Robert F.
Moran
a,
David
McKay
a,
Chris J.
Pickard
b,
Andrew J.
Berry
c,
John M.
Griffin
d and
Sharon E.
Ashbrook
*a
aSchool of Chemistry, EaStCHEM and Centre of Magnetic Resonance, University of St Andrews, St Andrews KY16 9ST, UK. E-mail: sema@st-andrews.ac.uk
bDepartment of Materials Science & Metallurgy, University of Cambridge, 27 Charles Babbage Road, Cambridge CB3 0FS, UK
cResearch School of Earth Sciences, Australian National University, Canberra, ACT 2601, Australia
dDepartment of Chemistry, Lancaster University, Lancaster LA1 4YB, UK
First published on 21st March 2016
The structural chemistry of materials containing low levels of nonstoichiometric hydrogen is difficult to determine, and producing structural models is challenging where hydrogen has no fixed crystallographic site. Here we demonstrate a computational approach employing ab initio random structure searching (AIRSS) to generate a series of candidate structures for hydrous wadsleyite (β-Mg2SiO4 with 1.6 wt% H2O), a high-pressure mineral proposed as a repository for water in the Earth's transition zone. Aligning with previous experimental work, we solely consider models with Mg3 (over Mg1, Mg2 or Si) vacancies. We adapt the AIRSS method by starting with anhydrous wadsleyite, removing a single Mg2+ and randomly placing two H+ in a unit cell model, generating 819 candidate structures. 103 geometries were then subjected to more accurate optimisation under periodic DFT. Using this approach, we find the most favourable hydration mechanism involves protonation of two O1 sites around the Mg3 vacancy. The formation of silanol groups on O3 or O4 sites (with loss of stable O1–H hydroxyls) coincides with an increase in total enthalpy. Importantly, the approach we employ allows observables such as NMR parameters to be computed for each structure. We consider hydrous wadsleyite (∼1.6 wt%) to be dominated by protonated O1 sites, with O3/O4–H silanol groups present as defects, a model that maps well onto experimental studies at higher levels of hydration (J. M. Griffin et al., Chem. Sci., 2013, 4, 1523). The AIRSS approach adopted herein provides the crucial link between atomic-scale structure and experimental studies.
In recent years, there has been growing interest in the use of theoretical calculations to provide a prediction of the NMR parameters and to validate structural models in conjunction with experimental measurements.3–5 In particular, the GIPAW6 approach, implemented in planewave density functional theory (DFT) codes such as CASTEP,7–9 has been applied to a range of materials, including organic pharmaceuticals, microporous solids, energy materials and supramolecular assemblies, with excellent agreement between experiment and calculation for many different nuclear species.3–5 The majority of studies have considered well-ordered crystalline solids, with initial structural models provided by prior diffraction-based experiments or previous computational work. Optimisation of the geometry can then be used to ensure that the lowest energy configuration of atoms is used. Generating an initial model is, however, more complex when considering disordered solids. The average structures that result from diffraction – often with fractional or partial site occupancies – cannot be used directly, and multiple structural models can have similar energies. Possible solutions to this problem include the simple substitution of one or more atoms into a known but ordered structure, thereby providing information on the changes in NMR parameters that are expected upon such substitution, or the generation of a set of similar structures with atoms placed on different sites, and the predicted NMR parameters from each then summed. Although these approaches have seen much success,3–5,10–16 they quickly become unfeasible as the amount and types of disorder increase resulting in an increasingly large number of possible structures. Furthermore, for some disordered systems atoms may not lie at well-defined sites, leading to a potentially infinite range of possible structures. Choices have to then be made about how possible models are generated, whether all possibilities are considered equally or whether only those that appear most “chemically reasonable” or have lower energy are selected. Any constraint placed on the structures considered may lead to bias in the results, or in the omission of potentially important motifs.
Recent advances in computing power, and improvements in the accuracy with which energies can be calculated at reasonable costs, have resulted in structure prediction, i.e., searching for stable structures of materials, becoming a rapidly-expanding field.17 At a fundamental level this provides a method for determining the most stable structure, or global minimum, but these approaches can also provide information on local energy minima and potential structural variants (of considerable interest for disordered solids). Computational searching can be easier and cheaper than experiments, and can also augment incomplete or ambiguous experimental data. Searches can also be performed under conditions that are difficult to reproduce experimentally. The ab initio random structure searching (AIRSS)18–21 approach utilizes first-principles DFT in a simple strategy for generating essentially random initial configurations. The simplest searches involve a randomly selected unit cell with the positions of the atoms then chosen randomly. First-principles methods are then used to relax each structure to an enthalpy minimum. The process is then repeated many hundreds or thousands of times. As the complexity of the structure increases constraints will need to be employed, perhaps using known initial cell sizes, restricting species to have a particular bonding environment within the final structure, or adding prearranged molecules or groups of atoms.19
Here, we utilize AIRSS to help investigate a particularly challenging experimental problem – the structural characterisation of hydrous wadsleyite. Fe free wadsleyite (β-Mg2SiO4, Fig. 1a) is the Mg end-member of wadsleyite (β-(Mg,Fe)2SiO4), a high-pressure mineral that is believed to be the principal component of the Earth's transition zone, between depths of ∼410 and 530 km.22–25 Over the last few decades, it has been recognised that the nominally anhydrous minerals (NAMs) that make up the Earth's mantle can hold significant quantities of hydrogen as hydroxyl units, typically referred to as “water”, as defects in the structure, resulting in significant effects on the physical and chemical properties of the Earth.26,27
Wadsleyite has received intense interest as a potential water reservoir, as it can incorporate up to 3.3 wt% H2O, a much greater amount than the other mantle minerals.22–25 The substitution of hydrogen into wadsleyite is charge balanced by the removal of a Mg2+ cation, but there is no fixed crystallographic site upon which the hydrogen is expected to be located. A variety of work has appeared in the literature attempting to address this problem, using both experimental approaches including XRD, neutron diffraction, FTIR and solid-state NMR, and also computation.28–42 The high pressures required for synthesis (14–15 GPa) present technical challenges and also limit the amounts of sample that can be produced. While there is some broad agreement on probable structural models, for example, that vacancies are located on Mg3 sites,28,36,37,40 there remain a number of different defect types proposed, some debate over which oxygen species are protonated and disagreement over the exact location of H atoms. It should be noted that in some previous computational studies potential docking sites were considered without explicit inclusion of cation vacancies.29,30 Although our recent NMR study40 did attempt to relate the experimental NMR parameters to those predicted by DFT, only a small number of initial structural models were used and their selection was (by necessity) biased by previous literature suggestions. In this work, we use AIRSS to provide a range of possible structural models (and information on their relative energies), independently of any prior experimental measurements. We have adapted the original AIRSS approach by starting from an initial anhydrous wadsleyite structure, removing a single Mg2+ cation and randomly placing two H+ within 3.0 Å of the vacancy (see Fig. 1b). In addition to providing possible structural models, as AIRSS also uses the CASTEP code, the NMR parameters for the proposed models can be directly calculated and compared to experiment, aiding spectral assignment and interpretation for this challenging material.
Initial optimised structures (819 examples) were ranked according to relative enthalpy (ΔH) with the lowest set to ΔH = 0.0 eV. This was then used to select structures for further study. These included the 25 most stable structures, then one structure for each of a series of increments (δΔH): δΔH = 0.002 eV up to ΔH = 0.5 eV; δΔH = 0.005 eV for 0.5 < ΔH < 0.6 eV; δΔH = 0.05 eV for 0.6 < ΔH < 1.0 and δΔH = 0.1 eV for each structure with a relative enthalpy above 1.0 eV. This selection process ensured a large number of relatively stable structures were sampled. In addition, a number of less stable structures are included to allow for comparisons across the full enthalpy range. These structures were then optimised using better converged values of Ecut (60 Ry) and k-point grid spacing (0.04 × 2π Å−1, resulting in 30 k-points). All internal atomic coordinates and the lattice parameters were allowed to vary. This produced a set of 103 “fully-optimised” structures. NMR parameters were then computed for these structures using the GIPAW6 algorithm within CASTEP at a level of theory consistent with the full optimisations. Post processing of computed NMR data was conducted using Python scripts extending the CCP-NC MagresPython module.46 Calculations were performed using a cluster at the University of St Andrews, consisting of 201 AMD Opteron processing nodes, each containing twelve cores, partly connected by Infinipath high-speed interconnects. Calculation wallclock times ranged from 8 to 48 h using 72 cores.
Calculations of the NMR parameters produce the absolute shielding tensor (σ) and the quadrupolar tensor (V). Diagonalisation yields the three principal components, σ11, σ22 and σ33 and VXX, VYY and VZZ. The isotropic chemical shift δiso, is given by (σref − σiso), where σiso is (⅓) Tr(σ) and σref is a reference shielding. Reference shielding values of 31.0, 320.1, 560.1 and 249.7 ppm were used for 1H, 29Si, 25Mg and 17O, respectively, determined by comparison of experimental and computational results for Mg(OH)2 (1H), MgO (25Mg) and anhydrous wadsleyite (29Si and 17O). From the principal components of σ (σ11 ≤ σ22 ≤ σ33), the magnitude (or span), Ω = σ33 − σ11, and the skew, κ = 3(σiso − σ22)/Ω, can also be determined. The quadrupolar coupling constant, CQ = eQVZZ/h and the asymmetry parameter, ηQ = (VXX − VYY)/VZZ, are obtained directly from the principal components of V. Q is the nuclear quadrupole moment (for which values of 199.4 and −25.6 mb were used for 25Mg and 17O, respectively).
Fig. 2 (a) Plot showing the relative enthalpy of all 819 (blue) and 103 systematically selected (red) AIRSS-generated structures of hydrous wadsleyite with an initial Mg3 vacancy. The relative enthalpy of the 103 structures, post DFT geometry optimisation is also shown (green). (b) Plot showing the 103 selected structures re-normalised according to the enthalpy after optimisation. The dotted line represents the relative enthalpy of the structural model derived from that proposed by Smyth.24 |
Upon inspection, the structures with the lowest enthalpy (representative structure, A (ΔH = 0.0 eV) shown in Fig. 3) correspond to those where only the O1 site is protonated. These structures contain hydroxyl groups staggered with respect to each other (dihedral angle, d(HOOH) = 104° in A), lying along O1–H⋯O4 vectors, hydrogen bonded to O4, in good agreement with previous studies.32–34 The next plateau of points shown in Fig. 2b at ΔH ≈ 0.5 eV, corresponds to structures containing adjacent protonated O1 and O3 sites, e.g., B, in Fig. 3. The protonation of a pyrosilicate oxygen site (O3) rather than an O1 site appears to be associated with a sizeable decrease in stability, perhaps unsurprisingly since, as a result, this structural motif contains a formally “underbonded” O1. Hydroxyl bond vectors are again orientated along the edges of the vacant octahedron, here O1–H⋯O4 and O3–H⋯O3 interactions are found with d(HOOH) = 78.4°. The protonation of an O4, rather than an O3 site, giving motif C, is associated with a further enthalpy increase of ∼0.1 eV. Here, O1–H⋯O4 and O4–H⋯O1 hydrogen bonding interactions are seen (HOOH dihedral angle of 40° in C), possibly suggesting a relationship between A and C through hopping of a proton across the O1–H⋯O4 hydrogen bond. The trend in relative enthalpy with respect to protonation site (O1–H < O3–H < O4–H) is associated with the nature of the newly-formed hydroxyl, with an increase in bond length, from 0.99 to 1.03 Å, and a decrease in Mulliken bond population, from 0.66 to 0.60, across the series (see Table 1). The latter indicates a decrease in bond covalency. While the range in relative enthalpy seen here is not insignificant (ca. 0.6 eV), this is thought to be due to the explicitly periodic nature of the model used, whereby protons in non-optimal environments (i.e., silanol groups) and optimal positions (i.e., O1-bound) are necessarily present in a 1:1 ratio, resulting in a large destabilisation.
Structure | O–H distance/Å (O site) | O–H bond population |
---|---|---|
A | 0.99 (O1) | 0.66 |
0.99 (O1) | 0.66 | |
B | 0.99 (O1) | 0.66 |
1.01 (O3) | 0.62 | |
C | 0.99 (O1) | 0.66 |
1.03 (O4) | 0.60 |
Test calculations were conducted considering larger periodic repeat units (derived from 2 × 2 × 1 supercells with 32 Mg2SiO4 formula units, therefore allowing 4 × Mg3 vacancies and 8 × H atoms; see ESI†), through which a reduced silanol to O1–H ratio of 1:7 was modelled. These suggest silanol groups seen experimentally may be present as “defects” within a structure dominated by O1–H hydroxyls, where silanol formation, though coming at a high enthalpic penalty, becomes feasible through the increase in entropy as the silanol to O1–H ratio decreases.
A periodic system entirely devoid of O1–H hydroxyls is considered to be unlikely. Such models, e.g., D in Fig. 3, exhibit enthalpies of at least 0.8 eV above O1–H hydroxyl-dominated congeners.
A previous study by Smyth suggested a structure for fully hydrous (3.3 wt%) wadsleyite, where 2 × Mg2 cations were substituted for 4 × H, with all protonation at the O1 sites.24 However, due to the ionic point charge model and symmetry constraints adopted in that study, O1–H hydroxyls were found to align with the z-axis (as defined in the present work), rather than forming hydrogen-bonding interactions.24 A model based on the Smyth structure, modified to represent the 1.6 wt% hydration level considered here, was generated and fully optimised (see ESI† for details). Its relative enthalpy (ΔH = 0.9 eV) is illustrated in Fig. 2b by the horizontal dashed black line. 81 of the 103 fully-optimised structures are lower in enthalpy than the modified Smyth model. Additionally, a FTIR study by Jacobsen et al.,36 based on wadsleyite with a hydration level of ∼1 wt%, only found evidence of Mg3 vacancies. Further studies corroborate the present findings in terms of the sites at which protonation is likely to occur. X-ray diffraction studies by Kudoh et al. suggested protonation of the O1 site, however with the hydroxyl orientation disordered over O1⋯O1, O1⋯O3 and O1⋯O4 vectors.31 Recent work by Deon et al.,37 predominantly based on FTIR spectroscopy, found that the O1 and O3 sites in the vacant Mg3-centered octahedron were the favoured sites of protonation. In a 1H MAS NMR and FTIR study of wadsleyite at low levels of hydration (<1.5 wt%), Kohn et al.35 found O1 protonation to be the dominant environment, while the lowest degree of protonation was associated with the O3 and O4 sites.
Fig. 5 Plot of calculated 17O CQ correlated against 17O δiso for all 103 fully-optimised AIRSS-generated structures of hydrous wadsleyite. |
The computed chemical shift values for these three clusters of points do not overlay exactly with the peaks observed in the experimental 1H MAS NMR spectrum of hydrous wadsleyite reported previously.40 However, the sample studied experimentally had ∼3% H2O, and it has been suggested that the preferred protonation sites vary with hydration level,35 and small errors owing to the DFT exchange–correlation functional and referencing cannot be ruled out. It should also be noted that although the periodic approach employed here may result in small changes in structure (and therefore shift) when compared with an isolated defect or more disordered structures, attempting to include this possibility in the AIRSS approach would lead to an unfeasibly large number of possible structures that could not realistically be considered on a reasonable timescale (but would provide a challenge for further future study). However, the relative calculated isotropic shifts are in reasonably good agreement with experimental measurements, suggesting that both O1/O1 and O1/O3 substitution occur, and the calculations confirm that the previous assignments of Mg–OH and Si–OH species are correct.40
Fig. 4b, which shows the variation of the 1H δiso with H⋯O distance, also shows three distinct clusters of points when only structures with low ΔH are included, although noticeably more scatter is present. Structures with particularly short hydrogen bonds exhibit high 1H δiso, showing that as the H–O and OH distances become similar, the proton is shifted downfield, a phenomenon that is well documented in the literature.47–49CQ demonstrates an almost linear dependence on the O–H bond length (Fig. 4c).
Despite the good overall agreement, five points lie noticeably below the trend line. These points correspond to structures containing protonated O1 sites where the hydroxyl bond vector is orientated away from the Mg3 vacancy (see ESI,† Fig. S2.2), in contrast to all other structures where the O1 hydroxyl vectors are co-linear with the O1⋯O4 octahedron edge (see Fig. 3). However, due to their high relative enthalpies (ΔH = 1.1 to 1.8 eV) such O1–H orientations are unlikely. The outlying point with the longest hydroxyl bond length and smallest CQ, 1.09 Å and 0.07 MHz, respectively, corresponds to the least stable structure found (ΔH = 3.7 eV). In this structure, both hydrogen atoms are located outside the Mg3 vacant octahedron. The small computed CQ may be due to one proton experiencing a relatively symmetrical, though unstable, local environment. As seen in the δiso data, the correlation between CQ and O–H bond length is noticeably stronger than that observed when the H⋯O hydrogen bond length is considered (Fig. 4d). Generally, CQ appears to increase with decreasing O–H bond length, corresponding to the loss of symmetry as the O–H and H⋯O bonds become more significantly different.
Despite the calculated 1H and 2H NMR parameters providing insight into the types of hydroxyls present in a given structure (i.e., Mg–OH or Si–OH), given the chemical similarity between the O3 and O4 non-bridging oxygen atoms in particular, this information is not sufficient to unequivocally determine the exact site of protonation, even when 1H isotropic chemical shift is plotted directly against 2H CQ (see ESI,† Fig. S5.1). In an attempt to clarify this ambiguity, 17O NMR parameters were also considered, as these are known to exhibit considerable change upon protonation.40,50 A plot of 17O CQ against δiso, with points coloured according to the crystallographic O site notation, and those denoting protonated O–H species distinguished by filled circles shown in Fig. 5. Regardless of oxygen site, hydroxyl oxygen centres exhibit an increased CQ (by ca. 2–5 MHz) relative to their non-protonated counterparts. The most significant changes are observed for O1, which also sees a large upfield shift in δiso (17O1: δiso ≈ 30–60 ppm, CQ ≈ 1–2 MHz; 17O1–H: δiso ≈ 0–50 ppm, CQ ≈ 5–8 MHz). Protonated O1–H sites further divide into two groups with δiso ≈ 0–20 ppm and 40–50 ppm, the latter group consisting of the five structures destabilised due to unusual orientations; these correspond to the same structures found to be outlying previously in terms of O–H and H⋯O distances (Fig. 4), suggesting that the 17O δiso is more sensitive to the nature of the hydroxyl group than the 1H δiso. Fig. 5 illustrates that a combination of 17O CQ and δiso is again sufficient to differentiate between Mg–OH and Si–OH groups, and provides support for the tentative assignment, to silanol groups, of the 17O signals observed at higher δiso experimentally.40 Unfortunately, ranges in silanol 17O NMR parameters show significant overlap. However, non-protonated silicate oxygen centres exhibit more distinct NMR parameters; in particular, O2 exhibits a relatively large CQ that is typically associated with bridging oxygen species.51
Since the 1H and 17O NMR parameters were found to be equally sensitive indicators of the protonation site and hydroxyl orientation, variations of 1H δiso against 17O δiso are shown in Fig. 6a, focussing on structures with H–O distances ≤1.10 Å. Points are denoted by protonation site and coloured according to their relative energy. Ranges in 1H and 17O δiso found in this ab initio study are larger than those observed experimentally.40 However, when relative enthalpy is taken into account (by only considering structures with ΔH < 1.5 eV), these ranges are narrowed significantly. This suggests that experimental spectra are unlikely to display significantly intense features at δiso > ∼10 and ∼60 ppm in 1H and 17O NMR spectra respectively. From Fig. 6a, extensive overlap in the ranges of δiso observed for both O3 and O4 sites is apparent. This suggests that neither 1H nor 17O chemical shifts (nor their combination) are suitably discriminatory for a clear distinction between protonated O3 and O4 sites.
Fig. 6b shows a plot of 1H vs.29Si δiso. The values are coloured according to Si⋯H interatomic distance to distinguish Si–OH (blue) and Si⋯HO (red) interactions, where the latter includes Si centres neighbouring either silanol or O1–H species. While this plot exhibits extensive scatter, the 64 most stable structures, shown in darker colours, are found to form distinct clusters. The three clusters of points with 1H δiso ≈ 5 ppm and 29Si δiso ≈ −77.5, −78.5 and −79.5 ppm correspond to structural motifs of type A (Fig. 3), where protonated O1–H hydroxyls interact with three inequivalent 29Si nuclei with Si⋯H distances between 2.4 and 4.0 Å. There are numerous clusters of points corresponding to structures containing protonated O1 and O3 sites (motif B, Fig. 3), with the points at 1H δiso ≈ 4 ppm corresponding to the O1 hydroxyls of these structures. The feature at ∼4 ppm is made up of two clusters of points, with the same δiso but different HSi interatomic distances, 2.9 and 3.5 Å, respectively. The blue points at 1H δiso ≈ 9 ppm and 29Si δiso ≈ 75 ppm correspond to protonated O3 sites, indicating that as the H⋯Si interatomic distance decreases, both hydrogen and silicon exhibit a downfield shift, in good agreement with previous work by Griffin et al.40 The remaining clusters at 1H δiso ≈ 9 ppm and 29Si δiso ≈ 77.0, 78.3 and 78.7 ppm represent interactions between silanol protons and 29Si nuclei belonging to other pyrosilicate units. This correlation was not observed experimentally,40 though this could be due to differences in relaxation parameters for different hydroxyl environments (i.e., Mg–OH or Si–OH). Indeed, there is evidence in the literature to suggest that silanol environments in minerals are often underrepresented due to fast T1ρ relaxation.51,52
Despite consideration of a restricted set of models (i.e., limited to Mg3 vacancies and 1.6 wt% hydration), the predicted NMR parameters agree well with previous multinuclear NMR experimental results on a sample with ∼3.0 wt% H2O, and are able to confirm the interpretation and assignment of a number of spectral features.40 The high 1H chemical shifts observed are shown to result from strong hydrogen bonding to nearby O species, and a linear correlation of 1H δiso and 2H CQ with hydroxyl O–H bond length and H⋯O hydrogen bond length is clear. The structure proposed by Smyth (modified to 1.6 wt% hydration) has a high relative enthalpy; ca. 1.0 eV above the most stable structure generated by AIRSS in this work. This appears to be due to the high-symmetry point charge model adopted by Smyth, which precludes the formation of stabilising hydrogen-bonding interactions.
Further work will seek to test these conclusions in progressively more complex models. This will include the investigation of higher levels of hydration (3.3 wt%; two Mg vacancies and four H per unit cell) and the assessment of alternative vacancies of Mg1, Mg2, Si sites and their combinations (in ordered unit cells). In addition, the systematic use of supercells will assess the effect of positional disorder of identified motifs. As the complexity of models used increases, the opportunity for comparison with the true disordered system arises, and at this point, the level of theory may be tested by the inclusion of, e.g., dispersion effects via appropriate correction schemes. Note, such steps lead to a highly combinatorial problem and the simultaneous assessment of each phenomenon is unlikely to be possible. As such, quantifying the degeneracy of any given arrangement of defects is an aim that may only be fulfilled by experiment.
The AIRSS philosophy of randomly generating candidates for unknown and disordered structures, combined with the prediction of solid-state NMR parameters (providing a direct link to experiment) heralds a step change in the characterisation of complex materials. We envisage opportunities for the application of this combinatorial approach to areas as diverse as energy materials, gas storage and separation, microporous frameworks, (supra)molecular self-assembly and catalysis.
Footnote |
† Electronic supplementary information (ESI) available: Further information on the structures generated by AIRSS, alternative structural models, supercell calculations, total enthalpies of all computed structures and further information on 1H/2H NMR parameters. Example input and all raw output files from AIRSS and CASTEP NMR calculations are also included. See DOI: 10.1039/c6cp01529h |
This journal is © the Owner Societies 2016 |