The increasing volume of big data and the rapid advancements in artificial intelligence have spurred a growing interest in the adoption of neuromorphic systems for visual perception sensors. This approach emulates the human brain to overcome the significant limitations of traditional von Neumann architectures, including high power consumption, poor efficiency and limited CMOS scaling^1,2. Inspired by the bio-sensory nervous system, the concept of in-sensor memory and computing technology has emerged, which integrates sensing, memory and computation functions into a single device. This provides a reliable technical solution for visual sensors^3,4. Optoelectronic synaptic devices play a crucial role in this context as they can directly perceive external light signals and carry out information processing during the optoelectronic conversion process. These devices have been primarily adopted to simulate the human vision system, achieving various visual preprocessing functions such as early image processing, edge computing, and image recognition⁵. Two-dimensional (2D) metal chalcogenides have emerged as promising candidates for optoelectronic synaptic devices due to their outstanding optoelectronic properties, ease of surface state modification, and excellent potential for nanoscale integration^6,7. These 2D metal chalcogenide sensors effectively integrate sensing, storage and computation functions, simplifying circuit structures and offering new opportunities and applications for neuromorphic visual simulations and microintegrated device development.

While significant researches have been conducted to explore the neuromorphic behaviors of individual synaptic devices of 2D metal chalcogenides, the realization of integrated artificial synapses at the chip level has been sparsely reported. This is primarily ascribed to the challenges existing in the preparation of large-scale 2D metal chalcogenides as well as the absence of industry-compatible nanofabrication and integration techniques suitable for these materials^8,9. Recently, we introduced an innovative allopatric defect-modulated Fe₇S₈@MoS₂ synaptic heterostructure which is featured with a defect-rich Fe₇S₈ core enclosed by a curved MoS₂ dome shell. In this structure, photocarriers generated in the MoS₂ shell are effectively captured and immobilized through the intrinsic defects of the adjacent Fe₇S₈ core. As a result, neuromorphic phototransistors with remarkable optically tunable synaptic behaviors were developed for potential artificial vision devices.

Building upon this achievement, the current work presents a sulfurization-seeding synergetic growth strategy, enabling large-scale and arrayed preparation of the Fe₇S₈@MoS₂ heterostructure. Furthermore, a three-dimensional (3D) vertical integration technique has also been developed for fabricating an arrayed optoelectronic synaptic chip. with the adoption of this chip, we have successfully simulated the visual persistence function of human eyes. Our synaptic devices excel in emulating the preprocessing functions of the human visual system, which contributes to significantly improved noise reduction and image recognition efficiency. This study is poised to significantly advance the field of optoelectronic synaptic devices, moving closer to the development of mature integrated visual perception chips.

A sulfurization-seeding synergetic growth strategy has been developed for the preparation of large-area Fe₇S₈@MoS₂ heterostructures on Si/SiO₂ substrates. As illustrated in Fig. 1a, a uniform Fe film was initially deposited on the silicon substrate with a 300-nm-thick SiO₂ insulation layer using magnetron sputtering¹⁰. Following a 1-h annealing process, nanoparticles were formed, serving as nucleation sites for MoS₂ growth. Subsequently, a large-area and high-quality MoS₂ film was obtained through chemical vapor deposition, with the nanoparticles sulfurizing into Fe₇S₈¹¹. This resulted in the formation of Fe₇S₈@MoS₂ heterostructures with a characteristic dome-like structure, where the MoS₂ thin layer wraps tightly around the Fe₇S₈ core¹². Patterned heterostructures of various sizes were fabricated adopting the method of photolithography and magnetron sputtering coating. The Fe nanoparticles in each pattern serve as nucleation sites for the formation of MoS₂, which leads to the uniform growth of MoS₂ on the square arrays with the sizes of 520 μm × 520 μm (Fig. 1b) and 1020 μm × 1020 μm (Fig. 1c). This demonstrates the continuity and uniformity of the products achieved through the sulfurization-seeding synergetic chemical vapor deposition^13,14. Within the selected area, as shown in Fig. 1b (50 μm × 50 μm square), optical microscopy revealed extensive MoS₂ growth, forming a continuous and uniform film between the square arrays of particles. The consistent thickness of this region was confirmed using Raman spectroscopy, with the in-plane and out-of-plane vibration modes of MoS₂ characterized by the $E_{2 \mathrm{~g}}^{1}$ and $A_{1 \mathrm{~g}}$ modes, respectively^15,16. The peak positions exhibited consistency, with a difference of 20 to 21 cm⁻¹, indicating that the MoS₂ shell on the nanoparticles exhibits a few-layer structure¹⁷. Raman spectra from the 15 spots in three directions on the square arrays show the same intensity as well as a constant phonon frequency of 20 to 21 cm⁻¹ (Supplementary Fig. S1), which further proves the uniform growth of the heterostructure over a large area. According to the atomic force microscopy (AFM) image of the Fe₇S₈@MoS₂ heterostructure, we can observe the particle with a radius of 20 to 35 nm (Fig. 1e) and the thickness of the lateral MoS₂ surrounding the nanoparticles is 1.6 nm (Supplementary Fig. S2). The X-ray photoelectron spectroscopy (XPS) spectrum of the Fe₇S₈@MoS₂ heterostructure sample is presented in Supplementary Fig. S3, exhibiting distinct peaks corresponding to Fe, O, C, Mo and S^18,19.

Fig. 2a illustrates the process of 3D vertical integration for an arrayed Fe₇S₈@MoS₂ optoelectronic synaptic device. The device consists of 10 × 10 individually addressable pixels, as depicted in Fig. 2b, showing photographic images and schematics of the array chip mounted on a printed circuit board (PCB), moreover, the interconnections within the circuitry are also displayed, together with the individual sensor cells that share a structure similar to the aforementioned phototransistor. Fabrication procedures are described in the Methods section. During the performance testing of optoelectronic devices, it is essential to construct an arrayed arrangement of metal electrodes on the sample surface, which could be achieved through direct laser writing and etching processes^20,21. The fabrication of Fe₇S₈@MoS₂ heterostructure optoelectronic synaptic arrayed devices is accomplished through direct laser writing and etching processes, as described in the Supplementary Fig. S4. Direct laser writing is employed to etch the MoS₂ layer to prevent short-circuits resulting from continuous MoS₂ growth. In order to test the imaging performance, the constructed array devices are wire-bonded and connected to the chip circuitry. The Fe₇S₈@MoS₂ heterostructure optoelectronic synaptic imaging chip is schematically presented in Fig. 2b. Gold wires are used to connect the large electrodes to the corresponding directional pins on the PCB board (Fig. 2c). After preparing the fourth-layer metal electrode layer, the sample is wire-bonded, and the tape applied to the edges of silicon wafer to provide a sufficient clearance for the gold wires so as to prevent contact with the edges, thereby avoiding device shorts. Fig. 2d showcases the bonded array devices, where the large electrodes connected via gold wires are further linked to circular vias on the upper PCB board through the board circuits. The final schematic diagram of the array device is displayed in Fig. 2e, which comprises three layers from top to bottom: the silicon wafer with array devices, the upper PCB board, and the lower PCB board with pins. Circular vias on the upper PCB board are connected to the lower PCB board through soldering with a spot welder, allowing external testing system connections via pins.

As shown in Fig. 3a, the I-V curves of the synaptic devices exhibit linear variation in the dark state and under different light power densities, indicating a good ohmic contact between the Bi/Au electrodes and the Fe₇S₈@MoS₂ heterostructure in the array device²². In order to investigate the uniformity of the array device, the dark current of each device in the 5 × 5 array and the photocurrent value after 1 s of ultraviolet (UV) light irradiation with illuminance of 7.15 mW/cm² were tested. As depicted in Fig. 3b, statistical results of the dark current for the Fe₇S₈@MoS₂ array device show that all the values are around 19.3 to 52.1 pA, which is consistent with the dark current performance of individual devices within the same order of magnitude. Furthermore, Fig. 3c presents the statistical values of the photocurrent for each array device, which exhibits an increase of more than one order of magnitude compared to the dark current. All the values are around 0.213 to 1.13 nA, indicating significant contrast between low-power, short-duration light illumination and the dark current. These results demonstrate the relative good uniformity of the prepared array device in terms of its photoresponse functionality under pulsed UV light.

The Fe₇S₈@MoS₂ array device exhibits excellent UV light synaptic tunability. In Fig. 3d, the photocurrent increases with increasing illumination intensity. This is attributed to the higher intensity of the excitation light, which leads to an increased number of carriers excited in MoS₂. Additionally, this heightened intensity renders more carriers trapped by defects, contributing to a gradual extension of the synaptic memory time^23,24. In Fig. 3e, the conversion from short-term plasticity (STP) to long-term plasticity (LTP) is achieved by increasing the frequency of light pulses. Since the rate of carrier release from defects remains constant, the current levels after the removal of light stimulation exhibit significant differences under varying frequencies^25,26. For example, after 10 s of cessation of light exposure, the current level at a high frequency (f = 0.5 Hz) remains much higher than that at a low frequency (f = 0.1 Hz). Moreover, under the stimulus of multiple consecutive pulses, the number of carriers captured by defects tends to be saturated, resulting in a slowdown in the increase of the current level after several pulses. This trend is demonstrated by the curve with 25 pulses in Fig. 3f. Simultaneously, as the rate of carrier release by defects remains constant, a large number of carriers fail to be released by defects after multiple pulses, leading to the formation of persistent photoconductivity and resulting in the transition from short-term memory (STM) to long-term memory (LTM) after multiple pulse stimulations²⁷. As shown in Fig. 3f, an increase in pulse number from 1 to 25 results in an increase in photocurrent intensity from 0.356 nA to 0.75 nA, which is accompanied by an increase in decay time²⁸. Also, under the short pulse width (50 ms to 100 ms), the device can still exhibit memory effects to the current (Supplementary Fig. S5). This transition from short-term to long-term memory can be achieved and simulated by varying the frequency and number of pulses^29,30. Overall, the Fe₇S₈@MoS₂ device offers several advantages that will be highlighted in the comparative analysis (Supplementary Table S1), including its simple structure that facilitates integration and short operation pulse widths. These features collectively contribute to the distinctive performance of our device in neuromorphic systems, underscoring its potential for advanced applications.

Human vision, as illustrated in Fig. 4a, combines sensing and processing functions. The heterostructure device is used to mimic the perceptual and preprocessing functions of the human retina, as shown in Fig. 4b. The retina preprocesses images while receiving light signals and converts them into physiological electrical signals. These signals are then forwarded to the visual cortex for more complex information processing. Biological synapses primarily facilitate signaling and information processing between neurons through electrical and chemical synaptic transmission, with information transfer from the retina to the brain occurring primarily via presynaptic and postsynaptic neurotransmitter release and reception. In image preprocessing, contrast enhancement and noise reduction are crucial for highlighting the main features of an image. A technique was formulated to convert the pixel intensity values of handwritten digit images into the corresponding intensity of pulses. This conversion ensures that the intensity of each pixel precisely influences the number of pulses, thus modulating the response of the device in a direct manner. The presence of noise within the images led to artificially inflated values for certain pixels, which could potentially compromise the accuracy of image recognition. A specific pulse threshold was implemented to mitigate this issue. Pixels with values beneath this threshold were deemed to be noise and were consequently excluded. Digital raw data from the MNIST test dataset were selected for simulation preprocessing, with each image measuring 28 × 28 pixels. Gaussian noise with a grayscale range of 0 to 1 was randomly added to the selected data. The noisy image data were preprocessed with noise reduction using a heterostructure device^31,32. Fig. 4c shows a real image, with the optical signal being converted to the device conductance and the output image being read based on voltage. All the information was subsequently normalized and sent to a neural network consisting of an input layer (784), a hidden layer (15), and an output layer (10) for image recognition^33-35. Fig. 4d compares the input and output images after preprocessing. The heterostructure device effectively highlights the main features of the letters and reduces the redundant data. The normalized photocurrent variations with the intensity of the pulse are shown in the Supplementary Fig. S6. Fig. 4e displays the recognition rate of the vision system with and without the heterostructure device and the original unprocessed dataset. When using the heterostructure device for image preprocessing, the overall recognition accuracy of the dataset before preprocessing is 75.67%, whereas the recognition accuracy of the dataset after photoelectric synaptic preprocessing is 91.25%. We further explored whether short operation pulse widths have an impact on image preprocessing. The Supplementary Fig. S7a shows the variation of current levels with light intensity under short pulse widths. Moreover, this characteristic is also utilized for image preprocessing and recognition operations. It is noteworthy that the recognition accuracy of our heterostructured system in the artificial neural network remained almost unchanged regardless of the pulse width used (Supplementary Fig. S7b). Compared with the image data without noise reduction pretreatment, this pretreatment process could significantly improve the recognition efficiency³⁶. Preprocessing of sensor arrays can effectively extract feature information and smooth out background noise signals, enabling the accurate recognition of visual information. This highlights the potential of synaptic devices for the development of neuromorphic vision systems.

The optoelectronic synaptic chip is also adopted to simulate the visual suspension phenomenon. As mentioned earlier, human eye imaging involves converting external light signals into neural signals through the eyes and transmitting them to the brain for processing and interpretation, as depicted in Fig. 5a³⁷. The visual images are formed by the brain processing and short-term storage of visual stimuli information³⁸. These images can be reproduced from memory within a certain period, known as visual suspension³⁹, as illustrated in Fig. 5b. The Fe₇S₈@MoS₂ photoelectric synaptic heterostructure can simulate the process of receiving and converting light into electrical (neural) signals. The imaging effect of a 5 × 5 array device of Fe₇S₈@MoS₂ was tested by irradiating the designated area of the device with UV light at a light power density of 7.15 mW/cm² for 1 s, with a source-drain voltage of 1 V. The device exhibits a clear imaging effect. Fig. 5c displays a clear contrast image of the letters “HUST”, with red indicating high current contrast and gray indicating low current contrast. In the area with light irradiation, the magnitude of the photoelectric current of Fe₇S₈@MoS₂ is 0.1 nA, whereas the current in the area without light irradiation ranges from 10 to 30 pA, showing a clear contrast between high and low currents. Fig. 5d shows the imaging pattern after removing the light irradiation, and the Fe₇S₈@MoS₂ optoelectronic synaptic device retains the patterns of “HUST”. The device can remember the images even after removing the light irradiation, achieving the functions of image recognition and memory simulation of the human eye. The duration of the light pulse used in the test is 1 s, representing short-term memory. Under this condition, the current level of the device decays rapidly, ranging from 70 to 190 pA, which results in quick forgetting. In order to achieve long-term memory, the imaging effect can be maintained for a longer time by increasing the light power density, which extends the pulse duration and increases the number of pulses among other factors⁴⁰. These results demonstrate that the integrated device of Fe₇S₈@MoS₂ photoelectric synaptic heterostructure could effectively simulate the imaging and memory characteristics of the human retina vision.

In summary, the current work highlights the growing significance of neuromorphic systems in visual perception sensors for addressing the limitations of traditional computing architectures, such as high power consumption and efficiency issues. The mature application of novel Fe₇S₈@MoS₂ heterostructures in optoelectronic synaptic devices was promoted by developing a sulfurization-seeding synergetic growth strategy to realize their large-scale arrayed preparation and a 3D vertical integration technique to realize the fabrication of the arrayed optoelectronic synaptic chip. Our synaptic devices demonstrate great capability in mirroring the preprocessing behaviors of the human visual system, yielding significantly improved noise reduction and image recognition efficiency. We have also realized the simulation of the visual persistence phenomenon of human eyes by using the arrayed chip. The study will significantly prompt the progress of optoelectronic synaptic devices towards mature integrated visual perception chips.

Supplementarymaterials Supplementary data to this article can be found online at https://doi.org/10.1016/j.chip.2024.100088.

Acknowledgments Yao Deng and Shenghong Liu contributed equally to this work. We also would like to thank the technical support from Analytical and Testing Center in Huazhong University of Science and Technology.

Funding This work was supported by the National Key R&D Program of China (2021YFA1200501), National Natural Science Foundation of China (U22A20137, U21A2069, 62202350), Shenzhen Science and Technology Innovation Program (JCYJ20220818102215033, GJHZ20210705142542015, JCYJ20220530160811027).

Declaration of Competing Interest--- The authors declare no competing interests.