Thermal and Chemical Integrity of Ru Electrode in Cu/TaO_x/Ru ReRAM Memory Cell

When positive voltage applied to the Cu electrode of a Cu/TaO_x/Pt is swept at a constant voltage ramp rate rr [V/s], the current will remain substantially zero until a critical voltage V_set is reached, at which a Cu conductive filament (CF) is formed connecting the Cu and Pt electrodes with each other, and the cell switches from a high resistive state (HRS) characterized by R_off (1–900 MΩ) to a low resistive state (LRS) characterized by R_on (70–15 kΩ), yielding a ratio of R_off/R_on ≈ 10³−10⁷. When the voltage is swept back, ohmic behavior is observed until a negative voltage V_reset is reached when the CF is ruptured and the current collapses to a very small value. A characteristic log-scale I-V switching characteristic of the Pt device is shown in Fig. 3. When the cell is set to an ON-state for the very first time, one speaks of forming operation, characterized by the forming voltage, V_form. The V_form is always larger than V_set, since the Cu filament has to be formed in its entirety and not partially whereas for a reset operation only a fraction of the filament has been ruptured but the entire filament not totally undone at the subsequent set operations. For Cu/TaO_x/Pt devices, we find distribution V_form with a mean V_form,m = 4.6 V and a standard deviation of σ = 0.6 V, the V_set distribution with a mean, V_set,m = 2.8 V, and standard deviation, σ = 0.58 V. The rupture of the CF is triggered mainly by Joules heating at a critical current I_reset = V_reset/R_on. The V_reset distribution of our Cu/TaO_x/Pt devices is characterized by V_reset,m = −0.9 V and σ = 0.32 V. These distributions apply roughly both to a multitude of devices as well as to a single device that has been switched repeatedly. In general, we observe that statistically the V_set threshold voltage tends to increase with increasing V_reset threshold voltage of the preceding reset operation. This dependence of V_set on the preceding V_reset is shown in Fig. 4.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The endurance of the Pt devices depends on the I_cc current during the set operation. For not too low I_cc (>10 μA) and not too high I_cc (<5mA), the device can be switched very often. Typically, in a few cases, the device was switched for 140 times and it could be still switched, but because of the lack of automatic probing station and limitation of student's physical endurance further switching cycles have been not attempted. We have also probed the retention of the devices as shown in Fig. 5. As seen in the Fig. 5 the retention of the off-state (characterized by R_off) and of the on-state (characterized by R_on) for Pt and Ru are very satisfactory. Larger off-state variations than on-state variations can be observed for the Ru devices but the noise does not compromise the retention capability as for R_off the resistance excursions are seen toward higher resistances. While the plateau of the R_off is at 200–300 M Ω, the excursions can reach GΩ range.

Zoom In Zoom Out Reset image size

In most instances, a compliance current I_cc is imposed, lest the devices be damaged during the set operation. No I_cc current limitation is applied during the reset operation. R_on of the LRS state depends on I_cc via the relation 1, where the exponent n for cation filaments is very close to unity.

The R_on–I_cc relation in Eq. 1 has been reported to be valid for numerous anode/electrolyte/cathode material systems^46–53 with n≈1. In Ref. 53 it has been shown that the constant A in Eq. 1 is universally correlated to the minimum SET voltage for all metallic conductive filaments reported so far. It is known⁵⁴ that the compliance current enforced externally by Keithley 4200-SCS leads to a transient overshoot current especially during the very fast last phase of filament formation, as Keithley circuitry is unable to respond instantly to large current changes. In this work, where different devices are compared with one another, the effect of current overshoot should be the same for all devices, and thus the relative comparisons should still be valid.

From the R_on vs I_cc plot shown in Fig. 6a, the constant A = 0.5 V and the exponent n = 1.01 are obtained by curve fitting. Here, the voltage sweep rate rr = 0.2 V/s has been used for the set operation. In Fig. 6b V_set is plotted as a function of the sweep rate. It is seen that the minimum V_set at low sweep rates corresponds to the constant A = 0.5V as predicted by the theory.⁵³ The slope of the plot was found out to be the minimum possible value of SET voltage. Fig. 7 shows the dependence of the forming voltage and on-resistance on three different voltage sweep rates of 0.1 V/s, 0.2 V/s and 2V/s. It can be seen that not only does the forming voltage increase with increasing ramp rate, but also the on-resistance increases with increasing ramp rate. The dispersion of R_on is significantly higher at a high ramp rate compared to the one at low ramp rate. According to Eq. 1 R_on can be varied by imposing different levels of the compliance current I_cc and voltage sweep rate. When the voltage sweep rate increases from 0.2 V/s to 2V/s, the constant A value of relation 1 increases from 0.5 V to 1.01 V.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Fig. 8 depicts the dependence of R_on on voltage sweep rate under different levels of compliance current (5μA, 100μA and 1mA). The range of voltage ramp rate is 0.1 V/s to 2 V/s. It can be seen that R_on is larger at lower I_cc, in agreement with Eq. 1. Under low compliance current (5 μA and 100 μA), R_on increases with increasing voltage sweep rate. However, in the case of high compliance current, R_on becomes independent with voltage ramp rate, indicating ohmic behavior of a robust and stable conductive filament.

Zoom In Zoom Out Reset image size

Typical switching cycles underlying the data in the previous diagrams are shown in Fig. 9. Here, for the sake of clearer representation, only even cycles of a 12 switching set-reset cycles are shown. The first set operation is done on a fresh cell and results in a forming voltage V_form = 3.2V. The subsequent set voltages lie between 0.7V and 2.1 V. The reset voltages are more tightly distributed and lie between −0.9V and −1.4V.

Zoom In Zoom Out Reset image size

Since filament rupturing is chiefly a thermal phenomenon, the heat conductivity of the inert electrode plays a significant role in determining the maximum temperature of the filament.

During the heating of a resistive cell A, the reset operation results in rupturing of the CF at a reset current I_res = V_res/R_on of typically a few of mA. Hence, most of the Joules heat Q_JH is deposited according to Eq. 2

where V_res is the reset voltage, t_res = V_res/rr is the reset time, the low reset ramp rate rr = 0.1V/s to maximize the heating during the reset.^55,56 It has been shown⁵³ that R_on depends on I_cc via R_on = A/Iⁿ_cc given by Eq. 1, with n≈1 and A≈0.5V for Cu/TaO_x/Pt devices. Depending on chosen values for I_cc and rr, Q_JH can vary from 3 to 60 μJ. The possibility of controlling the dissipated heat in the filament in terms of compliance current, I_cc, and ramp rate, rr will be important for the characterization and analysis of the electrical performance of the Pt and Ru devices.

With respect to the geometrical shape of the filament, we associate the cells with large R_on with a truncated cone with a sharp constriction of the top (see Fig. 10a), where the bulk of the R_on resistance is concentrated at the tip of the cone. Filaments with a sharp constriction at the top of the cone are easy to rupture since the maximum Joules heat is deposited at the tip (highest resistance) and leads there to a high local temperature which, in turn, leads to Cu out diffusion and formation of a gap in the filament which completes the rupture of the filament leading to the restoration of the HRS state. In contrast, the shape of Cu filament with a small R_on approaches that of a cylinder (Fig. 10b) and the temperature hot spot moves to the center (i.e. midway between the electrode interfaces) of the filament where it is much more difficult to reach high temperature (much lower resistance) and cause out-diffusion of a larger number of Cu atoms to form a gap. Experimentally, we find consistently that filament rupturing is more difficult with lower R_on resistance, i.e. formed at higher I_cc (see Eq. 1).

Zoom In Zoom Out Reset image size

In the set operation there are two fluxes responsible for the formation of the filament as shown in Fig. 10c at the tip of the filament there is a large voltage drop which creates high electric field between the filament tip and the Cu electrode, which allows for further transport of Cu⁺ ions from the Cu electrode to the filament even after establishing the initial connection between the Cu filament and the Cu electrode. The role of I_cc is to limit the resulting voltage drop and thus to set a limit on the electric field driving the Cu⁺ ion transport. When sufficient number of Cu atoms is deposited at the tip, the resistance of the filament decreases, reducing the electric field at the tip and bringing the formation of the filament to a halt at a given I_cc. When I_cc is increased, then the electric field at the tip increases proportionally triggering additional arrival of Cu⁺ ions until the resistance of the filament drops sufficiently to reduce the electric field and thus to halt further Cu⁺ ion transport. The second flux is the Cu atom diffusion flux weakening the base of the filament and leads to an increased R_on resistance. The diffusion effect can be gradual in so far as any weakening of the filament leads to increased R_on resistance which, in turn, may trigger the compensating constructive Cu⁺ electromigration flux.

The Cu atom diffusion flux out of the filament is impacted by the thermal conductivity of the inert electrode. For the otherwise same conditions, the inert electrode with high heat conductivity will be able to remove the heat at a higher rate than an electrode with low heat conductivity. The larger heat removal rate will result in lower attainable maximum temperature. Hence, a cell with inert electrode of low heat conductivity is bound to display enhanced Cu diffusion and structural weakening of the filament. The high temperature in the filament during the reset will trigger Cu diffusion near the base of the filament partly into the dielectric, partly along the TaO_x/inert electrode interface, and partly into the inert electrode if the inertness of the electrode is compromised. The net result of those diffusion components is the weakening of the base of the cone and a gradual transformation of the shape of the filament from a sharply cone-shaped into a more cylinder-shape filament as shown in Fig. 10b. Once the shape of the filament is sufficiently close to a cylinder-like shape the respective filament will be very difficult to be ruptured and the number of switching cycles comes to a halt.

Fig. 11a illustrates the SET voltage's dependence on ambient temperature and the RESET voltage's near independence on the ambient temperature. V_set and V_reset of five different cells are plotted when the ambient temperature is 20°C, subsequently increases to 85°C and reduces back to 20°C. It can be seen that the SET voltage at 85°C is reduced by ∼ 0.8V compared with the set voltages at room temperature. The reduction of V_set with increasing temperature can be explained by three different mechanisms in TaO_x. (1) The ionization rate of Cu, the rate for Cu ionizes to Cu ion (Cu → Cu⁺ + e⁻), increases with increasing temperature. (2) The mobility of Cu ions in TaO_x increases with increasing temperature. (3) The nucleation of Cu ions or the speed of forming the cluster of Cu at the interface of Pt electrode may also accelerate with increasing temperature, as all of those reactions are governed by Arrhenius law. We find that the RESET voltage is independent of temperature. The local temperature leading to rupturing of the filament has been reported⁵⁶ to be around 900°C and therefore slight change of ambient temperature has little impact on the high local temperature responsible for the partial dissolution of the filament. The dependence of the corresponding switching currents I_reset on ambient temperature is shown in Fig. 11b as a cumulative probability distribution of I_reset at room temperature and at 85°C. It can be seen that the distribution of I_reset at room temperature is much tighter than at 85°C.

Zoom In Zoom Out Reset image size

The metallic Cu filaments are well characterized by their temperature coefficient of resistance (TCR). For Cu filaments with R_on = 1 kΩ we find TCR(Cu) = 0.003 K⁻¹ (which is close to the TCR of bulk Cu of 0.0039 K⁻¹. In Figs. 12a and 12b the TCR measurement for a fragile filament (R_on = 11.2 KΩ) formed at I_cc = 10 μA and for a robust filament (R_on = 250 Ω) formed at I_cc = 10 mA is shown, for Pt and Ru device, respectively. For a fragile Cu filament (@I_cc = 10μA) we get TCR = 0.00235K⁻¹ (Fig. 12a) and for the robust filament TCR = 0.00362 K⁻¹ (Fig. 12c). Very similar TCR values are found for the Ru device: for a filament formed at I_cc = 10μA TCR = 0.00236K⁻¹ (Fig. 12b) and for robust filament TCR = 0.00350 K⁻¹ (Fig. 12c). The Cu bulk values have been verified on our Cu lines with following dimensions: thickness 150 nm, width 1 μm – 35 μm, and length 150 μm. The measurement of the TCR of the copper lines yielded consistently TCR values of 0.0388 K⁻¹ − 0.00396 K⁻¹, in good agreement with the TCR value of the TCR for bulk Cu of 0.0039 K⁻¹. This is also in agreement with a study of the electric properties of freestanding Cu nanowires reported by Y. Zhao et al.⁵⁸ who report that TCR for Cu nanowires decreases with decreasing of the diameter: at 230 nm TCR = 0.0038 K⁻¹, at 100 nm TCR = 0.0033 K⁻¹ and at 50 nm TCR = 0.0028K⁻¹. Also A. Bid et al.⁵⁹ reported TCR values between 0.004 K⁻¹ and 0.0025 K⁻¹ for Cu nanowires with diameters between 200 nm and 15 nm respectively. The metallic nature of the filament has also been deduced by Xiao et al.⁶⁰ using ab initio atomic calculations attributed the main contribution to the conductive path along Cu-Cu metallic bonds. As seen from Fig. 12 the TCR increases with decreasing R_on resistance. Thus in both devices the electric properties of the Cu filament are the same.

Zoom In Zoom Out Reset image size

The Cu/TaO_x/Ru devices have been characterized in the same way as the Cu/TaO_x/Pt devices. We find following distributions: for V_form a mean V_form,m = 7.3 V and a standard deviation of σ = 0.7V, for V_set distribution with a mean, V_set,m = 4.4 V, and standard deviation, σ = 0.9 V. The rupture of the CF is triggered mainly by Joules heating at a critical current I_reset = V_reset/R_on. The V_reset distribution of our Cu/TaO_x/Ru devices is characterized by V_reset,m = −3.4 V and σ = 0.7 V. Thus, all the critical voltages of the Ru devices are considerably higher than of Pt devices: ΔV_form = 2.7 V, ΔV_set = 1.6 V and ΔV_reset = 2.5 V. For V_form and V_set part of the difference can be explained by the difference of the work function between Pt and Ru of 1.65 eV, which agrees remarkably well with the observed differential for set voltages, ΔV_set = 1.6V. The remaining discrepancy of about 1V can be attributed to nearly 100% higher surface roughness of Pt (2.30 nm) than Ru surface roughness (1.20 nm), as seen in Fig. 2. Higher surface roughness generates higher local fields thus allowing for filament formation at lower applied voltages. The differential in set voltage between Pt and Ru of ΔV_set = 1.6V accounted solely by the work function difference, confirms this conclusion since for the set operation the surface roughness of the inert electrode is irrelevant as the gap in the filament near the Cu electrode is independent of the surface roughness of the inert electrode. In the set operation, only the gap in the ruptured filament (far away from the TaO_x/inert electrode interface affected by the surface roughness) has to be closed to restore the filament. The built-in field in case of Pt electrode is the work function difference between Pt and Cu divided by the thickness of the TaO_x dielectric, which in case of Pt devices comes to a quite high field of 6.8 × 10⁵V/cm. This field favors Cu⁺ ion transport toward the Pt electrode. The corresponding built-in field in case of Ru devices is 6 × 10⁴V/cm, i.e. one order of magnitude lower than for the Pt device. The field responsible for the formation of the Cu filament in its final stages has been estimated to be (1--2) × 10⁶V/cm.⁶¹ The calculation of the effective fields at V_set for Pt and Ru including the built-in potential, are, indeed, in this range and approximately the same: (Δϕ(Pt,Cu)/q + V_set,m(Pt))V/25nm ≈ (Δϕ(Ru,Cu)/q + V_set,m(Ru))V/25nm ≈ 1.8 × 10⁶V/cm, in excellent agreement with.⁶² This analysis clearly shows that for the filament formation the local electric field is the determining factor.

The difference between V_reset voltages of 2.5V cannot be explained by the work function difference between Pt and Ru, as the main mechanism of the dissolution of the conducting filament is the Cu atom outdiffusion at high temperatures near the filament constriction. We discuss the higher reset voltage for Ru devices in more detail further below as this issue appears to be related to the issue of Ru inertness and integrity. The high built-in field for the Pt device has been verified experimentally in the following way. We have set Pt and Ru devices under three different I_cc conditions of 10, 50, 100, and 250 μA and subsequently reset the devices. During the reset the ruptured gap in the filament is estimated to be ∼ 5–7 nm. This means that the built-in field in the gap would be for Pt devices 1.65V/7nm = 2.4 × 10⁶ V/cm and for Ru devices 0.1V/7nm = 1.4 × 10⁵ V/cm, i.e. more than one order of magnitude lower than for the Pt device. The built-in field for Pt device is high enough to spontaneously set the device. Indeed, experimentally we observe a spontaneous setting on a large fraction of devices after a wait time of 48 h (i.e. when the device is at 300K) after resetting the device that initially has been set at I_cc = 10μA and occasionally for those set at I_cc = 50μA. No spontaneous setting of the devices has been observed for I_cc = 100μA. Not a single instance of spontaneous setting of the devices has been observed for the Ru devices tested at the same conditions. Thus the spontaneous forming of a ruptured filament is facilitated by the high work function difference between Pt and Cu metal electrodes. Spontaneous set operations have been also observed during consecutive switching cycles of a Pt device. After a 5^th reset operation, we observed that the device was ON spontaneously, after a time of the order of tens of seconds. Here, the spontaneous formation capability of the CF has been augmented by the elevated local temperatures in the cell caused by the Joules heat dissipation leading to higher rate of Cu ionization and higher Cu⁺ ion electromigration mobility. This is another confirmation of the impact of work function difference between the active and inert electrode.

In Fig. 13, a typical set and reset bipolar operation for Cu CF in Cu/TaO_x/Ru device is shown, corresponding to Fig. 3 for the Pt device. The V_set = 3.5 V at I_cc = 10 μA and V_reset = −3 V. The bipolar switching cycles have been repeated on Cu/TaO_x/Ru devices. We observe that most of Cu/TaO_x/Ru devices is becoming not resettable after few set-reset operations. Thus in terms of endurance the Pt device is vastly superior to the Ru device. While the Pt devices could be switched open-ended with verified 140 switching cycles, the best Ru device had an endurance of maximally 12 switching cycles. Thus the endurance is the main weakness of Ru devices for the memory applications when compared to the Pt devices. The failure of the Ru devices after a few switching cycles is likely to be related to the geometrical shape of the Cu filament. In terms of retention, Ru devices show comparable retention as shown in Fig. 5 however, Ru retention curves are much noisier than the Pt retention curves for both the on-state and off-state. The noise of the Ru off-state curve could be attributed to the surface roughness. The noise of the Ru on-state curve is likely to be a consequence of degraded inertness of the Ru electrode. Note that during a retention test we apply a small constant voltage (of 0.05V) continuously for a long time. Although the currents are small even in the case of the on-state, the dissipated Joules heat accumulates over time and raises the temperature of the filament.

Zoom In Zoom Out Reset image size

As in the case of Pt devices, R_on for Ru devices also depends on the compliance current, I_cc. On a double logarithmic scale, on can see that R_on also decreases linearly with increasing I_cc conforming to the Eq. 1. As is shown in Fig. 14, the extracted parameters are A = 0.61 V and n = 1.08. The condition for the set operation is: voltage sweep rate rr = 0.2 V/s. It can be concluded that Cu/TaO_x/Ru devices has slightly higher constant A value than Cu/TaO_x/Pt device when the set operation is the same (rr = 0.2 V/s). As shown elsewhere⁵³ with the exponent of I_cc in the denominator of Eq. 3 being close to 1, the constants 0.5 V and 0.61 V can be interpreted as the lowest possible V_set voltages under which the Pt device and Ru device, respectively, can be set at a given voltage ramp rate. Lowest value for the constant A can be reached as a limiting case for very slow voltage ramp rates. The difference in the minimum set voltage extracted from R_on - I_cc characteristics confirms that the V_set voltage for Ru devices is higher than for Cu device.

Zoom In Zoom Out Reset image size

Since in Pt and Ru devices the filament is made of Cu atoms one would expect similar temperature coefficients of resistance. In Figs. 12b and 12c, we found TCR (Ru) = 0.00236 K⁻¹ and TCR(Pt) = 0.00235 K⁻¹ for similar values of R_on which, within the accuracy of our measurement, means that the values are identical. Therefore, we conclude that under a positive voltage stress applied to the Cu electrode, Cu conductive filaments are formed in both devices. Moreover, these values are consistent with TCR values of Cu CFs observed in many other devices. It has been observed that for both devices TCR increases with decreasing R_on reaching a value of 0.0035 K⁻¹ at R_on = 300 Ω as seen in Fig. 12c.

To summarize the characterization of Ru devices one can state that with the proviso that all critical threshold voltages, V_form, V_set, and V_reset are higher than for Pt devices which can be partly understood in terms of work function and in terms of differences in surface roughness difference between Pt and Ru, Pt and Ru devices behave very similarly. There are, however, two major differences: (i) V_reset is much higher in Ru than in Pt devices, and (ii) and the endurance in terms of the maximum number of sequential set-reset cycles for Ru (at most 12) is substantially lower (maximum than for Pt devices (more than 140), particularly when the devices are set at I_cc>100μA, i.e. for low R_on values. We address those two issues in the subsequent sections in more detail.

The major drawback of the Ru device is that while Pt devices can be switched repeatedly back and forth for at least 140 times, Ru devices are becoming non-resettable after several set-reset operations (less than 13) and sometimes after the first set operation when the set operation is performed at high I_cc. In some cases, it was even difficult to reset a high resistance Cu filament in a Ru device formed at I_cc as low as 5 μA. The failure of the Ru devices to switch after several switching cycles is likely to be related to the geometrical shape of the Cu filament.

It is known that even in Pt devices, it is difficult and sometimes impossible to reset the cell if the cell has been set at high I_cc. This phenomenon may be explained by different geometrical shape of the conductive filament. For low I_cc the shape of the CF can be approximated by a truncated cone as shown in Fig. 10a where the bulk of the CF's resistance resides in the tip of the cone. Since during the reset the power dissipated in the filament is equal to I_reset² × R_on, the highest temperature is reached in the constriction at the tip of the cone, where the rupturing of the filament is easy, as shown in Fig. 10a. At high I_cc (as seen in Fig. 13 R_on vs I_cc) the R_on decreases sharply by adding Cu atoms to the upper section of the cone. Therefore, the shape of the filament becomes more and more cylindrical as shown in Fig. 10b. In this case, the maximum temperature, in absence of any pronounced constriction, is reached in the middle section of the cylinder, where the low resistance filament is relatively strong. Hence, rupturing of the CF becomes difficult or impossible.

The cylindrical shape of the CF may be also obtained if the base of the cone in contact with Ru electrode is eroded, see Fig. 15a. As the bottom base of the CF is reduced while new Cu⁺ ions are added from the top during the set operation, the shape of the filament will tend to assume a cylindrical shape. There are several possible mechanisms for the erosion of the base of the cone: (i) Cu surface diffusion along the Ru interface to which has been alluded before in the context of the much larger V_form for Ru than for Pt devices, (ii) crystallization of Ru and out-diffusion of Cu along the Ru grain boundaries, (iii) Cu and Ru chemical reaction with oxygen and/or with Si. Those mechanisms are depicted in Fig. 17 and will be discussed further below.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

To keep the R_on constant, the loss of Cu atoms at the base of the cone has to be compensated by additional flux of Cu⁺ ions explaining the need for higher forming voltages in Ru devices in excess of the work function difference of 1.6V. Whereas, with a nearly perfect stopping power, the shape of the filament in the Pt device can be assumed to be conical with more or less sharp tip at the Cu electrode, the shape of the Cu filament in the Ru device is more cylinder-like especially if the Ru device is undergoing several set-reset switching cycles. These mechanisms appear to be very similar to the properties of Cu/TaO_x/Ta devices observed in Ref. 44.

The on-resistance of a conically shaped resistor is given by:

where h is height, a is the radius of the truncated cone base at the top, b the radius of the base at the bottom, and ρ is the specific resistivity of the filament, see also Fig. 14b. To demonstrate the difference between the hypothesized shapes for the Cu CF for the Pt and Ru device, we have constructed two cone shapes with the same overall CF resistance, R_on: we assume following parameters for the Pt device b_Pt = 3 nm and a_Pt = 0.5nm and for Ru device b_Ru = 1.76 nm and a_Ru = 0.85 nm, for a common height h = 25 nm. Based on Eq. 3, the resistances of Cu CFs for Pt and Ru devices are 15,923 Ω and 15,966 Ω respectively, which are almost the same values shown in Fig. 6a and Fig. 13 when the devices are set at I_cc = 10 μA. However, the geometrical shape is very different. Here, the resistivity of copper filament has been assumed to be ρ = 300 × 10⁻⁶ Ω·cm.⁶² This demonstrates that the two filaments of the same R_on resistance (enforced by the same I_cc current) may have substantially different geometric shape. The different shape of the filament has an impact on the reset operation and the endurance properties of the device.

To study the relative degradation of Cu/TaO_x/Ru vs Cu/TaO_x/Pt devices, we have manufactured two nominally identical Cu/TaO_x/Ru devices (where the layers of the cell proper have been manufactured at the same time) however embedded differently on the Si wafer. While the substrate for Ru device A is the same as for the Pt device, i.e. Ti(20nm)/SiO₂(730nm)/Si-wafer, the Ru device B is manufactured on the layer stack Ti(20nm)/TaO_x(30nm)/SiO₂(730nm)/Si-wafer. Thus, Ru device B has an additional TaO_x-30nm layer inserted between SiO₂ and Ti layers. The electrical characterization of the two Ru devices, A and B, has been guided by the hypothesis that the difference in electric performance between the two devices is caused by effects stemming from larger or smaller exposure to the Joules heating dissipated in the device. Therefore, the form/set operations have been performed at low and high I_cc currents, i.e. I_cc = 1 μA, 5 μA, 10 μA, and 50 μA. Clearly, once the filament is formed exposure to Joules heating effects at I_cc = 50 μA should be larger than at I_cc = 1 μA, see Eq. 2. To control the heating effects even further, also during the reset operation, we have varied the reset voltage ramp rate, rr, between 0.2V/s and 2V/s, see Eq. 2. In the reset operation the currents flowing through the cell are large; on the order of a few mA. Therefore, low ramp rate implies larger heating because of longer duration of exposure to high currents flowing through the filament, before the cell is ruptured at a reset current, I_reset. Thus, most heating would occur for I_cc = 50 μA and rr = 0.2V and least heating for I_cc = 1 μA and rr = 2.0 V/s, see Eq. 2.

In general, we find that the Ru device B with the additional TaO_x layer inserted between SiO₂ and Ti shows much better performance than the Ru device A (i.e. without that TaO_x layer) at all test conditions. However, the degree of the improvement in the switching properties of Ru cell B over cell A varies with the levels of applied compliance current and ramp rate. The highest improvement is found for I_cc = 50 μA and rr = 0.2V/s (high heat dissipation), and almost no difference in performance between the two devices can be observed for the condition I_cc = 1 μA and rr = 2.0V/s (small heat deposition). A typical successful forming, reset, and set I-V characteristics for device B are shown in Fig. 17. No I-V characteristics of comparable quality can be shown for the device A; its best behavior is shown in Fig. 16.

Zoom In Zoom Out Reset image size

There exists a significant difference in endurance between the two Ru devices differently embedded on the substrate. We find that at I_cc = 50 μA and rr = 0.2V/s the Ru device A does not display resistive switching behavior at all, i.e. its endurance is nil. The device A cannot be reset after the filament has been formed the very first time. However, when the reset ramp rate is increased fivefold to rr = 1V/s – thus reducing the Joules heat dissipated in the device - the device A can be neither set or, if the set operation is eventually successful, the device cannot be reset, i.e. the resistive switching cell has been permanently damaged. During the set operation, even if successful, no sharp set transition can be observed but rather a gradual transition to the conductive state characteristic for a dielectric breakdown. In contrast, device B with the additional TaO_x layer at the bottom shows some resistive switching behavior for a few cycles (maximum 6). Sharp formation and set characteristics are observed with V_set between 2.6 V and 4.0 V, however the subsequent reset attempts have proved unsuccessful. The frequency of devices B displaying resistive switching behavior is quite low at about 1% of the devices B tested. We conclude that high heating effects are detrimental to both devices. But, while device B displays some (if small) degree of resisting switching behavior (5-6 switching cycles), device A does not display resistive switching at all.

Keeping the reset ramp rate at 0.2V/s but now reducing the I_cc current from 50 μA to 1 μA, we observe improved resistive switching behavior for both devices with the frequency of the Ru devices B being twice as high (maximum 22 cycles) as for the Ru devices A (maximum 10 cycles). Thus, at low heat dissipated in the cell, the endurance of the Ru device B is 22 switching cycles while the Ru device A exhibits 10 switching cycles.

The switching behavior improves even more when we keep I_cc at 50 μA and increase the ramp rate from 0.2V/s to 2.0V/s for both devices. Now both devices show some resistive switching behavior, however, again, the frequency of resistive switching is at least three times higher for the device B (maximum 30 cycles) than for device A (maximum 11 cycles). Thus again the Ru device B shows a superior endurance over the Ru device A. Clearly, reduced thermal budget due to the faster voltage ramp rate improves resistive switching behavior of both devices. Table I shows comparison of the V_form, V_set and V_reset for both devices.

It can be seen that V_form for device A is higher 4.6 V < V_form(A) < 5.0 V than V_form, for device B 2.9 V < V_form(B) < 4.5V. A similar observation applies to V_set. This behavior could be explained by the assumption that Cu loss of the filament for device A is higher than for device B. It can also be seen that the reset and R_on values for devices A and B are comparable. This indicates that once a Cu filament is established the dynamics of the creation of a gap in the filament during the reset operation are the same.

Table II shows the values for the same quantities as in Table I but now for I_cc = 1 μA and rr = 2.0V/s. Comparing Table I and Table II we find the same trends described above in Table I. The frequency of the devices B tested that display resistive switching behavior is now very high at 100%, while the frequency for the devices A tested showing resistive switching behavior is about 2% for the same switching conditions. Manifestly, the least amount of heating produces best resistive switching behavior for device B.

We have measured device A and B set also at very high I_cc = 0.5 mA at which a robust filament with a low resistance R_on is being formed. We find the same trends as found at lower I_cc current levels. Interestingly, for rr = 2.0V/s the frequency of resistive switching is about the same at 60% for both devices. However, when we lower rr from 2.0V/s to 0.2V/s, i.e. allow for more Joules heating, the frequency of device A drops to 17% while the frequency of devices B stays constant at 60%. Thus again, Ru device B is much less sensitive to dissipated Joules heating than the Ru device A. This clearly demonstrates that the endurance of the device is directly related to the amount of heat deposited in the device.

Finally, we have compared two I-V characteristics of both devices at the same stress conditions with very low I_cc: I_cc = 1 μA and rr = 2V/s. The first measurement was taken on a fresh device. The second measurement was taken on a fresh device that has experienced resistive switching at I_cc = 50 μA and rr = 2V/s, prior to the measurement at I_cc = 1 μA. Thus in the second case the device has experienced considerable "pre-heating" at I_cc = 50 μA. We find that both devices A and B perform worse after they have been "preheated" at I_cc = 50 μA. While in the case of device B some of the devices exhibit resistive switching, there was almost no resistive switching for the device A. This means that at the operation at I_cc = 50 μA, some irreversible changes must have taken place within the memory cell. On the other hand, the same experiment with "pre-heating" at 1 μA did not affect the subsequent resistive switching of both devices demonstrating again that heat dissipated in the cell is the key factor in designing optimum endurance of the cell.

There is a sizeable circumstantial indirect evidence from the extant literature pointing to a possible cause of the degraded behavior of Ru device A vs Ru device B. During the forming and set operations a substantial local heating of the filament takes place. The maximum temperature has been estimated to be between 600°C and 1000°C.^57,63--65 Regner and Malen⁶⁶ have argued that because of the nm dimensions of the filament the thermal transport is non-diffusive and the differential Fourier equation of heat transport is grossly inadequate. The solution of the more adequate Boltzmann transport equation leads to much higher temperature estimates,⁶⁶ as high as 1000°C. Our hypothesis is that such high temperatures even at relatively short times of microseconds up to seconds can activate Si diffusion from the SiO₂ layer through the thin Ti layer into Ru layer, possibly accompanied by a TiO_x reaction. The Ru layer, in turn, can undergo, at such high temperature, local crystallization creating grain boundaries along which Si may readily diffuse.⁶⁷ Thus Si that may reach at the Ru/Ti interface, may induce Ru silicide reaction to Ru_xSi_y and/or it can travel along the Ru grain boundaries and trigger Cu silicide reaction Cu_sSi_t at the forming base of the Cu filament at the Ru/TaO_x interface.⁶⁸ Both silicides are known to be poor diffusion barriers for Cu atoms.⁶⁸ The possible degradation mechanisms of the Ru electrode are shown conceptually in Fig. 18. Those degradation mechanisms may also explain the low frequency of resistive switching behavior for the Ru device A when compared with the Ru device B. For both devices, the worse performance is found when the base of the Cu filament overlaps with a possible Ru grain boundary. In this case, the Cu atoms may be drained out of the filament through Cu diffusion along the grain boundary. The best case scenario for resistive switching is when the Cu filament does not overlap significantly or at all with a grain boundary. In that case the resistive switching behavior in both devices could be comparable. The non-overlap case may be responsible for the few instances when good resistive switching behavior can be observed for device A. In case of device A, the silicide reactions may lead to inclusions of Ru_xSi_y and Cu_sSi_t most likely along the grain boundaries and, hence, such inclusions widen the area at the Ru surface where Cu can easily diffuse into the electrode. Therefore, Ru device A should suffer from Cu in-diffusion more than device B. Both Ru and Cu silicide reactions, compromise the inertness properties of the Ru electrode and entail depletion of Cu atoms from CF into Ru electrode. The motivation of inserting TaO_x layer between Ti and SiO₂ layer in device B, has been to stop - or at least reduce - the Si diffusion into the Ru electrode. Our experimental results presented and discussed below seem to confirm it showing that in case of device B the degradation of the Ru electrode due to silicide reactions and Ru crystallization has been reduced, even if still short of a complete suppression. The TaO_x layer of only 30 nm deposited by PVD is known to form tantalum oxide with lower density and higher porosity than the stoichiometric Ta₂O₅ deposited by Atomic Layer Deposition (ALD). Thus the inserted 30 nm TaO_x layer may still not be a perfect diffusion barrier for Si, either. Third integrity issue is the adhesion of Cu atoms to Ru interface. It is known that Cu has a low wetting angle on Ru (43°)⁶⁸ and high affinity to Ru surface.⁶⁹ Therefore, Cu adheres strongly to Ru surface even at high temperatures as high as 600°C.^70,71 These findings argue strongly against significant Cu surface diffusion on Ru surface. This argument is further corroborated by the significantly higher surface roughness of Ru than that of Pt. Higher surface roughness is less conducive to surface diffusion. However, at 475°C and above Cu starts to diffuse through the energetically favorable inter-grain boundaries of Ru columnar microstructure.⁶⁷

Zoom In Zoom Out Reset image size

These three reliability issues may conspire to cause substantial degradation of the inertness properties of the Ru electrode and trigger enhanced Cu diffusion through the degraded Ru electrode. The existing Ru₂Si₃ may accelerate the Cu diffusion even further.^67,72 At 550°C, Cu can completely diffuse through a 20 nm Ru film, penetrate into the Si substrate, and form Cu₃Si by interacting with the substrate and eventually completely nullifies the barrier inertness properties of the remaining Ru layer.^67,69 The Cu₃Si is found to be morphologically inverted pyramid-shaped rectangular crystallites with orientation perpendicular to the substrate.^73,74 Such isolated crystals form a rough surface and induce a permanent damage to the Cu metallization in CMOS BEOL. Similar damage is most likely being done to the Cu filaments in our devices. Although, ideally both Ru and Cu are relatively inert on thermally stable SiO₂ even at high temperatures, the direct reaction with Si causes barrier failure at the same temperature for both Si and SiO₂ substrate.⁷⁵ Thus, the temperature at which Ru₂Si₃ forms may be the key triggering factor for the degradation of the inertness properties of Ru.⁷²

Given the degradation mechanisms described above, some mitigation strategies may be considered. One is a deposition of a thicker Ru electrode (presently at 50 nm). However, because of the columnar grain structure thicker Ru layer is not likely to be of much help. The other approach would be to limit the temperature of filament rupturing. This is difficult to achieve since the rupturing temperature appears to be an intrinsic property of the device. The formation of Ru₂Si₃ requires higher temperature than formation of Cu₃Si,⁷² but it can be reached readily during a filament formation or filament rupture. Indeed, it has been shown that the barrier functionality of Ru film fails completely at ∼700°C irrespective of the Ru layer thickness.⁷²

In conclusion, applying above observations to the circumstances encountered in our devices, it is plausible to argue that the insertion of the TaO_x layer between SiO₂ and Ti leads to a suppression of Si diffusion which leads, in turn, to a suppression of Ru and Cu silicide reactions. Since Ti is known to be an aggressive oxygen getter, the TiO_x at the interface of SiO₂ frees up Si atoms that can then readily diffuse into the Ru layer. The insertion of TaO_x layer between Ti and SiO₂ appears to act as an effective Si diffusion barrier.

To find a tangible evidence for the compromised inertness of the Ru we have considered direct physical analysis techniques SIMS and XRD. However, Secondary Ion Mass Spectroscopy (SIMS) is inadequate to measure copper or oxygen content in Ru electrode. The number of Cu atoms in a filament at the base of the cone interfacing the inert electrode can be estimated to be 50 atoms with a base of 50 nm². If of these atoms 1/10 or less would diffuse into Ru, most likely along the grain boundaries, it would result in a concentration of Cu atoms in Ru of at most 10¹² cm⁻³ in a tiny volume of less than 1000 nm³ with a density of the matrix material of typically 10²² cm⁻³, while other neighboring volumes with still much lower concentration. Such tiny average concentrations of copper or oxygen are well below the sensitivity of SIMS which is typically at most 10¹⁴ cm⁻³. Moreover, the matrix material, here the Ru electrode, is only 50 nm thick and typically 10 μm wide. SIMS cannot work on such thin and narrow layers because the sputtering yield would be far insufficient. Usually the SIMS crater when using Cs or O ions as the bombarding species is several hundred microns wide.

A more promising physical analysis technique is X Ray Diffraction (XRD) which is presented next.

We have performed extensive XRD studies on the substrates for Ru devices A and B, see Fig. 1 i.e. Cu(25nm)/Ru(50nm)/Ti(40nm)/SiO₂(611nm)/Si for device A (the same substrate as for the Pt device) and, Cu(25nm)/Ru(50nm)/Ti(40nm)/TaO_x(27nm)/SiO₂(611nm)/Si for Ru device B. In order to extract the significant XRD signals, we have manufactured following auxiliary samples 1) Si (Si wafer with native oxide), 2) SiO₂(611nm)/Si, 3) Ti(20nm)/Si, 4) TaO_x(27nm)/Si, 5) Ru(50nm)/Si, 6) Cu(25nm)/Si, 7)Cu(25nm)/SiO₂(611nm)/Si, 8) Ru(50nm)/SiO₂(611nm)/Si, 9) TaO_x(27nm)/SiO₂(611nm)/Si, 10) Ti(40nm)/SiO₂(611nm)/Si 11) Ti(20nm)/TaO_x(27nm) /SiO₂(611nm)/Si, and 12) Ru(50nm)/Ti(40nm)/SiO₂(611nm)/Si. The auxiliary samples serve the purpose to track the XRD peaks with added layers to the existing stack, and help identify background signals in order to extract the relevant signals pertinent to Ru devices. The XRD measurements have been performed on the aforementioned samples for three annealing conditions: i) unannealed, i.e. at 300K, ii) annealed at 600°C for 10 min, and iii) annealed at 900°C for 30 min. While an extensive study of this XRD study will be published elsewhere,⁷⁶ the results relevant to Ru devices A and B are summarized below. We observe a Ru crystallization peak at 42.15° of the XRD spectra on all samples that display a Ru layer after the samples at room temperature and after an anneal at 600°C, and 900°C. This is in agreement with XRD studies of similar layer systems reported in Ref. 77. XRD signals for samples of Ru/Si at 300K, Ru/Si after an RTP anneal at 600° C for 10 min, and for Ru/SiO₂/Si after an anneal at 600°C for 10 are shown in Fig. 19. It is seen that a Ru₃Si₂ crystallization peak appears at 600°C. We observe also on some samples an XRD peak (or shoulder observed next to a broad Si wafer peak) of ruthenium silicide Ru₂Si₃ at 44.7° only after 900°C anneal as shown in Fig. 19. This is in agreement with observations of Ru₃Si₂ XRD detection made in Ref. 68. We observe this peak on the following samples: Ru/Si, Ru/SiO₂/Si, Ru/Ti/SiO₂/Si, and Cu/Ru/Ti/SiO₂/Si. The strongest Ru₃Si₂ signal can be seen for the Ru/Si and Ru/SiO₂ samples. From comparison with Fig. 19 it can be seen that the XRD signal at 44.7° is absent after an anneal at 600°C and also for unannealed sample Ru/SiO₂/Si. However, for samples Ru/Ti/TaO_x/SiO₂/Si and Cu/Ru/Ti/TaO_x/SiO₂/Si the Ru₃Si₂ shoulder is either not there or very weak, as shown in Fig. 20. From Fig. 19 it is also seen that the Ru₃Si₂ peak at 41.7° disappears at 900°C, again in agreement with observations made in Ref. 77. We conclude that TaO_x layer inserted between SiO₂ and Ti acts as a Si diffusion barrier, and hence suppresses the Ru silicidation reaction.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

We do not observe any XRD peaks related to copper silicide reaction. The lack of a clear Cu₃Si signal may not necessarily eliminate this reaction altogether, but points, at least, to scarcity of the compound which could be formed along the Ru grain boundaries which is a confirmation of our assessment of Cu detectability by SIMS discussed before.

As Ru is easy oxidizable even in ambient atmosphere we were looking for a RuO_x XRD signal. It is well known that RuO₂ has characteristics XRD peaks at 2θ = 30°, 35°.^78,79 No RuO_x signal could be detected even on free Ru surface samples annealed at 600°C and 900°C in our samples. No direct oxidation of Ru at an elevated temperature in a furnace has been performed as RuO_x was not planned for part of the manufactured layer system shown in Fig. 1b and Fig. 1c. The lack of an XRD signal does not per se preclude existence of a native oxide. But it indicates that either the RuO_x is amorphous or that there is an insufficient quantity of crystalline RuO_x to produce an XRD signal.

The XRD study confirms that the inertness of the Ru electrode is impaired by Ru crystallization at temperatures around 600°C and due to Ru silicidation at 900°C. The XRD study corroborates not only the difference in electrical performance between Pt and Ru device but also among the Ru devices embedded on different substrates.