2.1 Graphene

Graphene, the first 2D material studied, possesses attributes such as high carrier mobility, elevated thermal conductivity, remarkable optical transparency, and robust thermal stability.⁴¹ Due to its zero bandgap, intrinsic graphene exhibits semi-metallic properties. Furthermore, graphene is known for its higher contact resistance and weaker vdW forces, both of which contribute to reduced operating currents when used as an electrode.⁵¹ For example, Yu et al. utilized a heterostructure consisting of graphene and MoS₂ to fabricate an optoelectronic transistor.⁵²

2.2 Transition metal chalcogenides

TMDs, such as MoS₂, WS₂, WSe₂, have attracted significant attention due to their large and adjustable bandgaps.^{41, 43} TMDs typically have an X-M-X structure, with three layers of atoms that exhibit notable anisotropy. In this structure, X represents a sulfur group element, while M denotes a transition metal element.⁴³ The optoelectronic adjustability, flexibility, and atomically smooth surfaces of TMDs make them suitable for high integration density and scalability, enabling rapid switching and reduced power consumption.^{53, 54} Most semiconductor TMDs have a bandgap that increases as the number of layers decreases. Importantly, the electronic band structure of TMDs has the potential to transition from an indirect to a direct bandgap, thereby enhancing the switch ratio of the fabricated devices.⁴³ Furthermore, their exceptional optical transparency and flexibility make them suitable for the fabrication of flexible transparent devices.

MoS₂ is a typical 2D material of the TMDs. It is abundantly available in nature and known for its exceptional properties, such as strong thermal stability and flexibility between layers.⁵⁵ Interestingly, when MoS₂ is reduced to a single layer, its bandgap transitions from an indirect to a direct configuration.⁵⁵ By exploiting this phase transition and its high switch ratio, researchers have successfully developed artificial synapses using MoS₂.⁵³ Li et al. observed switched characteristics in memristors as a result of the modulation of the MoS₂/metal Schottky barrier and the redistribution of vacancies within the MoS₂ channels.⁵⁶ Additionally, Liu et al. created optoelectronic synaptic devices by integrating 2D phosphorene with MoS₂.⁵⁷ Furthermore, Roy et al. found that the heterojunction formed by combining MoS₂ with graphene exhibits high sensitivity for photodetection and persistent photoconductivity.⁵⁸

2.3 Two-dimensional organic materials

2D organic materials offer great flexibility and can be easily processed in a solution. By adjusting their chemical structure and thickness, their response to external signals can be controlled, making them suitable for the large-scale production of low-cost and flexible synaptic devices. For instance, Fang et al. used solution epitaxy to create a photo-stimulated synaptic transistor using 2D organic semiconductor materials.⁴⁷ Additionally, the organic–inorganic hybrid perovskite has attracted significant research attention in the optoelectronic device field due to its advantageous properties, such as a prolonged carrier lifespan, diffusion length, and high light absorption coefficient.^{24, 59} Sun et al. used 2D layered perovskite to create synaptic photoconductors that can imitate synaptic behavior when exposed to light (Figure 2B).⁶⁰ They discovered that the photoelectric properties of 2D layered perovskites can be precisely adjusted through chemical composition engineering, presenting a new platform that fulfills the complexity necessary for neural morphology calculations.⁶⁰ Similarly, Liu et al. employed 2D metal–organic framework thin films as active layers to develop optoelectronic synaptic devices. These devices demonstrated fundamental synaptic activities such as LTP and STP, along with remarkable synaptic plasticity.⁶¹ In addition, Lai et al. combined 2D Ruddlesden–Popper phase layered perovskites with organic semiconductors to form type II heterojunctions. Their experiment revealed that these heterojunction systems can simulate fundamental properties of biological synapses, including EPSC, paired-pulse facilitation (PPF), and the transition from short-term memory (STM) to long-term memory (LTM) states.⁶² Moreover, Wang et al. integrated 2D inorganic and organic materials to fabricate an organic/inorganic hybrid heterostructure synaptic transistor, which not only demonstrated robust optical and electrical modulation but also exhibited STP and LTP characteristics (Figure 2C).⁶³

2.4 Other two-dimensional materials

In addition to the aforementioned 2D materials, there are other materials used in the production of optoelectronic synaptic devices. For example, Oh et al. utilized the optical memory phenomenon in hexagonal boron nitride (h-BN) to create a photoelectroactive synapse. This achievement was made possible by creating a heterojunction between h-BN and MoS₂. It is important to note that optical stimuli can either enhance or suppress changes in synaptic weights.⁶⁴ Organic–inorganic halide perovskites (OIHPs) demonstrate impressive optical properties and are widely considered promising candidates for artificial synapses.²⁴ The optical and electrical characteristics of 2D layered OIHPs can be readily manipulated by modifying their composition.²⁴ However, despite their exceptional optical properties, OIHPs only exhibit temporary changes in conductivity when exposed to light. Additionally, the conductivity rapidly decreases after the light is turned off due to the recombination of photoexcited charge carriers.²⁴ Park et al. developed an optical synapse by effectively combining a 2D layered OIHP with the transparent oxide semiconductor indium zinc tin oxide (IZTO). This optical synapse capitalizes on the high light absorption capacity of OIHP, enabling efficient stimulation by light spikes. The favorable band alignment between IZTO and OIHP also facilitates the efficient spatial separation of photoexcited charge carriers, resulting in nonvolatile changes in conductivity.²⁴ Furthermore, Xiong et al. successfully demonstrated the application of a 2D vdW heterostructure by combining black phosphorus (BP) with rhenium disulfide (ReS₂) in logic circuits (Figure 2D).⁶⁵

3.1 Two-dimensional heterostructures

In addition to individual 2D materials, heterostructures formed from these materials have been extensively researched due to their enhanced optoelectronic properties and diverse characteristics.^{48, 66, 67} The weak vdW interlayer forces enable the extraction of various 2D materials, allowing their reassembly into arbitrarily stacked heterostructures.⁶⁸ Heterostructures based on 2D materials exhibit excellent scalability and efficient photon–electron conversion efficiency, enhancing density.^{3, 20, 69} Compared to single materials, the energy band arrangement and carrier mobility of stacked 2D materials within heterostructures can be further adjusted, significantly improving the photoelectric properties.^{42, 53, 67, 70, 71} Through the precise design of the material arrangement and composition, it becomes possible to realize multifunctional applications of 2D materials. Moreover, this approach is expected to greatly enhance the functional integration density and reliability of optoelectronic synaptic devices for multifunctional applications.^{72, 73}

There are two types of heterostructures: vertical and lateral.^{42, 74} Vertical heterostructures involve the stacking of different 2D materials layer by layer in the vertical direction,⁴² with vdW forces between layers that are not limited by lattice matching.⁷⁵ For example, Zheng et al. produced high-quality and large-area WS_2(1–x) Se_2x/SnS₂ vertical heterostructures using a two-step chemical vapor deposition (CVD) method.⁷⁶ On the other hand, lateral heterostructures are formed by connecting different 2D materials on the same plane, typically through covalent bonds, resulting in sharp atomic-level transitions at the junction.⁴² For example, Duan et al. created compositionally controlled lateral heterostructures of MoS₂–MoSe₂ and WS₂–WSe₂ by manipulating the reactants in the vapor phase while growing 2D crystals. Their experimental results showed that the WS₂–WSe₂ heterojunction formed a lateral p–n diode with outstanding performance in terms of current rectification and photocurrent generation.⁷⁷ Additionally, Zhang et al. achieved the synthesis of lateral heterostructures of MoS₂–WS₂ and MoSe₂–WSe₂ using ambient-pressure CVD.⁷⁸

Heterojunctions are commonly classified into three types, namely type I, type II, and Type III, based on their electronic band structure, as depicted in Figure 3.^{79, 80} In type I heterostructures (Figure 3A), the minimum of the conduction and valence bands are both located within the same material.⁷⁵ Conversely, in type II heterostructures (Figure 3B), the minimum of the conduction band and the maximum of the valence band are situated in two distinct materials. Type II heterojunctions have garnered significant attention due to their remarkable light response characteristics and substantial rectification ratio. They effectively separate electrons and holes, thereby prolonging the exciton lifetime and regulating the photovoltaic and rectification properties.^{67, 74} For example, at the interface of MoS₂/GaSe, an inherent electric field arises due to the difference in energy between the valence and conduction bands. This leads to the formation of a space charge region. Additionally, potential barrier peaks are created to restrict the movement of electrons.⁸¹ When a forward voltage is applied to the source and drain electrodes, the strength of the inherent electric field decreases, leading to a decrease in the height of the potential barrier. This reduction facilitates the movement of electrons across the barrier, resulting in an increase in forward current. On the other hand, when a reverse voltage is applied, the strength of the built-in electric field intensifies. As a result, the energy of the electrons is no longer sufficient to overcome the potential barrier, leading to minimal current between the source and drain electrodes.⁵⁵ Type III heterostructures exhibit pronounced staggered gaps (Figure 3C) with no overlap between band gaps, leading to the formation of a discontinuous band gap. This specific heterostructure is extensively utilized in tunnel field-effect transistors.^{80, 82, 83}

The vast range of 2D materials is essential for the advancement of heterojunctions with various properties. For instance, in 2019, Wang et al. introduced a hybrid heterojunction synaptic transistor that combines MoS₂ and perylene-3,4,9,10-tetracarboxylic dianhydride (PTCDA).⁶³ The adjustable gate band configuration of the type II heterojunction enables the introduction of charge carriers into MoS₂. By employing mechanisms like separation, capture, and controlled release of carriers at the interface of the heterojunction, the vital functions of biological synapses are replicated.⁶³ In 2021, Luo conducted a study on synthesizing a heterojunction using gold nanoparticles (Au NPs) and MoS₂.⁸⁴ The addition of Au NPs to the heterojunction resulted in enhanced light absorption through plasmon effects. Consequently, the photoluminescence intensity increased, and a maximum carrier mobility of 20.04 cm²·V⁻¹ s⁻¹.⁸⁴ To address the problems of nonlinearity and conductivity saturation caused by the built-in electric field at the interface of 2D materials/electrodes, Hao et al. introduced a vertical heterostructure using thin films of MoS₂ and WO₃. The function of WO₃ is to serve as a stable anion reservoir, efficiently absorbing and releasing anions to enhance the linearity of conductivity changes.⁸⁵ In 2022, recognizing that laterally designed memory transistors pose challenges for integration, Wang et al. devised a vertical vdW heterostructure composed of graphene/CuInP₂S₆ (CIPS)/MoS₂/h-BN.⁸⁶ Serving as asymmetric electrodes, graphene and MoS₂ correspond to pre- and post-neurons, respectively. Experimental findings indicate that the bias voltage of the resistance switch closely aligns with the coercive voltage of the CIPS flake. This alignment suggests that the reversal of ferroelectric polarization drives the behavior of a resistance switch.⁸⁶ Recently, Syed et al. demonstrated the modulation of excitation and inhibition using a type II 2D heterostructure opto-memristive neuron. This allows for the nonlinear integration of excitatory and inhibitory postsynaptic potentials within a single device.⁸⁷

3.2 Working mechanisms

Optoelectronic synaptic devices that utilize 2D materials primarily function by capturing and releasing charges either within the materials themselves or at their interfaces. This enables their functional capabilities.³ These charge transfer mechanisms are influenced by various factors, including the density of charge capture points at the interface of the 2D material/substrate.^{3, 26, 48, 60} Figure 4A depicts the classification of optoelectronic synaptic devices powered by 2D materials. These devices can be grouped into several categories, such as intrinsic 2D material channels with or without functional oxide interfaces, 2D materials integrated with nanoparticles or quantum dots, 2D vdW heterostructures, and 2D ferroelectric materials.³ Among them, optoelectronic synapses based on 2D heterojunctions mimic synaptic behavior through the charge capture/release process that occurs at the interfaces of specific 2D layers within heterostructures.^{3, 88}

Using the heterojunction consisting of 2D MoS₂ and PTCDA as an example (Figure 4B), the combination of these materials creates a type II heterojunction. The band alignment of this particular heterojunction can be adjusted by the gate, allowing for the injection of carriers into MoS₂ through both electrical and optical means. Within the interface of the heterojunction, carriers can undergo separation, capture, and gradual release, giving rise to synaptic behavior such as STP and LTP.⁶³ When a specific gate voltage is applied, carriers within the MoS₂/PTCDA synaptic transistor heterojunction reach dynamic equilibrium, maintaining a stable drain–source current (I_ds).⁶³ Upon application of a positive gate voltage pulse, more electrons accumulate in the channel of MoS₂, resulting in an increase in I_ds. At the same time, the electric field at the gate weakens and the band bends, causing some electrons to transfer to PTCDA. Consequently, the number of electrons in MoS₂ decreases, leading to a decrease in I_ds. Since the migration rate is low, electrons in PTCDA slowly diffuse back to MoS₂, gradually restoring I_ds to the reference level within a few seconds, resulting in IPSC (Figure 4C–E). Conversely, when negative electrical pulses are applied, the direction of charge transfer reverses, resulting in EPSC, as illustrated in Figure 4F. Upon being stimulated by a laser pulse, a significant number of unbalanced electrons, created by the laser pulse, migrate from PTCDA to MoS₂. This migration results in a noticeable increase in the PSC as shown in Figure 4D. Once the laser pulse stops, the electrons gradually return to PTCDA, causing the corresponding PSC to return to the reference level. The behavior demonstrated in Figure 4G is a typical indication of STP known as PPF. Importantly, the process of electron transfer occurs in both directions within this inorganic–organic hybrid heterojunction, although with different time constants.

Various 2D materials exhibit distinct optical and electrical properties, and the configuration of devices also varies. This can result in different directions and speeds of electron transfer within heterojunctions when exposed to optical or electrical stimulation.⁸⁹ Precision in the design of device structures is essential for controlling the movement and recombination of photogenerated carriers.¹⁶ Enhanced light absorption and responsiveness allow for the creation of synaptic devices with heightened sensitivity and reduced power consumption.⁵⁵

3.3 Structures of optoelectronic synaptic device

In contemporary times, optoelectronic synaptic devices are categorized into vertical and lateral structures, as well as two-terminal and three-terminal configurations (Figure 5).^{90, 91} Table 1 demonstrates the impressive performance of optoelectronic synaptic devices based on 2D materials in achieving essential biological synaptic functions. By optimizing and balancing the materials, the device's performance can be significantly enhanced, resulting in features such as low power consumption and a higher PPF index. It is important to note that the PPF index plays a critical role in identifying and decoding time-resolved information. A higher PPF index indicates a greater memory capacity for detected optical signals, thus suggesting significant potential for minimizing information loss.^{7, 92, 93} In two-terminal configurations, the upper and lower electrodes represent the presynaptic and postsynaptic neurons, respectively. The external stimulus signal replicates the action potential of the presynaptic neurons, while the change in resistance value of the device indicates changes in synaptic weights.^{18, 94} Despite the simplicity of the two-terminal structure, which facilitates large-scale integration and compatibility with conventional CMOS technology, it often experiences instability. Additionally, its ability to simultaneously transmit signals and perform learning functions is often limited.^{16, 51, 95-98} In contrast, three-terminal structures demonstrate superior stability and controllability. They can collaborate with multiple signals to achieve scalable and low-power synaptic plasticity, despite facing challenges in large-scale production.^{33, 37, 42, 99} By adjusting the voltage on the third gate electrode, synapses with a three-terminal structure can independently modify the device's characteristics. For example, the gate electrode enables individual adjustment of synaptic weights, facilitating simultaneous signal transmission and learning functions.^{90, 95, 100, 101} Devices with this structure, commonly known as memory transistors, combine the characteristics of both memristors and transistors.¹⁸

3.4 Preparation methods

The ideal optoelectronic synapse is expected to have a fast response speed, low power consumption, a high signal-to-noise ratio, and the ability to simulate various synaptic functions such as STP, LTP, and spiking-timing-dependent plasticity. Achieving these functions depends on the efficient generation, transport, and recombination of photogenerated carriers within the device.¹⁶ The performance of optoelectronic synapses, especially those based on 2D material heterostructures, is significantly affected by the specific configuration of the heterostructure formed through the interaction of 2D materials.²⁰

Currently, the main methods used to prepare 2D materials include both top-down peeling and bottom-up growth approaches. Top-down methods involve mechanical and solution peeling techniques. Mechanical peeling, which uses tape, and solution peeling, which involves reacting the sample with a solvent to extract a 2D film, are both fast and straightforward.^{18, 42} In contrast, the bottom-up growth method involves directly depositing materials onto a substrate to create 2D films. This method includes various techniques, such as molecular-beam epitaxy, CVD and metal organic chemical vapor deposition. These approaches offer the advantage of producing uniform and high-quality thin films over large areas.⁴³ The fabrication of heterojunctions in 2D materials is mainly achieved through mechanical transfer stacking and epitaxial growth techniques. In the mechanical transfer stacking approach, 2D materials obtained through exfoliation are carefully layered onto another material with the assistance of a microscope. This method, known for being cost-effective and having minimal defect density, allows for the versatile integration of multiple 2D materials with different material compositions and stacking orientations.⁵⁵ In contrast, epitaxial growth involves using a 2D material as a substrate to facilitate the growth of another material on top of it.⁵⁵

Overall, further development and improvement of preparation processes are necessary to achieve the continuity and uniformity of 2D materials at larger sizes. Additionally, as the prepared 2D films must be transferred to the target substrates, it is imperative to ensure that the films remain intact and free of contamination during the fabrication of optoelectronic synaptic devices. Therefore, research on 2D material preparation methods and heterojunctions is significant for enhancing the performance of artificial synapses.¹⁰⁵

4.1 Image recognition

In traditional machine vision systems, there is a clear separation between the front-end image sensor and the back-end processor. The processor consumes significant energy and time in processing and recognizing the image data captured by the sensor.⁹¹ Optoelectronic synaptic devices offer a solution for emulating human visual functions and constructing machine vision systems through the creation of large-scale device arrays. These devices have the potential to streamline circuit design and reduce interconnection energy usage, thereby improving the overall efficiency of machine vision systems.^{6, 31, 90, 92, 107-109} For example, biological cone cells possess remarkable abilities to distinguish between strong light and colors, and they play a crucial role in the human visual system. Seo et al. integrated artificial synapses and light sensors onto the h-BN/WSe₂ heterostructure, creating one artificial cone cell for every three neurons.²⁹ Through this integration, they developed an artificial optical neural network (ONN) with 28 × 28 arrays, successfully simulating the color-mixed pattern recognition abilities of the human visual system.²⁹ By performing pattern recognition tasks, such as identifying target numbers from color-mixed images and optimizing synaptic weights through training on various datasets, the ONN achieved a recognition rate of over 90%. This recognition rate was significantly higher than that of neural networks composed of synaptic devices that lacked optical sensing functions.

To further reduce the power consumption, Zhu et al. introduced a new approach to create an ultra-low power artificial photonic synapse. They achieved this by using a BP/Cadmium sulfide (CdS) heterostructure, with CdS as the photosensitive layer and BP as the channel layer. The device was used to build a three-layer fully connected optoelectronic neural network capable of recognizing Modified National Institute of Standards and Technology (MNIST) handwritten digits (Figure 6A). With an impressive image recognition accuracy of up to 94.1%, this device also has remarkably low energy consumption ranging from 0.43 to 8.9 fJ per light spike and 25 fJ per electrical spike.⁸⁸

By using an artificial neural network (ANN), the neural morphological visual system can process visual data efficiently, similar to the visual cortex of the human brain. This system also reduces the amount of hardware and power consumption.¹¹⁰ Zhou et al. developed an image recognition system by utilizing a two-terminal optoelectronic resistive random access memory (ORRAM) synaptic device in ANN (shown in Figure 6B).¹¹⁰ The use of ORRAM preprocessing led to an impressive image recognition rate of 0.986, as demonstrated by the results. Yu et al. proposed a mechano-photonic artificial synapse based on a graphene/MoS₂ heterostructure. The device consists of a heterojunction-based optoelectronic transistor and a triboelectric nanogenerator. The optoelectronic synaptic behavior, such as persistent photoconductivity, postsynaptic photocurrent, and photosensitivity, can be effectively controlled by manipulating the charge transfer/exchange between the graphene and MoS₂ layers. They also simulated a three-layer ANN using the device to evaluate the impact of mechanical plasticity on image recognition accuracy. Through training, they discovered that the highest recognition accuracy could reach 92%. This type of mechano-photonic artificial synapse can easily be extended to multimodal sensory neurons, interactive photon synapses, artificial visual-somatosensory associative systems, and so on.⁵²

Optoelectronic synapses can be integrated not only with ONN and ANN for image recognition but also with convolutional neural networks (CNN) and spiking neural networks (SNN). Lee et al. aimed to improve the accuracy of CNN image recognition by using a photoelectric synaptic array with color discrimination function (Figure 6C).¹¹¹ Their experiment showed that the CNN, using its designed optoelectronic memristor (4-bit RGB image resolution), achieved an inference accuracy of approximately 94% (bidirectional) and 96% (unidirectional) when tested on the CIFAR-10 dataset. Additionally, the preprocessed input signal demonstrated a significant enhancement in RGB color resolution. Wang et al. employed SNN to evaluate optoelectronic devices based on the 2D b-AsP/MoTe₂ vdW heterostructure (Figure 6D), and the inference accuracy of the MIR MNIST dataset encoded by the device exceeded 96%.¹¹²

Islam and his colleagues have developed an optoelectronic synaptic device that can detect, store, and manipulate optical data across a broad range of wavelengths, from UV to infrared. This allows for comprehensive optical information processing within the electromagnetic spectrum.¹¹³ As a result, this device is capable of constructing and recognizing multicolor patterns. To replicate the various functions of the human visual system, Zhang et al. developed a heterogeneous neuromorphic imaging system. This system is based on the concave hemispherical human retina and utilizes synaptic transistors with exceptional sensitivity.¹¹⁴ The system is composed of a 10 × 10 transistor array that transforms monochromatic radiation into stereoscopic light with different depths using prefabricated masks. This simulation mimics the 3D imaging behavior of the retina and demonstrates significant potential in the fields of visual restoration and artificial intelligence. Choi et al. constructed a curved neural morphology image sensor array based on a MoS₂-organic heterostructure, which can achieve efficient image acquisition and recognition processes.¹¹⁵ This sensor array enables efficient image acquisition and recognition processes. By combining optoelectronic synaptic devices with various materials and structures in different neural networks, these devices can exhibit diverse performances, such as multiband light response, multicolor pattern recognition, and 3D imaging. These capabilities have promising applications in various fields.

4.2 Neuromorphic wearable electronic devices

Optoelectronic synaptic devices with neuromorphic computing capabilities have attracted considerable interest due to their potential application in wearable devices in the field of flexible electronics. 2D materials have been investigated for their thinness and impressive optoelectronic properties, making them excellent candidates for optoelectronic synaptic devices. These materials offer the potential for devices that are reliable, energy-efficient, and flexible.^116-121 Additionally, optoelectronic synapses based on 2D materials enable the construction of wearable electronic devices with biocompatibility and portability.³⁰

Visual injury receptors detect harmful stimuli or stimuli exceeding a certain intensity level. The optic nerve network then transmits pain signals to the cerebral cortex for advanced processing to avoid potential damage. Feng et al. successfully established a retinal-inspired artificial visual electronic skin with pain perception ability using flexible transistors. This innovative device can perform essential functions such as light response, memory retention, and visual pain calculation.¹²² Furthermore, it can execute painful photosensitization in various space–time color modes to prevent secondary visual damage (Figure 7A,B).¹²² After the activation of a visual nociceptor, the central nervous system receives, integrates, and transmits the pain information. Moreover, the skin of cephalopods serves as a flexible cognitive organ capable of perceiving, recognizing, and processing nerves in a distributed manner through complex biological processes. Leveraging this principle, creating an elastic and biaxially stretching device embedded with distributed neural and cognitive functions holds significant potential in fields such as skin repair surgery and wearable devices. Shim et al. utilized synaptic transistors made of elastic materials to develop a neural morphology recognition skin device. This device combines biaxial stretching and synaptic neural morphology functions to achieve image recognition, LTM retention, and fault tolerance.¹²³ Therefore, the advancement of wearable electronic devices in intelligent healthcare can significantly promote the application of artificial electronic skin.

Lee et al. successfully demonstrated optical wireless communication through the conversion of patterned optical stimuli into potential synaptic responses using organic optoelectronic artificial synapses. By assigning a distinct EPSC to each English letter, the device enables the generation of unique synaptic responses. Furthermore, when integrated with artificial muscle actuators, this device enables the development of an optical sensory neuromuscular electronic system, which presents a promising solution for enhanced optical wireless control of soft electronic devices.¹²⁴ Hou et al. introduced a new two-terminal optical synapse that uses pyrenyl graphdiyne/graphene/PbS quantum dots. Interestingly, even when bent and folded, the LTP and LTD characteristics of these optical synapses remain mostly unchanged, showing great fault tolerance. The straightforward structure and low-dimensional properties of this synapse make it ideal for incorporation into wearable electronic devices (Figure 7C).³⁰ Then, Li et al. fabricated optoelectronic synaptic devices on flexible substrates using MoS₂ material through a simple gate modulation technique (Figure 7D).¹²⁵ By using paired stimuli consisting of light and voltage pulses to integrate optoelectronic synapses and replicate the classical conditioned reflex Pavlovian dog experiment, they successfully achieved associative learning behavior. Hwang et al. conducted a study in which they created flexible artificial synapses with 2D heterojunctions. The results of their study showed that these synapses had excellent long-term enhancement and long-term cycling endurance on curved surfaces with a diameter larger than 10 mm or undergoing more than 700 bends.¹²⁶ Additionally, the energy consumption per single peak was measured at 0.55 nJ. The researchers aimed to build a neural network and observed that the flexible synapses maintained a high level of recognition accuracy even under curved conditions. This promising outcome suggests a potential avenue for the advancement of wearable devices. Notably, the development of large-scale flexible optoelectronic devices highlights the potential application of human neuromorphic systems in wearable electronic products.

4.3 Logical operations

Using optoelectronic synaptic devices made of 2D materials, researchers have successfully performed logic operations by taking advantage of the changes in conductance caused by the frequency, quantity, and intensity of input stimuli.³³ These capabilities make it possible to build large-scale computing systems and carry out computationally complex tasks with the help of collaborative optical inputs.^{30, 33}

Hou et al. demonstrated optical logic functions that are modulated by inputting light pulses of different wavelengths, using optical synapses with vertical heterojunctions.³⁰ In Figure 8A, signals A and B correspond to “0” and “1” for light off and on, respectively. To determine the logic function, modulation of the output current level is performed using the modulation input. The “OR” function is achieved when either input A or input B triggers it, surpassing the threshold, provided the modulation input is off. The “AND” function is attainable only when both inputs A and B are applied simultaneously, and the modulation input is activated, enabling the output current to surpass the threshold (Figures 8B,C). The “NOR” function is realized when either input A or B lowers the output current below the threshold once the modulation input is deactivated. Conversely, when the modulation input is activated and inputs A and B are applied simultaneously, the output current falls below the threshold, achieving “NAND.” Furthermore, the “XOR” function is achieved when the wavelengths of inputs A and B differ.

In 2023, Fu et al. introduced a new type of dual-terminal photomemristor that consist of a graphene/MoS_2-xO_x/graphene structure.¹²⁷ These devices are able to perform logic operations by responding to light, which is achieved through distinct memristor schemes involving electrical and optical stimulation. This approach includes a comprehensive logical framework that incorporates material implication and false operations.¹²⁷ As a result, photomemristors can serve as both logic gates and storage units, enhancing the functional versatility of neural networks.

4.4 Neuromorphic computing systems

A neuromorphic computing system that can emulate the information processing functions of the human brain shows promise in addressing the bottleneck faced by conventional computing systems that use von Neumann architecture. This system takes advantage of the parallel operation of synapses, which serve as both computing and memory units, allowing for high-speed operation and low-power consumption in information processing.^128-131 Synaptic devices are the fundamental elements in this neuromorphic computing paradigm, storing input information as electrical signals and processing data through synaptic weight updates.³³ The amount of effective conductivity states, as well as the linearity and symmetry of weight update behaviors, are critical factors that influence the accuracy and reliability of recognition in neuromorphic computing.¹³² Optoelectronic neural synapses offer advantages such as low crosstalk and robust interconnections, making it easier to achieve neuromorphic computation.^{6, 133}

Hou et al. implemented a three-layer ANN that included a hidden layer and utilized vertical heterojunction optical synapses.³⁰ In this network, the synaptic weight is determined by the conductivity difference between two equivalent optical synapses. They used supervised learning to train the network on handwritten numerals sourced from the MNIST. To assess the network's fault tolerance, different levels of noise were introduced (Figure 9B). Remarkably, the network shows robust fault tolerance, maintaining an accuracy of over 80% even when subjected to a noise ratio of 60% (Figure 9C). Additionally, the network demonstrates the strong fault tolerance in both bent and folded states (Figure 9D,E).

Han et al. incorporated ultra-thin h-BN into graphene/CsPbBr₃ quantum dots to serve as a carrier-regulating layer, thereby fabricating a three-layer heterostructure synaptic transistor.⁹³ The application of continuous light pulses leads to a gradual increase in conductivity countered by a subsequent decline induced by continuous electrical pulses (Figure 9A). The two-layer neural network connected with this device achieves a recognition accuracy of approximately 91.5% for handwritten digits. This network comprises 10 output neurons, 300 hidden neurons, and 784 input neurons. The input image is converted into an output value through a weight matrix. Subsequently, the weight matrix is adjusted through backpropagation of the error between the output value and the image label value, thus enhancing the accuracy of digital recognition. Furthermore, Fang et al. developed a three-layer ANN that uses optoelectronic dual modulation of dual-terminal synaptic devices. They achieved a facial recognition accuracy of up to 95%.⁹⁸ This highlights how the continuous improvement in recognition accuracy in neuromorphic computing systems is due to advancements in the materials and structures of optoelectronic synaptic devices.

Additionally, Li et al. developed a metal/semiconductor/metal (M/S/M) neuromorphic device using 2D metal sulfides as semiconductors. The goal was to create a convolutional neuromorphic network (CNN) for image processing and object detection.¹³⁴ By controlling the concentration of sulfur ions in the M/S junction, the device can modify the height of the Schottky barrier. This modification is achieved by intentionally creating sulfur vacancies and controlling their movement with an electric field. Importantly, the M/S/M device can disrupt the symmetry of the electric potential, making the architecture for object position detection in neuromorphic vision networks simpler.