In this section, we further discuss the utilization of engineering techniques detailed in previous sections, in relation to the use of memristor devices in applications where they function as artificial synapses, with extended characteristics. Section 4.1 describes bi-layered, doped and annealed memristors, utilized for electronic synapse applications. Specifically, synapse device characteristics, i.e. linear potentiation and depression, are quite important. To achieve these, the memristor device should possess bi-directional analog resistance states in both SET/RESET operation. Therefore, analog-state memristor engineering has been widely explored, including oxide bi-layered stacking, metal doping, post-annealing, and rapid thermal annealing. Section 4.2 discusses analog multi-state memristor devices and their synapse characteristics. The structural engineering of a cone-shaped n-ZnO memristive Schottky diode produced a very low-power multi-resistance state synapse device with frequency-related synapse functioning, in accordance with the Bienenstock, Cooper, and Munro (BCM) theory. Section 4.3 argues that both non-volatile and volatile memristor characteristics can be utilized for synapse devices. Specifically, memristor devices based on metal cation migrations, such as Ag+/Cu+, can be RS switched in two modes: non-volatile and volatile regimes. These can be utilized to emulate such important synaptic characteristics as short- and long-term potentiation (STP & LTP) and its transition, and spike-time dependent plasticity (STDP), all in a single memristor device, which has been applied in a real crossbar structure (10 × 10). Section 4.4 elaborates on volatile / non-volatile low-dimensional organic-based memristor devices utilized for synapse applications, i.e. paired-pulse facilitation (PPF), STP & LTP, STDP, etc. In addition, basic neuron device characteristics are demonstrated, i.e. delay-shoot-relax, leaky-integrate and fire (LIF) neurons, etc.
Bilayer, doped and annealed memristors for neuromorphic applications
For the realization of bioinspired neuromorphic computing, the emulation of biological synaptic functions is the crucial step. In order to simulate these synaptic functions using memristive devices, the memristors are required to exhibit specific switching behaviors. In general, devices exhibiting analog switching characteristics are the most suitable devices for accomplishing bio-realistic synaptic functions. In addition, volatile threshold switching characteristics are also utilized, to mimic synaptic plasticity. As discussed in section 3, the utilization of bilayer switching materials and the structural engineering of metal-oxide switching layers by means of various techniques, such as doping and annealing, are feasible and efficient approaches to modulating the switching behaviors of memristors. In this way, the switching characteristics of memristors can be tuned based on a given application[151, 152]. Wu et al. developed a novel methodology to achieve analog switching characteristics in an HfOx device. Here, a TaOx layer is used as a thermally enhanced layer (TEL), resulting in a bilayer structure HfOx/TEL in order to realize uniform analog switching, which is useful for neuromorphic computing. Li et al. proposed a method of realizing bidirectional analog switching with a wide dynamic range of weight modulation by stacking Ag-nanocluster-doped SiO2 on a TiO2 buffer layer. The SiO2:Ag/TiO2 bilayer device demonstrated various synaptic functions. Kim et al. compared the analog switching characteristics in CeO2 single-layer and ITO/CeO2 bilayer memristors. The bilayer device exhibited linear and symmetric synaptic weight changes, with superior long-term stability in terms of modulated synaptic weight, which is essential for neuromorphic systems. This improvement in analog switching characteristics is attributed to the added ITO layer, which acts as an oxygen ion reservoir for ion migration from the CeO2 layer during switching. In another study by Chen et al., abrupt switching in a TaOx-based memristor is tuned to gradual switching by inserting a WOx redox layer between the TaOx switching layer and the top electrode, effectively mimicking essential synaptic functions. Moreover, this synaptic linearity can be enhanced by engineering the switching layer via doping. The memristor device with Al-doped HfO2 (Al:HfO2) layer, proposed by Chandrasekaran et al., demonstrated a clear improvement in synaptic weight linearity, resulting in a learning accuracy of 91% with only 13 iterations. In comparison, the learning accuracy of the undoped pure device was 78%.
As stated above, the utilization of bilayer switching layers is a facile and efficient approach to modulating switching characteristics to the gradual switching required for neuromorphic applications. In one of our recent studies, bilayer ZrO2/ZTO-based memristor devices were fabricated to achieve stable and controllable analog switching behaviors for improved emulation brain function for neuromorphic computing. Here, the amorphous ZTO layer acted as an oxygen reservoir layer, resulting in greater control of the gradual switching characteristics. The TiN bottom electrode partially oxidized to form TiON at the bottom electrode interface, leading to a non-uniform distribution of oxygen vacancies in the oxide switching layer, as shown in the TEM image of the device in Fig. 9(a). The formation of the TiON interface layer was also confirmed by XPS and XRD analyses. In general, gradual multistate switching is achieved by controlling the compliance current or RESET-stop voltage during the SET or RESET processes, respectively. Gradual analog switching with multiple conductance states has been achieved in a ZrO2/ZTO bilayer device by controlling the SET compliance current and RESET-stop voltage simultaneously, as depicted in Figs. 9(b) and 9(c). The highly desirable linearity in relation to synaptic weight (conductance) modulation is confirmed by the long-term potentiation (LTP) and long-term depression characteristics illustrated in Fig. 9(d). Furthermore, short-term synaptic plasticity behavior is confirmed by the successful emulation of paired-pulse facilitation (PPF). As shown in Fig. 9(e), pulse interval-dependent synaptic weight modulation was observed for a pair of pulses with varied pulse intervals. Spike-timing-dependent plasticity (STDP) is an essential biological process in neurobiology, and is one of the Hebbian learning rules of synapses and neurons[153, 105]. The emulation of STDP behavior by the ZrO2/ZTO memristor confirms the applicability of bilayer memristors to neuromorphic applications, as shown in Fig. 9(f).
As mentioned above, spatial (device-to-device) and temporal (cycle-to-cycle) variations in the synaptic weight modulation process, due to stochastic switching behaviors, hinder the application of memristors to synaptic emulation, as depicted in Fig. 9(g). Therefore, extensive research is currently being undertaken to develop methods of addressing the weight update linearity issue. A recent study demonstrated reliable synaptic emulation with HfO2-based memristive devices by means of metal doping and post-deposition annealing procedures. As depicted in Fig. 9(h), Al-doping and post-deposition annealing methods were utilized, which enhanced the formation of oxygen vacancies in the HfO2 switching layer, thereby improving its switching characteristics. The device with a 16.5% Al doping concentration demonstrated superior switching properties, and was subsequently utilized to achieve gradual multilevel switching, as shown in Fig. 9(i). Having achieved reliable and controllable analog switching behavior, potentiation and depression characteristics with a near-linear behavior were duly demonstrated, as presented in Fig. 9(j). Furthermore, the successful emulation of STDP behavior was also confirmed in relation to the doped memristor devices, as shown in Fig. 9(k).
Furthermore, the structural engineering of switching layers via thermal annealing can also alter switching behavior. This is a simple and efficient approach to engineering the morphology of a switching material, using appropriate annealing procedures to obtain the required switching characteristics, as well as to enahnce stochastic switching behaviors. The structural engineering of a tantalum oxide-based memristor using rapid thermal annealing (RTA) demonstrated the coexistence of reliable digital and analog switching characteristics. The as-deposited Ta2O5–x switching layer was found to be amorphous, as shown in Fig. 10(a). To engineer the morphology of the switching layer, the RTA process was carried out for 60 s at the crystalline temperature of Ta2O5, i.e., 700 °C. The appropriate RTA process changed the morphology of the switching layer, rendering it favorable for reliable analog switching. For the annealed device, polycrystalline regions with different grain boundaries were observed in the Ta2O5 layer, as depicted in Fig. 10(b). The morphological transition from an amorphous to a polycrystalline structure in the Ta2O5 and Ti/Ta2O5 interface via a simple annealing procedure resulted in reliable and repeatable analog switching characteristics. The device without RTA exhibited bipolar switching with abrupt SET and gradual RESET processes, whereas the RTA-processed device demonstrated reliable bidirectional analog switching, as shown in Figs. 10(c) and 10(d). The annealed device also displayed excellent digital switching characteristics, in addition to an electroforming process which is useful for memory applications. Repeatable potentiation and depression behaviors were demonstrated, as shown in Fig. 10(e). By utilizing the reliable analog switching characteristics of the annealed device, spike-rate-dependent plasticity (SRDP) and spike amplitude-dependent current modulations were successfully demonstrated, as presented in Figs. 10(f) and 10(g). Spike height and inter-spike interval-dependent PPF and post-tetanic potentiation (PTP) behaviors were also confirmed. Moreover, Hermann Ebbinghaus' forgetting curves were replicated using the RTA-processed device. Device evaluations, performed by applying different numbers of pulses, showed different forgetting rates, as illustrated in Fig. 10(g). Shorter rehearsals resulted in faster forgetting, and longer rehearsals yielded slower forgetting rates.
Structural oxide storage synapse memristors
Neuroscience, identifies two types of synapses in the brain: electrical, and chemical. If electrical synapses are responsible for very primitive/instinctive functions, e.g., muscle contraction, the chemical synapses are, in contrast, responsible for highly complex cognitive functioning, making them a major target for in-depth research. Typically, the synapse is a small area connection, where one neuron’s dendrite is in close proximity to another neuron’s dendrite. The synapse is a communication channel, by means of which signaling information from one neuron can be transferred to another. Chemical synapses act at a slower speed compared to electrical synapses, due to electrochemical processes occurring in the synaptic cleft. Specifically, pre-synaptic neurons sending action potentials (APs) depolarize the synapse, causing it to open the channels to enable Ca2+ ions to flow, and permitting the further release of neurotransmitters, which bind with the receptors of the post-synaptic neuron, further, depolarizing the post-synapse, which starts to transmit its AP. It is worth noting that synapse behavior is highly analog in character, including many intermediate conductance states, depending on the strength of its potentiation or depression; more specifically, synapse conductance weight can be facilitated, depressed, exhibit paired-pulse facilitation (PPF) or short-term potentiation (STP), or remain in long-term potentiation (LTP) or long-term depression (LTD) states.
To meet the specifications of thi analog-type behavior in biological synapses, Sokolov et al. researched electronic synapses based on analog-type resistive switching behavior, occurring in a Schottky diode, operating a Pt/cone shape n-ZnO/SiO2–x/Pt-based interface resistive switching type memristor. A schematic of the structurally designed and fabricated synapse device is shown in Fig. 11(a). The pre-defined protrusions in the SiO2 buffer matrix were created via the BOE etching method. The wurtzite polycrystalline grains of the n-ZnO thin film were obtained via controlled ALD deposition. As displayed in Fig. 11(b), STEM images revealed cone-shaped protrusions in the SiO2 oxide matrix, filled with n-ZnO thin film. Subsequent EDS analysis showed the different stoichiometry of theALD-deposited n-ZnO thin film, being non-stoichiometric ZnO1–x oxide inside the cone, and stoichiometric ZnO oxide at the top interface, respectively. Analogous to the biological synapse, multi-level analog-type interface resistive switching was confirmed in the cone n-ZnO based memristor synapse device via extensive I–V characterization, indicating gradual resistive switching from HRS to LRS and vice-versa, as depicted in Fig. 11(c). Preservation of multi-level resistance states, i.e., retention characteristics, was thoroughly assessed, indicating multi-level synapse device capability up to 7 distinguished resistance states, as shown in Fig. 11(d). Susequently, the cone-shaped n-ZnO memristor demonstrated synaptic characteristics such as a transition from STP-to-LTP, which is achieved via identical spiking of the device, varying only the number of spikes applied, as displayed in Fig. 11(e). It is understood that low- or high-frequency spiking from pre-neuron to bio-synapse alters synaptic weight (conductance) less or significantly, respectively. The frequency-related synaptic characteristics of the memristor device are therefore of great significance. As shown in Fig. 11(f), lower or higher conductance can be achieved in the cone-shaped n-ZnO based synapse device via the low- or high- frequency modulation of applied identical spikes. Furthermore, the importance of previously applied spiking history, either HFS or LFS, was demonstrated for this device, as depicted in Fig. 11(g). Similarly, as in the bio-synapse, HFS stimulation causes potentiation of the cone-shaped n-ZnO based synapse device, after which LFS (~ 6 Hz) stimulation causes depression of the synapse; however, after other HFS and LFS spikes, the same LFS (~ 6 Hz) can create potentiation in the synapse device, demonstrating a bio-synapse-like spiking history-dependent synapse weight update. Fig. 11(h) displays learning-forgetting-re-learning synaptic behavior, realized in the cone-shaped n-ZnO based memristor. Firstly, learning occurs, corresponding to an HFS identical spike at ~60 Hz; subsequent LFS (~6 and ~1 Hz) memristor behavior correlates to the forgetting of learned information, prior to learning being strengthened by the reintroduction of HFS (~12 Hz) into the cone-shaped n-ZnO synapse device; this triad of frequency variations is described as learning, forgetting and re-learning synapse behavior, respectively. Other research teams have also investigated structural memory-synapse memristors. For example, Huang et al. investigated volatile and non-volatile memory behavior with threshold resistive switching in a cone-shaped patterned poly-TiOx/a-TiOx matrix with Ag doping, demonstrating dual memory switching/threshold switching behaviors. Its unique push-pull mechanism of switchable oxygen vacancies allowed for forming-free, low operating voltage (< 1 V) and a low self-compliance current of ~50 μA. Ling et al. used an organic cone-shaped polymer poly(N-vinyl carbazole - PVK) as a resistive switching storage oxide memristor, discovering such benefits in the cone-shape as the reduced randomicity of filament formation, shortened dynamic-gap zone (DGZ) contact, and decreased switching voltage with improved RS uniformity. Russo et al. researched structural nanorods of ZnO material, subjected to UV illumination, for multi-level memristor applications. They found that each different typesof interface, such as Ag/ZnO and Au/ZnO, the current amplification by UV Light, as well as the current decay constants, resulted in specific RS multi-level characteristics. Kim et al. studied a Cu cone-shaped cation source memristor with TiO2/TiN oxide storage. Their study showed the advantages of a Cu cone-shaped bottom electrode, in terms of superior switching performance, reliability, and achievement of the appropriate Cu cation concentration, as well as directed electric field focusing.
In neuroscience, the Bienenstock, Cooper, and Munro (BCM) theory describes the synapse, where the synaptic weight update strongly depends on the frequency of spiking occurrence, i.e., action potentials, from the pre-synaptic neuron. Therefore, the update of gradual conduction of the memristor should be higher at higher frequencies of stimuli than at lower frequencies, corresponding to the time response of charged ions such as VO, which shares similar dynamics with Ca2+ ion concentration. Note that higher frequency stimuli signifies smaller intervals between spikes; therefore, when charged VO ions are triggered, their high accumulation may occur, due to the lack of time for their relaxation to the device’s initial off-current state. Recently, many studies of memristors include the frequency-related, or spike-rate dependent plasticity (SRDP), characteristics of a given device. Li et al. reported synaptic plasticity and learning behavior in an Ag/conducting polymer(PEDOT:PSS)/Ta memristor. Its frequency-modulated characteristics, i.e., spike-rate dependent plasticity (SRDP), were assessed in detail by means of varied stimulation frequencies in the device, as shown in Fig. 12(a). For clarity, the same number of stimulating spikes (i.e.,10), were used, but with varied spike time intervals, to assess the polymer memristor for SRDP characteristics, observing that higher frequencies result in greater current alterations in the device, as displayed in Fig. 12(b). Li et al. studied a chalcogenide activity-dependent synaptic memristor with an Ag/AgInSbTe/Ag structure. The SRDP characteristics of the device were realized via post-spiking frequency modulation, where a lower firing rate induced depression (decreased current) and a higher firing rate induced potentiation (increased current) inthe device, as depicted in Fig. 12(c). The non-volatile property of the device, potentiated by 70 kHz and depressed by 30 kHz stimuli, was assessed via retention test, demonstrating up to ~2200 s resistance state stability, as shown in Fig. 12(d). Du et al. investigated a bio-realistic WOx-based second-order memristor with a variety of synaptic functions. Frequency related synaptic behavior, such as paired-pulse facilitation (PPF) was clearly demonstrated in the device, by applying double spikes with different time intervals, as displayed in Fig. 12(e). Furthermore, with 10 identical spikes, but different time intervals, a higher spiking frequency was found to trigger a larger conductance enhancement in the synapse device, as depicted in Fig. 12(f). Kim et al. reported an experimental demonstration of a second-order memristor with a Pd/Ta2O5–x/TaOy/Pd structure and were able to implement synaptic plasticity. The second-order memristor was realized solely via spiking activity, enabling a 2nd state-variable (temporal elevation and decay of conductance) and a 1st state-variable (constant modulation of conductance), respectively, as shown in Fig. 12(g). Within the two state-variable memristor synapse, a spiking frequency-related characteristic was also deployed, as depicted in Fig. 12(h). Yin et al. adapted the crystallite kinetics in an HfOy/HfOx-based memristor to investigate diverse synaptic plasticity behavior. The diverse crystallite phases of the HfOy/HfOx memory storage affected the RS behavior of the memristor, including the processes of extrusion/injection of oxygen vacancies, crystallite coalescence/separation, phase transformation, and crystal alignment, leading to homogeneous resistive switching in the device, as depicted in Fig. 12(i). Furthermore, the SRDP rule, i.e. frequency-related synapse characteristics, was demonstrated in an HfOy/HfOx based synapse device, as shown in Fig. 12(j). Finally, Xiong et al. reported a BCM rule-based second-order memristor with a tunable forgetting rate, based on a top electrode modulation of Pt-Al with Pt % doping, with STO memory oxide storage. The authors found that by engineering the Pt-Al top electrode, the current states of the memristor could be varied, as displayed in Fig. 12(k). This top electrode dependency of Pt % percentage doping in Pt-Al led to different frequency-related characteristics in the synapse device, including the demonstration of forgetting characteristics, as shown in Fig. 12(l).
Volatile memristor synaptic arrays for neuromorphic computing
The development of bio-realistic electronic devices capable of mimicking biological synapses is an essential step towards the development of efficient neuromorphic computing systems. As discussed in previous sections, nonvolatile analog switching characteristics are usually utilized for the replication of synaptic functionalities in memristive devices. However, in recent years, devices with volatile threshold switching characteristics have emerged, exhibiting promising switching behaviors for the successful emulation of synaptic functions[99, 97, 170]. In particular, CBRAM-based volatile diffusive memristors display the diffusive dynamics that are analogous to the dynamics of biological synapses. With switching mechanisms based on metal ion migration/diffusion, the switching dynamics of diffusive memristors closely resemble the dynamics of biological neurons and synapses. In addition, diffusive memristors offer energy-efficient switching characteristics with very low switching voltages, making these devices promising candidates for efficient brain-inspired computing. Most of the research focused on the emulation of the synaptic functions in the volatile diffusive memristors is carried out on single devices. However, the beauty of memristors is that they can readily be built into crossbar arrays which directly map artificial neural networks. Numerous simulations and experimental implementations of large-scale memristor synaptic arrays indicate the potential of these networks for brain-inspired computing[171-174].
In one of recent study, synaptic crossbar arrays were fabricated, with an atomic layer of deposited HfO2 functioning as the main switching layer. The utilization of existing semiconductor industry-compatible conventional materials such as HfO2 is more convenient and practical, owing to their compatibility with existing fabrication facilities, reliable switching properties, low cost, and easy fabrication processes. A top-view SEM image of the fabricated crossbar array, depicting a magnified image of a cross-point memristor cell, having a device size of 20 × 20 μm2, is shown in Fig. 13(a). A cross-sectional TEM image of a single memristor cell in the crossbar array is shown in Fig. 13(b), where the FFT analysis confirms the amorphous nature of the HfO2 switching layer. The Ag/HfO2/Pt synaptic device is perfectly analogous to a biological synapse in terms of their specific physical structures, where the Ag top electrode, HfO2 switching layer and Pt bottom electrode are analogous to the pre-synaptic neuron, synaptic cleft, and post-synaptic neuron, respectively, as depicted in Fig. 13(c). Excellent volatile threshold switching characteristics were realized, with a very low threshold voltage (0.15 V), confirming the energy-efficient switching characteristics of the device, as presented in Fig. 13(d). The threshold switching was realized by limiting the compliance current, which in turn limits the growth of the conductive filament. The weaker conductive filament undergoes self-rupture, thereby realizing volatile threshold switching behavior. Device operation with a higher compliance current yields thicker conductive filaments requiring a proper RESET process in order for rupture to occur, as shown in the inset of Fig. 13(d). During the application of a positive bias on the Ag top electrode, Ag+ ions diffuse into the switching layer, forming a conductive filament between the bottom and top electrodes. This diffusive behavior of Ag+ ions in the switching layer is clearly identical to Ca2+ dynamics in biological synapses. The diffusion and migration of Ag+ ions into the HfO2 layer, and the formation of the Ag conductive filament is confirmed by the TEM image shown in Fig. 13(e). A TEM analysis carried out on a device switched to LRS with a higher compliance current highlighted the conductive filament region, which was confirmed by the FFT patterns analyzed in both filament and non-filament regions. The crystalline nature of the filament region indicates the presence of Ag conductive filaments, which was further confirmed via energy-dispersive X-ray spectroscopy (EDS) analysis. The volatile threshold switching was exploited to mimic various synaptic functions. The resulting delay, SET, and self-RESET characteristics are depicted in Fig. 13(f), demonstrating that the self-relaxation time of the device was ~1 ms. By utilizing the coexistence of volatile and nonvolatile behaviors, the transition from short-term potentiation (STP) to long-term potentiation (LTP) was realized by varying the number of rehearsals during the programming process, as shown in Fig. 13(g). The essential STDP characteristics were emulated by utilizing the nonvolatile bipolar switching behaviors of the device, as depicted in Fig. 13(h). According to the brain memorization model of Atkinson and Shiffrin, information is transferred from short-term memory (STM) to long-term memory (LTM) based on the number of repetitions. Utilizing the STP to LTP transition behavior of the synaptic device, the psychological model of STM and LTM was demonstrated by image memorization into the crossbar array, as presented in Fig. 13(i). Three images were stored at different locations on the crossbar array, using a different number of pulses (reputations) for each figure. The rehearsals (repetitions) dependence of STM to LTM transition was confirmed.
Organic oxide-based synapse memristors
The physical limitations of conventional inorganic Si-based memory storage systems have led to widespread research into inorganic memory storage systems. Organic-based RS materials possess fascinating properties, such as low cost, high scalability, light weight, and high compatibility with roll-to-roll fabrication. Moreover, the optoelectronic properties of organic-based memristors can easily be modulated via a molecular design synthesis strategy. The advent of wearable electronics has also resulted in a high level of demand for flexible/soft capable memory devices, which can be realized using organic-based memristors. Overall, organic RS memories include polymers, 2D graphenes, organic small molecules, bio-based materials, metal-organic frameworks, organic-inorganic hybrids, and polyoxometalate (POM) molecules; in addition such materials may be refined at the nanostructural/nanomolecular level, i.e. quantum dots (QDs). By themselves, QDs can be considered as a new class of materials, possessing a combination of outstanding optical/electronic properties, together with low cost, structural stability, large-effective area, and simple solution-based processing capability. Specifically, QDs offer solution-processed fabrication with superior tuning properties. For example, nanometer scaled molecules can be tailored to specific composition, shape, size, and surface ligands, facilitating the engineering of bandgap, photoluminescence, self-assembly, and quantum confinement effects. Therefore, studying the thin film assembly of QDs for non-volatile/volatile memristor applications, with potential further applications in neuromorphic engineering represents the cutting edge in the field of RS memory.
Carbon-based materials are renowned for their low cost, mechanical flexibility, and eco-friendliness. Furthermore, new classes of materials, such as graphene quantum dots (GQDs) have attracted research interest due to their unique properties and potential applications, i.e., high chemical inertness, enhanced photoluminescence, and superior biocompatibility. Therefore, research into new materials such as nitrogen-doped graphene oxide quantum dots (N-GOQDs) as memristor storage, which are also applicable to bio-inspired electronics, seems a wise strategy. Sokolov et al. developed organic N-GOQDs thin film ionic conductor storage for memristor and synapse device applications. A TEM cross-section image of the fabricated device in Ag/N-GOQDs/Pt structure, with SAED in the inset, showing an amorphous phase of N-GOQDs, is depicted in Fig. 14(a). The N-GOQDs-based memristor, together with its electrical setup, and a schematic of Ag ion migration via functional groups such as –O, –OH, –NH in the N-GOQDs’ ionic storage conductor, is shown in Fig. 14(b). Threshold resistive switching (TS) characteristics with a low threshold voltage of ~0.3 V and a huge resistance window of ~107 were demonstrated by the N-GOQDs based memristor, as displayed in Fig. 14(c). The “firing” behavior of the TS switch was assessed via a single pulse (0.25 V/500 μs) and by long resistance state read-out voltage (~0.02 V), highlighting the delay-relax characteristics of the device. The time needed for Ag ions to migrate into N-GOQDs storage, i.e., the delay time, is followed by the thin Ag filament connecting the top and bottom electrodes, i.e., the ‘fire’ current; finally, after the pulse, current relaxation characteristics are observed, i.e., the Ag filament self-brakes, due to Rayleigh instability properties, as shown in Fig. 14(d). This “firing” process in the N-GOQDs-based memristor was assessed multiple times at ~1.2 × 104 to validate the repeatable behavior of the device, as displayed in Fig. 14(e). The synaptic behavior of the N-GOQDs-based storage memristor was also established. The frequency dependence, or so-called pair-pulse facilitation (PPF) phenomenon was verified in the device via 10 pulses with varied pulse interval timings, between 0.5 and 5 ms, revealing the strong frequency dependency of the device in relation to applied stimuli, as shown in Fig. 14(f). Short- and long- term potentiation (STP and LTP) synaptic characteristics were confirmed in the N-GOQDs based memristor via the application of a different number of pulses, e.g. 10 pulses for STP and 30 pulses for LTP, resulting in weak Ag filament formation and strong Ag filament formation, respectively. Moreover, STP-to-LTP transition was achieved in the device by applying consecutive trains of numbered pulses, and measuring the conductance state after each pulse trains, as depicted in Fig. 14(g). Finally, significant spike-timing-dependent plasticity (STDP) characteristics were exhibited by the N-GOQDs based memristor, as shown in Fig. 14(h). Briefly, STDP is a synaptic learning rule, which reflects the sign and magnitude of synapse conductance update, which is strongly dependent on the arrival timing of pre- and post-synaptic stimuli. When pre-synaptic stimuli come first, synaptic conductance increases, conversely, post-synaptic stimuli coming first results in decreased synaptic conductance. Firstly, higher current compliance is used to achieve bipolar resistive switching characteristics in the N-GOQDs, as shown in the inset of Fig. 14(h). Next, in order to achieve STDP, pulses of pre-spiking and post-spiking, arriving at the top and bottom electrodes of the device simultaneously, are applied, respectively. Therefore, when pre-spiking arrives at the N-GOQDs-based memristor, long-term potentiation (LTP) is induced; conversely, when post-spiking is delivered to the device, long-term depression (LTD) is realized.
Organic materials utilized as memristor memory storage and their derivatives, such as refining into quantum dots (QDs) separated nano molecules represent advances in resistive memory technologies and related bio-inspired electronics. For example, Ling et al. investigated a poly(N-vinyl carbazole) covalently bonded C60 polymer memristor device. PVK-C60 is a functional polymer, including carbazole (electron donors) and fullerene moieties (electron acceptors), likely to be responsible for the resistive switching characteristics of the ITO/ PVK-C60/Al memristor, as shown in Fig. 15(a). The RS behavior of the device displayed a huge resistance ratio of ~105, write/erase voltages of –2.8 V/3 V, and high resistance state read-out durability, with a potential for being further utilized as a synapse device. Kang et al. studied an organic block co-polymer (PVPCz59) based memristor with various morphologies, as displayed in Fig. 15(b). Unipolar RS behavior has been observed in devices based on lamellar structure, derived from block ratios of PVPCz and P2VP, respectively. The switching mechanism is associated with carbazole segments, which create conductive paths and can be switchable, also rendering this organic memristor suitable for synapse applications. Chen et al. demonstrated an organic memristor based on an egg albumen thin film, prepared by heat-denaturation of proteins, and sandwiched in an Al/Albumen/ITO structured device, as shown in Fig. 15(c). The memristor displayed a reliable RS property over 500 DC cycles, with an on/off current ratio of >103, and resistance states maintained over a long period of ~>104 s, together with further potential for application as a synapse device. Celano et al. reported a complete nanocellulose (nano paper) based memory as a bipolar RS memristor, as displayed in Fig. 15(d). This memory device attains single-use disposable characteristics (biodegradable), and is therefore, very safe for humans. The RS behavior of the device also demonstrated multi-level storage capability and scalability, up to a single nanofiber of 15 nm in size. Moreover, this device has a vast potential for application as a synapse device.
When a memristor device exhibits TS resistive switching characteristics, i.e., abrupt ‘firing’ current change behavior, it can also be utilized to produce leaky integrate and fire (LIF) neuron characteristics. In neuroscience, synapses and neurons are wired together; however, their functions are quite different, with synapses being responsible for information processing via synaptic weight tuning, whereas neurons guide the summarized conductance of nearby synapses and process it to neighboring areas of the neural network. Therefore, it is key that on-chip synapse-neuron systems, are capable of functioning as both synapse and neuron devices. As a bio-neuron, artificial neurons needed to exhibit such functions as automatic fire, leaky integration, and fast recovery. Fortunately, memristors based on TS resistive switching possess most of the characteristics of bio-neurons. For example, Wang et al. studied a core-shell InP/ZnS quantum dots (QDs) thin film-based memristor, which exhibited TS switching behavior, and was further utilized as an artificial neuron with leaky-integrate and fire (LIF) dynamics, as shown in Fig. 15(e). Displaying characteristics analogous to biology, the synapses (considered as capacitors) in the device connected to the LIF neuron, i.e., the TS memristor. Artificial neurons accumulate all conductance via synapses (capacitors); if the threshold voltage is met, the neuron fires its spike further to other parts of the neural network. This LIF neuron behavior was demonstrated via a simple capacitor plus TS memristor circuit, and according to the arriving voltage pulses, e.g. ~1.3–1.7 V, the device’s neuronal firing dynamics can also be varied.