Effect of Al Concentration on Ferroelectric Properties in HfAlOx‐Based Ferroelectric Tunnel Junction Devices for Neuroinspired Applications

Since HfOx‐based ferroelectric tunnel junctions (FTJs) are attractive compared to perovskite‐based FTJs and other emerging memory devices, they are being actively studied recently. They have advantages such as a simple metal–insulator–metal structure, complementary metal oxide semiconductor (CMOS) compatibility, non‐destructive operation, and low power consumption. Moreover, doped HfOx‐based FTJs are in the spotlight in terms of neuromorphic engineering as a way of advancing from the von Neumann structure. In particular, Al dopant is effective for inducing ferroelectric properties due to its smaller radius than that of Hf. The optimal concentration of Al varies depending on the device materials and the annealing conditions during deposition. Therefore, in‐depth research is required for neuromorphic applications. Herein, the properties of FTJ devices according to Al doping concentrations are analyzed. Subsequently, using the device with the highest remanent polarization, neuromorphic applications are implemented, including spike‐timing‐dependent plasticity (STDP), paired‐pulse facilitation (PPF), long‐term potentiation, and depression. The characteristics in different frequency regions are also studied to satisfy the demand for fast switching. Finally, the FTJ device is used as a physical reservoir in reservoir computing for efficient processing of time‐dependent inputs.


Introduction
Computing systems including memory devices with high integration and short latency periods are being studied to address increasing data demands. [1,2] Existing von Neumann structures have limitations in terms of data processing because computation and storage are separated. [3] Moreover, it is expected that charge-based DRAM and NAND flash memories will reach physical and technical limitations. [4][5][6][7] To ensure better memory performance, various emerging memory devices have been studied to satisfy conditions such as fast switching, high integration, and parallel operation. These new devices include resistive random access memory (RRAM), ferroelectric random access memory (FeRAM), phase-change random access memory (PcRAM), and spin-transfer torque magneto random access memory (STT-MRAM). Within the industry, the complementary metal oxide semiconductor (CMOS) compatibility of device materials is very important. [8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][23][24] Among these devices, ferroelectric tunnel junctions (FTJs) are attractive for various neuromorphic applications that could imitate the nervous system of the human brain to overcome the limitations of conventional von Neumann structures. The structure of FTJs is simple, whereby a ferroelectric thin film capable of electric tunneling is inserted between the top and bottom electrodes. Their advantages include a nondestructive reading process (unlike RRAM) and low power consumption. [25][26][27][28] In 2011, ferroelectric characteristics were discovered in a HfSiO x (HSO) thin film for the first time. Since then, HfO x -based FTJs have been studied actively and presented a new possibility for CMOS-compatible materials. [29] Existing FTJ devices based on perovskite materials, such as Pb(Zr, Ti)O 3 (PZT), BaTiO 3 (BTO), and SrBi 2 Ta 2 O 9 (SBT), are incompatible with the CMOS fabrication process and difficult to scale down due to their loss of ferroelectric properties when reduced to a few nanometer thicknesses. [30,31] Recently, it was discovered that doping HfO x films induces strong ferroelectricity which can form excellent orthorhombic phases even at a few nanomete thickness. These dopants enhance the crystallization and phase stability of the film which enhances its ferroelectric properties. Among the various dopants that have been studied with HfO x -based FTJs, Al is effective for inducing ferroelectric phase transitions, because it generates spontaneous mechanical stress during the postannealing cooling process due to its smaller atomic radius compared to Hf. [32] In HfAlO x (HAO), it is assumed that the Curie temperature is spatially ununiform due to the diffusion of Al atoms near the electrode interfaces. [33] The ferroelectricity of HAO-based FTJs is known to be the best characteristic when the concentration of Al relative to Hf is 2-5%. However, further research is required because this percentage depends on the device's material stacks and the annealing conditions. [34] For HfO x materials to have stronger ferroelectricity, the generation of a monoclinic phase (m-phase) should be suppressed and an orthorhombic phase (o-phase) should be induced. Moreover, during the cooling process after rapid thermal annealing (RTA), elongating the c-axis of a tetragonal phase (t-phase) by internal stress can promote a phase change to the o-phase. To generate o-phases in HfO x -based materials effectively, different methods have been proposed, such as high-pressure annealing (HPA) and stress engineering using the thermal expansion coefficients of various materials. [35][36][37][38][39] A fundamental approach for implementing neuromorphic engineering in nonvolatile memory is the design of devices that mimic the functions of the artificial synapse. Learning is based on long-term synaptic plasticity and spike-timing-dependent plasticity (STDP), which has been recognized as the basis for applications in memory devices. Synaptic plasticity can be emulated by nonvolatile memory devices, whose conductance can be changed continually based on the history of applied voltages or currents and remains over hours or days. Therefore, memristors are considered promising electronic materials for artificial synapses. Among memristors, FTJs offer several advantages for emulating synapses given their large on/off ratio, fast read and write/erase speed, low power consumption, and excellent endurance. By switching ferroelectric domains under an applied electric field, FTJs demonstrate a memristor-like response, producing discrete analog resistance levels appropriate for modulating synaptic weights. [25,40] In this study, the ferroelectric characteristics of HAO-based FTJs according to various Al doping concentrations were examined. To analyze the devices, a positive-up-negative-down (PUND) method was used to obtain the remnant polarization (P r ) values, and DC current versus voltage (I-V) measurements were conducted to extract the tunneling electroresistance (TER) ratio. In addition, grazing-angle incidence X-ray diffraction (GIXRD) and electrical film analyses were conducted for polarization reversal theory. Based on these analyses, various neuromorphic system applications were implemented, such as STDP, paired-pulse facilitation (PPF), potentiation, and depression, using the device with the highest P r value and easy conductance modulation capability. Furthermore, the characteristics within a high-frequency region were studied, which are important in terms of operating speed. Finally, the FTJ device was used as a physical reservoir to implement 16 states of 4 bits in reservoir computing, which is a computing framework that efficiently processes time-dependent and sequential inputs. Therefore, these critical traits of the device will have high potential in both energy-efficient nonvolatile memories and artificial neural network applications.

Experimental Section
A schematic of the device is presented in Figure 1a. To induce better stress on the HAO films, TiN was employed for the top electrode and n þþ Si substrate was used for the bottom electrode. An appropriate thickness of %8 nm for the HAO film was selected because when a ferroelectric layer is too thin, the o-phase formation could be degraded. Moreover, when the film becomes too thick, the film properties could be seriously reduced due to a partial bilayer. [41,42] The thicknesses of HAO films were fixed at 8 nm, although the A1 concentration varied between 2% and 6%. The complete device fabrication process is presented in Figure 1b. First, we started with a 6 inch (100) heavily doped n þþ Si substrate with a thickness of 625 AE 25 μm and a resistivity of 0.005 Ω cm. Prior to the ferroelectric layer deposition, a DHF (HF: H 2 O = 1:100) cleaning step was performed to completely remove any native oxide on the surface of the Si substrate. Next, HAO films were deposited by thermal atomic layer deposition (ALD) at a stage temperature of 380°C. The doping concentration of the 8 nm-thick HAO film was controlled by the number of Al 2 O 3 deposition cycles. Tetrakis (dithylmethyl amino) hafnium (TDMAHf ) and trimethyl aluminum (TMA) were respectively used for the Hf and Al precursors, with O 3 as the oxygen reactant. The deposition rates were 1.38 and  www.advancedsciencenews.com www.advintellsyst.com 0.9 Å/cycle for HfO 2 and Al 2 O 3 , respectively. The Al composition in the HAO film was determined by the ratios of Al to Hf, which were 1:25 (2%), 2:24 (4%), and 3:23 (6%) in this work. This process was repeated three times, ultimately depositing an 8 nmthick HAO layer. After the deposition of the HAO layer, a 100 nm-thick TiN layer was deposited on the HAO layer by physical vapor deposition (PVD) sputtering (AMAT, ENDURA 5500) to act as an electrode and a capping layer to contact and suppress volume expansion. Next, the RTA process was performed in a range of 400-800°C in N 2 ambient for 25 s. Finally, patterns with an area of 100 Â 100 μm were defined by lithography and reactive ion etching. A Keithley 4200-SCS parameter analyzer and GIXRD (X'pert Pro) were used to measure the electrical properties and analyze the ferroelectric phases, respectively. A field-emission transmission electron microscopy (FETEM, JEOL JEM-F200) image was used to observe the structure of the device (Figure 1c).

Results and Discussions
The polarization versus voltage (P-V ) loops with Al-to-Hf ratios of 2%, 4%, and 6% are presented in Figure 2a. The P-V curves of each device were obtained after 1000 cycles using the PUND method, with optimized pulse amplitudes for each device to achieve maximum 2P r values. If the applied voltage is too high for the device, defects (including oxygen vacancies at the interface) can affect the effectiveness of the Schottky barrier of the junction or form a filament inside the ferroelectric layer, facilitating the breakdown of the device. [43] In contrast, there will be insufficient switching if the applied voltage is too low. Therefore, the P-V curves of the three devices with different Al content were compared by measuring the condition at which their performance was optimum, based on the characteristics of each device. In the initial P-V loop, since the coercive voltage (V c ) was not the same at all domains of the ferroelectric layer (due to locally pinched ferroelectric domains), it was possible to obtain curves with relaxed pinching after sufficient cycling. [44,45] In addition, each V c in the positive and negative regions displayed asymmetric values. This was interpreted as being the result of either an internal bias due to the work function differences of the FTJ stack, fixed charges in the process of fabrication, or trapped charges in the cycling process. [46][47][48] Since there can be a difference in the applied voltage range between negative and positive biases, an open loop could occur if the bottom electrode is a semiconductor. [49] The 2P r values extracted from P-V curves are displayed in Figure 2b. When the Al concentrations of the devices were 2%, 4%, and 6%, the obtained 2P r values were 43.53, 14.58, and 4.49 μC cm À2 , respectively. When the concentration was greater than 2%, the 2P r values exhibited significant degradation. Moreover, when the Al concentration was too high, serious degradation of the ferroelectric characteristics was expected, even for pure HfO x without any dopants. When pure HfO x was used as a ferroelectric layer, sufficient o-phase could be induced if the thickness of the film was 7 nm (or less). However, if the thickness increased, it was difficult to achieve sufficient ferroelectric properties. [34,50] During the process of alternately depositing Al and Hf layers using ALD, as the Al layer becomes thicker, Al behaves as an insulating layer rather than a dopant layer. Hence, the characteristics of the film could be seriously degraded. In general, Al has a smaller radius than Hf and is known to induce mechanical stress after the annealing process. However, if the thickness of the Al layer increases excessively, this could result in a disconnection between each Hf layer located above and below the Al layer. Consequently, this disconnection between Hf layers could reduce the effective thickness of the entire   www.advancedsciencenews.com www.advintellsyst.com ferroelectric film. The thickness of the ferroelectric film also has a significant effect on the characteristics of the FTJ device, because if the film becomes too thin, the vertical direction is restricted, suppressing the formation of the o-phase. [41] Next, GIXRD analysis was used to understand the crystalline characteristics of the HAO thin films with different Al doping concentrations. The m (À111), o (111)/t (011), and m (111)-phase peaks in the three devices were observed. Figure 2c depicts the GIXRD spectrum of the devices with Al contents of 2%, 4%, and 6%. The diffraction peaks corresponding to the (111) planes of the noncentrosymmetric o-phase, which are responsible for the ferroelectricity of HAO films, are visible in the diffraction patterns of the three devices with a 2θ range of 26°-40°for the HAO film with various Al doping concentrations. Each Al doping concentration rendered a collective combination of different phases. Moreover, the peak intensity of the o-/t-phase decreased as the Al doping concentration increased with a corresponding increase in the t-phase ratio, representing the lack of transitional phase transformation from the t-phase to the o-phase. To separate the peaks between the o-and t-phases, Gaussian peak fitting was performed to distinguish the deconvolution of all peaks, as shown in Figure 2d. The results suggested that the o-phase had the highest peak in the 2% device. However, this should be compared through the ratio of the phases, because the m-and tphases are not negligible. Figure 2e illustrates the phase ratios of the o-, t-, and m-phases. Each phase ratio was calculated from the ratio of the integrated peak intensity corresponding to the o-, t-, and m-phase in GIXRD patterns. The highest o-phase ratio of the 2% device was consistent with the P r value mentioned previously. For the 4% and 6% devices, the t-phase was the dominant formation and both devices exhibited lower P r values. For the device with an Al concentration of 2%, the o-phase ratio was the highest, which is consistent with the results shown in Figure 2a,b. All of the devices with various Al concentrations were annealed at 800°C. To compare changes in P r values depending on the annealing conditions, the result of varying the annealing temperature for the 2% of Al concentration is displayed in Figure S1, Supporting Information. If the annealing temperature is below 800°C, this will be insufficient to induce phase transition. In contrast, if the annealing temperature is too high, the phase transition will degrade due to unexpected diffusion problems and thermal stress. [32] Next, we verified the device-to-device variability shown in Figure S2a,b, Supporting Information. We collected the I-V and P-V curves from ten different cells. Additionally, the coercive voltage and remanent polarization values were extracted from the P-V curves of the ten devices, as shown in Figure S2c,d, Supporting Information. We can infer that the devices exhibit excellent device-to-device uniformity.
To understand the cause of the different P r values in our three devices, we applied a short-pulse measuring technique based on the polarization reversal theory. [45] Ferroelectric films contain a mixture of phases. Among these phases, the o-phase contributes to the ferroelectric property. Therefore, the higher the o-phase ratio inside the film, the better the ferroelectric performance of the device. Assuming that the interfacial layer includes defects from the fabrication process and phases that do not contribute to switching, the capacitance of the dead layer (composed of the phases and defects) can be obtained using the polarization reversal theory. Moreover, since the ferroelectric capacitor acts as a resistance in the polarization switching process and causes a constant voltage drop, the interfacial capacitance (C i ) is considered a time-independent constant value; hence, the current flows according to the fixed time constant. [51,52] Here, the C i values correspond to the amount of interfacial nonferroelectric t-phases. These interfacial nonferroelectric compositions (t-phases) induce weak ferroelectricity in the HAO films. The time-dependent domain-switching current (I SW ) can be expressed as follows where t 0 , t sw , C i , and R L are the time of start domain switching, the time of the end of switching, the capacitance value of the interfacial dead layer, and the total resistance of the device, respectively. [52] For prepoling, a write pulse with a width of 10 μs and amplitude of À8 V was applied, followed by a read pulse of 10 μs with amplitudes of 6-8 V. Figure 3a,b,c respectively shows the data of the devices with Al concentrations of 2%, 4%, and 6%, which were measured under the same conditions. Starting from t 0 , I SW decreased exponentially and switched after the capacitor was fully charged. According to Equation (1), the current value at t 0 can be expressed as follows where V a is an applied voltage. [52] According to Equation (1), the results of fitting each I SW value after t 0 with exponential functions (depending on time) are displayed in Figure S3, Supporting Information. Next, after plotting I 0 SW as a function of V a , the results of fitting with linear functions are presented in Figure 3d, and the R L values could be derived through the slope of the functions. Finally, the C i values per unit area of each device were calculated (Figure 3e), where it can be observed that the device with 2% of Al doping concentration represented the largest C i value. Since the capacitance value is inversely proportional to the thickness of the layer, it can be concluded that the dead layer of the device with 2% of Al concentration was the lowest. This result exhibited the same trends with the P-V loops and GIXRD. However, the exact thicknesses of these interfacial layers are difficult to obtain through these calculations since it is only a rough measurement of nonferroelectric factors distributed throughout the film. [45,53] Nevertheless, it is still an effective way to compare the nonferroelectric factors of the devices presented in this work.
Since FTJs follow domain switching dynamics, gradual and analog switching is possible. [54][55][56][57] Figure 4a displays the equivalent circuit with a schematic that depicts the aspect of switching per domain in response to an external bias. If the applied voltage is not sufficient to align all domains in one direction, only some domains with low V c are switched. Hence, the film can be considered a combination of parallel domains with different resistance values. Figure 4b presents the changes in 2P r values as a function of the sweep voltage in the PUND method. As the sweep voltage increased, large 2P r values were obtained, and this domain-switching characteristic provided the possibility of utilizing the device for various applications. Band diagrams and alignment schemes of the polarization vectors when the TiN/HAO/n þþ Si FTJ was in the high-resistance state (HRS) and low-resistance state (LRS) are shown in Figure 4c. Generally, the screening potential (φ) of FTJ devices can be explained using the Thomas-Fermi model φðxÞ¼ 8 > > > < > > > : where d, σ s , δ top , and δ bottom are the barrier thickness, screening charges per area, and Thomas-Fermi screening length of the top and bottom electrodes, respectively. [43] In the band diagrams at the top of Figure 4c, the white dotted line indicates the average value of the potential barrier when each material is contacted. Since the top and bottom electrode materials had different screen lengths, an asymmetric electrostatic profile was formed (as shown in the band diagrams) and sufficient tunneling occurred.    Furthermore, if the bottom electrode is a semiconductor, it is possible to remove and form additional barriers due to the conversion of accumulation and deletion near the junction surface, meaning that the tunneling barrier can be modulated. [58] This can be advantageous because the tunneling phenomenon of electrons is exponentially dependent on the barrier potential and width (from a quantum mechanical perspective). [59] Figure 5a displays the hysteretic tunneling current responses obtained after cycling the device 10 times with 2% Al concentration. The current responses were double swept in a range from À3 to 7 V. In the negative region, it was revealed that the tunneling current was rectified by the depletion region of the bottom electrode, which can be expected to prevent sneak current problems in array structures. [49,60] Through cycling from an initial state and due to a wake-up behavior, which is affected by internal fields from ununiformly distributed defects, the TER ratios can improve gradually. [49,58] The TER ratio changes as a function of measured voltage during the cycling process are plotted in Figure 5b. Moreover, the TER values at each point were calculated by an inserted equation in the right corner of Figure 5b. The maximum TER ratios of 79 and 120 for the first and last cycles were measured at read voltages of 1.8 and 1.6 V, respectively. During cycling, the maximum TER ratio gradually increased and the read voltage required to extract the maximum TER value tended to decrease gradually. This tendency suggested that the device was free from wake-up effects during cycling, while simultaneously exhibiting improved performance with less power. [61] When the bottom electrode is a semiconductor, it is advantageous to secure a high TER ratio. [49,58,60] Table 1 shows the comparison of the different on/off (TER) ratios in other MF(I) S-structured FTJ devices. Figure 5c displays the tunneling current by changing the sweep range, starting from À3 V to 5-7 V. The HfO x -based ferroelectric film with polycrystalline characteristics operated based on domain switching dynamics. Hence, a larger number of domains were switched as the bias increased, and the possibility of controlling the on-current gradually increased. Figure 6a displays how the polarization vectors of the device with an Al concentration of 2% responded to the second pulse over time. [62] When the time interval between the two pulses was zero, the switched state by the first pulse was maintained. This indicated a suppressed current response, with only those components that did not contribute to switching. However, when the pulse interval was 1 s, the maximum value of the current returned to 92.70% of the initial set. Moreover, in the case of  Chouprik et al. [84] HZO (12 nm) Al 2 O 3 (2 nm) Ti/p þ Si 5 Ryu et al. [40] HZO (8.4 nm) Al 2 O 3 (1 nm) TiN/p þ Ge 14 Shekhawat et al. [85] HZO (6 nm) None TiN/p þ Ge 30 Goh et al. [25] HZO (12 nm) Al 2 O 3 (1 nm) TiN/p þ Ge 78 Shekhawat et al. [85] HZO (4.5 nm) None TiN/p þ Si 80 Mikheev et al. [86] HAO (8 nm) None TiN/n þ Si 120 This work www.advancedsciencenews.com www.advintellsyst.com 2 s, it returned to 97.26%. As the gap between pulses increased, it could be interpreted that local switching occurred, as the polarization vectors aligned by the first pulse were returned by the depolarization field (E dep ). If the bottom electrode is a semiconductor, E dep could occur due to incomplete screening of the surface-bound charges of the ferroelectric film and the fixed charge traps inside the device. [43] Moreover, E dep causes a significant change in the resistance value of the barrier after the reversal of the polarization in FTJ devices with asymmetric structures. [63] To assess the reliability of the device, endurance tests were conducted at different voltage ranges, as shown in Figure 6b. These tests were conducted through PUND measurements by applying voltage amplitudes ranging from 6 to 8 V with increments of 0.5 V.
The endurance of the HAO film increased with relatively lower voltage amplitudes, while the absolute value of P r increased with continuous cycling due to the wake-up effect. In specific voltage ranges, the device underwent three stages: discrete wake-up, stability, and fatigue. As the applied voltage increased, we observed that the device could not withstand the high voltage amplitudes and tended to suffer from an earlier fatigue stage or hard breakdown. This fatigue can be caused by the formation of leakage current paths due to the rearrangement (or formation) of defects, such as oxygen vacancies in the ferroelectric film during electrical cycling. Nevertheless, the device still achieved excellent endurance performance, enduring repeated cycling (10 7 cycles) at relatively lower applied voltage ranges. In Figure 6c, the V c values are plotted in a log scale that changes depending on the frequency when sweeping with the same pulse amplitude using the PUND method. An increase in frequency means a decrease in the pulse width of the PUND measurement; hence, a larger voltage is required for polarization reversal. [64] Since fast switching is a critical factor in terms of practical applications, it is important to analyze the characteristics in each frequency region. Here, measurement frequency and V c are closely related, and two models describe the switching dynamics of the ferroelectric layer: nucleation and domain wall motion and viscous flow motion. [53,65] In the low-frequency region, where the frequency was less than the crossover frequency ( f cr ), the nucleation and domain wall motion was dominant. In comparison, in the region where the frequency was greater than f cr , the motion mainly followed the viscous flow motion. [66] Nucleation and domain wall motion that causes domain wall transformation from one polarization state to another was suggested by Du et al. and follows Equation (4) [53,[67][68][69] Furthermore, viscous flow motion assuming that domain growth is limited has been reported by Orihara et al. and can be expressed as follows where D is an effective dimensionality of domains and k is an experimental parameter. Moreover, by extending a previous theory of Nattermann et al., we know that a dynamic crossover between these models causes two scaling regions, which is used for explaining magnet systems. [65,66,70] FTJs exhibit memristor-like responses through switching ferroelectric domains resulting in discrete conductance levels. www.advancedsciencenews.com www.advintellsyst.com Therefore, implementing analog multilevel states using the device is important for demonstrating synaptic device applications. [49] Figure 7a presents an identical pulse scheme in which 0.9 ms pulses with amplitudes of 7.7 and À0.9 V are applied 50 times. In addition, an incremental pulse scheme in which the amplitude gradually changes and becomes the same as the maximum amplitude of the identical scheme is presented in Figure 7b (pulse amplitudes: from 5 to 7.7 V and À1e À7 to À0.9 V). In both methods, the read pulse was determined as 2.3 V, and the conductance values of the device programmed in two ways exhibited significant differences, especially in the linearity of potentiation. However, not much difference in the linearity of depression is observed because of the asymmetric MFS structure. This asymmetry in the electrodes induces asymmetric switching behavior as a function of positive and negative bias. One of the reasons we can infer through the momentary drop in the depression is that the initial domain nucleation gives an initial conductance variation as the domain expands. [53,55,65,66] This mechanism may be responsible for the nonlinearity even when an incremental pulse scheme is used. Furthermore, our device has a strong short-term depression characteristic which is why even with a small negative amplitude of À0.9 V, the conductance tends to return to its original state quickly. This type of short-term depression is beneficial for reservoir computing where the boundary between high-and low-conductance states needs to be large. [71,72] Moreover, a gradual state change could be induced at a relatively low range of conductance levels. Accordingly, demonstrating that the device can be easily controlled by applied pulses proves that it can be used as a synaptic device. We demonstrated the applicability of the device with 2% Al concentration for synaptic device applications, from which various functions imitating the human brain have been reproduced.
In Figure 8b, we imitated an STDP function, which is in charge of an important process of learning and memory in the biological nervous system. [40,[73][74][75][76] The conductance weight, which represents the synaptic weight, can change with differences in the order and timing between pre-and postsynaptic spikes. However, individual spikes cannot elicit conductance changes, which can only be induced by overlapped spikes. [77] As shown in the pulse schemes in Figure 8a, if the prespike is before the postspike, this represents a positive conductance change, while the reverse indicates a negative change. The developed device demonstrated the ability of neuroinspired computing through the simulation of STDP learning. Figure 8c displays the transient current responses of the device when applying the programming pulses. The pulse conditions consist of voltage amplitudes of 8 V and a pulse width of 1 ms. As the interval time increases, the rate of current decay also grows, which resembles the short-term effect. Figure 8d shows the results of mimicking a PPF function, which is a short-term memory mechanism of the human brain. Since spontaneous state recovery occurs by E dep , as shown in Figure 6a, it is possible to check changes in conductance by the time interval between two pulses. The result of PPF was obtained as a ratio of the intermediate values of the current response through two pulses with a pulse width of 1 ms and an amplitude of 8 V. As the time interval between the pulses increased, the current value extracted from the second pulse became significantly lower than that from the first pulse. Therefore, the overall PPF data could be fit as an exponential function for the time interval. Similar to the short-term memory of the human brain, we demonstrated that the shorter the learning interval of the device, the better remembered what is to be learnt. [78][79][80] Finally, using the device as a physical reservoir, we implemented reservoir computing, which is a computing architecture  for the efficient processing of sequential and temporal inputs based on the biological nervous system. [71,72,81,82] Reservoir computing includes a reservoir (which receives input and transforms it high dimensionally) and a readout (which is used for training), as shown in Figure 9a. Since the device is nonlinear with respect to the input and has the characteristic of returning to a previous state over time, it is suitable for a physical reservoir in reservoir systems. Reservoir computing has significant advantages in terms of time and cost because only the readout is trained with a simple process, such as linear regression or classification. [83] Figure 9b displays the result of demonstrating 4 bits of 16 different states by inputting time-dependent pulses to the device.
To verify the repeatability of the system, each state was expressed as an average of the conductance values repeated ten times in one cell. The result of repeatedly executing each state is presented in Figure S4, Supporting Information. Since there is a distinguishing layer of at least 0.3 μS between the two states, 1 and 0 states can be distinguished, even when considering device variations. This result demonstrates the possibility of expanding to a parallel array structure, as shown in Figure 9c. Here, since each row is  www.advancedsciencenews.com www.advintellsyst.com responsible for 4 bits (depending on the input pulse trains), it is possible to implement 0-9 digits consisting of 5 Â 4 pixels by integrating these 5 rows. Furthermore, it is possible to expand the system for processing more bits in each row by inputting the 4-bit-pulse trains several times.

Conclusion
Herein, we analyzed the characteristic differences of TiN/HAO (8 nm)/n þþ Si devices according to Al doping concentration. When the Al concentration was 2%, the o-phase exhibited the highest ratio, meaning the optimum P r value and TER ratio could be demonstrated. As the Al concentration increased, the Al layer operated as an insulator rather than a dopant, resulting in a decrease in the effective thickness of the entire film. In addition, the frequency-dependent characteristics of devices with an Al content of 2% were measured, and two models suggested in previous studies were explained for the analysis. To derive the synaptic properties of the Al 2% device, we obtained repetitive potentiation and depression data using several pulse schemes. Based on these results, we successfully implemented synaptic functions (STDP and PPF), demonstrating the possibility of using the device as a synaptic device. Additionally, the device was utilized as a physical reservoir in reservoir computing for effective data processing, and 4 bits of 16 states were successfully implemented. Furthermore, the possibilities of repeated application of the 4-bit-pulse train or expansion into the array structure were demonstrated.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.