In today's digital era, artificial intelligence (AI) is rapidly becoming a key force driving technological progress. With breakthroughs in algorithms such as deep learning, AI has achieved remarkable success in fields like image recognition, natural language processing, and intelligent decision-making. However, despite the significant advancements on the software level, the development on the hardware level has been relatively lagging. To achieve more efficient and energy-saving intelligent systems, neuromorphic computing and artificial sensory systems have emerged, aiming to simulate the working principles of biological nervous systems to achieve processing capabilities and efficiency similar to the human brain.

The concept of neuromorphic computing was first proposed by professor Carver Mead of the California Institute of Technology in 1989^[1]. Its core idea is to use very large scale integrated circuit technology to simulate the electrical characteristics of biological neurons and synapses, and to construct hardware systems capable of performing complex computing tasks. Compared with traditional von Neumann computer architecture, neuromorphic computing systems have advantages such as strong parallel processing capabilities, low power consumption, and good fault tolerance, making them particularly suitable for processing high-dimensional data such as images and speeches. Artificial neurons synapses are the basic units of neuromorphic computing systems.

Artificial sensory systems, a significant application of neuromorphic computing, combine artificial sensory neurons with various sensors and artificial synapses to simulate the functions of biological sensory organs, enabling perception and response to the external environment^[2]. Biomimetic sensors can respond to various external stimuli. For visual emulation^{[3, 4]}, photodetectors capture external light and convert it into electrical signals. For olfactory emulation^{[5, 6]}, chemical sensors detect gas molecules. To mimic gustation^{[7, 8]}, electrochemical sensors identify substances dissolved in saliva. For mimicking tactility^[9−11] and hearing^[12], flexible electronic sensors and piezoelectric materials replicate the skin's sensation and the ear’s structure. However, the signals output by these sensors are mostly analog voltage or current signals, which cannot directly act on the subsequent synaptic devices and require artificial sensory neurons to process these signals to produce spike trains for processing in synaptic devices.

Currently, the widely studied neuron models include Hodgkin−Huxley (HH) neurons^[13], leaky integrate-and-fire (LIF) neurons^[14−16], and oscillatory neurons^[17]. Among these, HH neurons present the richest neuron dynamics, making their hardware implementation crucial for simulating the dynamic behavior of human brain neural networks. FitzHugh simplified the HH neuron model into a two-dimensional dynamic FitzHugh−Nagumo (FHN) model by omitting chaotic dynamics and burst firing^[18]. As a simplified version of the HH model, LIF neurons, although lacking rich neural dynamics, have a simple circuit structure, are easy to integrate at high density, and meet most neuromorphic computing requirements, thus gaining widespread attention. Unlike HH neurons and LIF neurons, oscillatory neurons lack biological plausibility, but due to the coupled oscillation of multiple oscillatory neurons and their similarity to the macroscopic electroencephalogram output of the human brain, they are of great significance for studying the coordinated resonance of a large number of neurons. Early artificial neurons were mainly based on complementary metal−oxide−semiconductor (CMOS) technology, implementing the integration, firing, and resetting functions of neurons through analog circuits. However, with technology advancements, there is a growing realization of the need for new types of electronic devices to more efficiently implement artificial neurons. These devices, such as new silicon-based devices and memristors, have unique physical properties and can more accurately simulate the dynamic behavior of biological neurons. At present, artificial sensory neurons mainly adopt the LIF model, which can also achieve oscillation functions and adjust their firing frequency.

In addition to the front-end sensors and the sensory parts of the artificial sensory neurons, artificial synapses are also needed to preprocess the perceived information and then send it to the brain. Artificial synapses have been built with virous electronic and optoelectronic devices^[19−21]. In 2015, Strukov et al.^[22] prepared excellent performance memristors based on double-layer metal oxides, adopting a cross-array structure to prepare a 12 × 12 memristor synaptic array, and solved the current interference problem in the array through the nonlinearity of the devices. In 2018, ferroelectric synaptic transistors^[23] with Hf_0.5Zr_0.5O₂ as the ferroelectric insulating layer were demonstrated, achieving 32 analog states, symmetric potentiation/depression, spike timing-dependent plasticity (STDP), and other synaptic functions. In 2019, a photoelectric transistor^[24] based on selenide nanosheets, which can achieve long-term potentiation (LTP) and long-term depression (LTD), was realized. In 2020, Zhuge et al. presented all-optically controlled memristors^[25−28] based on oxide semiconductors, which can simulate synaptic functions using only light stimili. In 2024, domain wall magnetic tunnel junctions^[29] were used to achieve all-spin synapses and activation function generators. These results on synaptic devices lay a good foundation for the construction of artificial sensory systems.

This review primarily focuses on the latest research advancements in artificial sensory neurons, covering methods for their hardware implementation and working mechanisms. In addition, it elaborates on the integration of artificial sensory neurons with sensors and artificial synapses across various sensory modes such as vision, touch, hearing, taste, and smell, along with progress in multimodal fusion perception systems. Finally, it discusses the current challenges in the hardware implementation of artificial sensory neurons and highlights their future development prospects. The specific contents are shown in Fig. 1.

The biological sensory system consists of five main senses: vision, touch, taste, smell, and hearing, each with specialized sensory receptors and sensory neurons^[40]. These sensory neurons are distributed among different nerves in the brain. The optic, olfactory, trigeminal, facial, vestibulocochlear, glossopharyngeal, and vagus nerve all contribute to perceiving the outside world (Fig. 2).

The optic nerve contains six types of sensory neurons, processed by the cone and rod cells for photoelectric conversion. Bipolar neurons receive signals from these cells and transmit them to downstream ganglion cells, regulating contrast and adaptation of visual signals. The transmission from photoreceptors to bipolar cells and from bipolar cells to ganglion cells is regulated by horizontal and amacrine cells, respectively. Ganglion cells are the final stop for visual information in the retina, integrating information in time and space before transmitting electrical signals through the optic nerve to the brain’s visual cortex for further processing. Information transmission between neurons primarily occurs through chemical synapses. In simple terms, the electrical activity of one layer of neurons triggers the release of specific chemicals, known as neurotransmitters, which influence the electrical activity of the next layer. Some neurotransmitters enhance the electrical activity of the subsequent neurons, while others inhibit it. For example, neurotransmitters from bipolar neurons can enhance the electrical activity of ganglion cells, whereas those from amacrine cells can inhibit it^[41]. The sensory neurons in the trigeminal nerve are mainly general somatic sensory fibers, responsible for superficial sensory perception, such as touch, pain, and temperature. The vestibular nerve transmits signals for position and balance, while the cochlear nerve is responsible for hearing. Taste is perceived by the combined efforts of the facial, glossopharyngeal, and vagus nerves. The facial nerve detects sweet, salty, and sour tastes on the front of the tongue, while the glossopharyngeal nerve senses bitter and sour tastes at the back. The vagus nerve contributes to taste perception in the throat and pharynx. The olfactory nerve is a purely sensory nerve, with primary neurons being bipolar. It allows the perception of external odors^[40].

Central to the development of artificial sensory neurons are artificial neurons. This section will discuss recent advances in artificial neurons based on emerging electronic devices like transistors and memristors.

Traditional silicon-based transistor neuron circuits are relatively complex, making it a research frontier and hot topic to streamline circuit structures for better scalability and energy efficiency. Kornijcuk et al.^[42] utilized the non-volatility of floating-gate transistors to construct an LIF neuron circuit. The entire neuron circuit can emit spike signals at frequencies below 100 Hz, with power consumption for a single spike emission under 30 pW. This study used floating-gate transistors to integrate synaptic signals, providing a new direction for simplifying neuron circuit structures. Researchers have also explored other methods to simplify neuron circuits or reduce energy consumption. For example, Kwon et al.^[43] designed a neuron circuit with an emission energy of 120 fJ per spike using dual-gate positive feedback transistors. Chatterjee et al.^[44] employed the highly scalable fin field-effect transistor to design an LIF neuron with a spike emission energy of 6.3 fJ. Despite these innovative efforts to achieve neuron circuits that are highly integrated and energy-efficient, as Moore's Law gradually loses effectiveness, neuron circuits based on traditional silicon transistors struggle to meet the growing computational demands of the current era.

In 2018, Han et al.^[45] developed a new type of LIF neuron based on an n−p−n structured two-terminal silicon device using the latch-up effect. This device realized the simple spike emission function of the LIF neuron but had a low emission frequency and high energy consumption. Later, Han et al.^[46] designed a 1B2T neuron using one bistable resistor (B) in parallel with two transistors (T), which reduced the pulse width and amplitude of the voltage spike, increased the neuron's maximum emission frequency, and reduced power consumption.

In 2021, Han et al.^[33] utilized vertically structured Si nanowires to construct a ring gate transistor, obtaining a 1T neuron device. Fig. 3(a) shows the device structure, where the top n⁺ type silicon serves as the source, the middle is p−type silicon, and the bottom n⁺ silicon acts as the drain. Fig. 3(b) illustrates the energy band structure of the device when the drain input current (I_in) is applied and V_G is −1.5 V. By applying a constant current (I_in = 20 nA) to the drain, the neuron can be made to emit or be inhibited by adjusting the gate voltage (Fig. 3(c)).

In 2019, Wan et al.^[47] utilized a sodium alginate bio-polymer electrolyte as the gate dielectric to construct an indium−tin−oxide transistor, as shown in Fig. 4(a). This device can emulate key characteristics of pain-perceptual nociceptors, such as pain threshold, memory of prior injury, and pain sensitization/desensitization. In 2021, Wan et al.^[48] constructed a vertical coplanar multiterminal flexible transient photogate transistor network with a 3 nm ultra-short channel (Fig. 4(b)). This network can emulate visual nociceptor sensitization triggered by spatiotemporal pattern stimuli (SPS) and spatiotemporal color stimuli (SCS).

Memristors are the fourth fundamental type of passive electronic component, in addition to resistors, inductors, and capacitors. They have a resistance dimension and exhibit nonlinear electrical properties distinct from ordinary resistors. The memristance value of a memristor can be continuously and reversibly adjusted under external signal stimulation and can maintain the changed state after the stimulus is removed. Memristors typically have a sandwich structure of top electrode−dielectric layer−bottom electrode, with advantages such as simple structure, easy high-density integration, low energy consumption, and fast read and write speeds.

According to their resistance variation mechanisms, memristors are classified into several categories, such as conductive filament type and purely electronic type^{[49, 50]}. The conductive filament type involves the migration of oxygen vacancies or metal ions, forming conductive filaments. The formation and rupture of these conductive filaments induce the resistance transition of the memristor. The resistance variation mechanism of the purely electronic type memristor is the capture and release of carriers at defects, such as oxygen vacancies^[27].

According to resistance retention, memristors can be divided into non-volatile and volatile types. The resistance of non-volatile memristors can be maintained for a long time and can be used to simulate the cumulative function in biological sensory neurons. With the assistance of reset circuits, comparators, and pulse generators, non-volatile memristors can implement LIF neurons^[51−53]. However, this artificial neuron circuit structure is complex, has high hardware consumption, and consumes a lot of energy.

Compared to non-volatile memristors, volatile memristors typically exhibit threshold switching (TS) characteristics, including conductive filament memristors and mott phase change memristors. When the applied voltage exceeds the threshold voltage (V_th), the device abruptly transitions from a high-resistance state to a low-resistance state. Conversely, when the applied voltage is below the holding voltage (V_h), the device automatically returns to the high-resistance state^[54] (Fig. 5(a)). This characteristic is highly suitable for simulating the self-restoring behavior of the neuronal membrane potential, avoiding the complex reset circuits used in neuronal circuits^[55]. Mott memristors^[56−59], ovonic threshold switching (OTS) memristors^[60−62], and conductive filament memristors^[63−65] can all achieve TS, and their neuronal circuit structures are consistent^[66]. Fig. 5(b) shows a neuron circuit based on a TS memristor^[54], where R_L is the voltage-dividing resistor, R_S is the test resistor, and C is the device's parasitic capacitance or an external capacitor. The voltage at node "2" exhibits oscillatory firing behavior, while the current at node "3" displays LIF-type spike firing behavior (Fig. 5(c)). Lee et al.^[66] systematically studied artificial neurons made from mott devices, OTS devices, and conductive filament memristors. The results show that the high-resistance state and switching time of the TS device respectively determine the leaky/non-leaky characteristics of the neuron and the type of activation function, while a higher low-resistance state and a lower threshold voltage can effectively reduce the neuron’s total power consumption (Fig. 5(d)).

In addition, Zhuge et al.^[36] demonstrated a bipolar TS (BTS) memristive neuron that can operate in both excitation and inhibition modes. In this BTS memristor, the V_th is increased by applying a negative voltage (RESET operation), and decreased again by applying positive voltages (Fig. 5(e)). Applying RESET voltages with different values results in different change amplitudes of V_th (Fig. 5(f)). It is inferred that this BTS phenomenon may be widespread in filamentary memristors.

Ferroelectric transistors, spintronic devices, and phase change devices also provide effective technical pathways for achieving artificial neurons that are highly integratable and energy-efficient. For instance, ferroelectric materials can generate spontaneous polarization within a certain temperature range, and the polarization direction can be reversibly altered by controlling the polarity of the external electric field^{[67, 68]}. Ferroelectric transistors, based on these materials, can simulate the dynamic behavior of neurons by matching the appropriate circuits. Chen et al. developed ferroelectric transistors that can automatically reset by controlling the crystallinity of the Hf_0.5Zr_0.5O₂ ferroelectric thin film, and subsequently constructed LIF neurons with frequency-adaptive firing functions^[69] (Fig. 6(a)) and LIF neurons without the need for capacitors^[70]. However, controlling the crystallinity of ferroelectric materials is challenging, making it difficult to meet the needs of large-scale device integration.

Compared to ferroelectric materials, in antiferroelectric materials, the adjacent dipoles within the material point in opposite directions under zero electric field, resulting in an overall remnant polarization strength of zero. When a sufficiently strong electric field is applied, the antiferroelectric phase transitions to a ferroelectric phase. However, once the electric field is removed, the ferroelectric phase returns to the antiferroelectric phase^[71]. Cao et al.^[30] used Hf_0.2Zr_0.8O₂ antiferroelectric thin films to construct antiferroelectric transistors (Fig. 6(b)). As shown in Fig. 6(c), by applying voltage pulses of different amplitudes on the gate, the device can simulate the dynamic accumulation and leakage behavior of neuronal membrane potential. After the neuron circuit is built, LIF firing behavior and the refractory period can be achieved.

Spintronic devices utilize the spin and magnetic moment of electrons to achieve reversible switching between high-resistance and low-resistance states^{[72, 73]}. By controlling the size of the magnetic domain within the device through current, the accumulation process of the neuronal membrane potential can be simulated. With the help of appropriate peripheral circuits, the functions of biological neurons can be realized. In 2023, Wang et al.^[38] used antiferromagnetic materials to construct a volatile spin−orbit torque magnetic tunnel junction (Fig. 6(d)). Under the stimulation of external magnetic fields and current, the magnetic domain in the antiferromagnetic body gradually grows from left to right, simulating the neuronal accumulation function (Fig. 6(e)). When the amplitude of the current pulse is zero, the residual Joule heat within the device regulates the built-in magnetic field and the mutual competition of the Ruderman−Kittel−Kasuya−Yosida coupling effect, causing the magnetic domain to grow from right to left, a process that can simulate the neuronal leakage function. Fig. 6(f) shows the device simulating LIF neuronal accumulation and leakage processes under current pulses.

Phase change devices are a kind of non-volatile device with a functional layer made of phase change material. When voltage is applied, the phase change material undergoes a transition from an amorphous to a crystalline state due to Joule heating, allowing the device to switch from a high resistance state to a low resistance state. Applying voltage again causes the crystalline state to melt back into an amorphous state, completing the transition from a low resistance state to a high resistance state. By controlling the excitation voltage pulse, only partial crystallization occurs each time, simulating the accumulation process of biological neurons^{[39, 74]}. Similar to ferroelectric and spin neurons, phase-change neurons require setting circuits, comparators, and pulse generators to achieve neuron membrane potential regression to the initial state, threshold comparison, and spike emission behavior^[52]. Due to the randomness of nucleation and growth in phase change materials, the emission behavior of phase change neurons is inherently unpredictable, laying a foundation for probabilistic inference calculations. Researchers have designed an automatically set phase-change neuron circuit^[75], reducing circuit complexity and energy consumption.

Human perception of the real world has tremendous advantages in efficiency, robustness, flexibility, and fault tolerance. The human sensory system comprises various senses that collaborate with the brain, enabling individuals to explore and gather information. This biological sensory system integrates the sensing, storage, and computation of sensory data, facilitating the efficient and low-energy collection and processing of external information. This perceptual processing architecture in humans inspires the development of new types of perceptual systems.

To simulate biological perception and processing methods, researchers have explored two neuromorphic perceptual architectures. 1) By interconnecting sensors with artificial synapses, they have constructed perceptual systems that integrate sensing, storage, and computation. This architecture has been used to achieve simple recognition of sensory information such as vision^{[76, 77]}, touch^{[24, 78, 79]}, hearing^[80], taste^[81], and smell^{[82, 83]}. This architecture is suitable for integrating sensors that can directly output pulse signals. 2) By connecting sensors, artificial sensory neurons, and synapses in sequence, various continuous stimuli from the real world can be transformed into neural spikes, which are then stored and computed by the synapses. This architecture fully simulates the sensory collection, integration, and output functions of biological sensory neurons and can adapt to various types of sensory signal inputs, thus receiving more attention^[84−88].

This section will discuss various artificial sensory neurons and their applications in terms of sensory types.

Traditional visual perception systems typically consist of cameras, memory units, and processing units. The camera senses light signals from the real world, capturing a series of images (frames) that are then sent to the processing unit for analysis. With the aid of algorithms, tasks such as target extraction and edge detection are completed. During this process, visual information is continuously shuttled between the memory and the processing unit, resulting in a large amount of redundant information, which makes the visual system inefficient and energy-intensive.

In the human visual system, photoreceptors (cones and rods) convert light signals into encodable electrical spike signals, which are preprocessed by bipolar cells and ganglion cells before being transmitted to the visual cortex of the brain. Through the collaboration and synchronization of multiple areas in the visual cortex, the brain obtains visual information such as the shape, color, position, direction, and contrast of the target. Inspired by the human visual system, neuromorphic visual systems have become a research hotspot.

Building visual neurons that convert visual information into spike signals is the first step in realizing a neuromorphic visual system. At present, integrating photoelectric sensors with artificial neuron devices is a common method to achieve the function of visual neurons. Wu et al.^[89] combined a Ta/InGaZnO₄ (IGZO₄)/Pt ultraviolet light sensor with an NbO_x-based oscillating neuron to construct an ultraviolet visual neuron (Fig. 7(a)). Because the resistance of the IGZO₄ sensor under ultraviolet light falls between the high and low resistance states of the NbO_x memristor, they can be directly connected to achieve neuronal oscillating emission. The IGZO₄ sensor was sequentially exposed to 365 and 254 nm ultraviolet light, causing the sensor resistance to decrease accordingly, while the output frequency of the NbO_x oscillating neuron gradually increased. Fig. 7(b) shows the oscillation wave of the neuron when 365/254 nm ultraviolet light is applied simultaneously. By employing a pulse coupled neural network (PCNN), image segmentation of an ultraviolet butterfly image was achieved (Fig. 7(c)). The limitation is that this visual neuron only responds to ultraviolet light, and to simulate the human visual system, it is necessary to further expand the wavelength response range of the photoelectric sensor.

Pei et al.^[35] integrated a TiN/PbS/ITO structure photoelectric synapse and an Ag/MoS₂/Ag/MoO_x/Ag conductive filament neuron to build a visible light perception visual neuron (Fig. 7(d)). The PbS photoelectric synapse's resistance decreases under different wavelengths, increasing the neuron's firing frequency. As shown in Fig. 7(e), this feature was used to simulate the dynamic behavior of the visual neuron's spike emission in a vehicle meeting scenario. As the oncoming vehicle approaches, the light intensity increases, causing the neuron's firing frequency to increase, alerting the vehicle to slow down or take evasive action.

Humans can quickly adapt to visual information at different distances (Fig. 7(f)), making the simulation of biological visual depth recognition an important research direction. Chen et al.^[90] utilized a photoresistor, an Ag/TaO_x/ITO structure LIF neuron, and an IGZO floating-gate transistor synapse to construct a visual perception neuron (Fig. 7(g)), successfully simulating the fatigue phenomenon of biological vision and the recognition of depth information (Fig. 7(h)). In this work, pulsed incident light is needed to irradiate the photoresistor, a feature that poses challenges in adjusting neuronal firing behavior under continuous light stimulation in real environments.

Ganglion cells are an important component of the biological visual nervous system, with their receptive field comprising an approximately circular central part and a ring-shaped peripheral part. Stimulating the central and peripheral parts produces opposite effects. Ganglion cells with increased peak frequency in the center and decreased peak frequency in the periphery are on-center type, while the opposite are off-center type. Through their cooperation, the biological visual system can perform contrast analysis and extract the shape of objects. Bao et al.^[91] used a memristor and a metal−oxide−semiconductor field-effect transistor (MOSFET) to simulate photoreceptor cells and ganglion cells, where the voltage changes of the memristor and MOSFET can respectively simulate neural spikes and membrane potential changes caused by light exposure. On this basis, researchers constructed a new type of tunable integrate-and-fire memristor neuron and achieved the functions of the two types of ganglion cells. Simulation results show that this neuron can extract shape information from images and convert it into spike frequency, achieving the computational function of the sensor.

Although integrating photoelectric sensors with artificial neurons can simulate visual neurons, this process inevitably increases circuit complexity and energy consumption. Therefore, optimizing the structure of artificial visual neurons remains a challenge that researchers are dedicated to addressing.

Photoelectric neurons can respond to ambient light, reducing the threshold voltage of the neuron device and increasing its firing frequency^[92−94]. Han et al.^[95] employed the latch-up effect to construct a single-device silicon-based photoelectric neuron, consisting of a floating body, drain−source electrodes, and a gate electrode (Fig. 8(a)). When exposed to light, the photogenerated electron−hole pairs accelerate the charge accumulation process, thereby reducing the threshold voltage and changing the amplitude and frequency of the voltage pulse output of the neuron. Fig. 8(b) shows the change of the neuron's output spike under a 100 nA input current and red light irradiation. In addition, green and blue light can regulate the firing frequency of the neuron (Fig. 8(c)). Moreover, the device can regulate the firing frequency through gate voltage and can inhibit the neuron. This capability has laid the foundation for simulating the lateral inhibition characteristics in human vision. Based on the light-controlled output characteristics of this neuron, researchers have constructed a single-layer perceptron, demonstrating letter recognition. However, this work did not develop a photoelectric neuron array with lateral inhibition.

To further expand the wavelength range of photoelectric neuron light response, Han et al.^[96] integrated an InGaAs-based n⁺/p/n⁺ structure photoelectric neuron with the MOSFET synaptic array to construct a 3D stacked vision array with integrated perception and computing, enabling the regulation of neuron firing frequency through light irradiation from visible to infrared wavelength. Wang et al.^[97] employed an Ag/black phosphorus/CsPbBr₃/ITO structure TS memristor to construct a 12 × 12 photoelectric memristor array (Fig. 8(d)). When exposed to 365, 450, and 520 nm light, the device shows a tunable threshold voltage. Based on this photoelectric memristor, the constructed photoelectric neuron, under the joint stimulation of a constant amplitude voltage pulse and continuous light signal, can increase the neuron's output spike firing frequency by increasing the light power density. Utilizing this feature, researchers have further achieved navigation of a machine car. Fig. 8(e) exhibits an integrated diagram of the photoelectric neuron and the machine car control panel. As the machine car approaches the light source, the light power density received by the memristor increases, thereby increasing the firing frequency of the neuron. When the firing frequency exceeds 2.5 Hz, the control circuit can guide the machine car to avoid a specified light source (Fig. 8(f)). However, this photoelectric neuron did not achieve the recognition of visible light color.

Wan et al.^[98] integrated the photoelectric sensor and TS memristor to create a vertical structure photoelectric memristor with the configuration ITO/IGZO/Ag/Ta₂O₅/ITO. Due to parasitic capacitance, the device can directly convert continuous light signals into voltage spike signals. Under the irradiation of 360, 405, and 532 nm light, the spike frequency output via the optoelectronic neuron varies in magnitude. Based on this feature, the team further identified mixed color graphics composed of these three wavelengths. The photoelectric neuron has a simple structure, high color discrimination, and can simulate the color recognition of human eye cone cells, providing a new idea for developing highly integrated neuromorphic visual sensors.

Compared to the insect visual system, the human visual system is more complex but relatively less effective at processing high-speed moving objects. This is because the refractory period characteristics of human neurons limit their high-frequency firing, while the graded neurons in the insect vision can respond to stimuli in a certain time domain with multiple levels of response. These neurons not only exhibit temporal accumulation but also can respond to external stimuli at any time. Inspired by graded neurons, Chai et al.^[99] used double-layer MoS₂ as the channel to build a photoelectric transistor (Fig. 8(g)). When the gate voltage is zero and the channel voltage is 0.1 V, under the illumination of a 660 nm light pulse, photoconductivity of the device shows a volatile multi-level response within a certain time domain, simulating graded visual neurons (Fig. 8(h)). On this basis, precise judgment of the motion direction of dynamic objects has been achieved (Fig. 8(i)). However, the neurons in this work cannot convert continuous external light signals into graded neuron spike signals.

Insects can also detect polarized light, which is common in nature but hard for the human eye to perceive. Jiang et al.^[100] proposed a new polarized neural transistor based on a two-dimensional ReS₂ phototransistor. They obtained the transmission characteristics of the transistor at different polarization angles with a polarization-sensitive photodetector system (Figs. 8(j) and 8(k)). Modulation of the transistor’s excitatory postsynaptic current (EPSC) can be achieved by changing the modulation voltage, thereby enabling memory-consolidating behavior in response to polarized light stimuli. This led to the application of polarimetric navigation and 3D visual polarimetric imaging. Building on this work, Jiang et al.^[101] developed a novel porous metal−organic−framework phototransistor with an anisotropic-ReS₂-based heterojunction, which exhibits notable polarization sensitivity and adaptability. Furthermore, through a polarization-electrical cooperation strategy, the polarization-sensitive visual adaptation with bottom gate control and environmental dependence was successfully realized.

At present, research on vision perception based on artificial neurons is still in the stage of principle device research, typically only achieving control of neuron firing frequency by light. Further in-depth exploration is still needed in areas such as array scale, integration of visual neurons and synapses, lateral inhibition, and frequency adaptation.

As the largest organ of the human body, the skin serves multiple functions, including protection, mechanical sensation, and temperature perception. The tactile receptors beneath the skin generate neural spike signals when stimulated by external pressure, which are then transmitted to the cerebral cortex via nerves, forming the sense of touch. Processing tactile information allows us to interact more effectively with the external environment.

Currently, the primary approach in bionic tactile research involves simulating the dynamic behavior of tactile receptors by integrating pressure sensors with sensory neurons. Zhang et al.^[102] developed artificial tactile neurons using NbO_x-based oscillating neurons and piezoelectric sensors (Fig. 9(a)). Piezoelectric sensors are a common type of pressure sensor. As pressure increases, the output voltage of the piezoelectric sensor also rises. By integrating the piezoelectric sensor with the NbO_x neuron, the voltage signal output by the sensor acts on the neuron, and after processing by the neuron, the analog voltage signal is converted into a neural spike. As shown in Fig. 9(b), as the output pressure of the piezoelectric sensor increases, the output frequency of the neuron also rises. It is worth noting that when the sensor's output voltage exceeds the neuron's limit, the neuron can achieve self-inhibition due to the large relaxation time of NbO_x itself, meaning it no longer outputs neural spikes (Fig. 9(c)). This feature can effectively prevent damage to the tactile neurons and extend their service life.

Simulating the sense of touch from a single touch point lays the foundation for further research into large-area contact perception. In human touch, neuronal cells integrate spatiotemporal spike information and transmit it to other neurons through synapses. This process allows the brain to decode and perceive tactile information, enabling us to identify crucial details like object shape from complex tactile stimuli. To replicate this brain function, Li et al.^[103] employed micro-pyramid structure piezoresistive sensors and NbO_x oscillating neurons to build tactile neurons (Fig. 9(d)). When external pressure is applied to the sensor, its resistance changes, producing a corresponding current. The current signal is fed into the NbO_x neuron, which, after integration, emits oscillation wave signals at different frequencies. By constructing an array of tactile neurons and simulating the spatiotemporal signal processing, researchers have used pulse-coupled neural network algorithms to recognize hand shapes and pressure distribution (Figs. 9(e) and 9(f)).

Similarly, Lee et al.^[105] combined OTS oscillating neurons and piezoresistive sensors to develop a tactile perception system for tumor recognition. Human skin is highly sensitive to external pressure and can easily perceive subtle pressures such as wind direction and speed. By employing triboelectric sensors and bistable resistive neurons, a tactile input neuron capable of sensing human exhalation and inhalation was created^[106]. Theoretically, integrating various types of pressure sensors with artificial neurons can achieve a richer tactile function, such as rapid and slow pressure adaptation^[107].

Compared to mott neurons and OTS neurons, conductive filament neurons can construct oscillating neurons, but due to the high randomness of neuronal firing behavior, their accuracy is relatively poor. However, the information integration and threshold transition characteristics of conductive filament neurons are suitable for simulating the function of nociceptors. In the body, when neurons receive external stimuli, they generate electrical signals and transmit them to nociceptors. If the stimulus exceeds the threshold of the nociceptor, an action potential is sent to the brain, resulting in a sensation of pain^[104]. The threshold characteristics of biological nociceptors can be simulated using conductive filament TS memristors. When the output voltage pulse of the pressure sensor is less than the TS neuron’s threshold voltage, the neuron will not release a spike signal; otherwise, it will output a spike signal. Another characteristic of nociceptors is non-adaptation, meaning that if nociceptors are subjected to continuous harmful stimuli, they will not stop firing spike signals to adapt to the stimulus but will continue to release spike signals. Under continuous stimulation of voltage exceeding the threshold, conductive filament memristors will form a continuous conductive filament and continuously output spike signals, which can simulate the non-adaptation characteristics of nociceptors. Sensitivity is another important characteristic of nociceptors, which can protect the injured area by enhancing pain.

Hyperalgesia and allodynia are two common pain behaviors, characterized by a reduced sensation threshold and increased pain sensation, respectively. By applying pulses exceeding the threshold voltage to the TS device, more metal can migrate into the device, simulating injured tissue. As shown in Fig. 9(g), as the damage value increases, the threshold voltage of the TS device decreases, similar to hyperalgesia. The increase in neuronal output current after stimulation is consistent with allodynia. Although conductive filament TS devices can simulate nociceptors, external stimuli may lead to the formation of highly stable conductive filaments, causing device failure.

Temperature perception is another crucial ability of nociceptors. When external temperature exceeds the threshold, nociceptors can fire spikes to alert the brain and prevent burns. By integrating thermoelectric modules with TS neuron modules, temperature nociceptors can be simulated. As shown in Fig. 9(h), when the heat source temperature is 40 °C, the nociceptor does not release a signal, indicating it is harmless. However, when the temperature exceeds 50 °C, the nociceptor can release a damage signal to alert the brain. Conductive filament LIF neurons constructed using ZrO_x^[108] and organic perovskite^[109] have also achieved similar functions.

Human skin can perceive cold, heat, pain, touch, etc., and can integrate and process information from multiple receptors. However, current tactile neuron research is still limited to the integration of single receptors, necessitating further research into the integration of multimodal perception and spatiotemporal information processing.

In the biological auditory system, sound waves vibrate the eardrum at specific frequencies and amplitudes, which are then transmitted through the ossicles to the cochlear hair cells and converted into electrical signals. The neural network's recognition of auditory signals requires the analysis of the frequency composition of sound waves (Fig. 10(a)). To simulate the biological auditory system, it is necessary to develop appropriate auditory sensors and neurons to perceive a wide range of sound wave frequencies. Currently, auditory sensors include electromagnetic auditory sensors, piezoelectric auditory sensors, and triboelectric auditory sensors. At present, there is relatively little research on combining auditory sensors and artificial neurons.

When both ears receive sound simultaneously, they can determine the direction of the sound while transmitting the information to the cerebral cortex. However, in simulating auditory functions, due to the slight distance differences between the two auditory sensors and the sound source, the sensor signals cannot be transmitted synchronously. To address this issue, in 2021, Zhou et al.^[110] used two piezoelectric auditory sensors and FHN neurons to appropriately scale the transformation of physical variables and circuit equation parameters, obtaining an auditory neuron model that can generate various firing patterns similar to those in biological neurons, such as the resting state, spike firing, burst firing, and chaotic firing. When two auditory neurons are stimulated by the same external sound waves, the two neurons can synchronously exhibit spike firing and burst firing, respectively. Furthermore, when appropriate noise is applied, the two neurons can be effectively synchronized.

The human auditory system can not only distinguish sound waves with frequencies between 20 and 20 000 Hz but also differentiate pitches with a frequency difference of less than 0.5%. Recognizing different sound wave frequencies has always been a key technology in audio signal processing. In 2023, Yun et al.^[111] developed an auditory neuron module with self-sensing capabilities by connecting a triboelectric auditory sensor with a bistable resistive neuron^[45] (Fig. 10(b)). The triboelectric auditory sensor inputs a current signal to the bistable resistive neuron, and when the current exceeds the device threshold, the neuron fires spikes at a certain frequency. When inputting sound waves of the same frequency but different decibels, the firing frequency of the neuron increases with the increase in sound decibels. By using two artificial auditory neurons and four MOSFET synapses, a single-layer artificial auditory spiking neural network (SNN) system was constructed (Fig. 10(c)). This auditory system can distinguish music played on a cello and a violin, both in the G3 pitch (Fig. 10(d)). This artificial auditory system uses an event-driven spike transmission scheme, which offers advantages in low power consumption and high efficiency.

Neuromorphic auditory systems have significant potential applications in fields such as disaster relief, cochlear implants, and intelligent diagnostics. Currently, the sensory system based on auditory neurons still needs in-depth research in miniaturization, portability, and biocompatibility.

From a physiological perspective, taste is usually classified into five basic sensations: sour, sweet, bitter, salty, and umami. These basic tastes can be combined to form various flavors. When a taste substance dissolves in saliva and interacts with the receptors of the taste cells, it is transformed into neural signals by the taste buds and then transmitted to the cerebral cortex, producing the corresponding taste (Fig. 11(a)).

To simulate the biological taste perception system, Han et al.^[112] integrated hydrogen ion and sodium ion sensors with 1T neurons to construct biomimetic sour and salty taste buds, serving as input neurons to classify and recognize vinegar and saline water (Fig. 11(b)). By connecting the pH-sensitive part with a MOSFET into an extended gate structure, a pH-sensitive artificial taste neuron was realized. The sour taste bud hardware consists of a pH sensing component and a 1T neuron, with the sensing component connected to the gate of the 1T neuron. As shown in Fig. 11(c), as the pH value of the taste quality decreases, H⁺ ions are bound by OH⁻ on the sensor surface, causing the sensing electrode surface to carry a positive charge. The neuron gate voltage increases with the rise of surface charge on the sensor. Due to the reduction of the potential barrier between the source and drain, the number of carriers injected from the source to the drain increases. Therefore, as the pH value decreases, latch-up voltage (V_latch) and latch-up current (I_latch) decrease. On this basis, an input current is applied to the drain, and the neural output frequency corresponding to different pH values varies (Fig. 11(d)), thus achieving the recognition of sour taste. Similarly, by connecting the sodium ion sensor^[113] to the gate of the 1T neuron, a taste bud hardware sensitive to salty taste is realized. When the Na⁺ sensitive carrier captures Na⁺, a membrane potential is generated due to the concentration difference at the membrane/solution interface. As the concentration of Na⁺ decreases, the neuron output frequency decreases accordingly (Fig. 11(e)). Finally, a single-layer taste SNN system was constructed by integrating taste bud neurons with synapses (Fig. 11(f)). As shown in Figs. 11(g) and 11(h), when the output current frequency of synapse A exceeds that of synapse B, it indicates the taste comes from vinegar, and the opposite is true for saline water. This taste perception neural system has the advantages of low power consumption and high integration compared to traditional taste perception systems. Although this work only simulated the neural transmission of sour and salty tastes, the integration of sweet^[114], bitter^[115], and umami^[116] sensors with neuron devices is also feasible.

Currently, because taste sensing has developed later than other sensory types, more in-depth research is needed in taste types, quantitative analysis, and mixed tastes. By combining neuron devices to simulate the biological taste system, the artificial taste system’s size, weight, and energy consumption can be effectively reduced, enhancing portability and ensuring rapid detection in fields like food safety, medicine, and heavy metal detection.

When olfactory receptors detect odors, the chemical reactions between them can output spike signals. These signals are preprocessed by intermediate neurons and then transmitted to the olfactory cortex of the brain for odor recognition (Fig. 12(a)).

To simulate the biological olfactory function in hardware, it is essential to convert signals from gas sensors into spike signals via neurons. Han et al.^[32] employed a chemical resistance gas sensor and a 1T neuron to build an olfactory input neuron (Fig. 12(b)). Fig. 12(c) shows the scanning electron microscope image of the 1T neuron device. When the gate voltage is zero, the output current exhibits a resistive change characteristic similar to that of a TS memristor due to the latch-up effect. As illustrated in Fig. 12(d), to operate the 1T neuron, a constant input current is applied to the drain, and its output voltage is measured. Upon applying the input current, the parasitic capacitor charges, similar to the neuron integration process. Due to the single-transistor latch effect, when the charge of the parasitic capacitor reaches the threshold, the output spike is emitted. By repeatedly charging and discharging, the oscillating characteristics of biological neurons can be mimicked. Fig. 12(e) shows that the firing frequency of the 1T neuron increases with the input current.

The gas sensing unit is based on SnO₂ and WO₃. In an oxygen environment, oxygen is adsorbed on the surface of the metal oxide, creating an electron depletion area. When gas is adsorbed on the sensing material’s surface, it reacts with the adsorbed oxygen, affecting the electron depletion area. Changes in this area lead to variations in the metal oxide’s resistance. As shown in Fig. 12(f), when NH₃ is introduced, the resistance of the SnO₂ and WO₃ sensors decreases with the increase of gas concentration. By connecting the sensor to the drain of the 1T neuron, an olfactory neuron is realized. When the SnO₂ sensor contacts NH₃, the sensor resistance decreases with increasing gas concentration, the current input to the 1T neuron drain rises, and the neuron output spike frequency increases (Fig. 12(g)). Using this mechanism, the recognition of four gases, namely, NH₃, CO, NO₂, and acetone, is achieved through a 4-layer SNN algorithm.

In addition, a single-layer olfactory SNN hardware system using olfactory neuron and synapses has been developed. This hardware enhances signal contrast and energy efficiency through lateral inhibition. When the gas sensor adsorbs volatile gases from Merlot and Shiraz wines, neurons connected to SnO₂ and WO₃ sensors are activated, producing frequency output in the corresponding synapses. If the frequency of synapse 1 is greater than that of synapse 2, it indicates Merlot wine (Fig. 12(h)); otherwise, it indicates Shiraz wine (Fig. 12(i)). This research successfully classifies targeted gases by simulating the biological olfactory system’s signal transmission mechanism. By realizing lateral inhibition, it mimics the olfactory system’s adaptability to prolonged exposure to specific gas environments.

The development of the neuromorphic olfactory system provides potential for portable or wearable gas sensors, with significant applications in medical diagnosis, public safety, environmental monitoring, and food safety. At present, research on the olfactory neuron-based sensory system is still in its early stages and needs further exploration regarding the types of gases that it can detect, quantitative analysis, and perception accuracy.

Multimodal fusion sensing in bionic sensory systems aids in comprehensively understanding object properties and making accurate judgments^[117−119]. In 2022, Liu et al.^[120] introduced a compact multimodal fusion spiking neuron (MFSN) structure that can achieve human-like multisensory perception. As illustrated in Fig. 13(a), MFSN integrates a pressure sensor to process pressure and an NbO_x-based memristor to detect temperature. The NbO_x memristor exhibits a distinct TS phenomenon (Fig. 13(b)), and the circuit schematic of MFSN, based on the pressure sensor R_p and the NbO_x memristor, is shown in Fig. 13(c). As the external pressure increases, the resistance value of R_p gradually decreases, and the spiking pulse frequency output by MFSN increases accordingly (Fig. 13(d)). For the NbO_x memristor, under a fixed input voltage of 4 V, higher sensed temperature results in lower spiking pulse amplitude and higher frequency (Fig. 13(e)). Subsequently, multisensory analog signals can be integrated into a single spike pulse sequence, and by decoupling the output frequency and amplitude, the pressure and temperature information can be distinguished from the fused spike pulse information, supporting multimodal tactile perception. This work has achieved the fusion between different senses in tactile perception but not integrated multiple senses.

Wan et al.^[119] developed a dual-mode artificial sensory neuron to achieve visual and tactile fusion. As shown in Fig. 14(a), the dual-mode sensory neuron collects light and pressure information from photodetectors and pressure sensors, respectively. The dual-mode signals are transmitted through ionic cables and converted into postsynaptic currents using synaptic transistors. By changing the time interval between the arrival of the two sensory signals at the synaptic transistor, the change in postsynaptic current (ΔEPSC%) can be modified. A significant change in ΔEPSC% is observed when the two signals arrive in a very short time, indicating signal synchronization (Fig. 14(b)). The sensory neurons can be activated at multiple levels, allowing manipulation of skeletal muscle tubes and mechanical arms (Fig. 14(c)). The average speed of the skeletal muscle tube varies with different ΔEPSC% changes (Fig. 14(d)). Finally, the fusion of touch and vision enhanced recognition capabilities in multi-transparency pattern recognition tasks. This work achieved fusion between different senses but only integrated touch and vision.

To simulate more complex biological activities, the fusion of multiple different senses needs to be studied. In 2021, Tan et al.^[37] reported a bio-inspired multi-sensory neural network that integrates different sensory neurons including vision, touch, hearing, olfaction, and taste (Fig. 14(e)). ITO/ZnO/Nb-doped SrTiO₃ photoelectric memristors receive light pulse signals, generating postsynaptic currents (PSCs). By integrating photodetectors with pressure sensors, the recognition of 26 English letters is achieved. The PSC from the visual part verifies the results of the tactile part. The tactile part outputs a total signal from each row of the pressure sensor to the photoelectric memristor, generating five PSCs. These five current values serve as inputs to the artificial neural network, ultimately recognizing handwritten letters (Fig. 14(f)). The integration of photodetectors and sound detectors combines visual and auditory information to determine when a person should cross the road. The signals from both detectors are superimposed and encoded into voltage pulses input into the memristor. A higher accumulated PSC value indicates the car is closer to the person (Fig. 14(g)). Finally, auditory and visual/olfactory/taste fusion learning were achieved. This work realized the mutual integration of various senses, providing ideas for multi-sensory multimodal fusion. However, given the use of photoelectric memristors, the transmission of tactile, auditory, olfactory, and taste signals needs to be converted into light pulse form first.

Research in the field of multimodal sensory fusion is still in its infancy, requiring further in-depth exploration of the integration of two or more different senses, methods of signal transmission, and expansion into various application fields. By integrating different sensors with sensory neurons, recognition accuracy in unimodal perception tasks can be enhanced, laying the foundation for developing artificial sensory systems.

Table 1 compares the performance of different sensory neurons. At present, most sensory neurons are oscillating or LIF neurons, and the sensory array size is small, typically capable of only simple tasks. Therefore, further research is needed in functional simulation, array size, recognition speed, algorithms, and other areas.

In recent years, research on artificial sensory systems has made remarkable progress, with its core goal being to simulate the complex functions of biological nervous systems to achieve efficient, low-energy perception and computation. By integrating advanced sensors, artificial neurons, and synapses, artificial sensory systems are gradually approaching the sophistication and efficiency of natural biological sensory systems. However, there is still a significant gap between current artificial sensory systems and the human brain in terms of power consumption, scale, and integration methods, making it difficult to meet the demands of complex functional applications. Artificial sensory systems based on artificial sensory neurons are still in the early stages of research, and numerous scientific and technological issues need to be overcome in terms of devices and systems.

At the device level, the implementation of artificial sensory neurons has evolved from complex circuits based on CMOS technology to the use of new types of neuromorphic devices, such as 1T neurons based on silicon nanowires^{[33, 121−123]} and memristors based on NbO_x^{[55, 124, 125]}. These new devices not only achieve basic neuron functions, such as the integrate-and-fire mechanism, but also simulate the self-restoring behavior of biological neurons by adjusting threshold and holding voltages, offering the potential to construct more efficient and energy-saving neuromorphic systems. However, further simplification of circuits is still needed, for instance, by designing devices that integrate sensors and neurons, like photoelectric neurons. Moreover, as the perceptual layer of the artificial sensory system, sensory neurons are the starting point of the entire system's information flow, and their accuracy, selectivity, and biological plausibility in perceiving real-world environmental signals can significantly affect the system's computational accuracy. Implementing advanced neuronal firing behaviors such as frequency-adaptive emission, stochastic resonance, and burst firing can help enhance signal detection. Additionally, artificial sensory neurons still require further optimization in terms of power consumption and stability.

For power consumption, neuroscience data indicates that the energy consumption of biological neurons when generating action potentials is very low, only in the femtojoule range, and it drops to the picowatt level in the resting state^{[126, 127]}. In contrast, the energy consumption of current electronic or photonic artificial neuron devices is relatively high, especially in the resting state. To reduce energy consumption, researchers are exploring various methods, including increasing the frequency of spike emission and reducing the working current. For instance, by employing mott transition and TS devices, which have rapid resistive transitions in the nanosecond range, the energy consumption of each spike emission can be reduced. Moreover, by improving device structures or adjusting material compositions, the threshold voltage and low resistance state current of the devices can be lowered^{[66, 128−130]}, further reducing energy consumption. It should be noted that reducing the working current may affect the switching speed of the devices; therefore, a balance between energy consumption and operating frequency must be considered during the optimization process.

Regarding stability, artificial neurons based on new types of neuromorphic devices frequently exhibit variations in electrical characteristics. These variations may stem from inconsistencies between devices or changes within a single device across different operational cycles. While this randomness has potential applications in probabilistic neuromorphic computing^[131−133], stability is crucial to ensure that devices can provide reliable and repeatable neuronal responses. Especially in sensory neuron applications that demand high precision, it is necessary to minimize electrical characteristic differences between devices and within individual device cycles to ensure that actual signals can be accurately converted into neural spikes. Enhancing the quality of the dielectric layer thin films is an effective way to reduce inherent randomness. For instance, using high-quality VO₂ epitaxial thin films can significantly reduce the electrical characteristic differences of a single device across different operational cycles^[134].

Apart from those common issues faced by artificial sensory neurons, each type has its own challenges. Visual neurons can adjust the emission frequency of devices using light, offering promising applications in areas like luminosity adaptation^[36], dynamic visual perception^[99], and polarization light vision^{[100, 101]}. However, the current array size is small, the range of sensitive wavelengths is limited, and they typically respond to only a few colors. Further research is needed to integrate synapses and neurons for complex functions. Currently, tactile neuron research is limited to single perceptions of temperature, touch and pain, whereas human skin can perceive multiple senses simultaneously, necessitating further study in multi-modal integration. Research on auditory neurons is limited, yet they hold significant potential in medical and rescue fields, offering a better balance between sensitivity and stability. Taste neurons still need to work with sensors, making it challenging to perceive and process external information in a single device, thus requiring more research to reduce power consumption and system complexity. Olfactory neurons mainly function with front-end sensors, so optimizing power consumption, size, and integration is possible. In addition, the number of gas types olfactory neurons can detect is a crucial indicator, warranting further research.

At the system level, researchers have been able to successfully integrate sensory neurons with sensors and artificial synapses, achieving perception and processing of external information. However, for complex real-life scenarios, perception through a single sense is not sufficient; often, a combination of multiple senses is required to accomplish tasks. Existing work in multimodal fusion perception has shown that by combining two senses to identify objects, the accuracy can be greatly improved, which is also more in line with our human habit of perceiving objects. The research on multimodal fusion perception is still in its infancy, with most efforts focused on the multi-sensory integration of a single sense, such as pressure and temperature in the sense of touch, and the integration between two senses, which is still far from the actual biological perception system. Moreover, processing the signals output by sensory neurons through artificial synapses is an indispensable part of the artificial sensory system. Currently, the connection between artificial synapses and sensory neurons is still at a simple interconnection stage, failing to reflect the advantages of high-integrated array parallel computing. Therefore, during the integration process, the compatibility of different components in terms of electrical and mechanical properties needs to be considered.

In summary, although artificial sensory neurons have made certain progress at the device and application levels, they still face many challenges. With continuous advancements in fields such as material science, nanotechnology, and integrated circuit design, artificial perception systems are expected to achieve leapfrog development. By simulating the complexity and diversity of biological nervous systems, intelligent systems capable of adapting to complex environments and performing complex tasks can be constructed. This will not only drive the innovation of artificial intelligence technology but also bring profound impacts to the development of human society.