Multi-Level Switching of Al-Doped HfO 2 RRAM with a Single Voltage Amplitude Set Pulse

: In this paper, the resistive switching characteristics in a Ti/HfO 2 : Al/Pt sandwiched structure are investigated for gradual conductance tuning inherent functions. The variation in conductance of the device under different amplitudes and voltage pulse widths is studied. At the same time, it was found that the variation in switching parameters in resistive random-access memory (RRAM) under impulse response is impacted by the initial conductance states. The device was brought to a preset resistance value range by energizing a single voltage amplitude pulse with a different number of periodicities. This is an efﬁcient and simple programming algorithm to simulate the strength change observed in biological synapses. It exhibited an on/off of about 100, an endurance of over 500 cycles, and a lifetime (at 85 ◦ C) of around 10 5 s. This multi-level switching two-terminal device can be used for neuromorphic applications to simulate the gradual potentiation (increasing conductance) and inhibition (decreasing conductance) in an artiﬁcial synapse.


Introduction
Resistive switching (RS) devices were originally conceived for memory applications [1,2] because of the advantages compared to traditional charges-based memories, such as their superior speed, low operating voltage, excellent scalability [3], and non-volatile and multi-level data storage abilities [4]. Currently, multi-level resistive random-access memory (RRAM) devices can be widely used for data storage [5,6], logical calculation [7,8], bionic neural network systems [9,10], and so on. They have also been studied as artificial synapses for applications in neuromorphic computation, featuring high parallelism [11][12][13] and the ability to perform adaptive learning [14]. RRAM devices can solve problems that the traditional Von Neumann paradigm is not able to deal with, such as incomplete defined inputs, as in real-world scenarios [15], or the implementation of adaptive learning algorithms [14].
Great efforts have been made to optimize both materials and programming techniques to design a device that can simulate biological synaptic behavior, i.e., gradually increase and decrease its conductance under suitable voltage excitation [13,15,16]. In order to facilitate the practical applications of multi-level RRAM devices, it is necessary to consider the materials of complementary metal oxide silicon (CMOS)-compatible devices. HfO 2 is used as a high-k gate dielectric in CMOS devices because of its high reliability, fast operating speed, and low power consumption [17,18]. Currently, analog switching with multi-level states and reliability is not available for RRAM. Therefore, researchers have come up with a variety of solutions, including improving the manufacturing process (doping in the HfO 2 layers [19][20][21][22][23], modulation of oxide/metal interfaces [9,24], post-deposition annealing [19], etc.) and optimizing the measurement conditions. For electrode materials, Hong et al. [9] suggested that the use of a combination of inert and active metals, such as Ti/Pt, could improve the reliability of the device. Traore et al. [24] found that a Ti/HfO 2 interface facilitates the formation of O defects from the interface region into Ti and suggest the interface region to be the precursor for defect formation. Previous studies [20][21][22][23] have shown that the doped HfO 2 RRAM devices are more likely to realize stable intermediateresistance states (IRS) together with high-resistance states (HRS) and low-resistance states (LRS) compared with pure HfO 2 devices. It was attributed to controllable oxygen vacancies (Vo)-based filament formation along the Al atoms diffused into the HfO 2 resistive switching film [21]. Wang et al. [4], Gao et al. [25], and Covi et al. [20] prepared multi-level RRAM by doping HfO 2 with Cu, Gd, and Al, respectively. However, they can only carry out a four-level storage state [4,25]. According to Zhou et al. [26], TiN/Al 2 O 3 /HfO 2 /Pt bilayer RRAM devices show seven analog reset states for both current scanning and voltage scanning operations with an on/off ratio of around 10. Woo et al. [27] explained that RRAM devices with an AlO x /HfO x bi-layer structure present 12 distinct analog multi-levels with on/off ratio over 10. Moreover, the different conductance states were achieved using the application of voltage pulses with different amplitudes [12,26,28] or widths [10,13,28] or compliance currents of different flexibilities [4] during the set program process. This will increase the operational difficulty in the above applications. Few scholars have realized that the device can reach a certain conductivity range accurately through a train of identical pulses [29], which has a simple structure that facilitates a highly integrated implementation and is therefore best suited for neuromorphic applications. In addition, it was reported that the reset operation usually features a gradual change in the device conductance. The resistive switching is inherently stochastic due to the randomness of the generation and the migration of oxygen vacancy based on the Kinetic Monte Carlo simulation proposed by Yu et al. [29]. The sudden change in conductivity in the set process requires an additional current compliance [12,15,30,31] to prevent device damage, which is the problem that must be overcome to realize the above scheme.
In this context, an Al-doped HfO 2 -based memristor is presented as a suitable device for an artificial synapse. We previously studied the mechanism of multi-stage resistance of a HfO 2 : Al memristor [23]. This time, by studying different switching dynamics in the pulsed regime, a scheme was obtained through which the device can be stabilized to the desired conductivity range by using a series of identical pulses. Firstly, the device was obtained by cyclic growth of HfO 2 and Al 2 O 3 in a certain proportion by an atomic layer deposition (ALD) reactor. Then, the resistive switching characteristics of the device were analyzed by direct current (DC) scanning and voltage pulses. Finally, by matching the appropriate pulse and initial conductance values, the device conductance could be gradually changed stepwise with the set pulses.

Materials and Methods
The samples studied here are Ti/HfO 2 : Al/Pt (Figure 1a) RRAM devices. Ti/Pt layers of 120 nm were successively deposited by sputtering at room temperature on a Si/SiO 2 substrate. Ti acts as the adhesive layer between the bottom electrode (BE) of Pt and the substrate. After this, a 20-nm HfO 2 layer was deposited on Pt/Ti/SiO 2 -coated silicon substrates in an ALD reactor at 300 • C [32]. Al-doped films were obtained by depositing one cycle of Al 2 O 3 for every 33 cycles of HfO 2 . The thicknesses of the thin films were measured using a spectroscopic ellipsometer (J.A. Woollam Co.SE M2000U, Lincoln, NE, USA), and the measured result was about 19.73 nm. Finally, a Ti layer of 50 nm acting as a top electrode (TE) and a Pt layer of 100 nm acting as covering layers were deposited by DC sputtering and patterned to form a 100 µm × 100 µm isolated square (Figure 1b). The Al atomic concentration was 1.81%. According to the chemical binding energy analysis of the film by X-ray photoelectron spectroscopy of the Theta 300 XPS system, the atomic concentration of Al was 1.81%. Figure 1c,d show the spectra of Al-doped and non-doped devices. The spectra were calibrated by C 1 s (284.6 eV) peak. For the HfO 2 film, the peak of Hf 4f can be fitted as a double peak. Through fitting, it can be seen that the ratio of Hf 4f 5/2 to Hf 4f 7/2 increased for the doped devices. Moreover, Al-doped devices show the shift of the Hf 4f peak to the higher binding energy and a distinct 74.3 eV peak of Al 2p. 4f7/2 increased for the doped devices. Moreover, Al-doped devices show the shift of the Hf 4f peak to the higher binding energy and a distinct 74.3 eV peak of Al 2p. The microscopic optical image of the top electrode of the device. We prepared devices with different areas ranging from 10 μm × 10 μm to 300 μm × 300 μm. In our work, we only discussed the case of 100 μm × 100 μm. XPS spectra of (c) Hf 4f and (d) Al 2p in Al-doped and non-doped devices.
Electrical measurements were performed using Agilent B1500A (Agilent, Santa Clara, CA, USA). The Source Measurement Unit (SMU) was used to perform the DC sweeps, and the Waveform Generator/Fast Measurement Unit (WGFMU) was used to perform pulse signals generation and measurement. The signal needle of the two modules contacted the TE of the device simultaneously, and the working module was selected at the operation interface of the testing equipment [25]. Then, the selected module generated corresponding electrical signals that were applied to the device. The bottom electrode was always grounded. Figure 2 shows typical DC I-V bipolar switching curves of Al-doped HfO2-based RRAM devices. In the set/reset process, multi-stage conductance was obtained by changing the compliance current and the stop voltage, respectively. It was found that both the set (off-to-on transition) process and the reset (on-to-off transition) process are gradual. This is because Al dopants reduce the formation energy of oxygen vacancies (Vo) and Vo are formed scattered near the dopants [33]. Therefore, multiple filaments are formed in the set process based on the uniform distributed Vo in the oxide layer [9,21,34]. It can also be seen that the set voltage increases obviously with the increase in the reset stop voltage. It shows that with the increase in the reset stop voltage, the more oxygen vacancies in the oxide layer are consumed, the more serious the damage to the conductive filaments. At this point, the set process requires a larger forward voltage to reform the conductive filaments in the oxide layer. In contrast, the reset voltage does not obviously change with the increase in the set limiting current. In the reset process, regardless of the original resistance, the conductive filament will break when the reverse voltage applied to the TE of the device reaches a certain value. This shows that the difference in conductive filaments in different low-resistance states is not due to the thickness of filaments, but the number of filaments [21,35]. The microscopic optical image of the top electrode of the device. We prepared devices with different areas ranging from 10 µm × 10 µm to 300 µm × 300 µm. In our work, we only discussed the case of 100 µm × 100 µm. XPS spectra of (c) Hf 4f and (d) Al 2p in Al-doped and non-doped devices.

Direct-Current Characteristics
Electrical measurements were performed using Agilent B1500A (Agilent, Santa Clara, CA, USA). The Source Measurement Unit (SMU) was used to perform the DC sweeps, and the Waveform Generator/Fast Measurement Unit (WGFMU) was used to perform pulse signals generation and measurement. The signal needle of the two modules contacted the TE of the device simultaneously, and the working module was selected at the operation interface of the testing equipment [25]. Then, the selected module generated corresponding electrical signals that were applied to the device. The bottom electrode was always grounded. Figure 2 shows typical DC I-V bipolar switching curves of Al-doped HfO 2 -based RRAM devices. In the set/reset process, multi-stage conductance was obtained by changing the compliance current and the stop voltage, respectively. It was found that both the set (off-to-on transition) process and the reset (on-to-off transition) process are gradual. This is because Al dopants reduce the formation energy of oxygen vacancies (Vo) and Vo are formed scattered near the dopants [33]. Therefore, multiple filaments are formed in the set process based on the uniform distributed Vo in the oxide layer [9,21,34]. It can also be seen that the set voltage increases obviously with the increase in the reset stop voltage. It shows that with the increase in the reset stop voltage, the more oxygen vacancies in the oxide layer are consumed, the more serious the damage to the conductive filaments. At this point, the set process requires a larger forward voltage to reform the conductive filaments in the oxide layer. In contrast, the reset voltage does not obviously change with the increase in the set limiting current. In the reset process, regardless of the original resistance, the conductive filament will break when the reverse voltage applied to the TE of the device reaches a certain value. This shows that the difference in conductive filaments in different low-resistance states is not due to the thickness of filaments, but the number of filaments [21,35].  Figure 2c shows the DC I-V characteristic curve during gradually increasing compliance current in the set process. It can be seen that the device has excellent resistance retention characteristics in the set process and the resistance changes gradually. Combined with Figure 2a, the resistance of the device after the set process of the DC sweep is mainly determined by the compliance current. The multi-step switching effect is demonstrated by increasing the current compliance step by step [36]. The Vo generated by the reapplied positive bias voltage forms more conductive filaments as the limiting current increases. Therefore, the set process may present as a gradual change in resistance. Figure 2d shows the DC I-V curve for the first 256 cycles of the device. It can be seen that the set/reset voltage barely changed, whereas the switching window is slightly reduced. This shows the excellent cyclic repeatability of the device.

Pulse Characteristic
Currently, we turn our attention toward the impulse response of the devices using WGFMU, which is the central exploration of this work. First, the device was set and reset using a square-waveform voltage pulse with a width of 100 μs. Figure 3 shows the resistance window (HRS/LRS for set is Rinitial/Rafter, and for reset is Rafter/ Rinitial) variation by applying a single voltage pulse of different amplitudes to a device with the same initial resistance. It can be seen from the figure that when the pulse width is constant, the increase in the voltage amplitude will cause a great variation in the conductance [28]. Both set and reset have a threshold bias level of around Vth ~ 0.45 V and Vth ~ -0.6 V, respectively. When the set voltage reaches 0.6 V, it is limited by current limitation, and the change of resistance is no longer obvious.  Figure 2c shows the DC I-V characteristic curve during gradually increasing compliance current in the set process. It can be seen that the device has excellent resistance retention characteristics in the set process and the resistance changes gradually. Combined with Figure 2a, the resistance of the device after the set process of the DC sweep is mainly determined by the compliance current. The multi-step switching effect is demonstrated by increasing the current compliance step by step [36]. The Vo generated by the reapplied positive bias voltage forms more conductive filaments as the limiting current increases. Therefore, the set process may present as a gradual change in resistance. Figure 2d shows the DC I-V curve for the first 256 cycles of the device. It can be seen that the set/reset voltage barely changed, whereas the switching window is slightly reduced. This shows the excellent cyclic repeatability of the device.

Pulse Characteristic
Currently, we turn our attention toward the impulse response of the devices using WGFMU, which is the central exploration of this work. First, the device was set and reset using a square-waveform voltage pulse with a width of 100 µs. Figure 3 shows the resistance window (HRS/LRS for set is R initial /R after , and for reset is R after /R initial ) variation by applying a single voltage pulse of different amplitudes to a device with the same initial resistance. It can be seen from the figure that when the pulse width is constant, the increase in the voltage amplitude will cause a great variation in the conductance [28]. Both set and reset have a threshold bias level of around Vth~0.45 V and Vth~−0.6 V, respectively. When the set voltage reaches 0.6 V, it is limited by current limitation, and the change of resistance is no longer obvious.  The typical dynamic current responses of the voltage pulse are shown in Figure 4. According to the reset process, it can be seen from Figure 4b that the response current of the device suddenly decreases under the impulse voltage. The results of the same reset voltage pulse applied in different resistance states of devices are also compared in the figure. The original resistance state of the device cannot change the mutation of the reset process, but only the time of the reset process. Combined with the above DC characteristics, it is shown that the thickness of several conductive filaments in the oxide layer is the same, and the fracture occurs simultaneously in the reset process. The gradual change in the conductance value during the reset process in the DC sweep is because the bias value applied to the device increases point by point. Multiple conductive filaments will gradually break from various relatively thin places until all the conductive filaments are broken. The conductance modulation of the reset process is suitable for applying voltage pulses with different amplitudes.
Compared to the reset process, the multi-stage set process is easier to implement. It can be observed from Figure 4a that the transient current of the device increases gradually under continuous constant voltage. This indicates that the change in resistance of the device during the set process is gradual. This, in combination with Figure 2a above, led to the conclusion that the resistance change process is the process through which the conductive filaments are formed in the oxide layer one by one. With the gradual switch performance of the memory unit, multi-bit data storage can be easily implemented under an optimized switching operation scheme. Thus, we utilized the gradual set process to obtain multi-stage resistance states.  The typical dynamic current responses of the voltage pulse are shown in Figure 4. According to the reset process, it can be seen from Figure 4b that the response current of the device suddenly decreases under the impulse voltage. The results of the same reset voltage pulse applied in different resistance states of devices are also compared in the figure. The original resistance state of the device cannot change the mutation of the reset process, but only the time of the reset process. Combined with the above DC characteristics, it is shown that the thickness of several conductive filaments in the oxide layer is the same, and the fracture occurs simultaneously in the reset process. The gradual change in the conductance value during the reset process in the DC sweep is because the bias value applied to the device increases point by point. Multiple conductive filaments will gradually break from various relatively thin places until all the conductive filaments are broken. The conductance modulation of the reset process is suitable for applying voltage pulses with different amplitudes. The typical dynamic current responses of the voltage pulse are shown in Figure 4. According to the reset process, it can be seen from Figure 4b that the response current of the device suddenly decreases under the impulse voltage. The results of the same reset voltage pulse applied in different resistance states of devices are also compared in the figure. The original resistance state of the device cannot change the mutation of the reset process, but only the time of the reset process. Combined with the above DC characteristics, it is shown that the thickness of several conductive filaments in the oxide layer is the same, and the fracture occurs simultaneously in the reset process. The gradual change in the conductance value during the reset process in the DC sweep is because the bias value applied to the device increases point by point. Multiple conductive filaments will gradually break from various relatively thin places until all the conductive filaments are broken. The conductance modulation of the reset process is suitable for applying voltage pulses with different amplitudes.
Compared to the reset process, the multi-stage set process is easier to implement. It can be observed from Figure 4a that the transient current of the device increases gradually under continuous constant voltage. This indicates that the change in resistance of the device during the set process is gradual. This, in combination with Figure 2a above, led to the conclusion that the resistance change process is the process through which the conductive filaments are formed in the oxide layer one by one. With the gradual switch performance of the memory unit, multi-bit data storage can be easily implemented under an optimized switching operation scheme. Thus, we utilized the gradual set process to obtain multi-stage resistance states.  Compared to the reset process, the multi-stage set process is easier to implement. It can be observed from Figure 4a that the transient current of the device increases gradually under continuous constant voltage. This indicates that the change in resistance of the device during the set process is gradual. This, in combination with Figure 2a above, led to the conclusion that the resistance change process is the process through which the conductive filaments are formed in the oxide layer one by one. With the gradual switch performance of the memory unit, multi-bit data storage can be easily implemented under an optimized switching operation scheme. Thus, we utilized the gradual set process to obtain multi-stage resistance states.
According to Figure 4a, the current of the set voltage at 0.55 V, 0.6 V, and 0.65 V are all gradients. However, the trends of the current curve are significantly different. According to Figure 3, the larger the set voltage, the larger the resistance window. However, it can be found from Figure 4a that the trend of current change is too fast under an excessive set voltage, which is not conducive to the multi-resistance state. Combining Figures 3 and 4a, when the set voltage is 0.6 V, there is a sufficiently large resistance change window, and the changing trend of current is balanced throughout the set process, which is the most suitable set pulse voltage.
Then, the excitation pulse width was reduced properly. Figure 5 shows the typical dynamic current response of a 0.6 V amplitude, 10 µs pulse width, and square-waveform voltage pulse. It can be seen that different initial resistance values have a large influence on the change in resistance values in the set process [37]. The three HRS listed in figure correspond to the reset stop voltages of −0.9 V (black), −1.0 V (red), and −1.1 V (blue). When the initial resistance is larger, the conductance changes more sharply. This may be the result of the complete destruction of the conducting filaments of the device under the extremely low conductivity [30,38]. This makes the oxygen vacancy more disorganized in the oxide layer. The oxygen vacancy concentration is very high in the local part of the filament structure, which is composed of oxygen vacancies. In the case of equivalent oxygen vacancies, at this point, the distribution of oxygen vacancies throughout the dielectric layer is more uniform than when the partially conducting filament is maintained. The oxygen vacancy concentration in the outer space of the conductive filament in the dielectric layer is much lower than in the higher conductivity state, in which the oxygen vacancy is more disordered. According to the transport model, when the applied electric field intensity is the same, the greater the concentration of ions, the greater the transport flux. Under the excitation of the set pulse, the oxygen vacancy can be recombined more quickly to form conductive filaments that are different from those before the reset process. When the device is in a high configuration, in which the conductive filament has just broken, the conductive filament retains its integrity except for the broken parts. During the set process, the conductive filaments are restored one by one at the break to form the original conductive filament.
Electronics 2021, 10, x FOR PEER REVIEW 6 of 10 reset process. The current range was set to 10 mA to protect the RRAM devices from permanently breaking down during the transient test.
According to Figure 4a, the current of the set voltage at 0.55 V, 0.6 V, and 0.65 V are all gradients. However, the trends of the current curve are significantly different. According to Figure 3, the larger the set voltage, the larger the resistance window. However, it can be found from Figure 4a that the trend of current change is too fast under an excessive set voltage, which is not conducive to the multi-resistance state. Combining Figures 3 and  4a, when the set voltage is 0.6 V, there is a sufficiently large resistance change window, and the changing trend of current is balanced throughout the set process, which is the most suitable set pulse voltage.
Then, the excitation pulse width was reduced properly. Figure 5 shows the typical dynamic current response of a 0.6 V amplitude, 10 μs pulse width, and square-waveform voltage pulse. It can be seen that different initial resistance values have a large influence on the change in resistance values in the set process [37]. The three HRS listed in figure correspond to the reset stop voltages of −0.9 V (black), −1.0 V (red), and −1.1 V (blue). When the initial resistance is larger, the conductance changes more sharply. This may be the result of the complete destruction of the conducting filaments of the device under the extremely low conductivity [30,38]. This makes the oxygen vacancy more disorganized in the oxide layer. The oxygen vacancy concentration is very high in the local part of the filament structure, which is composed of oxygen vacancies. In the case of equivalent oxygen vacancies, at this point, the distribution of oxygen vacancies throughout the dielectric layer is more uniform than when the partially conducting filament is maintained. The oxygen vacancy concentration in the outer space of the conductive filament in the dielectric layer is much lower than in the higher conductivity state, in which the oxygen vacancy is more disordered. According to the transport model, when the applied electric field intensity is the same, the greater the concentration of ions, the greater the transport flux. Under the excitation of the set pulse, the oxygen vacancy can be recombined more quickly to form conductive filaments that are different from those before the reset process. When the device is in a high configuration, in which the conductive filament has just broken, the conductive filament retains its integrity except for the broken parts. During the set process, the conductive filaments are restored one by one at the break to form the original conductive filament.  Therefore, it is particularly important to select and fix the resistance state before applying the pulse in order to obtain a specific stable resistance value in the set process. When the pulse width is 10 µs, the dynamic current of the device reaches saturation at Electronics 2021, 10, 731 7 of 10 about 30-50 pulse periods. This is not conducive to obtaining more different specific resistance values. It is necessary to further reduce the pulse width in the set process.

Multi-Level Switching
Finally, we chose the square-waveform voltage pulse signal with a width of 1 µs and an amplitude of 0.6 V as the set pulse. One cycle included 400 consecutive set pulses. To stabilize the initial resistance value in a certain range before the set process, the DC sweep was adopted to complete the reset process of the device. The reset stop voltage was set to −0.9 V. Device conductance is shown as a function of the applied voltage pulse number for ON switching (from HRS to LRS), as shown in Figure 6a. The conductance was measured at 0.1 V after each pulse. The set training process at several milestone cycles in the endurance test is also shown. No significant degradation was found after 500 cycles. Figure 6b shows the box-plot of the conductance of the device at different pulse periods for 100 different devices. After selecting 12 nodes in the set process of 400 pulses, the increment necessary for the pulse period to reach the next resistance state increases. It can be found that under the first 50 set pulses, the conductance of the device is less uniform, and then tends to be concentrated. Although the upper and lower limits of the conductance of each node overlapped partially, the interquartile ranges did not coincide. The set pulse with only a single voltage amplitude was realized to set the device in more than a dozen different resistance states.
Electronics 2021, 10, x FOR PEER REVIEW 7 of 10 Therefore, it is particularly important to select and fix the resistance state before applying the pulse in order to obtain a specific stable resistance value in the set process. When the pulse width is 10 μs, the dynamic current of the device reaches saturation at about 30-50 pulse periods. This is not conducive to obtaining more different specific resistance values. It is necessary to further reduce the pulse width in the set process.

Multi-Level Switching
Finally, we chose the square-waveform voltage pulse signal with a width of 1 μs and an amplitude of 0.6 V as the set pulse. One cycle included 400 consecutive set pulses. To stabilize the initial resistance value in a certain range before the set process, the DC sweep was adopted to complete the reset process of the device. The reset stop voltage was set to −0.9 V. Device conductance is shown as a function of the applied voltage pulse number for ON switching (from HRS to LRS), as shown in Figure 6a. The conductance was measured at 0.1 V after each pulse. The set training process at several milestone cycles in the endurance test is also shown. No significant degradation was found after 500 cycles. Figure 6b shows the box-plot of the conductance of the device at different pulse periods for 100 different devices. After selecting 12 nodes in the set process of 400 pulses, the increment necessary for the pulse period to reach the next resistance state increases. It can be found that under the first 50 set pulses, the conductance of the device is less uniform, and then tends to be concentrated. Although the upper and lower limits of the conductance of each node overlapped partially, the interquartile ranges did not coincide. The set pulse with only a single voltage amplitude was realized to set the device in more than a dozen different resistance states. The conductance of the device at different pulse periods for 100 different devices; the illustration shows actual data. The conductance was calculated by the Source Measurement Unit (SMU) (Vread = 0.1 V), and the set pulse by 0.6 V amplitude, 1 μs pulse width, and square-waveform voltage pulse (rise time = 50 ns). The 500-cycle endurance test outcome is notable. One cycle includes 400 set consecutive pulses.
For the data retention characteristic, 12 devices were set to preset conductance by respectively applying 1, 2, 5, 10, 20, 40, 70, 110, 160, 220, 300, and 400 set pulses and were kept at 85 °C [39]. Figure 7 shows the retention test of all levels for resistance at 85 °C. The conductance was read every hour with a voltage of 0.1 V. It can be seen that the resistance values of all states were stable and remained unchanged after around 10 5 s. The conductance of the device at different pulse periods for 100 different devices; the illustration shows actual data. The conductance was calculated by the Source Measurement Unit (SMU) (V read = 0.1 V), and the set pulse by 0.6 V amplitude, 1 µs pulse width, and square-waveform voltage pulse (rise time = 50 ns). The 500-cycle endurance test outcome is notable. One cycle includes 400 set consecutive pulses.
For the data retention characteristic, 12 devices were set to preset conductance by respectively applying 1, 2, 5, 10, 20, 40, 70, 110, 160, 220, 300, and 400 set pulses and were kept at 85 • C [39]. Figure 7 shows the retention test of all levels for resistance at 85 • C. The conductance was read every hour with a voltage of 0.1 V. It can be seen that the resistance values of all states were stable and remained unchanged after around 10 5 s.

Conclusions
The HfO2: Al RRAM devices proposed in this paper demonstrated a stable resistance state set/reset performance under DC sweep and gradual set resistance state performance under consecutive voltage pulses. It was found that the conductivity of the devices at high conductivity depends more on the number of conductive filaments than their thickness. The device can stably reach the preset resistance value by applying the same number of pulses in the setting process. Using a single voltage amplitude of pulse throughout the multi-stage resistance adjustment process can simplify the later storage system. These novel devices can be used for realizing complex logic functions. No data were lost over the course of more than a day at 85 °C. The Ti/HfO2: Al/Pt device is a good potential candidate for multi-level storage.

Conclusions
The HfO 2 : Al RRAM devices proposed in this paper demonstrated a stable resistance state set/reset performance under DC sweep and gradual set resistance state performance under consecutive voltage pulses. It was found that the conductivity of the devices at high conductivity depends more on the number of conductive filaments than their thickness. The device can stably reach the preset resistance value by applying the same number of pulses in the setting process. Using a single voltage amplitude of pulse throughout the multi-stage resistance adjustment process can simplify the later storage system. These novel devices can be used for realizing complex logic functions. No data were lost over the course of more than a day at 85 • C. The Ti/HfO 2 : Al/Pt device is a good potential candidate for multi-level storage.