## Introduction

The importance of oxygen point defects in dictating physical properties has been more and more widely recognized by the functional oxide community1,2,3,4. Oxygen defects, disguised by the name, can be actually used to advantageously enhance functionalities and device performances, ranging from electronic5,6,7, magnetic8,9,10 and multiferroic properties11 to energy storage and conversion applications2,12,13. Especially, there has been an increasing interest in utilizing oxygen point defects for tailoring electronic structures of oxides, due to the designability and reversibility of this approach14. Electrical switching enabled by oxygen defects has been proven promising for applications in neuromorphic computing due to the fact that oxygen defects allow actively tuning electrical conductivity on demand14,15. In the pursuit of a defect-tuning functionality in oxides, it was found that the effects of oxygen defects can go beyond what can be explained within the simple rigid band model and band-filling picture. In correlated oxide systems, changes in oxygen content can profoundly alter electronic structure and even trigger a metal–insulator transition (MIT). VO2-δ (δ denotes oxygen non-stoichiometry) is such a material system, wherein a small change in composition can lead to a large modulation in correlation effects16,17,18,19,20. Near-stoichiometric VO2 has 3d1 electron configuration and shows a MIT with a transition temperature Tc ≈ 340 K, accompanied by a structural transition from low-temperature M1 (monoclinic) phase to high-temperature R (rutile) phase. The MIT transition temperature can be lowered by 60 K using epitaxial strain imposed by substrates21,22. Interestingly, the MIT has been shown to be completely suppressed by inducing oxygen vacancy formation, either chemically17,23 or electrochemically16,24. Oxygen-deficient VO2-δ was shown to remain in the metallic R phase when cooled down to below the MIT transition temperature. This makes oxygen stoichiometry a knob for reversibly controlling the MIT in VO2.

Oxygen non-stoichiometries in functional oxide thin films are most commonly manipulated by changing the oxygen electrochemical potential via annealing in a certain oxygen partial pressure (pO2) or by applying electrical bias25. In contrast, a novel strategy based on interfacial oxygen defect migration induced by the mismatch of defect formation energy of two dissimilar oxides, is much less studied. This “oxygen diode” effect essentially utilizes the oxygen chemical potential gradient from the difference in defect formation energy (Ef) across oxide interfaces. In order to balance the oxygen chemical potential, a unidirectional flow of oxygen defects can be induced during the fabrication process of oxide heterostructures. Recently, this approach has been successfully applied to tune the charge carrier density of LaNiO3-δ (LNO) by using oxide capping layers with different oxygen vacancy formation energies26, and also to transform thin films of La0.67Sr0.33CoO3 (LSCO) from its perovskite form to the brownmillerite form La0.67Sr0.33CoO2.5 by capping thin film with Gd27. Since the underlying mechanism of this effect is not material specific, it should be feasible to apply this approach to a broader spectrum of oxide material systems.

In this work, we use a TiO2 capping layer to construct “oxygen diodes” of TiO2/VO2 heterostructures, in order to tune the MIT behavior of VO2. Interfacial oxygen vacancy migration from rutile TiO2 to VO2 is induced by the much higher oxygen vacancy formation energy (Ef) of TiO2 compared with VO2, as predicted by our ab initio many body diffusion Monte Carlo (DMC) as well as DMC-benchmarked density function theory (DFT + U) based calculations of Ef across this heterostructure. The formation of oxygen vacancies in VO2 induced by a TiO2 capping layer was also confirmed by an observed c-axis lattice expansion using X-ray diffraction (XRD), as well as by electron microscopy and X-ray photoelectron spectroscopy characterization, in agreement with our computational predictions. The MIT behavior in VO2 was drastically altered by the incorporation of oxygen vacancies. While a VO2 thin film without a TiO2 capping layer showed a sharp transition in resistivity vs. temperature, the transition was shown to be suppressed in VO2 with TiO2 capping layers. Our combined experimental and computational investigation on the TiO2/VO2 system underscores the effects of oxygen defect redistribution in oxide heterostructures on altering electronic structure and demonstrates how to use the ‘oxygen diode’ effect to control a MIT.

## Methods

### Thin film deposition

VO2 and TiO2/VO2 thin films were grown by using pulsed laser epitaxy (PLE) on TiO2 (001) single crystals (CrysTec, Germany). Growth temperature was fixed at 300 °C, and VO2 layer was grown in oxygen partial pressure (pO2) of 15 mTorr, while the TiO2 capping layers were grown in three different pO2, i.e. 10 mTorr, 15 mTorr and 20 mTorr. X-ray diffraction (XRD) was performed on the grown thin films by using a four-circle X-ray diffractometer (Panalytical X’Pert MRD).

### Transport measurement

Electrical contacts to thin film samples were made by using ultrasonic Al wire bonding. A Physical Property Measurement System (PPMS, Quantum Design) was used for measuring resistivity as a function of temperature by performing a warming up and cooling down cycle.

### STEM and EELS

Cross-sectional STEM specimens were prepared to see along the [100]TiO2 substrate direction using mechanical thinning and precision polishing followed by ion milling. High-angle annular dark-field (HAADF) imaging and STEM-EELS analysis were carried out in Nion UltraSTEM200 operated at 200 keV. For HAADF imaging, inner detector angle of 65 mrad was used, and the convergence semi-angle for the electron probe was set to 30 mrad.

### X-ray photoelectron spectroscopy (XPS)

XPS measurements were performed on a VO2 thin film sample and a VO2 thin film with a very thin TiO2 capping layer (~ 2 nm). The thin TiO2 capping layer was deposited by using the same condition as the LP sample (i.e., 10 mTorr pO2). The reason for choosing a thin thickness for the TiO2 capping layer is due to the shallow probing depth of XPS (estimated ~ 5 nm for V 2p spectra). XPS spectra of O 1s and V 2p were collected by using a system equipped with a monochromated Al Kα X-ray source ( = 1486.7 eV) and a multi-channel detector (Sigma Surface Science, Germany). XPS data collection, as well as XRD and STEM characterizations, were performed at room temperature.

### DFT calculations

Density functional theory (DFT) calculations were performed using the Vienna ab initio simulation package (VASP28). Electron exchange correlation was represented by the functional of Perdew, Burke, and Ernzerhof (PBE) of generalized gradient approximation (GGA29). The correlation effects were considered by using the DFT + U method30 with U values of (UV,UTi) = (4.0,5.0) eV as determined by scanning the two-parameter space using many-body quantum Monte Carlo (QMC) calculations (further details available in SI and “QMC calculation” section below). The ion–electron interaction was described with the projector augmented wave (PAW) method31. A cutoff energy of 400 eV was used for the plane-wave basis set. The internal coordinates were relaxed until the forces were lower than 0.01 eV/Å. The calculations were performed with spin polarization, and a ferromagnetic (FM) ordering was applied to the V sites.

DFT calculations were performed on 48-atom and 96-atom TiO2/VO2 supercells. The 48-atom supercell contains four VO2 layers and four TiO2 layers stacked along the rutile c axis (a = 6.4966 Å, b = 6.4966 Å, c = 11.5892 Å), while the 96-atom supercell contains eight VO2 layers and eight TiO2 layers (c = 23.1784 Å). Monkhorst–Pack k-point meshes of 6 × 6 × 3 and 6 × 6 × 2 were used for the 48-atom and 96-atom supercells, respectively. The O vacancy formation energy $${E}_{f}^{O}$$ was calculated using the equation: $${E}_{f}^{O}={E}^{tot}\left[Vo\right]-{E}^{tot}\left[perf\right]+{\mu }_{O}$$, where $${E}^{tot}\left[Vo\right]$$ is the total energy of the supercell with one oxygen vacancy, $${E}^{tot}\left[perf\right]$$ is the total energy of the pristine supercell, and $${\mu }_{O}$$ is the chemical potential of oxygen. We use a definition of $${\mu }_{O}$$ common in the literature32 as $${\mu }_{O}=\frac{1}{2}$$(2 $${E}^{tot}\left[O\right]+ {\varepsilon }_{O2}^{coh}$$) where $${E}^{tot}\left[O\right]$$ is the energy of an oxygen atom, and $${\varepsilon }_{O2}^{coh}$$ is the cohesive energy of oxygen molecule. From experimental data, we determined $${\varepsilon }_{O2}^{coh}$$ to be − 5.21 eV33.

### QMC calculations

Calculations using the Diffusion Monte Carlo (DMC)34 flavor of QMC were performed on 48 atom rutile TiO2/VO2 interfacial cells containing a neutral oxygen vacancy for all inequivalent sites. The V site magnetic moments were constrained to a ferromagnetic configuration. High quality pseudopotentials35 were used to represent the V, Ti, and O species. Fixed node errors were minimized by optimizing LDA + U36 parameters for both the V and Ti species separately. All other parameters of the trial wavefunction were optimized via the linear method37. All supercell results were averaged over a 2 × 2 × 2 Gamma centered supercell twist grid. The calculations were performed with the QMCPACK simulation code38 at the Argonne Leadership Computing Facility and all simulation workflows were driven with the Nexus workflow automation system39.

## Results and discussion

VO2 thin films and TiO2/VO2 heterostructures were grown on TiO2(001) substrates by pulsed laser epitaxy (PLE) in different oxygen partial pressures (pO2). The deposition conditions for the VO2 layers were fixed to a substrate temperature of 300 °C and 15 mTorr pO2, while the growth pO2 of the TiO2 capping layers was varied between 10 mTorr, 15 mTorr and 20 mTorr. Both the VO2 thin films and the TiO2 capping layers were ~ 15 nm thick. We refer to the three TiO2/VO2 heterostructures grown in different pO2 as LP, MP and HP (i.e., low, medium and high pressure). The different growth pO2 of the TiO2 capping layer effectively changes the concentration of formed oxygen vacancy in TiO2 during deposition. Figure 1 shows XRD scans on VO2 and TiO2/VO2 samples. The clearly resolved thickness fringes indicate the high sample quality. We observed a clear trend of c-axis lattice expansion following the direction of decreasing growth pO2 of TiO2 capping layer (i.e. HP < MP < LP). This “chemical expansion”40 indicates an increased oxygen vacancy concentration in LP and MP samples compared with bare VO2 and HP samples. Therefore, the XRD data is a strong proof of interfacial migration of formed oxygen vacancies in the TiO2 capping layer to the VO2 layer underneath. A lower growth pO2 introduces more oxygen vacancies into VO2 layer, which results a larger c-axis lattice spacing.

We also performed detailed structural and chemical analysis to verify the reduction of VO2 induced by TiO2 capping layers, as shown in Fig. 2. Figure 2a shows a Z-contrast (where Z refers to atomic number) high-angle annular dark-field (HAADF) scanning transmission electron microscopy (STEM) image of a TiO2/VO2 heterostructure grown on a TiO2 substrate. Clearly resolved atom columns and coherent interface again indicate the high sample quality. Figures 2b,c show electron energy loss spectra (EELS) of the Ti L2,3-edge of the TiO2 substrate and capping layer, as well as V L2,3-edge of LP and HP samples. The identical line shapes and peak positions in Ti L2,3-edge show the nearly indistinguishable chemical states between TiO2 substrate and capping layer. The well-resolved t2g and eg peaks indicate that Ti cations in the TiO2 capping layer have 4 + oxidation state with no appreciable reduction to 3 + , and do not change valence with decreasing deposition pO2. In contrast, a clear peak shift towards lower energy was observed in V L2,3-edge spectra comparing LP and HP sample, indicating a lower V oxidation state (high oxygen vacancy concentration) in the LP sample. EELS spectroscopic results on TiO2 capping layer and VO2 layer evidently reveal that the oxygen vacancies introduced by lowering growth pO2 of TiO2 capping layer migrates to the VO2 layer, resulting in the reduced V oxidation state. We believe the atomic-scale oxygen vacancy redistribution in VO2 could be affected and non-uniform along the thickness direction due to the vicinity effects associated with the surface and interface. Further investigation to determine both the quantity and spatial nonuniformity of oxygen vacancies would provide useful information to accurately understand the ion transport in VO2/TiO2 heterostructures.

The unidirectional oxygen vacancy migration from TiO2 to VO2 was also probed by using X-ray photoelectron spectroscopy (XPS), shown in Fig. 2d,e. Due to the shallow probing depth of XPS (~ 5 nm), we decreased the thickness of the TiO2 capping layer down to ~ 2 nm, while interestingly the oxygen diode effect was still apparent. VO2 is known to have overoxidized surface layers, therefore a V5+ oxidation state was present in the V 2p spectra41,42. Due to the existence of V5+ oxidation state, we used the average V oxidation state, represented by the intensity ratio of V4+ and V5+ (i.e., I(V4+)/I(V5+)), as a measure of oxygen vacancy concentration in the VO2 layer. As shown in Fig. 2d, the TiO2 capping lowered the average oxidation state of V, as indicated by a higher relative intensity of V4+ peak (higher I(V4+)/I(V5+)). On the other hand, the Ti 2p spectrum collected on TiO2 capping layer (Fig. 2e) can be fitted with only one single component of Ti4+, which is consistent with EELS results showing the near absence of oxygen deficiency in TiO2 capping layer. We would like to point out that, since the quantitative analysis of oxygen vacancies based on the local EELS data and surface sensitive XPS is challenging, we remain our conclusion be only qualitative.

We hypothesize that the increased oxygen deficiency is due to the oxygen diode effect, and results in strongly affecting the MIT of VO2, as shown by transport (resistivity ~ temperature plot) data in Fig. 3. We observed a clear MIT in the TiO2/VO2 HP sample with a sharp change in the resistivity at Tc of ~ 300 K. This is consistent with the XRD data showing that there is no difference between the c-axis lattice spacing of HP and bare VO2, indicating a nearly identical oxygen stoichiometry between these two samples. Contrarily, the sharp MIT was gradually suppressed with decreasing deposition pO2. While the MP sample showed a much lower ON/OFF ratio (ROFF/RON, defined as the ratio of resistance measured during warming up and cooling down exepriments at Tc, see Fig. 3b) compared with the HP sample, the resistivity hysteresis was completely suppressed in the LP sample. The observed suppression of the MIT is consistent with a previous report on oxygen deficient VO2-δ thin films17. We also note that our finding is consistent with a recent repot on VO2/TiO2 heterostructures43. By utilizing the proposed “oxygen diode effect” we could tune the MIT and thereby change the ON/OFF ratio by three orders of magnitudes. This therefore provides a novel route of controlling oxide electronic properties.

We further explored the mechanism of the unidirectional oxygen vacancy migration in TiO2/VO2 by performing ab initio many body diffusion Monte Carlo (DMC) as well as DMC-benchmarked density function theory (DFT + U) calculations. Because of two different 3d metal atoms, we used a hybrid U value of (UV,UTi) = (4.0,5.0) eV as benchmarked by surveying the two-dimensional parameter space using total-energies from many-body quantum Monte Carlo (QMC) calculations (further details about the calculations as well as benchmarking can be found in the SI). A 96-atom supercell consisting of TiO2/VO2 bilayers with an atomically ‘sharp’ interface was used in the DFT calculations to evaluate the (neutral) oxygen vacancy formation energy (Ef) in each oxygen atomic layer, as shown in Fig. 4. A gradient of Ef across the TiO2/VO2 interface was observed, with Ef decreasing from the TiO2 layer to the VO2 layer. Ef at oxygen layers away from the interfaces approaches the bulk Ef value of TiO2 (~ 5.8 eV) and VO2 (~ 4.6 eV). In spite of quantitative differences (Fig. S2, discussion in SI), that can be rationalized by comparing the underlying electronic-structure changes (Fig S3), the considerably lower (~ 1 eV) formation energy of oxygen vacancies in VO2 compared to TiO2 is further supported by benchmark DMC calculations performed in a smaller 48-atom supercell. These DMC calculations provide further support for a solid theoretical justification for the hypothesis that the TiO2/VO2 heterostructure behaves as an “oxide diode”, similar to the perovskite heterostructures26. Allowing effects of cation intermixing to model more realistic interface geometries, at the otherwise ‘sharp’ interface model, does not change this conclusion (Fig. S4). This implies that there is a strong chemical driving force for oxygen vacancy migration from TiO2 layer to VO2 layer, which is responsible for tuning the oxygen non-stoichiometry, and thereby achieving the suppression of the MIT, that we achieved by using TiO2 capping layers grown at different pO2 conditions. Even though we did not model the dynamic process of oxygen migration from TiO2 to VO2, we would like to point out that the growth pO2 of the TiO2 capping layer is the only parameter we altered. Since oxygen vacancies can be readily introduced into grown TiO2 by lowering growth pO2 (See Figure S5 in SI), the only feasible mechanism that responsible for the observed suppressed MIT is the oxygen vacancy mirgration we discussed above.

## Conclusions

In summary, by utilizing the “oxygen diode effect” in TiO2/VO2 heterostructure we have demonstrated tunability of the MIT and thereby changes in the ON/OFF resistivity ratio by three orders of magnitudes. This, therefore provides a novel route of controlling oxide electronic properties. Electronic-structure calculations as well as multiple characterization methods including XRD, STEM-EELS and XPS were used to prove the unidirectional oxygen vacancy migration across the interface from TiO2 to VO2. Our results highlight the importance of oxygen defect redistribution by design in oxide heterostructures on determining physical properties and functionalities.