Emerging functions of two-dimensional materials in memristive neurons

Neuromorphic computing (NC), considered as a promising candidate for future computer architecture, can facilitate more biomimetic intelligence while reducing energy consumption. Neuron is one of the critical building blocks of NC systems. Researchers have been engaged in promoting neuron devices with better electrical properties and more biomimetic functions. Two-dimensional (2D) materials, with ultrathin layers, diverse band structures, featuring excellent electronic properties and various sensing abilities, are promised to realize these requirements. Here, the progress of artificial neurons brought by 2D materials is reviewed, from the perspective of electrical performance of neuron devices, from stability, tunability to power consumption and on/off ratio. Rose up to system-level applications, algorithms and hardware implementation of spiking neural network, stochastic neural network and artificial perception system based on 2D materials are reviewed. 2D materials not only facilitate the realization of NC systems but also increase the integration density. Finally, current challenges and perspectives on developing 2D material-based neurons and NC systems are systematically analyzed, from the bottom 2D materials fabrication to novel neural devices, more brain-like computational algorithms and systems.


Introduction
With the development of intelligent society, the need for data-intensive technology such as cognitive processing, big-data analysis and Internet of Things (IoT) is boosting [1].However, the traditional von Neumann computer architecture, where the CPU and memory are separated, is inefficient for solving these data-intensive problems.That is because most of the time latency and energy consumption are generated in the system bus between CPU and memory rather than information process, which is noted as von Neumann bottleneck [2].With the advancement of Moore's law, the switching speed of transistors are increasing, further leading to the growing gap between processor and memory, which negates the effects of scaling down [3].As a result, physically bringing the computation and memory units closer or even together is a key solution for overcoming the memory wall, such as near-memory computing and in-memory computing [4].
Neuromorphic computing (NC) is a significant concept for future computing architecture proposed by Carver Mead, which is different from traditional computer architecture and holds the specific characteristic of in-memory computing.Inspired by the intelligent and super high efficiency of the human brain, NC aims to physically emulate the structure and working functions of the biological brain [5].The two fundamental building blocks of such a system are neurons and synapses [6].Synapses, the connections between neurons, can store and modulate the connecting synaptic weights, which avoids frequently loading synaptic weights from memory into processing unit, and thus saves latency and energy costs.Neurons, which play the role of integrating input information in both temporal and spatial domains, can significantly reduce storage occupation and expand the NC system's computing power under the same amount of hardware.Early NC systems were implemented by CMOS technology, where a simple artificial neuron with limited functions required tens of transistors, falling far short of ideal energy and area reduction [7,8].
Recently, memristors have been demonstrated to be able to emulate biological neurons, dramatically transforming the complicated circuits into a few devices and even a single device [9,10].In fact, the mentioned memristor here is not the commonly known non-volatile memristor but rather a volatile one.The volatile switching refers to the property where the current increases abruptly when the applied voltage is greater than a threshold voltage and sharply decreases to high resistance state (HRS) when the applied voltage is lower than the hold voltage, while non-volatile memristor can hold its on-currents and store its resistance with high endurance and multilevel.Since the volatile behavior is suitable for neuronal functions such as the Integrate-and-Fire behavior [11], memristors with volatile resistive switching are emphatically studied to emulate neurons.For instance, a volatile filaments growth in the dielectric can result in neuron behaviors: metal cations or oxygen vacancies migrate under electric fields, undergoing REDOX reactions, and form fragile conductive filaments.The growth and rupture emulate the integration and fire behavior of neurons, respectively [12,13].At the beginning, a large number of artificial neurons were based on traditional bulk materials, which was easy for fabrication and large-scale integration.However, neurons with 3D bulk materials often features unstable volatility and thus need additional circuits such as a resistor and a capacitor to achieve neuronal leaky integrate-and-fire (LIF) behavior [10,14,15].Other low-dimensional materials, such as 0D and 1D materials, are difficult to function as standalone memristive layers.0D materials, such as quantum dots and nanoparticles (NPs), typically require other nanomaterials working as substrates, such as MXene decorated with Ag-NPs or heterojunctions between quantum dots layers and other materials [16].1D materials, like carbon nanotubes, are often composite with NPs when used as memristive materials [17,18], and the random distribution of carbon nanotubes within the interlayer makes it challenging to control [17].
Instead, 2D materials have the potential to address these limitations.In physics, two-dimensional materials are rigorously defined as materials in which carrier motion is confined to a 2D plane.In fact, in most studies, the occurrence of resistive switching (filament growth, vacancy migration, charge trapping) takes place in the vertical direction of three-dimensional space [19].Single-layer 2D materials, as well as multi-layer 2D materials bound by weak van der Waals (vdW) interactions, can be employed in memristive neuron devices to improve their performance.In recent years, 2D materials have been widely introduced to artificial neuron devices and have demonstrated their superior characters.It is important to acknowledge that plentiful reviews have explored the advancements in neuronal devices, but only a limited number have specifically focused on the emerging functions of two-dimensional materials as well as the physical mechanisms.Most reviews have provided comprehensive analysis and summaries of the implementation of neuronal functionalities, starting from neuron models or working mechanisms.However, few studies have elevated the discussion to NC systems and their algorithmic theory and large-scale fabrication [6,[20][21][22].
In this article, we provide a review of artificial neuron devices and the system-level applications of neurons, with their progress facilitated by 2D materials as well as their physical mechanisms.Figure 1 illustrates the beneficial properties of 2D materials in achieving neurons and neuron-based systems.The dangling-bond-free and vdW gaps of 2D materials allows different mobility and permeability of atoms and ions through neuron devices, endowing the basic but paramount volatility of device [23,24].The atomic-level thickness and vdW gaps of 2D materials enable the realization of ultra-thin dielectric layers of memristor with a minimal tunneling probability.This can further reduce the threshold voltage and enhance integration density [25].The nanoscale imperfection and intrinsic anisotropy can optimize characteristics and enrich the functions in a simple way [26].Additionally, the easy heterogeneous integration of 2D materials provide more novel mechanisms and structures to realize artificial neurons, such as memtransistor and grain boundary-induced memristor [27,28].Furthermore, many 2D materials have been proven to be multisensory responsible, such as MXene's peculiar optoelectronic transport properties and graphene's tactile sensing.Therefore, 2D-material-based neurons also have the potential to sense the external information such as light, pressure, sound, humidity and more, which is essential for achieving biomimetic computing system.In conclusion, 2D materials allow neuron devices to bridge the gap between processing, storing and sensing in traditional computer architecture.With the assistance of large-scale synthesis and transfer of 2D materials, hardware-implemented neuromorphic systems feature higher integration density.Combined with 2D materials' flexibility, 2D-material-based neurons are promising to construct flexible, all-in-one (sensing, processing, storing) systems.
The following paper summarizes and analyzes the advantages in detail.In section 2, biological neuron and artificial neuron models are first introduced, followed by analysis of the advantages of the better performance brought by 2D materials in neuron devices, in terms of significant electrical performance metrics, such as on/off ratio, power consumption and stability.Subsequently, system-level applications such as NC and artificial perception system composed of neurons, are summarized, focusing on connecting Reproduced from [30].CC BY 4.0.Reprinted from [31], Copyright (2022), with permission from Elsevier.Reproduced from [32].CC BY 4.0.Reprinted with permission from [33].Copyright (2021) American Chemical Society.Reprinted with permission from [34].Copyright (2021) American Chemical Society.structures, working principles and specific utilization of neurons as well as 2D materials.Finally, the current challenges and future directions of 2D material-based neurons and neuromorphic systems are presented.

Biological neuron and neuron characteristics
As mentioned above, in both biological brains and NC systems, neurons serve as the computational elements that transmit information through discrete spikes in the time domain.Understanding the working mechanism of the nature of biological neurons can provide researchers with endless ideas to build NC systems with more intelligent and require less power.A biological neuron receives bioelectrical signals from pre-neuron synapses, and the input signals regulate the membrane potential through spatiotemporal integration.At rest, the extracellular side of a neuron cell has excess positive charge, while the intracellular side has excess negative charge, which is maintained by the insulating property of the lipid bilayer membrane [35].As depicted in figures 2(a) and (b), when the membrane potential exceeds the threshold potential of neuron due to the input stimulus, the ion channel (such as Na + channels) rapidly opens, and Na + diffuses through the channel, generating an output action spike close to the Na + Nernst potential (+55 mV).The diffusion of Na + is then immediately suppressed, and the efflux of K + is enhanced due to depolarization, resulting in repolarization with the membrane potential returning to its initial state (figure 2(c)) [13,36].
In order to construct a NC system, it is necessary to mimic these series of behaviors of biological neurons with electronic devices.Therefore, various models abstracted from neuron behavior have been proposed, including the Hodgkin-Huxley model, Oscillation model and LIF model [6].Among them, Hodgkin-Huxley model is regarded as the most similar to biologically-accurate models as it emulates neuron behaviors from the bottom, such as biophysical characteristics of neuronal membranes and the voltage-controlled ion channels [37].However, most of the hardware demonstration of Hodgkin-Huxley (HH) neurons consist of numbers of passive and active devices, which costs large areas and huge amounts of energy [9], and the model are too complicated to analyze and make any exploration [38].On the contrary, LIF model has been closely investigated and widely implemented by researchers due to its simplicity in theory and feasibility of implementing it using simple electronic components [21].Specifically, the LIF model depicts the electrical response of artificial neurons under a series of pulse trains stimulus.As shown in figure 2(d), the membrane voltage of the device accumulates continuously with each input pulse, and once the membrane voltage reaches a threshold, the device outputs a spiking signal.Then after a refractory period, wherein the device remains unresponsive to any form of stimulus, it subsequently begins to accumulate the inputs again and prepare for the next spiking [15].To achieve LIF neurons, 'Integration' and 'Fire' are the two basic behaviors should be implemented.Filaments growing in the interlayer under the applied electric field can imitate the integration of input electrical stimuli.Besides, accumulation of electrostatic charge through tunneling effect, crystallization caused by Joule heating [39] are also demonstrated to achieve integration of input voltage.As for firing, that is to form an output spike with devices turning form HRS to low resistance state (LRS) and back to HRS at once.The resistive switching can be achieved by filaments' growth and rupture, as well as phase change mechanism.

Implementation of neuron in single device
Memristor, with a simple form of a sandwich-like, two-terminal structure but a dynamically reconfigured storage layer, has become the popular principle of artificial neuron [40].Not only because the simple and scalable structure, but also for the resemblance of ions dynamics to biological neuron, in which ions will migrate inside the interlayer under the electrical inputs, imitating the ions diffusion through the neuron membrane [41].
However, during the early developing time of memristor-based artificial neurons, most of the devices (either 3D bulk material based or 2D material based) had to be connected with auxiliary components such as capacitances and resistors (figure 3(a)), which cost huge area and consumed unnecessary power [10,15,42,43].In Bian's work, a W/Ag: ι-car/Pt threshold switching memristor (TSM) was built and tested to reveal the inside mechanism.The TSM features two threshold voltages under DC mode, corresponding to the voltage changing from HRS to LRS and from LRS back to HRS, respectively.In figure 3(b), when stimulated by a series of input pulses, the TSM device was triggered by a voltage pulse (1 V/1.2 s), during which the TSM was rapidly switched from HRS to LRS.However, it does not promptly return to the HRS spontaneously, resulting in an output that resembles a step function rather than an individual spiking pulse [15].It is this disparity (spiking or step) that results in connecting extra capacitance in parallel with memristors to achieve neurons' Integrate-and-Fire behavior.Figure 3(c) is the voltage variation across the capacitor in figure 3(a) [10], from which we can learn that the role of capacitance is to moderate the inputs of TSM based on its continuous accumulation ability.The parallel capacitance helps TSM return to HRS in time, leading to a spiking output.In most of the memristor-based neurons, including the above one, conductive filaments grow and rupture inside the interlayer, resulting in device transforming between HRS and LRS.Therefore, finding ways to grow easy-ruptured filaments can facilitate the timely change from LRS to HRS, also known as volatile switching behavior, and thus allows neurons free from connecting a large-area capacitor in parallel.Recently, more and more research has realized LIF neuron in single memristor device and most of them are based on 2D materials, which can facilitate the volatile switching of device.2D materials, working as the interlayer, feature high particle mobility and easy defect engineering, being able to modulate filaments' thickness, and allow LIF neurons achieved in single device.
Higher mobility of particles in the interlayer, which construct the filaments can lead to weak filaments growth and volatile behavior.Wu et al fabricated a monolayer MoS 2 LIF neuron with a lateral structure as shown in figure 4(a).To demonstrate the advantage of monolayer MoS 2 , researchers conducted a control experiment without MoS 2 and found that Ag filaments could not rupture automatically.The research drew a conclusion that the lateral monolayer MoS 2 rendered Ag filaments forming on its surface and filaments were easily ruptured due to the high Ag atom mobility in monolayer MoS 2 [44].Research has shown that the diffusion barriers for Ag atoms on the surface of monolayer MoS 2 and between bilayer MoS 2 are lower than the diffusion barrier in bulk HfO 2 [45,46].Since it is the dynamic equilibrium of particle migration, such as the competition between Ag migration driven by e-field and thermal diffusion driven by the concentration gradient, that lead to volatile behavior [47], 2D materials, especially MoS 2 , are more conducive to the natural fracture of conductive filaments.
Interfacial engineering can also enable control over the growth of filaments and 2D materials can achieve interfacial engineering through simple methods such as wet transfer or chemical vapor deposition growth [49].Artificial LIF neuron implemented by Zhang et al employed oxidized 2D material HfSe 2 as dielectric layer as well as the interfacial engineering material.Figure 4(b) shows the structure of device and it is obvious to see that the interface between the embedded ultrathin Ti electrode and HfSe 2−x O y was blurred, indicating that a TiO x layer formed due to the oxygen absorption from the oxidized 2D layer.Most importantly, the product TiO x layer would barrier the migration of Ag + and limit the forming location and diameter of filaments, ensuring the weak filament growth.The introduction of two-dimensional materials has enabled more reliable volatile behavior with vulnerable filaments, as demonstrated in figure 4(c) [29].
Furthermore, altering specific forming particles can also facilitate the fragility of filaments.In Zeng's work, the oxygen vacancy defects migrated in MXene and accumulated around Ag filaments, forming a fragile CF (figure 4(d)) [48].In addition, several research also mentioned a lower compliance current and a thicker interlayer would weaken the filaments, ensuring the achievement of single device neuron [30,42,47].Besides neurons in single device, researchers have also paid much attention to optimize other electrical characteristics of neuron devices with the assistance of 2D materials.Electrical characteristics such as threshold voltage, on/off ratio, integration period, power consumption and sensory function are summarized in table 1, together with materials and mechanisms.The important electrical metrics of the artificial neurons and their current status has also been plotted in figure 5.Under such horizontal comparison, the advantages of 2D materials in promoting device performance can be well analyzed, and the following subsections are mainly focused on them.

Implementation of stability and tunability
As for neuron devices, stability guarantees the accuracy and efficiency of NC, which has usually been demonstrated through statistical analysis of threshold voltage (V th ) distribution.Tunability, represented through the gate-voltage-controlled V th distribution, can enrich the function of neurons.Threshold voltage, is the key indicator of neurons' stability and tunability.Under DC testing mode, where input voltage raises from 0 V to a higher voltage, when it reaches threshold voltage, the device will change from HRS to LRS.Under pulse testing mode, threshold voltage is the input voltage value boundary between having spiking responses and no spiking responses, representing the lowest operating voltage.
In NC system, information is transmitted through pulse trains with specific amplitude [54].Once the threshold voltage of neuron varies drastically, the input signals will stochastically activate the neuron, leading to wrong data transfer.Therefore, implementing neurons with high stability is essential to further application in neuromorphic system.Solutions assisted by 2D materials have been investigated as follows.In Yu et al's work, neuron based on MXene decorated with Ag NPs showed a 3-times decrease in standard deviation of distribution of threshold voltage than undecorated neuron (figure 6(a)).This phenomenon was  Continuously optimizing these four metrics can lead to a smaller array footprint, lower power consumption, more bio-inspired computational capabilities, and enhanced computational reliability.Other reference information: Au/Al2O3/Graphene [28], Cu/MXene/Cu [36], Ag/MoS2/Au [43], Ag/MoS2/HfAlOx/CNT [30], Ag/Ag NPs/MXene/ITO [31], Au/MoS2/PEO/Pt [34].
attributed to that Ag NPs between multi-layered MXene could facilitate the formation of Ag conductive filaments, leading to a uniform process of filaments growing [31].Zhang et al neuron embedded with Ti layer demonstrated a more concentrated distribution of V th (figure 6(b)), owing to the synergistic effect of diffusion barrier and point contact of migrating Ag atoms.The TiO x layer would restrict the number of oxidized Ag atoms and induce the growing position, unifying the filaments growth in two dimensions of both thickness and position [29].Modulated by particle decoration and interface engineering, the formation of conductive filaments signifies higher cycle-to-cycle, device-to-device stability, which was beneficial to computation accuracy of neuromorphic system [55].
Apart from controlling the standard deviation of distribution of threshold voltage, which is associated with device stability, a tunable distribution of threshold voltage has also been a focal point of research, which presents the controllability of device stochasticity.The device level tunable stochasticity, allows the probabilistic spiking neurons viable for stochastic neural network as well as hardware security [56].The certain probability distribution such as sigmoid functions was proven to enable improved error rates, compared to deterministic implementations [57].Therefore, approaches to control the probability distribution of threshold voltage have been widely explored.Yan et al overcame the challenge of tunable statistical distribution of neurons threshold voltage by applying a three-terminal, gate modulated memristor based on tin oxide/MoS 2 heterostructure, as sketched in figure 6(c).The inherent random motion of oxygen ions leaded to the uncertain set voltage, while the insertion of MoS 2 layer made it possible to modulate the contact-energy barrier between MoS 2 and SnO x with gate bias, effecting the movement of oxygen ions and hence resulting in tunable probability distribution of threshold voltage (figure 6(d)).A Boltzmann machine (BM) was then constructed to solve combinatorial optimization problem, proving the advantage of tunable stochastic neuron [32].
The reliability and controllability of threshold voltage distribution is one of the major challenges for memristor-based-neurons toward commercialization [58].Apart from the above-mentioned material engineering and gate-structure, more solution should be proposed to better control the filament growth, especially taking advantage of 2D materials' characters, for instance, vdW gaps and dangling-bond-free.What's more, modulating parameters of fabrication process is another way to propel the stability of artificial neurons, which is still short of study [59].

Trade-off between high on/off ratio and low power
2D materials, working as the resistive switching interlayer, can not only contribute to fragile filaments, implementing volatile neurons, but also determine neurons' resistance and operating current, which highly affect the on/off ratio and power consumption of devices.Neurons with high on/off current ratio can preserve classification accuracy in spiking neural network (SNN) [60,61] and enhance the reliability of in-memory computing, when programming a single state under variations in multi-bit configurations [62].Wang et al demonstrated a 2D TiO x -based neuron as shown in figure 7(b), with on/off ratio up to 10 9 , which is 10 3 times greater than other neurons, especially with the ultrathin, large bandgap 2D interlayer [51].To achieve similar switch ratios and volatility as in traditional oxide-based memristor, the dielectric layer needs to be thicker than 10 nm and requires specific selection of anode metal and oxide species [63,64].
[51] and two other neurons with 10 6 high on/off ratio can cost nearly hundreds of microwatt for one spike (figure 7(a)) [29,42].As the size and device density of functional memristor networks increase, a practical memory or computing system may require billions of neurons and synapses [40].When building such large-scale network with a limited energy consumption demand alike human brain, these superhigh on/off ratio neurons will stand at a disadvantage.Therefore, some research aims to achieve low-power neuron devices.
Wang et al fabricated a flexible SNN networks with synapses and low-power neurons.An energy cost of 1.9f Joule for one spike was at least three orders of magnitude lower that of biological neurons and other reported artificial neurons.As figure 7(c) demonstrated, the conductive filaments between two electrodes are very thin, especially in the MoS 2 layer, resulting in a low fire-current at picoampere level.A control experiment was also conducted where the Ag/HfAlO x /CNT memristor was tested and showed 10 000 higher power consumption, proving the physics that MoS 2 ensured tenuous Ag filaments growth [30].Moreover, the low current could restrict the redox reaction and migration of Ag atoms, furtherly limiting the diameter of filaments [67], achieving a stable low power consumption (figure 7(d)).
Based on the analysis, we conclude that, in situations where the HRS resistance or off-state current is fixed, the trade-off between on/off ratio and power consumption should be considered at the application and system level (figure 7(a)), taking pivot spots of network and spiking rate of neurons into account to construct lower power and higher efficiency system.Although 2D materials have been demonstrated for achieving high switching ratios and ultra-low power consumption respectively, we believe there is still room for further development of 2D material-based memristors to achieve a better balance between on/off ratio and power consumption.Upon comparing table 2, we observed that compared to conventional HfO 2 memristors, memristors with 2D materials as the memristive layer exhibit an increase in off-state current, which can be attributed to factors such as bandgap, memristive layer thickness and metal-electrode contact [68].Therefore, engineering on the electrical characteristics of 2D materials and metal-electrode contact [69] to reduce HRS current should also be considered as one of the future research directions.

Implementation of optoelectronic sensory neurons
With the development of autonomous systems in a myriad of fields such as robotics, IoT and self-driving vehicles, the demand for perception system comprising sensor and computational networks with less area and power consumption has been soaring [70][71][72].Research above mainly focused on mimicking electrical signal processing in biological neuron, while in-sensor computing neuron devices were neglected.The burgeoning sensory neuron can directly transform raw data from the real world to electrical spiking signals, instead of using conversion circuit such as an analog to digital converter (ADC) or other hardware components for cascade connection with the von Neumann architecture [20].Sensory neurons, should be able to integrate external stimuli such as light, stress, heat, etc., with electrical inputs to generate output electrical pulses.In order to realize this in-sensor computing device, the intermediate dielectric layer must possess both sensing capability and resistive switching properties.Such strict demand brings challenges to the implementation, resulting in the restricted sensory functions, that most of the sensory neurons are opto-sensing, hardly for tactile, auditory and so on.Therefore, this section focuses on implementation of optoelectronic neurons.
As for the existing solution, 2D materials have played the key role of the sensory interlayer.MXene, for instance, features fascinating optical and electrical properties comprising transparent, saturable absorbance and superconductivity, which can also be tuned through altering its thickness and the intercalations of chemical functional groups [73].The ultraviolet irradiation producible oxygen vacancies in oxidized MXene (O-MXene) can also transform light information into carriers' generation and migration changes [74].TMDs materials also have been widely used for optoelectronic devices, among which MoS 2 has been of particular interest due to its tunable bandgap and charge carrier mobility.The appropriate band gap and trapping center in the energy gap of MoS 2 brings extraordinary photo-induced catalytic ability to MoS 2 nanosheets, leading to implementation of high photoresponsivity [75][76][77].The following cases demonstrated the photoelectronic mechanisms in detail when the above 2D materials were applied into neuron devices.
Zeng et al fabricated an Ag/O-MXene/SiO 2 /Si optoelectronic neuron by introducing an O-MXene layer.Figure 8(a) depicts the structure of device and the energy band diagram under 365 nm UV light irradiation.Oxygen vacancy defects in TiO 2 , which was generated by MXene oxidation, could absorb photon energy and produce electrons, facilitating the reduction of Ag cations.As a result, the higher intensity of 365 nm UV light stimuli, the less time needed for optoelectronic neuron to fire (figure 8(b)), which proved to be light sensitivity [48].In Zhang et al's work, MXene was also introduced due to its excellent electrical and optical properties [78].The easily O-MXene surface brought in oxygen vacancies, resulting in a high electronegativity and ultimately accelerating the diffusion of Ag + .The introduction of oxygen vacancies and Ag NPs was also proved to broaden the spectrum in visible band.In such foundation, this optoelectronic device could integrate electrical as well as optical input signals of different spatial and temporal distribution, which approximated further to biological neuron, for most neural activities were activated by synergy of electrical and optical signals with high spatiotemporal dependence [31].
Charge traps in energy band of 2D materials also has been applied to optoelectronic neuron devices.Chen et al came up with an optoelectronic graded neuron, as sketched in figure 8(c).Different from spiking neuron aforementioned with refractory period, graded neuron exhibited multi-level responses in terms of magnitude of current depending on temporal summation of the input optical stimulation, which was illustrated in figure 8(d).This phenomenon owed to the shallow charge trapping center in MoS 2 film, which would trap and release carriers in a shorter time than deep trapping center to allow graded neuron characteristics in temporal domain instead of synaptic plasticity behavior [76,79].Although the optical response characteristics and principles of 2D materials have been extensively investigated in traditional optoelectronic devices [80], their integration with resistive switching principles enables the development of more biomimetic and intelligent optoelectronic devices.
In summary, the combination of sensing and memristive properties in 2D materials enables the realization of optoelectronic neurons.In contrast to traditional CMOS circuits that require separate and redundant sensing, transmission conversion, processing and memory modules, optoelectronic neurons as well as all-in-one devices based on 2D materials offer significant circuit scale reduction [81,82] while enabling the implementation of diverse bio-inspired functionalities.

Implementation of NC based on 2D materials
NC, inspired by the intelligent and super high efficiency human brain, aims at physically emulating the structure and working functions of biological brain [83].In human brain, there are approximately 10 11 neurons interconnecting each other, forming a massive network, which allow us to possess the ability of perception, learning and decision making [84].To pursue the high efficiency of human brain, neuron arrays have been built based on the implementation of artificial neuron devices, in which the LIF behavior and stochasticity nature are prominent among various types of devices.Thus, neuron based SNN utilizing LIF behavior and stochastic neural network deploying stochasticity have been vigorously developed.Although the existing 2D-based NC systems are limited, the introductions of 2D materials have indeed broken through several challenges of hardware-implemented NC systems.Moreover, 2D materials' easy planar integration, vertically stacking integration and multi-sensitive properties [85] will further facilitate the performance of hardware implemented neuromorphic systems.Even so, there are still many unresolved problems and fields that 2D materials have not covered.Considering that NC systems based on traditional materials have developed more completed, those cases without the introduction of 2D materials are also analyzed as a reference for 2D-based NC systems to develop.

SNN
SNN, as one of the main forms of NC system, imitates the way our brains recognize and study in a more sophisticated way than traditional artificial neural network (ANN) does.The basic structures of both ANN and SNN are the same, where a crossbar array with synapses located at each cross point and neurons at the output points as sketched in figure 9.These two fundamental units work cooperatively to realize learning and recognition of neural network.Synapse array works as weight matrix and neurons play the role of nonlinear activation functions [86].However, there are some fundamental differences between ANN and SNN.(1) information in ANN is in form of continuous-valued electronic signals while neurons and synapses in SNN communicate through spike trains.(2) SNN expands information processing into both spatial and temporal domains while ANN can only process information in one moment time.These differences can not only promote learning ability but also reduce power consumption of neural network [54,87,88].
As SNN works, it is necessary to acquire information from the external environment and convert it into electrical signals, and this requires the involvement of artificial sensory neuron.Spiking trains from neurons will then be fed into synapses in parallel sequence.Synapses renew their conductance by spike timing dependent plasticity (STDP) learning rule as spiking trains pass through.Finally, they sum up in both time and spatial domain to stimuli neurons.Here, we require artificial neuron that is stable, controllable, and capable of processing both excitatory and inhibitory inputs.Different spiking frequency of neurons represents the classification results [91,92].The collaboration and interaction of neurons with synapses can achieve unsupervised learning, which has more extensive application than supervised network (such as ANN) for it can learn through data sets without labels, just like human exploring the unknown [93].With the performance of both neurons and synapses developing, more and more hardware-implemented SNN with 2D materials have been demonstrated, which not only resolved hardware problems, but promoted the integration density and learning ability as well.
The 2D material family, with a rich band gap distribution, can cover from metalloids, semiconductors to insulators and the bandgaps can also be modulated by the number of layers [94].This diversity and tunability promote the design of novel neural devices to address biomimetic challenges in system applications.In human brain, one neuron can integrate both excitatory and inhibitory stimulus from multiple synapses, which was impossible for traditional SNN hardware to realize, for synapses and neurons can hardly process and generate signals with both positive and negative levels in simple devices, corresponding to excitatory and inhibitory stimulus respectively.This missing multi-connection between synapses and neurons would limit the learning efficiency of unsupervised learning network.
In Won et al's work, a multi-terminal floating-gate memristor (MT-FGMEM) was fabricated to solve this problem.As figure 10(a) illustrated, the MT-FGMEM utilized graphene as the float gate, h-BN as the tunneling layer and MoS 2 as conductive channel, taking full advantage of the various band gaps and easy stacking of 2D materials.Multi-terminal and tunneling effect can transform positive (negative) signals to positive (negative) charges, and the floating gate can generate both positive and negative electrical levels by simply integrating the tunneling charges.This resulted in the implementation of both spatiotemporal integration of neurons and spiking-based multi-level memory modulation (equivalent to STDP) of synapses (figure 10(b)).A 9 × 3 SNN array consisting of the above neurosynaptic connection was then constructed and used for pattern classification, facilitate a large-scale fabrication of hardware unsupervised SNN (figure 10(c)) [95].Indeed, this research still has limitations as the implementation of neuronal functionalities often requires additional comparator units, which can occupy significant space.Additionally, the large-scale manufacturing of two-dimensional materials, especially h-BN, poses challenges in terms of high production costs, technical difficulties.As a result, the array scale of these materials remains relatively small.
The optoelectronic sensory neurons based on 2D materials, have the potential to expand the functions of SNN, realizing multimodal and multi-task learning.
Although multimodal learning SNN have not came out, a multimodal ANN based on multi-sensory memristors that combined visual data sensing and relative humidity sensing was demonstrated by Wang et al which could bring inspiration to the development of 2D material-based multimodal SNN.The interlayer of the multimodal memristor was a MXene nanosheets/ZnO NPs heterostructure (figure 10(d)), among which ZnO NPs were photo-active and MXene-ZnO heterojunction would absorb water molecules, restricting the growth of the oxygen vacancy based conductive filaments (figure 10(e)).A multimodal ANN (figure 10(f)) built on this memristor successfully imitates perception systems of human eyes, where the recognition accuracy was strongly affected by the environment such as relative humidity condition [89].
Multi-task learning has proved its potential to develop future intelligent system, such as obstacle avoidance and object tracking in autonomous driving [96].In conventional software implemented neural network, multi-task learning is achieved by introducing a new loss function, which jointly combine two target tasks [97].When it comes to hardware network, neuron devices will be capable for multi-task learning as long as appending more input ports to the device, jointly modulating the membrane potential.Therefore, neurons with the ability of integrating electrical and optical signals are proved theoretically to be capable of multi-task learning.Yu et al further demonstrated it with experiment and simulation.An SNN array was constructed with synapses and light sensitive neurons, in which Ag NPs decorated MXene layer could absorb light and reduce the threshold voltage.Both electrical and optical inputs can be integrated to stimulate the fire which will extend another dimension without extra hardware costs, realizing multi-task learning [31].
Introducing 2D materials can also build SNN with higher integration density and flexibility.In traditional hardware implemented SNN, synapses and neurons are in heterostructure, where synapses are based on memristors and neurons are implemented by CMOS, which causes limitation of integration densities and mismatch of device performance [98].Enlighted by the fact that devices based on non-volatile and volatile memristors can realize synaptic and neural functions respectively, Wu et al fabricated a reconfigurable memristor textile network (figure 10(g)), where synapses and neurons are based on totally the same structure.The Ag/MoS 2 /HfAlO x memristor, as the fundamental building block, could convert from a synapse into a neuron simply by reducing the compliance current under 10 µA.Traditional bulk material HfAlO x could only produce stable and thick filaments, while the high mobility of Ag atoms in MoS 2 leaded to fragile filaments growth [44].The stack of these two materials allowed the transition between non-volatile behavior and volatile behavior [30].And the low-dimensional materials realized a flexible hardware-SNN for intelligent temperature modulation (figure 10(h)).This research provides insights into a feasible future direction for SNN fully based on memristors.However, challenges remain in compliant current control and the network lacking the ability for autonomous learning.
Currently, there is no widely recognized and feasible hardware structure for SNN that can achieve multi-connections between synapses and neurons while supporting both excitatory and inhibitory inputs.In addition to the challenges of large-scale growth and poor uniformity of 2D materials, the scale of hardware-based SNN with full memristors is still limited to around 50 devices [95].

Stochastic neural network
Apart from the significant character of spatiotemporal integration of stimuli, which performs a vital role in SNN, neurons also show rich stochastic property, which contributes to the implementation of stochastic neural network.The stochasticity nature of neurons owes to the ensembles of random movements of atoms and ions in the interlayer (figure 11(a)) [99,100].Indeed, neurons used in stochastic neural networks require controlled and predictable randomness rather than complete chaos.In stochastic neural network, neurons with controllable stochasticity can help in tackling complex optimization problems, which would cost huge energy and hardware occupation in conventional deterministic computers.2D materials, with the easy multi-terminal regulation, can offer simplicity in gate modulated stochastic neuron structure.Through adding gate bias to modulate the probability distribution of specific device's characteristics into an ideal curve, such as sigmoidal functions, stochastic neurons can be harnessed for probabilistic computing and BM.
Probabilistic computing (p-computing) was demonstrated realizing integer factorization and invertible logic based on spintronics technology [102], which shares similar concepts with quantum computing but can avoid decoherence and cryogenic operation requirement of quantum computing.The individual p-bit unit corresponding to one bit data, is the stochastic building blocks of p-computing, whose outputs take on the external input, following a probabilistic distribution [103].As a stochastic device, neuron have also been demonstrated to implement p-bit unit, although without utilizing 2D materials.Woo et al fabricated p-bit unit consisting of a volatile memristor and a comparator.Under pulse train input with different voltages, the memristor fires in a different rate, in accord with neuron LIF behavior, which leaded to average voltage variation within a period of time.The relationship curve between V in and time-averaged V out followed approximately the sigmoidal functions (figure 11(b)), which was essential for the probabilistic output.Cost functions were defined according to logic operations to calculate the feedback input to the output p-bit.Through calculating the ratio of time-average V out and reference voltage V DD , probabilities of p-bit to be zero and one state were determined, realizing a one-bit p-computing.Moreover, in terms of inverted logic and complex logic such as full adder, p-computing were also realized and it achieved results in one shot (figure 11(c)), leaving out the training process in machine learning, which would consume more power [101].Although the neuron device in this research was not based on 2D materials, p-bits principle and peripheral circuits design were still precious enlightenment.Furthermore, the future introduction of 2D materials may simplify the extra circuits and realize a better device-to-device uniformity, which is still needed for a more accurate calculation result, according to researchers.
BMs consisting of stochastic processing units as well, contributes a lot to solving combinatorial optimization problem, for it can carry out operations in parallel, resulting in a speedup over the sequential simulated annealing algorithms [104,105].Yan et al made a breakthrough of realizing tunable statistical distribution of neurons set voltage by applying a three-terminal, gate modulated memristor based on tin oxide/MoS 2 heterostructure, which was also mentioned in section 2.3 (figure 6(c)).The probability of device threshold voltage increases when top electrode voltage increases (figure 11(d)), and the relationship follows a sigmoidal distribution, which can be modulated by gate voltage.The tunable stochastic device was then constructed into a BM circuit with other CMOS peripheral components (figure 11(e)), which was able to implement different annealing strategies to solve an NP-hard combinatorial optimization problem.Four different annealing strategies with their changing process of energy, also known as effective temperature were shown in figure 11(f), which could be modulated by tuning V g , in order to reach optimal convergence efficiency [32].The hardware implemented simulated annealing process in the BM can much reduce the occupation of processing unit and memory when compared to conventional computing architecture, achieving lower power consumption and effective problem-solving [106].
This section has summarized the application of stochastic neurons, providing ideas for future application of neurons' tunable stochasticity with the assistance of 2D materials.

Towards artificial perception system based on 2D materials
Different from conventional sensory system, which is composed of numerous sensors, large signal processing circuits and memories based on von Neumann architecture [107], artificial perception system utilizes simple Visual perception of a nervous memristor-based system imitating the human visual system.(d) Schematic of the artificial perception system: Left is the flowchart of the meeting control system operating process; Right is the evolution of parameters in the meeting control system during the meeting process.Reprinted with permission from [33].Copyright (2021) American Chemical Society.devices, such as neurons and synapses, to detect environmental signals and at the same time transform the information into NC-available spiking trains [108].By incorporating in-sensor neural processing, neural sensors can be further combined with SNN to implement learning, recognition and even reaction.What's more, the artificial perception system can be particularly sensitive to the needed information such as illumination in specific wavelengths, rather than detecting and memorizing all information before processing useful data, as conventional sensory systems do, which inevitably increases the hardware cost and energy consumption [109,110].Existing perception systems can be divided into two types based on the combination mode of sensors and neural devices.One type integrates sensors into synapses and neurons, forming a more compact structure and direct transformation of information; The other type connects well-developed sensors to converters then to neural devices, which can break through the limitation of materials both capable of electrical interlayer and sensing, achieving perception system with wider sensitivity [86].
Sensory neurons and synapses have been invented by applying sensing materials to the resistive interlayer of devices, enabling modulation of device conductance by environmental signals [70].By connecting them with other neural devices, a perception system with all-in-one (sensing, storing and processing) functions can be realized.An integrated visual perception system was constructed by Yu et al to emulate conditional response, which could react to light illumination with a robotic hand after advanced training.When light was applied to the MXene-based neuron in figure 12(a), the neuron would integrate electrical and optical signal spatiotemporally and form spikes to control the robotic hand to make a fist (figure 12(b)), accomplishing the sensing and processing functions [31].Pei et al also built a memristor-based artificial visual perception nervous system (AVPNS), where artificial neuron was connected with a light-sensitive synapse enabled by PbS quantum dots (figure 12(c)).The system was demonstrated in the circumstances of autonomous driving during meeting cars.As figure 12(d) shows, when the input light intensity enhanced, the resistance of synapse would reduce and the pre-neuron current increased, which further resulted in a more rapid frequency of output pulse trains.By detecting the output frequency and distance between two cars, the AVPNS would modulate the driving speed to enhance safety when two cars passed each other [33].
The aforementioned research focuses on the feasibility of utilizing neuron and synapse devices to achieve artificial perception systems.However, these systems are currently limited in scale, as a single sensory device can only perceive extremely limited information, such as a beam of light.Such systems are primarily suitable for functional demonstrations.To enable practical applications, it is necessary to develop arrays of sensory neurons with a large-sale synthesis 2D materials, that can convert a complete image into a series of electrical pulses for subsequent learning and recognition tasks (figure 13(a)).This will pave the way to constructing visual perception system processing real-time full view.MXene-based neurons fabricated by Zeng et al (figure 13(b)) were constructed into a 64 × 64 array, being able to perceive a clear moving trajectory of illumination (figure 13(c)) due to the integration and all-or-nothing nature of artificial neuron [48].Chen et al fabricated an action recognition system based on optoelectronic graded neuron array, which was analyzed in section 2.3.The neuron sensor of this kind was proven to be able to record and show the contour of the trajectory in the visual field, fusing the spatiotemporal information of a series of frames into compressive images with agility and high accuracy.Furthermore, by modulating V g , the bioinspired neuron sensor could seize trajectory with different speed, capable of implementing action recognition with various motion, which was demonstrated in figure 13(d) [79].The specific visual neuron array features the nature of capturing moving trajectory of objects, and is able to be connected directly to SNN without signal converter, which is promising for the future autonomous driving and robotic motion.
As materials for interlayer of neuron that are both capable of filaments growing and sensing are limited, connecting separated sensors based on 2D materials with neural devices is the other strategy to realize in-sensor neural processing, especially with more abundant sensing ability, such as tactus [112][113][114] and auditory [115].Tan et al fabricated an artificial tactile nerve by connecting a MXene based tactile sensor with an ADC-LED circuit and a synaptic photomemristor (figure 13(e)).Once the MXene sensors were applied to external pressures, the output voltage would increase.The change in voltage signal would then be transformed into optical spikes by the ADC-LED circuit and the spiking frequency was related to the pressure intensity.The artificial nerve was then constructed into an array, which was able to recognize Morse code, braille characters and object movement (figure 13(f)) [111].By connecting a tactile sensor to the back gate of a MoS2-based light-sensitive memtansistor [116], Sadaf et al implemented a visuotactile neuron for multisensory integration.Coupling both the vision and tactus, the sensory neuron can achieve a higher sensitivity as well as some special biomimetic functions for multisensing.It is enlightening that they have developed a spike encoder circuit that converts the analog current response from the neuron into digital spikes, which allows the module be cascaded with neural network in an easier and more accurate way [90].Combining sensors with neurons enables simple electrical levels to be transformed into spiking, which is compatible with NC.Therefore, the perception system can be designed and trained into a more intelligent and energy-saving one, meeting the demand and promoting the evolution of robotics, IoT and autonomous vehicles [117].Bottom is in term of materials including taking full advantage of 2D materials' characteristics, such as sensory functions (left) and dangling-bond-free (middle), and developing large-scale synthesis and integration method with BEOL of CMOS (right).Middle is in term of neuron devices, pursuing more biomimetic functions (left) as well as simplify device's structure (right).Top is in term of neuromorphic systems based on neurons, with more algorithm and architecture to be explored.

Summery, perspectives and challenges
In the last few years, we have witnessed striking progress on artificial neuron based on two-dimensional materials as well as the various application of neurons in system level.Outstanding features of 2D materials have been fully demonstrated in diverse neuron devices, neuron-based NC and perception systems.In terms of devices, vdW gaps of 2D interlayer can facilitate the rupture of conductive filaments, leading to reliable volatility of neurons.Easy assembly of heterojunctions and interfacial engineering between 2D materials and bulk materials can further promote neuron performance such as lower power and better uniformity.Moreover, optoelectronic properties of 2D materials enable the construction of neuron sensor, realizing in-sensory computing.The easy planar integration and easy vertically stacking integration of 2D materials allow neuron devices to be fabricated into array, forming SNN with synapses, stochastic neural network and all-in-one perception system.
However, there still exits a large room for researchers to further explore, as summarized in figure 14.In terms of 2D materials, much more advantages have not been demonstrated in neurons, which limit their electrical performance and sensory function [118].For instance, sensory neurons based on 2D materials with tactus, audition are still missing, while 2D materials such as transition metal dichalcogenides have been demonstrated to be sensitive to strain [119,120].At the core of integrating sensors with neural devices, harnessing the full potential of 2D materials to realize all-in-one neurons holds the promise of creating perception systems that are comparable to or even surpass human capabilities.Additionally, vdW integration of 2D materials can enable the creation of a series of artificial heterostructures and superlattices with atomically clean and electronically sharp interfaces highly disparate materials, due to their dangling-bond-free nature, allowing efficient electron tunneling, few interfacial trapping states and various device structure with more abundant electrical properties [121][122][123].
When it comes to fabrication technology, there is an urgent need for fabricating large-area 2D materials, especially those that are compatible with BEOL of CMOS fabrication.In the last few years, several methods were found to integrate large-scale 2D material on CMOS chips by transferring the already-grown monolayer materials at room temperature [51,[124][125][126][127], which will limit the yield and is difficult to realize automation fabrication.Instead, it is promising to directly grow 2D materials on CMOS circuits with a low thermal-budget, similar to the demonstration of Zhu et al innovative metal-organic chemical vapor deposition fabrication process [128].Realizing CMOS-compatible BEOL integration of more and more 2D materials with low-thermal-budget, pursuing wafer-scale uniformity and better electric properties of 2D materials are the strategic way to accomplish commercialization of 2D-material-based neuromorphic system.3D integration of 2D material-based memristors and CMOS circuit will not only increase the integration density, but also give full play to each other's potential [126].
As for neuron devices, far more biomimetic functions have not been widely realized, such as multi-input with both inhibitory and excitatory [129], tunable spiking [14,130] or HH model with simple devices [130,131], which can further enhance the power reduction and data processing capacity.Lately, Won et al have realized a spatiotemporal integration neuron based on floating-gate memristor with multiple terminals.Positive (negative) inputs could cause holes (electrons) tunneling through dielectric layer to the floating gate, leading to excitatory-inhibitory and spatiotemporal integration behavior, which paved the way to fabricating multi-connection between synapses and neurons, much closer to human brain structure [95].When compared to conventional MOSFET in CMOS circuits, robust basic block which can act as both a neuron and a synapse still remains to be explored, in the field of neuromorphic system.This shortage will limit the large-scale integration, structure flexibility as well as functional diversification of neuromorphic system, causing NC system being inferior to the mature CMOS system.
Last but not least, at system level, since the memristive system has intrinsic learning and signal-processing ability, deepgoing study of human brain and significant innovation of algorithm are required to narrow the gap between 'artificial intelligence' and 'human intelligence' [83,132,133].

Figure 2 .
Figure 2. Biological neuron behavior and LIF model.(a) Schematic of the biological neuron.(b) Schematic drawing of how membrane potential is modulated by ions movement through ion channels [13].John Wiley & Sons.[© 2018 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim].(c) Different phases of the action potential in a nerve cell.Reproduced from [35].CC BY 4.0.(d) LIF neuron output current spikes under a train of input voltage pulses, showing integration, fire and refractory period.© [2022] IEEE.Reprinted, with permission, from [15].

Figure 3 .
Figure 3. Artificial neurons based on traditional bulk materials with extra capacitors.(a) Schematic illustration of neuron achieved by one memristor and 2 R1C circuit.© [2018] IEEE.Reprinted, with permission, from [10].(b) Volatile switching process under pulse mode of a single threshold switching memristor.© [2022] IEEE.Reprinted, with permission, from [15].(c) The voltage variation across the capacitor with circuit structure in figure (a).The voltage increases when the TSM is in the HRS and decreases when in the LRS.© [2018] IEEE.Reprinted, with permission, from [10].

Figure 4 .
Figure 4. Artificial neurons in single device implemented with 2D materials.(a) Schematic of a planer artificial neuron with MoS2 working as the memristive interlayer [47].John Wiley & Sons.[© 2020 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim].(b) Cross section TEM characterization of Ag/Ti/HfSe 2−x Oy device.It can be found that a TiOx layer formed due to the interfacial engineering enabled by 2D material.(c) Resistive switching model with embedded TiOx layer, where the barrier layer and point contacts work together to restrict the filaments.© [2023] IEEE.Reprinted, with permission, from [29].(d) Mechanism diagram of Vo 2+ involvement in the formation and rupture of CF. © [2023] IEEE.Reprinted, with permission, from [48].

Figure 6 .
Figure 6.Stability and tunability of artificial neuron enabled by 2D materials.(a) The distribution histogram SET voltage of Ag NPs undecorated and decorated device, proving the enhanced stability.Reprinted from [31], Copyright (2022), with permission from Elsevier.(b) Hundred-cycle V th and V hold statistical distribution map, demonstrating the comparison of stability between devices without/with Ti layer, namely, the Ag/HfSe 2−x Oy device and the Ag/Ti/HfSe 2−x Oy device.© [2023] IEEE.Reprinted, with permission, from [29].(c) Schematic of the heteromemristive device, with MoS2 enabled gate-tunability.(d) The probability distribution of SET voltage under different gate voltages, showing exponential-class sigmoidal distribution function.Reproduced from [32].CC BY 4.0.

Figure 8 .
Figure 8. Optoelectronic neuron device based on 2D materials.(a) Schematic of mechanism of threshold switching behavior assisted by UV light (365 nm in wavelength, 100 µW cm −2 in light intensity) irradiation.Device structure (left), consisting of Ag (top electrode), O-MXene (white interlayer), SiO2 (light slate gray interlayer) and Si (Dim gray bottom electrode).(b) Variation of firing time of OM-AOM under dark and UV light with different intensities.© [2023] IEEE.Reprinted, with permission, from [48].(c) Schematic of the MoS2 memtransistor enabling light sensing, processing and memorizing.(d) Artificially graded neurons for encoding temporal vision.A graded neuron can respond to sequential stimulation with nonlinear temporal summation characteristics.Reproduced from [79], with permission from Springer Nature.

Figure 9 .
Figure 9. Schematic of basic SNN structure and the significant role neurons play in.Input information from external environments will be first collected by the sensory neuron array and transformed into spikes.2D materials can work as both sensory materials and memristive materials.Under the stimulation of spike trains, the synapses will modulate their conductance following the STDP rule, processing the input spikes with various amplitudes.Finally, the processed spikes over a period of time are integrated by neurons and result in the different firing frequency, which indicates the classification results.And with the assistance of two-dimensional materials, these neurons can achieve enhanced functionality and process more complex application [89].John Wiley & Sons.[© 2021 Wiley-VCH GmbH].Reproduced from [90].CC BY 4.0.Reproduced from [79], with permission from Springer Nature [28].John Wiley & Sons.[© 2022 Wiley-VCH GmbH] [47].John Wiley & Sons.[© 2020 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim].Reprinted with permission from [34].Copyright (2019) American Chemical Society.

Figure 10 .
Figure 10.Progress of hardware-implemented SNN achieved with 2D materials.(a) Cross-sectional schematics, and operation principle of MT-FGMEM.(b) Schematic of basic synapse-neuron assembly for unsupervised learning process by synaptic STDP and neuronal LIF functions.(c) Optical image of neurosynaptic array with neuron-FG and synapse.Reproduced from [95].CC BY 4.0.(d) Schematic illustration of the flexible and multimodal sensing memristive device.(e) Schematic illustration of a proton-mediated resistive switching mechanism of the MXene-ZnO memristor.The protons can combine with the VÖ through the electrostatic force, and restrict the formation of the oxygen vacancy-rich bridge in the memristor.(f) Schematic illustration of in-sensor low-level processing by utilizing the MXene-ZnO memristor to sense the information, suppress/filter the noise, and specialize the features [89].John Wiley & Sons.[© 2021 Wiley-VCH GmbH].(g) Schematic image of textile memristor network, including top-layer device with synaptic plasticity and bottom-layer with neural functions.Middle is the structure of fiber-based memristor consisting of Ag/MoS2/HfAlOx/CNT.Right is the scanning electron microscopy (top) and cross-sectional transmission electron microscopy (bottom) images of fiber-based memristor.(h) Photograph of the fiber-based intelligent heating memristors.Scale bar, 1 cm.Reproduced from [30].CC BY 4.0.

Figure 11 .
Figure 11.Stochastic neural network based on artificial neuron.(a) Schematic of the random motions of ions, which is the base cause of stochasticity in neuron devices.(b) Predictable randomness of neurons.Average Vout as a function of V in with a sigmoidal fitting curve and the inset shows a circuit diagram of the volatile memristor based p-bit unit.(c) Schematic of forward and inverted half-adder operations achieved by stochastic neurons.Reproduced from [101].CC BY 4.0.(d) The set process under different V TE of the tin oxide/MoS2 heteromemristors.(e) Schematic of the BM circuit blocks with tin oxide/MoS2 heteromemristors as the artificial neurons.(f) Experimentally obtained energy evolution in the BM optimization process for the four different strategies.Reproduced from [32].CC BY 4.0.

Figure 12 .
Figure 12.Artificial perception systems based on optoelectronic neural devices.(a) The schematic diagram of AOHN consisting of a resistive layer between Ag/ITO electrode.(b) Schematic of the artificial perception system: the real-time optical photographs of robotic hand during the dynamic training process.Reprinted from [31], Copyright (2022), with permission from Elsevier.(c)Visual perception of a nervous memristor-based system imitating the human visual system.(d) Schematic of the artificial perception system: Left is the flowchart of the meeting control system operating process; Right is the evolution of parameters in the meeting control system during the meeting process.Reprinted with permission from[33].Copyright (2021) American Chemical Society.

Figure 13 .
Figure 13.Sensory neural arrays based on 2D materials.(a) Schematic of large-scale optical sensing array, consisting of sensory neurons based on a wafer-scale integrated 2D materials.(b) Device structure of an oxidized MXene-based artificial optoelectronic memristor (OM-AOM).(c) The output trajectory processed by the OM-AOM array.© [2023] IEEE.Reprinted, with permission, from [48].(d) The output current maps of graded neuron for faster motion with a speed of 5 ms −1 for (i) Vg = 0 and (ii) Vg = +3 V.The optoelectronic graded neuron array can perceive the fast-moving scene by increasing Vg from 0 to +3 V.Reproduced from[79], with permission from Springer Nature.(e) An artificial tactile nerve, where external pressures applied to the e-skin change the resistance of MXene in the flexible pressure sensor.The ADC-LED circuit, consisting of a ring oscillator, an edge detector and an LED, receives the voltage signal from the MXene sensor and initiates optical spikes with coded pressure information.The optical spikes are transmitted to a synaptic photomemristor (OE synapse), which integrates and processes the spikes into a PSC to decode and memorize the pressure information.(f) Input-output of the tactile system showing the correlation among pressure, ADC outputs and post-synaptic currents (PSCs).Reproduced from[111].CC BY 4.0.

Figure 14 .
Figure 14.A pyramid-shaped schematic of critical aspects to further develop artificial neuron devices and neuromorphic systems.Bottom is in term of materials including taking full advantage of 2D materials' characteristics, such as sensory functions (left) and dangling-bond-free (middle), and developing large-scale synthesis and integration method with BEOL of CMOS (right).Middle is in term of neuron devices, pursuing more biomimetic functions (left) as well as simplify device's structure (right).Top is in term of neuromorphic systems based on neurons, with more algorithm and architecture to be explored.

Table 1 .
Comparison of the device characteristics of 2D materials-based neurons.

Table 2 .
Comparison of high resistance state (HRS) of artificial neurons with different memristive interlayer.