The rapid advancement of information technology has revealed significant limitations in traditional von Neumann computer architecture, particularly in processing big data and complex computational tasks. The primary bottleneck stems from the separation of computing and storage units, resulting in increased latency and energy consumption^[1−4]. To address these challenges, researchers are exploring alternative computing paradigms, notably in-memory computing, exemplified by neuromorphic computing. This approach mimics the neural networks of the human brain, facilitating efficient in-memory computation by reducing the need for data transmission and simulating neuronal functions^[5−7]. Neuromorphic computing shows great potential in handling complex tasks such as image recognition and speech processing. Neurons and synapse are two basic building elements for constructing neuromorphic systems. considering numerous number of neurons in brain, developing high compact artificial neuron with dynamic neuronal firing patterns are demanding. The use of artificial neurons in neuromorphic systems enhances parallel processing capability and improves the recognition of intricate data patterns, making significant inroads into various fields^[8]. In applications like speech recognition and natural language processing, these systems enable intelligent assistants and translation services to operate more accurately and naturally, thereby enhancing user experience^[9].

To develop high-efficiency and high-precision neuromorphic computing systems, emerging memories are playing an increasing role, such as resistive random-access memories^{[5, 10−12]}, phase-change memories^{[13, 14]}, ferroelectric memories^{[15, 16]}, magnetic random-access memories^{[17, 18]}, and so on. Each technology offers distinct advantages, contributing to the creation of more efficient, low-energy, and highly integrated neuromorphic computing systems. Among these, memristors are particularly promising for neuromorphic applications^{[6, 19−24]}. Volatile memristors have attracted a lot of attentions due to their two-terminal configuration, high scalability, and dynamic switching behavior, where especially Mott^[25] and diffusive memristors^[26] serve as effective artificial neurons by mimicking key properties of biological neurons^[27]. They simulate neuronal firing behavior—wherein neurons respond to accumulated input and generate output action potentials—by reflecting activation levels through changes in resistance states. This capability allows for the optimization of neural network performance and the realization of reconfigurable and adaptive computing systems^{[28, 29]}. Importantly, the implementation of complex neural network operations at the hardware level based on volatile memristors minimizes data transmission between storage and processing units, thereby reducing energy consumption and improving computational efficiency^[5]. The unique attributes of volatile memristors—such as dynamic switching behaviors, fast switching speed and low energy—are particularly significant for applications with stringent energy requirements, including edge computing and mobile devices.

This article reviews and analyzes cutting-edge research on volatile memristor-based artificial neurons and discusses future prospects for both volatile memristor-based artificial neurons and neuromorphic computing.

Leaky integrate-and-fire (LIF) is a fundamental model emulating the behavior of biological neurons^[30]. These neurons are known for their ability to produce an all-or-nothing output spike when the membrane potential reaches a critical threshold, akin to the action potential observed in biological neurons^[31]. The firing of a neuron is governed by the aggregation of input signals; once the accumulated potential hits a specific threshold, the neuron emits a spike discharge. The neuron gathers charge through the integration of input signals, and when this charge reaches a certain threshold, a spike is triggered. Following this, the neuron enters a refractory period, after which it begins to accumulate charge again in readiness for the subsequent spike discharge. In 2017, it was demonstrated that Mott insulator materials could be employed to construct LIF neurons. These neurons integrate input pulses and emit current pulses, replicating the emission behavior of a neuron upon reaching a threshold^[32]. Mott memristor-based neurons operate at lower voltages and currents, featuring a simple double-ended structure, rapid response times, and high energy efficiency. This is attributed to the Mott insulators' ability to switch between stable insulating and metastable conducting states.

Mott memristor is a type of memristor based on Mott insulator materials (e.g., NbO_x^[33], VO₂^[29]) with remarkable conductivity change properties for neuromorphic computing and intelligent information processing systems. Mott insulators are composed of strongly correlated electronic systems^[34], where electrons may be localized due to Coulomb repulsion between electrons, in turn leading to cleavage of energy bands and formation of insulating states. On roof of this theory, conductivity of Mott insulators changes drastically in response to subtle changes in external conditions such as pressure^[35], voltage^[36], and temperature^[37], realizing rapid switching from insulation state to conduction state. This unique property makes Mott insulator-based memristors show great potential in constructing artificial neurons and synapses, especially in simulating the dynamic behavior of biological nervous systems^[38].

Chen et al. successfully reduced the electroforming voltage and increased the activation yield of the threshold switching behavior by precisely controlling the oxygen content in the NbO_x film. Furthermore, the group proposed a novel stacking method, inserting an NbO_y layer with high oxygen content between the low-oxygen NbO_x layer and the bottom electrode^[8]. Since the oxygen content of the NbO_y layer is higher than that of the NbO_x layer, the charge carrier concentration in the NbO_y layer is higher under the action of the electric field, which increases the local electric field strength. Further, the enhanced local electric field contributes to the formation of more stable and uniform conductive channels in the NbO_x layer, which reduces electrical resistance changes due to uneven electric field distribution or oxygen vacancy migration during cycling operation, thereby improving the cycling stability of the device. Meanwhile, by controlling the thickness and oxygen content of the NbO_y layer, the electric field distribution and the formation of conductive channels can be precisely regulated, resulting in better consistency in the electrical performance of each device. Through experiments and physical modeling, the researchers gained an in-depth understanding of the effect of oxygen content on the performance of NbO_x devices, which provides a methodology for the large-scale fabrication of highly stable NbO_x devices.

Nevertheless, maintaining the quality and consistency of the NbO_x and NbO_y layers during mass production (extremely precise oxygen content control process) is a common challenge, as even small manufacturing variations may affect the performance of the final device. In short, the resistance switching mechanism of Mott memristors involves the localization and delocalization of electrons, which is achieved by changing the concentration of oxygen vacancies. At high concentrations of oxygen vacancies, electrons in Mott memristors are more mobile, exhibiting a metallic state; whereas at low concentrations of oxygen vacancies, the movement of electrons is restricted, exhibiting an insulating state. The potential of Mott memristors in neural robotics is experimentally demonstrated for the first time in 2020^[39]. The NbO_x-based artificial spiking afferent nerve (ASAN) proposed by Zhang et al. is capable of converting analog signals from sensors into spiking signals in a spiking neural network, which is a key interface for building self-aware neural robots. It was demonstrated that the ASAN spiking frequency is proportional to the stimulus intensity when receiving stimuli of different intensities and decreases the spiking frequency when encountering extremely high stimuli, simulating the protective inhibitory behavior of biological neurons under strong stimuli. The high integration, low power consumption, dynamic threshold switching characteristics, and diffusion dynamics of the ASAN provide a new way of thinking about the simulation of biological neurons and brain-like computation. In 2023, Yuan et al. used Mott memristor to construct efficient LIF neurons and adaptive LIF neurons and integrated them into a decision-based long and short-term memory pulse neural network to enable detection tasks in arrhythmia classification and seizure detection^[40]. This provides further new ideas for the development of neural robotics engineering. To further improve the possibility of performance-enhancement, Li et al. attempted to integrate NbO_x Mott neurons with TiO_x-based artificial dendrites to construct a dendritic neuron unit that mimics biological neurons^[41]. This dendritic neuron unit is capable of efficiently processing spatiotemporal information by mimicking the nonlinear spatiotemporal filtering and integration functions of biological neurons. The group also demonstrated how to integrate this dendritic neuron unit with artificial synapse arrays, proving that the accuracy and energy efficiency of computation can be significantly improved by integrating dendritic functions in neuromorphic devices exemplified by Mott memristors.

Moreover, based on the internal thermal dynamics of Mott insulators, Kim et al. proposed a neuron integrated on a flexible organic substrate capable of encoding information within a single device through the temporal dynamics of the internally generated heat and processing the information through heat transfer between multiple devices^[42]. These Mott neurons demonstrate 18 bionic neuron behaviors and frequency-based damage sensing behaviors capable of spatiotemporal communication between devices via heat for graph optimization problems without additional electrical energy conversion. This work demonstrates the potential of utilizing natural thermal processes for computation, enabling functionally dense, energy-efficient, and radically novel proto-linguistics for hybrid physical computation. Park et al. increased the threshold voltage and reduced the thermal capacitance of the system by introducing gold nanodots into NbO_x Mott memristors, resulting in a 6-fold increase in the pulse amplitude of the memristors without the need for additional external circuitry components^[43].

Inspiringly, in 2021, Mott activated neurons are being implemented in full hardware as the activation function of ReLU in neural networks, verifying that Mott activated neurons achieve significant improvements in energy and area efficiency compared to traditional analog CMOS and analog-to-digital convertors^[44]. However, the stochastic nature of artificial neurons, as previously discussed, may amplify computational uncertainty and the complexity of decoding algorithms. This complexity can make the prediction and reproducibility of neural network outcomes challenging, potentially reducing the efficiency of information processing. To address these potential issues, it is prudent to consider strategies for dynamically adjusting thresholds and weights, thereby enhancing the robustness of neuromorphic computing systems against random spiking patterns. Further research should focus on developing neural network models tailored to specific spiking patterns, integrating experiments with computational simulations for algorithm-hardware co-optimization, and deepening our understanding of how random spiking patterns affect system performance. Notably, these random spiking patterns also represent potential high-quality entropy sources in the realm of hardware security. This insight paves the way for neuromorphic computational-cryptography cross-applications, emulating the random firing phenomena of biological neurons.

Volatile memristors based on conducting filaments transition from a high-resistance state to a low-resistance state by being applied a voltage that causes the active metal atoms (Ag^[45], Cu^[46], etc.) or oxygen vacancies^[47] in the memristor to migrate to form conductive filaments that spontaneously dissolve when the voltage is removed, returning the device to a high-resistance state. This property allows the memristor to mimic the dynamic behavior of biological neurons and does not require an additional reset phase and can be operated with a unipolar voltage, which simplifies circuit design and reduces energy consumption.

In 2018, Zhang et al. proposed an artificial neuron based on a threshold-switching memristor (Ag/SiO₂/Au)^[48], which successfully modeled the key functions of biological neurons, including all-or-nothing spike discharges, threshold-driven spikes, periods of ineligibility, and intensity-modulated frequency responses. Lin et al. constructed LIF neurons exhibiting good threshold-switching properties and stability using memristors with the structure Ag/TaO_x/Si^[49], and found that an increase in oxygen vacancies significantly reduces the threshold voltage for neuronal activation, the holding voltage, and the probability of occurrence of an undeserved period. This work provides a simple and effective strategy for developing artificial neurons with tunable performance.

After exploring memristor-based implementations of single LIF neuron models, our vision is extended to more complex and integrated neuronal structures. These structures are not only capable of modeling basic LIF functionality, but also enable richer neuromorphic computational functionality through innovative circuit design and material integration. Wang et al. innovatively combined dynamic capacitors with transistors to construct a novel neural transistor that mimics the soma and axon functions of neurons^[50]. In particular, the dynamic capacitor consists of a memristor and capacitor structure based on the structure Pt/SiO_xN_y:Ag/Pt, which is capable of voltage-driven capacitance switching to mimic the ion channel dynamics of neurons. This structure not only consumes low power because it expresses signals through voltage rather than current, but also theoretically consumes no power when a constant voltage is output from the signal source. In addition, this neural transistor is able to provide signal amplification to drive the propagation of the signal through the multilayer network, which is lacking in conventional CMOS neuromorphic circuits.

In addition to common oxide-based memristors, new materials have likewise been explored for the construction of LIF neurons. Ge_0.3Se_0.7-based artificial neurons generate electric potentials similar to those of nanocells when operating in a nonequilibrium state^[51]. The generation of these electrochemical potentials is related to the chemical potential gradient of the tried materials (SiO₂−GeS_x, GeSe_x−AgI), the size effect and the non-stoichiometric nature of the electrolyte, which together influence the dynamic behavior of the neuron model. A LIF bio-inspired neuron based on perovskite memristor exhibits superior amplitude-frequency response characteristics, showing potential for application as a filter^[52].

With the increasing amount of redundant information on the edge end-side in the IoT era, memristors-based artificial neurons are needed to be extended to more advanced cognitive tasks such as speech recognition, marking an important step in understanding and modeling the brain's processing of complex sensory information. Milozzi et al. implemented artificial neurons by mimicking the way the human auditory system processes sound, using conductive filaments in the memristors and applying them to sound recognition (Fig. 1)^[26]. The team further showed how these memristors can be used to implement logarithmic integration and pitch mapping of sound signals, which are key features that mimic the way the human cochlea processes sound signals. These volatile memristors exhibit random, dynamic, and adjustable conductance changes that can be used to simulate neuronal LIF behavior. Specifically, volatile memristors, upon receiving a series of voltage pulses (simulating neural impulses), change their conductance state according to the frequency and amplitude of the pulses, thus enabling the encoding of the time and frequency characteristics of the input signal. This approach not only demonstrates the potential of memristor in processing time-series signals, but also proves their effectiveness in building energy-efficient, high-density neuromorphic systems, especially for applications such as speech recognition.

The Hodgkin−Huxley (H−H) neuron model describes the nonlinear differential equations of electrophysiological phenomena in neuronal cell membranes, which directly reflect the opening and closing of ion channels in the cell membrane and their relationship to changes in membrane potential^[53]. In modeling H−H neurons, volatile memristors are used to simulate the dynamics of ion channels, including the opening and closing of ion channels and the resistance changes associated with them^[54]. Complex model configurations require that volatile memristors are able to exhibit complex dynamic behaviors to match the ion channel properties in the H−H model.

The neuristor, constructed on the foundation of Mott memristors, is capable of demonstrating all-or-nothing pulse emission and signal amplification—crucial attributes for simulating neuronal functionality^[25]. Research has illuminated the potential of VO₂-based Mott memristors in emulating the behavior of biological neurons and in neuromorphic computation. These memristors exhibit all three classes of excitability and a broad spectrum of dynamics observed in biological neurons, including various spiking behaviors such as sustained spiking, bursting, phasic spiking, and mixed-mode spiking, as depicted in Fig. 2(a)^[29]. Additionally, these neurons possess intrinsic stochasticity^{[29, 55, 56]}, enabling them to display stochastic spiking patterns akin to those of biological neurons. This characteristic could provide distinct advantages for executing complex computational tasks, such as Bayesian inference^[57]. For example, sustained spiking ensures that decision making or categorization is maintained under continuous input, bursting spiking plays an important role in the accumulation of evidence during decision making, and phasic spiking is relevant for processing time-sensitive tasks^{[29, 58]}. In addition, VO₂ artificial neurons have been preliminarily shown to be advantageous in terms of size and power consumption, and they exhibit good size and power scaling properties, which implies potential in maintaining bio-competitive energy efficiency and area^[29].

Facing with the big data and IoT era where computing power of common processing units is insufficient, the information processing efficiency and cost of integrated circuits are the two main challenges that neuromorphic computing systems need to focus on improving. Yi et al. discussed the scalability of VO₂ neurons, noting that since they were fabricated based on thin-film structures, it was theoretically possible to vertically stack repetitive pairs of synaptic cores to map directly to the cerebral cortex (Fig. 2(b))^[29]. This suggests the possibility of constructing fully-memristor neuromorphic computers that resemble the structure of the cerebral cortex. They also explored the thermal management issue of VO₂ neurons, noting that the critical temperature of the Mott transition can be increased by doping, thus alleviating the rigors of thermal management. In detail, increasing the critical temperature through doping allows the system to operate stably over a wider temperature range without worrying about exiting the desired phase state, enhancing the stability and reliability of neuromorphic computing systems, especially in environments where temperature fluctuations are common. Due to the breached thermal limitations, more VO₂ neurons can be integrated into a single neuromorphic computing system, significantly enhancing the scalability of the system. The epitaxial VO₂-based artificial neurons proposed by Yuan et al. achieve excellent device consistency through high crystal quality and calibrated resistors, and successfully encode a variety of physical signals such as pressure, light intensity, temperature, and curvature through multiple sensors^[59]. These neurons demonstrate efficient performance and wide application potential in modelling bio-sensory systems and implementing gesture recognition. This further demonstrates that artificial neuron systems based on Mott memristors not only have significant advantages in terms of biomimetic and computational efficiency, but also provide a new direction for the future development of neuromorphic computing technology. Besides, the area efficiency of integrated circuits is also a concern in the IoT era.

Moreover, Yang et al. successfully designed and fabricated a fully integrated H−H neuron circuit based on Mott memristor, which is not only capable of simulating a wide range of complex biological neuron firing behaviors, but also achieves dynamic threshold adjustment and steady-state plasticity based on resistance changes^[60]. These features based on Mott memristor enable the integrated circuit to show great potential in simulating biological neuron functions and constructing higher-order neuromorphic systems.

To sum up, the nonlinear dynamic properties of the Mott memristor, such as the metal−insulator transition and the Joule heating effect, enable it to better simulate voltage-gated ion channels in the H−H neuron model, especially in simulating the generation and propagation of action potentials, which is crucial for the implementation of bioheuristic and neuromorphic computations. In contrast, the diffusive memristor, although capable of exhibiting fast switching characteristics and high switching ratios, is inferior to the Mott memristor in simulating more complex neural dynamic properties than the LIF neuron model.

One artificial visual neuron proposed by Li et al. in conjunction with first-time spike timing coding, which relies on the timing of the first spike in a neuron to encode information, employs In₂O₃ photoelectronic synaptic transistors sensitive to ultraviolet (UV) light and encoding neurons based on NbO_x Mott memristors^[61]. The choice of In₂O₃ is due to its photovoltaic properties, allowing the neuron to convert light signals into electrical signals, thus mimicking the photoreceptive function of biological synapses. NbO_x memristors, on the other hand, are known for their threshold switching characteristics, which enable them to replicate the all-or-none firing behavior of biological neurons. The combination of the two has achieved a complete artificial visual neuron with ultra-high endurance, ultra-fast response speed (ultra-low power consumption), and excellent photosensitivity, where the In₂O₃ photoelectronic synaptic transistor serves as one photoreceptor and the NbO_x Mott memristor serves as one biological neuron. This combination can encode visual information at different spike frequencies and with precise, energy-efficient time-to-first-spike encoding, similar to the multiplexed data encoding scheme in biological vision systems. Considering the practicality of neuromorphic hardware, the trade-offs of this combination need to be discussed. For instance, whether the integration of In₂O₃ and NbO_x materials will introduce negative interface states, whether the ultra-high endurance of NbO_x memristors will conflict with the photoelectric conversion capability of In₂O₃ transistors, and whether the response speeds of the two units will limit each other, and so on. In particular, such artificial neurons exhibit a high degree of consistency with real-world data in hardware-implemented pulsed neural networks, which is a significant advantage for application scenarios such as steering and speed prediction of self-driving vehicles under complex conditions. Similarly, an ion-mediated spiking neuron based on the Mott memristor has been proposed to mimic the signalling of biological neurons^[62]. This neuron is capable of generating electrical signals in response to changes in ion concentration, specifically targeting changes in the concentration of sodium ions and spiking from physiological to pathological levels. This study amply demonstrates the great potential of Mott amnesia-based neurons to be used in engineering medicine.

As the physical size of integrated circuits shrinks to the atomic level, traditional CMOS technology is facing increasing challenges, including quantum tunneling effects, thermodynamic limitations, power consumption issues, etc.^[63]. Notably, Beyond CMOS focuses not only on improving computational performance, but also encompasses a comprehensive consideration of energy efficiency, integration, multifunctionality, and environmental impact, which is the key to the semiconductor industry's continued growth^[64−66]. Optoelectronic devices play a crucial role in Beyond CMOS, utilizing the high-speed transmission properties of photons and the logic control capabilities of electrons to achieve a significant increase in information processing speed and a substantial reduction in energy consumption.

The integration of optoelectronic neurons with reservoir computing enables the construction of efficient, high-throughput optoelectronic hybrid computing systems that excel in processing massively parallel information. These systems are particularly suited for real-time signal processing and high-speed data transmission applications. This combination also promotes the development of new neural network architectures capable of simulating more complex dynamic behaviors, closely mimicking the functionality of retinal neural systems, and providing new opportunities for intelligent vision-information processing. Sun et al. developed a photoelectronic reservoir computing system that efficiently processes complex information by directly exploiting the response of memristors to both electrical and optical signals (see Fig. 3)^[9]. The system’s dual-mode operation allows it to receive and process photoelectronic signals from the environment without additional sensors or data converters, similar to photoreceptors in biological vision systems. During a language learning task, a reservoir consisting of these memristive neurons transforms the input Korean sentence into a high-dimensional state that can be classified and recognized by the connected readout layer. The nonlinear properties and memory effects of the memristive neurons play an important role in capturing the time-series features of the input signal and achieving accurate classification. In a related study, Tan et al. proposed an all-photonic neuron-based reservoir computing system inspired by retina’s structure, which demonstrated remarkable performance in dynamic machine vision tasks by integrating photonic neuron arrays and readout networks^[67]. The system leverages the dynamics of the memristor array to embed past motion frames as hidden states into present frames and achieves accurate recognition of past motion and prediction of future motion through machine learning algorithms. This motion processing capability within the sensor eliminates redundant data streams and facilitates real-time perception of moving objects in dynamic machine vision. This research theoretically advances the academic understanding of the application of photonic neurons and memristor arrays in dynamic vision processing.

Although the combination of optoelectronic neurons with reservoir computing systems opens up new possibilities for sensing information processing, this integration faces many specific challenges and limitations. Firstly, the optoelectronic dual-mode operation of optoelectronic neurons increases the complexity of the system, requiring precise control and synchronization. Second, the long-term stability and reliability of these systems are key issues, as optoelectronic components can be affected by light fatigue and material degradation. Besides, consistency between different optoelectronic devices is crucial to enable large-scale integration, but there are still problems with variability in the manufacturing process. Notably, the nonlinear optical response and timing-dependent properties of memristors provide a basis for the development of new optical reservoir computation models, which, together with the wavelength dependence of different materials that can be used to differentiate and process optical signals of different wavelengths and modes, enable reservoir computing system to better simulate the dynamic behavior of the biological visual nervous system.

The time-surfaces represent the spatiotemporal context within a given spatial radius around the sensor's incoming events at a given time history^[68]. Further, time-surface neurons refer to an artificial neuron model capable of processing spatiotemporal information. This model is inspired by the way neurons process information in biological nervous systems and aims to enhance the extraction of spatiotemporal features by computational systems by simulating the dynamic behaviors of neurons. Wu et al. have attracted a lot of attention due to their novel method of extracting spatiotemporal audio features using dynamic time-surface neurons that are based on typical volatile memristors with unique material designs and configurations (Fig. 4)^[69]. The core component of the study, the 1TFT1R (one thin-film transistor and one memristor) features adjustable temporal response characteristics with tunable temporal kernel decaying, referring to the ability to control the rate at which the neuron's response to an input signal diminishes over time. This feature is pivotal as it allows the generation of time surfaces that facilitate efficient spatiotemporal feature extraction from audio event signals. This is achieved through the use of a memristor with a unique material stack, Au/LiCoO₂/Au, which forms a "sandwich" configuration. In this setup, the memristor's resistance can change dynamically in response to electrical signals, allowing the system to modulate the duration of the neuron's activation. The adjustability aforementioned is crucial for mimicking the varied response times observed in biological neurons and for fine-tuning the processing of time-sensitive data in neuromorphic applications. Moreover, it is worth noting that the 1TFT1R configuration provides absolute hardware-level tunability, allowing the system to tailor its performance to the specific requirements of diverse applications. In addition, the proposed memristor array configuration enables efficient encoding of audio signals and provides high-quality inputs for subsequent audio signal processing and recognition. In general, hardware-implemented time-surface neurons based on memristor arrays demonstrate the potential of memristor-based speech recognition systems for real-world applications, including significantly improved classification accuracy, computational efficiency, and robustness to noise. Moreover, this study addresses the significant challenge of energy consumption, a critical hindering factor in practical applications, particularly in von Neumann architectures that are limited by the "memory wall" issue. The proposed system improves information processing efficiency, thereby mitigating this bottleneck and paving the way for the practical application of neuromorphic systems.

The conclusion and future outlook are concisely illustrated in Fig. 5. While the architecture and algorithms of volatile memristors-based neuromorphic systems are crucial, the reliability of individual devices is a necessary prerequisite for realizing high-performance neuromorphic computing. Ensuring the reliability of a single memristor can mitigate the deterioration of the training effect caused by the abnormal variability induced by the unit in the neural network. Diffusive filament-based volatile memristors and Mott memristors are used as artificial neurons in neuromorphic computing hardware by virtue of their large switching ratios and excellent uniformity among devices, respectively. Nevertheless, improving the uniformity of diffusive volatile memristors remains an ongoing challenge, while the high operating currents of Mott memristors introduce reliability issues that limit the scalability of the hardware^{[28, 70]}. In addition to conventional endurance performance^[71], it is crucial to subject volatile memristors to cycling fatigue tests that replicate the membrane potentials of biological neuron to ensure functional reliability with reasonable stochastic spiking behaviours. Based on high-reliability volatile memristors, the four major aspects mentioned in this paper need to be considered extensively in terms of development directions and challenges as described below. To advance the development of volatile memristive systems, four key areas require in-depth consideration.

The advancement of neuromorphic computing through Mott and diffusive memristor-based artificial neurons offers significant potential, with cutting-edge research demonstrating their superior stability and reliability. These technologies enable the development of high-density, large-scale systems capable^{[28, 29, 72]} of handling complex computational tasks such as video action recognition and deep reinforcement learning^{[73, 74]}. Their stochastic and adaptive properties are particularly advantageous for simulating decision-making processes under uncertainty, which is essential for applications in industrial IoT and autonomous driving. In addition, edge computing requires rugged, compact hardware with sufficient storage, rich connectivity options, a wide power range, and hardware that meets performance requirements^{[75, 76]}. High-density integration and high reliability give it an edge in meeting these hardware requirements, especially in application scenarios that require high reliability and performance, such as industrial IoT and autonomous driving. Also due to the nanoscale nature of Mott memristors and diffusive memristors, they can be used to build high-density integrated neuromorphic systems with high efficiency, which opens up the possibility of realizing large-scale neural networks^[25]. However, ensuring the long-term durability and stability of these memristors is crucial to maintain their electrical characteristics without degradation, necessitating in-depth research into the physical mechanisms of Mott memristors, including material optimization and device structure refinement.

MRAM and PCM, as new types of memories, are mainly applied in the field of nonvolatile memory^{[13, 77]}. And Mott memristor and diffusive memristor, as emerging neuromorphic computing elements, have received more research attention in recent years due to their unique physical properties and potential to simulate the behavior of biological neurons^[78−82]. Although MRAM is compatible with CMOS technology, it has a small gap between its high and low resistive states, and it is not easy for it to require signal processing and analog-to-digital converter realize high-density integration. Meanwhile, PCM has relatively low durability and high write current, which limit its application in the field of artificial neurons.

Furthermore, in terms of extensive feasibility of applications, the combination of volatile memristor-based reservoir computing and time-surface neurons holds great potential for developing highly efficient and low-energy-consuming intelligent computing systems^[69]. The integration combines the high-dimensional dynamic properties of the reservoirs with the spatiotemporal information processing capabilities of time-surface neurons, effectively addressing tasks involving static and dynamically changing spatiotemporal data. However, significant challenges remain, including the design and implementation of large-scale memristor-based time surface neuron arrays that can be seamlessly integrated within reservoir models. Additionally, new fundamental algorithms (e.g., impulse neural networks) need to be developed to fully utilize the spatio-temporal properties of these neurons for enhanced information processing.

Last but not least, aligning volatile memristor-based artificial neurons with CMOS processes is vital for the large-scale application of neuromorphic computing. The advancement of low-dimensional materials and the integration of volatile memristors with silicon semiconductor technologies are key to overcoming the limitations of traditional CMOS technology^[83]. These efforts are crucial for developing neuromorphic systems capable of processing complex data with high spatial and temporal resolution, paving the way for the next generation of integrated circuits and intelligent information processing systems.