In the era of big data and artificial intelligence (AI), there is a pressing need for prompt sensing, memory, and preprocessing of vast amounts of data generated in our daily life^{[1, 2]}. Traditional machines based on Von Neumann architectures face significant challenges in efficiently managing these extensive datasets due to the segregation of memory and processing units, which leads to issues such as data redundancy, access delays, and high energy consumption^{[3, 4]}. To address these challenges, there is an active development of neuromorphic computing systems based on artificial synaptic devices, aiming to harness the advantages of the human brain, including in-memory computing, massive parallelism, and low energy consumption^[5−7]. By leveraging these capabilities, neuromorphic computing systems hold great promise for effectively processing the ever-expanding big data generated by various optoelectronic devices within the Internet of Things^{[8, 9]}.

In human neural system, information is transmitted through synapses. Therefore, artificial synaptic devices that mimic synaptic functionalities are potential building blocks for artificial neural networks. Initially, synaptic devices utilized electrical pulses as stimulating signals to emulate synaptic functionalities due to their relative ease of fabrication, rapid switching speeds, and high integration density^[10]. However, the advent of optogenetics in neuroscience has facilitated the incorporation of light into synaptic devices, providing advantages such as ultrafast processing, high bandwidth, and reduced crosstalk in comparison to electrical signals^[11]. In addition, optically stimulated synaptic devices are capable of integrating visual sensing, signal processing, and image memory^[12]. This is significant for neuromorphic computing, considering that humans primarily acquire information through visual input. Therefore, substantial efforts have been devoted to develop optically stimulated synaptic devices utilizing various materials, including Si nanocrystals^{[13, 14]}, metal oxide^{[15, 16]}, organic semiconductors^{[17, 18]}, perovskite^{[19, 20]}, and two-dimensional (2D) materials^[21]. These devices have successfully emulated fundamental synaptic functionalities and neural activities at room temperature.

As humanity increasingly engages in the exploration and interaction with extreme environments characterized by high temperatures and intense radiation, the urgency to replace human labor with AI systems becomes increasingly pronounced^[22−25]. This poses a challenge regarding the heat resistance of artificial synaptic devices^[26]. For example, synaptic devices integrated onto the detection robots operating on the surface of Vernus must be capable of enduring temperatures exceeding 300 °C^[27]. However, the heat resistance of most synaptic devices based on materials such as metal oxide^[28−32], organic semiconductors^[33−38], perovskite^{[39, 40]}, and 2D materials^{[41, 42]} is insufficient to fully meet the demands of neuromorphic computing and intelligent applications in such extreme environments.

Currently, high-temperature synaptic devices are being developed using wide bandgap semiconductors such as SiC^{[43, 44]}, GaN^[45], GaO_x^[46], and GaPS₄^[47] due to their better thermal stabilities. Among these materials, SiC is considered to be an ideal candidate for high-temperature synaptic devices due to its high electron mobility^[48], superior thermal conductivity^[49], excellent thermal stability^[50], and well-controlled n-type and p-type doping^[51]. Moreover, the diameter of commercially available 4H-SiC wafer with low defect state has surpassed 8 inches, indicating that SiC is conductive to the fabrication of highly integrated synaptic devices in wafer scale^[52]. It has already been utilized as the active layer in ultraviolet photodetectors capable of operating at temperatures above 450 °C^[53−56].

In addition, both electrically and optically stimulated synaptic devices based on SiC have been developed to emulate synaptic functionalities and achieve neuromorphic applications even at high temperatures. The active layers of these devices comprise amorphous or single-crystal SiC thin films^[57−59] and SiC nano wires^{[43, 60]}. This review aims to summarize the recent advancements in SiC-based synaptic devices. A schematic diagram outlining the contents of this review is depicted in Fig. 1. First, an in-depth discussion is conducted regarding the categories, working mechanisms, and structural designs of these devices. Subsequently, several application scenarios for SiC-based synaptic devices are presented. Finally, a few perspectives and directions for their future development are outlined.

SiC-based synaptic devices utilize active layers comprising amorphous, single-crystal thin films^[57−59] or one-dimensional nano wires^{[43, 60]}. While amorphous SiC films offer simpler fabrication, single-crystal thin films, despite their higher cost and processing complexity, exhibit superior charge carrier mobility and thermal stability. SiC nanowires, with their inherently high surface-to-volume ratio and diameter-confined domain effects, facilitate axial electron transport, leading to enhanced photoelectric conversion efficiency, improved charge transport, and reduced leakage current. However, challenges remain in achieving uniform, scalable nanowire growth and the large-scale integration of nanowire-based optoelectronic synaptic devices^[61]. Electrically stimulated SiC synaptic devices have been demonstrated using both amorphous and single-crystal thin films, whereas optically stimulated devices predominantly utilize single-crystal and nanowire SiC architectures.

In biological synapses, the action potential on a pre-synaptic neuron facilitates the release and migration of neurotransmitters from the pre-synaptic membrane to the post-synaptic membrane. In electrically stimulated synaptic devices, electrical pulses serve as the pre-synaptic neuron and modulate channel conductance so as to emulate synaptic functionalities. The underlying modulation mechanisms primarily include ion migration^{[57, 58]}, ferroelectric polarization^[41], and the capture and release of charge carriers by defects^[62]. SiC-based electrically stimulated synaptic devices typically utilize electrical pulses to induce the migration of metal ions, leading to the formation and rupture of metallic conductive filaments within the SiC active layer. This process effectively modulates the device’s conductivity and enables the exhibition of synaptic characteristics.

For instance, in 2021, Yan et al. developed a 4H-SiC memristor based on the structure of Ag/4H-SiC/Pt, with Ag and Pt electrodes connected to the positive terminal and ground of the device, respectively (Fig. 2(a))^[57]. The resistance states of the device could be adjusted through the application of varying electrical pulses. When an increasing positive voltage was applied to the Ag electrode, the conduction path of Ag continued to grow until an Ag atom bridged to the Pt electrode, resulting in a transition from a high-resistance state (HRS) to a low-resistance state (LRS) (Fig. 2(d)). Conversely, if the applied electric field was diminished, the conductive filament would break, and the device would revert to the HRS. In addition, continuous application of positive or negative voltage would either thicken or dissolve the conductive filament, endowing the device with bipolar resistance switching capabilities and enabling it to simulate the synaptic functions, such as spike-timing-dependent plasticity (STDP) and paired-pulse facilitation (PPF) (Figs. 2(b) and 2(c)).

In 2022, Huang’s group reported a Si₇C₃ memristor based on the structure of Cu/Si₇C₃/W, wherein the Cu top electrode served as the active metal layer and the W electrode functioned as the inert electrode (Fig. 2(e))^[58]. Upon the application of positive electrical pulses to the memristor, Cu was chemically dissolved into the Si-rich SiC thin film, resulting in the formation of conductive filaments (Fig. 2(f)). This process enhanced the memristor conductance and enabled modulation of the synaptic weight. As illustrated in Figs. 2(g)−2(i), continuous electrical pulses with varying intervals, amplitudes, and durations were applied to the memristor, which achieved the effective mimicking of the spike-frequency-dependent plasticity (SRDP), spike-voltage-dependent plasticity (SVDP), and spike-duration-dependent plasticity (SDDP), closely resembling the functionalities of biological synapses.

In addition to reactive metals such as Ag and Cu, Al^[63] and AgI^[64] are also promising electrode materials for the fabrication of conductive filament-type memristors utilizing SiC as the active layer.

In optically stimulated synaptic devices, light is employed in conjunction with electrical pulses to modulate the channel conductance and synaptic weight. The underlying modulation mechanisms include the capture and release of carriers by heterostructures^{[37, 65, 66]} or defects^{[46, 67]}, ionization and deionization of oxygen vacancies^[68], as well as optically induced phase transition in active materials^[4]. The incorporation of light as a stimulating signal offers several advantages, including a wide bandwidth, reduced crosstalk, low energy consumption, and enhanced signal transmission speeds^{[7, 8, 12]}. In addition, optically stimulated synaptic devices endow neuromorphic systems with a range of versatile applications, such as wireless transmission, image memory, polarimetric measurements, motion recognition, and stereo vision^{[5, 69−71]}. SiC-based optically stimulated synaptic devices are normally achieved based on the mechanism of capture and release of photogenerated carriers by heterostructures or defects.

Heterostructures are commonly employed to achieve non-volatile performance of synaptic devices^{[47, 72]}. The trapping of photogenerated holes and electrons in distinct materials enables these carriers to be stored for extensive periods without undergoing immediate recombination, which consequently results in a gradual decay of the photocurrent and the storage capabilities of the devices^{[73, 74]}. To date, organic semiconductor or metal oxide have been utilized to form heterojunction with SiC, thereby enabling the synaptic behavior of SiC-based synaptic devices.

For instance, in 2023, Yang et al. designed an ultraviolet optoelectronic synaptic transistor based on the heterostructure of 4H-SiC/PVK/P3HT (Fig. 3(a))^[59]. Upon exposure to light stimuli, the photogenerated electrons were effectively trapped in the 4H-SiC layer, while the photogenerated holes were transported and injected into the P3HT channel layer (Fig. 3(b)). PVK that acted as the electron blocking layer and hole transport layer significantly elongated the lifetime of photogenerated holes in the P3HT channel and facilitated the achievement of non-volatility of the synaptic transistor. The efficient separation of the photogenerated carriers at the heterojunction interface led to a distinct photogating effect, which contributed to the device’s non-volatility and light-stimulated synaptic performance. The device successfully emulated a variety of biological synaptic functionalities and neural activities, including PPF, SDDP, spike-number-dependent plasticity (SNDP), SRDP, and learning-forgetting-relearning behavior (Figs. 3(c)−3(g)). Moreover, the EPSC triggered by 400 optical spikes didn’t decay completely even 10⁴ s after the cessation of stimulation, thereby proving the superior retention capability and formation of long-term memory (LTM) in the synaptic transistor (Fig. 3(h)).

In 2023, Shen’s group fabricated an optoelectronic synaptic device based on the structure of ITO/PMMA/3C-SiC nano wires/ITO (Fig. 4(a))^[60]. As the irradiation time increased, a significant number of photogenerated carriers were created in the deeper regions of the 3C-SiC nano wires (Fig. 4(b), stage Ⅰ & Ⅱ). The photogenerated electrons became trapped within the conduction band due to the band alignment mismatch between 3C-SiC and PMMA, leading to a pronounced photogating effect and an extended detrapping time for the trapped electrons after the cessation of light stimulation (Fig. 4(b), stage Ⅲ & Ⅳ). As a result, the device was capable of mimicking various synaptic functionalities and neural activities, such as SDDP, SNDP, and classical conditioning of Pavlov’s dog (Figs. 4(c)−4(e)).

However, the utilization of organic semiconductors has restricted the high-temperature applications of aforementioned SiC-based synaptic devices due to the issues such as thermally induced oxidation and disruption in molecular packing within organic materials at elevated temperatures. In contrast, Wei’s team selected the oxide semiconductor NiO to construct 3C-SiC@NiO core-shell nanowire networks, thereby developing an optoelectronic memristor capable of operating at high temperatures up to 200 °C (Fig. 5(a))^[43]. The application of positive or negative voltage facilitates the rupture or formation of O²⁻ conductive filaments between the ITO electrode and NiO layer, subsequently decreasing or increasing the device’s conductance. Upon light stimulation, photogenerated electrons and holes accumulated in the conduction band of the 3C-SiC layer and the valence band of the NiO layer, respectively (Fig. 5(b)). The efficient spatial separation of the photogenerated carriers at the heterojunction interface prolonged the relaxation time of the photocurrent, leading to a persistent photoconductivity effect. As a result, the memristor successfully simulated various synaptic functionalities and neural activities under electrical or optical stimulation, including long-term potentiation/depression (LTP/LTD), SNDP (200 °C), and learning-forgetting-relearning behavior (200 °C) (Figs. 5(c)−5(e)).

In addition to the capture and release of photogenerated carriers by heterostructures, optically stimulated synaptic devices can also be constructed based on the mechanism of trapping and detrapping of photogenerated carriers by defects. Recently, Pi’s group fabricated 4H-SiC optoelectronic synaptic devices through the introduction of high-concentration deep-level defects via electronic irradiation (Fig. 6(a))^[44]. As illustrated in Fig. 6(b), upon optical excitation, electrons at the defect levels of 4H-SiC absorbed photo energy and transition to the conduction band. After the optical excitation stopped, the photogenerated electrons in the conduction band may be recaptured by defects. Hence, the lifetime of the photogenerated carriers in the 4H-SiC device was extended, which enabled the device to mimic various synaptic functionalities even at elevated temperatures up to 327 °C, such as PPF, SNDP, and SRDP (Figs. 6(c)−6(e)). The superior thermostability of the devices can be attributed to the inherent robustness of 4H-SiC material. Its wide bandgap (about 3 times that of Si, and 1.4 times that of 3C-SiC), large thermal conductivity (about 3.3 times that of Si, and 1.2 times that of 3C-SiC), and exceptionally strong Si−C bond results in a high melting point (about 2700 °C), stable physicochemical properties, and resistance to structural degradation at elevated temperatures^[75].

A biological brain can parallelly perform complex computations through its neural network. To simulate such functionalities, it is imperative to integrate logic functions into synaptic devices, which serve as the fundamental building blocks of neuromorphic computing systems. Achieving logic functions in synaptic devices typically necessitates precise inputs that are encoded in both temporal and informational domains. Fig. 7(a) presents the information integration in the synaptic device with input applied at both ends and output obtained at one end. When light spikes (365 nm, 5 s) were applied to two 3C-SiC nano wire synaptic devices, a series of spike logic response could be generated under the regulation of the read voltage on the output electrode. At a low read voltage of 0.1 V, the output relationship of synapses 1 and 2 consistently exhibited an "AND" logic function (Figs. 7(b) and 7(c)). This behavior was observed only when both driving synapses were activated and the response current surpassed a threshold set at −8 nA. Conversely, at a higher read voltage of 0.5 V, the output relationship transitioned to an "OR" logic function when both synapses were activated, and the response current exceeded the same threshold of −8 nA.

The successful emulation of visual synapses indicates that visual signals can be effectively utilized to transmit information, providing significant advantages in the field of signal transmission and processing. Shen et al. demonstrated the capability of a 3C-SiC nano wire synaptic device to distinguish alphabet Morse code^[60]. Their findings revealed that the devices could respond to light signals representative of the International Morse code, with each letter trigging a distinct EPSC amplitude response. As illustrated in Figs. 8(a) and 8(b), they encoded the message "hello" and "world" to realize continuous information transmission and recognition using the 3C-SiC synaptic device. It was observed that each letter could be recognized by analyzing the sum (Fig. 8(c)) and endpoint (Fig. 8(d)) of the EPSC amplitude peak values. A comparative analysis of the SUM and END modes demonstrated that each Morse code could be accurately identified, thereby facilitating the encrypted transmission and recognition of data.

As a typical wide bandgap semiconductor, SiC is an ideal candidate for high-temperature-resistant UV optoelectronic synaptic devices due to its strong UV light absorption, robustness against UV radiation^[76], and high thermostability^[49]. SiC has been employed to achieve UV image learning and memory at elevated temperatures, highlighting its application potential in harsh-environment neuromorphic UV visual systems, such as fire alarms, nuclear combustion monitoring, and planetary exploration^{[4, 46, 77−79]}. Fig. 9(a) depicted the conductance response of a typical 3C-SiC nano wire synaptic device stimulated by different number of 365 nm light pulses at 200 °C^[43]. Figs. 9(b) and 9(c) exhibited the conductance responses of the device 5 and 10 s after stimulation, respectively. To visually verify the image learning and memory capabilities of the devices, three letters, "S", "U", and "T", were stimulated in a 5 × 3 device array using one, five, and ten light pulses, respectively (Figs. 9(d)−9(f)). An increase in the number of stimulations resulted in enhanced learning and memory efficiency regarding the images. Fig. 9(f) demonstrated that the image of "T" could be clearly remembered even 10 s after the stimulation at 200 °C, highlighting the application potential of SiC-based optoelectronic synaptic devices in harsh environments.

In the era of big data and the internet of things, the demand for the quick processing and transmission of large volumes of high-resolution images has significantly increased. Consequently, the necessity for data compression while maintaining acceptable image quality has become increasingly crucial. Recently, color quantization and data compression were achieved through the utilization of a self-organizing map (SOM) neural network based on a 4H-SiC device array (Fig. 10(a))^[44]. During each training epoch, the network learns from input RGB color vectors and updates the weight vectors in accordance with the specified learning rate. This learning rate is determined by the rate of change of the weight vectors within the array. Upon completion of the training process, the image is reconstructed using the trained weight vectors. To evaluate the results, the authors measured the quantization error by averaging the Euclidean distance between each input color vector and the trained weight vector of the nearest competitive node. A lower quantization error indicates a reduced loss in image quality subsequent to quantization, thus signifying a more effective quantization outcome. Fig. 10(b) displays that the quantization error gradually decreased as the number of competitive nodes increased. In addition, the quantization error of the network was found to be independent of temperature, indicating the stability of the color quantization capability of the array across varying temperature conditions. The visual processing results for the SOM network at different temperatures are presented in Fig. 10(c). As the number of competitive nodes increased from 6 to 12, the image quality improved; however, noticeable differences remained when compared to the original image. As the number of competitive nodes increased to 24, the output images closely approximated the input image. Moreover, its color space was compressed from about 16 000 000 to just 24, highlighting the promising potential for the application of neuromorphic visual systems based on 4H-SiC optoelectronic synaptic devices.

In summary, we have discussed the state-of-the-art SiC-based electrically and optically stimulated synaptic devices (Table 1). SiC-based electrically stimulated synaptic devices typically utilize electrical pulses to induce the formation and rupture of metal conductive filaments within the SiC active layer, thus modulating the device’s conductivity and synaptic weight. Their optically stimulated counterparts generally rely on the capture and release of photogenerated carriers facilitated by heterostructures or defects. These synaptic devices are capable of mimicking essential synaptic functionalities and neural activities, such as EPSC, PPF, SDDP, SNDP, STDP, SRDP, LTP/LTD, learning-forgetting-relearning behavior, and classical conditioning of Pavlov's dog. Moreover, they demonstrate significant potential for a variety of neuromorphic applications, including logic functions, wireless transmission, high-temperature image learning and memory, as well as high-temperature color quantization. Specifically, optically stimulated synaptic devices based on 4H-SiC exhibit high heat resistance, with the capability to operate at temperatures as high as 327 °C^[44]. However, current advancements may not fully meet the requirements of harsh-environment neuromorphic applications. To this end, several challenging issues need to be addressed.

(1) Higher heat resistance. As a typical wide bandgap semiconductor, SiC is an ideal candidate for high-temperature-resistant optoelectronic devices^[80]. 4H-SiC-based photodetectors have been demonstrated to operate at temperatures exceeding 450 °C^[53−56], indicating considerable potential for enhancing the maximum operating temperature of SiC-based synaptic devices. This underlines the need for further research and advancements in the design of device structures. The utilization of multi-layer metallization schemes (e.g., Cr/Pt for 450 °C operation^[56], TiW/Al for 550 °C^[53], and Ti/Al for 600 °C^[81]) maybe effective strategies for enhancing the robustness of metal-SiC contacts and increasing the maximum operating temperature of the devices. On the other hand, current SiC-based optically stimulated devices primarily employ organic semiconductors or metal oxides to create heterojunctions with SiC so as to facilitate the separation of photogenerated carriers and extend their lifetime. However, the moderate thermal stability of organic semiconductors and metal oxides may compromise the overall heat resistance of the devices. Utilizing SiC homojunctions would be an effective strategy for the fabrication of SiC-based optically stimulated synaptic devices with superior heat resistance. In addition, the incorporation of nano-micro structures on the SiC surface could enhance the photosensitivity of the optically stimulated devices under harsh conditions, thereby further increasing the maximum operating temperature of the devices^{[82, 83]}.

(2) Fully optical modulation. In a biological neural network, a synapse is either excitatory or inhibitory in response to external stimuli. Thus, to accurately emulate biological synaptic functionalities for the development of artificial neural network, it is essential to create both excitatory synaptic devices and inhibitory synaptic devices. In the context of optically stimulated synaptic devices, it is suggested that the devices should realize the conversion of light into both positive and negative current signals^[84−88]. However, existing SiC-based optically stimulated synaptic devices rarely exhibit a negative photoresponse, thereby restricting their bandwidth, processing speed, and integration density^[9]. Therefore, it is essential to develop SiC-based synaptic devices that can achieve bipolar synaptic behaviors through optical means.

(3) Higher integration density. As synaptic devices are ultimately intended for use in artificial neural networks, their fabrication must be optimized to facilitate integration. The fabrication process of 4H-SiC optoelectronic devices is compatible with conventional silicon technology. Moreover, the diameter of 4H-SiC wafer with low defect state has surpassed 8 inches^[52]. However, the integration density of existing 4H-SiC synaptic devices remains insufficient for constructing complex neural networks and simulating versatile neuromorphic applications. Hence, it is imperative to intensify the use of micro-nano fabrication techniques in future development.