Recent Progress in Multiterminal Memristors for Neuromorphic Applications

The essential step for developing neuromorphic systems is to construct more biorealistic elementary devices with rich spatiotemporal dynamics to exhibit highly separable responses in dynamic environmental circumstances. Unlike transistor‐based devices and circuits with zeroth‐order complexity, memristors intrinsically express some simple biomimetic functions. However, with only two‐terminal structure, precise control of operation principles to ensure large dynamic space, improved linearity and symmetry, multimodal operation as well as high‐order complexity is hard to achieve in a traditional memristor owing to its limited degree of freedom. Therefore, multiterminal memristors including both concrete terminals and virtual terminals (light, pressure, gas, ferroelectric polarity, etc.) have been proposed to obtain precise modulation of memristive characteristics. This review focuses on the recent progress in multiterminal memristors and their neuromorphic applications. The operation principle, application of multiterminal memristor on neuromorphic computing in different scenarios, existing challenges and future prospects are discussed.

construct more biorealistic elementary devices with rich spatiotemporal dynamics to exhibit highly separable responses in dynamic environmental circumstances. The next frontier in advancing computing performance is to incorporate the dynamical and adaptive capabilities of natural and biological systems into hardware level. [21][22][23] Historically, complex biomimetic functions with multiple dynamical equations are simulated utilizing transistor-based circuits including central processing unit and graphics processing unit. Unlike transistor-based devices and circuits, memristor intrinsically express some simple biomimetic functions. With the advance of materials science and electronic engineering, the physical implementation of memristors in 2008 by HP lab make this approach much easier. [24] Due to their intrinsic dynamics, high speed, low power consumption, high density, and scalability, memristors are especially appropriate for working in neuromorphic applications. The conductivity value determines the information memristor stores and then serves as synaptic weights to store information and process input signals in neuromorphic computing systems. By collecting accumulated output current of memristor in a cross array, matrix vector multiplication can be naturally accomplished in one step utilizing Ohm's law and Kirchhoff 's law. Therefore, high parallelism can be easily attained, accelerating computation without latency and power consumption between memory and computing cells. However, there are still nonideal device characteristic observed in the two-terminal memristor, which poses challenges for its further application in versatile biomimetic emulation. i) The high nonlinearity/asymmetry, low dynamic range, and high variability of two-terminal memristor decreases the recognition accuracy of the memristorimplemented neural networks. ii) Most two-terminal memristors exhibit first-order behavior with one variable, exploration of memristors with multiple variables by engineering electrophysico-chemical processes of the switching material ensures the emulation of biomimetic functions with higher-order complexity. iii) Multimodal memristor is highly demanded since biorealistic in-sensor computing system highly necessitates artificial neuromorphic devices, which are able to simultaneously sense, memorize, and compute. [25][26][27] Theoretically, the resistive switching (RS) of memristor is typically originated from the formation and rupture of conductive filaments (CFs), valance change, charge trapping or manipulation of the carrier transport barrier under external electrical stimuli. With only two-terminal structure, precise control of above-mentioned principle to ensure large dynamic space, improved linearity and symmetry, multimodal operation as well as high-order complexity is hard to be achieved in a traditional memristor owing to its limited degree of freedom. [25] Therefore, multiterminal memristor including both concrete terminal and virtual terminal (light, pressure, gas, ferroelectric polarity, etc.) has been proposed to obtain precise modulation of memristive characteristics. Dual and multiple modulatory terminals will also endow memristor with the capacities to minimize cross-talk in crossbar configuration, and implement sophisticated tasks from multiple inputs in learning-capable neuromorphic systems. [6] In addition, introducing multiple tunable terminals of memristors could be a promising technique toward scalable and efficient in-sensor computing systems. For instance, photonic memristors are capable of optically storing and processing data with unprecedented bandwidth, high speed, and low energy consumption provided by the optical domain. [28] In this review, we focus on the recent progress in multiterminal memristors and their neuromorphic applications (Figure 1). First, we review the operation principles of twoterminal and multiterminal memristors (Section 2). Based on the in-depth understanding of physical principles, the recent advances in multiterminal memristor including virtual terminal-and concrete terminal-devices are detailed discussed (Sections 3 and 4). Furthermore, we attempt to provide valuable the application of multiterminal memristor on neuromorphic computing in different scenarios (Section 5). In the end, we give a brief summary of the existing challenges hindering the development of multiterminal memristors and envisage the future prospects of these novel electronic devices.

Memristor Fundamentals
The nonlinear electronic element, memristor, was first proposed by Leon Chua in 1971 and then HP Labs successfully physically implemented a memristor prototype device. [24,29] A typical two-terminals memristor possesses a metal-insulator-metal structure. The physical structure of the switching is reconfigured by the applied electrical field, which highly depends on the history of stimuli, ensuing the controllable multiresistance states of memristor (Figure 2a). [30,31] For example, the physical reconstruction process of filamentary memristor is mainly induced by internal ion migration and redistribution, and the change of conductivity is related to the formation and fracture of CFs in the switching layer. According to the response time, physical process can be classified as emergent or progressive type, corresponding to digital and analog memristors, respectively. [26] Digital memristor devices only have two distinct discrete resistance states where low/high resistance state (LRS/HRS) correspond to logic 1/0 ( Figure 2b). The on operation is defined as a transition from HRS to LRS, while the off operation is defined as a relaxation from LRS to initial HRS. While the analog memristor shows a continuously controllable resistance state under voltage scans, which can be applied as multibit memory (Figure 2c). [32] Physical mechanism of memristor can be mainly divided into three categories: [33] filamentary type: [34][35][36] the migration of cation/anion in the switching layer is driven by electrical field so that formation and rupture of CFs is controlled the between two terminals. [37] The resistance state depends on the size and stability of CFs. Nonfilamentary type: [38,39] resistance state is determined by the interface Schottky/tunneling barrier arising from carrier trapping/detrapping or ion migration modulation in the switching layer. The carrier trapping/detrapping-based memristors will alter the distribution of trap energy level in the energy band during RS operation, resulting in various resistance states in the RS layer and the space charge limited current effect. [40,41] Phase-change type: [42,43] different from local ion migration-based memristor, thermal effect generated by www.advelectronicmat.de external electrical stimulation induces phase-change between the amorphous phase (HRS) and the crystalline phase (LRS). Memristive characteristics can also be originated from other effects such as ferroelectric effect, magnetoelectric effect, which still demand abundant research to improve the performance.
Due to its strengths of fast operation speed, simple structure, low power consumption, and high integration, memristor has shown great potential in the field of next generation storage and computing technology. [44] Date processing unit and memory unit are separated in traditional von Neumann architecture, thus the repeated data shuffling between processor and memory limits processing capacity and poses latency issue as well as power consumption inefficiency. For purpose of building a high parallelism and high performance computing system, the concept of in-memory computing has been proposed. [45] In addition to data storage capacity, memristor can essentially execute Ohm's law and Kirchhoff 's law to accelerate the vector-matrix multiplication (VMM). Therefore, memristor is a kind of technology that blurs the boundary between memory and computation. Digital and analog memristors have distinct application fields. Digital memristors are frequently used in digital logic gates where the states of Boolean logic 1 and 0 are mapped to the LRS and HRS of memristors. The digital memristor based logic circuit shows the characteristic Figure 1. Summary of device structures and neuromorphic applications of concrete and virtual multiterminal memristors. According to the working mechanism of the device, the concrete multiterminal memristor can be divided into filamentary type, nonfilamentary type, phase-change type, and ferroelectric type. Virtual multiterminal memristors are classified into photonic memristor, pressure-tunable memristor humidity-tunable memristor, and multimodal memristor based on the various virtual terminal response stimuli. Multiterminal memristors introducing complexity of computing further facilitate neuromorphic applications of artificial synapses, brain-inspired computing, biomimetic emulation, and in-sensor computing. www.advelectronicmat.de of logic gates and latchs. As for analog memristor, its crossbar arrays are utilized to realize multiplicative computation in a single clock cycle, which strongly accelerates the computing speed. When voltage stimulus input to the row of memristor arrays, analog memristor units act as weights for the matrix dot product operation, the current of each memristive unit is the product of input voltage and device conductance (multiplication operation) obeying to Ohm's law, the column output current of memristor array is the accumulated current of corresponding node (accumulation operation) obeying to Kirchhoff 's law. High parallelism and high throughout make analog memristors ideal for accelerating computation-intensive applications such as hardware artificial neural networks (ANNs). [46]

Fundamentals of Concrete Multiterminal Memristor
Due to concrete multiterminal memristor introduces extra electrical stimulation terminal to the device on the basis of traditional two-terminal memristor, its operational mechanism of the RS is still reliant on the electrically induced dynamic variations mentioned above: filamentary type, nonfilamentary type, phase-change type, and ferroelectric type.
The concrete multiterminal memristor encourages the further evolution of neuromorphic system. [47,48] At the device level, by modulating the electrical stimulation of extra terminal, it breaks first-order behavior of most two-terminal memristor and renders device biorealistic such as switching synaptic learning schemes. [49,50] At the large-scale array level, traditional two-terminal memristors typically incorporate an additional electrical element like a transistor or a selector to serve as a switch to select individual nodes and to minimize sneak current between adjacent nodes. This design undoubtedly increases the footprint and complexity. By contrast, gate-tunable multimemristor is integrates the RS and selection functions within a single device structure. [49,51] For example, Feng et al. demonstrated a multiterminal selfselective memtransistor crossbar array. [52] Depending on the I-V characteristics of the memtransistor, a set voltage is applied to the bit line (drain terminal) of the selected node, the word line (gate terminal) is either grounded or positively biased, and the source line (source terminal) can thus collect the current of selected nodes. Conversely, the word line of unselected nodes is floated or negatively biased, resulting in almost no current flowing through unselected nodes, thereby achieving the goal of minimizing crosstalk and sneak current. Furthermore, the integration density of 3D architectures with three-terminal memtransistor is significantly higher than that of conventional cross array like one-transistor one-resistor (1T1R). [53,54] Despite the fact that multiterminal memtransistor exhibits significant benefits in terms of device characteristics (nonlinearity/asymmetry, dynamic range, variability, etc.), it still confronts key obstacles at the circuit and software levels. Most memtransistors have been restricted to a few large-scale demonstrations while 1T1R can be easily integrated into the logic process for the applications of system-level chips.

Fundamentals of Virtual Multiterminal Memristor
Virtual multiterminal memristor refers to its resistance state can be modulated by light, pressure, gas, ion, and other stimulus sources. [55] At present, according to the type of stimuli, virtual multiterminal memristors can be divided into four categories: photonic memristor, pressure-tunable memristor, humidity-tunable memristor, and multimodal memristor.
When one considers that the additional virtual terminal regulation is no longer purely dependent on electrically induced dynamic variations, the fundamental process underlying RS grows increasingly complex. [28] Synergistic coupling of electrical, photonic, and ionic reactions produces the photonic memristor's resistance to switch. [3,56] According to the different light effect, the physical mechanisms of photonic memristors are classified into the interfacial barrier modulation, charge trapping/detrapping, filament formation and annihilation, chemical reaction/conformation change. Humidity-tunable memristor typically selects a water-molecularly sensitive material as its active layer. The presence of water molecules will alter the material's physical and chemical properties. The RS processes of pressure-tunable memristor are primarily governed by piezoresistive effect, piezoelectric effect, and triboelectric effect.
Additional terminal regulation brings the well-modulated operating current (I HRS and I LRS ), SET/RESET voltages, and additional variables for emulating biological functions with high-order complexity (e.g., temperature, build-in electrical field, interfacial defects, polarization). For instance, photonic memristor could be a promising technique toward scalable and efficient in-sensor computing systems as a result of the high bandwidth and ultrafast transmission characteristics of photons. Moreover, virtual multiterminal memristors are capable of more sophisticated memory and learning capabilities than regular sensors. Inspired by the human sensory system (vision, touch, etc.), the creation of virtual multiterminal memristor enables neuromorphic system to possess high-level cognitive functionalities for accurate evaluation and comprehensive understanding of the real world. [57] Accordingly, the emergency of multiterminal memristors presents new potential for different application scenarios.

Multiterminal Filamentary Memristor
Applying external electrical stimulation to the switch layer of multiterminal filamentary memristors allows for dynamic reconfiguration of its internal state. The electric field can drive the migration of cation/anion in the switching layer followed by the formation and rupture of CFs connecting the electrodes. CFs consist of interstitial metal atoms (cation type), oxygen vacancies (anion type) or both (dual ionic type). [33] Based on solid-state physically evolving networks (PENs), Yang et al. designed Ag nanoclusters that can self-organize in an electric field, evolve as a network over time, and produce filaments that bridge separate terminals. A multiterminal system's fundamental operation is depicted in Figure 3a. Under the influence of external stimuli, evenly dispersed Ag clusters www.advelectronicmat.de voluntarily create fiber bundles linking various terminals, which can be used to regulate information flow and serve as the basis for a neuromorphic computer system. [58] In contrast to RS, which is driven directly from several terminals, recent work on modulating memristor behavior by adding a third terminal has also produced promising results. By introducing a gate electrode to the Pt/Cu 2-x S/Pt parallel structured memristor, Mou et al. created a three-terminal memristor resembling the transistor structure (as shown in Figure 3b). Due to the residual charges of metal nanowires (NWs), the gate field effect can influence the SET/RESET voltages by affecting the growth route of the filaments through repulsion and attraction (Figure 3c). It is important to note that the conductance in the "on" and "off" states rises with a positive gate voltage because Cu 2-x S appears to be n-type, and the positive gate voltage induces the concentration of free carriers on the surface of positively charged NWs. However, the device's overlap between the drain-gate and source-gate areas will result in parasitic capacitance, which will weaken the device's characteristics while strengthening the gate control effect. [59] Yang et al. reported a three-terminal memristor with mechanically exfoliated 2D III-VI semiconductor GaSe as the active switching layer. Modulating carrier density through the third terminal lowers the switching voltage of the device (Figure 3d). Due to the low migration energy of the intrinsic Ga vacancy in the p-type GaSe layer, ultralow set and reset electric fields of ≈3.3 × 10 2 V cm −1 can be obtained. [60] In addition, Haque and Mativenga utilized the ion migration characteristics of halide perovskite CH 3 NH 3 PbI 3 (MAPbI 3 ) to prepare a transistor-structured memristor with gate electrode that can continuously adjust the LRS across several orders of magnitude; [61] Farronato et al. reported a three-terminal memristor based on ultrashort channel molybdenum disulfide(MoS 2 ) multilayers where the gate bias controls the HRS independently. [62] The aforementioned multiterminal memristor makes use of the transistor structure to achieve gate-tunable memristive behavior. However, putting the source and drain electrodes laterally will result in a larger device area, which is disadvantageous for large-scale integration.  [58] Copyright 2015, Wiley-VCH. b) Schematic illustration of the three terminal Pt/Cu 2-X S/Pt memristor structure. c) The typical I-V curves of the three terminal Pt/Cu 2-X S/Pt memristor. The SET/RESET voltages increased when positive gate voltage was applied. Reproduced with permission. [59] Copyright 2016, Elsevier. d) Gate tunable RS behaviors of multiterminal memristor based Ag/GaSe/Ag. Reproduced with permission. [60] Copyright 2018, Elsevier. e) Schematic of three-terminal memristor with vertical structure. Reproduced with permission. [63] Copyright 2018, IEEE. f) Diagrams showing the function of an inserted thin Ti layer. Ti's oxidation into TiO X provides a potential barrier between the TaO X layer and the gate, effectively reducing leakage current. Reproduced with permission. [64] Copyright 2018, Elsevier. g) Band diagrams of multiterminal ReSe 2 /Graphene memristor. Reproduced with permission. [65] Copyright 2020, Elsevier. h) A schematic of the tin oxide/MoS 2 heteromemristor. Reproduced with permission. [66] Copyright 2021, Springer Nature.

www.advelectronicmat.de
The vertical structure of the multiterminal memristor offers greater scalability compared with the planar structure, which is more appropriate to be integrated into neuromorphic circuits. As shown in Figure 3e, Herrmann et al. reported a three-terminal memristor with a vertical structure based on SrTiO 3 (STO) that was capable of adjusting the oxygen vacancy content in STO by gate bias with a continuously controlled conductance state. The challenges of two-terminal memristors with limited resistance states and low margin were overcome by separating the read and write terminals for producing highly precise resistance states. [63] As reported by Wang et al., a TaO X -based three-terminal memristor with a similar structure was further optimized through inserting a thin Ti layer between the gate and the switching layer. The thin Ti can be oxidized to a TiO X layer within the first few SET processes, which promotes the drift of oxygen vacancies under small V g and creates a barrier directly between the switching layer and the gate, thereby significantly reducing the gate leakage current ( Figure 3f). [64] Another approach is to optimize the gate electrode of the three-terminal 2D materials-based memristor in order to form a heterojunction with the switching layer, and then adjust the memristive behavior by regulating the height of the heterojunction barrier. Rehman et al. reported a low-power gate tunable memristor with a copper/rhenium diselenide (ReSe 2 )/graphene structure. The barrier height of ReSe 2 /graphene heterojunction controls electron flow, which in turn controls the neutralization of Cu ions. It could be modulated by the gate voltage, which is capable to modulate the Fermi levels of Graphene, barrier height of the ReSe 2 /graphene, and subsequent SET voltage and ON/OFF ratio (Figure 3g). [65] A similar mechanism has also been incorporated in tin oxide/MoS 2 heteromemristor ( Figure 3h). [66]

Multiterminal Nonfilamentary Memristor
The further scaling down of devices was significantly hindered by the stochastic nature of filament formation and rupture in filamentary memristors. Unlike filamentary memristors, Schottky/tunneling barriers of interfacial memristors can be modulated via interface reactions to induce RS without the filament formation/rupture process, gaining great attention in recent years due to their excellent stability. [33] Various types of multiterminal interfacial memristors have also been developed recently, among which 2D materials are promising switching materials owing to their unique optoelectronic properties, great mechanical flexibility, and outstanding thermal stability.
In 2015, Sangwan et al. first realized the memristive behavior in parallel structures based on polycrystalline sulfur-deficient monolayer MoS 2 , a 2D material grown by chemical vapor deposition (Figure 4a) where different switching behaviors can be observed with different grain boundary (GB) topologies. Planar structure conquered the bottleneck that the gate tunability is limited arising from the strong screening of traditional vertically structured memristor and thus greatly improved the gate tunability. [67] Subsequently, by employing multiple-GBs within one channel, the dependence of devices on single-grain boundary topology is reduced, which greatly reduces the device-to-device variation. The grain-boundary-mediated defects migration induces the modulation of Schottky injection and subsequent RS (Figure 4b). The I D -V D switching hysteresis evaluated at multiple gate voltages confirms gate tunability with four orders of magnitude. Furthermore, the devices exhibit large switching ratios, high cycling endurance, and long-term retention time. It is worth mentioning that switching the direction of the resistance state can be varied by a high gate voltage, as shown in Figure 4c. [68] Although the devices exhibit excellent gate tunability, the operating voltage is relatively large (the drain and gate voltages often require tens of volts), which may hamper the application of the devices in neuromorphic systems. In addition, the difficulty in controlling the directional distribution of GBs and the defect concentration also limits the mass production of devices. In recent years, different scientific research teams have proposed various optimization methods to tackle these problems. Wang et al. used a high-K gate dielectric (HfO 2 ) ( Figure 4d) to greatly reduce the operation voltage of the devices with retaining wide gate-tunability and excellent analog RS behavior. As shown in Figure 4e, by applying a small drain-source voltage (V DS = 0.1 V) to read the LRS and HRS resistance values of the devices, a switching ratio higher than 10 4 is obtained. [69] On the other hand, Feng et al. proposed that optimized annealing terrace-assisted growth prompts epitaxial growth of MoS 2 with oriented grains, reducing the geometries of individual grains while achieving very low switching voltage (V SET = 0.42 V) and switching energy (20 fJ bit −1 ). Figure 4f shows the eight levels of data storage capability of device after gate and drain modulation. [52] Despite the capacity to modulate resistance, the multiterminal interfacial memristors' orientation and distribution of GBs remain random and uncontrollable. This makes it difficult to scale the production of devices. In 2019, Jadwiszczak et al. lined irradiated MoS 2 with a focused probe of a helium ion microscope where the irradiation area would generate a nanoscale defect-rich region, as shown in Figure 4g. This optimization scheme essentially modifies the local structure of the material, thereby regulating the defect distribution. [70] Wang et al. fabricated the devices based on MoS 2 /MoS 2-x O δ GBs via lithography-free direct-laser-writing control, which artificially created designed boundaries (Figure 4h). The devices also exhibit a low power consumption with 20 pJ. [71] Recently, Nguyen et al. reported a reliable method for producing MoS 2 thin films with a large surface area and high quality by spraying and then sulfuring MoS 2 colloidal ink, as shown in Figure 4i, providing an alternative approach for large-scale production of devices. [72] Besides, as shown in Figure 4j, Lee et al. recently designed a double-gated memristor by introducing a second gate terminal where the double gate could refine the regulation of the memristors to provide a novel path to optimize the traditional memristor circuits, especially for the integration of large-scale crossover circuits. [49] Our group fabricated threeterminal memristors using mechanically exfoliated singlecrystal WSe 2 . Due to mechanical exfoliation, charge trapping/ detrapping is the main resistance-switching mechanism of single-crystal materials involves between adjacent dielectrics www.advelectronicmat.de Figure 4. a) Schematic of gate-tunable MoS 2 memristor, a vertex inside the channel formed by the intersection of two GBs connected to one of the electrodes. Reproduced with permission. [67] Copyright 2015, Springer Nature. b) Schematic showing the energy band of a multiterminal memristor at the four RS stages. c) I-V characteristics of multiterminal MoS 2 memristor at different gate voltages. Reproduced with permission. [68] Copyright 2018, Springer Nature. d) Schematic of the MoS 2 memristor structure, 20 nm HfO 2 was used as the high-K gate dielectric. e) Diagram showing HRS and LRS resistances at various gate voltage biases. Reproduced with permission. [69] Copyright 2019, Wiley-VCH. f) Schematic showing the eight levels of data storage after different gate and drain voltages modulation. Reproduced with permission. [52] Copyright 2021, American Chemical Society. g) The multiterminal MoS 2 memristor fabricated via irradiation. Reproduced with permission. [70] Copyright 2019, American Chemical Society. h) An illustration of MoS 2 /MoS 2-x O δ GBs fabricated by lithography-free direct-laser-writing technology. Reproduced with permission. [71] Copyright 2021, Wiley-VCH. i) Schematic showing the process of large surface area MoS 2 thin film fabrication and post sulfurization. Reproduced with permission. [72] Copyright 2021, Wiley-VCH. j) Schematic of the dual-gated MoS 2 memristor structure. Reproduced with permission. [49] Copyright 2020, Wiley-VCH. k) Schematic illustration of the switching mechanism in terms of charge trapping and detrapping at the WSe 2 /SiO 2 interface. Reproduced with permission. [73] Copyright 2021, Wiley-VCH.

www.advelectronicmat.de
instead of the Schottky barrier mediated by GBs reported in polycrystalline MoS 2 memristor (Figure 4k), and this trapping is also gate-regulated, showing gate-tunable memristive behavior. [73]

Multiterminal Phase-Change Memristor
The RS of multiterminal phase-change memristor depends on the structural phases change of material. In 2019, Zhu et al. utilized reversible storage and release of lithium ions to control the transition of structural phase of multilayered MoS 2 to achieve tunable memristive behavior. The electric-field-driven migration of Li + ion along the in-plane (IP) direction results in local changes in Li + ion concentration, facilitating the transition to metallic 1T′ phase (semiconducting 2H phase), as shown in Figure 5a. The high IP mobility of Li + ions ensures ion-exchange-based coupling of coplanar devices. The authors also fabricated a multiterminal memristor in which two or more memristors share one common electrode in order to enable synaptic cooperation and synaptic competition in neuromorphic systems (Figure 5b). The shared grounded electrode G, resembling a postsynaptic terminal connected to a dendritic, is there. The limited supply of Li + ions in Li X MoS 2 thin films is comparable to the competition for limited plasticity-related proteins in human brain networks, allowing devices to emulate synaptic competitive behavior. In Li x MoS 2 devices, where the potentiation of one synaptic device can be facilitated by the potentiation of neighboring synaptic devices, synaptic cooperative effects have also been verified. [74] Iqbal et al. fabricated a three-terminal phase change memristor by employing the nanoscale Mott metal-to-insulator (MIT) transition properties of multiphase VO X . The SET voltage of memristor based on MIT transition is smaller than Figure 5. a) Schematic illustration of modulated Li + ion migration causing local 2H-1T′ phase transitions. b) An illustration of the multiterminal memristor, four terminals share one gate electrode. Reproduced with permission. [74] Copyright 2018, Springer Nature. c) I-V characteristics of VO X memristor without and with gate voltages. Reproduced with permission. [75] Copyright 2021, Wiley-VCH. d) Illustration of the multiterminal phasechange memristor nonvolatile conductance (ΔW) and volatile conductance (ΔF). Reproduced with permission. [76] Copyright 2022, Springer Nature. e) I-V characteristics of multiterminal ferroelectric memristor at different top-gate biases. Reproduced with permission. [77] Copyright 2020, IOP Publishing. f) Schematic illustrating the energy band of a memristor to depict the ferroelectric working mechanism when the drain bias is positive. g) Typical I-V curves showing positive gate bias significantly inhibits resistance switching, while negative gate bias restores it. Reproduced with permission. [78] Copyright 2019, Wiley-VCH. h) The schematic and AFM topography image of a multiterminal α-In 2 Se 3 ferroelectric memristor. Reproduced with permission. [80] Copyright 2021, Wiley-VCH.

www.advelectronicmat.de
conventional memristors, as shown in Figure 5c. The mechanism of the three-terminal memristor is complicated owing to the synergistic of the Mott MIT transition and the capacitive effect caused by the interface reaction between VO X and SiO 2 to create VSiO X . The device also exhibits ultralow power consumption (199.5 fJ) and fast switching time (35 ns). [75] It is worth mentioning that Sarwat et al. make full use of the phase change characteristic and inherent semiconductor of the phase change materials to demonstrate the four-terminal memristor based on Ge 15 Sb 85 . Through the planar electrodes of the devices, structural changes can be obtained by controlling the atomic arrangement of the material, and the difference in conductance between different structures enables the device to achieve a nonvolatile RS effect, providing an opportunity to simulate long-term potentiation (LTP). While intrinsic semiconductivity allows the volatility of device without altering the material structure, which makes it possible to emulate shortterm potentiation. As shown in Figure 5d, the device exhibits better promise for more sophisticated artificial neurosynaptic computations by employing two mechanisms to emulate mixed-plasticity synapses. [76]

Multiterminal Ferroelectric Memristor
By sandwiching a ferroelectric layer between two electrodes, a typical ferroelectric memristor conducts continuous resistance switching depending on the poling bias history of the ferroelectric material. These ferroelectric memristors are under vigorous studies because of their exceptional ON/OFF switching ratios, simple device designs, and extended retention times. Dragoman et al. reported a three-terminal memristor based on graphene-ferroelectric HfO 2 /Ge-HfO 2 /HfO 2 by combining the ferroelectricity and floating gate concepts. Through altering the polarization of ferroelectric structure and introducing a bandgap into graphene, the top gate regulates the ON/OFF states of the device. The flow of carriers in graphene can be controlled by both the top and back gates. The charge storage of the ferroelectric intermediate layer-based Ge-HfO 2 is primarily responsible for the memristive behavior. The devices possess the low power operation with drain voltages (−2 to +2 V) and gate voltages less than 9 V compared with the most 2D MoS 2 multiterminal memristor (Figure 4c), as shown in Figure 5e. [77] Most commonly reported ferroelectric memristors are based on out-of-plane (OOP) polarization of bulk ferroelectric materials to control the tunneling effect for RS. The limitations of bulk ferroelectric materials cannot meet the requirements of three or more terminals to achieve complex calculations. In 2019, Xue et al. reported a three-terminal memristor with an adjustable gate based on a 2D van der Waals ferroelectric semiconductor material, α-In 2 Se 3 . The memristive behavior of the device is originated from the asymmetric modulation of the two Schottky barriers by the drain voltage, which affects the ferroelectric IP polarization, as shown in Figure 5f. In addition, gate bias may flip ferroelectric domains and change channel conductance by integrating special inherent IP and OOP polarizations, as shown in Figure 5g. [78] The unusual structural asymmetry between the IP and OOP polarizations of α-In 2 Se 3 permits polarization switching to occur under the influence of electric fields in either direction, allowing for freedom of device design. [79] On this premise, the team conducted more research in 2021 and built a six-terminal memristor, as shown in Figure 5h. In order to manage channel current, the degree of polarization of distinct channels is not only determined by their corresponding electrodes but also by manipulating the polarization of materials at the various terminals. High pattern recognition accuracy was achieved by using multiple terminal devices in both supervised and unsupervised learning, which has considerable promise for sophisticated neuromorphic computing. [80]

Virtual Multiterminal Memristor
In contrast to the concrete multiterminal memristor described in the previous section, the virtual multiterminal memristor does not reserve a solid input terminal for direct transmission of external stimulus. In general, the state of the active layer of virtual multiterminal memristor is closely correlated with its external environment and history. Here, we focus on the microworking mechanism of four types of virtual multiterminal memristor: light, pressure, humidity, and multimodal.

Interfacial Barrier Modulation
The memristive characteristics can be determined by the interfacial barrier formed at the interface with distinct Fermi levels. The modulation of interfacial barrier can be classified as either the heterojunction/homojunction interface or the Schottky barrier between the semiconductor and metal. The generation of free carriers derived from the photovoltaic effect and separation of photogenerated electron-hole pairs, boosting ion migration or carrier trapping, and tunable interfacial barrier essential for RS behavior.
Hu's group reported an optoelectronic homojunction memristor (Figure 6a), which demonstrates all-optically SET/ RESET characteristics. [81] The homojunction system is consist of oxygen-deficient InGaZnO (O D -IGZO) and oxygen-rich  [82] The accumulation of oxygen vacancies at the CeO x /ZnO boundary changes the barrier height of the CeO x /ZnO junction under light irradiance. Furthermore, Wang et al. developed a CsFAMAPbIBr photoelectric memristor with increased HRS and LRS current under light illumination. [83] Defects near the perovskite/TiO 2 contact neutralized by photogenerated carriers in the perovskite, therefore the barrier height is going to vary. Das et al. studied an Ag/copper(II) phthalocyanine embedded in poly-methyl methacrylate www.advelectronicmat.de (CuPc@PMMA)/indium-tin-oxide (ITO) device with optical response characteristic where excess charge carriers derived from visible light excitation fill all the reap sites, causing device possesses a low external voltage to perform RS operation. [84] Besides the heterojunction/homojunction system, photoinduced RS behaviors also exhibit at the metal/semiconductor interfaces. Li et al. exhibited the Ag/Cs 3 Bi 2 Br 9 /ITO device with electrical SET/optical RESET characteristics. [85]  Reproduced with permission. [90] Copyright 2021, American Chemical Society. d) Polarization switching by combining effect of light and E IMP electric field. Reproduced with permission. [97] Copyright 2021, Springer Nature. e) Schematic illustration of the proposed photonic RS mechanism. f) Typical I-V characteristics of the Ag/STO/CPB/Au device. Reproduced with permission. [99] Copyright 2022, Wiley-VCH. g) State transition from LRS to HRS under a light pulse. Reproduced with permission. [102] Copyright 2021, American Chemical Society. h) Schematic representations of the valence change caused by light. i) Fully light-modulated synaptic modification. Reproduced with permission. [87] Copyright 2021, Wiley-VCH. j) The switch in Mo's valence state coincides with the resistance state transition. The green and blue balls stand in for Mo 5+ and Mo 6+ , respectively. Reproduced with permission. [104] Copyright 2019, Springer Nature. k) FORMING behavior and RS characteristics of Al/GO-TiO 2 /ITO memristor without/with UV photoinduced reduction. Reproduced with permission. [105] Copyright 2018, Wiley-VCH. l) The nanocomposite containing conductive gold NP routes is optically expanded. Reproduced with permission. [106] Copyright 2019, Wiley-VCH.

www.advelectronicmat.de
The laser illumination process activates electrons liberate from the trapping locations and the Schottky barrier returns to the initial state. Moreover, Zhou et al. designed a phototunable CH 3 NH 3 PbI 3-x Cl x memristor device where captured photogenerated holes result barrier lower and lead to the photoassisted low voltage switching. In this process, the Fermi level is pushed toward valence band. [86] Shan et al. utilized Ag ions valence change switching under vis light irradiation, which induces the modulation of the Ag/TiO 2 interfacial barrier. [87] Lv et al. reported to utilize NIR to modulate SET/RESET voltage and working current, which are the result from potential barrier regulated by light absorption at the Ag and active layer interface. [88] Yin et al. reported an electrical and optical dual gatetunable memristor based on mechanical exfoliation of threelayer MoS 2 . The surface state of the MoS 2 /Au interface traps photogenerated carriers, lowering barrier height and enhancing device conductance. [89]

Charge Trapping/Detrapping
The charge carriers generated under light illumination may be trapped in localized defects or trapping layers, in which effect the electrical transfer process, hence modifying the RS behaviors.
Zhu's group designed and fabricated an optoelectronic memristor with framework of ITO/poly(3,4-ethylenedioxythiophene)/poly(styrenesulfonate)/copper(I) thiocyanate (CuSCN)/ poly(9,9-dioctylfluorene-co-N-(4-butylphenyl)diphenylamine) (TFB)/CdSe/ZnS quantum dots (QDs)/ZnO nanoparticles (NPs)/Ag. [90] As shown in Figure 6b, CuSCN layer captures photogenerated holes to generate a built-in bias voltage which facilitating the transport between CuSCN and TFB layer, thereby changing the conductance of device. When holes fill all traps, device current will reach a saturation value (Figure 6c). Kumar et al. exhibited an ultraviolet-tunable memristor. [91] When the device is illuminated with UV light, photogenerated electrons will trapped at the defects states located in the antimony-doped tin oxide layer. Similarly, our group utilized upconversion NPs to trap photocarriers for accelerating the trapping process of SET modulation. [92] Poddar et al. utilized the photogenerated electrons to boost the reduction of Ag + cations for introducing additional Ag NPs during the SET operation. [93] The Lee's group developed a heterostructure optical memristor consisting of MoS 2 /hexagonal boron nitride (h-BN)/ graphene. [94] The photoinduced holes tunnel through the heterostructure system and reach the graphene layer where the restricted electrons could be neutralized. Qiu et al. suggested an InGaZnO (IGZO) multiterminal memristor that employs naturally oxidized Al 2 O 3 and ion gel as gate stacking dielectrics. [95] Under the action of the additional electric field generated by the positive charges trapped in the Al 2 O 3 layer and the external gate voltage, the photogenerated charges are trapped at the IGZO/ion gel interface, allowing the devices to store information in both optical and electrical modes. Nguyen et al. fabricated multiterminal memristors using Al 2 O 3 -encapsulated Te/ReS 2 heterostructure. [96] The device exhibits not only electrically and optically tunable memristive behavior, but also a distinctive rectification effect of the p-n junction due to the combination of the excellent optical and electrical properties of p-type Te and n-type ReS 2 with the charge trapping effect exhibited by Al 2 O 3 encapsulation. The resistance level of ferroelectric tunnel junctions (FTJs) could also be modulated by light irradiation. Long et al. reported an optically controlled ferroelectric memristor with a structure of Pt/BaTiO 3 (BTO)/ STO/La 2/3 Sr 1/3 MnO 3 . The combine of the imprint electric field (E IMP ) and their photovoltaic response can assist to turn the polarization, as shown in Figure 6d. [97] In FTJs, polarization switching varies the electrostatic energy landscape and the value of resistance under application of modest electric fields. The polarization of the ferroelectric material depends on defect migration, thus its polarization state can be suppressed by filling defects with photogenerated carriers that are the product of photo absorption in BTO.
On the other hand, photoillumination provides extra energy to release the trapped charge carriers. Pei et al. proposed a QDsbased photoelectric memristor as the part of artificial visual perception nervous system. [98] The trapped electrons within the interfacial defects can be detrapped after light illumination, leaving positively charged empty traps with additional potential and required built-in bias voltage to pursue the Fermi level equilibrium at the heterojunction interfaces. As a result, the height of the barrier will drop, which could improve the probability of electron tunneling and reduce the resistance of device. Besides, Li et al. utilized laser to excite the trapped electrons and therefore induce device switching back to the HRS state. [85]

Filament Formation and Annihilation
Typically, cation migration is observed in memristors that utilize Ag or Cu as the anode material while anion migration is found in the memristor with switching layer composed of oxygen, iodine, and sulfur ion elements. [3] The light irradiation accelerates the migration of both cations or anions. Guan et al. reported a photonic RS memory with a dual-band response among UV-vis based on Ag/STO/CsPbBr 3 (CPB)/Au architecture. [99] Intensity and wavelength of light enable to alter the RS parameters of memristor because of the modulation of depletion region under photo irradiation. The STO/CPB layer absorbs photons and generates electron-hole pairs under exposure to UV light. The flow of Ag + ions impeded by internal bias, which originated from the charge trapping process under the applied external voltage, as illustrated in Figure 6e. However, the photo irradiation can weaken the internal bias field, which accelerates the formation of Ag filaments. Due to the excellent UV response of the active layers, a substantial number of photoinduced carriers can be created to induce higher current of HRS (Figure 6f). Consequently, more than three distinct OFF states can be distinguished in the device by adjusting wavelength of photo.
The mechanism of MAPbI 3 -based memristors is related to the formation/annihilation of CFs consisting of iodide vacancy (V I ) flaws. In the initial eletroforming process, a high voltage is generally applied to generate vacancy defects which are essential to activate the memristor. However, during this process, the current overshoot causes excessive CFs expansion and lowers the RS performance. Zhao et al. proposed a straightforward www.advelectronicmat.de photo-assisted electroforming approach to address the overriding current issue where light irradiation could reduce the transfer barrier, photoconductivity and subsequent electroforming voltage (electric SET process). [100] Similarly, Ham et al. carried out the investigations on the influence of photo irradiation on synaptic plasticity and underlying mechanism of organolead halide perovskite (OHP)-based memristor. [101] Notably, the onset limitation of the synaptic plasticity is reduced in OHP synaptic device as a result of the rapid movement of an inherently present iodine vacancy during photo irradiation.
By contrast, it has also been demonstrated that light-assisted diffusion of charged ions facilitates the annihilation of CFs (RESET process). It is generally accepted that photoillumination enables to accelerate the recombine behavior between iodide/bromide vacancies and iodide/bromide ions. Liu et al. demonstrated a CsPbBr 2 I-based memristor with optoelectronic operation. [102] The recombination of iodide or bromide vacancies (ions) results the breaking of CFs and the blockage of conduction channels that plays a crucial role in the sudden decrease of conductivity during the photoinduced erase operation ( Figure 6g). The HRS of memristor with nonvolatile characteristic can be accelerated by the destruction of vacancies. Furthermore, Zhu and Lu developed a planar Ag/MAPbI 3 /Ag memristor. [103] The two-terminal electrodes linked by CFs consisting of V I − , which prone to destruct owing to the recombination process under the photostimuli.

Chemical Reaction/Conformation Change
Photo irradiation may generate valence change, which results in RS behavior in the memristors. Shan et al. developed a completely light-modulated memristor based on the chemical reactions of Ag molecules inside an nanocomposite layer. [87] As shown in Figure 6h, heated electrons are excited and subsequently transmitted to the conduction band of the TiO 2 at the interface of the Ag-TiO 2 nanocomposite material under exposure to visible light. Importantly, this process occurs the oxidation of Ag NP (Ag 0 → Ag + + e − ) and lowers the Schottky barrier at the Ag/TiO 2 boundary, thus enhancing electronic conductivity. By contrast, electrons can be excited by UV light to move from the valence band to the conduction band of TiO 2 , resulting in the photoreduction of Ag + (Ag + + e − → Ag 0 ). As shown in Figure 6i, the memristor displays completely lighttunable synaptic potentiation and depression in response to visible and UV light. Furthermore, Zhou et al. demonstrated an optoelectronic memristor with a structure of Pd/MoO x / ITO where the valance of Mo is related to the UV light. [104] As illustrated in Figure 6j, the conduction band of MoO x is activated by the UV light penetrates from the electrode. The photoinduced carriers and protons assist active layer to carry out chemical reactions. The switching of memory state is attributed to the Mo ions valence state (6 + or 5 + ), which shows higher conductivity at lower valence states. Similarly, Hu et al. realized an all-optically controlled analog memristor based on bilayered IGZO structure. [81] For the bilayered IGZO device, light induces the conversion of the V O s at the interfacial barrier area into V O 2+ s, which narrows the barrier and increases memconductance.
Furthermore, photo irradiation may initiate chemical reactions to modify the conformation of a material and its corresponding resistance. It is well known that ratio of sp 2 /sp 3 in graphene oxide (GO) determines the quantity of V O -like defects. Consequently, the working properties of GO memristor can be adjusted by modifying the size and proportion of reduced GO(RGO)-domains. Zhao et al. exhibited the process of photocatalytic reduction facilitated by TiO 2 to improve the RS characteristics of Al/GO-TiO 2 /ITO memory cells. [105] The RGO-domain size is increased by the photocatalytic reaction occurring at the interface of TiO 2 NPs under UV stimuli, thereby lowering the oxygen migration barrier and, resulting the forming process with lowest SET voltage (Figure 6k). Furthermore, Jaafar et al. adopted the percolation thresholds in segregated nanocomposite material for realizing memristive behavior. [106] Using optical expansion of a nanocomposite with conducting gold NP channels, the memristor can be optically switched from the HRS to LRS (Figure 6l).

Pressure-Tunable Memristor
Except for exploiting the optoelectronic properties of materials for constructing a memristor with light programming as the third virtual terminal, more and more research teams have committed to adjust the mechanical energy of the memristors by external pressures in recent years. [107] Through physical mechanisms such as the piezoresistive effect, piezoelectric effect, and triboelectric effect, adding force as the third virtual input terminal variable of the conventional two-terminal memristor affects the physical/electrical characteristics of the devices. [108] Pressure-tunable memristors are capable of more sophisticated memory and learning capabilities than regular pressure sensors, which can directly transform external pressure into processed electrical signals. The inclusion of virtual terminals gives conventional two-terminal memristors an additional degree of flexibility, which has significant implications for neuromorphic applications.
Many groups have integrated various functional devices in recent years to accomplish pressure-manipulation of the conductance state of memristor. Our group simulated an artificial tactile sensing system by combining a gold (Au)-coated micropyramid structured piezoresistive sensor with a nafion-based organic memristor, the schematic diagram of which is shown in Figure 7a. Pressure stimuli are converted into electrical pulses using piezoresistive sensors, which are subsequently sent to the memristor for memorizing and processing. Distinct input pressure intensities can regulate different conductance states of memristor. [109] However, this form of integration is unable to resolve the data redundancy issue produced by a high frequency of data transfer between the two modules for in situ perception. Zhu et al. used the polydimethylsiloxane (PDMS) film embedded with silver nanowires (Ag NWs) as the functional layer of the piezoresistive sensor and SiO 2 as the switching layer of the memristor. The pressure sensor is integrated at the device level by using the Ag top electrode of the memristor as its bottom electrode (Figure 7b). The variation of the piezoresistive sensor resistance influence the voltage across the memristor, which further regulates the conductance www.advelectronicmat.de Figure 7. a) Schematic diagram of an artificial tactile sensing system consisting of a gold (Au)-coated micropyramid structured piezoresistive sensor and a nafion-based organic memristor. Reproduced with permission. [109] Copyright 2019, Wiley-VCH. b) Schematic diagram of pressure-tunable memristor by combining pressure sensing with SiO 2 memristor. c) Typical I-V characteristics of the pressure-tunable memristor without pressure and with pressure. The device can only be programmed and erased when pressure is applied. Reproduced with permission. [110] Copyright 2015, Wiley-VCH. d) Schematic illustrating the multilayer structure of the HPPMS. e) Schematic illustrating the process of the memristor's piezoexcitation conversion under the effect of the pressure. Reproduced with permission. [111] Copyright 2021, Elsevier. f) Schematic illustration of GaN piezotronic memristor structure. g) I-V curves of GaN memristor. The forward I-V curve's current minimum drifted away from zero voltage, indicating the presence of an internal field in the piezotronic memristor. Reproduced with permission. [112] Copyright 2016, Wiley-VCH. h) Schematic illustrating the compressive stress could modulate the V N conductive channels at the knot of the GaN microwire. Reproduced with permission. [114] Copyright 2020, American Chemical Society. i) Typical resistive swiching I-V characteristics of memristor under various bending radius. Reproduced with permission. [115] Copyright 2018, Wiley-VCH. j) Schematic showing the memristor structure of Au/ZnO/Au with virtual terminal of pressure. k) I-V characteristics of the pressure-regulated memristor. Reproduced with permission. [116] Copyright 2020, Elsevier.

www.advelectronicmat.de
state of memristor and enable storage of pressure information, as shown in Figure 7c. The integrated devices combine high sensitivity, fast switching times, high endurance, and nonvolatile memory characteristic. [110] Son et al. exploited the colloidal synthesis of MoS 2 nanosheets for large-area integration of piezoresistive sensors and memristors. [67] Except for the piezoresistive effect, a high-resolution pressure piezomemory system (HPPMS) was reported by Jiang et al. in 2021. As shown in Figure 7d, the multilayer HPPMS device is constructed by integrating piezo-NWs (ZnO) and RS material (MoO 3 ). By applying external pressure on the device, piezoelectric dipoles of ZnO NWs can be adjusted to trigger the valence change of Mo in the MoO 3 layer and the conductance of device for memorizing the nonvolatile pressure information (Figure 7e). It is worth noting that the device has a rectifying function owing to the Schottky barrier between ZnO and Au. The system's pixel size is only 60 nm, which supports the high resolution of the whole system. [111] Liu et al. developed the piezotronic memristor by employing the piezoelectric properties of 1D bambo-like GaN semiconductors (Figure 7f). Under bias voltage, the drift and diffusion of defects and doping successfully modified the defect boundary areas near the bamboo knots energy barrier, resulting in RS behavior. The amount of compressive strain (from 0% to −0.76%) causes strain-induced piezoelectric polarization, which in turn influence the SET voltage of the device (from 1.42 to 2.15 V) (Figure 7g). [112] Hua et al. reported the piezotronic sensory memory device based on the same material whose resistance states can be modulated by the piezotronic effect induced generation of nitrogen vacancies (V N ), which functions as electron trap sites to form/dissolve conductive channels on GaN NWs (Figure 7h). The input of external strain programmed the device to nonvolatile HRS while the application of electrical voltage reset the device to LRS. [113] The piezoelectric artificial synaptic device based on this material has also been reported by Hua et al. [114] Wang et al. designed a flexible volatile memristor based on a nanocomposite film with an elastic composite film (composed of an elastic insulating matrix PDMS and conductive Ag NWs) sandwiched between two planar Au electrodes. Under mechanical bending tests, the devices still demonstrated obvious RS behavior without fluctuation of on-off ratio, but the set voltage of the devices increased as the bending radius decreased (Figure 7i). [115] The horizontal structure of Au/ZnO/Au with virtual terminal of pressure designed by Wang et al. is shown in Figure 7j. ZnO/ polyvinyl alcohol (PVA) hybrid heterojunction was fabricated by spin-coating ZnO with PVA and calcium chloride (CaCl 2 ) polymer ionic glue to provide dual modulation of electricity and pressure. The application of applied electric fields controls the separation and polymerization of charged ions in PVA film, generating a built-in electric field. The charge-coupling effect between PVA/ZnO layers under the synergistic effect of external and built-in internal electric fields manipulates the conductivity of the ZnO channel. When pressure is applied to the surface of PVA film, the thickness difference between the film position above the two electrodes and the middle position decreases progressively, resulting in a change in the dielectric constant of PVA film at various positions. The conductivity of the ZnO channel is influenced by the pressure-regulated electric field, according to the connection between the dielectric constant, voltage, distance, and electric field intensity. Figure 7k shows the I-V curve of the pressure-regulated memristor. In addition, the device successfully simulates the pressure-regulated synaptic plasticity. [116]

Humidity-Tunable Memristor
The physical and chemical characteristics of memristor can be modified through the adsorption and desorption of water molecules therefore moisture can be utilized as virtual terminal to control the memristive characteristics through materials engineering and device design.
The tyrosine-rich peptides, such as the Tyr-Tyr-Ala-Cys-la-Tyr-Tyr (YYACAYY,Y7C), exhibit strong proton conductivity and redox capabilities. Song et al. demonstrated a tyrosine-rich peptide-based memristor that the operational mechanism combines the transfer of protons and electrons. [117] The humiditytunable mechanism is shown in Figure 8a. By creating a tyrosyl radical and losing a proton, the phenolic hydroxyl group of tyrosine helps to decrease Ag ions. Electrons donated by Ag atom induce the synthesis of tyrosine at the adjacent tyrosyl radical. The formation of Ag CFs is attributed to the reaction between tyrosine and Ag ions. The I-V characteristics of the peptide film under varied RH conditions further confirm the proton conduction effects (Figure 8b). RH variation regulates the device working parameters with an on/off ratio over 10 4 (Figure 8c). GB permits rapid penetration of ions and therefore ideal for fast entry of water molecule, which facilitates the humidity-controlled RS of MAPbI 3 . Zhang et al. utilized this material to investigate the humidity effect on memristor such as RS properties. [118] The memristor performs optimally between 5% and 75% RH. At 90% RH, the resistance of OFF state gradually reduces, and after 20 cycles, the memristor could be useless (Figure 8d). Water-induced lowering of the iodide motion barrier is accountable for the increased conductivity of the MAPbI 3 layer with rising RH. The Raman and X-ray diffraction analysis further confirm that the process probably is derived from the intercalation of water molecules into the crystal lattice with the weakened PbI bond (Figure 8e). Similarly, Albano et al. utilized Hong Kong University of Science and Technology (HKUST)-1 structure with capability of strong water absorption ability from the environment and sensitivity of memristor under high RH levels where electronic structure of HKUST-1 can be tuned by water molecules. [119] Fundamental researches on memristors primarily focus on the interaction between oxygen vacancies and charges within the metal-oxide toward realizing memristive characteristics. However, the reaction among protonic defects, oxygen vacancies, and charge carriers are non-negligible. In a N 2 /H 2 18 O tracer gas environment, Heisig et al. conducted isotope labeling studies (Figure 8f) to investigate the behaviors of oxygen and water molecules during state transition of Pt/STO/Nb:STO memristor. [120] Figure 8g evidently shows that the moisture level has a direct impact on the resistance transition. The decreased resistance of the HRS under N 2 /H 2 O atmosphere can be described by solubility of oxygen vacancy rich perovskites in water where the water is responsible for filling oxygen www.advelectronicmat.de vacancies and substitutional hydroxide ions also be generated. Similarly, Zhou et al. proposed the humidity impact on RS of Ag/TiO x /Ti structured memristor. [121] The H 2 O adsorption plays crucial roles in the H 2 O-related redox reactions, thus altering the surface state of the TiO x nanobelt makes it simple to modify its energy band configuration. After adsorption of water, several redox reactions such as the production of the OH − will take place and the conductivity of memristor is consequently enhanced. Moreover, Messerschmitt et al. studied the moisture effect on RS of the STO by varying the surface to bulk ratio and oxide surface exposure. [122]

Multimodal Memristor
Integration of multisensory information is the fundamental function of the human perceptual system. The human brain is capable of synthesizing the information of several senses in order to respond rapidly and precisely in complicated surroundings. Inspired by the human multimodal perception system, the next-generation artificial sensory system should be able to interpret multimodal environmental data in order to achieve more complex cognitive activities. [123] A possibly plausible strategy for realizing this system is to explore multimodal memristor whose memristive characteristic can be simultaneously modulated by either concrete terminal or two or more virtual regulatory terminals such as light, pressure, gas, humidity, etc.
In 2021, our group fabricated a multimodal bipolar nonvolatile memristor with an ITO/MXene-ZnO/Al structure on a flexible PDMS substrate (Figure 9a), which could be concurrently controlled by UV light and humidity stimuli. The SET voltage decreases as the incident light intensity rises where high-intensity UV light exposure can induce transition from an HRS to  [117] Copyright 2020, Springer Nature. d) The memristor's endurance performance at varied RH values. e) Raman spectra of the MAPbI 3 film exposure to 25% and 75% RH, respectively. Reproduced with permission. [118] Copyright 2021, American Chemical Society. f) Experimental setup for isotope labeling experiments. g) The effects of the atmosphere on the properties of RS behavior. Reproduced with permission. [120] Copyright 2018, Wiley-VCH.

www.advelectronicmat.de
an LRS since MXene is capable of trapping photogenerated electrons over an extended period of time, and the photogating effect results in the formation of oxygen vacancy filaments (Figure 9b). Self-assembled ZnO NPs enhance the heterostructure's contact area between the two electrodes, which increases ion adsorption and diffusion behavior. Water molecules can be adsorbed on MXene-ZnO heterojunctions through double hydrogen bonds at high humidity. The hydrolysis of surface functional groups raises proton concentration, and the electrostatic interaction between protons and oxygen vacancies limits the growth of oxygen-vacancy CFs and breaks their stable state. Competition between humidity-induced rupture and photoassisted formation of oxygen vacancy filament modulates the conductance state and V SET /V RESET of device (Figure 9c). [16] The UV-sensitive piezomemristor with RS layer of MoO X , a top electrode of Pd NP/Li-ITO, and a bottom electrode of Au/ZnO which can be jointly modulated by light and pressure has also been recently reported by Jiang et al. Photogenerated carriers can migrate to the MoO X conduction band while the protons (H + ) could be produced simultaneously to change the valence state of Mo ions. The internal electric field is generated by the piezoelectric polarization charges of ZnO NWs along the direction of strain when pressure is applied, which induces the diffusion of Li + ions into the MoO X layer from the top electrode, modulating the resistance of device. The two approach of valence modulation are synergistic under the stimuli of both UV light and pressure, as shown in Figure 9d. [124] As shown in Figure 9e, in 2022, Duan et al. reported a volatile virtual multiterminal memristors with an Au/Ti/VO 2 /Ti/Au structure through which both pressure and temperature information can be encoded. The authors utilize the inherent thermal sensitivity of VO 2 materials and the effect of voltage divide through piezoresistive sensor and memristor as the main operation principle of the multimodal virtual terminal devices. Traditionally separated tactile and temperature signals can be coupled in a VO 2 -based memristor to identify multimodal tactile/temperature information. Figure 9f shows the I-V characteristics of the VO 2 volatile memristor at different temperatures ranging from 284 to 306 K. [125]

Artificial Synapses
Simulating biological synapses with emerging electronic devices is a novel field of research which is broadly acknowledged as the initial stage in constructing hardware for brainlike computers and artificial systems. [126] Thus far, various forms of memristors have been presented to imitate synaptic functions. Wang and co-worker demonstrated an intelligent neuromorphic tactile synapse. [127] The difference in excitatory postsynaptic potential (EPSP) difference between the two adjacent pressure stimulations (ΔEPSP) is positively correlated with the number of pressure stimuli which achieved maximum amplitude at the 10th press and decreased with further increased number of stimuli (Figure 10a). Subsequently, this group presented a piezotronics synapse to realize strain sensing c) The current of multimodal memristor modulated by humidity and UV irradiance. Reproduced with permission. [16] Copyright 2021, Wiley-VCH. d) Schematic illustration of the switching mechanism in terms of change of Mo ions between Mo 6+ and Mo 5+ modulated by pressure and light. Reproduced with permission. [124] Copyright 2020, Elsevier. e) Schematic illustration of VO 2 -based multimodal memristor both pressure and temperature information can be encoded. f) Typical resistive swiching I-V curves of VO 2 memristor under various temperature. Reproduced with permission. [125] Copyright 2022, Wiley-VCH.

www.advelectronicmat.de
and synaptic functions. [114] To boost synaptic weight updating capabilities with pulse trains, the piezotronics effect in wurtzite GaN is introduced (e.g., 330% augmentation at a compressive stress of −0.36%).
To promote development of efficient neuromorphic vision, photonic memristors combing sensing and processing functions have been explored. Shan et al. developed an photonic memristor with light-manipulated synaptic plasticity in response to visible and ultraviolet light. [87] As schematically illustrated in Figure 10b, visible and UV light pulses were utilized as the virtual gating stimuli to trigger and suppress the electrically induced synaptic plasticity, respectively. In addition, Ng et al. showed an associatively responsive photonic synapse in which a boost in input amplitude, frequency, and repetitions lead to a higher learning level (maximum current) and stored memory (weight variations) (Figure 10c). [128] Moreover, Liu et al. realized copper phosphide nanoribbons (Cu 3 P NRs)-based memristors with optically and electrically modulated artificial synaptic functions. [129] Under the same light illumination, the first duration of 1.2 s is required to increase the conductance from 0.04 to 0.19 mS while the second duration of 0.5 s is demanded to augment the conductance from 0.09 to 0.19 mS, revealing that the lightenhanced resistance switching behavior may replicate the cognitive characteristics of learning, forgetting, and backtracking.
The change of humidity can also regulate synaptic behavior. Song et al. demonstrated a proton-regulated control of tyrosinerich peptide memristor (Figure 10d). [117] The drain current responses to a single pulse exhibit a typical EPSC behavior of biological synapses. The time-scaled electrical characteristics under different RH were studied as shown in Figure 10e. In Figure 10. Multiterminal memristors for artificial synapses. a) The press number is related to ΔEPSP. Details of the ΔEPSP between two EPSP signals are shown in the inset. Reproduced with permission. [127] Copyright 2020, American Chemical Society. b) Light-gated electrically driven synaptic plasticity. Reproduced with permission. [87] Copyright 2021, Wiley-VCH. c) By adjusting the intensity, frequency, and repetition number of the laser training pulses, memory state from short-term to long-term memory can be achieved. Reproduced with permission. [128] Copyright 2021, Wiley-VCH. d) Diagrams showing the proton activation mechanism in artificial (bottom) and biological (top) synapses. Red and green represent presynaptic spikes and EPSCs, respectively. e) Artificial synapses' plasticity is proton-dependent. A major learning process only manifests when both voltage and RH inputs are active, and the synaptic response is severely inhibited by lowering RH. Reproduced with permission. [117] Copyright 2020, Springer Nature. www.advelectronicmat.de region I at low RH, applied voltage input shows negligible influence on EPSC. While EPSC began to increase as the increase of RH owing to increased proton conduction under high RH.

Functional Neurons
The multimodal memristor, enabling the identification of various external stimulus signals have applied for neuromorphic device by adding more faithful biological features including dynamic changes and self-adaptability, which is feasible to capture, identify, and extract dynamic information from complicated environments.
The lobula giant movement detector (LGMD) may identify quickly to an impending object and activate escape strategy.
LGMD can rapidly respond to looming object and efficiently triggers escape behavior. Our group utilized light-mediated memristor featured with threshold switching to realize the artificial LGMD visual neuron. [130] The biomimetic compound eye composed of artificial LGMD visual neuron is capable of extensive field-of-view scout and averting collision which in a nonmonotonic type in response to approaching object (Figure 11a).
Analogously, serving as one of the primary roles of the biological retina, the preprocessing of visual scenes is essential for subsequent parallel and hierarchical processing, as it may do feature processing and categorization. Yang et al. utilized memristor based on perovskite to enhance contrast and noise processing capabilities at the hardware level, therefore creating a self-powered artificial retina sensing platform. [131] Furthermore, Figure 11. Multiterminal memristors for functional neurons. a) An illustration of birds recognizing targets from various angles and distances. Reproduced with permission. [130] Copyright 2021, Springer Nature. b) Artificial multimodal neural network's operational diagram. Using the patterned pressures as the letter "N" on this flexible device, cognition-response storage, and effective sustainability are achieved. Illustrations of the capabilities for long-time storage (of 2 months) and electrical erasure. Reproduced with permission. [124] Copyright 2020, Elsevier. c) Normal-state responses are produced by voltage triggers with mild amplitudes (V gs = 1 V), whereas sensitized responses are produced by high-amplitude, strong voltage shocks (V gs ≥ 2 V) that simulate damage. Similar to biological nociceptors, increased amplitude of the noxious stimuli increases the current response (hyperalgesia) and decreases the incubation time/threshold (allodynia). Reproduced with permission. [132] Copyright 2020, Springer Nature. d) Operational flow diagram of the artificial multisensory system. Reproduced with permission. [57] Copyright 2021, Springer Nature.
www.advelectronicmat.de by detecting the spatial distribution state of pressure in realtime and collaborating with the storage function, the memristorbased bionic electronic skin is capable of accomplishing various intelligent work. Jiang et al. demonstrated a novel electronicskin, which is capable of pressure sensing and memorizing (Figure 11b). [124] Only pixels that have been squeezed by the rubber pen exhibit RS from HRS to LRS. Ultralow damping allows the memristor to retain the "P" form on the readout picture after being exposed to ambient temperature for two months. Similarly, Hua et al. utilized the GaN microwire to fabricate piezotronics sensory memristor. It will store the input strain as a state of resistance when the strain stops to be applied. [113] Wu et al. demonstrated tactile sensory memristor, which could generate pressure-activated electrical impulses without external power owing to the utilization of nanogenerator technology. [127] In order to interpret lower-order information and transmit information from the central nervous system to the rest part of the body and enable an adaptive response to the environment, the peripheral nervous system contains a high density of receptors and various types of nerves. John et al. proposed gathering pertinent data from the robot's skin to embed intelligence into the sensor nodes. [132] Figure 11c demonstrates how the escape reflex of the robotic arm was triggered by pain perception and the connection between nociception and tactile perception. Nociceptors preserve a greater threshold level under normal circumstances. By contrast, the threshold of nociceptors is decreased in the injured condition to protect the damaged location. Similarly, our group used a ferroelectric memristor with two operational modes and a repeatable threshold switching system to simulate active nociception and blocked nociception. [133] The adaptive and cognitive capabilities of distributed processing in biological hierarchy enable it to efficiently assess complicated multimodal data. For instance, in the superior colliculus of the midbrain, multisensory neurons immediately integrate spikes from plenty of senses to activate neuronal reactions to multimodal environmental stimuli. Tan et al. constructed an artificial crossmodal system that combines artificial vision, touch, hearing, and simulated smell and taste sensations were also presented. [57] In addition to perceiving, encoding, sending, decoding, processing, storing, and interpreting multimodal information, the hierarchical and cognitive system (Figure 11d) is also capable of multimodal recognition and imagination through crossmodal learning for autonomous robotic applications.

Artificial Neural Network
The synaptic array with analog weight updating capability for hardware-implemented ANN is the most prevalent ways to realize classification, recognition, prediction, cognition, and decision making. In general, ANNs are made up of input neurons, hidden neurons, and output neurons that are all connected by synaptic weights (conductivity values). During the functional simulation, the neurons in the input layer will accept the pixel values of features and assign them to the corresponding input vectors before converting them to output values via hidden neurons (weight matrix). The difference between the output and input values is used to modify the synaptic weights in the network using a weight updating method based on the algorithm (e.g., back-propagation (BP)) to realize object identification (Figure 12a). [134] However, owing to the physically separated logic and memory units of conventional von Neumann design, retrieving millions of synaptic weight values from the main memory induces the huge latency burden of implementing ANNs. [135] Recent studies have shown that the vital step of implementing ANN, VMM, can be greatly accelerated by memristor crossbar arrays, eliminating data shuttling between memory and computing cells.
While the high nonlinearity/asymmetry, low dynamic range, and high variability of two-terminal memristor decreases the recognition accuracy of the memristor-implemented neural networks. Therefore, multiterminal memristor including both concrete terminal and virtual terminal (light, pressure, gas, ferroelectric polarity, etc.) has been proposed to obtain precise modulation of memristive characteristics. Dual and multiple modulatory terminals control synaptic dynamics, providing a more flexible learning scheme such as heterosynaptic plasticity to increase the accuracy of neural networks. [49,[136][137][138][139][140] For instance, our group simulated two essential characteristics of Bienenstock, Cooper, and Munro learning rules in 2021 using a thriplet-spike-time-dependent plasticity (STDP) model with sliding frequency threshold and enhanced depression effect based on third terminal modulation. [141] It is worth mentioning that increasing the number of device conductance states can be obtained by third terminal modulation, offering the ability to perform weight assignment based on k-means clustering during VMM operations, which is more accurate than conventional uniform weight quantization. [135] In 2022, Sarwat et al. designed multiterminal memristor combined the nonvolatile resistive storage characteristics originated from phase-change and the volatile resistive characteristics modulated by the inherent field effects to functions as a synapse for building the short-term spike-timing-dependent plasticity (ST-STDP) framework with sequential learning capability. As shown in Figure 12b, ANN was used to classify running image frames with integrating the transient plasticity unit (F m,n ) to the conventional single-layer neural network in order to temporarily increase the number of synaptic weight corresponding to the input during inference. The long-term and transient plasticity of the synapses provides compensatory effects. As shown in Figure 12c, multiple distinct but associated images of boy-girl identification task were input into the neural networks. However, the network employing solely LTP (red trace) could only recognize the image in the first frame and fail to recognize associate subsequent frames with the boy. The ST-STDP framework (blue trace) successfully recognizes the initial frame and establish the relationships between moving frames in order to accurately identify all of the images. [76]

Recurrent Neural Network
Traditional ANNs that input data passes through input, hidden, and output layers show weak capability to process the sequence www.advelectronicmat.de information. On contrast to ANNs, recurrent neural networks (RNNs) derived from Hopfield neural networks (HNNs) proposed by John Hopfield in 1982, [142] are able to accept input from both other neurons and their own output, allowing it to flawlessly accomplish the task of sequence learning. In reality, they sometimes converge to a local minimum during computation, despite the fact that continuous HNNs should, in theory, converge to a global minimum of the energy function. RNN may jump out of the local minimum if it can temporarily accept the outcome with the rising energy function during iteration. Boltzmann machine (BM) and simulated annealing essentially revolves on the idea of adding probability to the network by decreasing energy function with a high probability rather than a specific probability. In 2022, Sarwat et al. introduced instantaneous weight fluctuations into the synaptic array to simulate annealing through the third terminal of the memristor which solved the conventional HNN parasitic local minimum problem and hastened the system's convergence to the best solution, speeding up the solution of combinatorial optimization problems. [76] It is challenging to properly regulate the probability distribution relating to the randomness of memristors since cutting-edge computational method always uses rich stochastic properties of memristors to imitate biological neural networks. In 2021, Yan et al. used a reconfigurable threeterminal heterostructure memristor to realize exponentialclass sigmoidal distributions that matches the Fermi-Dirac distribution in physical systems. As shown in Figure 12d, the P ss,t < 2 s is the probability of the device being successfully SET within 2 s, and V TE0 is the bias voltage applied to the device when the device will successfully SET at 50% probability. When the gate voltage is fixed, the success probability of SET operation is sigmoidal distributed with V TE . The Fermi level and charge density of MoS 2 are altered when the gate voltage is applied and the transition of the sigmoidal distribution is sharper when the gate voltage is higher. It is possible to further simplify this distribution to an analogous Fermi-Dirac distribution with a gate voltage-dependent "temperature" effect. The multiterminal memristor function as a reconfigurable neuron in the BM, a recurrent artificial neural network that resembles the thermodynamics of a real physical system. By using a BM to simulate annealing, the "cooling" strategy can be tuned to achieve the optimal convergence  [76] Copyright 2022, Springer Nature. d) P ss,t < 2s is sigmoidal distributed with V TE under different gate voltages. e) A diagram showing the typical evolution of different "cooling" strategies regulated by the gate during the BM optimization process. Reproduced with permission. [66] Copyright 2021, Springer Nature. f) Schematic illustration of RNN. g) The multiterminal memristive network maps spatiotemporal inputs to feature space and then analyzed by the read-out layer. h) Direct visualization of the reservoir state represented by the network filaments map after each stimulation time frame. Reproduced with permission. [144] Copyright 2021, Springer Nature.

www.advelectronicmat.de
based on reconfigurable statistical behavior of multiterminal memristor. Figure 12e shows the typical evolution of different "cooling" strategies regulated by the gate during the BM optimization process, demonstrating the impact of different "cooling" strategies on the BM efficiency. [66] Reservoir computing (RC), one of the unconventional computing framework, is derived from the Echo State Network [143] designed by Jaeger and the liquid state machine designed by Maass. As shown in Figure 12f. RC is an algorithm originally developed from the RNN. The learning complexity is drastically reduced compared to other RNN. The input signals in RNN are not a single point but a sequence that can reduce the network size according to the performance of devices. However, only the weights between the reservoir layer and output layer need to be trained in RC system, which greatly reduces the training cost. And another convenience is that RC does not require stable states or attractors in the reservoir. Milano et al. developed a highly interconnected multiterminal memristive network as a physical reservoir in 2021 by suspending drop-casting Ag NWs as switching materials. This network maps spatiotemporal inputs to feature space, which is then analyzed by a conventional memristor array as a read-out layer, as shown in Figure 12g. The memristor with multiterminal configuration ensures nonlinear dynamics, fading memristor memory, and temporal processing of numerous spatial inputs. In Figure 12h, the formation of filaments in the network during the calculation effectively visualized the evolution of the reservoir's internal state. It is clear that the reservoir with self-organized memristive architecture is more consistent with the topological structure and emergent behavior of biological neurons. The system can accurately recognize Modified National Institute of Standards and Technology (MNIST) digits by 90.4% after 60 000 training epochs. [144]

Spiking Neural Network
Different from DNN which uses analog signals to implement algorithm, spiking neural networks (SNNs), the third-generation of neural networks, is capable of simulating the brain communication via discrete spike signals which is highly analogue to human brain. [145] SNNs handle information in event-driven and asynchronous modes with high energy efficiency. STDP learning rules are also more biologically realistic than the BP algorithm typically utilized in deep neural networks to actualize synaptic learning rules. According to the STDP rule, the synaptic weight increases or decreases as the preneuron spikes applied before or after the postneuron spike. [25] However, the multiterminal memristor-based spiking neural network hardware implementation is still in its infancy.
In 2021, Xue et al. developed the spiking neural network for unsupervised learning based on α-In 2 Se 3 multiterminal memristors. After 30 epochs of training, the accuracy for the MNIST handwritten digit identification task ultimately stabilized at 89%, which was marginally lower than the accuracy of convolutional neural networks built by the same device. Even if SNNs show benefits of low power consumption and spatiotemporal information processing for unsupervised learning, it still has the drawback of not ensuring excellent performance. [80] Inspired by biological systems, Yuan et al. designed reconfigurable three-terminal memristor synapse. The switching of LTP and long-term depression (LTD) is controlled by gate pulses rather than the polarity of drain pulses. Gate reconfigurability of the device leads to different learning curves and simplified STDP, facilitating unsupervised learning of simulated SNNs. [51]

In-Sensor Computing
The traditional sensory systems from the sensors, analogue-todigital converter (ADC), a memory unit to computation units can be implemented based on conventional complementary metal-oxide-semiconductor (CMOS) technology where the sensors convert the external stimulus (e.g., light, pressure, chemical analytes, etc.) into a voltage signal, ADCs binarize the data after amplification and denoising, memory store the binary information and computation units perform final recognition and classification. However, high latency and inefficient power consumption originated from the data shuffling between memory and computation units restricts its further application for motion detection, object recognition, navigation, activity recognition, etc. The memristor with additional virtual terminal couples the sensing, memorizing, and computing functions in the one unit, which provides a feasible solution for the realization of the in-sensor computing system. [45] At present, there are relatively few studies on olfactory and auditory in-sensor computing based on multiterminal memristor, this section primarily highlights the research that constructed artificial visual, tactile, and multimode perception systems based on multiterminal memristor units.

Visual In-Sensor Computing System
As one of the most important channels for acquiring biological information, vision can accurately detect and recognize objects in complex environments. Biological visual systems with hierarchical organization including retina, optic nerve, lateral geniculate nucleus, and visual cortex could easily accomplish complex visual tasks. By encoding optical signals as nerve spikes through the retina and processing them through the visual cortex, the biological vision system is capable of high speed, and ultralow energy consumption. Memristor controlled by virtual optical terminals have the capacity to mimic visual imaging processes in biological systems. In 2019, Zhou et al. designed memristor featured with a nonvolatile time-dependent response to the optical pulses with different intensity and pulse width, which enable device to simulate artificial synapses with optically controllable synaptic plasticity. Moreover, the spike current and retention time increase with increased light intensity. The higher the brightness, the greater the accumulation effect of the device, allowing image contrast enhancement. Subsequently, in the task of recognizing the letters "P", "U", and "C", the preprocessed information (enhanced contrast, smoothed noise, highlighted image features) attained from memristor array further input into the network for training (Figure 13a) with the recognition rate of 98.6% after 1000 training epochs (Figure 13b). [104] In 2021, Wang et al. employed white light modulated memristor to build an in-sensor computing artificial vision system. After 1000 training epochs, the face recognition accuracy reached 0.867. [146] In 2021. Shan et al. designed an artificial memristor synapse with fully light-modulated synaptic plasticity with response to UV and visible light of 300-800 nm. Visible light-induced LTP and ultraviolet lightinduced LTD plasticity ensures the device to enhance contrast based on LTP, and minimizes long-term noise based on LTD. The preprocessed image was further sent to the photoelectric memristor neural network for learning and recognition. After 300 training epochs, the recognition rate reach 98%. In this study, the memristor implements both low-level and high-level processing functions of the in-sensor computing system. [87] In recent years, more and more virtual-terminal memristors with volatile and nonlinear photogating characteristics have been utilized to construct the reservoir layer of in-sensor reservoir computing. In 2021, Sun et al. developed a 2D tin sulfide (SnS) memristor-based photoelectric dual-mode reservoir to map complicated timing sequences to high-dimensional space since the device possess a nonlinear response and fading memory to both electrical and optical stimuli (Figure 13c). The accuracy of the SnS memristor-based in-sensor computing for recognizing Korean sentences reaches 91%. [147]

Tactile In-Sensor Computing System
Human is able to establish tactile memories that indicates how we interact with the world around us by sensing external pressure by under-skin tactile receptors, transmitting the Figure 13. a) Schematic illustration of visual in-sensor computing system based on the ORRAM array. b) Comparisons of the letters recognition rate without and with ORRAM-based in-sensor computing system. Reproduced with permission. [104] Copyright 2019, Springer Nature. c) Schematic diagram of SnS-based memristor array driven by multiple electrical and optical inputs. Reproduced with permission. [147] Copyright 2021, AAAS. d) Comparison of force-images before and after preprocessing with the HPPMS. Reproduced with permission. [111] Copyright 2021, Elsevier. e) Schematic illustration of tactile in-sensor computing system based Au/ZnO/Au virtual-gate memristors. Reproduced with permission. [116] Copyright 2020, Elsevier. f) Schematic of multimodal in-sensor computing system based on MXene-ZnO based memristors to sense the information, filter the noise, and specialize the features. Reproduced with permission. [16] Copyright 2021, Wiley-VCH. g) Schematic illustration of conventional in-sensor RC system, two reservoir layers responds distinct signal. h) Schematic illustration of the in-sensor RC system employed a single reservoir to process various types of sensory signal. g) A diagram showing the in-sensor RC system sensing the digit in a complex environment. h) Comparison of the digit-recognition rate under different modes. Reproduced with permission. [148] Copyright 2022, Springer Nature. www.advelectronicmat.de process tactile signals via the nervous system, and processing the information by neural network of brain. In the field of human-machine interface and flexible robots, pressure-sensitive memristor has the potential to serve as the fundamental component of tactile in-sensor computing systems.
Jiang et al. reported an HPPMS with the basic pixel of 60 nm by integrating piezo-NW sensors and memristors to enable external pressure sensing and signal processing capabilities. Figure 13d shows the corresponding input and output images, confirming the preprocessing capability of HPPMS to highlight the main characteristic of letters and reduce background noise. After 2533 training epochs, the accuracy of the neural network reached 99%. [111] By utilizing the pressure modulation properties of Au/ZnO/Au virtual-gate memristors, Wang et al. used a 10 × 10 memristor array to monitor applied pressure and preprocess tactile images, which were then fed into a three-layer neural network for learning and recognizing (Figure 13e). [116]

Multimodal In-Sensor Computing System
In a complex environment, the human perception system could simultaneously store and process multiple forms of perceptual data. Our group reported the multimodal memristor with structure of ITO/MXene-ZnO/Al and a competitive relationship between the formation of light-assisted formation of CFs and humidity-induced rupture of CFs is the main operation principle of device. Humidity significantly affects the processing effect during the image perception preprocessing task (Figure 13f). The recognition accuracy is 75.44% under humidity of 0-20% and it is 82.96% under humidity of 40-60%. [16] In 2022, Liu et al. demonstrated the α-In 2 Se 3 multiterminal memristor with tunable relaxation time under light and back-gate voltage modulation, contributing the emulation of heterosynaptic plasticity. The input signals are collected by various sensors and mapped to the high-dimensional feature space by various reservoirs in the conventional in-sensor RC system. Each reservoir layer responds to a distinct signal type, while the readout layer integrates high-dimensional feature for final recognition (Figure 13g). The study employed a single reservoir to simultaneously process various types of sensory signal to implement multisensory fused RC (Figure 13h). The left panel of image is covered, which can only be sensed by tactile sensor, whereas the right panel of image digit on the right half can be sensed via visual signals (Figure 13i). Compared to single reservoir structure with the different reservoir layer for different modalities, mix-reservoirs with single reservoir layer for different modalities show increased accuracy of 84.9% (Figure 13j). [148]

Summary and Outlook
Owing to great advances in learning algorithms and the improved computing capability of CMOS technology, the artificial intelligence has experienced enormous growth. Compared with traditional transistor-based CMOS circuits based on static elements with zeroth-order complexity, memristors have attracted significant recent attention for implementing hardware-based artificial intelligent due to their rich spatiotemporal dynamics, fast switching time, small footprint, low power consumption, and multibit storage. Therefore, memristors can realize simple biomimetic function such as STDP based on drift and diffusion of ions, vacancies, electrons or defects which is highly analogue to the biological counterparts. However, memristors with simple two-terminal structure typically still suffers nonideal device characteristic including the high nonlinearity/asymmetry, low dynamic range, high variability, and especially one variable, which is incapable to. Simply engineering multi layered structures or modulating material stoichiometry of two-terminal memristor is hard to emulate biomimetic functions with higher order complexity and implementing multimodal operation since the operation principle is closely related to underlying physical mechanism including ion migration, electron transport, thermal transport, spin flipping, phase change, etc.
The RS of multiterminal memristor can be coupled with gate-modulated Schottky/tunneling barriers via charge trapping, interface reactions, vacancy/ion transport, ferroelectric domain switching, photovoltaic/photogating effect, etc. Therefore, the multiterminal memristors are capable to employ either concrete or virtual electrode to improve the tunability of the switching behavior for implementing on-demand learning rules. A summary of device metrics for two-/multiterminal memristor is provided in Table 1. Multiterminal memristors have shown great potential as the building blocks of neuromorphic computing and in-sensor computing to break through the von Neumann bottleneck. However, explicit research on materials, devices, and systems level are required for its further application. First, variability, stochasticity, and low robustness to material/device defects are the main obstacle of commercialization of multiterminal memristor. Variation of setting memristor to the target conductance induces the low distinction of different states and low bit precision of neuromorphic computing. Efforts in introducing novel operation principle via materials and device engineering, such as constrain the pathway of ion migration, and precise control the interfacial barrier etc. could be employed to decrease the device variability.
Second, in-sensor computing mimicking the biological sensory system is expected to simultaneously perceive a wide variety of sensory information (visual, auditory, tactile, etc.). Previous works attempted to integrate different kinds of sensors, which are structurally complex. Although developing multimodality in a single multiterminal memristor is desirable, measurement accuracy decreases stemming from signal interference and calibration is required whenever the stimuli changes. Therefore, more attempts should be made toward decoupling the different signals during the operation. For example, exploring intrinsic material property to distinguish different stimuli, ion relaxation dynamics is favorable since it is not dependent on the dimension of device via materials engineering.
Third, multiterminal memristor exhibits a structural limitation and a low array density on line-design compared to the crossbar array of two-terminal memristor in a complicated circuit configuration, which results in low energy efficiency and processing speed of the whole system. Optimization of device structure is the efficient way to facing this challenging. For example, multiterminal memristor with vertical structure featuring a remote gate control via ion gel or other kinds of dielectric coupling can be employed, which is extendable to a crossbar array structure. For the device configuration, switching channel can be placed at every cross point of the top and bottom terminals, www.advelectronicmat.de and the dielectric and additional gate layer is deposited on them, ensuring the high performance and high density integrated circuits.
The in-depth investigation of QDs, 2D materials, and other emerging electronic materials has provided a promising route for multiterminal memristor to enhance its performance continuously. QDs possess controllable electronic structure at the atomic scale, as well as low cost, low-power switching, and biocompatibility. Based on their unique tailor capabilities (electrical, optical, chemical, mechanical, etc.), QDs enable to implement solution-based manufacturing technology such as rapid switching speed, multiple data storage capabilities and specific optical regulation terminal. 2D materials possess various bandgaps, stacking capabilities, and physical/chemical properties, which allow broad options for the construction and optimization of memristor. 2D material-based memristors with atomic thickness possess the potential for device minimization and rapid heat dissipation, thereby ensuring comparable energy consumption to that of biological synapses and neurons.
Developing multiterminal memristor-based neuromorphic system demands coherently collaborative research across various disciplines (physics, materials science, electrical engineering, biology, computer science, etc.) at different levels (devices/materials, circuits, system, architecture, algorithm). With better understanding of the different coupling mechanism in multiterminal memristor, new device structure should be developed to promote hardware-implemented. This Review looks at the structure, operation principles, progress, and challenges of multiterminal memristor with both concrete gate and virtual gate, both as building blocks for neuromorphic computing and in-sensor computing. Force-image recognition [111]