An oxygen-rich, tetrahedral surface phase on high-temperature rutile VO$_2$(110)$_\text{T}$ single crystals

Vanadium dioxide undergoes a metal-insulator transition (MIT) from an insulating (monoclinic) to a metallic (tetragonal) phase close to room temperature, which makes it a promising functional material for many applications, e.g. as chemical sensors. Not much is known about its surface and interface properties, although these are critical in many of its applications. This work presents an atomic-scale investigation of the tetragonal rutile VO$_2$(110)$_\text{T}$ single-crystal surface and reports results obtained with scanning tunneling microscopy (STM), low-energy electron diffraction (LEED) and X-ray photoelectron spectroscopy (XPS), supported by density-functional theory-based (DFT) calculations. The surface reconstructs into an oxygen-rich (2$\times$2) superstructure that coexists with small patches of the underlying, unreconstructed (110)-(1$\times$1) surface. The best structural model for the (2$\times$2) surface termination, conceptually derived from a vanadium pentoxide (001) monolayer, consists of rings of corner-shared tetrahedra. Over a wide range of oxygen chemical potentials this reconstruction is more stable than the unreconstructed (110) surface as well as models proposed in the literature.


Introduction
Vanadium(IV) dioxide, VO2, undergoes a first-order metal-to-insulator transition (MIT) at a critical temperature TC of ~340 K (67°C), where the lattice changes from the monoclinic structure M1 of the semiconducting/insulating phase (T<TC, distorted rutile, space group P21/C) into a tetragonal (rutile) structure of the metallic phase (T>TC, space group P42/mnm) [1]. The MIT shows a hysteresis of several K in heating-and-cooling cycles. During the phase transition, an intermediate strain-or doping-induced monoclinic phase M2 has been observed, together with a metastable triclinic phase T occurring between M1 and M2 [2,3]. The physical mechanism behind the complex phase transition of this strongly correlated oxide is still controversially discussed; an overview is given in Ref. [4].
VO2 is technologically interesting as its MIT occurs near room temperature, and the change in resistivity by several orders of magnitude is accompanied by changes in optical (NIR transmittance), thermal and magnetic properties. The MIT can be tailored to ultrafast switching in the range of ~500 femtoseconds [5,6], and, recently, ~26 fs were reported by Jager et al. [7]. Moreover [4], lattice strain [8], also induced by cation or hydrogen doping [9][10][11], electric current, electric field gating, and irradiation with light [12] modify the MIT and shift its critical temperature even closer to room temperature. The properties and tuneability of the MIT in VO2 are employed in applications, e.g., in memristive devices [13], optical modulators [14], gas sensors [15], field-effect transistors [16] or smart window coatings [17,18].
Since surface properties play an important role in many of these applications, several recent studies have characterized VO2 surfaces, predominantly using supported thin and ultrathin films. It was revealed [19,20] that tensile (compressive) strain along the rutile c-axis imposed by a lattice mismatch with the substrate is correlated with an increase (decrease) in transition temperature. Recent evidence points towards a selvedge that is not just simply a bulk termination: Thin VO2(110) films grown on RuO2(110) and TiO2(110) substrates show an oxygen-rich (2×2) surface termination [21,22]. This is in agreement with earlier studies, which indicated an increased concentration of oxygen atoms at the surface not only under ambient conditions [23], but also under reducing conditions comparable to ultrahigh vacuum (UHV) [24].
Moreover, DFT calculations [22] revealed that rutile surfaces are lower in energy compared to the monoclinic counterparts, and also that oxygen-rich reconstructions reduce the occupation of surface 3d states which is an important driving parameter for the MIT [25]. This is in agreement with recent experimental studies [22,26], which did not reveal any evidence of the bulk structural transition to the monoclinic phase at the surface.
These recent results call for a detailed knowledge of the atomic-scale structure of VO2 surfaces. Such information can best be gained by applying complementary surface science techniques to single crystals and combining these experiments with the DFT calculations. This work focuses on the (110)T (tetragonal) surface of VO2 above TC, where the high electrical conductivity enables such experiments. It starts with a thorough bulk characterization of the structure and purity of the VO2 single crystals, grown from the melt. The experiments were quite challenging; the MIT is connected with a massive structural change. The rigid sample mount required for the surface science measurements can easily lead to a fracture of the samples during the phase transition; to avoid this, the crystals were kept at elevated temperature throughout most of the experiments.
LEED confirms the (2×2) periodicity observed in earlier works [21,22]. X-ray photoemission spectroscopy (XPS) supports that the surface is a vanadium oxide phase enriched with oxygen compared to the bulk and excludes a possible segregation of impurities. Atomically resolved scanning tunneling microscopy (STM) studies show that the (2×2) structure consists of an adlayer, i.e., a modified layer at the very top of the crystal. A detailed analysis of these results constrains the possible structural models.
From the DFT perspective, vanadium dioxide is a challenging material due to strong electron-electron correlations. Benchmark calculations show that standard DFT functionals cannot correctly describe bulk properties like the electron density, relative phase stability, band gap and magnetic ordering for both the rutile and monoclinic phases at the same time [27,28]. Nevertheless, DFT was successfully employed to characterize the surface structure of several VOx-derived surfaces. For example, Schoiswohl et al. [29] characterized the formation of ultrathin VOx structures on a metallic substrate with a combination of atomically resolved STM and ab initio calculations, which provided additional information such as the exact stoichiometry or the atomistic structure. Klein et al. [30] investigated V2O3 and V5O14 structures on Pd(111) and confirmed the structures obtained with DFT experimentally by LEED-I(V).
DFT calculations were also employed in the structural and energetic characterization of V2O5 and V6O13(001) surfaces [31].
In this work, DFT is used to interpret the experimental observations and a novel, tetrahedrally-coordinated surface phase is the best fit to the experimental results.

Materials and Methods
Samples: The VO2 single crystals were grown from melt using V2O5 powder kept in an Ar flow at 1000 °C for 120 h in a quartz (SiO2) crucible. The solid black crystals are shaped as needles or thin plates of approximately 2-10 mm length, 1-4 mm in width and <1 mm in thickness, exposing a flat and reflective top side; see Figure 1(a). Larger crystals often consist of several smaller needles grown together. The crystals are brittle and break easily during mounting. Figure 1(b) shows a crystal tightly mounted on a Ta sample plate with a single Ta spring (1 mm wide) for STM measurements. A thin Au foil is placed between the crystal and sample plate to improve thermal contact to the rough back side of the crystal. The crystal broke into several pieces after a week of measurements and sample transfers inside the UHV chamber. Panel (b) shows one of the larger crystals prepared for XPS measurements. It is placed inside a recess on a molybdenum sample holder; the crystal is gently held by a 0.2 mm Chromel wire spotwelded to the edges of the recess. A Chromel-alumel (type K) thermocouple is connected to the crystal. were particularly high on a crystal that had been previously sputtered in UHV, mounted with a Ta spring on a Ta sample plate with an Au foil underneath the sample (Fig. 1b).
Structural Characterization: X-Ray diffraction (XRD) was done with a SuperNova, Dual, Cu at zero, AtlasS2 diffractometer (Rigaku Oxford Diffraction). The purpose of these measurements was not only to confirm the bulk structures above and below the transition temperature, but also to determine the surface orientation of the shiny crystal side. Since this machine requires small samples, it was necessary to cut the original VO2 single crystals into appropriate pieces. Several crystals were investigated, both asgrown and after the XPS measurements (discussed below), which included crossing the phase transition a few times and heating in UHV to 650 °C. As expected, the treatment in UHV does not influence the bulk structure, although it introduced twin formation. In order to keep track of the prominent shiny side of the crystals, which was used in the other experiments, this side was colored using a white varnish before cutting. This made it possible to identify this facet even in case of irregularly shaped fragments. A selected fragment is shown in Fig. 2 software package [32], and unit cell determination confirmed the presence of the known monoclinic modification at 25 °C, while the tetragonal form was present at 110 °C. The faces of the crystal were indexed using the Crystal Shape Tools implemented in CrysAlisPro.
In addition to the study of the crystal shown in Fig, 2(a,b), Laue back-reflection analysis was applied to some crystals prior to the STM measurements employing a Siemens Kristalloflex instrument in air and at room temperature with wavelengths from 0.25 to 2.50 Å (30 kV, 18-20 mA, 10 min measurement time). The crystals were first aligned perpendicular to the incident beam/X-Ray gun and afterwards tilted by ~4° to shift the central diffraction spots onto the detector. The diffraction pattern was simulated using the software OrientExpress.

Surface measurement procedures:
For the measurements in UHV, the crystals were inserted into the chambers via a small side chamber without bake-out/heating, to minimize the number of phase transitions. It was found that each phase transition shortens the lifetime of the crystals significantly. However, even when minimizing the transitions to two, i.e., cooling to room temperature after the crystal growth and the first heating in UHV (thereafter the crystal is kept above TC), they eventually break along grain boundaries, most likely due to mechanical stress during transfers between different positions in the vacuum chamber and the forces applied by the mounting process. Thus, after the first heating above the MIT temperature, the VO2 crystals were always held at 200-250 °C, i.e., also during sputtering, LEED and XPS measurements, as well as overnight and at 80 °C inside the STM. In both the STM/LEED chambers, the surface of the VO2 crystals was initially cleaned by several cycles of sputtering (1 keV Ar + ions, ~2.5 µA/cm 2 , 10 min) and annealing in UHV (10 min). In the chamber with the variable-temperature STM, the annealing temperature was determined as >600 °C measured on the crystal and ~700 °C measured on the Ta sample plate using a pyrometer with an emissivity set to 0.8. In the LT-STM system, the annealing temperature of ~600 °C was measured via a thermocouple. The surface was daily refreshed either by a full cleaning cycle or by annealing in-between the STM measurements. Variations in the sample preparation, which did not change the surface according to STM examination, included: Annealing at 600 °C in 2×10 −6 mbar O2 combined with cooling in O2 until the temperature decreased to 300 °C; annealing in 1×10 −8 mbar H2 at 450 °C, and dosing 10 Langmuir (L, 1 L = 1.33×10 −6 mbar·s) of water into the STM at 80 °C. In the XPS chamber, the crystals were sputtered for 60 min (1 keV Ar + ions, ~8 µA/cm 2 ) followed by annealing at 350-700 °C for 10 min. The samples were sputtered prior to each annealing step to 'reset' the history of the surface.

DFT Calculations: All calculations were performed with the Vienna ab initio
Simulation Package (VASP) [34]. The projector-augmented wave (PAW) [35] method was employed for treating core electrons. For oxygen 6 valence electrons (2s 2 2p 4 ) and for vanadium 13 valence electrons (3s 2 3p 6 3d 4 4s 1 ) were expanded in a plane-wave basis set with an energy cut-off set to 500 eV. As the present study is focused on the metallic rutile phase, we chose a density functional theory-based description using the meta-GGA SCAN functional [36]. This functional was reported [37] as the best compromise between computational cost and accuracy in terms of lattice parameters and relative phase stability for the rutile and monoclinic VO2 phases. In the present paper, all calculations were performed assuming non-magnetic VO2 systems. While a recent study suggested a better description of the surface energies [22] with spin-polarized calculations, we found that the influence on the current presented results is rather small.
A detailed discussion of these findings will be presented in a forthcoming publication together with a critical assessment of the performance of various DFT functionals in this system [38].
The Brillouin zone was sampled with a Γ-centered Monkhorst-Pack grid [39], using 6×6×9 k-points for the bulk rutile phase. For surface calculations the k-points grid was adjusted to obtain a comparable sampling of the surface Brillouin zone. Ionic relaxations were stopped when all residual forces became smaller than 10 −2 eV/Å. All slabs were calculated with lateral cell dimensions corresponding to the optimized bulk lattice constants with a separating vacuum layer kept at 15 Å.
While the surface energies of the (1×1) terminations were calculated from the linear interpolation of total slab energies ranging from five to eight layers, the surface free energies of the off-stoichiometric (2×2) surface terminations were calculated using five-layer slabs with symmetric top and bottom surfaces and the bulk energy derived from the (1×1) slabs. All simulated STM images were generated in the Tersoff-Hamann approximation [40] with a bias voltage of +2 eV. To obtain more accurate description of the charge density in these calculations, the energy cut-off was increased by 30% with respect to the other calculations.

Diffraction Results
Before discussing the diffraction results it is useful to recall the crystallographic relationship between the monoclinic and tetragonal phases. The lattice parameters of both phases are summarized in Table 1. The transition of a crystallographic plane characterized by (hkl)T to a monoclinic (M) plane is described in ref. [41] by  Table 1: Structural parameters of the VO2 phases below and above the transition temperature TC = 67 °C. [41,42] The orientation of the shiny side of the crystals was determined by XRD (both, the monoclinic and tetragonal phase), Laue back-reflection (RT, monoclinic phase) as well as LEED (200 °C, tetragonal phase). In XRD, see Figure 2

Scanning Tunneling Microscopy: Results and Analysis
The surface structure observed in STM evolved with the annealing temperature after the sputter treatment. While the (1×1) structure is observed after mild annealing at ~560 °C, different superstructures emerge at higher temperatures. Annealing the crystal at ~560 °C (measured with a pyrometer on the crystal; 650 °C on the Ta plate) leads to a rather rough surface with terraces of only 5-10 nm size. On the terraces, the structure of rutile VO2(110) is observed. This is shown in Fig. 4, where the atomic rows of the (110)T surface running in [100] direction are clearly visible, in particular in the right image. The absence of mirror symmetry in the island shape and the high step density suggest screw dislocations due to an STM tip crash nearby. After annealing at >600 °C (as measured with the pyrometer on the VO2 crystal), the surface exhibits terraces that are larger than after annealing at 560 °C, see  direction are usually decorated with three different features, which do not depend on the imaging contrast of the (2×2) structure. They are all imaged as single protrusions sitting in the 'wide' sites, i.e., between the pairs of narrow-spaced rows, seemingly in a 4-fold hollow site with respect to the protrusions of the (2×2) structure. The first species is an additional dot between the double-rows with a similar apparent height as the row maxima. These features often occupy every other '4-fold hollow site' along the rows, leading to flower-like features (white ellipses in Figure 6(a-c)). The second feature is a protrusion in the same site, but with a fuzzy appearance (indicated by white arrows in Fig. 6(c)) including 'fuzziness' of the four neighboring protrusions of the (2×2) structure. Both species are stable during image acquisition and do not diffuse at 80 °C. Finally, there are also a few very bright protrusions close to the sites that can be taken by the weaker maxima discussed here. All these features can be used to identify the wide spacing of the (2×2) double rows even if the rows appear equidistant as in Fig. 6(c). White lines at the top parts of  layer ( Figure 5(d)). Figures 5(a, b) show a very high coverage of this structure, where almost all of the (2×2) is covered. This structure is tentatively assigned to Cs impurities segregating to the surface, which were found in the XPS spectra discussed below. The Cs content differs from crystal to crystal, and it was not possible to change or remove it by excessive sputtering and annealing cycles including sputtering at 600 °C.

X-ray Photoelectron Spectroscopy: Chemical Composition
The surface and near-surface region of larger crystals was systematically investigated at different annealing temperature after cleaning by sputtering. The aim was to see  the VO2 crystals were grown in. Possibly, some crystals grew directly attached to the crucible wall, i.e., the SiO2 is predominantly at the edges or backside of the crystal, which were also visibly discolored on some samples. Si was not detected in the LA-ICP-MS (trace analysis) across several other samples, but these samples (some of them had been previously used for STM) were selected to be visually without fault. Spectra obtained by this procedure essentially display the same oxidation states as in the tetragonal phase but with a drastically different peak shape due to the different screening in the metallic and insulating phases [46]. A plot comparing the tetragonal and monoclinic spectra is provided in the Supplemental Material ( Figure S4).

DFT: The (110)T (1×1) and (011)T (1×1) terminations of rutile VO2
The slab calculations were performed at the optimized bulk-like lattice constants presented in Table 2. In the case of the rutile bulk, both the c/a ratio and the volume of  Table 2: Calculated parameters of the rutile unit cell and surface energies σ of VO2, using the meta-GGA (SCAN) functional.
The stability of the unreconstructed low-index facets of the rutile phase, (110)T and (011)T was evaluated by calculating their respective surface energies (Table 2) finding the (110) surface to be more stable by 36 meV/Å 2 .
The bulk-terminated, relaxed surfaces of rutile VO2(110)T and VO2(011)T are shown in Figures 9(a,b), respectively. In calculated STM images of the VO2(110)T terminated surface, Figure 9(c), the twofold-coordinated oxygen atoms appear as straight, bright chains with a distance of 6.45 Å. Note that this appearance is different from that of the well-known TiO2(110) surface, where the bridging oxygen rows appear dark, [47] but the same as that of RuO2(110), which is also metallic [48]. Both, the appearance of rows and their separation is in agreement with experimental STM images of VO2(110)T (1×1) in Figures 4 and 5(a). The bright features on the VO2(011)T termination, Figure   9(d), are formed by both vanadium and oxygen atoms, resulting in zigzag chains.

Developing a Surface Model
Evidently, the experimentally observed (2×2) phase cannot be obtained by a simple variation of the bulk structure due to the lower symmetry. To investigate possible reconstructions of the bare VO2(110)T surface, a simulated annealing technique was employed. A subsequent relaxation (with the SCAN functional) shows that even for the bare VO2(110)T surface, the calculated ground state has its symmetry lowered due to a buckling in the topmost layer, resulting in a (2×1) superstructure (Fig. 10  The result is a hexagonal ring of vanadium tetrahedra as shown in Figure 11(b). A similar structure consisting of corner-sharing up-and down-pointing VO4 tetrahedra has already been confirmed for vanadium oxide on a Pd(111) surface [30]. For the unsupported model layer (Figure 11(b)), the lateral distance between the oxygen atoms at the top of the tetrahedra along [1−10] is 3.8 Å. When this layer is supported by the rutile (110) surface as shown in Figure 11(c) and (d), the distance is slightly decreased to 3.7 and 3.6 Å, respectively. In both cases, the surface has an overall stoichiometry of V4O13.  the ring termination is bound in a corner-sharing fashion, i.e., with just with one oxygen bond, forming a purely tetrahedral termination, and leaving half of the undercoordinated O atoms of the substrate unterminated. These atoms, colored in orange, can be also partly or fully removed, which leads to V4O12 and V4O11 surface stoichiometries. In Figure 11 To explain the additional features in the STM images, other metastable ringtype structures were also explored. In these structures, an additional vanadium tetrahedron was added between the rows, which changes the stoichiometry of the surface layer (V5O14 and V5O15); these structures are less stable than those in Fig. 11(ch). Details are discussed in the Supplemental Material.

Stability of Surface Phases
To evaluate the stability of various surface terminations, we plotted the surface free energy as a function of the oxygen chemical potential as it is described in Ref. [49] see Figure 12. The black, horizontal line represents the stoichiometric buckled VO2(110)T surface. Green lines denote (2×2) supercells of this buckled surface with 1, 2 or 4 additional O atoms adsorbed in a vanadyl configuration on top of the fivefold coordinated V. For the latter two cases, our preferred structures agree with the models of an earlier DFT study by Mellan et al. [24]. Decreasing the coverage from 1/2 to 1/4 ML (1 adsorbed oxygen atom) every second oxygen atom is removed from the remaining oxygen row. Blue, orange, pink and red lines mark the oxygen-rich ringsuperstructures, including V4O13, V5O14 (zig-zag and ring), and V5O15 stoichiometries as depicted in Figures 11(d,f), S5(a,b) respectively. Furthermore, gray dashed lines represent the reduced V4O13 ring structure depicted in Figure 11(c) where the undercoordinated oxygen atoms in the subsurface layer (colored in orange) were subsequently removed from both rows, leading to the V4O12 and V4O11 surface stoichiometries. Structures and simulated STM images of additional ring terminations are presented in the Supplemental Material. The plot also shows the stability limit of the VO2 phase with respect to the vanadium pentoxide as a vertical black solid line defined as the enthalpy of the following reaction: 2VO2 + ½ O2 ® V2O5. For calculating this phase boundary, the experimental heats of formation of the VO2 and V2O5 phases with respect to vanadium metal [50,51] were used. It should be noted that the calculated phase boundary strongly depends on the chosen functional and spin-treatment, due to the peculiarities of an appropriate treatment of the VO2 phase. These dependencies will be discussed in detail in a forthcoming publication [38].
Over a wide range of chemical potentials, the ring structure with V4O13 ( Figure   11(d)) and the zig-zag with V5O14 ( Figure 11(f,g)) surface stoichiometry are the most stable configurations. An unreconstructed, buckled VO2(110)T surface, partially covered with O atoms would be stable under strongly reducing conditions (oxygen chemical potential less than −2.05 eV).

Discussion
This work clearly shows that the lowest-energy state of the VO2 surface in a wide range of chemical potentials is a reconstruction, distinctly different from a bulk-terminated surface. The unreconstructed rutile (110)T termination is found only after mild annealing of a sputter-treated surface. After equilibration at higher temperatures, an adlayer with a double-row superstructure is observed. While the (2×2) periodicity is consistent with previous reports [21,22]; the simple models proposed earlier [24] that invoke only adsorption of excess O are neither supported by the STM measurements nor by DFT calculations.
The DFT models explain the main features of the STM images. The calculations showed that the aligned, bright spots (separated by 3.6 Å) in Figure 6(b) are related to the ring structure, which is the most stable surface termination at an oxygen chemical potential of −1.54 eV and higher. The model assigns the experimental double rows in Figs. 6 (a, b) to O atoms at the apex of VO4 tetrahedra. In the range of chemical potentials between −1.54 and −2.05 eV, another ring structure is more stable that exhibits the zig-zag pattern similar to the features observed in Figure 6(d).
The stability of the ring terminations, especially at chemical potentials corresponding to higher oxygen pressures, is related to both the fact that the ring structure contains more oxygen than the adsorption phases and also to the close relationship of the ring structure to a vanadium pentoxide monolayer whose surface energy is only 11 meV/Å 2 according to our calculations. This relationship is not only structural -as we pointed out, the ring structures were derived from a V2O5 monolayer -but also evident in the electronic structure. As shown in Figure 13, the projected density of states (pDOS) onto vanadium and oxygen atomic orbitals of the ring phase compares well with the V2O5 bulk pDOS. The graphs show that, unlike the VO2 phase, V2O5 as well as the V4O13 ring display a band gap where the V 3d band is separated by 1.9 and 2.2 eV from the O 2p band, respectively. It should be noted that the calculated band gap for the V2O5 phase underestimates the experimental band gap of 2.2-2.4 eV [52][53][54] and, therefore, we also expect a similar underestimation for the surface phase. Second, the double rows in the experiment are always aligned with respect to the neighboring row like as in Figs 6(a,b), but our present model also allows the hexagonal rings that form the double-row pattern to be shifted by half of the unit cell in [001] direction, which is not observed in the experiment. The effect that would restrict the observed structure just to the aligned pattern is not evident from the DFT model; probably the alignment is caused by the entities forming the additional bright spots between the double rows. Furthermore, the experimental zig-zag row always neighbors a rectangular row (see Fig. 6d), which is not captured by our model because of the limited periodicity of the considered (2×2) supercell.
Nevertheless, the ring structures with the V4O13 and V5O14 surface stoichiometries can be seen as a structural basis that explains the most prominent features of the experimentally observed surface reconstruction. However, it should be pointed out that a huge amount of potentially stable surface structures exists, with only subtle energy differences between them. A further exploration remains a challenging task for the near future, see e.g. Ref. [38].

Summary
In summary, this work reports a comprehensive study on the surface properties of VO2 single crystals, employing imaging, diffraction and spectroscopy techniques complemented by DFT calculations. The crystals exhibit the expected bulk structures below and above the MIT, and the most stable surface, assigned as (0−1−1) and (110) in the monoclinic and tetragonal structure, respectively, was investigated in detail.
Impurities show strong variations even within individual samples, but are not affecting the described results (except for a minority superstructure with c(4×2) symmetry that is attributed to Cs atoms). XPS in grazing (surface sensitive) and normal (more bulk sensitive) emission shows that the surface is oxygen rich. The distinctly different XPS peak shapes reflect the drastic change in electronic structure that accompanies the MIT.
The most prominent outcome of the present work is a description of a (2×2) surface phase that is structurally distinct from the VO2 bulk. The proposed ring models are based on corner-sharing tetrahedra and pyramids structurally and electronically similar to V2O5 layers, implying that surface atoms are in a markedly different environment than the bulk atoms, likely with profound influence on the surface properties, such as oxygen adsorption or the temperature of the metal-insulator transition of nanoparticles of this interesting material.

Supplemental Material
An oxygen-rich, tetrahedral surface phase on high-temperature rutile VO2(110)T single crystals

Bulk characterization: Trace Analysis by LA-ICP-MS
The quantification was done using NIST612 with 51 V as an internal standard. Table S1 contains the averaged value measured across five samples. Some elements are very inhomogeneously distributed in specific samples; see also the standard deviation given in        panels (a, b)).

DFT: Modified ring terminations
Since the ring termination with the surface stoichiometry V4O13 described in the main text is only one out of many possibilities for such a structure, we also studied its modifications. The most promising candidates are described in this section. The original ring structure with the surface stoichiometry V4O13 can be adapted by adding another vanadium tetrahedron between the rows as is shown in Figure S5(a-b), forming V5O14 and V5O15 rings. The structure in panel (a) with the surface stoichiometry V5O14 connects the rings with another vanadium tetrahedron that is bound to the subsurface layer. The vanadium rings are connected in this case by a single V-O bond to the subsurface layer, as in the ring termination in the Figure 11(c). The V5O14 configuration is, therefore, structurally similar to the SrTiO3(110) termination that is shown as the (3×1) surface structure in ref.
[S1], but with two major differences. The first difference is related to the bonding of the superstructure to the bulk termination. In the case of VO2, vanadium tetrahedra are bound only with a single V-O bond to the VO2 bulk, i.e., this superstructure is more open compared to the titanium tetrahedra that terminate the SrTiO3(110) surface. Secondly, some vanadium tetrahedra are additionally oxidized with another oxygen atom, which results in a disconnection from the (110) surface and the formation of vanadyl bonds on top, as is seen by comparing the side views of the structures in the Figure S5.
Simulated STM images also show patterns that are close to the experimental findings. The shorter distance between the topmost oxygen atoms of the ring structure on the left side is reduced in size from 3.6 Å to 3.3 Å which causes a small overlap between the bright spots. The additional oxygen atom in the structure in panel (b) forms a weak protrusion that sits between the double rows. Figure S5: Top views, calculated STM images and perspective views of the relaxed modifications of two ring terminations. In panel (a) the V4O13 rings are connected via an additional tetrahedron that is bound to the substrate, leading to an overall V5O14 stoichiometry. In panel (b) one connecting tetrahedron per (2×2) cell is flipped, resulting in adsorption of an additional oxygen.