IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 1 A Low-Power Nonvolatile Switching Element Based on Copper-Tungsten Oxide Solid Electrolyte Michael N. Kozicki, Member, IEEE, Chakravarthy Gopalan, Student Member, IEEE, Muralikrishnan Balakrishnan, Student Member, IEEE, and Maria Mitkova Index Terms—Copper electrodeposition, nonvolatile memory devices, Raman spectroscopy, resistance change, solid electrolyte, tungsten oxide, X-ray photoelectron spectroscopy. I. INTRODUCTION T by way of electrodeposition at low voltage and current reduces the resistance of the electrolyte by several orders of magnitude. Our earlier work in this area concentrated on two-terminal devices, with active area as small as 40 nm, based on a silver doped chalcogenide glass solid electrolyte film sandwiched between a silver anode and an inert cathode [5]. However, we have recently added electrolytes consisting of transition metal oxides combined with copper to our studies and it is this work that we describe in this paper. is a particularly desirable base Tungsten trioxide material for a solid electrolyte as it is compatible with back-end-of-line (BEOL) processing in CMOS integrated circuits where tungsten metal already plays a significant role. Tungsten oxide formation techniques are compatible with semiconductor processing and include wet chemical and plasma oxidation, and physical vapor deposition. As we have demonstrated, coupling this simple oxide with copper leads to a solid electrolyte with desirable qualities for electrochemical device operation. Copper is a particularly appropriate mobile ion choice in the context of integration into CMOS processes as many high performance ICs already use copper, along with tungsten, in the upper levels of metal. Fig. 1 illustrates how the integration of such a switching element could be achieved. In this scheme, selected tungsten plugs which are normally used to form the connections between one level of metal and another are used as the lower electrode in the devices and the copper interconnect is the upper electrode. The solid electrolyte is placed between the two electrodes and is switched between high and low resistance states by the application of appropriate voltages on the electrodes. Note that during processing, only one additional mask is required over the normal CMOS logic to define which plugs have switching devices and which are allowed to remain as simple connections between interconnect layers and so the approach is inherently low in incremental cost. This paper focuses on the nature of copper-doped tungsten oxide and on devices comprising this material. A comparison is made between the quasi-static electrical characteristics of devices based on plasma-grown and deposited oxides, the copper being added by photodissolution in both cases. IE W EE eb P Ve ro o rs f ion Abstract—We describe the materials aspects and electrical (Cu WO3 ) Cu switching elements. characteristics of W These materials are compatible with back-end-of-line processing in CMOS integrated circuits where both tungsten and copper already play a significant role. Devices based on Cu WO3 solid electrolytes formed by photodiffusion of copper into tungsten oxide switch via the electrochemical formation of a conducting filament within the high resistance electrolyte film. They are able to switch reversibly between widely spaced nonvolatile resistance states at low voltage ( 1 V) and current ( 10 A). Electrical characterization revealed that devices consisting of plasma-grown oxides have a variable initial threshold voltage and poor retention, whereas devices based on deposited oxide exhibit a stable switching threshold and good retention, even at elevated operating temperature ( 125 C). This difference in behavior was attributed to the observation that the copper tends to oxidize in the plasma-grown oxide whereas the copper in the deposited oxide exists in an unbound state and is, therefore, more able to participate in the switching process. HERE is a considerable driving force within the semiconductor industry to create scalable elements that can switch between widely spaced nonvolatile resistance states at low power. Such elements could find widespread application in next generation memory and reconfigurable logic [1]–[3]. Whereas there are several new technologies that show promise in this respect, they typically lack complete scalability, i.e., they have a physical size, programming voltage, or programming current that is excessive for high density systems at the 65-nm node and beyond. Many are also difficult to integrate due to the complexities associated with the additional processing steps and masking levels required and, therefore, will be expensive to add to highly scaled CMOS. One potential approach to this problem involves switching elements, known as Programmable Metallization Cell (PMC) devices, which utilize the reduction of nanoscale quantities of metal ions in solid electrolyte films [4]. The formation of a robust but reversible conducting pathway II. DEVICES BASED ON SOLID ELECTROLYTES AND OXIDES Manuscript received March 30, 2006; revised May 1, 2006. This work was supported by Axon Technologies Corporation. The review of this paper was arranged by Associate Editor B. G. Park. The authors are with the Center for Solid State Electronics Research, Arizona State University, Tempe, AZ 85287-6206 USA (michael.kozicki@asu.edu). Color versions of Figs. 2–12 are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TNANO.2006.880407 In our earlier work [4], [5], we presented the formation and operation of PMC devices composed of a silver doped chalcogenide glass solid electrolyte film (e.g., Ag–Ge–Se or Ag–Ge–Se) sandwiched between a silver anode and an inert cathode. For an applied bias in excess of a few hundred mV, an electron current from the cathode reduces silver ions in the 1536-125X/$20.00 © 2006 IEEE 2 IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 resistivity which were originally attributed to homogeneous modifications or an unspecified breakdown mechanism. These explanations were dismissed by Dearnaley and colleagues [8], who developed a filamentary model for the changes in resistance. It was claimed that since the anode insulator interface is not completely smooth, there will be places where the electric field is locally high, resulting in the formation of a conductive filament by field induced migration of material from the anode into the oxide. The metallic filament propagates through the insulator and thereby reduces the resistance of the oxide layer. Subsequent switching in this “electroformed” structure was thought to be via electrically stimulated rupture and healing of the filamentary connection at locations along its length. Thurstans and Oxley [9] proposed a model which took other characteristics of the electroformed structure, in particular the observed negative differential resistance, into account. Their model features a conduction mechanism that depends on trap-controlled thermally activated tunneling between isolated segments of a ruptured metallic pathway. Recently, Baek et al. [10] used such models to describe the unipolar switching characteristics of their transition metal oxide-based devices. In these structures, nickel oxide is switched to a low resistance state using a relatively high voltage/low current “set” pulse and returned to a high resistance state using a low voltage/high current “reset” pulse of the same polarity. Although the effect is attributed to charge trapping mechanisms in the electroformed oxide, it is perhaps more likely to be due to the rupture and healing of a metallic pathway. We have observed similar unipolar programming in PMC devices in which a metallic filament is formed by the electrodeposition of the anode metal in the electrolyte at a particularly high critical voltage but can be disrupted using a current pulse of the same polarity [11]. In this case, the erase (reset) current must be in excess of the current used to write (set) the device so that the current carrying capacity of the filament is exceeded and it is thereby severed, however, the erase voltage must be below that required for electrodepostion of the metal so that the break does not spontaneously re-close. Note that for the devices described in [4], the voltage threshold for electrodeposition is sufficiently low that a reverse bias is required to erase the device; a forward current in excess of the original programming current limit leads to further electrodeposition and the device resistance falls to a lower stable value rather than rising to the off state. The consideration of tungsten oxide as a base glass for the electrolyte in PMC devices led to a number of obvious questions in the context of the above discussion. The main issue was whether the devices would behave like the chalcogenide-based variants or more like the metal-insulator-metal structures described by Oxley. To address these issues, we engaged in a program of material analysis and device characterization, described in the following sections. IE W EE eb P Ve ro o rs f ion electrolyte while an equivalent number are injected from the anode via oxidation of the silver. A metal-rich pathway forms in the electrolyte film via this electrodeposition process. The level and duration of the ion current above the electrodeposition threshold determine the quantity of Ag electrodeposited and, hence, the resistance of the pathway; the resistance of the structure falls by many orders of magnitude even for applied currents as low as a few microamperes. The low resistance electrodeposit is electrically neutral and stable, leading to retention in excess of 10 years [1], [4], however, applying a bias with opposite polarity will break the link by a combination of current-induced rupture and electrochemical oxidation of the electrodeposit. Note that when a reverse bias is applied, the break cannot “heal” by electrodeposition as there is a net removal of metal from the electrolyte in a process which is opposite to that which formed the metal filament in the first place. The reverse ion current flows until the previously deposited Ag has been oxidized and returned to the silver electrode. Thus, the resistance increases again until the high resistivity of the solid electrolyte is achieved. The oxidation/reduction process is fast (tens of nanoseconds or less for write and erase), consumes little power (in the microwatt range), and can be cycled well beyond conventional nonvolatile memory technology [4], [5]. The electrical characteristics of PMC devices comprising chalcogenide-based solid electrolytes result from the unique nanostructure of the electrolyte material. We previously established that the dissolution of Ag into Se- or S-rich base glasses produces a ternary that is a combination of a dispersed superionic nano-crystalline Ag-rich phase within a glassy insulating Ge-rich phase [6], [7]. The Ag ion-containing regions are typically less than 10 nm in diameter and are separated by a few nanometers of the insulating phase [5], [6]. It is this particular nanostructure that allows the films to exhibit a resistivity of 100 cm or more (which yields a high device off resistance) while retaining good ion availability and mobility. The accessibility of mobile Ag ions for electrodeposition throughout the electrolyte allows the devices to switch rapidly as the growing electrodeposit will always have a local source of ionic metal to feed its formation. The reduced ions are replaced via the ion current from the anode and so the electrolyte does not become depleted of ions following switching and contains “excess” metal within the electrodeposited pathway. Indeed, it is this addition of metal into the already metal ion-rich but nano phase-separated electrolyte that leads to both the reduction in resistance of the structure and the nonvolatility of the on state. To complement our work on chalcogenide-based PMC devices, we have recently examined elements that are more closely related to the materials present in modern integrated circuits. Tungsten metal and its oxide and Cu are obvious choices in this respect. There has been widespread interest in the use of a variety of oxides in switching/memory applications for a number of years. Several models exist that describe the electrical phenomenon seen in thin oxide films sandwiched between two metal layers but as yet no complete or consistent model describes all the observations. Films from tens to thousands of nanometer thick when subjected to certain critical electric fields have been observed to undergo permanent large reductions in III. TUNGSTEN OXIDE CHARACTERISTICS The electrical properties of have been widely investigated because of the existence of electro- and photo-chromic effects in materials based on this oxide [12]–[14]. It is a wide is bandgap dielectric—the electronic gap of amorphous KOZICKI et al.: LOW-POWER NONVOLATILE SWITCHING ELEMENT BASED ON COPPER-TUNGSTEN OXIDE SOLID ELECTROLYTE be characterized. Note that in the fabrication of device structures (discussed later), a patterned tungsten electrode was used and the excess metal following photodiffusion was left in place on the surface of the electrolyte to form part of the anode top electrode. Characterization of these films was performed using Raman spectroscopy and X-Ray Photoelectron Spectroscopy (XPS). Raman spectra were performed in the micro Raman mode using a 514.5 nm Ar+ laser, 100 objective, with an accumulation time of 100 s and a power of 6 mW. The experimental curves were fitted using Labcalc software. X-ray Photoelectron Spectroscopy (XPS) was carried out with high-resolution XPS X-ray excitation and constant analyzer energy with (Mg 8 eV pass energy, take-off angle of 90 , 240 W power). The energy calibration of the spectrometer was performed using a gold plate fixed to the sample. Raman data related to plasma oxidized films are presented in Fig. 2. Since the films are predominantly amorphous, only breathing modes resulting from the short-range order appear. The intensity of Raman signals from the films is inherently weak because they are rather thin but the deconvolution of the results permits collection of meaningful information relating to the vibration of the W-O bond. This revealed a mode at 280 cm (bending mode), very weak modes at 815 and 930 cm and one at 998 cm (stretching film is illuminated with light, a mode at mode). After the , suggesting 450 cm appears that could be due to some reduction has occurred [22]. Illumination with light evidently causes some crystallization of the films, an effect that is manifested in the appearance of a peak at 719 cm , bond. Such a structure has also characteristic of the been observed by Delichere et al. [23] who relate it to a higher order of structural organization. When Cu is photodiffused into the films, modes at intermediate frequency appear (440 and 460 and cm ) that could be attributed to the vibrations of modes [24]. This is significant as combinatorial electrochemical synthesis [24] has demonstrated that the formation of Cu oxides is possible following introduction of Cu in the films. The XPS data of plasma oxidized films show that they are indeed influenced by previous exposure to light Fig. 3(a) and (b). Gaussian deconvolution of the W4f spectrum of the initial and films Fig. 3(a) shows well resolved doublet caused by spin-orbit coupling with binding energies of 35.85 and 37.94 eV, which correspond to a typical 6 state of W. However illumination with light causes a small shift of this douand blet towards lower energies Fig. 3(b). The show binding energies of 35.71 and 37.8 eV that could be an indication of some reduction of the oxide. We could not observe a well expressed simultaneously occurring W4f double peak of as was reported by Sun et al. [25] since the amount of the 5-valent W is small. We suggest that this is related to the fact that there is no water included in our films. However the shift is certainly an indication of the influence of light on our films. The introduction of Cu causes significant changes in the XPS spectra Fig. 4. The Gaussian deconvolution suggests the appearance of 4 binding energy values that could correspond to the 6 and 5 state of W, as expected. However, they are somewhat shifted from the standard data for these particular doublets. It is diffi- IE W EE eb P Ve ro o rs f ion 3.25 eV [15]—but a large range of conductivity has been reported [16]. This is due to the fact that W can exist in different valence states when oxidized, thus forming oxides with different stoichiometry and, hence, with different electrical conductivity. Different valence states can also be created by illumination with UV light with energy exceeding 3.3 eV [17]. In can be altered by the introaddition, the conductivity of duction of metal ions and this is of particular interest in the context of the current work [18]. As discussed earlier in this paper, we have used silver as the mobile ionic species in the base glass (and as the oxidizable electrode) in chalcogenide-based PMC devices due to its excellent electrochemical qualities but copper is also a good candidate due to its high ionic mobility in a variety of materials, including oxides [19]. To combine metals and thereby form a stable solid electrolyte, we can with take advantage of the photosensitivity of the oxide. This effect more commonly manifests itself as photochromism which can be explained by a double charge-injection model [20], in which and electrons into a lattice forms a tunginjection of sten bronze structure , . Note that can in tungsten oxide hydrate or be partially reduced to in 12-tungstates by electrons occupying the empty orbital is typically in due to the electron localization effect. photochromic but can actually be a variety of mobile ions, . So, if the film is prepared using a dry including reaction and a very thin UV transparent Cu film is evaporated on top, illumination can cause the formation of a bronze. formation is concerned, although wet chemAs far as ical or plasma oxidation of tungsten metal are simple room temperature processes, the thickness of the layer formed is limited to a few nanometers by the diffusion of the oxidant species through the growing oxide. It is, therefore, necessary to use the evaporation of tungsten oxide to form thicker films. To grow thin tungsten oxide films on tungsten metal, we elected to use plasma techniques as we discovered that the growth of layers by wet chemical means was difficult to control and led to material with highly variable characteristics. In addition, the wet grown oxides also contain species that would possibly interfere with the incorporation of Cu. The surface of 100-nm-thick tungsten glow dismetal films was oxidized using a nitrous oxide charge at an RF power of 250 W and a substrate heater temperature of 300 C for 10 min. in a PlasmaQuest plasma processing system. The flow rate was 20 sccm with 1000 sccm of helium carrier gas. The precise thickness of the grown oxide was difficult to ascertain due to the roughening of the underlying tungsten caused by the growth process but it was estimated using ellipsometry and capacitance-voltage measurements to be in the order of 3 to 4 nm thick. To achieve metal diffusion, 25-nm-thick Cu layers were deposited by high vacuum evaporation and the metal was photodiffused into the oxide using the 405-nm broadband UV light source of a Karl Suss MJB 3 mask aligner for 14 . This wavelength was min at a power density of 4.5 [21], where chosen as it is near the absorption edge of the the light causes maximal change in the material. Blanket samples were fabricated for material characterization and the excess surface metal was removed using a short treatment with iron nisolution so that only the Cu in the oxide would trate 3 4 IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 IE W EE eb P Ve ro o rs f ion Fig. 1. Schematic cross section of an integrated circuit, showing position of oxide switching elements on selected tungsten via plugs. Interconnect layers M1 and M2 are separated by IMD. Through vias provide M1 to M2 connectivity. WO Fig. 2. Raman analysis of formed by plasma oxididation of W (a) as synthesized; (b) after illumination with UV light; (c) photo-doped with Cu. cult to fully interpret these results but we believe that they could be related to distortions in the amorphous structure of the films and formation of clusters [26] in which some of the diffused Cu is oxidized and the clusters become part of the overall structure. This leads to the occurrence of bond overlapping and a shift in the binding energies. Overall, even though the signal-to-noise ratio is rather poor, the data related to Cu does indeed indicate its presence in an oxidized state within the tungsten oxide matrix (see inset in Fig. 4). We now turn to the deposited oxide samples. The evaporation source was high purity tungsten oxide (99.99%) in 3–12 mm sintered yellow-green pieces (Cerac, Inc.). Using high vacuum thermal evaporation, 50- or 100-nm thick films were deposited on the substrates described above. Following this, 25 nm of Cu without breaking the deposition was evaporated on the system vacuum. The Cu was then photodiffused into the oxide for 14 min and the excess metal was etched using the techniques described earlier. The analysis of these films revealed some simwas clearly ilarities with the plasma-grown oxides in that present as before in the as-deposited films, indicating the forma, but the way the copper was incorporated in the detion of posited oxide appeared different. Both Raman and XPS analysis of the films gave little or no evidence of the formation of Cu-ox. We ides after copper is photodiffused into the deposited WO Fig. 3. XPS spectra of formed by plasma oxididation of W (a) as synthesized; (b) after illumination with UV light. are currently investigating this in more detail but it would appear that whereas at least some of the photodiffused Cu becomes oxidized in the plasma-grown oxide to form a two phase (copper oxide/tungsten oxide) system, it exists in an unbound state in the deposited material. This could be due to the fact that the deposited oxides are in a highly stable state and do not readily react with the diffusing copper, whereas the presumably substoichiometric plasma-grown oxides are much more reactive. In any case, this dissimilarity is extremely significant in the context of the observed differences in the electrical characteristics 5 IE W EE eb P Ve ro o rs f ion KOZICKI et al.: LOW-POWER NONVOLATILE SWITCHING ELEMENT BASED ON COPPER-TUNGSTEN OXIDE SOLID ELECTROLYTE WO Fig. 4. XPS spectra of formed by plasma oxididation of W after Cu diffusion in the oxide film; inset—XPS spectrum of Cu-containing species in the oxide film. of devices based on plasma-grown and deposited oxides, as described in the following sections. IV. DEVICE FABRICATION AND ELECTRICAL CHARACTERIZATION A. Devices Based on Plasma Grown Oxide Tungsten oxide films were incorporated in PMC device structures using oxygen plasma treatment of a tungsten metal bottom electrode as described in the previous section and with copper as the mobile species and oxidizable top electrode. Schematics of typical test device structures are shown in Fig. 5 (Fig. 5(a) represents devices based on grown oxide whereas Fig. 5(b) is the deposited oxide variants described in the following section). A 100-nm-thick tungsten layer was deposited by chemical vapor deposition on silicon dioxide grown on silicon substrates. The tungsten was covered with 100 nm of silicon dioxide by chemical vapor deposition and via (through) holes were defined in this dielectric using optical lithography and plasma etching. The exposed tungsten in the holes was oxidized using the plasma oxidation technique described in the previous section. A 25-nm-thick Cu metal layer was deposited using high vacuum thermal evaporation and photodiffused into the using a 405-nm broadband UV light source in a Karl Suss MJB . The excess 3 mask aligner at an intensity of 4.5 Cu metal was not removed to ensure the presence of a Cu-rich source on the electrolyte surface. A 50-nm-thick Cu metal layer was then deposited using high vacuum thermal evaporation and patterned using a lift off process to complete the oxidizable electrodes of the devices. Note that this device configuration was absolutely necessary to observe switching regardless of how the base glass for the electrolyte was formed; a lack of Cu in the top electrode results in no measurable switching activity. The Fig. 5. Cross-sectional schematics of oxide-based switching devices. The active device area is the tungsten-electrolyte-Cu stack with area defined by the dielectric. (a) Device based on grown diameter of the through hole in the oxide. (b) Device based on deposited oxide. SiO probe pad window in the to the tungsten layer was formed using another lithography step in conjunction with wet-chemical etching. The quasi-static characteristics of the test devices were obtained by connecting the device electrodes via tungsten probes held in micromanipulators in a probe station to a semiconductor parameter analyzer (Agilent 4155C). Voltage doublesweeps were carried out starting at maximum reverse bias (tungsten electrode positive, copper electrode negative), sweeping to an appropriate forward voltage (tungsten electrode negative, copper electrode positive), and sweeping back again to the reverse bias starting point. Fig. 6 shows a resistance-voltage plot from our earlier work on large (1– m diameter) devices with a Cu/plasma-grown electrolyte and Cu oxidizable electrode [4]. Three consecutive sweeps are shown from 0.75 V to 1.0 V to 0.75 V for a current limit of 0.5 A. The maximum off resistance and the transition for the first sweep is in excess of to the low resistance state occurs at 0.9 V at which point the current reaches its compliance limit. The current stays at the compliance until the negative-going sweep reaches 0.25 V, , more at which point the measured on resistance is 500 than four orders of magnitude lower than the off state. Note that the apparent rise in resistance between 0.25 and 1 V is an artifact of the current compliance control in the measurement instrument. The appearance of the lower threshold following switching is typical of all PMC devices and represents the minimum voltage required for electrodeposition once the process has been initiated and a metal-rich region has formed in the electrolyte. Electrodeposition within the oxide will continue as long as the voltage drop is greater than 0.25 V IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 IE W EE eb P Ve ro o rs f ion 6 Fig. 6. Resistance-voltage plot of a 1.0-m-diameter device with Cu/plasmaelectrolyte and Cu oxidizable electrode. grown Fig. 7. Resistance-voltage plot of a 240-nm-diameter device with Cu/plasmaelectrolyte and Cu oxidizable electrode. grown for this physical configuration. Note that the decreasing resistance due to the electrodeposition effect increases the current flowing through the device until the compliance is reached. At this point, the voltage drop falls to the threshold and the electrodeposition stops. The on resistance is, therefore, defined by the electrodeposition threshold divided by the current limit . It is clear from the plot of Fig. 6 that the first sweep is different from the following two. The off resistance is apparently reduced by an order of magnitude following the first sweep and the switching threshold moves from 0.9 V in the first sweep toward 0.7 V in the latter cycles. The same switching threshold lowering effect was seen in all devices tested. In our experience, elevated switching threshold is a result of “subsaturation” of the electrolyte with metal ions. We believe that the reduced amount of mobile copper in the doped plasma-grown oxides is due to the two-phase nature of this electrolyte, in which some of the copper is in the form of an oxide and, therefore, cannot readily take part in the electrodepostion process. It is likely that the first voltage sweep injects additional Cu from the anode to increase the unbound copper concentration, thereby lowering the off resistance and write threshold for the following sweeps. However, since this injection of Cu by the applied field will only increase the concentration of metal along the volume of the filamentary pathway (where the field is highest) and not in the rest of the film, the metallic connection formed will be susceptible to dissolution by metal diffusion into the surrounding electrolyte due to the high copper metal concentration gradient. This was confirmed by resistance-time measurements which showed an upward drift in the on-state resistance, which returns to the high resistance off state after a few hours at room temperature and after only a few minutes at 70 C. The erase transition threshold occurs at 0.2 to 0.3 V in this device, with full erase occurring beyond 0.5 V. The sudden increase in resistance at the erase transition threshold is due to the current-induced breaking and voltage-induced electrochemical oxidation of the conducting pathway, leading to a device resistance that is dominated by the high resistance of a portion of now unbridged electrolyte. The subsequent resistance rise (and any steps therein) is a result of the removal by electrochemical oxidation of the remaining electrodeposited metal. Note that the drop in off state resistance for increasing voltage magnitude in either direction is due to the leakage current in the electrolyte. Fig. 7 shows the <Au: Define R-V?> (R-V) characteristic of a smaller Cu/plasma-grown electrolyte device with a 240-nm-diameter for a 0.5- m current limit. This shows many of the characteristics of the 1- m device but the write thresholds are all considerably higher (around 2 V initially and decreasing toward 1.7 V with increasing sweep number). The post-electrodeposition thresholds are also higher than before (close to 0.7 V), leading to higher on resistances as defined by the programming current relationship given earlier . We believe that these higher thresholds are due to considerably lower Cu diffusion into the oxide in the small area devices during illumination. The optical attenuation of the reduced intermetal dielectric (IMD) and the dimension hole in the nonconformal coverage of the copper in this reduced dimension feature will lead to less light reaching the interface of the metal and the oxide and this will reduce copper photodiffusion. The lower copper diffusion, in addition to the oxidation of the Cu that does enter the oxide, results in the higher switching threshold. The lower Cu concentration does little to alter the erase initiation threshold, which occurs between 0.2 and 0.4 V and the maximum off resistance of over (the limit of the measurement system) is achieved beyond 0.5 V as before. This is expected as the erase initiation occurs due to the current-induced breaking of the conducting filament and the nature of the electrolyte has little effect on this. Even though the devices based on plasma-grown oxide were not ideal as switching elements, they showed the essential characteristics of PMC structures, with a large decrease in resistance for a low write threshold and programming current and a return to a high resistance state at a low negative bias. However, if they are to be used in embedded memory or as switching elements in reconfigurable logic, more work is required to improve their stability. WO WO KOZICKI et al.: LOW-POWER NONVOLATILE SWITCHING ELEMENT BASED ON COPPER-TUNGSTEN OXIDE SOLID ELECTROLYTE 7 B. Devices Based on Deposited Oxide IE W EE eb P Ve ro o rs f ion The substrates chosen for the deposited oxide-based device experiments were similar to those described in the previous section. In this case, after the dielectric was etched to form the via layers holes and expose the underlying tungsten electrode, 50 or 100 nm thick were deposited under high vacuum using an electron beam evaporator. Following this, 25 nm of Cu was without breaking the deposition system evaporated on the vacuum and the Cu photodiffused into the oxide for 14 min using the technique described earlier. Once again, the excess Cu was not removed from the surface. To complete the device, an additional 50 nm of Cu was deposited and the electrolyte/top electrode stack was patterned using optical lithography and lift off. A schematic of the device structure is shown in Fig. 5(b). The electrical characterization using the set-up described earlier was performed at room temperature and at a range of elevated temperatures using a temperature controlled chuck on the probe-station. Fig. 8(a) gives a typical current-voltage plot of a 240-nm-didevice, comprising a 50-nm deameter posited oxide. This was programmed at 1 A and exhibited a write threshold of approximately 0.4 V and a lower electrodeposition threshold of 0.1 V, at which point the negative-going sweep drops below the compliance current level. The device returns to its high resistance state at 0.2 V. Fig. 8(b) shows the resistance-voltage plot for the same device, showing an avand an on resistance erage off resistance in the order of 2 . Note that the on resistance is dependent on proof 100 gramming current as before with the on resistance equal to the lower threshold (0.1 V in this case) divided by the programming current limit. Since this lower threshold for electrodeposition is quite invariant, the on resistance is also highly reproducible, typically to within 5% of the mean value for a given programming current. These results are similar in form to those obtained with the plasma-grown oxide except that the write threshold is not only lower for the much thicker deposited oxide but it is also much more consistent, largely devoid of the higher initial threshold effect; i.e., subsequent sweeps result in a switching threshold that is very close to the first sweep value, with a median of 0.4 V and a standard deviation of approximately 50 mV. In addition, the strong dependence of the write threshold on device diameter that was seen in the grown oxide devices is not present in the deposited oxide variants. This suggests that the deposited oxide in both the smaller and larger diameter devices is equally saturated with “free” metal following the photodissolution step, which was evidently not the case in the grown material. Note that the lower off resistance, which has an average for the 240-nm devices, is also indicative of value around 2 higher mobile metal content than the grown oxide devices. This is consistent with our material characterization work which suggested that the photodiffused copper formed a bronze with the deposited tungsten oxide and, therefore, would be unbound and mobile. Fig. 9 is the resistance-voltage behavior of a 5- m-diameter deposited oxide device (100-nm-thick ) at operating temperatures of 25 C, 70 C, and 135 C for a 10- A programming current. The device switched from the high resistance range to 0 Fig. 8. Electrical behavior of a 240-nm-diameter Cu WO based device formed from a 50-nm-thick oxide. (a) Current–voltage plot of a device programmed at 1 A. (b) Resistance voltage plot for the same device. around 10 at all temperatures, with a lower electrodeposition threshold of 0.1 V as before. The 10 lower on resistance is a direct result of the 10 higher programming current than the case shown in Fig. 8 (i.e., in this case on resistance is given ). The apparent slight rise in write by threshold with temperature is probably due to the increase in the probe contact resistance due to thermal expansion. The most significant difference with temperature is the erase threshold, which clearly drops from the room temperature initiation value of 0.2 V as the temperature increases. This is consistent with PMC device operation as the increased thermal energy reduces the current required to break the conducting link. Fig. 10 gives multiple sweep resistance-voltage plots of a 240-nm-diameter ) for a 100- A programming curdevice (100-nm-thick rent at 135 C. The 100- A programming current gives an on , as expected . resistance close to 1 With the exception of the first sweep, device operation is consistent with the larger device of Fig. 9 at 135 C, with a write threshold close to 0.4 V, a lower threshold of 0.1 V, and an erase IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 IE W EE eb P Ve ro o rs f ion 8 Fig. 9. Resistance-voltage plots at different operating temperatures of 5-m -diameter devices formed from 100 nm of evaporated and photo-doped with Cu. WO Fig. 10. Resistance-voltage plots at 135 C for a 240-nm-diameter device comphotodoped with Cu. prising 100 nm of evaporated WO initiation threshold of 0.1 V. The apparent higher threshold of 0.6 V for the first sweep is once again thought to be due to probe resistance/thermal expansion effects as it was not repeatable (and, therefore, was not the same mechanism as that seen in the grown oxide devices). It is important to note that in the results shown in Figs. 8 to 10, the on resistance is independent of device diameter and electrolyte thickness. Insensitivity to device diameter is expected when switching is due to the formation of a filamentary conducting pathway which is considerably smaller in width than the electrolyte (device) diameter, The lack of sensitivity to electrolyte thickness is due to the fact that the on resistance depends on electrodeposition threshold and programming current limit, neither of which are influenced by device geometry. The on state in the above devices is also quite ohmic, which suggests that the filament is indeed metallic in nature, with good (low barrier) contact to the tungsten and copper electrodes. WO Fig. 11. Data retention behavior of devices fabricated using Cu-doped films programmed at 10 A and 1 V and probed using a 100 mV measurement signal. Another means for determining the suitability of the deposited oxide for nonvolatile switching applications was measurement of retention of the devices over a range of temperatures. For this test, the devices were programmed using a 10- A current and the resistance measured using a 100 mV dc probe signal at logarithmic intervals of time. The test was conducted at room temperature and also at 70 C, and 135 C. The results revealed that devices fabricated using photodiffused films showed good state retention even at Cu-saturated 135 C, as shown in Fig. 11. This particular test was terminated at around 12 hours but in all the samples evaluated, there was no indication of an upward drift in resistance. Indeed, in most cases there was actually a small decrease in resistance for the test conditions used, even though the high off state was maintained. Finally, the cycling endurance of the device structure was assessed. A 500-nm-diameter device with electrolyte was a 100-nm-thick Cu-photodoped deposited repeatedly written and erased at 135 C using a programmable pulse generator. The pulse generator could be programmed to provide variable numbers of 100 ms 1 V (write) followed by 100 ms 1.5 V (erase) pulse sequences. The voltages and pulse lengths used were considerably in excess of what was required to program the devices to ensure that they were fully written and erased during each cycle, as confirmed by single pulse measurements and digital oscilloscope screen captures taken during cycling. A series resistor was used to limit the on current through the device to approximately 5 A. Fig. 12 gives quasi-static resistance-voltage plots for voltage sweeps from 1 V to 1 V to 1 V with a current limit of 5 A. The plots were taken after 1, 10, 100, 1000, and 10 000 cycles, as indicated in the figure. All curves below 10 000 cycles show good repeatability of the off resistance, write threshold, and on resistance, all to within (averaged 10% of the approximate median values of 400 , respectively. The erase porover 0–0.3 V), 0.4 V, and 20 tions of the individual curves (from 0 to 1 V) appear erratic but the devices are actually in the off state following this part KOZICKI et al.: LOW-POWER NONVOLATILE SWITCHING ELEMENT BASED ON COPPER-TUNGSTEN OXIDE SOLID ELECTROLYTE WO the Cu is not bound and is, therefore, more able to participate in the switching process. We assume that this difference is a result of the plasma-grown oxide being somewhat substoichiometric compared to the stable deposited oxide but additional analysis work is required to establish the exact mechanism of copper incorporation. The devices based on deposited oxide had reproducible electrical characteristics, with critical parameters such as off resistance, write threshold, and on resistance being within 10% or less of their median values. The control of on resistance by write current limit was evident in these devices to 1 were demonstrated, and on resistances from 100 suggesting that multi-level storage (two or more bits per cell) might be possible with these devices. Retention and endurance seemed promising for these early prototypes. Of course, more material optimization and electrical characterization is required to improve write and erase speed and other important factors such as endurance before these elements can -range devices show promise as be utilized. However, these an easily integrated switching element as they can use existing materials and compatible processing and involve as little as one additional masking step over standard advanced CMOS. IE W EE eb P Ve ro o rs f ion + Fig. 12. Effects of write-erase cycling (each cycle was 100 ms at 1 V, 100 ms at 1.5 V) at 135 C of a 500-nm-diameter device comprising 100 nm of photodoped with Cu. The device resistance-voltage plot was evaporated taken after various numbers of cycles as indicated in the plot. 0 9 of the sweep and can be subsequently fast cycled. The plot corresponding to 10 000 cycles shows a much lower off resistance and even though the device switches at the expected threshold on resistance, it cannot be cycled further. and attains the 20 The failure mechanism that causes the “stuck bit” condition is unknown but is currently under investigation. V. CONCLUSION To conclude, we have demonstrated the feasibility of fabricating solid-state switching elements using tungsten metal, a solid electrolyte based on tungsten oxide with photodiffused copper, and a copper electrode. These devices exhibit a high off resistance and can be switched at low voltage to an on resistance state that is independent of device geometry (diameter and electrolyte thickness) but is strongly governed by the write current. The device can be returned to its high resistance state by the application of a small reverse bias. The electrical characteristics, and in particular those of devices comprising deposited tungsten oxide, are extremely similar to Programmable Metallization Cell devices based on chalcogenide electrolytes. We can, therefore, conclude that the same low energy electrochemical processes that result in the formation and removal of a filamentary conducting pathway in the chalcogenide devices are also responsible for the switching in the copper-doped oxide devices. The exact composition of this conducting pathway is currently unknown but the experimental evidence, which includes the ohmic nature of the on state, suggests that it is the result of the localized presence of Cu metal within the oxide. Electrical characterization has revealed that devices consisting of plasma-grown oxides have a decreasing initial threshold voltage and poor retention, whereas devices based on deposited oxide posses a stable switching threshold and good . retention, even at elevated operating temperature This difference in behavior was attributed to the observation that the copper tends to oxidize in the plasma-grown oxide and is, therefore, not in a mobile state whereas the copper-doped deposited oxide is “bronze-like,” comprising a structure in which ACKNOWLEDGMENT The authors would also like to thank C. Poweleit for preparation of the Raman spectra and T. Karcher for generation of the XPS results. The authors gratefully acknowledge the help of the staff and use of the facilities within the Center for Solid State Electronics Research and the Center for Solid State Science at Arizona State University. REFERENCES [1] G. Müller, T. Happ, M. Kund, G. Y. Lee, N. Nagel, and R. Sezi, “Status and outlook of emerging nonvolatile memory technologies,” in IEDM Tech. Dig., 2004, pp. 567–570. [2] T. Sakamoto, H. Sunamura, H. Kawaura, T. Hasegawa, T. Nakayama, and M. Aono, “Nanometer-scale switches using copper sulfide,” Appl. Phys. Lett., vol. 82, pp. 3032–3034, 2003. [3] S. Kaeriyama, T. Sakamoto, H. Sunamura, M. Mizuno, H. Kawaura, T. Hasegawa, K. Terabe, T. Nakayama, and M. Aono, “A nonvolatile programmable solid-electrolyte nanometer switch,” IEEE J. Solid-State Circuits, vol. 40, no. 1, pp. 168–176, Jan 2005. [4] M. N. Kozicki, C. Gopalan, M. Balakrishnan, M. Park, and M. Mitkova, “Non-volatile memory based on solid electrolytes,” in Proc. 2004 NonVolatile Memory Technology Symp. (NVMTS), 2004, pp. 10–17. [5] M. N. Kozicki, M. Park, and M. Mitkova, “Nanoscale memory elements based on solid-state electrolytes,” IEEE Trans. Nanotechnol., vol. 4, no. 3, pp. 331–338, May 2005. [6] M. N. Kozicki, M. Mitkova, J. Zhu, and M. Park, “Nanoscale phase separation in Ag–Ge–See glasses,” Microelectron. Eng., vol. 63, pp. 155–159, 2002. [7] M. Mitkova, M. N. Kozicki, H. C. Kim, and T. L. Alford, “Thermal and photodiffusion of Ag in S-Rich Ge-S amorphous films,” Thin Solid Films, vol. 449, pp. 248–253, 2004. [8] G. Dearnaley, G. Stoneham, and D. V. Morgan, “Electrical phenomenon in amorphous oxide films,” Rep. Prog. Phys., vol. 33, pp. 1129–1191, 1970. [9] R. E. Thurstans and D. P. Oxley, “The electroformed metal-insulatormetal structure: a comprehensive model,” J. Phys. D: Appl. Phys., vol. 35, pp. 802–809, 2002. [10] I. G. Baek, M. S. Lee, S. Seo, M. J. Lee, D. H. Seo, D.-S. Suh, J. C. Park, S. O. Park, H. S. Kim, I. K. Yoo, U.-I. Chung, and J. T. Moon, “Highly scalable non-volatile resistive memory using simple binary oxide driven by asymmetric unipolar voltage pulses,” in IEDM Tech. Dig., 2004. 10 IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 Michael N. Kozicki (M’84) was born near Edinburgh, Scotland, and received the B.Sc. and Ph.D. degrees from the University of Edinburgh in 1980 and 1985, respectively. He was with Hughes Microelectronics Ltd. (now Raytheon Systems Ltd.) before joining Arizona State University (ASU) in 1985, where he is currently Professor of Electrical Engineering. He i holds several key patents in integrated ionics, including those involving the storage of data by electrodeposition in solid electrolytes. He is also a founder of Axon Technologies Corp., an ASU spin-out company formed to develop and license solidstate ionic technologies, and has served as Director of ASU’s Center for Solid State Electronics Research. Chakravarthy Gopalan (S’06 <Au: Please confirm year?>) received the B.S. degree in electronics and communication engineering from Madurai Kamaraj University, Madurai, India, in 1999, and the Ph.D. degree in electrical engineering from Arizona State University, Tempe, in 2005. He is currently working at Spansion, <Location?; please update affiliations footnote on p.1 if necessary>. His current research interests include passive array design and fabrication of programmable metallization cell memory devices as well as flash memory IE W EE eb P Ve ro o rs f ion [11] W. C. West, “Electrically erasable non-volatile memory via electrochemical deposition of multifractal aggregates,” Ph.D. thesis, Arizona State Univ., Tempe, May 1998. [12] E. Ozkan, S.-H. Lee, C. E. Tracy, J. R. Pitts, and S. K. Deb, “Comparison of electrochromic amorphous and crystalline tungsten oxide films,” Solar Energy Mater. Solar Cells, vol. 79, pp. 439–448, 2003. [13] C. O. Avellaneda and L. O. S. Bulhoes, “Intercalation in WO(3) and WO3: Li films,” Solid State Ion, vol. 165, pp. 59–64, 2003. [14] A. Antonaia, M. C. Santoro, G. Fameli, and T. Polichetti, “Transport mechanism and IR structural characterization of evaporated amorphous WO films,” Thin Solid Films, vol. 426, pp. 281–287, 2003. [15] K. Miyake, H. Kaneko, M. Sano, and N. Suedomi, “Physical and electrochromic properties of the amorphous and crystalline tungsten oxide thick films prepared under reducing atmosphere,” J. Appl. Phys., vol. 55, pp. 2747–2753, 1984. [16] M. Gillet, K. Aguir, C. Lemire, E. Gillet, and K. Schierbaum, “The structure and electrical conductivity of vacuum-annealed WO thin films,” Thin Solid Films, vol. 467, pp. 239–246, 2004. [17] C. Bechinger, S. Herminghaus, and P. Leiderer, “Photoinduced doping of thin amorphous WO films,” Thin Solid Films, vol. 239, pp. 156–160, 1994. [18] R. C. Agrawal, M. L. Verma, and R. K. Gupta, “Electrical and electrochemical properties of a new silver tungstate glass system: x[0:75AgI : o:25AgCl] : (1 x)[Ag O : WO ],” Solid State Ionics, vol. 171, pp. 199–205, 2004. [19] L. Chernyak, K. Gartsman, D. Cahen, and O. Stafsudd, “Electronic effects of ion mobility in semiconductors: semionic behaviour of CuInSe ,” J. Phys. Chem. Sol., vol. 56, pp. 1165–1195, 1995. [20] B. W. Faughnan, R. S. Grandall, and P. M. Heyman, “Electrochromism in WO amorphous films,” RCA Rev., vol. 36, pp. 177–197, 1975. [21] K. Gesheva, A. Szekeres, and T. Ivanova, “Optical properties of chemical vapor deposited thin films of molybdenum and tungsten based metal oxides,” Sol. Energy Mater. Sol. Cells, vol. 76, pp. 563–576, 2003. [22] S.-H. Lee, H. M. Cheong, C. E. Tracy, A. Mascarenhas, D. K. Benson, and S. K. Deb, “Raman spectroscopic studies of electrochromic aWO ,” Elelctrochim. Acta, vol. 44, pp. 3111–3115, 1999. [23] P. Delichere, P. Falaras, M. Froment, A. Hugot-Le Goff, and B. Aguis, “Electrochromism in anodic WO films 1: preparation and physicochemical properties of films in the virgin and colored states,” Thin Solid Films, vol. 161, pp. 35–46, 1988. [24] H. Rosen, E. M. Engler, T. C. Strand, V. Y. Lee, and D. Bethune, “Raman study of lattice modes in the high-critical-temperature superconductor Y-Ba-Cu-O,” Phys. Rev B, vol. 36, pp. 726–728, 1987. [25] S. H. Baeck, T. F. Jaramillo, C. Brändli, and E. W. McFarland, “Combinatorial electrochemical synthesis and characterization of tungstenbased mixed-metal oxides,” J. Comb. Chem., vol. 4, pp. 563–568, 2002. [26] M. Sun, N. Xu, Y. W. Cao, J. N. Yao, and E. G. Wang, “A nanocrystalline tungsten oxide thin film: preparation, microstructure, and photochromic behavior,” J. Mater. Res., vol. 15, pp. 927–933, 2000. [27] T. Nanba, M. Ishikawa, Y. Sakai, and Y. Miura, “Changes in atomic and electronic structures of amorphous WO films due to electrochemical ion insertion,” Thin Solid Films, vol. 445, pp. 175–181, 2003. 0 solutions. Murali Balakrishnan (S’) received the B.S. degree in electronics and communication engineering from the University of Madras, Chennai, India, in 1999, and the M.S. degree in electrical engineering from Arizona State University, Tempe, in 2003. He is currently working toward the Ph.D. degree at Arizona State University. His current research interests include fabrication and characterization of programmable metallization cell memory devices. Maria Mitkova received the M.S. and Ph.D. degrees from the Technological University, Sofia, Bulgaria, in 1970 and 1976, respectively. She was Professor at the Technological University and Bulgarian Academy of Sciences and has worked in the R&D of Ovonic memory devices in the Institute for Microelectronics. Since 1997, she has worked in the USA at the University of Cincinnati, Cincinnati, OH, and Arizona State University, Tempe. Her interests are in the field of amorphous semiconductors, their characterization and application as optical and electronic memory media. She has specialized in a number of Ag-containing chalcogenide systems. IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 1 A Low-Power Nonvolatile Switching Element Based on Copper-Tungsten Oxide Solid Electrolyte Michael N. Kozicki, Member, IEEE, Chakravarthy Gopalan, Student Member, IEEE, Muralikrishnan Balakrishnan, Student Member, IEEE, and Maria Mitkova Index Terms—Copper electrodeposition, nonvolatile memory devices, Raman spectroscopy, resistance change, solid electrolyte, tungsten oxide, X-ray photoelectron spectroscopy. I. INTRODUCTION T by way of electrodeposition at low voltage and current reduces the resistance of the electrolyte by several orders of magnitude. Our earlier work in this area concentrated on two-terminal devices, with active area as small as 40 nm, based on a silver doped chalcogenide glass solid electrolyte film sandwiched between a silver anode and an inert cathode [5]. However, we have recently added electrolytes consisting of transition metal oxides combined with copper to our studies and it is this work that we describe in this paper. is a particularly desirable base Tungsten trioxide material for a solid electrolyte as it is compatible with back-end-of-line (BEOL) processing in CMOS integrated circuits where tungsten metal already plays a significant role. Tungsten oxide formation techniques are compatible with semiconductor processing and include wet chemical and plasma oxidation, and physical vapor deposition. As we have demonstrated, coupling this simple oxide with copper leads to a solid electrolyte with desirable qualities for electrochemical device operation. Copper is a particularly appropriate mobile ion choice in the context of integration into CMOS processes as many high performance ICs already use copper, along with tungsten, in the upper levels of metal. Fig. 1 illustrates how the integration of such a switching element could be achieved. In this scheme, selected tungsten plugs which are normally used to form the connections between one level of metal and another are used as the lower electrode in the devices and the copper interconnect is the upper electrode. The solid electrolyte is placed between the two electrodes and is switched between high and low resistance states by the application of appropriate voltages on the electrodes. Note that during processing, only one additional mask is required over the normal CMOS logic to define which plugs have switching devices and which are allowed to remain as simple connections between interconnect layers and so the approach is inherently low in incremental cost. This paper focuses on the nature of copper-doped tungsten oxide and on devices comprising this material. A comparison is made between the quasi-static electrical characteristics of devices based on plasma-grown and deposited oxides, the copper being added by photodissolution in both cases. IE Pr EE int P Ve ro o rs f ion Abstract—We describe the materials aspects and electrical (Cu WO3 ) Cu switching elements. characteristics of W These materials are compatible with back-end-of-line processing in CMOS integrated circuits where both tungsten and copper already play a significant role. Devices based on Cu WO3 solid electrolytes formed by photodiffusion of copper into tungsten oxide switch via the electrochemical formation of a conducting filament within the high resistance electrolyte film. They are able to switch reversibly between widely spaced nonvolatile resistance states at low voltage ( 1 V) and current ( 10 A). Electrical characterization revealed that devices consisting of plasma-grown oxides have a variable initial threshold voltage and poor retention, whereas devices based on deposited oxide exhibit a stable switching threshold and good retention, even at elevated operating temperature ( 125 C). This difference in behavior was attributed to the observation that the copper tends to oxidize in the plasma-grown oxide whereas the copper in the deposited oxide exists in an unbound state and is, therefore, more able to participate in the switching process. HERE is a considerable driving force within the semiconductor industry to create scalable elements that can switch between widely spaced nonvolatile resistance states at low power. Such elements could find widespread application in next generation memory and reconfigurable logic [1]–[3]. Whereas there are several new technologies that show promise in this respect, they typically lack complete scalability, i.e., they have a physical size, programming voltage, or programming current that is excessive for high density systems at the 65-nm node and beyond. Many are also difficult to integrate due to the complexities associated with the additional processing steps and masking levels required and, therefore, will be expensive to add to highly scaled CMOS. One potential approach to this problem involves switching elements, known as Programmable Metallization Cell (PMC) devices, which utilize the reduction of nanoscale quantities of metal ions in solid electrolyte films [4]. The formation of a robust but reversible conducting pathway II. DEVICES BASED ON SOLID ELECTROLYTES AND OXIDES Manuscript received March 30, 2006; revised May 1, 2006. This work was supported by Axon Technologies Corporation. The review of this paper was arranged by Associate Editor B. G. Park. The authors are with the Center for Solid State Electronics Research, Arizona State University, Tempe, AZ 85287-6206 USA (michael.kozicki@asu.edu). Color versions of Figs. 2–12 are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/TNANO.2006.880407 In our earlier work [4], [5], we presented the formation and operation of PMC devices composed of a silver doped chalcogenide glass solid electrolyte film (e.g., Ag–Ge–Se or Ag–Ge–Se) sandwiched between a silver anode and an inert cathode. For an applied bias in excess of a few hundred mV, an electron current from the cathode reduces silver ions in the 1536-125X/$20.00 © 2006 IEEE 2 IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 resistivity which were originally attributed to homogeneous modifications or an unspecified breakdown mechanism. These explanations were dismissed by Dearnaley and colleagues [8], who developed a filamentary model for the changes in resistance. It was claimed that since the anode insulator interface is not completely smooth, there will be places where the electric field is locally high, resulting in the formation of a conductive filament by field induced migration of material from the anode into the oxide. The metallic filament propagates through the insulator and thereby reduces the resistance of the oxide layer. Subsequent switching in this “electroformed” structure was thought to be via electrically stimulated rupture and healing of the filamentary connection at locations along its length. Thurstans and Oxley [9] proposed a model which took other characteristics of the electroformed structure, in particular the observed negative differential resistance, into account. Their model features a conduction mechanism that depends on trap-controlled thermally activated tunneling between isolated segments of a ruptured metallic pathway. Recently, Baek et al. [10] used such models to describe the unipolar switching characteristics of their transition metal oxide-based devices. In these structures, nickel oxide is switched to a low resistance state using a relatively high voltage/low current “set” pulse and returned to a high resistance state using a low voltage/high current “reset” pulse of the same polarity. Although the effect is attributed to charge trapping mechanisms in the electroformed oxide, it is perhaps more likely to be due to the rupture and healing of a metallic pathway. We have observed similar unipolar programming in PMC devices in which a metallic filament is formed by the electrodeposition of the anode metal in the electrolyte at a particularly high critical voltage but can be disrupted using a current pulse of the same polarity [11]. In this case, the erase (reset) current must be in excess of the current used to write (set) the device so that the current carrying capacity of the filament is exceeded and it is thereby severed, however, the erase voltage must be below that required for electrodepostion of the metal so that the break does not spontaneously re-close. Note that for the devices described in [4], the voltage threshold for electrodeposition is sufficiently low that a reverse bias is required to erase the device; a forward current in excess of the original programming current limit leads to further electrodeposition and the device resistance falls to a lower stable value rather than rising to the off state. The consideration of tungsten oxide as a base glass for the electrolyte in PMC devices led to a number of obvious questions in the context of the above discussion. The main issue was whether the devices would behave like the chalcogenide-based variants or more like the metal-insulator-metal structures described by Oxley. To address these issues, we engaged in a program of material analysis and device characterization, described in the following sections. IE Pr EE int P Ve ro o rs f ion electrolyte while an equivalent number are injected from the anode via oxidation of the silver. A metal-rich pathway forms in the electrolyte film via this electrodeposition process. The level and duration of the ion current above the electrodeposition threshold determine the quantity of Ag electrodeposited and, hence, the resistance of the pathway; the resistance of the structure falls by many orders of magnitude even for applied currents as low as a few microamperes. The low resistance electrodeposit is electrically neutral and stable, leading to retention in excess of 10 years [1], [4], however, applying a bias with opposite polarity will break the link by a combination of current-induced rupture and electrochemical oxidation of the electrodeposit. Note that when a reverse bias is applied, the break cannot “heal” by electrodeposition as there is a net removal of metal from the electrolyte in a process which is opposite to that which formed the metal filament in the first place. The reverse ion current flows until the previously deposited Ag has been oxidized and returned to the silver electrode. Thus, the resistance increases again until the high resistivity of the solid electrolyte is achieved. The oxidation/reduction process is fast (tens of nanoseconds or less for write and erase), consumes little power (in the microwatt range), and can be cycled well beyond conventional nonvolatile memory technology [4], [5]. The electrical characteristics of PMC devices comprising chalcogenide-based solid electrolytes result from the unique nanostructure of the electrolyte material. We previously established that the dissolution of Ag into Se- or S-rich base glasses produces a ternary that is a combination of a dispersed superionic nano-crystalline Ag-rich phase within a glassy insulating Ge-rich phase [6], [7]. The Ag ion-containing regions are typically less than 10 nm in diameter and are separated by a few nanometers of the insulating phase [5], [6]. It is this particular nanostructure that allows the films to exhibit a resistivity of 100 cm or more (which yields a high device off resistance) while retaining good ion availability and mobility. The accessibility of mobile Ag ions for electrodeposition throughout the electrolyte allows the devices to switch rapidly as the growing electrodeposit will always have a local source of ionic metal to feed its formation. The reduced ions are replaced via the ion current from the anode and so the electrolyte does not become depleted of ions following switching and contains “excess” metal within the electrodeposited pathway. Indeed, it is this addition of metal into the already metal ion-rich but nano phase-separated electrolyte that leads to both the reduction in resistance of the structure and the nonvolatility of the on state. To complement our work on chalcogenide-based PMC devices, we have recently examined elements that are more closely related to the materials present in modern integrated circuits. Tungsten metal and its oxide and Cu are obvious choices in this respect. There has been widespread interest in the use of a variety of oxides in switching/memory applications for a number of years. Several models exist that describe the electrical phenomenon seen in thin oxide films sandwiched between two metal layers but as yet no complete or consistent model describes all the observations. Films from tens to thousands of nanometer thick when subjected to certain critical electric fields have been observed to undergo permanent large reductions in III. TUNGSTEN OXIDE CHARACTERISTICS The electrical properties of have been widely investigated because of the existence of electro- and photo-chromic effects in materials based on this oxide [12]–[14]. It is a wide is bandgap dielectric—the electronic gap of amorphous KOZICKI et al.: LOW-POWER NONVOLATILE SWITCHING ELEMENT BASED ON COPPER-TUNGSTEN OXIDE SOLID ELECTROLYTE be characterized. Note that in the fabrication of device structures (discussed later), a patterned tungsten electrode was used and the excess metal following photodiffusion was left in place on the surface of the electrolyte to form part of the anode top electrode. Characterization of these films was performed using Raman spectroscopy and X-Ray Photoelectron Spectroscopy (XPS). Raman spectra were performed in the micro Raman mode using a 514.5 nm Ar+ laser, 100 objective, with an accumulation time of 100 s and a power of 6 mW. The experimental curves were fitted using Labcalc software. X-ray Photoelectron Spectroscopy (XPS) was carried out with high-resolution XPS X-ray excitation and constant analyzer energy with (Mg 8 eV pass energy, take-off angle of 90 , 240 W power). The energy calibration of the spectrometer was performed using a gold plate fixed to the sample. Raman data related to plasma oxidized films are presented in Fig. 2. Since the films are predominantly amorphous, only breathing modes resulting from the short-range order appear. The intensity of Raman signals from the films is inherently weak because they are rather thin but the deconvolution of the results permits collection of meaningful information relating to the vibration of the W-O bond. This revealed a mode at 280 cm (bending mode), very weak modes at 815 and 930 cm and one at 998 cm (stretching film is illuminated with light, a mode at mode). After the , suggesting 450 cm appears that could be due to some reduction has occurred [22]. Illumination with light evidently causes some crystallization of the films, an effect that is manifested in the appearance of a peak at 719 cm , bond. Such a structure has also characteristic of the been observed by Delichere et al. [23] who relate it to a higher order of structural organization. When Cu is photodiffused into the films, modes at intermediate frequency appear (440 and 460 and cm ) that could be attributed to the vibrations of modes [24]. This is significant as combinatorial electrochemical synthesis [24] has demonstrated that the formation of Cu oxides is possible following introduction of Cu in the films. The XPS data of plasma oxidized films show that they are indeed influenced by previous exposure to light Fig. 3(a) and (b). Gaussian deconvolution of the W4f spectrum of the initial and films Fig. 3(a) shows well resolved doublet caused by spin-orbit coupling with binding energies of 35.85 and 37.94 eV, which correspond to a typical 6 state of W. However illumination with light causes a small shift of this douand blet towards lower energies Fig. 3(b). The show binding energies of 35.71 and 37.8 eV that could be an indication of some reduction of the oxide. We could not observe a well expressed simultaneously occurring W4f double peak of as was reported by Sun et al. [25] since the amount of the 5-valent W is small. We suggest that this is related to the fact that there is no water included in our films. However the shift is certainly an indication of the influence of light on our films. The introduction of Cu causes significant changes in the XPS spectra Fig. 4. The Gaussian deconvolution suggests the appearance of 4 binding energy values that could correspond to the 6 and 5 state of W, as expected. However, they are somewhat shifted from the standard data for these particular doublets. It is diffi- IE Pr EE int P Ve ro o rs f ion 3.25 eV [15]—but a large range of conductivity has been reported [16]. This is due to the fact that W can exist in different valence states when oxidized, thus forming oxides with different stoichiometry and, hence, with different electrical conductivity. Different valence states can also be created by illumination with UV light with energy exceeding 3.3 eV [17]. In can be altered by the introaddition, the conductivity of duction of metal ions and this is of particular interest in the context of the current work [18]. As discussed earlier in this paper, we have used silver as the mobile ionic species in the base glass (and as the oxidizable electrode) in chalcogenide-based PMC devices due to its excellent electrochemical qualities but copper is also a good candidate due to its high ionic mobility in a variety of materials, including oxides [19]. To combine metals and thereby form a stable solid electrolyte, we can with take advantage of the photosensitivity of the oxide. This effect more commonly manifests itself as photochromism which can be explained by a double charge-injection model [20], in which and electrons into a lattice forms a tunginjection of sten bronze structure , . Note that can in tungsten oxide hydrate or be partially reduced to in 12-tungstates by electrons occupying the empty orbital is typically in due to the electron localization effect. photochromic but can actually be a variety of mobile ions, . So, if the film is prepared using a dry including reaction and a very thin UV transparent Cu film is evaporated on top, illumination can cause the formation of a bronze. formation is concerned, although wet chemAs far as ical or plasma oxidation of tungsten metal are simple room temperature processes, the thickness of the layer formed is limited to a few nanometers by the diffusion of the oxidant species through the growing oxide. It is, therefore, necessary to use the evaporation of tungsten oxide to form thicker films. To grow thin tungsten oxide films on tungsten metal, we elected to use plasma techniques as we discovered that the growth of layers by wet chemical means was difficult to control and led to material with highly variable characteristics. In addition, the wet grown oxides also contain species that would possibly interfere with the incorporation of Cu. The surface of 100-nm-thick tungsten glow dismetal films was oxidized using a nitrous oxide charge at an RF power of 250 W and a substrate heater temperature of 300 C for 10 min. in a PlasmaQuest plasma processing system. The flow rate was 20 sccm with 1000 sccm of helium carrier gas. The precise thickness of the grown oxide was difficult to ascertain due to the roughening of the underlying tungsten caused by the growth process but it was estimated using ellipsometry and capacitance-voltage measurements to be in the order of 3 to 4 nm thick. To achieve metal diffusion, 25-nm-thick Cu layers were deposited by high vacuum evaporation and the metal was photodiffused into the oxide using the 405-nm broadband UV light source of a Karl Suss MJB 3 mask aligner for 14 . This wavelength was min at a power density of 4.5 [21], where chosen as it is near the absorption edge of the the light causes maximal change in the material. Blanket samples were fabricated for material characterization and the excess surface metal was removed using a short treatment with iron nisolution so that only the Cu in the oxide would trate 3 4 IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 IE Pr EE int P Ve ro o rs f ion Fig. 1. Schematic cross section of an integrated circuit, showing position of oxide switching elements on selected tungsten via plugs. Interconnect layers M1 and M2 are separated by IMD. Through vias provide M1 to M2 connectivity. WO Fig. 2. Raman analysis of formed by plasma oxididation of W (a) as synthesized; (b) after illumination with UV light; (c) photo-doped with Cu. cult to fully interpret these results but we believe that they could be related to distortions in the amorphous structure of the films and formation of clusters [26] in which some of the diffused Cu is oxidized and the clusters become part of the overall structure. This leads to the occurrence of bond overlapping and a shift in the binding energies. Overall, even though the signal-to-noise ratio is rather poor, the data related to Cu does indeed indicate its presence in an oxidized state within the tungsten oxide matrix (see inset in Fig. 4). We now turn to the deposited oxide samples. The evaporation source was high purity tungsten oxide (99.99%) in 3–12 mm sintered yellow-green pieces (Cerac, Inc.). Using high vacuum thermal evaporation, 50- or 100-nm thick films were deposited on the substrates described above. Following this, 25 nm of Cu without breaking the deposition was evaporated on the system vacuum. The Cu was then photodiffused into the oxide for 14 min and the excess metal was etched using the techniques described earlier. The analysis of these films revealed some simwas clearly ilarities with the plasma-grown oxides in that present as before in the as-deposited films, indicating the forma, but the way the copper was incorporated in the detion of posited oxide appeared different. Both Raman and XPS analysis of the films gave little or no evidence of the formation of Cu-ox. We ides after copper is photodiffused into the deposited WO Fig. 3. XPS spectra of formed by plasma oxididation of W (a) as synthesized; (b) after illumination with UV light. are currently investigating this in more detail but it would appear that whereas at least some of the photodiffused Cu becomes oxidized in the plasma-grown oxide to form a two phase (copper oxide/tungsten oxide) system, it exists in an unbound state in the deposited material. This could be due to the fact that the deposited oxides are in a highly stable state and do not readily react with the diffusing copper, whereas the presumably substoichiometric plasma-grown oxides are much more reactive. In any case, this dissimilarity is extremely significant in the context of the observed differences in the electrical characteristics 5 IE Pr EE int P Ve ro o rs f ion KOZICKI et al.: LOW-POWER NONVOLATILE SWITCHING ELEMENT BASED ON COPPER-TUNGSTEN OXIDE SOLID ELECTROLYTE WO Fig. 4. XPS spectra of formed by plasma oxididation of W after Cu diffusion in the oxide film; inset—XPS spectrum of Cu-containing species in the oxide film. of devices based on plasma-grown and deposited oxides, as described in the following sections. IV. DEVICE FABRICATION AND ELECTRICAL CHARACTERIZATION A. Devices Based on Plasma Grown Oxide Tungsten oxide films were incorporated in PMC device structures using oxygen plasma treatment of a tungsten metal bottom electrode as described in the previous section and with copper as the mobile species and oxidizable top electrode. Schematics of typical test device structures are shown in Fig. 5 (Fig. 5(a) represents devices based on grown oxide whereas Fig. 5(b) is the deposited oxide variants described in the following section). A 100-nm-thick tungsten layer was deposited by chemical vapor deposition on silicon dioxide grown on silicon substrates. The tungsten was covered with 100 nm of silicon dioxide by chemical vapor deposition and via (through) holes were defined in this dielectric using optical lithography and plasma etching. The exposed tungsten in the holes was oxidized using the plasma oxidation technique described in the previous section. A 25-nm-thick Cu metal layer was deposited using high vacuum thermal evaporation and photodiffused into the using a 405-nm broadband UV light source in a Karl Suss MJB . The excess 3 mask aligner at an intensity of 4.5 Cu metal was not removed to ensure the presence of a Cu-rich source on the electrolyte surface. A 50-nm-thick Cu metal layer was then deposited using high vacuum thermal evaporation and patterned using a lift off process to complete the oxidizable electrodes of the devices. Note that this device configuration was absolutely necessary to observe switching regardless of how the base glass for the electrolyte was formed; a lack of Cu in the top electrode results in no measurable switching activity. The Fig. 5. Cross-sectional schematics of oxide-based switching devices. The active device area is the tungsten-electrolyte-Cu stack with area defined by the dielectric. (a) Device based on grown diameter of the through hole in the oxide. (b) Device based on deposited oxide. SiO probe pad window in the to the tungsten layer was formed using another lithography step in conjunction with wet-chemical etching. The quasi-static characteristics of the test devices were obtained by connecting the device electrodes via tungsten probes held in micromanipulators in a probe station to a semiconductor parameter analyzer (Agilent 4155C). Voltage doublesweeps were carried out starting at maximum reverse bias (tungsten electrode positive, copper electrode negative), sweeping to an appropriate forward voltage (tungsten electrode negative, copper electrode positive), and sweeping back again to the reverse bias starting point. Fig. 6 shows a resistance-voltage plot from our earlier work on large (1– m diameter) devices with a Cu/plasma-grown electrolyte and Cu oxidizable electrode [4]. Three consecutive sweeps are shown from 0.75 V to 1.0 V to 0.75 V for a current limit of 0.5 A. The maximum off resistance and the transition for the first sweep is in excess of to the low resistance state occurs at 0.9 V at which point the current reaches its compliance limit. The current stays at the compliance until the negative-going sweep reaches 0.25 V, , more at which point the measured on resistance is 500 than four orders of magnitude lower than the off state. Note that the apparent rise in resistance between 0.25 and 1 V is an artifact of the current compliance control in the measurement instrument. The appearance of the lower threshold following switching is typical of all PMC devices and represents the minimum voltage required for electrodeposition once the process has been initiated and a metal-rich region has formed in the electrolyte. Electrodeposition within the oxide will continue as long as the voltage drop is greater than 0.25 V IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 IE Pr EE int P Ve ro o rs f ion 6 Fig. 6. Resistance-voltage plot of a 1.0-m-diameter device with Cu/plasmaelectrolyte and Cu oxidizable electrode. grown Fig. 7. Resistance-voltage plot of a 240-nm-diameter device with Cu/plasmaelectrolyte and Cu oxidizable electrode. grown for this physical configuration. Note that the decreasing resistance due to the electrodeposition effect increases the current flowing through the device until the compliance is reached. At this point, the voltage drop falls to the threshold and the electrodeposition stops. The on resistance is, therefore, defined by the electrodeposition threshold divided by the current limit . It is clear from the plot of Fig. 6 that the first sweep is different from the following two. The off resistance is apparently reduced by an order of magnitude following the first sweep and the switching threshold moves from 0.9 V in the first sweep toward 0.7 V in the latter cycles. The same switching threshold lowering effect was seen in all devices tested. In our experience, elevated switching threshold is a result of “subsaturation” of the electrolyte with metal ions. We believe that the reduced amount of mobile copper in the doped plasma-grown oxides is due to the two-phase nature of this electrolyte, in which some of the copper is in the form of an oxide and, therefore, cannot readily take part in the electrodepostion process. It is likely that the first voltage sweep injects additional Cu from the anode to increase the unbound copper concentration, thereby lowering the off resistance and write threshold for the following sweeps. However, since this injection of Cu by the applied field will only increase the concentration of metal along the volume of the filamentary pathway (where the field is highest) and not in the rest of the film, the metallic connection formed will be susceptible to dissolution by metal diffusion into the surrounding electrolyte due to the high copper metal concentration gradient. This was confirmed by resistance-time measurements which showed an upward drift in the on-state resistance, which returns to the high resistance off state after a few hours at room temperature and after only a few minutes at 70 C. The erase transition threshold occurs at 0.2 to 0.3 V in this device, with full erase occurring beyond 0.5 V. The sudden increase in resistance at the erase transition threshold is due to the current-induced breaking and voltage-induced electrochemical oxidation of the conducting pathway, leading to a device resistance that is dominated by the high resistance of a portion of now unbridged electrolyte. The subsequent resistance rise (and any steps therein) is a result of the removal by electrochemical oxidation of the remaining electrodeposited metal. Note that the drop in off state resistance for increasing voltage magnitude in either direction is due to the leakage current in the electrolyte. Fig. 7 shows the <Au: Define R-V?> (R-V) characteristic of a smaller Cu/plasma-grown electrolyte device with a 240-nm-diameter for a 0.5- m current limit. This shows many of the characteristics of the 1- m device but the write thresholds are all considerably higher (around 2 V initially and decreasing toward 1.7 V with increasing sweep number). The post-electrodeposition thresholds are also higher than before (close to 0.7 V), leading to higher on resistances as defined by the programming current relationship given earlier . We believe that these higher thresholds are due to considerably lower Cu diffusion into the oxide in the small area devices during illumination. The optical attenuation of the reduced intermetal dielectric (IMD) and the dimension hole in the nonconformal coverage of the copper in this reduced dimension feature will lead to less light reaching the interface of the metal and the oxide and this will reduce copper photodiffusion. The lower copper diffusion, in addition to the oxidation of the Cu that does enter the oxide, results in the higher switching threshold. The lower Cu concentration does little to alter the erase initiation threshold, which occurs between 0.2 and 0.4 V and the maximum off resistance of over (the limit of the measurement system) is achieved beyond 0.5 V as before. This is expected as the erase initiation occurs due to the current-induced breaking of the conducting filament and the nature of the electrolyte has little effect on this. Even though the devices based on plasma-grown oxide were not ideal as switching elements, they showed the essential characteristics of PMC structures, with a large decrease in resistance for a low write threshold and programming current and a return to a high resistance state at a low negative bias. However, if they are to be used in embedded memory or as switching elements in reconfigurable logic, more work is required to improve their stability. WO WO KOZICKI et al.: LOW-POWER NONVOLATILE SWITCHING ELEMENT BASED ON COPPER-TUNGSTEN OXIDE SOLID ELECTROLYTE 7 B. Devices Based on Deposited Oxide IE Pr EE int P Ve ro o rs f ion The substrates chosen for the deposited oxide-based device experiments were similar to those described in the previous section. In this case, after the dielectric was etched to form the via layers holes and expose the underlying tungsten electrode, 50 or 100 nm thick were deposited under high vacuum using an electron beam evaporator. Following this, 25 nm of Cu was without breaking the deposition system evaporated on the vacuum and the Cu photodiffused into the oxide for 14 min using the technique described earlier. Once again, the excess Cu was not removed from the surface. To complete the device, an additional 50 nm of Cu was deposited and the electrolyte/top electrode stack was patterned using optical lithography and lift off. A schematic of the device structure is shown in Fig. 5(b). The electrical characterization using the set-up described earlier was performed at room temperature and at a range of elevated temperatures using a temperature controlled chuck on the probe-station. Fig. 8(a) gives a typical current-voltage plot of a 240-nm-didevice, comprising a 50-nm deameter posited oxide. This was programmed at 1 A and exhibited a write threshold of approximately 0.4 V and a lower electrodeposition threshold of 0.1 V, at which point the negative-going sweep drops below the compliance current level. The device returns to its high resistance state at 0.2 V. Fig. 8(b) shows the resistance-voltage plot for the same device, showing an avand an on resistance erage off resistance in the order of 2 . Note that the on resistance is dependent on proof 100 gramming current as before with the on resistance equal to the lower threshold (0.1 V in this case) divided by the programming current limit. Since this lower threshold for electrodeposition is quite invariant, the on resistance is also highly reproducible, typically to within 5% of the mean value for a given programming current. These results are similar in form to those obtained with the plasma-grown oxide except that the write threshold is not only lower for the much thicker deposited oxide but it is also much more consistent, largely devoid of the higher initial threshold effect; i.e., subsequent sweeps result in a switching threshold that is very close to the first sweep value, with a median of 0.4 V and a standard deviation of approximately 50 mV. In addition, the strong dependence of the write threshold on device diameter that was seen in the grown oxide devices is not present in the deposited oxide variants. This suggests that the deposited oxide in both the smaller and larger diameter devices is equally saturated with “free” metal following the photodissolution step, which was evidently not the case in the grown material. Note that the lower off resistance, which has an average for the 240-nm devices, is also indicative of value around 2 higher mobile metal content than the grown oxide devices. This is consistent with our material characterization work which suggested that the photodiffused copper formed a bronze with the deposited tungsten oxide and, therefore, would be unbound and mobile. Fig. 9 is the resistance-voltage behavior of a 5- m-diameter deposited oxide device (100-nm-thick ) at operating temperatures of 25 C, 70 C, and 135 C for a 10- A programming current. The device switched from the high resistance range to 0 Fig. 8. Electrical behavior of a 240-nm-diameter Cu WO based device formed from a 50-nm-thick oxide. (a) Current–voltage plot of a device programmed at 1 A. (b) Resistance voltage plot for the same device. around 10 at all temperatures, with a lower electrodeposition threshold of 0.1 V as before. The 10 lower on resistance is a direct result of the 10 higher programming current than the case shown in Fig. 8 (i.e., in this case on resistance is given ). The apparent slight rise in write by threshold with temperature is probably due to the increase in the probe contact resistance due to thermal expansion. The most significant difference with temperature is the erase threshold, which clearly drops from the room temperature initiation value of 0.2 V as the temperature increases. This is consistent with PMC device operation as the increased thermal energy reduces the current required to break the conducting link. Fig. 10 gives multiple sweep resistance-voltage plots of a 240-nm-diameter ) for a 100- A programming curdevice (100-nm-thick rent at 135 C. The 100- A programming current gives an on , as expected . resistance close to 1 With the exception of the first sweep, device operation is consistent with the larger device of Fig. 9 at 135 C, with a write threshold close to 0.4 V, a lower threshold of 0.1 V, and an erase IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 IE Pr EE int P Ve ro o rs f ion 8 Fig. 9. Resistance-voltage plots at different operating temperatures of 5-m -diameter devices formed from 100 nm of evaporated and photo-doped with Cu. WO Fig. 10. Resistance-voltage plots at 135 C for a 240-nm-diameter device comphotodoped with Cu. prising 100 nm of evaporated WO initiation threshold of 0.1 V. The apparent higher threshold of 0.6 V for the first sweep is once again thought to be due to probe resistance/thermal expansion effects as it was not repeatable (and, therefore, was not the same mechanism as that seen in the grown oxide devices). It is important to note that in the results shown in Figs. 8 to 10, the on resistance is independent of device diameter and electrolyte thickness. Insensitivity to device diameter is expected when switching is due to the formation of a filamentary conducting pathway which is considerably smaller in width than the electrolyte (device) diameter, The lack of sensitivity to electrolyte thickness is due to the fact that the on resistance depends on electrodeposition threshold and programming current limit, neither of which are influenced by device geometry. The on state in the above devices is also quite ohmic, which suggests that the filament is indeed metallic in nature, with good (low barrier) contact to the tungsten and copper electrodes. WO Fig. 11. Data retention behavior of devices fabricated using Cu-doped films programmed at 10 A and 1 V and probed using a 100 mV measurement signal. Another means for determining the suitability of the deposited oxide for nonvolatile switching applications was measurement of retention of the devices over a range of temperatures. For this test, the devices were programmed using a 10- A current and the resistance measured using a 100 mV dc probe signal at logarithmic intervals of time. The test was conducted at room temperature and also at 70 C, and 135 C. The results revealed that devices fabricated using photodiffused films showed good state retention even at Cu-saturated 135 C, as shown in Fig. 11. This particular test was terminated at around 12 hours but in all the samples evaluated, there was no indication of an upward drift in resistance. Indeed, in most cases there was actually a small decrease in resistance for the test conditions used, even though the high off state was maintained. Finally, the cycling endurance of the device structure was assessed. A 500-nm-diameter device with electrolyte was a 100-nm-thick Cu-photodoped deposited repeatedly written and erased at 135 C using a programmable pulse generator. The pulse generator could be programmed to provide variable numbers of 100 ms 1 V (write) followed by 100 ms 1.5 V (erase) pulse sequences. The voltages and pulse lengths used were considerably in excess of what was required to program the devices to ensure that they were fully written and erased during each cycle, as confirmed by single pulse measurements and digital oscilloscope screen captures taken during cycling. A series resistor was used to limit the on current through the device to approximately 5 A. Fig. 12 gives quasi-static resistance-voltage plots for voltage sweeps from 1 V to 1 V to 1 V with a current limit of 5 A. The plots were taken after 1, 10, 100, 1000, and 10 000 cycles, as indicated in the figure. All curves below 10 000 cycles show good repeatability of the off resistance, write threshold, and on resistance, all to within (averaged 10% of the approximate median values of 400 , respectively. The erase porover 0–0.3 V), 0.4 V, and 20 tions of the individual curves (from 0 to 1 V) appear erratic but the devices are actually in the off state following this part KOZICKI et al.: LOW-POWER NONVOLATILE SWITCHING ELEMENT BASED ON COPPER-TUNGSTEN OXIDE SOLID ELECTROLYTE WO the Cu is not bound and is, therefore, more able to participate in the switching process. We assume that this difference is a result of the plasma-grown oxide being somewhat substoichiometric compared to the stable deposited oxide but additional analysis work is required to establish the exact mechanism of copper incorporation. The devices based on deposited oxide had reproducible electrical characteristics, with critical parameters such as off resistance, write threshold, and on resistance being within 10% or less of their median values. The control of on resistance by write current limit was evident in these devices to 1 were demonstrated, and on resistances from 100 suggesting that multi-level storage (two or more bits per cell) might be possible with these devices. Retention and endurance seemed promising for these early prototypes. Of course, more material optimization and electrical characterization is required to improve write and erase speed and other important factors such as endurance before these elements can -range devices show promise as be utilized. However, these an easily integrated switching element as they can use existing materials and compatible processing and involve as little as one additional masking step over standard advanced CMOS. IE Pr EE int P Ve ro o rs f ion + Fig. 12. Effects of write-erase cycling (each cycle was 100 ms at 1 V, 100 ms at 1.5 V) at 135 C of a 500-nm-diameter device comprising 100 nm of photodoped with Cu. The device resistance-voltage plot was evaporated taken after various numbers of cycles as indicated in the plot. 0 9 of the sweep and can be subsequently fast cycled. The plot corresponding to 10 000 cycles shows a much lower off resistance and even though the device switches at the expected threshold on resistance, it cannot be cycled further. and attains the 20 The failure mechanism that causes the “stuck bit” condition is unknown but is currently under investigation. V. CONCLUSION To conclude, we have demonstrated the feasibility of fabricating solid-state switching elements using tungsten metal, a solid electrolyte based on tungsten oxide with photodiffused copper, and a copper electrode. These devices exhibit a high off resistance and can be switched at low voltage to an on resistance state that is independent of device geometry (diameter and electrolyte thickness) but is strongly governed by the write current. The device can be returned to its high resistance state by the application of a small reverse bias. The electrical characteristics, and in particular those of devices comprising deposited tungsten oxide, are extremely similar to Programmable Metallization Cell devices based on chalcogenide electrolytes. We can, therefore, conclude that the same low energy electrochemical processes that result in the formation and removal of a filamentary conducting pathway in the chalcogenide devices are also responsible for the switching in the copper-doped oxide devices. The exact composition of this conducting pathway is currently unknown but the experimental evidence, which includes the ohmic nature of the on state, suggests that it is the result of the localized presence of Cu metal within the oxide. Electrical characterization has revealed that devices consisting of plasma-grown oxides have a decreasing initial threshold voltage and poor retention, whereas devices based on deposited oxide posses a stable switching threshold and good . retention, even at elevated operating temperature This difference in behavior was attributed to the observation that the copper tends to oxidize in the plasma-grown oxide and is, therefore, not in a mobile state whereas the copper-doped deposited oxide is “bronze-like,” comprising a structure in which ACKNOWLEDGMENT The authors would also like to thank C. Poweleit for preparation of the Raman spectra and T. Karcher for generation of the XPS results. The authors gratefully acknowledge the help of the staff and use of the facilities within the Center for Solid State Electronics Research and the Center for Solid State Science at Arizona State University. REFERENCES [1] G. Müller, T. Happ, M. Kund, G. Y. Lee, N. Nagel, and R. Sezi, “Status and outlook of emerging nonvolatile memory technologies,” in IEDM Tech. Dig., 2004, pp. 567–570. [2] T. Sakamoto, H. Sunamura, H. Kawaura, T. Hasegawa, T. Nakayama, and M. Aono, “Nanometer-scale switches using copper sulfide,” Appl. Phys. Lett., vol. 82, pp. 3032–3034, 2003. [3] S. Kaeriyama, T. Sakamoto, H. Sunamura, M. Mizuno, H. Kawaura, T. Hasegawa, K. Terabe, T. Nakayama, and M. Aono, “A nonvolatile programmable solid-electrolyte nanometer switch,” IEEE J. Solid-State Circuits, vol. 40, no. 1, pp. 168–176, Jan 2005. [4] M. N. Kozicki, C. Gopalan, M. Balakrishnan, M. Park, and M. Mitkova, “Non-volatile memory based on solid electrolytes,” in Proc. 2004 NonVolatile Memory Technology Symp. (NVMTS), 2004, pp. 10–17. [5] M. N. Kozicki, M. Park, and M. Mitkova, “Nanoscale memory elements based on solid-state electrolytes,” IEEE Trans. Nanotechnol., vol. 4, no. 3, pp. 331–338, May 2005. [6] M. N. Kozicki, M. Mitkova, J. Zhu, and M. Park, “Nanoscale phase separation in Ag–Ge–See glasses,” Microelectron. Eng., vol. 63, pp. 155–159, 2002. [7] M. Mitkova, M. N. Kozicki, H. C. Kim, and T. L. Alford, “Thermal and photodiffusion of Ag in S-Rich Ge-S amorphous films,” Thin Solid Films, vol. 449, pp. 248–253, 2004. [8] G. Dearnaley, G. Stoneham, and D. V. Morgan, “Electrical phenomenon in amorphous oxide films,” Rep. Prog. Phys., vol. 33, pp. 1129–1191, 1970. [9] R. E. Thurstans and D. P. Oxley, “The electroformed metal-insulatormetal structure: a comprehensive model,” J. Phys. D: Appl. Phys., vol. 35, pp. 802–809, 2002. [10] I. G. Baek, M. S. Lee, S. Seo, M. J. Lee, D. H. Seo, D.-S. Suh, J. C. Park, S. O. Park, H. S. Kim, I. K. Yoo, U.-I. Chung, and J. T. Moon, “Highly scalable non-volatile resistive memory using simple binary oxide driven by asymmetric unipolar voltage pulses,” in IEDM Tech. Dig., 2004. 10 IEEE TRANSACTIONS ON NANOTECHNOLOGY, VOL. 5, NO. 5, SEPTEMBER 2006 Michael N. Kozicki (M’84) was born near Edinburgh, Scotland, and received the B.Sc. and Ph.D. degrees from the University of Edinburgh in 1980 and 1985, respectively. He was with Hughes Microelectronics Ltd. (now Raytheon Systems Ltd.) before joining Arizona State University (ASU) in 1985, where he is currently Professor of Electrical Engineering. He i holds several key patents in integrated ionics, including those involving the storage of data by electrodeposition in solid electrolytes. He is also a founder of Axon Technologies Corp., an ASU spin-out company formed to develop and license solidstate ionic technologies, and has served as Director of ASU’s Center for Solid State Electronics Research. Chakravarthy Gopalan (S’06 <Au: Please confirm year?>) received the B.S. degree in electronics and communication engineering from Madurai Kamaraj University, Madurai, India, in 1999, and the Ph.D. degree in electrical engineering from Arizona State University, Tempe, in 2005. He is currently working at Spansion, <Location?; please update affiliations footnote on p.1 if necessary>. His current research interests include passive array design and fabrication of programmable metallization cell memory devices as well as flash memory IE Pr EE int P Ve ro o rs f ion [11] W. C. West, “Electrically erasable non-volatile memory via electrochemical deposition of multifractal aggregates,” Ph.D. thesis, Arizona State Univ., Tempe, May 1998. [12] E. Ozkan, S.-H. Lee, C. E. Tracy, J. R. Pitts, and S. K. Deb, “Comparison of electrochromic amorphous and crystalline tungsten oxide films,” Solar Energy Mater. Solar Cells, vol. 79, pp. 439–448, 2003. [13] C. O. Avellaneda and L. O. S. Bulhoes, “Intercalation in WO(3) and WO3: Li films,” Solid State Ion, vol. 165, pp. 59–64, 2003. [14] A. Antonaia, M. C. Santoro, G. Fameli, and T. Polichetti, “Transport mechanism and IR structural characterization of evaporated amorphous WO films,” Thin Solid Films, vol. 426, pp. 281–287, 2003. [15] K. Miyake, H. Kaneko, M. Sano, and N. Suedomi, “Physical and electrochromic properties of the amorphous and crystalline tungsten oxide thick films prepared under reducing atmosphere,” J. Appl. Phys., vol. 55, pp. 2747–2753, 1984. [16] M. Gillet, K. Aguir, C. Lemire, E. Gillet, and K. Schierbaum, “The structure and electrical conductivity of vacuum-annealed WO thin films,” Thin Solid Films, vol. 467, pp. 239–246, 2004. [17] C. Bechinger, S. Herminghaus, and P. Leiderer, “Photoinduced doping of thin amorphous WO films,” Thin Solid Films, vol. 239, pp. 156–160, 1994. [18] R. C. Agrawal, M. L. Verma, and R. K. Gupta, “Electrical and electrochemical properties of a new silver tungstate glass system: x[0:75AgI : o:25AgCl] : (1 x)[Ag O : WO ],” Solid State Ionics, vol. 171, pp. 199–205, 2004. [19] L. Chernyak, K. Gartsman, D. Cahen, and O. Stafsudd, “Electronic effects of ion mobility in semiconductors: semionic behaviour of CuInSe ,” J. Phys. Chem. Sol., vol. 56, pp. 1165–1195, 1995. [20] B. W. Faughnan, R. S. Grandall, and P. M. Heyman, “Electrochromism in WO amorphous films,” RCA Rev., vol. 36, pp. 177–197, 1975. [21] K. Gesheva, A. Szekeres, and T. Ivanova, “Optical properties of chemical vapor deposited thin films of molybdenum and tungsten based metal oxides,” Sol. Energy Mater. Sol. Cells, vol. 76, pp. 563–576, 2003. [22] S.-H. Lee, H. M. Cheong, C. E. Tracy, A. Mascarenhas, D. K. Benson, and S. K. Deb, “Raman spectroscopic studies of electrochromic aWO ,” Elelctrochim. Acta, vol. 44, pp. 3111–3115, 1999. [23] P. Delichere, P. Falaras, M. Froment, A. Hugot-Le Goff, and B. Aguis, “Electrochromism in anodic WO films 1: preparation and physicochemical properties of films in the virgin and colored states,” Thin Solid Films, vol. 161, pp. 35–46, 1988. [24] H. Rosen, E. M. Engler, T. C. Strand, V. Y. Lee, and D. Bethune, “Raman study of lattice modes in the high-critical-temperature superconductor Y-Ba-Cu-O,” Phys. Rev B, vol. 36, pp. 726–728, 1987. [25] S. H. Baeck, T. F. Jaramillo, C. Brändli, and E. W. McFarland, “Combinatorial electrochemical synthesis and characterization of tungstenbased mixed-metal oxides,” J. Comb. Chem., vol. 4, pp. 563–568, 2002. [26] M. Sun, N. Xu, Y. W. Cao, J. N. Yao, and E. G. Wang, “A nanocrystalline tungsten oxide thin film: preparation, microstructure, and photochromic behavior,” J. Mater. Res., vol. 15, pp. 927–933, 2000. [27] T. Nanba, M. Ishikawa, Y. Sakai, and Y. Miura, “Changes in atomic and electronic structures of amorphous WO films due to electrochemical ion insertion,” Thin Solid Films, vol. 445, pp. 175–181, 2003. 0 solutions. Murali Balakrishnan (S’) received the B.S. degree in electronics and communication engineering from the University of Madras, Chennai, India, in 1999, and the M.S. degree in electrical engineering from Arizona State University, Tempe, in 2003. He is currently working toward the Ph.D. degree at Arizona State University. His current research interests include fabrication and characterization of programmable metallization cell memory devices. Maria Mitkova received the M.S. and Ph.D. degrees from the Technological University, Sofia, Bulgaria, in 1970 and 1976, respectively. She was Professor at the Technological University and Bulgarian Academy of Sciences and has worked in the R&D of Ovonic memory devices in the Institute for Microelectronics. Since 1997, she has worked in the USA at the University of Cincinnati, Cincinnati, OH, and Arizona State University, Tempe. Her interests are in the field of amorphous semiconductors, their characterization and application as optical and electronic memory media. She has specialized in a number of Ag-containing chalcogenide systems.