Probing switching mechanism of memristor for neuromorphic computing

In recent, neuromorphic computing has been proposed to simulate the human brain system to overcome bottlenecks of the von Neumann architecture. Memristors, considered emerging memory devices, can be used to simulate synapses and neurons, which are the key components of neuromorphic computing systems. To observe the resistive switching (RS) behavior microscopically and probe the local conductive filaments (CFs) of the memristors, conductive atomic force microscopy (CAFM) with the ultra-high resolution has been investigated, which could be helpful to understand the dynamic processes of synaptic plasticity and the firing of neurons. This review presents the basic working principle of CAFM and discusses the observation methods using CAFM. Based on this, CAFM reveals the internal mechanism of memristors, which is used to observe the switching behavior of memristors. We then summarize the synaptic and neuronal functions assisted by CAFM for neuromorphic computing. Finally, we provide insights into discussing the challenges of CAFM used in the neuromorphic computing system, benefiting the expansion of CAFM in studying neuromorphic computing-based devices.


Introduction
It is well known that there is an inevitable degradation in the overall computing performance due to the von Neumann bottleneck induced by the separated memory and computing unit when processing complicated data and reducing power consumption [1][2][3][4][5]. In contrast, mimicking the working mechanism of the human brain is an effective way to process comprehensive and complex information such as sound localization [6][7][8], image recognition [9][10][11], tactile perception [12], and so on. Therefore, neuromorphic computing gradually attracted researchers' attention in recent years [13][14][15][16].
Neuromorphic computing systems perform data storage and processing by mimicking the human brain's biological structure, including synapses and neurons, using electronic devices [17][18][19][20][21][22][23][24]. The concept of memristors has been proposed by Leon O. Chua in 1971 [25], in which they represent the relationship between magnetic flux and electric charge. Normally, it is an essentially non-linear resistor with memory function, and the corresponding resistance can be changed by controlling the voltages. The memristors can store data by defining high/low resistance state as '0'/'1', which has been recognized as one of the potential electronic devices in high-density memory technology and neuromorphic computing [26][27][28][29].
Due to its resistive switching property, one application of memristors is the artificial synapse. The resistance of the memristors represents the strength of the connection between neurons, namely synaptic weight [30][31][32][33][34]. By applying external voltages, the resistance state will change corresponding to the plasticity of synapses, such as short-term synaptic plasticity (STSP) mainly including short-term potentiation (STP) and short-term depression (STD), long-term synaptic plasticity (LTSP) including long-term potentiation (LTP) and long-term

Conductive atomic force microscopy
CAFM is developed based on atomic force microscopy (AFM). The working principle of AFM is that the contact force between an ultrasharp tip at the end of a cantilever and a sample leads to the cantilever deflection using an optical or a piezoresistive system [69].
The magnitude of the interaction force between atoms is related to the distance between the tip and the sample. According to the literature [74,75], for distances larger than 0.5 nm, the detected forces belong to attractive, which means the noncontact mode. On the contrary, for distances smaller than 0.3 nm, the interaction forces are repulsive, which can be called the contact mode where the tip physically contacts the sample. In contact mode, the inevitable friction between the surface of the sample and the tip of the probe is much larger when obtaining a higher vertical resolution. So, the tapping mode that combines the benefits of both of them has been developed [76,77], which is the improvement of the noncontact mode in that the resolution is between contact mode and noncontact mode, significantly reducing the probability of destruction by lateral forces.
Concerning most AFM, it is common to use an optical system to detect deflection (figure 1(a)) [78,79]. When the tip is close enough to the sample, the cantilever deflects and the laser spot is deflected. However, the interaction between the tip and sample is undetectable when the tip is far from the surface, therefore, the cantilever has no deflection and the laser spot does not move. By analyzing the position of the laser spot on the photodiode, the force can be quantified, and information about the topological structure can be obtained. In addition, AFM can also use the pressure sensor at the tip cantilever to detect. When the cantilever beam is deflected, the surface height can be quantified with the change of the resistance of the piezoelectric sensor to obtain sample information, avoiding installing a photodiode above the tip.
However, AFM cannot obtain the electrical characteristics of the sample surface due to the non-conductive probe, which limits the application of AFM. For example, AFM cannot be used to directly detect the formation, dissolution, or distribution of local CFs under different resistance states. In contrast, CAFM technology can detect electrical characteristics of the sample with its conductive probe, such as spatial conductivity distribution, charge transport and distribution, and so on at the nanometer scale [80]. Figure 1(b) displays the basic structure of CAFM [79]. Different from the AFM, the three new elements are the conductive nanoprobe, preamplifier, and voltage source. The voltage source is used to apply a potential difference between the tip and the sample holder, and a preamplifier can convert the current signal into voltage signals that can be read by the computer. Using this setup, when a potential difference is imposed between the tip and the sample, an electrical field is generated, which results in a current flowing from the tip to the sample, monitoring the local electrical properties of the samples. Therefore, compared with AFM which can characterize topological structures at a nanometer scale, almost all CAFM can make it possible to collect topography and currents flowing between the tip and the sample simultaneously and independently. The first time to obtain electrical information in an AFM is 1993 [81]. For some time after its invention, CAFM is usually used to study ultra-thin dielectrics (especially high dielectric constant materials [62]) and some related phenomena. With the continuous development of technology, nanomaterials (such as quantum dots [82], nanowires [83], 2D materials [33], etc) and electronic phenomena (such as piezoelectricity [84], photoelectricity [85], ferroelectricity [86], etc) are also analyzed. In sum, CAFM can be used to collect topographic and current maps simultaneously and independently. The topographic map is obtained by measuring the deflection of the cantilever at each location of the sample, and the current map is built by recording the current using a current-to-voltage preamplifier [87]. CAFM offers the possibility to measure the electronic properties of electronic devices, such as transistors, memristors, photodetectors, and so on.

Methodologies of observation based on CAFM
As discussed above, CAFM can provide morphological and electrical information of samples which may play a critical role in contributions to the research on the RS mechanism of memristors. So far, there are three main methods of CAFM used to observe RS in memristors based on related research in the past [62,88].
The first method is to program a set or reset process at the device level, etch the top electrode and then scan on the exposed surface of the dielectric stack using CAFM (figures 2(a)-(c)). This method can accurately analyze the CFs generated in the two resistance states and reveal the internal conduction mechanism. In 2005, Choi et al performed this method for the first time. Some capacitors based on the structure of Al/TiO 2 /Ru were switched either to the high resistance state (HRS) or low resistance state (LRS) under a bias voltage. After a nitric acid solution was used to etch the Al top electrodes, CAFM was used to obtain the surface morphology and local conductivity of TiO 2 films, showing the correlation between RS and the formation/rupture of conducting spots [89]. However, this method requires an etching way to remove the top electrode with particularly little damage on the CFs and the surfaces of the dielectric layer. In past reports, wet etching, dry etching, and focused ion beam (FIB) were reported as the most common etching methods [89][90][91]. But it will inevitably cause non-negligible damage to the dielectric layer and CFs in the process [92]. Due to the etching process required, this method can The subsequent read scanning is performed on the exposed dielectric layer by the CAFM tip to find local conductivity changes. (d) The bias voltage is applied through the CAFM tip at a random position to induce the point set process. (e) After the set process, the read process is performed at the same location to analyze the resistance state of the dielectric layer. (f) Subsequent set process can be implemented by moving the tip to another random location. (g) The local area scan uses the CAFM tip under a bias near the V set to achieve the small-scale set process. (h) The read scanning is applied for the local conductivity analysis. (i) The reverse bias voltage is applied through the CAFM tip to achieve the reset process. Based on the process (g) and (i), one can realize arbitrary resistive distribution in the selected area.
only provide statistical and non-reproducible information on the properties of the CFs in the HRS and LRS, but fail to analyze the properties of a single filament in both HRS and LRS states.
The second method is that the CAFM tip is placed directly on the dielectric stack. By applying voltage through CAFM tip at a random point position, the CF can be observed and the properties of individual CF can be analyzed (figures 2(d)-(f)). In other words, it is completely certain that the collected current only corresponds to the CFs under the CAFM tip. Waser et al analyzed the local conductance with tip contact on the SrTiO 3 directly using CAFM firstly as a nanoelectrode, addressing single dislocations crossing the surface with adequate spatial resolution [93]. When applying an I-V curve, the experiment requires multiple sweeps of random locations to obtain the switching behavior to represent the entire dielectric [94]. Lanza et al reported similar work and analyzed local I-V curves with a voltage bias applied by the CAFM tip on the HfO 2 layer at different locations, and summarized that RS was only observed at the electrically leaky sites located at the grain boundaries [95]. Muenstermann et al scanned the surface of SrTiO 3 randomly, obtaining conductivity images that displayed many small and well-randomly distributed conducting spots. With the whole image scanned, all the pixels were achieved sequentially for getting statistical information [96].
The last method is to apply a bias voltage near the set voltage at a specific area by scanning with the CAFM tip to achieve the set process. The read and reset process can also be achieved by similarly local scanning and a suitable voltage (figures 2(g)-(i)). Yoshida et al measured tip-induced electrical RS, explaining the cause of the large change in conductance as one of the earliest experiments to use this method [97]. But there are some drawbacks to RS caused by the tip. For example, a small voltage applied by CAFM may not be enough to induce set and reset processes. In addition, the current in CAFM is not aware of the actual current between the sample and the probe. The true current flow of the tip/sample system is not limited by the equipment, which may cause dielectric damage. To solve these two problems, the SPA was introduced [98][99][100][101], which was first reported by Blasco et al in 2005 [102]. The good performance of the conductive tip is proved by integrating a CAFM and a SPA under a high current. Moreover, a sufficiently stable tip is also required to reduce the impact of tip wear on the experiment because the tip of the probe directly contacts the electrolyte layer, such as the diamond coating tip, graphene coating, etc.
In summary, each of the above three methods has advantages and disadvantages under different situations. The method of inducing RS at the device level by CAFM provides statistical and more representative information on CFs in HRS and LRS. But CFs cannot be studied cyclically, because additional set/reset at the device level is impossible with the top electrode removed. Due to the etching of the top electrode, ineluctable damage is made to the dielectric and/or CF, and analyzing of device-to-device variability is complex and timeconsuming. CAFM tip can be placed directly on the dielectric stack to induce RS without requiring etching the top electrode. By local point scanning randomly, RS can be observed in an in situ way and information can be obtained on a single CF with the research area down to 1-100 nm 2 . In addition, by collecting current maps, CAFM can carry out statistical analysis on certain areas and currents. But it's important to note that the unstable tip and almost inevitable dielectric pollution result in a limited amount of cycles by using tip scanning on the sample. The current flowing through the tip/sample system needs to be limited to prevent dielectric damage. So, appropriate methods need to be selected based on the actual situation.

Mechanism of memristor probed by CAFM
Memristor is normally designed as metal/insulator/metal (MIM) in a planar or vertical orientation [103][104][105][106]. Atoms, ions, holes, etc in the dielectric layer will move to change the resistance of the dielectric layer by applying electrical bias, thereby a switch between high resistance and low resistance is formed. Generally, mechanisms based on MIM structure can be divided into two different types [107][108][109][110]. One is that RS occurs uniformly in the dielectric layer, named distributed RS, and the other is a local phenomenon that occurs in small regions on the order of ∼10 nm 2 , named filamentary RS [55,107,108,[111][112][113][114]. CAFM is more suitable to study dielectric layers in which the RS is based on the formation and disruption of one/several CF/CFs. CAFM provides direct evidence demonstrating the filamentary switching mechanism toward understanding the possible RS mechanism in the device. Local filamentary RS mechanisms reported can be mainly divided into electrochemical metallization (ECM) and valence change mechanism (VCM) [23].
ECM mainly consists of the key dielectric layer sandwiched between the active electrode and the inert electrode [115][116][117][118]. With a positive bias applied to the active metal electrode, the active metal is oxidized into metal ions, dissociating from the metal electrode. The ions can diffuse into the dielectric layer and move to the inert metal, where they are reduced and form filaments. Therefore, the CFs connect the top and bottom electrodes with the resistance at LRS. On the contrary, when the reverse voltage is applied, HRS is formed with CFs breaking.
Celano et al explored the ECM based on the structure of Cu/Al 2 O 3 /TiN by using the special tomography technology of CAFM which gives specific detailed 3D surface information ( figure 3(a)). The typical memristive switching properties are shown in figure 3(b). The CAFM tomography image is obtained by integrating slices which are obtained by scanning the surface of Al 2 O 3 at different depths under a constant load force after removing the top electrode. The image shows that a tapered CF is formed with the narrow end close to the inert electrode. In figure 3(c), the internal mechanism is explained based on the CAFM tomography image. The Cu atoms are oxidized with a positive voltage applied to the Cu electrode. Due to the oxygen vacancies generated in Al 2 O 3 by a high electric field, Cu cations move in a short distance in the dielectric before they are reduced by capturing the electrons. Therefore, Cu cations are reduced to Cu near the active electrode, resulting in the CF growing from the active electrode to the inert electrode and forming LRS. When the negative voltage is applied, the CF breaks due to the process of electrochemistry and joule heating, and the HRS of the device is achieved [115].
Similarly, Hu et al showed a highly controllable memristor of Cu/annealed ZnS/Pt with an ultralow SET voltage and demonstrated the internal ECM conductive mechanism. Figure 3(d) is an AFM and CAFM diagram of the OFF and ON states. The small peak marked by the blue arrow in the ON-state AFM image is residual Cu after removing the Cu top electrode. According to the corresponding CAFM image, an increased current is observed at this location indicating that only one complete Cu filament is formed. Based on this, figure 3(e) illustrates the ECM of memristor based on Cu/annealed ZnS/Pt. Under a positive bias voltage to Cu through the tip, once the filament is formed between Cu and Pt by the migration of Cu ions, the device is switched to the ON state, as shown in the CAFM image of ON-state in figure 3(d). By applying an appropriate negative voltage, the CF breaks and reaches the OFF state due to the Joule heat-induced external diffusion or oxidation of copper [119].
In addition, RS can also occur through a VCM. Compared with ECM, the dielectric layer of VCM is sandwiched between inert metal electrodes [120][121][122][123]. The typical dielectric layer is a metal oxide. Taking oxygen ion as an example, the distribution and defects of oxygen atoms in metal oxide will affect the formation of filament. When the device is driven by an electric field, oxygen atoms are reduced to oxygen anions, creating vacancies at the original oxygen sites, which become highly conductive paths. With the recombination of the oxygen anions and vacancies, low conductive paths are generated.
Qian et al demonstrated a transparent ITO/WO 3 /ITO memristor that can achieve good memristive properties (figures 4(a)-(b)). The CAFM image shows the OFF and ON states of the WO 3 /ITO glass structure with a diamond-coated tip as the top electrode. When the device is in its initial state by the CAFM tip with read voltage applied, the current is evenly distributed across the region. In contrast, by increasing the voltage by the CAFM tip to the SET voltage, the device switches to the ON state with CFs of tungsten. Combining the CAFM The oxygen ions at the top electrode interface will migrate back to recombine with the oxygen vacancies in the WO 3 layer with the help of the reverse voltage, resulting in the rupture of the filaments and the realization of a HRS (RESET process). This experiment provided general guidance on the switching mechanism of valence state changes in fabricated metal oxide memristors [124].
Additionally, Aleksandra et al carefully studied the RS behavior of the memristor with the structure of Ru/TaO x /Ta and investigated the localized conduction paths of the TiN/Ru/TaO x /Ta device. The CAFM tip is applied to the partially oxidized Ta layer to obtain overlapped topographic and local current images of LRS, as well as 3D representations of corresponding overlapped images. It can be seen that local conduction mainly occurs on the side of prominent grains, indicating that CFs are formed in local field enhancement regions. As shown in figure 4(d), when Ta is deposited on the TaO x layer, oxygen vacancies are created at the TaO x /Ta interface. With a positive bias applied to the top electrode, oxygen vacancies migrate and accumulate locally towards the bottom electrode interface to form CF on the side of prominent grains. Thus, LRS was achieved when CF reached the top electrode. Conversely, under the negative bias polarity, oxygen ions may migrate back towards the bottom electrode to recombine with the oxygen vacancies which switches the device to HRS with the CF rupturing [125].
In addition to a single RS mechanism, there are memristors with both mechanisms [126][127][128]. Chang et al proposed Ag/Ta 2 O 5 /Pt device with typical bipolar characteristics and high stability (figure 5(a), with Ta 2 O 5 as the dielectric layer and Ag as the active electrode. The I-V curve in figure 5(b) shows the forming, set, and reset process of RS behavior. The SET process is analyzed by in situ TEM, EDS, and EELS. By applying a positive bias voltage to the Ag electrode. The CFs formed in the Ta 2 O 5 layer from the Pt electrode extended to the middle of the Ta 2 O 5 layer could be observed. EDS and EELS also showed the existence of oxygen vacancies, indicating that oxygen vacancies can migrate during the formation of Ag CFs. So, the forming of compound CFs enables the device to achieve the SET process. In addition, with a negative bias voltage applied, the RESET process can be achieved. CFs were observed based on 3D scanning images of HRS obtained by using CAFM tomography technology, proposing abrupt and progressive CFs fractures [129]. So, the fracture of CFs has been divided into two steps in the RESET process, with the partial dissolution of CFs at 0.6 V and complete fracture of CFs at 1.2 V. The figure 5(c) illustrates the specific process of CFs formation. In the initial state of the device without bias voltage, oxygen atoms are uniformly distributed in the Ta 2 O 5 layer. When an external electric field is applied to the device, the Ag electrode is oxidized to generate Ag cations, which migrate to the Pt electrode and obtained electrons at the Pt electrode. Reduced Ag atoms form a filament. However, the Ag ion concentration is limited in low humidity conditions, resulting in the formation of only half of the Ag filament. As the voltage continued to increase and the local electric field (part of the Ag filament produced) increased, VCM occurred, forming an oxygen vacancy filament that is connected to the rest half of the Ag filament ( figure 5(c)). The dual-filament mechanism enabled the memristor to be switched in different humidity environments, broadening the electronic applications of memristors [130].
Khot et al proposed an Ag/a-BN/Pt memristor device and explained that there may be two internal mechanisms at the same time by using CAFM, namely ECM (redox reaction of Ag) and VCM (movement of boron vacancy which is similar to oxygen vacancy). As shown in figure 5(d), the current images measured by CAFM during the high and low resistance states show that different resistance states have different current levels and different filament formation conditions. The surface of a-BN exhibits a current of picoampere magnitude in a HRS and a current of the order of nano amperes in a LRS. Meanwhile, the device in the LRS can have multiple CFs. Based on the above, the internal mechanism is explained. When the external electric field is applied to the device, the Ag electrode oxidizes to Ag cations, which migrated to the Pt bottom electrode, reducing to silver atoms to form a CF. In addition, the boron vacancy uniformly distributed in the a-BN layer is transferred to the Ag top electrode due to its net negative charge forming CF. On the contrary, when the CAFM-assisted negative bias voltage is applied to the top electrode of Ag, the Ag ions migrate to the Ag electrode, and the boron vacancy flows to the Pt electrode, resulting in the breakage of the CF. Therefore, the device completed dual RS [131].

CAFM-assisted neuromorphic computing
As mentioned above, artificial neural networks can simulate biological nervous systems by using memristors [137][138][139]. Figure 6(a) reported a two-terminal memristor used for mimicking biological synapses [140]. Compared with traditional electrical test techniques, i.e., probe station + semiconductor parameter analyzer, the main advantage of CAFM is that by using the tiny conductive AFM probe (radius<20 nm), we are able to obtain typical MIM memristive cell with nano sizes of ∼50 nm 2 , and achieve in situ detection of the synaptic and neuronal behaviors by CAFM [141,142]. It is noted that in the above MIM cell, the conductive tip is served as a top electrode, and the contact area (between the tip and insulating layer) equals the effective size of a single device. By this approach, we could not only avoid complicated fabrication processes of nanosized memristive devices but also promptly evaluate their synaptic behaviors or performance in a more reliable way (i.e., in situ). The tip of CAFM can be used not only to apply external voltage but also as a top electrode to more effectively obtain some important parameters of memristors in neuromorphic systems in terms of endurance, retention, cycle-to-cycle, device-to-device variability, and more. To express more data intuitively, here we made a table about concrete parameters of different devices obtained by using CAFM (table 1)  would be improved with the probe acting as the top electrode by using CAFM and finite element simulations (figure 6(d)). In the case of probe contact, the filament would be positioned and well confined, suggesting that point contact could ultimately provide a unique way to design high-performance devices [143]. Recently, Hui Fei et al provided a new method to explore nano-synaptic plasticity at the nanoscale. Based on the metal/ graphene oxide (GO)/metal structure, they constructed nanoscale electronic synapses in two configurations. One was to place a bulk Pt tip on the surface of a GO/Au/SiO 2 wafer to form a Pt/GO/Au/SiO 2 device, and the other was to place a GO-coated bulk Pt tip on the surface of the Cu film, forming Pt/GO/Cu/SiO 2 nano synapses (figure 6(e)). By applying a series of pulses using CAFM and connecting SPA, volatile and non-volatile RS were applied to achieve synaptic functions. Figure 6(f) shows the simulation effect of paired-pulse facilitation (PPF) by applying PVS. Figure 6(g) reveals that the Pt/GO/Au exhibited stable LTP and LTD, which is necessary for LTSP. EPSC response is also exhibited in Pt/GO/Au nano-synapses in figure 6(h), which represents that peak potential or action potential is transmitted from presynaptic neuron to postsynaptic neuron [132].
The biological neurons can receive, process, and transmit signals through the change of cell membrane potential. When the membrane potential reaches the threshold by receiving external stimulation in a given time interval, the ions start to move with the ion channel opened, and the neuron will produce an output spike to achieve signal transmission [144]. After that, the membrane potential returns to the equilibrium state. The dynamic behavior of this neuron potential can be described by the leaky integrate-and-fire (LIF) model [145][146][147]. In figure 7(a), the threshold behavior of the volatile memristor corresponds to the threshold behavior of the ion channel, and the capacitor in parallel with the memristor represents the lipid membrane. When the charging of capacitor makes the voltage across the memristor reach the threshold voltage, the device is turned on to allow resistance state, and then the capacitor discharges to produce an output spike [148]. CAFM can directly detect the threshold switching behavior for neuron circuits.
For example, Liu et al proposed a volatile RS device based on copper nanoparticles (CuNPs) modified with alkyl dithiophosphate (DDP) ( figure 7(b)). The ramp voltage stress (RVS) was applied to the sample by using the Pt tip of CAFM ( figure 7(c)). The device exhibits threshold switching characteristics, and its SET and RESET   When applying continuous input voltage pulses, they simulated the electrical behavior of LIF neurons. Figure 7(e) shows that, after several pulses are applied at the input, the output produces a voltage spike [133]. In addition, Chen et al used CAFM to achieve the excellent volatile threshold RS based on Ag/h-BN/Ag memristor device, as shown in figure 7(f), which can be used as a low-power integrated LIF artificial neurons in spiking neural networks due to the self-reset process of threshold RS devices with high I LRS /I HRS and low energy consumption [149].

Prospects and challenges
In summary, this article summarizes the features of CAFM and its application in neuromorphic computing devices. We introduce the basic principles of CAFM and discuss the methods of using CAFM to observe the switching mechanism of memristors, further expounding the conductive working mechanism. According to the type of CF, it can be divided into ECM, VCM, and a combination of these two types. Finally, we introduce several typical synaptic and neuronal behaviors realized by using the CAFM probe technology, and the probe can be directly used as the top electrode, which provides an idea for characterizing neuromorphic computing devices by CAFM in a localized area. As discussed in this work, although CAFM has played a certain role in memristor and neural network systems, CAFM technology still faces great challenges which are still needed to be solved.
(1) Improve temporal resolution of CAFM; the ultra-high temporal resolution is required to understand the specific details of switching dynamics in memristive devices, especially for in situ characterization techniques, requiring an ultrafast tracing ability where many physicochemical processes are involved.
(2) Compatibility of CAFM with other techniques; when the internal mechanism of the memristor involved is complicated, it is necessary to combine different techniques or different instruments to better reveal the mechanism behind, such as the integration of multi-probe CAFM into SEM is beneficial to improve detection performance and accuracy without increasing the overall complexity and cost.
Therefore, we need to further optimize the CAFM technology and expand the application range of CAFM. For example, applying mechanical stress by using CAFM technology in flexible electronic devices is helpful to study new nano phenomena. In addition, CAFM is expected to be used to collect electrical current diagrams at different vertical heights and build 4D images to more intuitively understand the internal switching process. Moreover, CAFM could be combined with other probe technologies for multifunctional in situ nanofabrication  [149]. Copyright (2020) Springer Nature. Reproduced from [149], Copyright © 2020, The Author(s), under exclusive licence to Springer Nature Limited. and characterization. Based on the discussion above, it is necessary to develop the applications of CAFM in memristors and small-scale synaptic devices to further realize how memristors respond to specific stimuli, which can contribute to the overall function of the system. These efforts will make vital contributions to the realization of high-density, high-capacity, and reliable neuromorphic computing devices in future.