Advances in memristor based artificial neuron fabrication-materials, models, and applications

Spiking neural network (SNN), widely known as the third-generation neural network, has been frequently investigated due to its excellent spatiotemporal information processing capability, high biological plausibility, and low energy consumption characteristics. Analogous to the working mechanism of human brain, the SNN system transmits information through the spiking action of neurons. Therefore, artificial neurons are critical building blocks for constructing SNN in hardware. Memristors are drawing growing attention due to low consumption, high speed, and nonlinearity characteristics, which are recently introduced to mimic the functions of biological neurons. Researchers have proposed multifarious memristive materials including organic materials, inorganic materials, or even two-dimensional materials. Taking advantage of the unique electrical behavior of these materials, several neuron models are successfully implemented, such as Hodgkin–Huxley model, leaky integrate-and-fire model and integrate-and-fire model. In this review, the recent reports of artificial neurons based on memristive devices are discussed. In addition, we highlight the models and applications through combining artificial neuronal devices with sensors or other electronic devices. Finally, the future challenges and outlooks of memristor-based artificial neurons are discussed, and the development of hardware implementation of brain-like intelligence system based on SNN is also prospected.

Original content from this work may be used under the terms of the Creative Commons Attribution 4.0 licence.Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.

Introduction
Artificial intelligence and semiconductor integrated circuit are regarded as representative of the extreme technologies, whose manufacturing processes are one of the most cuttingedge science.Recently, the attractive ChatGPT has elevated the significance of artificial intelligence and big-data analysis to a new level.The data-intensive tasks depend on the high-speed central processing unit (CPU) and the compatible large-capacity memory technology [1].However, the power of algorithms that carries by the integrated circuit hardware is critically restricted by the separated architecture of the processing unit and memory unit, namely the Von Neumann bottleneck [2][3][4].In addition, the widely known 'memory wall' issue caused by the processor/memory performance gap also retards the overall performance of the computing system [5][6][7].Researchers have developed high-bandwidth memory and hybrid memory cubes technologies to improve the data throughput of memories and minimize the physical distance from the processing unit [8,9].However, the pursuit of the performance matching between the processing unit and memory unit has never stopped.
Actually, the brain-like computing is an important source of inspiration and a core objective of artificial intelligence, which has the potential to completely break through the traditional computing architecture [10].The human brain contains ≈10 11  neurons and 10 15 synapses, and can synchronously compute and store information through the synergistic effect between neurons and synapses [11][12][13].The neurons usually accomplish the computing function by integrating the inputted signals from pre-neurons, while the synapse is the memory unit, the weight of which could be continuously adjusted with the spiking signals from the pre-neurons or post-neurons (synaptic plasticity) [14,15].Especially, the power dissipation of human brain is only approximately 20 watts.The component that models the connection and information transmission between biological neurons is the artificial synapse.Artificial synapses are employed in artificial neural networks to communicate the output of one neuron to another.Artificial synapses typically contain a weight that represents the effectiveness of information transmission between two neurons.To improve the network's performance, the weight can be modified during the learning process.The artificial neuron, on the other hand, receives input signals, does weighted summation, and then processes the output using an activation function.In most cases, artificial neurons have several inputs and just one output.And the output of an artificial neuron is frequently activated to produce a nonlinear response, allowing the neuromorphic network to better adapt to complex data patterns.Brain-like computing system could mimic the working mechanism of human brain, implementing fast, reliable and low-cost computing techniques, which is therefore expected to become the next-generation computing architecture.
The most promising architecture for implementing the brain-like computing system is the spiking neural network (SNN), which is widely known as the third generation of neural network [16,17].This SNN architecture possesses excellent spatiotemporal information processing capability and high biological plausibility.Further, the SNN system transmits information through the spiking action of neurons, guaranteeing that this network architecture has a fundamental advantage of low energy consumption, which is just appropriate to realize high-efficiency brain-like computing.To develop a SNN through semiconductor integrated circuit technology, both of the artificial neuron and synapse need to be prepared, responsible for the computing part and storage part, respectively.Up to now, almost all of the typical biological functions of synapses have been well imitated using various electron devices, or complementary metal-oxide-semiconductor (CMOS) circuits, which have improved the hardware development of brain-like intelligence [18].While for the biological neurons, multifarious technologies have also been utilized to implement them, such as CMOS integrated circuits, ferroelectric devices and phase change devices [19][20][21][22].Although these technologies have demonstrated significant improvements of artificial neurons accomplishments, it is still essential to study their miniaturization and functional integrating capacities.
A biological neuron mainly consists of a cell body, axon, and dendrites, which is the smallest information processing unit [23,24].The neuron can efficiently integrate the signals from the pre-neurons through dendrites, and then transmit the spiking pulse signals to post-neurons.The cell body determines the specific response in the manner of excitatory or inhibitory potential signals [25].Memristors, possessing the advantages of simple structure, low power consumption, CMOS processing compatibility and gradually changing conductance, are frequently introduced to develop artificial neurons, which could greatly simplify the complexity of the circuit system [26][27][28].And the memristor utilizes its own nonlinear characteristics and harnesses the complex resistive switching (RS) properties to achieve dynamic simulation of complex artificial neurons.Yang et al showed that the neuromorphic systems have great potential in achieving new paradigms of computation, and there is a need for memristors with rich dynamics and nonlinearity to support these systems [29].Kumar et al reported that the material properties decide different orders of complexity in memristor devices and artificial neuron systems [30].The first-order memristor exhibits volatile behavior in its electrical characteristics.Several RS mechanisms can lead to volatile behavior in devices, such as the Mott transition, volatile defect migration, and tunneling.The threshold switching (TS) in TiO 2 is modeled as a firstorder process with the internal temperature as the state variable, utilizing Joule heating and the nonlinearity of electronics to achieve volatile RS characteristics [31].By connecting a volatile TS in parallel with a capacitor, a second-order memristor for neuronal mimic can be realized.The oscillatory behavior of NbO x is a second-order process with the charging and discharging of internal temperature and capacitance as the state variables, enabling volatile RS and self-oscillation behavior [32].Through careful design and the stoichiometric ratio of NbO x , a third-order memristor for neuron mimic is achieved, incorporating the internal temperature, capacitance, and dynamic Mott transition as variables [33].A single device realizes volatile RS behavior and dynamic behaviors resembling 15 types of neurons.Then we discuss in detail the memristor-based artificial neurons from three perspectives (Materials-Models-Applications).As shown in figure 1, various materials can be utilized to prepare memristive devices for mimicking biological neurons, such as transition metal oxides [34], ferromagnetic materials [35], organic polymers materials [36] and two-dimensional (2D) materials [37].Taking advantage of the interesting electrical behavior based on these materials system, such as analog switching, mott transition, TS and polarization process, several artificial neuron models and functions could be implemented including Hodgkin-Huxley (HH) model [38,39], leaky integrate-and-fire (LIF) [36,40,41] model and integrate-and-fire (IF) model [42,43].In addition, various sensory functions including tactile perception, optical perception, motion perception and even brain-machine interface application could be successfully developed through combining artificial neuronal devices with various kinds of sensors [44][45][46][47].
In the past few years, several kinds of artificial neurons constructed with memristive devices have been reported, which have significantly improved the development of artificial neural networks in the field of artificial intelligence [48].Furthermore, there are several review papers about the research progress on memristor-based artificial neurons.Choi et al summarized recent developments in memristor-based artificial neurons, and highlighted the hardware implementation strategies [49].Lee et al have published a comprehensive review paper about the development of synaptic and neuronal electronics devices based on 2D materials [50].More recently, Li et al reviewed the memristor based artificial neurons and their function expansion in neuromorphic sensing and computing and sensing [51].These literature reviews mainly focus on the neuronal materials, devices and neuromorphic computing system.Different from them, we emphatically present the potential applications of the artificial neurons except from neuronal materials, devices, mechanism, and models, aiming at giving some inspirations for future research on brain-like intelligence.Firstly, we emphatically introduce various materials that can be used to fabricate memristive neurons.Subsequently, we summarize the physical mechanism of memristive devices for implementing artificial neurons.Next, we focus on the neuronal models through taking advantage of the RS nonlinear characteristics and some simple circuits.Then, some fascinating reports of the multisensory neuron and the brain-machine interface based on the memristive neurons remarkable progresses of memristive neurons for multisensory neuron and brain-machine interface are presented.Finally, the remaining challenges about the implementation of memristor based artificial neurons are summarized, and the development of the further neuromorphic computing and brain-like intelligent system is also outlooked.
Artificial neurons serve as the foundation for deep learning models and are used for training and analyzing large-scale complicated data in machine learning and deep learning.The use of artificial neurons allows computers to better interpret and analyze human language, propelling technological developments such as voice assistants, intelligent translation, and automatic text synthesis.And artificial neurons are vital in the development of autonomous systems and robots.Robots can sense and process information from their surroundings more naturally by imitating the activities of biological neurons and neuromorphic networks, allowing them to perform more complicated actions and decision-making.Furthermore, artificial neurons have a wide range of applications in medical image processing, disease prediction and diagnosis, and other domains.Their importance in better analyzing and interpreting medical data, boosting diagnosis accuracy, and improving treatment effectiveness will be critical in the future for the medical and healthcare industries.

Materials and mechanism
Biological neurons receive signals from other neurons via the dendrites, and transmit electrical signals to other neurons via the axon [54][55][56].The action potential and the resting potential of the neuron represent the firing and rest states of the neuronal signal, respectively [57].Biological neurons play a fundamental role in mediating brain activity [58,59].Circuits based on CMOS or transistors are typically used to stimulate biological neurons, and these implementations require many active components to achieve adequate functionality of artificial neurons [60][61][62].Similarly, Memristive neurons have attracted substantial interest because of their peculiar characteristics, such as their implementation using Mott memristors, phase-change memory (PCM), 2D materials, and magnetic random access memory [63][64][65].In the following sections, we will review ongoing research on advanced memristors for the implementation of artificial neurons with a focus mainly on materials, RS mechanisms, models, and applications.
The RS performance based on conductive filaments (CFs) represents one of the important memristor mechanisms [66][67][68][69][70], as shown in figure 2(a).Two-dimensional materials have been reported to exhibit phase transition and TS behavior.However, the existing literature on SNNs exclusively reports on threshold RS devices based on Ag-based two-dimensional materials.Therefore, we will take Ag-based two-dimensional materials as a starting point to investigate the RS mechanism induced by Ag-based electrochemical switching.Generally, Ag is often used for the electrochemical reaction process.The RS of an electrochemical-type memristor often involves three steps: (1) anodic oxidization; (2) migration; (3) cathodic reduction.When a positive voltage is applied, Ag undergoes oxidation at the positive electrode, and Ag ions migrate under the influence of the electric field towards the negative electrode, where they are reduced back to Ag atoms, forming a conductive filament.Conversely, when a negative voltage is applied, the conductive filament will rupture.The RS performance based on CFs can also be influenced by inherent properties of the memristive materials, such as spontaneous variations of CF properties resulting from the Joule heat effect and minimum energy positions [71][72][73][74][75]. Inherent properties, such as thickness, crystallinity, and component structures, combined with operating environments can be used to control the position and configuration of CFs.The instability of the CFs leads to the volatile characteristic of the device.Most memristive devices used to build artificial neurons are volatile.The development of memristors based on 2D materials has shown great significance.The high performance of high on/off ratio, low SET voltages (V SET ), fast switching speed, and low power and energy consumption has been achieved separately in different memristors based on 2D materials.More importantly, 2D materials exhibit potential ability to scale down in size owing to their ultrathin 2D layered structure, which fails to be achieved in traditional metal oxide thin films.When the thickness of traditional metal oxide thin films is decreased to a nanoscale, leakage dramatically increases.Further, the RS mechanism of Ag-based memristors is the oxidation and reduction process of Ag atoms and Ag + ions.Therefore, from the perspective of mechanism, the near-atomic scale film thickness greatly reduces the barriers to Ag ions migration.Moreover, the photoelectric properties of Ag-based 2D materials are of great significance for the development of optical vision neurons.Therefore, 2D materials based electrochemical metallization (ECM) architecture is expected to implement high-performance artificial neurons, which possess a variety of outstanding characteristics, such as high energy efficiency, high integration capability and excellent scalability.Artificial neurons based on memristors from 2D materials can be realized through the formation and rupture of CFs [37].The Ag ions are initially produced under positive electrical stimulation.Then, the Ag ions move towards the bottom electrode and are reduced.The filaments can undergo spontaneous rupture resulting from the Joule heat effect and minimum energy positions, as mentioned above.This occurs when the applied voltage is not strong enough.Therefore, artificial neurons based on 2D materials offer a simple solution for constructing efficient neuromorphic computing devices for SNN.
The action and resting potential of biological neurons are coordinated by potassium (K + ) and sodium (Na + ) channels on both sides of the cell membrane [76][77][78].Mott memristors are electronic analogues of the ion channels found in biological neurons [34,53].Oxide materials such as NbO 2 or VO 2 exhibit insulating properties, where electrons are in a localized state, resulting in a high-resistance state at low temperatures.However, these materials experience a Mott transition at high temperatures, where electrons move from localized to delocalized states, greatly increasing electrical conductivity and resulting in a low-resistance state.The intricate interactions between electrons inside the lattice structure and electron-electron interactions are thought to be the cause of the Mott transition in these materials.This transition, which is extremely significant, could lead to the use of Mott transition materials in industries including sensors, switch devices, and memory devices.As shown in figure 2(b), Kumar et al showed that the artificial properties mimicked by Mott memristors possess most of the known neuronal dynamics associated with biological tissues [33].Also, Pickett et al reported that Mott memristors with a simple structure exhibit important neuron models [53].In these reports, the Mott phase transition is driven by external voltages in insulator-to-metal transition materials.Mott two-terminal memristors based on NbO 2 and VO 2 materials frequently exhibit current-controlled negative differential resistance (NDR) behavior [79][80][81][82][83].This behavior is briefly described as follows: the device undergoes RS when sufficient current is driven through the device to locally heat some of the material under positive voltage applied.The memristive device will undergo phase transition once the Joule heat exceeds the transition temperature, thus producing a conductive channel through the device that bridges the two Reproduced from [35], with permission from Springer Nature.(d) Top: Phase-change materials.Bottom: Phase change mechanism (Joule heating).Reproduced from [41], with permission from Springer Nature.Reproduced from [52], with permission from Springer Nature.electrodes.In this manner, memristors can be switched from a high resistance state (HRS) to a low resistance state (LRS).Subsequent to this fast switching event, the resistance gradually reverts to HRS with decreasing sweeping voltage.The functionally dynamical resistance behavior of the Mott transition is very similar to that of the HH ion channels.Simulation of this physiologically realistic behavior has also been investigated intensively as a promising research field for future integration into neural computing.
Ferroelectric materials are regarded as those carrying the most potential for memristor development, as they can perform a wide range of artificial neuronal models [84,85].The phase transition that occurs in ferroelectric materials is related to their unique crystal structure and charge arrangement.Ferroelectric materials' positive and negative charges are positioned in a particular way in the crystal structure, creating a polarized electric dipole moment.As a result, ferroelectric materials display phase transition properties, undergoing polarization reversal in the presence of an external electric field.In particular, ferroelectric materials often have non-centrosymmetric crystal structures, in which the arrangement of cations and anions is asymmetrical.Due to this noncentrosymmetricity, the material has an internal electric dipole moment that is stable in one direction even in the absence of an external electric field.However, if this ferroelectric material is exposed to an external electric field, it may alter the direction of the electric dipole moment, reorganizing the positive and negative charges and leading to polarization reversal.Ferroelectric materials can undergo a phase transition from a polarized state to a non-polarized state under the direction of an electric field thanks to this polarization reversal, which causes substantial changes in their physical properties at the phase transition point.Recent experiments on switching of ferroelectric materials have demonstrated many of their advantages, such as low-energy consumption, high speed, and nonvolatile nature [86][87][88].Ferroelectric materials such as BiFeO 3 and Hf 0.2 Zr 0.8 O 2 have been developed for mimicking artificial neurons.Cao et al harnessed the dynamic characteristics of anti-ferroelectric field-effect transistors (AFeFET) built using Hf 0.2 Zr 0.8 O 2 materials to realize artificial neurons, as shown in figure 2(c) [35].The operation of triggering a transformation between anti-ferroelectric (AFe) and ferroelectric (Fe) domains serves to model a property of artificial neurons called 'leaky integrate and fire'.First, the FE orthorhombic phase (ophase) domains nucleate with the applied continuous-voltage pulse stimuli, leading to the transformation from AFe tetragonal phase (t-phase) domains.Consequently, the polarized charges of the AFe layer unit gradually accumulate.With further application of gate pulses, the Fe o-phase domains induced by the electric field grow, and the overall output current suddenly increases.Under these conditions, artificial neurons generate an output spike which may be regarded as 'fire' behavior.After firing, the Fe (o-phase) domains induced by the electric field transform back to AFe (t-phase) domains without gate pulses.Over a short period of time, the AFeFET layer will spontaneously return to HRS as a consequence of the phase change, reverting the neuron to its resting potential.The switching of a ferroelectric material using voltagedriven mechanisms confers significant savings in terms of switching energy requirements [89][90][91].At the same time, the key functions of biological neurons, such as all-or-nothing spiking of action potentials, the existence of a bifurcation threshold to a continuous spiking regime, signal gain, and the presence of a refractory period, are all successfully mimicked.
Ferroelectric materials have a key place in electronics and memory technology because of this special phase transition process.Ferroelectric materials' phase transition behavior can be changed by manipulating external electric fields, making it possible to perform very efficient logical and storage operations.Ferroelectric materials are also widely used in sensors, radio equipment, acoustics, and other industries.In order to advance cutting-edge technologies and gadgets in the electronics industry, these materials provide great potential for breakthroughs in a variety of electrical and electromechanical applications.
PCM materials are regarded as promising materials for data-storage applications and field effect transistors, which are used in rewriteable data storage and multilevel data storage [92][93][94][95][96]. PCM materials are examples of emerging nonvolatile electronic memory devices, and are frequently used to mimic artificial neuron [97,98].Phase change materials can undergo significant changes in their physical properties from one phase to another phase at specific temperatures, pressures, or other external conditions.The internal structure and properties of the material are rearranged throughout this phase transition process, resulting in noticeable changes to its macroscopic properties.Phase transitions, which alter the crystalline structure of phase change materials, can occur in solid-state materials.Energy absorption or release during these phase transitions frequently results in changes in the material's temperature.Phase change materials display a variety of traits, such as temperature sensitivity, the ability to absorb or release energy, and great sensitivity to environmental factors.Due to their special characteristics, they are useful in a wide range of applications, including temperature control, sensors, and memory devices, where their capacity to flip between several phases enables effective energy storage and management.As shown in figure 2(d), Sung et al developed a PCM device to emulate intrinsic synaptic mechanisms [41].Tuma et al have demonstrated a stochastic phase-change neuron.PCM materials exist in both amorphous and crystalline phases, which can alternate to simulate different RS processes [52].The principle of phase change is as follows: PCM devices switch from the amorphous phase to the partial crystalline region as a consequence of Joule heating produced by the application of sufficiently high voltage pulses.After this event, the crystalline region will revert to the amorphous phase via a glass transition once the pulse is stopped.This continuous phase change is analogous to changes in neuronal membrane potential.The neuronal membrane potential is kept in the phase configuration in a phase-change neuron.Utilizing crystal growth dynamics, it is possible to program cells into various intermediate amorphous/crystalline structures, simulating the development of the neuronal membrane potential.Therefore, phase change materials are one of the important materials to realize artificial neurons.
Additional mechanisms and applications have been proposed for oxide-based and organic-based memristors, such as the formation of valence change, charge trapping/detrapping, or the modulation of the interfacial barrier.Detailed descriptions of these mechanisms have already been presented by several prior reports [99][100][101].These physical mechanisms have also been proposed to explain the switching behavior of memristors, providing several methods for mimicking artificial neurons.Neuronal mechanisms can be simulated using various models, such as H-H, Morris Lecar [102][103][104], Izhikevich [105,106], LIF, and IF.We provide a detailed summary of neuronal models and working mechanisms from representative studies in section 3.

Artificial neuron functional models
The LIF/IF model and the H-H model are both mathematical models used to describe the activity of neurons.Their relationship with neural functionality lies in the fact that through these models, we can better understand and study the neuronal activity patterns under different conditions, including the firing behavior of neurons, the formation and propagation of action potentials, and the neuronal response to different input signals.These models are of significant importance for research and applications in various fields such as optical perception neurons, tactile perception neurons, and brain-machine interfaces.
The H-H model was proposed in 1952 by Alan Hodgkin and Andrew Huxley [107][108][109].It represents a seminal contribution for explaining the generation and propagation of action potentials [110][111][112][113].The H-H model is one of the most important and most realistic models in neuroscience [114].This model captures the mechanisms of voltage-gated Na + and K + ion channels as they mediate current flow across the cell membrane, accounting for action potential generation [39,53].The basic circuit topology of a two-channel active memristor neuron can be used to emulate the neuronal dynamics associated with this model.Pickett et al developed a Mott memristor with composition, structure, and mechanism similar to an artificial neuron (figure 3(a)) [53].The H-H model circuit was introduced using the traditional connected way.In this approach, a voltage-gated Na + channel is emulated via a negatively voltage-biased memristor device.The memristor is then connected in parallel with the film capacitor C 1 (C 2 ), and in series with the load resistor R L1 (R L2 ).Therefore, the voltage polarity in the two channels is opposite, similar to the Na + and K + channels in the H-H model.Using a related approach, Yi et al reported the neurons built with nanoscale VO 2 active memristors possess all three classes of excitability and most known biological neuronal dynamics (figure 3(b)) [39].Furthermore, their implemented neurons are intrinsically stochastic.Understanding the nervous system's computational capacity and its competitive advantage in obtaining extraordinarily high neuronal densities on biological substrates is largely dependent on the H-H model for action potential generation in biological axons.
The LIF/IF neuron model supports bio-plausible neuronal activity, alongside a biophysical explanation for this electrical activity [42,43,[115][116][117][118].A simple device or circuit with threshold properties is generally sufficient to implement IF ch as resting potential, depolarization, hyperpolarization, and repolarization [119,120].Three different connected circuits can be designed to fully mimic the behavior of biologicand LIF neurons.It is well-known that the electrical activity of biological neurons is multifaceted, spanning various modes sual neurons.We discuss each one separately below.
The first example of LIF model circuit involves volatile memristors and comparators.Yang et al reported a secondorder memristor for reproducing the integrated functionality of artificial LIF neurons [40].As shown in figures 4(a), a digital comparator was used to set a threshold value.The detailed operation of this neuronal circuit involves several processes.According to the relative distribution of synaptic weights, input pulses integrate both temporally and spatially during the integration process.When V out is lower than the reference voltage (V ref ), the divided voltage gradually leaks out throughout the leak process.V out exceeds V R during the firing process, causing the pulse generator P to produce an action potential (fire spiking).Second-order memristors have a number of benefits, but their capacity to support two-state variables that capture both long-term and short-term dynamics is their most important benefit.This enables second-order memristors to function in a manner that is similar to how biological systems function.More recently (in 2020), Yu et al introduced a nanoscale third-order circuit element, including the speed of Mott transition, temperature, and charge on an internal capacitor [34].Their device involves a simple network without transistors, designed to act as an artificial neuron, and constructed using third-order elements that can perform logical operations.This work paves the way towards very compact and densely functional neuromorphic computing primitives, which also provide energy-efficient validation of neuroscientific models.
The second example involves the controlled firing response of LIF neurons using different circuit parameters.Sung et al described an RC circuit that uses electrons, a parallel capacitor, and a bottom TS layer, which are each analogous to a biological system's voltage-gated ion channel, lipid bilayer, and charged ions [41].The capacitor separates charges and creates a potential difference, as seen in figure 4(c), which is comparable to the integration process provided by artificial neurons.The TS memristive device modulates sudden current increases associated with capacitor discharges generated when the voltage reaches the threshold value (V th ), thus simulating a firing process.It is worth noting that the charging and discharging behavior of the parallel capacitor can be used to simulate artificial neurons.Furthermore, each small (micro-nano scale) artificial neuron device constructed around multi-memristor coupling effects can completely mimic the electrical activity of biological neurons [121][122][123].
The third example involves a neuron model based on single integrated element.Lashkare et al exhibited a PCMO memristive device for IF neuron, as shown in figure 4(d).And the PCMO based neuron in SNN yields software-equivalent classification accuracy [42,43].Also, recently, Bian et al introduced a stacked memristive device supporting both analog and TS behaviors for artificial LIF neurons (figure 4(b)) [36].This device represents the experimental realization of a LIF neuron based on a single Au/WO x /W/Ag doped ι-carrageenan/Pt memristor.
This single nanoscale device performed both the integration and fire functions.All-or-nothing spiking, threshold spiking, a refractory period, and strength-modulated frequency response, among other essential neuronic functions of a biological neuron, were also successfully imitated.In these work, one single memristor was used to accurately control the emulation of the LIF/IF neuron, demonstrating great potential for constructing high-density and low power consumption SNN for neuromorphic computing.(Table 1 gives an overview of the neuronal models with different memristive materials.)

Tactile perception
A nerve cell known as a perception neuron is one that recognizes and reacts to outside signals [127,128].The peripheral nervous system's receptors allow perception neurons to receive information, which they then transform into electrical impulses [129,130].These electrical signals mediate the transmission of sensory information to perceptual processes, such as temperature sensing, light detection, and other perception functions.Artificial perception systems consist of sensory neurons that can perceive temperature, pressure, and other sensory signals, which is very important for mimicking human sensory behavior [131][132][133][134][135]. Therefore, perception neurons carry great potential for exploring the external environment [136][137][138].
In artificial perception systems, tactile perception is one of the important perceptual functions.For instance, Zhang et al developed an artificial spiking mechanoreceptor system (ASMS) based on Mott memristors that detects mechanical stimuli and generates appropriate responses [44].The tactile perception system of biological neurons is equipped with a natural protective mechanism for avoiding damage from excessive external stimulation.In figure 5, NbO x -based memristors are chosen as a part of artificial synapses for signal input.NbO x -based memristors are highly integrated dual-terminal devices that exhibit temperature-driven NDR behavior, serving as the foundation for voltage-sweeping  dynamic TS.The NDR behavior of the memristor is primarily temperature-dependent, meaning that the higher the current flowing through the circuit, the greater the Joule heat generated by the memristor.As a result, the relaxation time of the entire system becomes longer, thereby dominating the oscillation period.Furthermore, utilizing NbO x devices with lower threshold voltages, smaller windows, and smaller parasitic capacitance in the testing circuit can further reduce the energy consumption of the circuit.Figure 5(a) shows a self-powered ASMS that contains a pressure sensor, resistor, capacitor, and memristors.The pressure sensor is a piezoelectric device that can generate electricity in response to pressure or mechanical stress signals.This property relies on changes in electric polarization caused by the deformation of atomic structures.When harmless stimuli are applied, the amount of generated electricity is proportional to the intensity of the applied pressure, so this quantity can be encoded in the form of signal impulses produced by the pressure sensor within the ASMS.At the same time, the firing frequency of the neurons can be adjusted to reflect the intensity of harmless stimuli.However, the neuron will stop firing to protect the ASMS from injury when noxious stimuli are applied.Figures 5(b) and (c) show experimental results from the artificial mechanoreceptor system.A sinusoid-like voltage signal was generated in response to the pressure signal input, which was then converted into output pulse spiking.Neuronal firing rate increases with increasing pressure signal until the system engages the inhibitory mechanism just described (figure 5(d)).Therein, figures 5(c)-(f) magnifies specific sections of figure 5(b) to emphasize relevant features of this dataset.The artificial perception system can also perceive pressure, and recognize the structure of the external environment [139][140][141][142][143]. Flexible wearable devices are important for various applications across diverse areas such as sport, health, and medical treatment, and are currently the focus of international research [144][145][146].In the area of human movement monitoring, the main deformation modes associated with skin surfaces are bending deformation and tensile/compressive deformation [147][148][149][150]. Figure 6 reports an artificial sensory neuron based on pulsed laser deposition-grown epitaxial VO 2 memristors.In constructing complex artificial sensory neuron systems, stability and repeatability of the memristors are essential.Therefore, the use of epitaxial-grown VO 2 significantly enhances the stability of the memristors and the uniformity of the fire operation.The epitaxial VO 2 memristor is connected in parallel with a capacitor, and this structure is subsequently connected in series with a load resistor, R L , forming the artificial neuron device.Depending on the frequency and magnitude of the input signal, the oscillation frequency of the output neuron spikes can be altered.A stable artificial neuron system is crucial for the integration of subsequent sensory sensors.Yuan et al developed a spike-based neuromorphic perception system that contains a curvature sensor and RC circuit for monitoring bending deformations (curvature and bending angle), as shown in figures 6(a)-(c) [45].The resistance of the curvature sensor increases with increasing curvature.As a result, the divided voltage of RC circuit is decreased and neuronal firing frequency is reduced.A system for experimentally encoding hand gestures into various spike frequencies is known as a spike-based neuromorphic perception system.Depending on how each finger bends, different neurons fire at different rates.This is achieved by attaching a curvature sensor to each finger, thus supporting the recognition of different gestures.This application demonstrates that the pulse-based neuromorphic gesture perception system can effectively recognize gestures.Therefore, artificial sensory neurons are expected to play an important role in promoting the development of neurorobotics, perception, and neuromorphic computing.
In human biological systems, the perception nervous system provides a connection between the external environment  stimuli.The ability to differentiate between different types or levels of sensory inputs is crucial for accurate and reliable detection and interpretation.(2) Scalability and integration: scaling up memristive-based sensory systems to large arrays or networks while maintaining their performance and functionality can be challenging.The integration of multiple sensors and the management of interconnectivity between them also pose difficulties.(3) Standardization and compatibility: establishing standardized protocols and interfaces for memristivebased sensory systems is necessary for interoperability with other devices, systems, or platforms.Ensuring compatibility and seamless integration can facilitate broader adoption and utilization.Therefore, continued advancements in device design, fabrication techniques and system integration will contribute to improving the performance, reliability, and capabilities of memristive-based sensory systems.

Optical perception
The visual cortex processes information, which is subsequently transported to other parts of the brain for further analysis and utilization [152,153].The visual cortex's neurons are critical for executing the early processing of visual information [154][155][156][157]. Wang et al reported a lobula gigantic movement detector (LGMD) neuron that can aid in robot navigation.The artificial LGMD neuron may be made using the simple circuit depicted in figure 7(a) given the optically modulated TS behavior of FLBP-CsPbBr 3 [46].The TSM is a dual-terminal volatile memristor switch with a stack structure of Ag/FLBP-CsPbBr 3 heterojunction/ITO.(1) During the voltage scanning process, the active metal electrode material, Ag, can be doped into the FLBP-CsPbBr 3 layer.The formation of CFs in the memristor is a 'threshold' event determined by the net Ag content within a given filament volume, leading to the formation of stable threshold behavior.(2) Furthermore, due to the optoelectronic coupling and charge transfer in the FLBP-CsPbBr 3 heterojunction, the TS performance can be effectively controlled by optical input.Therefore, the formation and rupture of Ag CFs exhibit a non-monotonic response to the light flux field, similar to the escape response of the LGMD neuron.By utilizing optically mediated threshold switch memristors, artificial LGMD visual neurons can be realized.In this circuit, the TSM device is in series with a resistor, and the two components are in parallel with a capacitor.Under the application of continuous voltage pulses, the voltage on the capacitor was recorded as output voltage.The current leakage through the TSM is negligible, and the capacitor is charged (figure 7(b)).By examining the artificial neuron's charge loop and discharging loop individually, it is possible to understand how spike signals occur.Because the voltage predominantly declines across the TSM device due to the voltage dividing effect, the Cp can be initially charged via the charging loop.The TSM shifts from HRS to LRS (R on ∼ 1kΩ) when the voltage of C p exceeds its threshold, and the capacitor discharges via the discharging loop, resulting in the artificial neuron firing.R and C have a significant impact on the firing rate, as demonstrated in figures 7(c) and (d).A larger R decreases input current, slows the buildup of charge, and delays the firing event while a lower C speeds up the integration process.Different R 1 circuit configurations can be implemented to construct artificial LGMD neurons.When the overall output current increases suddenly, the artificial neuron would generate an output spike, which could be called 'fire' behavior.The TSM device does not revert back to the initial state immediately.Rather, the voltage slowly decreases though pulses applied.The duration of this relatively slow decrease in current signal may be regarded as the refractory period (figure 7(e)).The LGMD neuron device can respond to optical input and encode the associated information into spikes.Light-mediated spike behavior was measured by applying a fixed pulse train under UV-light irradiance at different power levels.As shown in figure 7(f), the spiking frequency strongly depends on light intensity.Figure 7(h) displays the relationship between UV irradiation power and TSM conductance.The spiking frequency follows a trend that mirrors the increase in UV irradiance (figure 7(g)).The spiking frequency decreases as a function of light wavelength.The pulse control system that implements the collision avoidance function of the robotic car demonstrates the potential of integrating artificial LGMD neurons into the circuit.The decision process of the robot car along the motion trajectory is shown in figures 7(i) and (j). Figure 7(k) also illustrates the behavior of the robot car as it approaches an external stimulus.The firing frequency of the LGMD neurons increases as the robot automobile approaches the light source.When the pulse frequency reaches the highest frequency, the robot car turns away from the light source to avoid an impending collision.Depending on pulse frequency, the neuron can thus deliver different directional decisions in different situations.This principle of fully integrated TSM neurons can be extended to implement complex learning capabilities in systems that are also energy efficient.
Similarly, drawing inspiration from the locomotion of flying insects, the processing of dynamic visual information can be applied to various fields, including the Internet of Things, robotics, and automation.Recently, Chen et al have successfully developed a biomimetic insect visual system using a graded neural network [158].This achievement highlights the effectiveness of the biomimetic approach in improving the perception of motion in visual systems.The use of MoS 2 as a conductive channel in phototransistors, utilizing shallow charge trapping centers, allows for the simulation of graded neuron functionality in a biomimetic fruit fly model.This device exhibits time encoding characteristics similar to graded insect neurons.
Based on the temporal characteristics of biomimetic graded neurons, optoelectronic devices can effectively encode temporal information.By utilizing optical sensor arrays that can encode spatiotemporal visual information and display motion trajectory contours, dynamic movements can be perceived.The biomimetic sensor array outputs spatiotemporal motion information by nonlinearly combining inter-frame information, effectively displaying the entire motion trajectory contour from left to right.When the motion occurs in the opposite direction, the sensor array can perceive the reverse contour of the motion.Artificial graded neurons can also perceive approaching/retreating movements.After training, the accuracy of action recognition based on biomimetic visual sensors can reach 99.2%, which is significantly higher than the accuracy of traditional image sensors.This type of biomimetic visual sensor provides an effective solution for rapid action recognition with limited computational resources.

Brain-machine interface
Artificial neurons play an important part in BMI technology.Artificial neurons are the essential units of artificial neural networks, and successful algorithms and models are intended to analyze brain inputs and achieve BMI by imitating and comprehending the working principles of biological neurons.Artificial neural networks, in particular, can be used to decode brain signals, converting measured brain activities such as electroencephalography into specific commands or control signals, for example, to control the movement of robots or assist people with disabilities in regaining motor functions.The efficient computing of artificial neurons has aided in the achievement of highly individualized and precise brain-machine interfaces in BMI technology.
Biomaterial-based memristors (bio-memristors) are used in BMI to mimic the functionality of biological neurons, enabling direct interaction with the brain.By utilizing biological memristors as key components of artificial neurons, the following purposes can be achieved: Signal transmission and interpretation: bio-memristors can control the flow of current by adjusting their resistance, facilitating information transfer with neurons.They can interpret neural signals from the brain and convert them into understandable forms, allowing the BMI to comprehend and respond to brain activity.
Neural signal simulation and modulation: bio-memristors can emulate the electrical activity and signal transmission processes of neurons, generating electrical signals similar to neural signals in the brain.These signals can be transmitted to external electronic devices or mechanical devices, enabling the control of external equipment and facilitating actions such as muscle contraction and movement.
Bio-memristors are often adopted to emulate the biological functions of neurons, and are used in the construction of SNNs [47,[159][160][161].Artificial neurons built from bio-memristors not only operate on the scale of biological action potentials, but also exhibit temporal integration properties that are similar to those observed in biological neurons [162][163][164].In terms of programming voltage and current, bio-memristors function at power levels comparable to those of biological neurons.In order to establish a connection between biological neurons and artificial neurons, the implementation of bio-voltage is a crucial challenge.Biomaterials (both natural and synthetic) are used in medical applications to support, enhance, or replace damaged tissue or a biological function [165][166][167][168]. Fu et al introduced these protein nanowires into memristors to reduce the switching voltage [164].Protein nanowires are considered as the preferred material for achieving low-voltage operations, which significantly reduces the energy consumption associated with fire process similar to biological neurons.Protein nanowires are biological materials that present several advantages compared with other memristive nanowire materials [169][170][171][172]. First, protein nanowires are 'green' and sustainable electronic materials produced by microbes or fabricated in vitro with microbe-inspired designs.Second, they present non-uniform diameters with periodic surface structures along the main axis.Furthermore, they are characterized by high conductivity, and protein nanowires can support nucleation of metals.Therefore, it is a computer-based interlocking system that implements an optional man-machine interface via bio-memristors.Protein nanowire memristors possess integrated, threshold, and self-recovery functionalities, making them artificial neurons with biovoltage characteristics.Additionally, the inherent biocompatibility and biodegradability of protein nanowires offer possibilities for realizing BMI.This concept of voltage modulation optimization may have broader applications in memristor-based artificial neurons [173].As shown in figure 8(a), vertical memristors (Ag/SiO 2 /Pt) were embedded in protein nanowire film [47].The most important characteristic of these bio-memristors is their ultra-low setting voltage (70 ± 3 mV), which is within the range of biological action potentials (figure 8(b)).The protein nanowires also play a key role in supporting the RS mechanism.The associated switching process relies on a characteristic mechanism supported by CFs called ECM.The RS of ECM-based memristors can be divided into three stages: Ag oxidation, Ag + migration, and Ag + reduction.Ag + ions can be readily reduced to Ag nanoparticles by Geobacter sulfurreducens in the environment, indicating that protein nanowires can facilitate the Ag redox process [164].Therefore, protein nanowires may facilitate Ag metallization during filament formation and reduce V th .In addition, protein nanowires can be used as active sensors.Active sensing devices require an external source of power to operate.Breathing signals are used as an external source to stimulate the memristive device.Planar memristors with a protein nanowire film are deposited on a pair of interdigitated electrodes.A potential inconvenience of this approach is that quick changes of local humidity resulting from breathing-induced non-uniform moisture adsorption may cause instant electric spikes (figure 8(c)).The smaller power consumption makes it possible to drive neuromorphic circuits at small sensor sizes (<1 cm 2 ), a key ingredient for achieving the level of compact integration that is desirable for wearable devices and microsystems.Devices that rely on planar protein nanowires can produce continuous and spiking signals, enabling different functional interfaces.For example, optical stimulations have been successfully converted to active voltage signals that were able to drive the neuronal firing (figure 8(d)).These findings show the degree of adaptability that can be reached in constructing selfsustaining, multifunctional BMI that can perceive and process many stimuli.These totally environmental-driven interfaces are fundamentally different from prior systems, which either required external powering or limited to select stimulus types that matched the signals provided by the adopted memristors.
Protein nanowires are flexible fibrous structures that can be mimicked and fabricated on flexible substrates, making them suitable for the preparation of wearable devices.By utilizing action potential signals (e.g.50-120 mV) as input signals, biological computing can respond to a broader range of environmental stimuli.This necessitates the requirement of low operating voltage for protein nanowires.The use of protein nanowire-based memristors offers several advantages.Firstly, these memristors can be driven by signals lower than 100 mV, enabling the realization of biologically relevant neural morphological circuits and signal processing.Secondly, devices fabricated with protein nanowires can harvest electrical energy from ambient humidity, providing both signals and power for computation.Thirdly, protein nanowires can serve as sensing elements for electronic sensors.The material properties of protein nanowires, including renewability, biocompatibility, and eco-friendliness, form the foundation for unified sensory and computational functions within biological systems, making them key to advancing integrated neuromorphic interfaces.Figure 8(e) illustrates the development of an integrated skin-wearable system responsive to respiration.In this application, respiration was used as the spiking stimulus.The bioamplitude neuromorphic component was powered by a small sensor measuring 0.3 cm 2 .The sensor was linked to a protein nanowire memristor and a capacitor-based artificial neuron.The sensor and artificial neuron work together to form the front-end afferent circuit, which may process physiological data for additional processing and decision-making.To carry out front-end decisions, several back-end circuits or efferent controllers can be used.This neuromorphic interface is very similar to its biological counterparts, in which sensing and decision are achieved without an external power source.This brain-machine interface offers substantial improvements with relation to size and power consumption.This review highlights the role played by bio-memristors at the interface Reproduced from [47], with permission from Springer Nature.
between neuromorphic systems and neural systems with biological organization.
The BMI is a pathway for the interaction between the human brain/nervous system and external electronic devices.It serves as a cornerstone in various technological fields such as monitoring and decoding brain activities, treating neurological disorders, and developing intelligent prosthetics.The BMI relies on electrical signals to interpret brain signals, which are primarily derived from the membrane potential changes of stimulated neurons.In living organisms, communication and transmission of information between neurons primarily occur through neurotransmitters such as dopamine and acetylcholine.The human brain's functions, including decision-making and emotion regulation, are closely related to neurotransmitters.Additionally, neurons possess neuroplasticity, which means they can dynamically regulate the weights of their connections based on the strength of stimuli.This molecular basis of neuronal plasticity is considered essential for human cognitive behaviors such as memory and learning.Therefore, to effectively interact with neurons, ideal artificial neurons should also possess similar adaptive capabilities.
To achieve this, Wang et al have developed an artificial neuron based on chemically mediated neurotransmission to facilitate bidirectional interaction in BMI [174].This artificial neuron is capable of interpreting chemical information in neurotransmitters and adaptively quantifying the release of neurotransmitters, specifically dopamine.It consists of three components: a dopamine electrochemical sensor, a memristor for signal processing and emulating synaptic plasticity, and a dopamine release module, as shown in figure 9(a).
(1) Dopamine Electrochemical Sensor: the dopamine electrochemical sensor, utilizing the excellent conductivity of graphene and carbon nanotubes, employs a graphene/carbon nanotube composite as the sensing element.It offers high sensitivity, a wide detection range, as well as good stability and selectivity, meeting the requirements for detecting dopamine concentrations in the synaptic cleft.(2) The RS memristor based on the doping effect of Ag nanoparticles enables the synchronous realization of shortterm memory on the order of milliseconds and long-term memory exceeding 10 4 s.This implies that it provides a foundation for system adaptivity.(3) By utilizing the gel actuation controlled by the current output of the memristor, the dopamine release module can achieve adaptive and quantitative release of dopamine molecules within the temperature range of 37.5 • C-45 • C, employing the Joule heating effect combined with the temperature-sensitive properties of the gel.
As shown in figure 9(b), in the relevant chemical information flow, it primarily follows the following three steps: Firstly, the dopamine chemical signal is converted into a current signal.Secondly, with the aid of the current signal, the resistance state of the memristor is adaptively modulated.Thirdly, utilizing the Joule heating effect, the device stimulates the release of dopamine chemical signals, transmitting them to adjacent neurons, enabling bidirectional interaction of chemical information.
In the interface between artificial neurons and living neural cells in organisms, this chemically mediated artificial neuron is capable of successfully releasing dopamine and stimulating PC12 cells to open ion channels.In a biologically mixed interface, the artificial neuron acts as a 'chemical bridge' to transmit information to PC12 cells, achieving chemical communication similar to interneurons.To further validate the ability of artificial neurons to mimic human motor neurons in controlling muscle contraction, the artificial neurons were used to activate a robotic hand and the hind legs of mice.The results showed that when dopamine stimulation was applied, the robotic hand performed a gripping motion and triggered the bending of the hind legs in the live mice, as shown in figures 9(c)-(f).Therefore, BMI can bridge the gap between humans and machines by interpreting and transmitting neurophysiological information.This is a crucial process for neuronal rehabilitation, semi-robotic construction, and ultimately consciousness detection and control.

Challenges and outlooks
In this paper, we have comprehensively reviewed recent progress in memristor-based artificial neurons from materials to functional applications.Various materials (transition metal oxides, ferromagnetic materials, organic polymers materials and 2D materials) and the corresponding physical mechanism (analog switching, mott transition, TS and polarization process) that are utilized to prepare artificial neurons have been firstly discussed.Then, several artificial neuron models, including H-H model, LIF model and IF model, have been discussed, and their advantages and potential applications have been respectively analyzed in detail.Finally, we emphatically present the potential applications of the artificial neurons once combining various kinds of sensors, such as tactile perception, optical perception, motion perception and even BMI.Although memristor-based artificial neurons have significantly promoted the development of artificial neural network and even brain-like intelligence, it can be seen that these research and development works are still at the exploration stage and facing various challenges.
Memristive materials and devices: first, the requirements of artificial neurons for constructing brain-like neural network mainly include scalability, integration potential, high thermal/humidity stability, low energy consumption and ultrafast data processing speed.And the electrical properties of memristive devices depend on the corresponding materials characteristics.Transition metal oxide materials are widely studied for the reliable micro-structure, CMOS process compatibility and high-performance electrical parameters, which are most promising as the memristive neural materials.However, it is still necessary to develop more neural materials to enrich the electrical and dynamic behaviors, which is beneficial to accomplish more complicated tasks.2D materials, such as h-BN and MoTe 2 , have been recently studied for realizing neuron behaviors owing to their remarkable advantages (atomic thickness, dangling-bond-free surfaces, and lattice mismatch free), which are attractive for implementation of highly energy-efficient and ultra-high integrated brain-like neural networks compared to conventional materials.In addition, it would be surprising if the whole neuron behaviors could be accomplished in single memristive device.
Neuronal connections in planar two-dimensional materials necessitate establishing physical connections between memristive components.Implementing physical connections between them is difficult, because electrode lithography and two-dimensional materials have nanometer-scale thicknesses.The physical connections with regard to traditional bulk materials could be easily achieved through some methods like lattice matching, but these methods are not applicable to two-dimensional materials.The connections between devices may also be limited by the conductivity and performance of flat two-dimensional materials.It may be difficult to construct intricate neural connections using twodimensional materials because some of them have strong conductivity in the horizontal direction (inside the plane) but poor conductivity in the vertical direction (between planes).
Models and Applications: in terms of the implementation of brain-like intelligence system, it is the foundation that constructs a certain scale of SNN.Thus, the primary issue is to accomplish the interconnecting architecture between the artificial neurons and synapses.Although the related works have been reported which indeed have improved the development of SNNs [175], there are still a large number of problems need to be solved.First, the integration approach between the artificial neurons and synapses is an important issue, which actually should not be directly connected in series.Secondly, it is another important consideration that the spiking pulse generated from the pre-neurons possesses the capacity to modulate the weight of synapse, which means that these artificial neurons and synapses could interact cooperatively.Therefore, the parameter matching relation between them needs to be investigated thoroughly.Finally, it will be exciting to utilize the SNN hardware to solve practical issues, such as mathematical reasoning, time series prediction, intelligent identification and task classification.In addition, it is still necessary to introduce more dimensions, such as force, temperature, humidity, optics and magnetism, to further enrich the perception capacity of artificial neurons and synapses, which is beneficial to implement the brain-like sense-computation integration and even sense-memory-computation integration.We believe that the neural network architecture composed of artificial neurons and synapses could highlight the advantages in the near future through continuous advances in memristive materials, models and applications.
Apart from the classical neuronal models, there are expanded neuronal models that aim to better simulate the complex behavior of biological neurons and achieve more advanced neuronal computations.Models like the LIF model and H-H model require complex external circuit parameters.However, when it comes to a neuron implemented using a single memristive device, deriving a simple set of mathematical formula models may be challenging.It is difficult to provide a simple mathematical model for memristive devices since they exhibit non-linear electrical characteristics and RS effects.Nevertheless, in the context of neuromorphic computing, researchers continue to investigate and create simplified models that can successfully replicate the fundamental properties and behavior of memristor-based neurons.
In addition, according to a variety of applications, it is possible to imitate more behavior of neurons by using threeterminal devices.This method enables a better comprehension of the dynamic properties of neurons as well as the reaction and behavior of various neuron models under particular circumstances.The accuracy and performance of neural computing models can be improved by adjusting the parameters of neurons to the actual properties of biological neurons.Artificial intelligence and neural computing now have new avenues for exploration.It not only advances our understanding of neuron function but also offers fresh approaches and techniques for building more flexible and intricate neuromorphic networks.
Finally, reservoir computing utilizes an artificial neuron pool to perform computations on temporal data and may face the following challenges: First, setting connection weights: the connection weights between artificial neurons in the reservoir need to be appropriately set.This involves the initialization and adjustment of weights to ensure that the reservoir has the desired dynamic range and information processing capability.Determining the appropriate weight values is a challenge that requires experimentation and optimization through adjustments.Then, the dynamic range and nonlinearity of artificial neurons are crucial for the performance of the reservoir.The reservoir needs to have a sufficient dynamic range to capture various patterns and variations in the input data.Additionally, the selection of an appropriate nonlinear activation function is a key issue to ensure that the reservoir can generate rich nonlinear dynamics.The development of efficient and powerful artificial neurons, along with advancements in understanding the dynamics of reservoir computing systems, will continue to drive the progress of brain-like neural network.
In the future, this method is anticipated to become more crucial as a result of ongoing technical breakthroughs in the fields of neuroscience and artificial intelligence.It has the potential to make a significant contribution to the study of neural dynamics and the creation of cutting-edge artificial intelligence systems.

Figure 1 .
Figure 1.Graphic summary of recent advances in research on the development of artificial neurons using memristors, spanning the implementation of neuronal components, their functions, and their applications.

Figure 5 .
Figure 5. Perception neuron.(a) Pressure receptors, as tactile sensors, are connected to oscillatory neurons to form an artificial sensory system.(b) Experimental data from the artificial spiking mechanoreceptor system (ASMS).(c)-(f) A magnified diagram of (b).The protective inhibition behavior can be observed in (c).Reproduced from [44], with permission from Springer Nature.

Figure 6 .
Figure 6.Perception neuron.(a) Schematic of artificial spiking curvature perception neuron.(b) The neuron exhibits spiking responses triggered by different bending radii.(c) Artificial spiking curvature perception neurons response under different gestures.Reproduced from [45], with permission from Springer Nature.

Figure 7 .
Figure 7. Optical signal processing.(a) Schematic diagram of the electrical circuit underlying the artificial LIF neuron.(b) The charging process of a capacitor caused by an input voltage.(c) and (d) The number of pulses varies with R and C. (e) Output neuronal firing from LIF neuron.(f) Relationship between laser power and spike frequency.(g) The LGMD response increases initially and then decreases with the increase in time.(h) The relationship between the number of spiking pulses and the light power varies.(i) A flowchart of the motion trajectory of a robot car containing an artificial neuron circuit (j) Schematic diagram of measurement equipment associated with the model car.(k) Schematic of the decision-making process engaged by the robot car with laser signal processing ability.Reproduced from [46], with permission from Springer Nature.

Figure 8 .
Figure 8. Flexible protein nanowire devices.(a) Schematic illustrations of Ag/dielectric/Pt memristors arrays on a polyimide substrate.Transmission electron microscope (TEM) image of protein nanowires.(b) I-V characteristic curves associated with different current compliances (I CC ).(c) The protein nanowire sensor with a planar structure.Breath-driven current response.(d) Circuit diagram of input stimulation to the optical sensing element of an artificial neuron that emulates an optical afferent circuit.Relationship between membrane potential (Vm) and current in the artificial neuron obtained by inputting an optical signal.(e) Integrated wearable neuromorphic interface.Reproduced from[47], with permission from Springer Nature.

Figure 9 .
Figure 9. Brain-machine interface.(a) Schematic diagram of biological hybrid neural interface, in which artificial neurons communicate chemically with biological neurons to complete exchange and transmission of information.(b) Artificial neurons, including three parts, the detailed content is introduced in this review.(c) Schematic diagram of the connection between artificial neurons and mouse sciatic nerve.(d) and (e) Leg bending angle of mice leg under different concentrations of dopamine stimulation.(f) Response angle and different concentrations of dopamine stimulation of the leg injury.Reproduced from [174], with permission from Springer Nature.

Table 1 .
Comparison of neuronal models and functions with different memristive materials.