Introduction

In an environment where data demand is rapidly increasing, a breakthrough is needed in computing performance limitations due to serial processing of CPU and memory [1]. It is necessary to change the computing structure and improve the materials of the memory device to solve the memory wall. Neuromorphic computing architecture is emerging as a structural solution to the bottleneck. The neuromorphic computing system mimics the neuron and synapses of the human brain [2,3,4]. This system is suitable for the process of complex and unstructured information. First of all, to implement neuromorphic computing, it is necessary to understand how the human brain processes information. The human brain includes numerous synapses and neurons, and learning and memory of information proceed through parallel chemical interactions. Information processing and memory capabilities vary depending on various factors such as the size, holding time, and a repetition time of external signals and stimuli [5,6,7].

Among various memories, the RRAM exhibits a fast switching speed and a low operating voltage [8,9,10,11,12,13,14,15]. In addition, RRAM could be implemented in a simple structure such as a metal-oxide-metal (MIM) with various structural expandability [16,17,18,19,20,21] such as the connection of transistor with each memory cell, an array structure, and a 3D vertical structure.

The switching of RRAM occurs by the formation and rupture of filament in an insulator between the metals [22,23,24,25,26]. The resistance of RRAM is varied through a conductive filament composed of oxygen vacancy in the insulator existing between the top electrode (TE) and bottom electrode (BE) and has two basic switching states (high and low) to process the data storage process [20, 27,28,29,30]. In the case of the high-resistance state (HRS), a low current flows in HRS, and in the case of the low-resistance state (LRS), it means a state has low resistance and good conductivity. Accordingly, the on/off state could be monitored through the read voltage. The repetition of set and reset processes cause the device to move back and forth between the HRS and LRS, which can be described as a memory that stores 0 and 1 from a digital perspective.

In this paper, the gradual resistive switching is conducted on Pt/Al2O3/TaN device, including Al2O3 high-k dielectric [31,32,33,34,35], which was deposited by atomic layer deposition (ALD) equipment. The characteristics using basic DC current sweep and on/off endurance characteristics were measured, and the suitability of neuromorphic devices was also measured through synaptic measurement, including potentiation, depression, PPD, EPSC, and STDP.

Experiments

Pt/Al2O3/TaN device was fabricated as follows. Firstly, TaN as BE was deposited by the sputtering system on SiO2/Si wafer. A 5-nm-thick Al2O3 film was deposited by the ALD process. In the ALD process, TMA precursors and O3 were used at stage temperature 450 °C. Then a 100-nm-thick Pt as TE was deposited by evaporator in which the top pattern was formed in a circular pattern by using a shadow mask with a diameter of 100 µm. For the measurement environment, all measurements were performed at room temperature and ambient atomic pressure. Electrical data were measured using the Keithley 4200-SCS semiconductor parameter ultrafast module and in pulse mode using a 4225-PMU ultrafast module.

Results and Discussion

Figure 1a shows the schematic illustration of the fabricated Pt/Al2O3/TaN device. In Fig. 1b, the cross section of the Pt/Al2O3/TaN RRAM device is inspected by a transmission electron microscope (TEM). The thickness of the Al2O3 insulator layer deposited by the ALD system is about 5 nm. In Fig. 1c, energy dispersion X-ray spectroscopy (EDS) mapping of each element was performed to investigate possible chemical interactions. EDS mapping shows the spatial distribution of elements in Pt/Al2O3/TaN. EDS maps of Pt, Al, O, Ta, and N elements were collected in the area shown in the electronic image. A region where O and Ta overlap is observed, indicating The TaON interface layer between the Al2O3 insulator and TaN BE is formed by a chemical redox reaction between the TaN BE and the lower Al2O3 layer due to the strong oxygen binding of TaN [36,37,38]. Because of the formation of the TaON interface layer by extracting oxygen from the Al2O3 layer by TaN, better switching characteristics could be exhibited according to the formation of the oxygen vacancy near the TaON/Al2O3 interface [38].

Fig. 1
figure 1

a Schematics image of Pt/Al2O3/TaN device, b cross-sectional TEM image, c EDS mapping images of Pt, Al, O, Ta and N elements collected from the area indicated in the TEM image of the Pt/Al2O3/TaN device

Full size image

In order to confirm the TaON layer, the X-ray photoelectron spectroscopy (XPS) spectra fittings were conducted. Figure 2a shows the Al 2p XPS spectra in which peak intensity is located at 75 eV for Al–O bonding [39]. Figure 2b and c shows Ta 4f and N 1s XPS peak for the TaON layer. In Fig. 2b, small peaks exist at higher binding energy than general Ta 4f XPS peaks. This indicates that the binding Ta–O or Ta–Al energy also affected the Ta 4f XPS peaks with binding Ta–N energy [40, 41]. From Fig. 2c, through combination with oxygen, N 1s XPS peak shows more biased to higher binding energy than the normal N 1s peak [42]. As a result, a thin layer of TaON exists between the Al2O3 insulator and the TaN BE.

Fig. 2
figure 2

XPS spectra of a Al 2p, b Ta 4f and c N 1s of the device

Full size image

Next, we investigate two types of bipolar resistive switching by DC sweep. All of the above I–V characteristics were measured at a step voltage of 0.01 V. Representative feature of this device shows forming-free characteristics in Fig. 3a [43]. The set process has similar I–V curves as the forming process, and the set process occurs at − 2 V or higher, and the reset process is induced by applying a 2.75 V. This is referred to as a deep reset curve. At this time, the on/off ratio is about 45,000 based on the read voltage of 0.5 V, which is a characteristic due to a large band gap of Al2O3. Set shows abrupt behavior, and in the reset process, it shows a curve that returns to the HRS state with a stepwise drop from 1 V or higher to 2.7 V. In the case of Fig. 3b, unlike Fig. 3a, it can be implemented by adjusting reset voltage less than 2.75 V. This is referred to as a partial reset curve, and the on/off ratio at this time is about 13 at the read voltage of 0.5 V. Compared to the I–V curves with fully reset, the I–V curves with partial reset process shows more gradual characteristics in the set and reset processes. Both I–V characteristics have self-compliance characteristics [44]. The method of connecting the two differences in Fig. 3a and b can be confirmed by a continuous DC sweep in Fig. 3c. The deep reset occurs when the larger voltage is applied, indicating that the strength of the reset can be controlled by the voltage adjustment. The current flows in the HRS induced by the partial reset and an additional reset occur, which lowers the current level due to additional filament decomposition. Figure 3d exhibits a possible switching mechanism of partial reset (left) and deep reset (right) curves. As confirmed in Fig. 2, Al–O bonding has higher binding energy than that Ta–O bonding. This suggests that switching depends on the TaON layer when the small electric field is applied and on the Al2O3 layer when it is a large electric field. Thus, oxygen ions formed between TaN and TaON affect the conduction mechanism of the device and are estimated to result in MLC characteristics [3, 45, 46]. Gradual partial reset with MLC occurs in the TaON layer within the − 2.2 V region. However, the more electric field induces the filament decomposition inside the Al2O3 and causes the abrupt current decrease during the reset process. In Fig. 3e, HRS and LRS were confirmed in the read operation of 0.5 V to demonstrate state uniformity. Since the filament decomposition depends on the magnitude of reset voltage, HRS varies more severely than LRS. Also, more decomposition demands more set voltage to re-form the filaments. Variation of set voltage is shown in Fig. 3f and it varied from − 1.25 to − 0.75 V in accordance with the previous reset cycle process.

Fig. 3
figure 3

Bipolar resistive switching of Pt/Al2O3/TaN device: a deep reset, b partial reset, c process of inducing from partial reset to deep reset, and d deep and partial reset mechanism schematic diagram. Uniformity of e HRS and LRS, f set voltage

Full size image

In Fig. 4a, the endurance characteristics were also measured for partial I–V conditions using pulse for 105 cycles. It shows that HRS and LRS can be switched even at 105 or more times. In Fig. 4b, it is the result of performing the retention test for each I–V characteristic including partial and deep resets. HRS and LRS were measured at the read voltage of 0.15 V, and both states were maintained for 104 s. These results show the Pt/Al2O3/TaN device has good non-volatile memory properties. Multi-level cell characteristics are very beneficial for practical applications such as high-density memory and neuromorphic device [43, 47, 48]. Figure 4c shows a reset process by increasing the reset voltage by 0.2 V for each cycle. Through this process, as the reset voltage increases, multiple HRS is achieved. In Fig. 4d, based on the reset voltage at the boundary between the partial reset and deep reset, the reset process was repeatedly measured while increasing 0.025 V from 1.8 to 2.35 V. It could be verified that the current level gradually decreases, and this could prove the existence of various multi-level states.

Fig. 4
figure 4

a Endurance characteristic of the device for 105 cycles and b each deep and partial HRS and LRS retention characteristics for 104 s. MLC characteristics of c different reset voltages, d consecutive incremental voltage repetition

Full size image

The property of conductance quantization [49,50,51,52] was confirmed. This is thought to be due to the quantization effect of conductive filament during the reset process. When the conductive filament is well controlled, it is possible to implement more state and higher density memory through this phenomenon. As shown in Fig. 5a, this phenomenon can be observed when the conductive filament is modified in atomic units. The step voltage of 0.002 V and delay time of 0.3 s every step is used to observe quantization in multiple cycles, and only elemental disruption of the filament was measured during the reset process. The conductance quantum, represented by the symbol G0, is the quantized unit of electrical conductance. It is defined by the elementary charge e and Planck constant h as G0 = 2e2/h = 7.74809 × 10–5 S. The device takes an integer multiple of G0 or an intermediate value between integers. In the end, LRS is changed to HRS. The statistical analysis is essential through multiple cycles [53,54,55,56,57]. Figure 5b shows the histogram plotting, and it can be seen that even in various conductance steps, there is a high tendency near a multiple of G0 or a half multiple [58,59,60,61]. It is noted the values between 0.5G0 and 3G0 are distinctly distinguishable. It may be necessary to make the conducting filament smaller by means of a method such as making the device smaller in order to distinguish the quantized values. Pulse measurements were performed in Fig. 5c and d to describe the quantized conductance [62,63,64]. Conductance calculated with the voltage of 0.5 V was induced by adding write pulses at 0.5 s intervals. In Fig. 5c, an incremental write pulse increased by − 25 mV from − 0.7 to − 1.775 V was used and the abrupt set operation occurred at a voltage of − 1 V or higher. The conductance in HRS increases more than 10G0 at a time due to the abrupt characteristic in the set region. This characteristic was also confirmed in the I–V curve in the inset image of Fig. 5c, which MLC implemented by limiting compliance current. In contrast, conductance quantization with the erase pulses composed by 25 mV from 1.5 to 2.175 V were ranged of about G0. From those two different conductance ranges show that it is more ease to implement MLC during reset process due to the clear state distinction.

Fig. 5
figure 5

a Quantized conductance characteristic using DC measurement on reset part, and b histogram of quantized conductance levels. Quantized conductance characteristic using pulse measurement c in set region, d in reset region

Full size image

A neuromorphic computing system can be implemented using multi-level cells in Pt/Al2O3/TaN devices. As shown in Fig. 6a, the conductive filament connecting the TE and BE of RRAM can be expressed very similarly to the human’s biological system [43, 65,66,67]. In order to confirm the suitability of neuromorphic computing, pulse measurements were conducted. In Fig. 6b, conductance control is continuously performed through 5 cycles of potentiation and depression by applying the pulses. Potentiation and depression were set to − 1.15 V and 1.3 V, respectively, and both pulse widths were set to 10 sµ. From the I–V characteristic of the set process, relatively abruptness in the potentiation can be confirmed. It could be verified that the depression part has a more gradual characteristic. Moreover, we demonstrate more gradual and symmetric resistance-change characteristics by controlling the voltage amplitude of pulses in Fig. 6c [68, 69]. Each 6 potentiation and depression segments are used to increase and decrease the conductance. The voltage varied from − 0.9 to − 1.4 V for potentiation and from 1.35 to 2.85 V for depression. Figure 6d shows MNIST pattern recognition simulation results by using the conductance results of Fig. 6b and c [70, 71]. The result of using Fig. 6c shows higher accuracy for each epoch. In other words, pulse improvement measurement provides a better learning process.

Fig. 6
figure 6

a Schematic illustration of similarity between human synaptic neural structure and RRAM device structure. Potentiation and depression characteristics: b consecutive same voltage pulse, c single incremental voltage pulse. d MNIST pattern recognition simulation results of (b) and (c)

Full size image

Synaptic functions, such as PPD, EPSC, and STDP measurements, were performed to determine suitability for the neuromorphic application [72,73,74]. Figure 7a shows the device's PPD measurement data, the ratio change between two pulses was confirmed when the seven different intervals were used. Synaptic weight changed with the time interval ranging from 20 μs to 5 ms between two consecutive depression pulses. The amount of synaptic weight change was expressed as ΔW = (A2 − A1)/A1 × 100 (%). As a result, the current responded by the second pulse decreases as the interval increases, indicating that the device is suitable for implementing STP. Figure 7b illustrates conductance changes before and after giving five identical write pulses and summarizes them with pulse amplitudes. As the voltage amplitude increases, both potentiation and depression have a larger synaptic weight change. Continuous stimulation raises EPSC; the degree of weight strengthening can be adjusted according to the amplitude. The strength of connections between neurons in biological synapses can be controlled by STDP. Therefore, if we can elucidate the detailed mechanisms of biological synaptic action and imitate the action behavior, it will be possible to mimic the energy-efficient processing of the human brain. Figure 7c explains the configuration of the STDP protocol. When the pre-spike signal and post-spike signal, which vary with the interval, were applied to the biological synapses, the weight was changed and implemented according to the learning behavior. This process was mimicked on the memristor in the same way. The pulse protocol in Fig. 7d was used for the measurements. The same pre and post-signal were composed, but the different shape of pulses was finally configured and applied according to the interval. Since the final pulse configuration was different, the synaptic weights over time had different weight changes, as shown in Fig. 7e [2, 75, 76]. In general, the shorter the absolute time of the spike time difference, the greater the change in conductance change like a biological synapse.

Fig. 7
figure 7

a PPD measurement, b EPSC data according to amplitude. STDP characteristics: c Schematic for measurement imitation between synaptic neural structure and RRAM, d pulse authorization for STDP measurement at Δt = 60 μs and e result of STDP measurement

Full size image

Conclusions

As a result, the MLC characteristics and quantized conductance were confirmed through the Al2O3-based RRAM device deposited with ALD, and excellent biological characteristics were investigated through pulse measurement. DC I–V bipolar switching characteristics were verified through DC measurement, and it was verified that switching characteristics of two different characteristics could be easily controlled only by adjusting a voltage. Multi-levels in various cases were confirmed by varying the amount of voltage that adjusts different characteristics, and the conductance quantization phenomenon was also confirmed within the reset section and pulse measurements. This MLC phenomenon was connected with pulse measurement to measure potentiation and depression, and it was possible to maximize MLC characteristics through voltage control of each segment. Including PPD and EPSC, through the measurement of STDP, the change in the conductance weight of the device was confirmed by imitating the synapse. In conclusion, the MLC characteristics of the device and the suitability of neuromorphic computing were successfully completed.