)—C(14)	1.4346 (2)
C(1)—C(2)	1.4283 (2)	C(13)—H(13)	1.0830 (1)
C(1)—C(14)	1.4293 (2)	C(15)—C(16)	1.5326 (2)
C(2)—C(3)	1.4374 (2)	C(15)—C(17)	1.5298 (2)
C(2)—C(7)	1.4427 (2)	C(15)—H(15)	1.0990 (1)
C(3)—C(4)	1.3735 (2)	C(16)—H(16A)	1.0770 (2)
C(3)—H(3)	1.0830 (1)	C(16)—H(16B)	1.0770 (1)
C(4)—C(5)	1.4223 (2)	C(16)—H(16C)	1.0770 (2)
C(4)—H(4)	1.0830 (2)	C(17)—H(17A)	1.0770 (2)
C(5)—C(6)	1.3671 (2)	C(17)—H(17B)	1.0770 (2)
C(5)—H(5)	1.0830 (2)	C(17)—H(17C)	1.0770 (2)
C(6)—C(7)	1.4317 (2)	C(18)—C(19)	1.5303 (2)
C(6)—H(6)	1.0830 (2)	C(18)—C(20)	1.5324 (2)
C(7)—C(8)	1.3953 (2)	C(18)—H(18)	1.0990 (2)
C(8)—C(9)	1.3948 (2)	C(19)—H(19A)	1.0770 (2)
C(8)—H(8)	1.0830 (2)	C(19)—H(19B)	1.0770 (2)
C(9)—C(10)	1.4329 (2)	C(19)—H(19C)	1.0770 (2)
C(9)—C(14)	1.4422 (2)	C(20)—H(20A)	1.0770 (2)
C(10)—C(11)	1.3675 (2)	C(20)—H(20B)	1.0770 (2)
C(10)—H(10)	1.0830 (2)	C(20)—H(20C)	1.0770 (2)
C(11)—C(12)	1.4203 (2)
S(1)—P(1)—C(1)	116.282 (5)	C(12)—C(13)—C(14)	121.850 (13)
S(1)—P(1)—C(15)	106.631 (5)	C(12)—C(13)—H(13)	118.722 (14)
S(1)—P(1)—C(18)	112.686 (5)	C(14)—C(13)—H(13)	119.405 (13)
C(1)—P(1)—C(15)	106.640 (6)	C(1)—C(14)—C(9)	119.537 (12)
C(1)—P(1)—C(18)	105.915 (6)	C(1)—C(14)—C(13)	124.346 (13)
C(15)—P(1)—C(18)	108.325 (7)	C(9)—C(14)—C(13)	116.067 (12)
P(1)—C(1)—C(2)	119.088 (10)	P(1)—C(15)—C(16)	110.941 (9)
P(1)—C(1)—C(14)	121.855 (11)	P(1)—C(15)—C(17)	109.515 (10)
C(2)—C(1)—C(14)	118.923 (13)	P(1)—C(15)—H(15)	108.137 (10)
C(1)—C(2)—C(3)	124.492 (12)	C(16)—C(15)—C(17)	111.205 (12)
C(1)—C(2)—C(7)	119.604 (12)	C(16)—C(15)—H(15)	107.927 (11)
C(3)—C(2)—C(7)	115.864 (12)	C(17)—C(15)—H(15)	109.033 (12)
C(2)—C(3)—C(4)	121.841 (13)	C(15)—C(16)—H(16A)	111.526 (12)
C(2)—C(3)—H(3)	120.845 (13)	C(15)—C(16)—H(16B)	107.849 (12)
C(4)—C(3)—H(3)	117.197 (13)	C(15)—C(16)—H(16C)	112.226 (12)
C(3)—C(4)—C(5)	121.259 (13)	H(16A)—C(16)—H(16B)	109.143 (13)
C(3)—C(4)—H(4)	118.686 (14)	H(16A)—C(16)—H(16C)	108.426 (13)
C(5)—C(4)—H(4)	120.040 (13)	H(16B)—C(16)—H(16C)	107.558 (12)
C(4)—C(5)—C(6)	118.973 (13)	C(15)—C(17)—H(17A)	111.911 (13)
C(4)—C(5)—H(5)	119.830 (14)	C(15)—C(17)—H(17B)	108.674 (13)
C(6)—C(5)—H(5)	121.163 (14)	C(15)—C(17)—H(17C)	110.618 (13)
C(5)—C(6)—C(7)	121.183 (14)	H(17A)—C(17)—H(17B)	108.147 (14)
C(5)—C(6)—H(6)	122.055 (14)	H(17A)—C(17)—H(17C)	109.398 (14)
C(7)—C(6)—H(6)	116.747 (14)	H(17B)—C(17)—H(17C)	107.978 (14)
C(2)—C(7)—C(6)	120.370 (13)	P(1)—C(18)—C(19)	111.991 (10)
C(2)—C(7)—C(8)	119.745 (13)	P(1)—C(18)—C(20)	110.875 (10)
C(6)—C(7)—C(8)	119.883 (13)	P(1)—C(18)—H(18)	104.162 (10)
C(7)—C(8)—C(9)	121.380 (14)	C(19)—C(18)—C(20)	110.088 (12)
C(7)—C(8)—H(8)	118.903 (15)	C(19)—C(18)—H(18)	109.329 (12)
C(9)—C(8)—H(8)	119.688 (14)	C(20)—C(18)—H(18)	110.243 (12)
C(8)—C(9)—C(10)	119.813 (13)	C(18)—C(19)—H(19A)	110.150 (13)
C(8)—C(9)—C(14)	119.904 (13)	C(18)—C(19)—H(19B)	108.725 (13)
C(10)—C(9)—C(14)	120.278 (13)	C(18)—C(19)—H(19C)	112.698 (13)
C(9)—C(10)—C(11)	121.216 (14)	H(19A)—C(19)—H(19B)	108.639 (14)
C(9)—C(10)—H(10)	117.152 (15)	H(19A)—C(19)—H(19C)	108.289 (14)
C(11)—C(10)—H(10)	121.624 (14)	H(19B)—C(19)—H(19C)	108.248 (13)
C(10)—C(11)—C(12)	118.911 (13)	C(18)—C(20)—H(20A)	112.231 (13)
C(10)—C(11)—H(11)	121.307 (15)	C(18)—C(20)—H(20B)	107.824 (13)
C(12)—C(11)—H(11)	119.764 (14)	C(18)—C(20)—H(20C)	111.032 (13)
C(11)—C(12)—C(13)	121.450 (14)	H(20A)—C(20)—H(20B)	109.378 (14)
C(11)—C(12)—H(12)	119.316 (13)	H(20A)—C(20)—H(20C)	108.099 (14)
C(13)—C(12)—H(12)	119.227 (14)	H(20B)—C(20)—H(20C)	108.195 (14)