Transition-metal-oxide-based resistance-change memories

Among the different materials used for realizing a new solid-state nonvolatile memory technology, transition-metal oxides are being pursued with increasing intensity to provide devices with high areal densities and low power consumption (for an example, see Reference [1]). One concept involves the use of the phenomenon of ferroelectricity occurring in numerous perovskite oxides. Here, switching the orientation of the ferroelectric polarization is used to modulate the charge in a memory cell structure [2]. Alternative memory cells comprising perovskite oxides can be realized for which an electrical stimulation in the form of voltage or current pulses establishes two or more defined resistive states of the memory device. Most of the oxides under investigation are insulators in the initial state. Hence, an electroforming step [a nondestructive (soft) breakdown] driving the material into a conducting state is required to obtain the resistance-switching memory operation. A microscopic explanation has not yet emerged to describe the soft nondestructive electrical breakdown and the memory switching in binary and ternary oxides such as the perovskites. Current-induced bistable resistance effects or voltage-controlled negative resistance phenomena in certain compounds, such as Nb₂O₅, TiO₂, Ta₂O₅, and NiO [3–6] investigated long ago and selected perovskites investigated recently, exhibit strong similarities in current–voltage (I–V) characteristics from the macroscopic down to the nanometer scale [7–9]. This suggests that a common physical model may be applicable.

Extended defects such as dislocations [10] and changes in the oxygen-vacancy concentration [11, 12] are considered responsible for the conducting state achieved by electroforming, while local reduction/oxidation processes have been proposed as an explanation for the resistance-switching mechanism [12–14]. Other models propose modified interface properties [15–17] or local inhomogeneities in the conduction path [13, 18], as the physical origin of the resistance changes in these materials. On the other hand, a phenomenological approach involves a nonpercolating domain structure as the origin of these changes [19].

While most research is done on thin-film oxide layers, a few reports on memories are based on bulk single crystals [10, 20–23]. The Cr-doped single crystals used for this work exhibit the same memory behavior as thin films and other oxides under investigation. Thus, they are used as a model memory system to study the role of both lattice defects controlled by dopants and possible interface effects in the electroforming procedure and the memory operation. The purpose of this paper is to provide a deeper insight into the mechanisms and the technical feasibility of the memory.

Figure 1(a) displays current–voltage characteristics of a typical Cr-doped SrTiO₃ single crystal memory cell, which is schematically shown in Figure 1(b). The memory cell exhibits a hysteresis in the current–voltage characteristics, that is, a bistable resistance state. A low- and high-resistance state of about 4 kΩ and 400 kΩ, respectively, is obtained for a small bias voltage. These two different states persist after removal of the applied electrical bias. Cr-doped SrTiO₃, therefore, holds potential for nonvolatile memory applications.

Figure 1

Essentially, an identical bistable resistance behavior is found for CMOS (complementary metal-oxide semiconductor)-integrated memory cells. A one-transistor and one-resistor (1T–1R) memory cell architecture was fabricated in CMOS, featuring 180-nm lithography. The memory cell size is 240 nm, and the resistor is 50-nm-thick Cr-doped SrTiO₃ with a technologically relevant Cu bottom and a Pt top electrode. Figure 1(c) shows current–voltage characteristics of such a CMOS-integrated Cr-doped SrTiO₃ memory cell. A low- and high-resistance state of about 50 kΩ and 5 MΩ, respectively, is obtained for a small bias voltage.

Cr-doped SrTiO₃ single crystals grown in air by the floating-zone method [21] were annealed at 1,150°C in a reducing gas mixture (95% Ar, 5% H₂) to make adjustments for a specific condition with respect to the Cr valence. Pt electrodes with a thickness of 50 to 100 nm and with a typical separation of 500 to 50 μm are deposited on (100) faces of polished crystals with either planar or parallel-plate capacitor geometry. The former enables optical access to the volume between the electrodes during the electroforming and memory operation with the electric field predominantly in-plane.

Thin-film memory test samples were performed by radiofrequency sputtering from a SrTiO₃ target doped with 0.2% Cr. On top of a patterned bottom electrode, a polycrystalline oxide layer was deposited and covered with a patterned top electrode. The electrodes were patterned by UV (ultraviolet) or e-beam lithography and liftoff technique. The oxide layer was not patterned. The metals for bottom (Cu or Pt) and top electrode (Pt) were deposited by thermal evaporation.

Electron-paramagnetic resonance (EPR) spectra were taken with a Bruker EMX** spectrometer system at 9.4 GHz (X-band).

X-ray absorption experiments were performed at the LUCIA (Line for the Ultimate Characterizations by Imaging and Absorption) French/Swiss Beamline of the Swiss Light Source (SLS), Paul Scherrer Institut in Switzerland and at the High-Brilliance X-ray Spectroscopy Beamline ID26 of the European Synchrotron Radiation Facility (ESRF) in Grenoble, France. The beam spot size at the sample was 5 × 5 μm². The sampling depth was approximately 3.5 μm. The data were recorded while the device was in the low-resistance state. The temperature dependence of the resistance was measured with a Physical Property Measurement System from Quantum Design, using an ac current of 10 μA at 30 Hz. Thermal images were taken with an infrared microscope equipped with a Hamamatsu infrared camera.

The luminescence experiments were performed using an optical multichannel analyzer covering a spectral range of 300 to 1,000 nm and with a photon-counting Si avalanche photodiode for the wavelength-integrated signal in the 400- to 1,060-nm range. These measurements were conducted under ambient conditions using crystals annealed in reducing atmosphere.

In pure SrTiO₃, the formation of oxygen vacancies during the reduction process leads to filling of Ti 3d conduction band (CB) states, and an insulator-to-metal (I–M) transition takes place when sufficiently high electron doping is established (~10¹⁸ electrons/cm³). In Cr-doped SrTiO₃, however, our reduction process is not sufficient to introduce enough oxygen vacancies to create a conducting state. EPR data (Figure 2) shows that upon reduction of Cr-doped (0.2 mol%) SrTiO₃, a fraction of the Cr dopant changes its valence from 4+ to 3+ (see the inset in Figure 2), indicating that electron doping occurs at the Cr dopant sites instead of filling the Ti 3d CB [24]. The energy levels of Cr³⁺ and Cr⁴⁺ are located in the bandgap of the SrTiO₃ host [20, 23], and therefore, the crystals remain insulating. A reduced SrTiO₃ containing 0.001 mol% of Cr was used as a reference to estimate the concentration of spin centers (Cr³⁺), which is proportional to the double integral of the EPR signal. At this low concentration, one can assume that upon reduction, all of the Cr atoms are converted to the 3+ valence state. The quasi-linear behavior of the Cr³⁺ signal versus the Cr concentration indicates a strong correlation between the Cr doping and the amount of oxygen vacancies [see Figure 2(a)].

Figure 2

Because of the high initial resistance, the electroforming procedure must start at high electric fields on the order of 10⁴ V/cm in order to initiate a current flow in the nanoampere range. As continuous carrier injection leads to an increasing current with time, the voltage is decreased in steps to lower the power level and prevent irreversible damage. The final state in which stable resistive switching is established is typically reached in a few seconds for thin films and up to a few hours in single crystals. This process can be accelerated by photoexcitation in the visible light range.

It was assumed [8, 20] that the Cr³⁺ center acts as a source of carriers to initiate the electroforming. Proof of the photoexcitation of electrons involving the Cr site is provided by simultaneous current–voltage and EPR measurements under illumination at selected wavelengths. Illuminating the crystals at the beginning of the electroforming process results in a drastic increase of current as well as a simultaneous decrease of the Cr³⁺ signal [see Figure 2(b)]. The initial state is rapidly recovered when irradiation stops. Excitation at various wavelengths (355, 633, and 780 nm) confirms that electron transfer to the CB takes place at sub-bandgap energies above 1.86 eV (670 nm) [25], a condition also met for irradiation in the visible range. That this process indeed involves the Cr dopant is corroborated by the spectra of the photoluminescence and electroluminescence (EL) signals, the latter excited with bias voltage pulses at relatively low currents. Both signals show the occurrence of a dominant line at approximately 790 nm, with a full width at half maximum of 45 nm [26]. This emission corresponds to the R-line, characteristic of charge-transfer processes involving Cr³⁺ in an octahedral crystal field [27, 28]. It results from the excitation of an electron from Cr³⁺ to the CB, leaving the Cr in the tetravalent state. When trapped by Cr⁴⁺, A CB electron will form an excited Cr³⁺ state, which subsequently relaxes via the ²E → ⁴A_2g transition to the ground state [25, 29]. The R-line emission, observed both in the delayed luminescence at zero bias at the early stages of voltage-induced stressing and in the EL, can be taken as proof of the contribution of the Cr³⁺ impurity gap states in promoting the I–M transition in SrTiO₃.

After preparation, the memory cells based on single crystals and films of Cr-doped SrTiO₃ are insulators unless they are treated with a reducing agent. In an electroforming process, the resistance is decreased on voltage-induced stress. Figure 3(a) shows the resistance of a 180-nm-thick Cr-doped SrTiO₃ film with a 1-mm² electrode area at various stages of the electroforming process. The forming process was conducted in steps of 1-minute duration, applying a stress voltage of 1 V at 285 K. The resistance decreased by roughly one order of magnitude within 12 steps of this rather mild stress condition. The electroforming process can be accelerated significantly using higher stress voltage. However, this slow procedure was chosen in order to analyze the electroforming mechanism using temperature-dependent resistance measurements R(T). After each forming step,  a temperature sweep was  performed measuring R(T) [Figure 3(b)]. R(T) follows a 1/T^¼ dependence in the early stage of the electroforming process, indicating a variable-range hopping mechanism [11]. The slope of the temperature dependence is associated with the density of localized hopping sites. The observed decreasing slope with progressive electroforming steps corresponds to an increasing number of these localized sites being responsible for the increase in conduction. Fully formed devices exhibit a metallic temperature dependence of the conduction (see the discussion in the next section).

Figure 3

To obtain detailed information on the microscopic nature of the conduction increase in Cr-doped SrTiO₃ single-crystal memory cells, x-ray absorption near-edge spectra (XANES) at the Cr and Ti absorption K-edges and micro-x-ray fluorescence (XRF) maps were collected. For x-ray absorption at the Cr K-edge, the dipole transition promotes an electron from the Cr 1s core state to an empty 4p band state. The energy position of the Cr K-edge provides information on the valence, whereas the absorption fine structure is characteristic of the spatial environment of the absorbing atom.

Figure 4(a) displays the Cr K-edge XANES underneath both electrodes; the anode is labeled (A) and the cathode is labeled (C). A spectrum of nonconditioned Cr-doped SrTiO₃, labeled (R) for reference, was taken 200 μm away from the electrodes. Underneath the anode, the position of the Cr K-edge (measured at half the edge jump) shifted by 1.9 ± 0.2 eV; this binding-energy shift is proof that the Cr valence underneath the anode is increased. Underneath the anode, the Cr K-edge XANES resembles the K-edge of Cr⁴⁺, whereas the XANES at the cathode and reference positions resembles the K-edge of octahedrally coordinated Cr³⁺. (A comparison of the K-edge XANES of Cr with different valences and spatial environments can be found in Reference [21].) Also shown in Figure 4(a) is the difference between Cr K-edge spectra taken at the electrodes and the reference position. This contrast has a peak at 6,007.3 eV for Cr underneath the anode. Essentially no contrast is found at the cathode, that is, no change of the binding energy occurs. The electronic state and coordination of Cr remain essentially unchanged in the volume probed by the beam.

Figure 4

Figure 4(b) shows the Cr K-edge XANES taken in the conducting path in the vicinity of the interfaces at the anode (PA) and the cathode (PC). Both spectra exhibit no energy shift compared with the K-edge of Cr³⁺ and essentially identical main-edge features. For energies in the pre-edge regime, however, a significant increase of absorption is found. This pre-edge region is most sensitive to structural distortions. This pronounced absorption increase at the Cr pre-edge in the conducting path originates from oxygen vacancies, located at octahedra surrounding the Cr ions [12]. Also included in Figure 4(b) is the contrast of the Cr K-edge in the conducting path. Both spectra (PA) and (PC) have a maximum in the K-pre-edge at 6,004.3 eV.

To characterize the lateral distribution of the electronic states of the Cr ions, XRF maps were taken near the electrodes. Figures 4(c) and 4(d) display the XRF maps taken at 6,007.3 and 6,004.3 eV, respectively. The Cr fluorescence was normalized with the Ti fluorescence to correct for the absorption of the Pt electrodes. The map taken at 6,007.3 eV [Figure 4(c)] has maximum contrast between a Cr valence of 3+ and 4+. The color scale represents the average valence of Cr in the probed volume, ranging from predominately Cr³⁺ (blue) to predominately Cr⁴⁺ (red). Only in the far-field region of the anode is a Cr valence of 4+ found. The Cr valence is 3+ underneath the cathode and between the electrodes of the memory cell. In other words, in the active area of the memory cell (i.e., between the electrodes), only Cr³⁺ is found. Therefore, probably the Cr valence change Cr³⁺ to Cr⁴⁺ is not of critical relevance for the conditioning of Cr-doped SrTiO₃ memory cells.

Figure 4(d) is the XRF map taken at 6,004.3 eV in the Cr pre-edge region of the XANES spectrum. At this energy, a pronounced contrast, indicative for oxygen vacancies V_O located at octahedra surrounding the Cr ions, is found for the conducting path of the memory cell. This shows that the conditioning process introduced a path of oxygen vacancies in the memory cell. These oxygen vacancies provide free carriers in the Ti 3d band, leading to metallic conduction.

At the completion of the electroforming process, the resistance typically decreases to the kΩ range, for which the I–V loops exhibit the typical hysteretic characteristics (see the inset of Figure 5) of the resistive memory switching. EL measurements conducted during the hysteresis loops reveal that light emission occurs only in one branch of the loop, namely where the switching from the low-resistance to the high-resistance state takes place [Figure 5(a)]. As the transition occurs while the current is hardware limited by the compliance function of the voltage source, the actual true voltage applied is the relevant parameter for the detection of the switching and the correlation with the EL event. The wavelength-integrated EL signal and the true voltage measurement during a series of I–V loops are shown in Figure 5(b). The abrupt changes in the V curve versus the t curve in the negative branch of the hysteresis indicate resistance changes that coincide with the onset of a strong EL. Apparently, a threshold value either in the power density or in the applied field needs to be reached to initiate the process and maintain the EL until the voltage drops from its plateau.

Figure 5

As seen in Figure 5(a), the EL spectrum covers a wide wavelength range, suggesting inhomogeneous broadening. It is composed of two salient bands, one of them centered at the position of the R-line, already observed prior to the electroforming. This indicates that the EL arises from a dynamic process involving trapping of carriers at the Cr⁴⁺ centers with subsequent radiative transitions to the ground state. Therefore, in the conducting state, the charge-transfer processes via the Cr bandgap states play a significant role in the electronic transport. The second band is a new emission band appearing in the 900- to 1,000-nm range. The intensity of this band becomes predominant in the fully electroformed sample. One could invoke Joule heating as a possible origin of this emission. Spectral analysis of the radiation emitted from crystals heated up to 980°C indicates, however, that the spectral features cannot be accounted for by black-body radiation alone. The latter band can, thus, be assigned to the ⁴T₂ → ⁴A_2g transition, which because of strong lattice coupling appears red shifted in relation to the ⁴A_2g → ⁴T₂ absorption band. In our samples, the red shift is approximately 300 nm, comparable to that found in other oxides [30–33]. The emergence of this band indicates that during electroforming, the maximum of the EL intensity shifts from the ²E(t³_2g) → ⁴A_2g(t³_2g) (R-line) to the ⁴T₂(t²_2ge¹_g) → ⁴A_2g(t³_2g) transition. This effect can be accounted for by a nonreversible field-induced distortion of the Cr-occupied octahedral site. Such a change of balance from the R-line to the ⁴T₂ → ⁴A_2g transition has been observed in experiments of pressure-controlled distortion of the octahedral site [33]. XAS experiments, on the other hand, give evidence of electric-driven modification of this site in Cr-doped SrTiO₃, where an oxygen-vacancy migration into confined regions of the crystal is revealed (see Figure 4). The relationship between changes of the EL signal and the modifications of the octahedral lattice site occurring with progressing electroforming can, thus, be rationalized by considering that the R-line involves the t³_2g orbitals whose lobes point between the O ligands, whereas the ⁴T₂ → ⁴A_2g transition involves the e_g orbitals, represented by the x² − y² and z² 3d wavefunctions [34], whose lobes point toward the O ligands. Hence, both octahedral distortions [24, 30, 35, 36] as well as vicinal oxygen vacancies can strongly affect the relative 3d intrashell transition probabilities of the Cr dopant atom.

Figure 6(a) shows the temperature dependence of the resistance of a single-crystal memory cell in the low- and the high-resistance state. The decrease of the resistance upon cooling indicates that a metallic path exists between the electrodes for both resistance states. The resistance obeys a quadratic temperature dependence [21]. This T²-dependence shows that the electron–electron scattering dominates the electron–phonon scattering processes, resembling electron-doped SrTiO₃, for example, found for La_1-xSr_xTiO₃ [37]. In addition, Figure 6(a) shows that R_high(T) ∝ R_low(T), where R_high(T) and R_low(T) are the resistances of both states. This would be consistent with a change of cross-section of the metallic region upon resistance switching.

Figure 6

Figure 6(b) displays an infrared thermal micrograph of the single crystal memory cell. The micrograph was taken while applying an electrical current of +5 mA at a bias voltage of approximately 30 V, that is, approximately 150 mW of power dissipated in the memory cell. The false-color image reflects the temperature distribution of the memory cell. The temperature elevates in a laterally confined path with a width of about 5 μm between the electrodes. The majority of the power is dissipated near the anode electrode, reflected by the hotspot, the temperature of which is estimated to be several hundred degrees centigrade. This indicates that the local resistance is highest in the vicinity of this anode electrode. Combining the results of XRF [Figure 4(d)] and the thermal image [Figure 6(b)], we propose that the resistance switching in Cr-doped SrTiO₃ originates from an oxygen-vacancy drift toward or away from the electrode that was used as an anode during the conditioning process.

The time traces of applied voltage and measured current of a macroscopic and a nanoscale Cr-doped SrTiO₃ memory cell are shown in Figures 7(a) and 7(b), respectively. The macroscopic and the nanoscale cells have electrode areas of 250 × 250 μm² and 200 × 200 nm², respectively. Both cells consist of a Cu bottom electrode, a sputtered Cr-doped SrTiO₃ thin film, and a Pt top electrode [38]. The macroscopic cell requires about 1 mA over a duration of 60 ms for programming, whereas for the nanoscale cell, 50 μA for 130 ns is sufficient. Although the area and programming timescale by six orders of magnitude, the current of these two samples only differs by one order of magnitude. The ratio of the read currents after the write and erase pulses and, hence, the resistance ratio R_reset/R_set are on the order of 10 for both samples. The voltage required to drive the necessary programming current thin-film Cr-doped SrTiO₃ cells is typically on the order of 1 V, independent of the electrode area. However, this particular set of nanoscale cells defined by e-beam lithography required 4–5 V due to an additional series resistance.

Figure 7

Figure 8(a) shows the development of the reset and set states (R_reset and R_set) as a function of the number of write and erase cycles. This endurance test was performed on a thin-film Cr-doped SrTiO₃ memory cell with a cross-point geometry. The overlap area of bottom and top electrode lines was 200 × 200 nm². The sample sustained 10⁵ programming cycles using symmetrical write and erase pulses of ±4 V and 2-μs duration. Whereas R_set remained close to its initial value of 50 kΩ, R_reset decreased during cycling from 650 kΩ to 110 kΩ. The programming currents I_write and I_erase stayed in a range of 50–100 μA. However, the microscopic origins of the degradation mechanism, the occurring resistance fluctuations, and the bit-to-bit variability are not fully understood and need further investigation. In Figure 8(b), the data retention of an adjacent cell of the same dimensions is shown. R_set and R_reset are measured at room temperature over 10⁵ seconds after a single write and erase pulse (±4 V, 2 μs), respectively. Both R_reset and R_set change by about 2% within this timeframe.

Figure 8

Further decrease of the cross-point area to 30 × 50 nm² led to a lowering of the programming current to less than 10 μA. The switching time, however, remained unchanged.

We investigated nonvolatile memory based on Cr-doped SrTiO₃. We used electrical transport, thermal microscopy, EL, and x-ray absorption measurements to investigate the microscopic nature of the electroforming and resistance-switching processes.

The observation of electrically stimulated light emission in Cr-doped SrTiO₃ provides evidence of dynamic processes involving trapping and subsequent radiative decay of electrons at the Cr dopant site. The observation of EL occurs only when a memory cell is switched from the low-resistance to the high-resistance state.

X-ray absorption experiments proved that the microscopic origin of the electroforming (i.e., the insulator-to-metal transition) is the creation of oxygen vacancies. The Cr acts as a seed for the localization of oxygen vacancies, leading to a statistically homogeneous distribution of charge carriers within the conducting path. From the combination of laterally resolved micro-x-ray absorption spectroscopy and thermal imaging, we propose that the resistance switching in Cr-doped SrTiO₃ originates from an oxygen-vacancy drift toward or away from the electrode that was used as an anode during the conditioning process.

Nanoscale Cr-doped SrTiO₃ memory cells exhibit short programming times (≤100 ns) and low programming currents (<100 μA). We have confirmed an endurance of 10⁵ write and erase cycles. Moreover, the weak area dependence of the current and the ease of integration with CMOS technology demonstrate the large potential for future scaling toward high-density nonvolatile memory.

We thank W. Riess and R. Allenspach for discussions, M. Tschudy, H. P. Ott, and K. Wasser for technical assistance, and M. Schwarz and A. Jakubowicz, Bookham AG, Zurich, Switzerland, for support with infrared microscopy. Part of this work was performed at the Swiss Light Source, Paul Scherrer Institut, Villigen, Switzerland. Fabio La Mattina gratefully acknowledges the support of the Swiss National Science Foundation.

**Trademark, service mark, or registered trademark of Bruker BioSpin in the United States, other countries, or both.

Received September 24, 2007; accepted for publication October 21, 2007; Published online July 2, 2008.