Irregular Resistive Switching Behaviors of Al2O3-Based Resistor with Cu Electrode

In this work, we examined the irregular resistive switching behaviors of a complementary metal–oxide–semiconductor (CMOS)-compatible Cu/Al2O3/Si resistor device. X-ray photoelectron spectroscopy (XPS) analysis confirmed the chemical and material compositions of a Al2O3 thin film layer and Si substrate. Bipolar resistive switching occurred in a more stable manner than the unipolar resistive switching in the device did. Five cells were verified over 50 endurance cycles in terms of bipolar resistive switching, and a good retention was confirmed for 10,000 s in the high-resistance state (HRS) and the low-resistance state (LRS). Both high reset current (~10 mA) and low reset current (<100 μA) coexisted in the bipolar resistive switching. We investigated nonideal resistive switching behaviors such as negative-set and current overshoot, which could lead to resistive switching failure.


Introduction
Since significant memory characteristics were first reported in metal oxides such as NiO and TiO 2 in 2005, resistive random-access memory (RRAM) memory has been extensively studied [1][2][3][4]. Early studies were conducted with a focus on unipolar resistive switching (URS) [1][2][3][4]. Among many metal oxides, NiO has been shown to have the most stable URS properties [1]. URS has the advantage that a diode can be used as a selection element in a cross-point structure because switching occurs at one polarity [5]. For a unipolar-type memory device, the first method that was introduced involved forming a diode through a high-temperature process, and then depositing a memory device on it. However, in the case of this structure, the scalability of stacking into two layers and four layers is insufficient due to the high-temperature process of the selection device. Moreover, URS has a disadvantage in that the reset current is too high for joule heating, or that the variation in switching parameters, such as the set and reset voltage, and high-resistance state (HRS) and low-resistance state (LRS) in the cycle-to-cycle and cell-to-cell, is large [1]. On the other hand, the phenomenon of bipolar resistive switching (BRS) in metal oxides has also been reported [6][7][8][9][10][11][12][13][14][15][16][17]. The set and reset processes occur at opposite polarities. The reset process is mainly induced by the electric field. The valence change model and the conductive bridge model are the most common bipolar switching types, and these are determined by the metal top electrode. Cu and Ag can diffuse through the insulating layer, and the conducting filament is formed by the cation ion [18][19][20][21]. On the other hand, the oxygen vacancies are modulated in the oxide and the conductance can be changed by the applied voltage [22]. It should be noted that HfO 2 , Al 2 O 3 , and Ta 2 O 5 achieved much more stable resistive switching behaviors [23][24][25][26] and higher endurances than the URS did. Ovonic threshold switching (OTS) [27] has recently become the most competitive selector element that provides bidirectional nonlinear function for RRAM. Therefore, BRS is now a more dominant research area in the RRAM community than URS is.
Conductive bridge memory has the advantages of a fast switching speed and lowcurrent operation, but unstable resistive switching can occur due to the current overshoot and negative-set behavior [28].
In this work, a Cu/Al 2 O 3 /Si device was fabricated and its resistive switching behaviors were investigated. The Al 2 O 3 /Si stack was verified by XPS analysis; the analysis confirmed similar I-V characteristics with bipolar resistive switching from several cells. The basic memory device properties, such as cycle-to-cycle variation, endurance, and retention, were tested. Moreover, nonideal resistive switching behaviors like the current overshoot and negative-set behavior were studied as well. Finally, we demonstrated the resistive switching by pulses in nonideal cases.

Materials and Methods
The Cu/Al 2 O 3 /Si memory device was fabricated using the following process: First, a highly doped Si substrate as the bottom electrode was formed by ion-implantation, wherein the dose was 5 × 10 15 and the energy was 40 keV. Next, a 3.5 nm-thick Al 2 O 3 resistive switching layer was deposited using an atomic layer deposition (ALD) system using precursors of Al(CH 3 ) 3 (TMA) and O 3 at a chamber temperature of 350 • C and a pressure of~1 Torr. A 100 nm-thick Cu top electrode was patterned using a shadow mask including circular patterns with a diameter of 100 µm and was deposited by a thermal evaporator. Here, Ag paste was applied over the Si for easy tipping. The electrical properties of the DC I-V curves were measured using a Keithley 4200-SCS semiconductor parameter analyzer (SPA). The transient characteristics were measured in pulse mode using a 4225-PMU ultrafast module (Keithley Instruments, Cleveland, OH, USA). During the DC and pulse measurements, a bias voltage and pulse were applied to the Cu layer while maintaining the ground on the Ag that was directly connected to the Si bottom electrode. XPS depth analysis was conducted using a Nexsa (Thermo Fisher Scientific, Waltham, MA, USA) with a Microfocus monochromatic X-ray source (Al-Kα (1486.6 eV)), a sputter source (Ar + ), an ion energy of 1 kV, a sputter rate of 0.3 nm/s for SiO 2 , and a beam size of 100 µm. Figure 1a illustrates a schematic of the Cu/Al 2 O 3 /Si stack examined in this work. We verified the material and chemical compositions of the Al 2 O 3 /Si stack before the electrical measurement characteristics. Figure 1b,c show the XPS spectra, including raw data and fitting curves, of Al 2p and Si 2p, respectively. Al 2p was the signal detected from the Al 2 O 3 layer and Si 2p was the signal detected from the substrate. The peak was centered at 76.05 eV in Al 2p, which corresponds to the Al-O bonds [36]. The peak of Si was centered at 99.41 eV, which indicates that the Si peak was from the Si-Si bonds [37]. Conductive bridge memory has the advantages of a fast switching speed and lowcurrent operation, but unstable resistive switching can occur due to the current overshoot and negative-set behavior [28].
In this work, a Cu/Al2O3/Si device was fabricated and its resistive switching behaviors were investigated. The Al2O3/Si stack was verified by XPS analysis; the analysis confirmed similar I-V characteristics with bipolar resistive switching from several cells. The basic memory device properties, such as cycle-to-cycle variation, endurance, and retention, were tested. Moreover, nonideal resistive switching behaviors like the current overshoot and negative-set behavior were studied as well. Finally, we demonstrated the resistive switching by pulses in nonideal cases.

Materials and Methods
The Cu/Al2O3/Si memory device was fabricated using the following process: First, a highly doped Si substrate as the bottom electrode was formed by ion-implantation, wherein the dose was 5 × 10 15 and the energy was 40 keV. Next, a 3.5 nm-thick Al2O3 resistive switching layer was deposited using an atomic layer deposition (ALD) system using precursors of Al(CH3)3 (TMA) and O3 at a chamber temperature of 350 °C and a pressure of ~1 Torr. A 100 nm-thick Cu top electrode was patterned using a shadow mask including circular patterns with a diameter of 100 μm and was deposited by a thermal evaporator. Here, Ag paste was applied over the Si for easy tipping. The electrical properties of the DC I-V curves were measured using a Keithley 4200-SCS semiconductor parameter analyzer (SPA). The transient characteristics were measured in pulse mode using a 4225-PMU ultrafast module (Keithley Instruments, Cleveland, OH, USA). During the DC and pulse measurements, a bias voltage and pulse were applied to the Cu layer while maintaining the ground on the Ag that was directly connected to the Si bottom electrode. XPS depth analysis was conducted using a Nexsa (Thermo Fisher Scientific, Waltham, MA, USA) with a Microfocus monochromatic X-ray source (Al-Kα (1486.6 eV)), a sputter source (Ar + ), an ion energy of 1 kV, a sputter rate of 0.3 nm/s for SiO2, and a beam size of 100 µ m. Figure 1a illustrates a schematic of the Cu/Al2O3/Si stack examined in this work. We verified the material and chemical compositions of the Al2O3/Si stack before the electrical measurement characteristics. Figure 1b,c show the XPS spectra, including raw data and fitting curves, of Al 2p and Si 2p, respectively. Al 2p was the signal detected from the Al2O3 layer and Si 2p was the signal detected from the substrate. The peak was centered at 76.05 eV in Al 2p, which corresponds to the Al-O bonds [36]. The peak of Si was centered at 99.41 eV, which indicates that the Si peak was from the Si-Si bonds [37].   Figure 2a shows the current-voltage (I-V) characteristics, including the forming, set, and reset processes, for the Cu/Al 2 O 3 /Si device. The forming process activated the device for resistive switching. Here, the compliance current of 100 µA was used to ensure that the device was not destroyed during a current surge. The forming process induces oxygen vacancies and the Cu filament in the Al 2 O 3 layer, resulting in a decrease in resistance. It is well known that the oxygen vacancies, as well as the Cu filament, can contribute to the increase in conductance [38].

Results and Discussion
tions in LRS occurred from the current overshoot during the forming and set processes. The LRS current could be controlled well even though the same compliance current was applied to the device during the cycle. Figure 2d shows the endurance cycle in which the LRS and HRS resistances were extracted at a read voltage of 0.2 V. The bouncing point did not show a certain tendency, and the LRS and HRS could be divided more than 10 times, even in the worst case. The I-V curves and endurance properties were similar in the other four cells in Figure S1. Figure 2e shows the retention properties of HRS and LRS; HRS was run for 10,000 s. HRS slightly fluctuated over time, but the on/off ratio was sufficiently maintained, indicating that the Cu/Al2O3/Si device is suitable for nonvolatile memory application. Next, the reset process occurred with a negative bias sweep, wherein the resistance increases by reducing the oxygen vacancies and the size of the Cu filament in the insulator. Meanwhile, the set process occurred by sweeping in a positive direction to decrease the resistance again. In this way, the set and reset processes can be performed repeatedly. Figure 2b shows the statistical distribution, including the forming, set, and reset voltages. The average forming voltage of 2.95 V was slightly higher than the set voltage of 2.59 V. This is because there were few oxygen vacancies in the initial state, so achieving a soft breakdown of the device by the forming process requires a larger voltage than doing so by the set process. The average reset voltage was −0.998 V. There were also cycles in which the current was reduced by several steps. Therefore, the reset voltage is based on the point at which the first current decreased. Figure 2c shows the cumulative probability of the low-resistance state (LRS) and the high-resistance state (HRS) in the cycle-to-cycle. Variations in LRS occurred from the current overshoot during the forming and set processes. The LRS current could be controlled well even though the same compliance current was applied to the device during the cycle. Figure 2d shows the endurance cycle in which the LRS and HRS resistances were extracted at a read voltage of 0.2 V. The bouncing point did not show a certain tendency, and the LRS and HRS could be divided more than 10 times, even in the worst case. The I-V curves and endurance properties were similar in the other four cells in Figure S1. Figure 2e shows the retention properties of HRS and LRS; HRS was run for 10,000 s. HRS slightly fluctuated over time, but the on/off ratio was sufficiently maintained, indicating that the Cu/Al 2 O 3 /Si device is suitable for nonvolatile memory application.
Next, we closely investigated the I-V curves for nonideal resistive switching. High-LRS currents and low-LRS currents are classified by two typical curves (Figure 3a). The high current by the current overshoot can be the nonideal case. The reset did not occur in some instances of switching, due to the overgrowth of the conducting filament caused by the Cu filament and oxygen vacancies. A reset occurred on the return sweep voltage, as shown in Figure 3b. However, if only a single sweep had been applied, the state would have been fixed at LRS. The reset process for the high current could occur by Joule heating. Figure 3c shows the unipolar switching of the Cu/Al 2 O 3 /Si device in which the set process and the reset process occurred at the same polarity, as shown in Figure 3c. The unipolar and bipolar resistive switching are both observed in Al 2 O 3 -based RRAM in the literature. The reset current in the unipolar resistive switching is higher than that in the bipolar resistive switching. Our device showed similar results [39]. The results confirm that the high LRS current in the bipolar type was similar to the LRS current in the unipolar type using log-log fitting (Figure 3d). The two cases had a similar slope of 1 following Ohmic conduction. This indicates that a strong conducting filament could be formed in the two cases. On the other hand, the low LRS current in the bipolar type had two distinctive slopes following the space-charge-limited current (SCLC) mechanism. Ohmic conduction occurs in the low-voltage regime in which thermally generated carriers are more dominant than the injected carrier. The higher voltage regime can be explained by trap-limited SCLC in which the total current is determined by the injected carriers.   The other case of nonideal switching is negative-set behavior. Figure 4a shows I-V curves of the normal reset (black curve) and the breakdown after reset (red curve). The negative-set behavior can induce the resistive switching failure. The repeated breakdown can cause a permanent breakdown, after which it is stuck in LRS. It is impossible to return to HRS by an additional reset process after negative-set behavior. The possible filament evolution of negative-set behavior is illustrated in Figure 4b-d. LRS was completed by the set process in which the Cu filament was formed by the diffusion of Cu ions from the top electrode (Figure 4b). The conducting filament was ruptured by the reset process, causing the device to change from LRS to HRS (Figure 4c). The current increased as a result of connecting the filament from the movement of the Cu ion when a larger voltage was applied. Here, it was difficult to stop the voltage before negative-set, due to the variation in reset voltage. Similar negative-set behaviors of the conductive bridge randomaccess memory (CBRAM) type have been reported in the Cu/ZrO 2 /Pt [40] device and the Ag/ZrO 2 /Pt device [41]. Finally, we demonstrated the pulse operation for the set and reset processes without negative-set behavior. Figure 5a

Conclusions
In summary, we studied the irregular resistive switching of a Cu/Al2O3/Si device, which can cause reliability issues. The Al2O3 layer and Si substrate were detected using an XPS depth profile. The cell-to-cell cycle and device-to-device characteristics were examined by scanning DC I-V curves. Due to the thin Al2O3, there was not much difference between the forming voltage and the set voltage. The variation in LRS was caused by the current overshoot in the set process. The variation in HRS occurred as a result of the uneven reset process. The LRS and HRS were maintained for 10,000 s. The uncontrolled high LRS current was similar to the LRS current in unipolar resistive switching, which was verified by the current-voltage fitting. Negative-reset behavior was confirmed, and a possible Cu filament model was presented. Finally, the increase and decrease in the current were performed by pulses without nonideal effects.

Conclusions
In summary, we studied the irregular resistive switching of a Cu/Al 2 O 3 /Si device, which can cause reliability issues. The Al 2 O 3 layer and Si substrate were detected using an XPS depth profile. The cell-to-cell cycle and device-to-device characteristics were examined by scanning DC I-V curves. Due to the thin Al 2 O 3 , there was not much difference between the forming voltage and the set voltage. The variation in LRS was caused by the current overshoot in the set process. The variation in HRS occurred as a result of the uneven reset process. The LRS and HRS were maintained for 10,000 s. The uncontrolled high LRS current was similar to the LRS current in unipolar resistive switching, which was verified by the current-voltage fitting. Negative-reset behavior was confirmed, and a possible Cu filament model was presented. Finally, the increase and decrease in the current were performed by pulses without nonideal effects.