Threshold switching stabilization of NbO₂ films via nanoscale devices

Over the past 50 years, silicon-based electronics and computational devices have undergone dramatic improvements as the size and energy usage of individual transistors have both been reduced. However, as the dimensions of a transistor approach fundamental quantum mechanical limits, there has been an increasingly frantic search for new materials and computational architectures that could replace, or supplement, CMOS-based circuits. One potential path that has been shown to benefit specific types of computations includes biomimetic device elements that have similar electronic behavior to the spiking behaviors seen in nerve cells.^1–3 

Key to developing so-called “neuromorphic” circuitry that can be transferred to a production environment is developing materials and devices that are robust and can be engineered reliably. Many of the current state-of-the-art neuromorphic materials depend on an electroforming step, in which a crystalline filamentary structure is formed in an otherwise amorphous material by passing a relatively large current between two electrodes that sandwich the neuromorphic material. This process can result in an electronic device that exhibits the typical spiking behavior that is seen in nerve cells. However, the electroforming step is stochastic in nature and may be difficult to use beyond fundamental studies. Therefore, a more robust and reliable technique for fabricating crystalline materials that exhibit neuromorphic electrical properties is required in order to intelligently and purposefully design circuit elements.

One specific material that exhibits neuromorphic electronic properties is crystalline niobium dioxide, which undergoes a phase transition from an insulator at temperatures below 800 $°$C, to a conductor at higher temperatures.^4–6 However, deposition of $NbO2$ is challenging, because the energetically preferred oxidation state at room temperature for niobium is $Nb2O5$⁠. It is well understood that the electroforming step creates a reduced filamentary structure of $NbO2$ that connects top and bottom electrodes, where that structure is embedded in the amorphous $Nb2O5$ material.³ Researchers have also grown crystalline $NbO2$⁠, which does not require the stochastic electroforming step.^7,8

$NbO2$⁠, once electroformed, has a dramatic S-type negative differential resistance that has been explained using both a Mott-type metal–insulator transition⁹ as well as a Poole–Frenkel model of Joule heating and conduction.¹⁰ Crystalline films that do not require an electroforming step undergo a Peierls transition between the tetragonal and the rutile $NbO2$ phases.^7,8 Thin films of the material show threshold switching behavior with on and off-state resistances (or currents) that can change by more than an order of magnitude.¹¹ 

Film growth most often occurs as $Nb2O5$⁠, which then requires either electroforming or annealing (or both) to change the material into conducting $NbO2$⁠. Micron-sized crossbar devices show filamentary behavior after electroforming, with filaments either in the center or at the edges of the devices, with different filament conductivities.¹² The switching behavior is also complicated by the choice of contact material, as different contact materials can give very different behaviors, from symmetric switching, asymmetric switching, switching on one polarity only, integrated selector memory switching (1S1M) and hybrid selector-memory switching (1S1R).^13,14 Often devices made directly from $Nb2O5$ and $NbO2$ both give gradually changing IV curves without strong threshold switching.¹⁵ 

Many recent studies have focused on reducing the stochastic nature of the threshold switching in $NbO2$⁠, while at the same time trying to retain large differences in the on and off state resistances and large differences in the threshold voltage (⁠$Vth$⁠) and hold voltages (⁠$Vh$⁠). Using smaller device sizes has been shown to increase both $Vth$ and $Vh$⁠, with edge conduction for device sizes of diameter less than 600 nm, such that the resistance $R∝1/d$⁠, the diameter, instead of $R∝1/A$⁠, the area.¹⁶ Adding titanium as a dopant to $Nb2O5$ makes the switching more stable and reliable, though it also appears to reduce both $Vth$ and $Vh$⁠.¹⁷ In the cases where the devices were without strong threshold switching to begin with, adding titanium has little effect on the performance beyond reducing the resistance of the films.¹⁵ 

Growing $NbO2$ directly is more difficult, but requires no electroforming step, except to break through the $Nb2O5$ oxide that grows when the sample is placed into air.¹⁸ Other researchers have found that annealing even at low temperatures below the nominal transition temperature can create crystalline $NbO2$ and reduce and eventually eliminate the electroforming voltage, without changing any of the threshold switching behavior.^7,8,19 Others have removed electroforming step by depositing multiple $NbOx$ layers with different oxygen concentrations before the top electrode with some success, though the switching is not stable.²⁰ 

In this work, we have deposited thin films of $NbOx$ using physical vapor deposition. Sputtering targets containing different concentrations of titanium were used in order to grow niobium oxide films with a homogeneous distribution of titanium, and the resulting material properties of the films were fully characterized. Electrical devices with lateral dimensions as small as 50 nm and as large as 15 $μ$m were created and measured, allowing a direct comparison between devices with a variety of dimensions. We also studied a thermal crystallization process that allowed us to directly compare electrical transport between crystalline material and the more conventional electroforming process. In all cases, a number of devices were created and measured, with statistical results presented in this study.

All $NbO2$ films were deposited in a Kurt J. Lesker PVD75 system via reactive magnetron sputtering using targets purchased from Kurt J. Lesker. The deposition occurred in an oxygen poor environment at 25 $°$C at 3% $O2$ flow in Ar atmosphere at 3 mTorr with a target power density of 6 $W/cm2$⁠. This process yields sub-stoichiometric amorphous $Nb2O5$ after the deposition which can be crystallized into the metastable allotrope $NbO2$⁠. The deposition rate and surface roughness of the annealed polycrystalline $NbO2$ thin film was measured via x-ray reflectivity (XRR).

Three binary alloyed $Nb1−xTix$ targets were purchased with Ti concentrations of 0, 5, and 10%. Thin films with a thickness of 25 nm were deposited with the respective Ti sputtering targets. Films with the different Ti dopings grew at 3.7, 4.0, and 4.3 nm/min, respectively.

X-ray photoelectron spectroscopy (XPS) was measured with a Kratos Axis Ultra DLD, which has a monochromatic Al-k$α$ x-ray source, and hemispherical analyzer. Survey spectra were measured with a pass energy of 100 eV, and high resolution spectra were measured with a pass energy of 20 eV. XPS data were analyzed with the CasaXPS software package.

Atomic force microscopy images were taken with a Park System AFM in a tapping mode. The measurements were performed in air. We used the open-source software package Gwyddion to analyze our data and to create high-quality images.

X-ray diffraction (XRD) was measured with a Bruker D8 Advance equipped with a Cu-K$α$ x-ray source. The tool was operated in grazing incidence XRD mode (GI-XRD) at an incident angle of $0.5°$ to limit interference with the Si substrate, moreover, the substrate was rotated by a few degrees to further avoid substrate interference. The spectrum was collected from 20 to $75°$ 2$Θ$ at a rate of 10 s per $0.02°$ yielding one spectrum in 7.64 h.

The SUNY Poly nanoscale test structure was utilized to fabricate nanoscale $NbO2$ devices that are electrically measurable. This structure implements a W interconnect layer with a minimum pitch of 180 nm below a TiN electrode that serves as the bottom electrode for two-terminal devices. The TiN electrode is subtractively structured and embedded in $Si3N4$ insulation. The available TiN electrode sizes range from $50×50nm2$ to $170×170nm2$ allowing for a size dependent evaluation of thin film characteristics. The electrical contact to the TiN pillar is created through the W interconnect layer for which an opening in the SiN insulation is created via an additional RIE process after the TiN electrode is finished. An $NbOx$ layer is deposited on top of the TiN electrode and subsequently capped in situ with 10 nm Ir and 100 nm W. Ir serves as the inert electrode, while W improves the electrical contact from the probe contact point to the device. A micron-sized top electrode and device was structured via contact lithography and a final RIE process. A schematic of the sample is shown in Fig. 3(a). A crystallization anneal of the nanoscale devices in nitrogen atmosphere at 1 atm pressure and 700 $°$C for 5 min was performed to yield polycrystalline $NbO2$⁠.

Larger crossbar structures were fabricated at the Cornell Nanoscale Science and Technology Facility with 50 nm-thick Pt bottom electrodes ranging in width from 2 to 15 $μ$m on $SiO2$⁠. Blanket $NbOx$ deposition occurred on top of the bottom electrode and then capped with a top electrode of a 10 nm adhesion layer of Ru followed by 50 nm of either Au or Pt. Device patterning used standard liftoff photolithography and RIE for cleaning and to etch through the $NbOx$ layer to reach the bottom electrode. The schematic of the cross devices is shown in Fig. 3(d). The $NbOx$ films in the crossbar devices were not thermally annealed after deposition and formed a native oxide layer of $Nb2O5$ between the film and the top electrode.

The nanoscale devices were characterized on a 300 mm semiautomatic probe station utilizing a B1500A semiconductor parameter analyzer from Keysight with a custom automation script for resistive and selector devices. The script executes 100 IV cycles from 0 to 4 V with a compliance current of 1 mA. We measured 10 devices per device size.

The microscale crossbar devices were measured using a Keithley 2400 SourceMeter in a voltage control mode with a compliance current of 2 mA. IV curves were initially swept from $−$1 to 1 V, with some sweeps reaching voltages as high as $±$5 V. Each device underwent at least 50 sweeps from positive to negative voltage. Measurements of the nanoscale and microscale devices were conducted at room temperature.

The as-grown materials were characterized with XPS, XRD, and AFM. XPS survey scans and high resolution spectra were measured for blanket film depositions using similar deposition conditions as on the device substrates. For all of the samples, the survey scans indicate the presence of Nb, O, adventitious C contamination, and for samples deposited with a Ti target, a small amount of Ti.

High resolution scans for binding energies corresponding to the niobium 3D doublet can be found in Fig. 1(b). The undoped sample has Nb peaks at around 207.7 and 210.4 eV, which corresponds to the doublet for $Nb2O5$ (+5 oxidation state). Typically, when $NbO2$ is exposed to atmospheric conditions, it will form a thin surface oxide of $Nb2O5$ due to the interaction with atmospheric oxygen.^6,21–23 However, for $NbO2$ bulk films, the surface oxide is typically limited to the top 1.5 nm, and XPS is expected to also reveal a peak around 206.1 eV, which corresponds to the underlying $NbO2$⁠.¹⁸ The lack of a significant secondary peak implies that either Nb is exclusively $Nb2O5$⁠, or that the surface oxide is thicker than the probe depth of XPS (⁠$∼$2 nm).

For the samples that were synthesized with Ti targets, the high resolution Nb spectra indicated the presence of both $Nb2O5$⁠, and $NbO2$⁠. We presume that the Nb in the +4 oxidation state (the smaller doublet at around 206.1 and 208.8 eV) is beneath a thin surface oxide, as has been reported elsewhere.^6,21,22

For all three samples, the concentration of titanium was extracted based on fits to the high resolution spectra for Nb, O, and Ti. We note that the Nb to Ti ratio in the sputtering target is not reflected in the films themselves, for reasons not fully understood at the time of publication. Possible explanations include kinetic factors related to the deposition as well as different sputtering cross sections for Nb and Ti. The sputtering target has not been independently measured as part of this study.

The AFM images look qualitatively similar between all three of the samples. A representative image can be found in Fig. 1(a). RMS roughness values for the three different Ti concentrations are 0.34, 0.45, and 0.52 nm, for 0, 1.0%, and 1.8% Ti, respectively.

The as-grown films are amorphous. X-ray diffraction results presented in Fig. 2 show that atmospheric pressure anneals in nitrogen gas as high as 600 $°$C create only slight crystalline $NbO2$ peaks. Upon annealing at 700 $°$C for 5 min in $N2$⁠, we see the emergence of polycrystalline $NbO2$⁠, with peaks that align well with expectations. Similar to other results, we find $NbO2$ growth well below the Peirls phase transition temperature $≈800°$C.¹⁹ We chose to anneal our films at 700 $°$C, which is the lowest temperature where clear $NbO2$ peaks form. Lower temperature anneals may enable compatability with traditional CMOS architectures.

To compare the effects of our two device architectures, we prepared multiple devices. The device schematics can be seen in Figs. 3(a) and 3(d).

Typically, as-grown $NbOx$ films are amorphous with niobium mostly in the +5 oxidation state (⁠$Nb2O5$⁠). In the as-grown films, an electroforming step is required to create a conducting pathway of crystalline $NbO2$ for current conduction. This electroforming step is not required for as-grown crystalline $NbO2$ films (e.g., layer-by-layer growth via molecular beam epitaxy)⁸ or for thermally annealed films. Thermal annealing between 500 and 800 $°$C promotes growth of crystalline $NbO2$⁠, as shown in Fig. 2, and an electroforming step is not required for these films.^7,19 Representative IV curves for the nanoscale devices can be seen in Figs. 3(b) and 3(c). Because the nanoscale devices were annealed at 700 $°$C, there was no electroforming step required.^7,19

Thermal annealing creates bulk crystalline $NbO2$⁠, which is highly conductive. Thus, thermally annealing the large-area microscale devices would effectively short the device. As a result, an electroforming step was required in the microscale devices (not shown).¹⁷ We electroformed our microscale devices at about $≈$1.8 V, but the electroforming voltage varied from as small as 1 V to as much as 15 or 10 V. Filament formation the microscale devices is most likely via dielectric breakdown, given the low conductivity of the films and IV-curves consistent with those shown in Ref. 12. Overall, our devices show similar behavior to others in the literature.^15,19

The smallest nanoscale devices show clearly asymmetric switching behavior, with positive biases leading to threshold switching behavior and negative biases without switching entirely. This behavior is due to the asymmetry in the material stack, and is similar to the unipolar or the integrated selector memory switching (1S1M) shown in other non-uniform stacks.^13,24,25 Because the nanoscale devices do not show the memristive threshold switching behavior at negative biases, we focused on the nanoscale devices under a positive bias only.

Researchers have spent considerable effort trying to reduce the stochastic nature of the threshold switching in $NbOx$ devices. In this work, we systematically examined different device sizes, characterizing them by their threshold and hold voltages. The threshold and hold voltages are marked on the IV-curves in Fig. 3. These voltages are obvious when the switching is clear, as in Fig. 3(c). For IV curves similar to Figs. 3(e) and 3(f), the threshold and hold voltages are the points where the change in slope along the IV curve is the greatest.

The threshold voltages are presented as a box and whisker plot in Fig. 4 for $n$ different IV curves from $m$ devices (approximately 100 IV curves per device). The hold voltages for the nanoscale devices (not shown) did not appreciably vary, with $Vh≈1.0$ V.²⁶ From Fig. 4 we can see that the smallest nanoscale devices have $Vth−Vh≳1$ V, making the smallest nanoscale devices promising candidates for use in neuromorphic computing.

The microscale crossbar devices require an electroforming voltage to form a conducting path (not shown), common in the literature. Once the conducting path is formed, the threshold and hold voltages are symmetric with respect to bias and consistent across the different device sizes, as shown in Figs. 3 and 4, with hold voltages of $Vh≈0.7$ V.²⁷ These results indicate that the microscale devices are all similar, despite the different active areas of the devices. We infer that these devices form filamentary paths through the material, whose formation is driven by dielectric breakdown,¹² and that once formed, the size of the filament is roughly similar despite the varying device sizes. Similar filamentary behavior independent of device size has been reported previously.¹⁴ In addition, our unannealed, electroformed devices do not show robust threshold switching, similar to others in the literature.^15,19

We can further investigate the bulk vs filamentary behavior by examining the devices far from the switching behavior. We measured the resistance of our devices at 0.4 V, much below the hold and threshold voltages in our devices, plotted in Fig. 5.

The microscale devices show little change in resistance even as the device area increases, again supporting the creation of filaments in the material to conduct the current.¹⁴ Our results match those often reported, where electroforming creates a conducting $NbO2$ filament. We expect these filaments to be the same conductivity and size, thus the resistance of the device does not change, despite the factor of $∼$100 increase in device area.

In Fig. 5, our results in the smallest nanoscale films show that the resistance scales inversely with the device area, $R≈ρdA−1$⁠, where $ρ$ and $d$ are the resistivity and thickness of the film, respectively. This is in contrast with earlier work that saw that nanoscale devices vary with the circumference rather than the area,¹⁶ and in contrast with other work that found different conductivities in different sized devices,¹² however, these earlier works both required an electroforming step. This electroforming step is stochastic by nature. In our nanoscale devices, we avoid the stochastic electroforming step and find that the resistance scales inversely with area. Given that our films are similar in thickness, we can conclude that the conductivity in our device areas are uniform and constant—we observe a bulk response in our nanoscale devices. Thus, we can conclude that our nanoscale devices are crystallized to an extent that the conduction path is a bulk $NbO2$ effect (rather than percolation), even if the material is not fully crystallized. These results are in stark contrast to previous work.¹⁹ Our results clearly show that both annealing and small device sizes are necessary to create stable and reproducible resistivities and threshold switching.

The largest nanoscale device (⁠$170×170nm2$⁠) deviates from the expected behavior of $R∝A−1$⁠. When comparing the devices, either the IV curves in Fig. 3 or the threshold voltages in Fig. 4, we see a gradual change as the nanoscale devices begin to behave in a similar manner as the microscale crossbar devices. This change occurs as the nanoscale devices reach the size of the filaments (roughly 300 nm).¹² We speculate that the Joule heating in the $170×170nm2$ device anneals the larger device in situ such that it behaves similar to the electroformed filaments in the microscale crossbar devices. These results indicate that there is a limit to the size of the devices: “nanoscale” alone cannot insure reproducible threshold switching—the size of the device must be smaller than the size of electroformed filaments.

The inclusion of titanium as a dopant in $NbO2$ films has shown some promise in its ability to stabilize the threshold switching in niobium oxide films. Our results in $2×2μm2$ crossbar devices are shown in Fig. 6, for films with Ti doping of 0, 1.0, and 1.8% (Fig. 1).

The change in threshold and hold voltages are shown as a box and whisker plot in Fig. 6. We see no significant change in the threshold or hold voltages upon addition of titanium. In devices similar to ours, where the films are a stoichiometric mix of $NbO2$ and $Nb2O5$⁠, researchers report both gradual threshold switching and no significant change in the threshold voltages upon the addition of as much as 7% Ti.¹⁵ In contrast, microscale devices based on stoichiometric $Nb2O5$ show stabilized switching and reduced $Vth$ and $Vh$ with as little as 0.7% Ti.¹³ 

The reason for these contrasting results with Ti doping are unclear. It is not clear if Ti-doped, crystalline $NbO2$⁠, nanoscale devices will also show threshold switching stabilization, or if threshold switching stabilization only occurs in Ti-doped $Nb2O5$ device. Our studies of the effect of titanium doping in the nanoscale devices are ongoing.

We have studied the effect of growth and annealing conditions as well as device size on the stabilization of the threshold switching in as-deposited sub-stoichiometric $Nb2O5$⁠. We have found that annealing at 700 $°$C, well below the $NbO2$ transition temperature of 900 $°$C, crystallizes the film into $NbO2$⁠. Nanoscale devices with the annealed $NbO2$ film show bulk crystallization and uniform resistivity, with large threshold voltages, and are excellent candidates for memristive devices in neuromorphic computing.

The microscale crossbar devices show clear evidence of filament formation after electroforming, with resistances that do not change as the device area is increased. The low threshold voltages make these devices poor candidates for memristor applications. The addition of titanium as a dopant, which has stabilized and improved the reliability in as-deposited $Nb2O5$ devices,¹⁷ did not improve switching in our sub-stoichiometric films.

Finally, we are able to see the transition from the nanoscale, with bulk crystalline $NbO2$ devices, to the microscale, with filament growth, in the largest nanoscale device, $170×170nm2$⁠. These devices, roughly the same size as the filaments created by electroforming, have resistances lower than expected from $R∝A−1$ and IV curves that nearly resemble the microscale devices. We posit that internal Joule heating self-anneals the largest nanoscale device to create behavior similar to the electroformed filaments.

This research was supported by the National Science Foundation (NSF) Grant Nos. DMR-2103197, DMR-2103185 and the Air Force Research Laboratory Grant No. 1152303-1-83972.

This work was performed in part at the Cornell NanoScale Facility, a member of the National Nanotechnology Coordinated Infrastructure (NNCI), which is supported by the National Science Foundation (Grant No. NNCI-2025233) and made use of the Cornell Center for Materials Research Shared Facilities, which are supported through the NSF MRSEC program (No. DMR-1719875). The nanoscale templates were fabricated at the Albany Nanotech Complex via NY CREATES and Dr. Sandra Schujman at NY CREATES assisted with x-ray measurements.

The authors have no conflicts to disclose.

M. C. Sullivan: Conceptualization (equal); Data curation (equal); Formal analysis (equal); Funding acquisition (equal); Investigation (equal); Methodology (equal); Project administration (equal); Supervision (equal); Validation (equal); Visualization (equal); Writing – original draft (equal); Writing – review & editing (equal). Zachary R. Robinson: Conceptualization (equal); Formal analysis (equal); Funding acquisition (equal); Investigation (equal); Methodology (equal); Project administration (equal); Resources (equal); Supervision (equal); Validation (equal); Writing – review & editing (equal). Karsten Beckmann: Conceptualization (equal); Data curation (equal); Formal analysis (equal); Funding acquisition (equal); Investigation (equal); Methodology (equal); Validation (equal); Writing – review & editing (equal). Alex Powell: Investigation (equal). Ted Mburu: Data curation (equal); Visualization (equal). Katherine Pittman: Investigation (equal). Nathaniel Cady: Funding acquisition (equal).

The data that support the findings of this study are available from the corresponding author upon reasonable request.