Demonstration of Synaptic Characteristics in VRRAM with TiN Nanocrystals for Neuromorphic System

To efficiently develop an extremely intensive storage memory, the resistive random‐access memory (RRAM), which operates by producing and rupturing conductive filaments, is essential. However, due to the stochastic nature of filament production, this filamentary type resistive switching has an inherent limitation, which entails the unpredictability of the driving voltage and resistance states. Several strategies such as doping, research into multilayer stacks, and interface engineering, are suggested to tackle this challenge. This work fabricates a CMOS‐compatible TiN/HfOx/TiN‐NCs (nanocrystals)/HfOx/TiN RRAM to implement analog resistive switching and advance the development of the synaptic device. Specifically, atomic force microscopy (AFM), scanning electron microscopy (SEM), and transmission electron microscopy (TEM) are utilized to observe the formation of TiN nanocrystals, which play a crucial role in the enhancement of resistive switching. By comparing HfOx–based RRAM devices with and without NCs, the DC I–V curves, retention, endurance, and switching speed are properly examined. Interestingly, it is found that the TiN/HfOx/TiN‐NCs/HfOx/TiN device is more appropriately utilized as an artificial synapse in neuromorphic systems mainly due to its stable and reliable resistive switching properties. Finally, this work demonstrates well‐controlled resistive switching 3D vertical RRAM with TiN‐NCs, which is particularly suitable for high‐density memory.


Introduction
The human brain, one of the most complex objects, is capable of high-dimensional cognitive activities such as generalization, decision-making, cognition, and prediction with a issues entailing individual defect diffusion and instability, which causes the fluctuation of switching values during consecutive switching cycles. [16][17][18] More precisely, cycle-to-cycle and deviceto-device variations are determined by the filament shapes, location in the insulator, and the internal structure of the RRAM devices. [19] Moreover, the stochastic characteristic of filament formation during switching operations constitutes a major hindrance to the mass production stage, which must operate stably in most of the cells. [20,21] Nevertheless, to mitigate cycle-to-cycle and device-to-device variation, several approaches have been reported, including the use of a multilayer and the doping and insertion of nanoparticles into the switching layer. [11,[22][23][24] Recently, it has been demonstrated that an HfO x -based device incorporating Au nanocrystals achieves stable switching and fast speed. [25] In the ITO/PVPy-Au NPs/Al device, these nanocrystals have been demonstrated to be suitable for neuromorphic systems by imitating synaptic functions with a high on-off ratio. [26] The gradual formation and rupture of the conductive filament in ALD titanium nitride nanoparticles by a local electric field indicate its reliable endurance properties and analog resistive switching. [27] Stable switching is achieved in the Ag/ZrO x /Cu-NC/Pt device by the persistent formation of conductive filaments with the Cu nanocrystal on the bottom electrode. [28] Ru nanodots (NDs) were introduced into the TiO 2 layer in which Ru-NDs played two roles, field enhancer, and field suppressor, demonstrating enhanced RS behavior. [29] Furthermore, for high-density storage memory and the integration of artificial synaptic devices into the neuromorphic system, the multilevel characteristic is vital, which may be implemented via gradual synaptic weight change with external stimulation. [30][31][32] Specifically, most memristor elements perform a binary transformation that abruptly flips between a high-resistant state (HRS) and a low-resistance state (LRS). In particular, memristors having an analog switching behavior that gradually changes resistance states between HRS and LRS in response to external stimuli are suitable for neuromorphic computing. [33][34][35] Synaptic operations on RRAM devices include potentiation, depression, spike-rate-dependent plasticity, and paired-pulse facilities. [36][37][38] In this study, the fabricated TiN/HfO x /TiN-NCs/HfO x /TiN RRAM devices were verified using energy-dispersive X-ray spectroscopy (EDS) and transmission electron microscopy (TEM). Additionally, the HfO x -based devices without NCs are prepared to compare and evaluate the I-V characteristics, retention, and endurance in DC mode. Furthermore, the resistance conversion while varying the pulse width is examined to determine the switching speed. Finally, we investigate the changes in synaptic weights according to interval time, pace, and amplitude as well as long-term potentiation/long-term depression (LTP/LTD) features to simulate biological synapses.

Results and Discussions
We initially examined at how TiN nanocrystals introduced between switching layers influences resistive switching properties in a single planar cell (100 × 100 μm square pattern). The crystalline and thickness of TiN/HfO x /TiN-NCs/HfO x /TiN RRAM devices were examined by the TEM in Figure 1a. Specifically, the polycrystalline states (yellow circle) can be partially identified in a 2-nm-thick TiN, which proves that TiN-NCs are well formed in Figure 1b. Further, energy dispersive spectroscopy (EDS) was performed to analyze the atomic concentration of each layer. Particularly, the presence of Ti, N, Hf, and O was detected (Figure 1c). It should be noted that Ti and N elements are observed in the middle of the HfO x layer. More specifically, the atomic force microscopy (AFM) and scanning electron microscope (SEM) images of the sample are shown in Figure 1d,e. Precisely, the TiN-NCs were formed by a high-temperature ALD deposition environment and annealed for a brief amount of time, during which the melted film formed nanocrystals of about 20 nm average diameter in an amorphous TiN nanolayer as a result of the strain effect. [39,40] Furthermore, control devices were prepared to analyze how the nanocrystals affect resistive switching. The composition and thickness of the devices are presented in Table 1. A single cell structure is employed for all three devices, including the control devices which are not included nanocrystals. Descriptively, the HfO x thicknesses of D1 and D2 are 10 and 20 nm, respectively, without nanocrystals; while D3 includes TiN-nanocrys.tals inside the HfO 2 layer. We initially examined how a TiN nanocrystal introduced between switching layers affects resistive switching properties in a single planar cell (100 μm × 100 μm square pattern). The formation and the initial reset operation of three devices after forming process are shown in Figure 2a. Particularly, the current increases quickly at a certain voltage in D1 and D2. However, D2 does not exhibit good resistive switching after the forming probably due to the current overshoot. Meanwhile, D3 has a different shape in which the current gradually increases. Specifically, the oxygen vacancies distribution before forming can affect the resistive switching depending on the presence of TiN-NCs. The distribution of the forming voltage over 15 cells is shown in Figure 2b. More precisely, a compliant current of 10 μA was adopted to prevent the devices from permanent switching failure. The forming voltages of D2 and D3 are ≈−8.16 and −8.68 V, respectively, indicating that the TiN-NCs do not affect the forming voltage significantly, rather, the thickness of the HfO x layer is the dominant factor for the forming voltage.
In the case of D2, conductive filaments (CFs) are first formed with both −7.8 and −8.7 V, whereas most of the cells in D3 are formed between −8.5 and −8.8 V. Since ALD-based TiN-NCs exist in the forming voltage distribution, a uniform forming process is accomplished. Figure S1, Supporting Information, illustrates the forming curves, where a negative voltage is applied from −3 V while increasing the maximum sweep voltage by steps of 0.025 V. In Figure S1, Supporting Information, the current of D3 is gradually increased while the negative voltage is steadily increased, indicating that the filament is created in phases. Meanwhile, the current of D1 and D2 rapidly increases at −9.25 and −5 V, respectively, indicating that TiN nanocrystals are not placed in those devices. Figure 2c,d shows the bipolar resistive switching characteristics of D1 and D3 after an electroforming process at 10 μA. Interestingly, the set and reset operation of D1 and D3 occurs fluently, unlike D2. Descriptively, compared to the electrical properties of D2 in Figure S2, Supporting Information, the resistive switching properties including memory window and yield rate were enhanced when the nanocrystal was inserted. In particular, the current level is greatly reduced, and the HRS value of D2 is close to the LRS value of D3. When comparing the resistance state  distributions in 10 cells of the two devices, Figure S3, Supporting Information, shows that the cell-to-cell variation is enhanced when the nanocrystals are placed in both resistance states. Furthermore, the switching models of HfO x -based RRAM and NCs-inserted HfO x -based RRAM are illustrated in Figure 3. Specifically, due to the non-uniform composition profile of the oxide layer, which results in a high local concentration of oxygen vacancies, this layer serves as a reservoir for defect injection during the filament formation and set/reset operation. [41,42] As a result, in the initial state, a conductive filament made of oxygen vacancies is created and transitioned to the low resistance state between the electrodes at the same time in Figure 3a. Particularly, the oxygen ions drive to the top electrode and interact with the oxygen vacancies, causing the filaments to be destroyed when the opposite polarity of the reset voltage is given in Figure 3b. Figure 3c,d show the switching mechanism of the RRAM with TiN-NCs. In contrast to Figure 3a,b, the conductive filaments are connected or disconnected near the TiN nanocrystals. Therefore, pseudo-straight conductive filaments are only created in the region between the nanocrystals, as opposed to the rupture and regeneration of the branched conductive filaments, which results in a significant fluctuation of driving voltage and resistance states. [43] To verify the reliability of RRAM with NCs, retention, and endurance tests were conducted as depicted in Figure 4. Specifically, both D1 and D3 devices exhibited outstanding retention properties while reading the resistance states at a read voltage of 0.2 V for 10,000 s in Figure 4a,b. In the endurance test, it can be seen that the D3 with NCs has improved the resistive switching characteristics. Meanwhile, the D1 device without NCs exhibits poor endurance characteristics such as the collapse of the memory window after fewer than 100 cycles (Figure 4c). On the other hand, the D3 achieves higher stable endurance characteristics during 300 cycles as depicted in Figure 4d.
Next, the switching speed of D1 and D3 is investigated to examine the effect of NCs on the devices. Specifically, the resistance of the devices is monitored by gradually changing the pulse width (20 ns, 50 ns, 100 ns, 1 μs, 10 μs, 100 μs, and 1 ms). Here, the set and reset voltages are fixed at −3 and 3 V, respectively; and HRS and LRS are measured with a read voltage of 0.2 V both before and after applying the set or reset voltage. Figure 5a,b shows the corresponding resistance values in five cells when switching from HRS to LRS, where the blue and pink boxes represent the value of HRS and LRS, respectively. Comparing the HRS (blue) before the application of the set pulse, D3 maintains a consistent  resistance state while the HRS of D1 is not uniform from cell to cell. As shown in Figure 5a, D1 does not switch to the LRS when a pulse having a pulse width of 100 μs or less is applied, while in D3, a set failure occurs at the pulse width of 100 ns (Figure 5b). Additionally, the remarkable variation in resistance states that occurs in proportion to the pulse width indicates the viability of integrating multi-level features. Similarly, the reset process is conducted in Figure 5c,d. Precisely, D1 reaches a reset failure at 100 μs in Figure 5c. D3 could not transition to the HRS when a pulse with a width of 10 μs or less is applied (Figure 5d). This demonstrates that the TiN nanocrystals' insertion into the oxide could help the oxygen vacancies to drive such that resistive switching can be faster and more reliably switched with small variations.
The programming curves of Figure 2d are re-plotted and fitted to examine the carrier transport to clarify the switching mechanism of D3 in Figure 6a. Specifically, the LRS and HRS are linear dependent for the ln(I) versus V 1/2 plot. When the temperature is increased from 298.15 to 338.15 K, the resistance is significantly where A* is the effective Richardson constant, T is the absolute temperature, q is the magnitude of the electronic charge, ϕ B is the barrier height, is the high frequency relative dielectric constant, and k is the Boltzmann constant. [44][45][46] To prove the ln(1/T 2 )sqrt(V) relationship fitting, the Schottky conduction equation is transferred to: where ( is the slope value and q B kT is the intercept of the straight line. The extracted barrier height is changed from 12.64 to 9.24 when the device is changed from HRS to LRS. [44] Since the energy barrier is lowered, thermally activated electrons are injected into the conduction band of the oxide, changing the resistance subjected by the Schottky conduction. [45,47] As illustrated in Figure 7a, the human brain is composed of neurons and synapses that play a role in the interactions between two adjacent neurons. Specifically, the synapse acts as the weighted transfer of spikes from a pre-neuron to a postneuron to convey information. Neurotransmitters are released from all synapses in biological synapses in response to stimulation, and they diffuse to post-neurons to produce stimulation. Its configuration is mirrored by and performs similarly to two terminal electrical RRAM. [48] In particular, when the impulse is given to the top electrode, the bottom electrode verified that a resistance change occurs as a result of the rearranged electrons and oxygen vacancies of the switching layer. [49] Synaptic www.advancedsciencenews.com www.advmatinterfaces.de plasticity is the term used to describe the process of considering conductance as synaptic weights and modifying weights through pulse simulation during pre-synaptic and/or post-simulation stimulations. [50,51] Furthermore, to emulate fundamental synaptic functions, pulse analyses were used to assess the synaptic plasticity characteristics of artificial synaptic devices. Specifically, the multi-level characteristics are regarded as the most significant function to potentiation and depression features in synapses among the various and essential synaptic functions. [52,53] Particularly, receiving stimulus impulses strengthen the link between the two following neurons, which will cause potentiation and depression. [50] The LTP/LTD characteristics are shown in Figure 7b, and conductance changes were induced by applying 50 identical pulse signals. To imitate the LTP behavior, successive pulses of −2.2 V/1 ms were used to increase the conductance. Precisely, negative pulses with a pulse height of 2.5 V and a pulse width of 1 ms were used to mimic the LTD behavior. By performing five additional cycles with the same potentiation and depressing pulse trains, the repeatability from cycle to cycle was demonstrated. The excitatory postsynaptic current (EPSC) of the response of the device to a stimulation train of inter-spike time, frequency, and intensity is shown in Figure 7c-e. First, paired-pulse facilitation (PPF) is a sort of short-term synaptic plasticity that reflects responsiveness to the second pulse when the first and second pulses are close enough. Specifically, with an inter-spike time difference of 100 s to 1 ms, two presynaptic spikes (−1.7 V/10 μs) are delivered to the presynaptic. A minor postsynaptic response is produced by the first spike, which is followed by a larger response at the second spike. Indicator PPF defines: In Figure 7c, the pulse interval-which is the product of the current to the second spike and the magnitude of the first spikedetermines the PPF. Particularly, as the inter-spike period lengthens, the PPF index value gradually drops. The initial spike causes oxygen ions to remain somewhat close to the TiN nanocrystal interface, while the second spike causes oxygen ions to grow even more, leading to PPF in the TiN-NC interface.
Pulse amplitude-dependent conductance modulation is shown in Figure 7d. Specifically, each of the six applied pulse trains has a distinct pulse height but maintains the same pulse width (10 μs) and pulse interval (40 μs). More precisely, the six trains are built with −1.5, −1.7, −2, −2.2, −2.5, and −2.8 V for synaptic weight loss changes, while pulse height is increased by 0.2 V from 2.4 to 3.4 V for growth. The synaptic weight associated with changing conductance has the following definition: where W t is the conductance before the pulse trains and W i is the conductance after the pulse train. It can be seen that the synapse weight varies as the voltage amplitude increases in both set and reset when 5 cycles of the configured pulse are repeated. In addition, we assess the spiking-rate-dependent plasticity (SRDP), which is the response potency of the post-spike that depends on the pre-spike stimulus frequency (Figure 7e). Particularly, different intervals (20 μs, 50 μs, 100 μs, 500 μs, 1 ms, and 5 ms) are configured for the same five pulse trains with voltage sizes of −2.3 or 2.5 V. Before and after the stimulus pulse, a read voltage of 0.2 V is used to investigate whether the synaptic weight declined or increased. Specifically, the synaptic weight computation confirms that it decreased as the pulse interval increased, just like the spike amplitude-dependent plasticity (SADP).
To implement the RRAM with TiN-NCs inserted in 3D Vertical RRAM, we verified that it not only has stable resistive switching characteristics but also functions as a synaptic element. The mimetic diagram and cross-section of the element into which the TiN-NCs are formed are shown in Figure 8a,b. Specifically, the vertical resistive random-access memory (VRRAM) was fabricated using the same stack as a single cell, and the process method has been presented in Figure S4, Supporting Information. VRRAM is formed on the side of the plane electrode (BE) in the trench hole with ALD-deposited switching layers, in contrast to single cells with a planar structure, where bottom electrods, switching layers with TiN nanocrystal are deposited in turn, and only top electrodes are patterned. A trench hole with a diameter of 10 μm that is made by etching a multilayer that alternately deposits TiN plane electrodes and electrical isolation layers (SiO 2 ), defines memory cells in a 3D VRRAM device. M2 refers to a plane electrode positioned in between SiO 2 insulating layers, whereas M1 refers to a plane electrode deposited directly on the passivation SiO 2 layer. Fabricated 3D device is depicted in Figure 8c by a TEM image and the EDS should be examined as shown in Figure S5, Supporting Information. TiN nanocrystals have been encapsulated between the HfO x switching layers as shown in Figure 1a.
The electrical properties in two layers of a 3D vertical RRAM are presented in Figure 8d,e. In each of the cells of M1 and M2, the current steadily increases when a negative bias is applied, switching from HRS to LRS, and on the other hand, switching back to HRS by the application of a positive bias. Similar to a single cell, a two-layer cell exhibits analog switching behavior with a progressive shift in current. When comparing Figure 2d and Figure 8e, the HRS current in VRRAM (152 μA) is significantly higher than the single cell (55 μA), and as a result, the on/off ratio of VRRAM is 1.85 on average, which is around 10 times lower than the single cell (10.4). Figure S6, Supporting Information shows a forming curve, and a filament is formed at a higher voltage as compared to the single cell of Figure 2a. It is attributed to the reduced cell size and array structure including long metal lines and many cells. Particularly, the HRS current increases sharply with the forming process, which reduces the on/off ratio and has a limited memory window.
The potential and depression characteristics for 5 cycles of VRRAM are shown in Figure 9a. Because of the device performance degradation brought on by the modification to the aforementioned VRRAM structure, the change in conductance decreases as the cycle goes on. However, it is clear that the conductance gradually changes, just like in the LTP/LTD characteristics of a single cell. The retention characteristics of the cells in each layer are displayed in Figure S7, Supporting Information, which also exhibits a low on/off ratio while sustaining each resistance state for 10,000 s. It indicates that non-volatile characteristics are  similar to those of single devices and enable TiN-NCs insertion in 3D VRRAM to improve memory characteristics.
A pattern recognition simulation is performed to compare the potentiation and depression results of VRRAM with a single cell. Figure 9c depicts the Deep Neural Network (DNN) structure built with the MNIST (Modified National Institute of Standards and Technology database). A neural network system evaluates the LTP/LTD properties and assesses its ability for pattern recognition. In a neural network, an artificial weight is developed using the measured conductance values. A sample image of 28 × 28 pixels is utilized, and the input node number is 784, which equals the number of pixels in the MNIST binary image. The input data is sent to the output layer after a nonlinear transformation has been carried out in the three hidden layers between the input layer and the output layer while taking the weight parameter into account. The output layer, which consists of 10 nodes numbered 0 through 9, produces the result. Figure 9d displays the outcomes of the MNIST pattern recognition simulation utilizing the 3D VRRAM conductance data ( Figure 9b) and a single cell (Figure 7). Since 3D VRRAM yields approximation results compared to single cells, it can implement synaptic properties suitable for neuromorphic systems even in memories of vertical structures.

Conclusions
In this study, we enhanced the electrical properties of the TiN/HfO x /TiN-NCs/HfO x /TiN RRAM device such as the retention, DC endurance, and switching speed. In addition, we demonstrated scalability with a 3D vertical RRAM structure embedded with NCs. By controlling the stochastic characteristics of filament production in the switching layer, ALD-deposited TiN-NCs alleviated cycle-to-cycle and device-to-device variation. Additionally, the device exhibits more stable switching and a faster transition speed due to the TiN-NCs. Further, the conduction process in the switching process has been thoroughly studied. Specifically, artificial synaptic features are emulated by varying conductance with electric pulse schemes. This simulated synaptic functions such as spike-rate-dependent plasticity, paired-pulse facilitation, and spike-amplitude-dependent plasticity, proving its potential as an essential component of the upcoming neuromorphic system. Finally, this process is extended to a 3D vertical RRAM structure with TiN-NCs to prove the possibility of highdensity storage memory and synaptic devices.

Experimental Section
Single Cell: The HfO x -based RRAM devices were prepared by inserting ALD-deposited TiN NCs between the switching layers. Additionally, this work manufactured the RRAM of HfO x without NCs for the control devices. Initially, a 100-nm-thick TiN as the bottom electrode (BE) was deposited on the SiO 2 /Si substrate by reactive sputtering, following a sulfuric acid-peroxide mixture (SPM) and hydrofluoric acid (HF) cleaning process. Specifically, the atomic layer deposition (ALD) (NCD, Lucida M300PL-O) was used for the growth of a 10-nm-thick HfO x thin film using trimethylaluminum (TDMAHf) as a precursor and ozone (O 3 ) as a reactant. The sequence proceeds as follows: TDMAHf feeding → N 2 gas purge → O 3 feeding → N 2 gas purge. This was repeated for 97 cycles at the 350°C stage temperature to deposit HfO x , atomic-by-atomic. After a 2-nm-thickness of TiN nanocrystal layer deposition by ALD with TDMATi and NH 3 , heat treatment was performed with rapid thermal annealing (RTA) for 30 s at 400°C. Overall, a 10-nm-thick HfO x was grown in the above method to cover the TiN-NCs after the formation of TiN nanocrystals. Finally, a 100nm-thick TiN top electrode (TE) was deposited by radio frequency (RF) sputtering after photolithography. More precisely, the lift-off process of TE was performed with a square pattern of 100 × 100 μm. Specifically, the DC I-V and pulse characteristics of the device were measured using Keithley 4200 SCS, a semiconductor parameter analyzer, and a 4225-PMU ultrafast module. In addition, the TiN BE was grounded during all measurements, and the bias voltage was applied to the TiN TE.
3D Vertical RRAM: The following procedures were used to fabricate the TiN/HfO x /TiN-Nanocrystals/HfO x /TiN VRRAM device. Four layers were generated by repeatedly depositing plane electrodes for multilayer cells and insulating layers for electrical isolation. Utilizing RF reactive sputtering and PECVD, respectively, the TiN plane electrode and SiO 2 dielectric layer were alternatively deposited. The trench holes were patterned and etched to produce multilevel RRAM cells. SiO 2 layers were dry etched with CF 4 /Ar plasma and TiN layers were dry etched with Cl 2 /Ar plasma using a reactive ion etching system (Oxford RIE 80 Plus). Then, the switching layers were deposited by thermal ALD (NCD, Lucida M300PL-O). TEMAHf was used as a precursor and ozone (O 3 ) was used as a reactant at 350°C for depositing a 10-nm-thick HfO x layer. After depositing using ALD with TDMATi (precursor) and NH 3 (reactant), a 2-nm-thick TiN nanocrystal layer was created by RTA (Rapid thermal annealing) at 400°C for 30 s. The nanocrystal layer was then covered with a HfO x switching layer, which was deposited under identical conditions as before. A 200-nm-thick TiN electrode was deposited by RF reactive sputtering. The contact pads were patterned and etched so that ground could be applied to the plane electrode. Contact etching was performed in the same manner as hole etching, with TiN using Cl 2 /Ar plasma and SiO 2 using CF 4 /Ar plasma. SiO 2 on the top floor was etched for the contact pads of M1 and M2, after which the contact pad of M2 was covered with photoresist and only the contact pad of M1 is opened by sequentially etching TiN and SiO 2 layers.

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