A novel photoreceptor for neural network visual perception and its preparation method and application

Abstract

This invention provides a preparation method and application of a novel memristor for neural network visual perception. Using ultrasonic atomization deposition, molybdenum disulfide nanosheets are asymmetrically deposited on a defective graphene electrode through a mask on a silicon dioxide substrate. Heating then induces partial oxidation of the graphene and molybdenum disulfide, forming an asymmetric photovoltaic-effect graphene oxide / oxygen-doped molybdenum sulfide / graphene oxide device, which serves as a novel memristor for neural network visual perception. This invention proposes a novel device, preparation method, and application that modulates the photoresponsivity of devices at both ends by memorizing the electrical flux flowing through the device.

Description

Technical Field

[0001] The present invention relates to a novel two-terminal device for neural network visual perception integrating sensing, storage and computing, and in particular to a novel photoreceptor for neural network visual perception and a preparation method and application thereof. Background Art

[0002] In fundamental and core areas critical to national security and overall development, the implementation of strategic scientific plans and projects, targeting cutting-edge fields such as artificial intelligence, integrated circuits, brain-inspired science, and optoelectronic chips, and implementing a series of forward-looking, strategic national science and technology projects is a pressing need for the strategic development of the country's 14th Five-Year Plan and the 2035 Vision. In recent years, with the rapid development of artificial intelligence fields such as informationized warfare and intelligent robotics, the demand for processing wide-spectrum image information has continued to grow, placing higher demands on intelligent perception technology. The development of high-performance, intelligent artificial intelligence visual perception chips is a strategic advantage for establishing a new generation of intelligent, autonomous recognition and decision-making systems and developing wide-spectrum information acquisition technologies for unmanned spaces such as deep sea and deep space.

[0003] Although the spectral range and response frequency of modern photodetectors have far exceeded those of the human eye, the existing intelligent imaging system is based on a separate architecture for image acquisition, storage, and recognition, which requires a large amount of data to be moved at each level. This places extremely high demands on the computer's computing power and data storage speed. This demand further highlights the huge gap between computing speed and information reading and writing speed (the "storage wall problem"). Figure 1a ), which slows data processing while also generating significant power consumption. With a deeper understanding of the human visual system, we've discovered that preprocessing the retina, optic nerve, and visual cortex, and extracting features from high-resolution, redundant image data, significantly reduces image feature information during transmission. Research has shown that image preprocessing significantly improves the recognition efficiency of back-end deep neural networks, shortens image recognition time, and reduces power consumption. This provides a new mathematical and physical paradigm for low-power, highly dynamic, autonomous decision-making neural network visual perception.

[0004] like Figure 1b As shown in the figure, in order to simulate the hierarchical structure and biological functions of the photoreceptors and bipolar cell layers of the human eye, the device needs to achieve controllable positive and negative light responses. At present, the devices that achieve positive and negative light responses are mainly divided into two categories: the first category is based on metal-semiconductor-metal double Schottky junctions, GaAs-based VSPD's (b) Si-based VSPD's, which realize image preprocessing; the second category is photodetectors and photoelectric memories based on bipolar two-dimensional materials.

[0005] As early as 1990, Yoshikazu Nitta et al. from Mitsubishi Electronics Research Laboratories in Japan used a GaAs-based Figure 2a ) and Si ( Figure 2b Variable Sensitivity Photo-Detector cells (VSPD's) have achieved ultra-high-speed image preprocessing. The principle is to change the depletion layer width of the metal-semiconductor-metal double Schottky junction by bias voltage to form built-in electric fields of different sizes and directions, thereby obtaining adjustable positive and negative photoconductivity, and successfully realizing image preprocessing, such as Figure 2b shown.

[0006] In recent years, with the rise of gate voltage-controlled bipolar two-dimensional materials to form PN junctions, it has become possible to adjust the positive and negative photoresponse rates. Professor Thomas Mueller of the Vienna University of Technology proposed an artificial visual neural network solution: through a PN junction detector array with adjustable photoresponse rate, the recognition and encoding of letter patterns can be achieved, such as Figure 2c As shown in the figure, a visual neural network detector composed of a two-dimensional WSe2 PN junction modulated by field effect has been successfully used for image target recognition. This study modulated the electron and hole concentrations in different regions of the two-dimensional WSe2 by gate voltage, forming a built-in electric field with adjustable direction and size, achieving continuously adjustable positive and negative light response rates, and based on this, proved the high efficiency of this artificial visual neural network. At the same time, since this type of artificial visual system is based on The multiplication and accumulation operations were completed, and the process of digital-to-analog conversion in the existing system was avoided through the integration of sensing and computing. The ultra-fast recognition of the target within 10ns was completed, providing a new idea for ultra-high-speed target recognition. Professor Miao Feng and others from Nanjing University adjusted the binding of photogenerated carriers in the WSe2 channel in the h-BN and Al2O3 insulating layers by gate voltage [7], and achieved the regulation of positive and negative light response rates. This study simulated Gaussian difference, Laplace operator, etc. by regulating the positive and negative light response rates, and achieved a variety of convolution image preprocessing effects, such as edge enhancement and contrast enhancement. Since this type of specific convolutional neural network does not rely on external training and has good versatility, it can well imitate the image preprocessing effects in the retina, such as Figure 2d .

[0007] Although the above-mentioned adjustable photodetectors are expected to realize neural network visual perception systems, existing devices still have great room for improvement in terms of device power consumption, photoresponsivity, quantum efficiency, weight storage method, and response rate.

[0008] 1) Photodetectors with adjustable photoresponsivity based on bipolar materials:

[0009] a) Photoresponsivity of PN junction detection device based on dual-gate regulation (10 -3 -10 -2 The device has extremely low A / W level and quantum efficiency, making it difficult to operate under weak light intensity. On the other hand, due to the limitation of the depletion layer width, the device can only operate at high speed in a very small pixel. If the double-gate is used to control the PiN junction to expand the channel length, the response speed of the device will be reduced. The response speed of devices based on the grating effect is still in the millisecond level, which is difficult to apply to high-frequency, high-dynamic application scenarios such as national defense security, autonomous driving and other fields.

[0010] b) The light response rate weight of three-terminal and four-terminal devices is heavily dependent on the regulation of external gate voltage, which not only increases the difficulty of high-density integration but also generates continuous power consumption.

[0011] 2) Sensitivity-tunable photodetectors (VSPDs) based on metal-bulk-metal double Schottky junctions:

[0012] a) The plasticity of the device's photoresponsivity comes from the modulation of the double Schottky junction by the bias voltage. During convolution visual imaging, a large operating bias voltage (5V-15V) needs to be maintained, resulting in high power consumption.

[0013] b) The photoresponsivity weight cannot be stored non-volatilely by designing a floating gate-like structure, which not only leads to high power consumption but also makes it difficult to improve the device structure.

[0014] Due to the inherent similarities between two-terminal memristor systems and biological synapses, crossbar arrays based on memristor systems can efficiently perform multiplication and accumulation operations at the hardware level. Furthermore, the simple two-terminal, vertical device structure greatly facilitates high-density integration. Therefore, the emergence of memristor systems provides the hardware foundation for integrated storage and computation decision-making systems. However, memristors can only regulate conductance, enabling both "storage and computation" without "sensing" functionality, and it is difficult to generate non-volatile negative conductance. To achieve efficient neural network visual perception, it is necessary to develop new device structures that can achieve the integration of "sensing, storage, and computation." Summary of the Invention

[0015] The first purpose of the present invention is to address the problems existing in the prior art of neural network visual perception systems in terms of device power consumption, weight storage method, and response rate, and to propose a method for preparing a new type of memorized photoresponder for neural network visual perception that regulates the photoresponse rate of devices at both ends by memorizing the electrical flux flowing through the device.

[0016] To this end, the above-mentioned purpose of the present invention is achieved through the following technical solutions:

[0017] A method for preparing a novel memristor for neural network visual perception involves asymmetrically depositing molybdenum disulfide nanosheets on a defective graphene electrode through a mask on a silicon dioxide substrate using ultrasonic atomization deposition. Heating then causes partial oxidation of the graphene and molybdenum disulfide to form an asymmetric graphene oxide / oxygen-doped molybdenum sulfide / graphene oxide device with a photovoltaic effect, which can be used as a novel memristor for neural network visual perception.

[0018] Furthermore, the preparation method of the molybdenum disulfide nanosheets includes the following steps: step 1, dispersing 0.6 g of molybdenum disulfide powder in 60 mL of isopropanol and 40 mL of deionized water; homogenizing the dispersion obtained in step 1 in a sealed bottle, and exfoliating it by ultrasonic bath in an ice bath and circulating water; step 3, after the exfoliation step in step 2 is completed, centrifuging the dispersion to remove the unexfoliated material, and collecting 50% of the suspension at the top for a second centrifugation; step 4, re-dissolving the precipitate obtained after separation in 32 mL of methanol and 8 mL of deionized water to obtain a high-concentration MoS2 NS, and then adding 40 mL of deionized water to obtain the final highly concentrated molybdenum disulfide nanosheet ink.

[0019] Furthermore, the selected graphene electrode has more grain boundaries and defect-rich double-layer graphene islands to achieve partial oxidation of graphene and molybdenum disulfide when heated; the double-layer graphene islands are dark islands with a diameter of 4-5 microns.

[0020] Furthermore, the substrate with the graphene electrode was preheated at 210 degrees Celsius for 20 minutes, and then molybdenum disulfide nanosheets were deposited on the graphene electrode at a deposition temperature of 210 degrees Celsius for 70 minutes.

[0021] Furthermore, the heating to induce oxidation of the graphene and the molybdenum disulfide includes heating at 160 degrees Celsius for 15 minutes in air and in vacuum, respectively.

[0022] The second purpose of the present invention is to provide a new type of neural network visual perception photoreceptor prepared by the above-mentioned preparation method and its application in response to the shortcomings of the existing technology.

[0023] To this end, the application of the novel memorized photoresponse for neural network visual perception includes an excitation process and a reset process, wherein the excitation process includes: the novel memorized photoresponse for face-type neural network visual perception, using the stimulation of light and bias voltage, drifts the graphene oxide of the photovoltaic cell to oxygen vacancies near the interface and then reduces it to graphene, and further oxidizes the graphene oxide of the photovoltaic cell on the other side, ultimately obtaining two opposing photovoltaic cells to generate a larger photocurrent, thereby improving the device's conductance and the photocurrent under zero bias;

[0024] The reset process includes applying a reverse bias, and the reduced graphene is replaced by MoS under the bias. 2-y O y The oxygen ions drifting from the middle are oxidized into graphene oxide; the further oxidized graphene on the other side is reduced, which reduces the difference in photocurrent between the two opposing photovoltaic cells, causing the photocurrent of the device of the present invention to decrease under zero bias, thereby achieving photocurrent reset of the entire device.

[0025] Compared with existing technologies, the present invention proposes a novel concept of a memristor, a novel device that modulates the photoresponsivity of a two-terminal device by memorizing the electrical flux flowing through it. This two-terminal memristor with tunable conductance and photoresponsivity broadens the applicability of memristive systems and enables efficient neural network visual perception. By utilizing ion migration to control the spatial position and thickness of the depletion layer, plastic positive and negative short-circuit currents are obtained at zero bias, thereby achieving locally stored, plastic positive and negative photoresponsivity. Compared to three- and four-terminal structures with gate modulation, the two-terminal memristor is expected to enable the fabrication of vertical devices in the future. Through a thicker semiconductor absorption layer and a shorter "channel" length, it achieves efficient absorption of photons and rapid extraction of currents, simulating the hierarchical structure and biological functions of photoreceptor and bipolar cell layers. Furthermore, through the non-volatile distribution of ions and vacancies, the photoresponsivity state is locally stored, further enhancing the intrinsic relevance of memristive systems to the biological retina and neural synapses, providing a broader theoretical and experimental basis for the application of memristive systems to artificial intelligence and highly dynamic autonomous decision-making systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 (a) shows the changes in CPU and GPU computing power and the demand for computing power by artificial intelligence algorithms; Figure 1 (b) shows the neurobiological structure and function of the human retina in existing technologies.

[0027] Figure 2 (a) shows GaAs-based VSPD's in the prior art; Figure 2 (b) shows Si-based VSPD's in the prior art; Figure 2 (c) shows a visual neural network detector composed of a two-dimensional WSe2 PN junction modulated by field effect in the prior art, which has been successfully used for image target recognition; Figure 2 (d) shows a two-dimensional WSe2 photoconductive device based on field effect modulation in the prior art, which realizes image preprocessing.

[0028] Figure 3 (a) shows the preparation of graphene oxide (GO x ) / oxygen-doped molybdenum sulfide (MoS 2-y O y ) / graphene oxide (GO x) device flow chart; Figure 3 (b) is the absorption spectrum data of the prepared MoS2 solution; Figure 3 (c) is the Raman data of the prepared MoS2 nanosheets; Figure 3 (d) is the SEM image of defective CVD graphene, and the inset shows the Raman data before and after deposition; Figure 3 (e) is the MoS 2-y O y SEM image of the film, with the inset showing the corresponding energy dispersive spectroscopy (EDX) data.

[0029] Figure 4(a) is GO x / MoS 2-y O y / GO x Figure 4 (bd) is the current-voltage data of the photoelectric characteristics. Figure 4 (bd) is the low light response state (LRS) and different high light response state erase data (HRS1; HRS2; HRS3), the voltage sweep order is: 0V – 2V –0V – -2V – 0V; 0V – 3V – 0V – -3V – 0V; 0V – 6V – 0V – -6V – 0V. Figure 4 (e) is the GO x / MoS 2-y O y / GO x Characterization of erase and write cycle characteristics between different photoresponse states; Figure 4 (f) is GO x / MoS 2-y O y / GO x Photocurrent data under different photoresponse states and different bias voltages.

[0030] Figure 5(a) is GO x / MoS 2-y O y / GO x Device structure diagram; Figure 5 (b) is a thin GO x / MoS 2-y O y / GO x Figure 5(c) shows the current-voltage data of the photoelectric characteristics. The spheres and parallelograms represent: carbon-gray spheres, oxygen-red spheres, MoS (2-y) O y- blue parallelogram; Figure 5(d) shows the equivalent circuit in Figure (c) and the device's evolution during excitation (the size of the cell is qualitatively represented by the size of the image). Figure 5(e) shows the evolution of Raman data for the redox state of the graphene electrode as the memristor switches between HPS and LPS. Figure 5(f) shows the ratio of Raman scattering intensities as the device's photoresponse states switch between HPS and LPS.

[0031] Figure 6 (a) is a schematic diagram of the human visual perception system; Figure 6 (b) is a schematic diagram of using a memristor to achieve image blur. Figure 6 (c) simulates the Chinese Academy of Sciences logo on a GO-based x / MoS 2-y O y / GO x The image preprocessing process of the memristor under different photoresponse states and parallel combinations of different devices. From left to right: the original image of the Chinese Academy of Sciences logo; the parallel combination image of the memristor; and the images preprocessed with different operators. DETAILED DESCRIPTION

[0032] The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.

[0033] The present invention provides a method for preparing a novel memristor for neural network visual perception. Molybdenum disulfide nanosheets are asymmetrically deposited on a defective graphene electrode through a mask on a silicon dioxide substrate using an ultrasonic atomization deposition method. Heating causes partial oxidation of the graphene and molybdenum disulfide to form an asymmetric graphene oxide / oxygen-doped molybdenum sulfide / graphene oxide device with a photovoltaic effect, which is used as a novel memristor for neural network visual perception.

[0034] The present invention is achieved through the following technical solutions:

[0035] MoS2 nanosheets (liquid-exfoliated MoS2 nanosheets - solution: 60% isopropyl alcohol & 40% deionized water) were asymmetrically deposited on the defective graphene electrode in air through a mask by ultrasonic atomization deposition. The deposition temperature was 210 degrees Celsius for 90 minutes. The deposition was then heated at 160°C for 15 minutes in air and vacuum, respectively. The long heating caused the oxidation of graphene and MoS2, and finally formed asymmetric graphene oxide (GO). x ) / oxygen-doped molybdenum sulfide (MoS 2-y O y ) / graphene oxide (GO x ), because of the asymmetry of the contact area and the contact barrier at both ends, the device shows a photovoltaic effect.

[0036] Specifically, a MoS2 solution was prepared: First, 0.6 g of MoS2 powder (<2 μm, 99%, Sigma-Aldrich) was dispersed in 60 mL of IPA (isopropyl alcohol) and 40 mL of deionized water. The dispersion was homogenized in a sealed bottle (110 mL) and exfoliated for 5 hours in an ultrasonic bath (Green Sonic 2000, Woosung Ultrasonic CO.LTD) with ice and circulating water. After the exfoliation step, the dispersion was centrifuged at 3000 rpm for 30 minutes at 295 K to remove any remaining loose material, and the top 50% of the supernatant was carefully collected. After collecting 250 mL of the above suspension, the suspension was centrifuged at 8500 rpm for 40 minutes at 295 K. All of the supernatant was then removed, and the precipitate was redissolved in 32 mL of methanol and 8 mL of deionized water to obtain a highly concentrated MoS2 NS, and 40 mL of deionized water was added. Finally, 80 mL (32 mL methanol and 48 mL deionized water) of highly concentrated MoS2 NSs were labeled as MoS2 solution.

[0037] In this technical solution, the defective graphene electrode grown by chemical vapor deposition is first transferred to a silicon dioxide substrate, and then a 10-micron channel is obtained by photolithography and oxygen plasma treatment. Figure 3a As shown, we deposited MoS2 nanosheets asymmetrically in air through a mask by ultrasonic atomization liquid deposition. Figure 3b (c) shows the absorption spectrum and Raman data of MoS2 nanosheets) deposited on a graphene electrode at 210°C for 90 min, followed by heating at 160°C for 15 min in air and vacuum, respectively. On the one hand, liquid-phase grown MoS2 has many defects, while on the other hand, our graphene has many grain boundaries and defect-rich double-layer graphene islands ( Figure 3d During the long-term heating in air, the graphene and MoS2 were partially oxidized, and finally asymmetric graphene oxide (GO) was formed. x ) / oxygen-doped molybdenum sulfide (MoS 2-y O y ) / graphene oxide (GO x ).

[0038] The novel photoreceptor for neural network visual perception in the present invention includes an excitation process and a reset process in its application. The excitation process is as follows: the entire device is equivalent to two opposing Schottky photovoltaic cells. Under the stimulation of light and bias voltage, the graphene oxide in one photovoltaic cell is drifted to the oxygen vacancies near the interface and reduced to graphene, increasing the transport efficiency of photogenerated carriers. The graphene oxide in the other photovoltaic cell is further oxidized, reducing the transport efficiency of photogenerated carriers. Ultimately, the two opposing photovoltaic cells generate a larger difference in photocurrent, thereby improving the device's conductance and photocurrent under zero bias.

[0039] The reset process includes: when a reverse bias is applied, the reduced graphene is 2-y O y The oxygen ions drifting in are oxidized into graphene oxide, which reduces the efficiency of photogenerated carrier transport. The further oxidized graphene on the other side is reduced, which improves the efficiency of photogenerated carrier transport. Ultimately, the photocurrent difference between the two opposing photovoltaic cells is reduced, causing the photocurrent of the device to decrease under zero bias, thus achieving photocurrent reset of the entire device.

[0040] This invention proposes the novel concept of a memristor, a novel device that modulates the photoresponsivity of two devices at either end by memorizing the electrical flux flowing through it. It exhibits polymorphic properties, generating varying positive and negative photocurrents through bias voltage control. This device simulates the hierarchical structure and biological functions of the photoreceptor and bipolar cell layers, enabling image preprocessing functions such as Gaussian blurring and contour enhancement. It provides a new physical device for feature extraction from high-resolution redundant image data, reducing power consumption for image recognition tasks. It also provides an integrated "sensing, storage, and computing" execution device for a new mathematical and physical paradigm of low-power, highly dynamic autonomous decision-making neural network visual perception.

[0041] In order to achieve efficient neural network visual perception and expand the scope of application of memristive systems, the present invention proposes for the first time a new concept of a two-terminal memristor with adjustable photoresponse rate: a memristor has a nonlinear photoresponse rate with the same dimension as the photoresponse rate, but can change with the history of the input current or voltage, and remember the electric flux flowing through the device.

[0042] Although the memristor has a similar device structure to the memristor, unlike the memristor that changes the conductance by memorizing the electrical flux flowing through the device, the memristor regulates the photoresponsivity of the device by memorizing the electrical flux flowing through the device, and has a different dimension from the memristor. This application aims to use ion migration to regulate the photoresponsivity of the two-terminal devices, and obtain a plastic photocurrent under zero bias, thereby achieving a locally stored, plastic photoresponsivity. It simulates the hierarchical structure and biological functions of the photoreceptor and bipolar cell layers, and completes the local storage of the photoresponsivity state through the non-volatile distribution of ions and vacancies, further increasing the intrinsic correlation between the memristor system and the biological retina and neural synapses, and providing a broader theoretical and experimental basis for the application of memristor systems to realize artificial intelligence and highly dynamic autonomous decision-making systems.

[0043] Tests and results

[0044] Through the above preparation method, asymmetric graphene oxide (GO x ) / oxygen-doped molybdenum sulfide (MoS 2-y O y ) / graphene oxide (GO x ) for neural network visual perception, Figure 3d , e show the SEM images of Graphene and MoS2 respectively, Figure 3d The inset Raman data after deposition shows a higher I D / I G and D′ peak, which proves the oxidation of graphene. Figure 3e The energy dispersive X-ray diffraction (EDX) data in the illustration show that MoS 2-y O y formation.

[0045] To test the device performance, we used a solar simulator to test the device's current-voltage characteristics at room temperature and pressure (the voltage sweep rate was 0.5 V / s). Figure 4a shows that when using a solar simulator at 56 mW / cm 2 The brightness of GO x / MoS (2-y) O y / GO x When the device is illuminated, a photocurrent is observed at zero bias, which is due to the asymmetric contact area ( Figure 3a ) and contact barriers. Figure 4bIt shows that the current switches from 0.30μA to 1.23μA at a set voltage of about 1.60V, realizing the switching from low photoresponsivity state (LPS) to high photoresponsivity state (LPS); when a negative bias voltage is applied, the structure switches from HPS1 to LPS at a reset voltage of 1.05V. In the zero bias state, the photocurrent of LPS is 0.01μA and the photocurrent of HPS1 is 0.08μA. Similarly, we stimulate the photoresponse state of the device by different bias voltages, such as Figure 4c As shown in Figure d, the device forms multiple photoresponse states under the stimulation of different bias voltages. HPS2 and HPS3 show photocurrents of 0.1μA and 0.16μA respectively under zero bias, proving the realization of multiple photoresponse states. Figure 4e Data showing repeated reading and writing of multiple photoresponse states demonstrates the stability of the device. Figure 4f The photocurrents at different bias voltages in different photoresponse states are shown. In LPS, the device has a photocurrent of 0.01 μA at 0 V bias; in HPS1, HPS2, and HPS3, the device has a photocurrent of 0.08 μA, 0.1 μA, and 0.16 μA at 0 V bias, respectively.

[0046] The novel light responder for neural network visual perception of the present invention is equivalent to two opposing photovoltaic cells ( Figure 5a ), in order to explore the working principle of the device, we prepared a 10nm thick MoS (2-y) O y The memory photoresponse device, Figure 5b The device's excitation and reset processes are shown. Its working principle is based on the interface state GO x Redox reaction: When a positive voltage is applied under light conditions ( Figure 5c In the figure, a positive voltage is applied from the right side to the left side of the device), the electric field at SC2 is determined by GO x Pointing to MoS (2-y) O y , with the help of light, MoS (2-y) O y Oxygen ions in GO x Migration and oxidation reaction occur, making GO x is further oxidized; the electric field at SC1 is composed of MoS (2-y) O y Point to GO x , with the help of light, the MoS (2-y) O y Oxygen vacancies in GO x Migration and reduction reaction occur, making GOx Reduction to Graphene. The literature proves that the carrier mobility of graphene decreases after oxidation and increases after reduction. Therefore, through the reduction of graphene at SC1 and the oxidation at SC2, Figure 5c In the left figure of d, the ability of SC1 photovoltaic cell to collect carriers is improved, while the ability of SC2 photovoltaic cell to collect photogenerated carriers is reduced, which is reflected in the device test results, that is, the photocurrent under zero bias is improved, and the change from low light response state to high light response state is completed ( Figure 4b ,c,d Forward writing process). The process corresponding to Reset is the opposite. Through the reverse voltage, the reduced graphene at SC1 is oxidized, and the oxidized graphene at SC2 is reduced. From the circuit point of view, the ability of SC1 photovoltaic cell to collect carriers is weakened, while the ability of SC2 photovoltaic cell to collect photogenerated carriers is improved, completing the transition from high light response state to low light response state ( Figure 4b ,c,d negative erasing process). Figure 5e , f shows the Raman data of the device at SC1. With the excitation process, the ratio of the D peak and the G peak decreases, and the ratio of the 2D peak and the G peak increases, which is a typical feature of the reduction of graphene oxide. After reset, the ratio of the D peak and the G peak increases, and the ratio of the 2D peak and the G peak decreases, which means that graphene is oxidized. These data show that our mechanism is based on the redox reaction of graphene.

[0047] By connecting memristors of different polarities in parallel, we can obtain a variety of positive and negative photoresponse states. We simulate these memristors into a 3×3 array, which can simulate the biological receptive field (RF) of the human retina controlled by different photoresponse states. A multiplication and accumulation operation is performed on all the photocurrents of each device group from the simulated array: Figure 6 simulates the Chinese Academy of Sciences logo on the GO-based x / MoS 2-y O y / GO x Image preprocessing of a memristor under different photoresponse states and bias voltages. The first column shows the original image of the Chinese Academy of Sciences logo. The second column shows the different photoresponse states and corresponding readout voltages required by different operators, as indicated by the red boxes. By setting the corresponding photoresponse states and readout voltages, we simulated the weight matrices of different operators in the blue boxes in the third column. The fourth column shows the processed image. From top to bottom, Gaussian blurring, image inversion, and Laplacian-based image contour extraction are simulated.

[0048] The novel photoreceptor for neural network visual perception exhibits polymorphic properties. By regulating the bias voltage, different positive and negative photocurrents are generated, simulating the hierarchical structure and biological functions of the photoreceptor and bipolar cell layers. This enables image preprocessing functions such as Gaussian blurring, outline enhancement, and image inversion. This device provides a new physical device for feature extraction from high-resolution redundant image data, reducing power consumption for image recognition tasks. This device provides an integrated "sensing, storage, and computing" execution device for a new mathematical and physical paradigm of low-power, highly dynamic autonomous decision-making neural network visual perception.

[0049] The above-mentioned specific implementation methods are used to illustrate the present invention and are only preferred embodiments of the present invention, rather than limiting the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit of the present invention and the scope of protection of the claims shall fall within the scope of protection of the present invention.

Claims

1. A method for preparing a novel photoreceptor for neural network visual perception, characterized by: Using ultrasonic atomization deposition on a silicon dioxide substrate, molybdenum disulfide nanosheets are asymmetrically deposited on a defective graphene electrode through a mask. Heating causes partial oxidation of graphene and molybdenum disulfide, forming an asymmetric graphene oxide / oxygen-doped molybdenum disulfide / graphene oxide device with photovoltaic effect. The asymmetry of the contact area and contact potential barrier at both ends shows the photovoltaic effect of the device, which can be used as a new type of photomemory responder for neural network visual perception.

2. The method for preparing the novel photoreceptor for neural network visual perception according to claim 1, characterized in that: The preparation method of the molybdenum disulfide nanosheets includes the following steps: step 1, dispersing 0.6 g of molybdenum disulfide powder in 60 mL of isopropanol and 40 mL of deionized water; homogenizing the dispersion obtained in step 1 in a sealed bottle, and exfoliating the dispersion by ultrasonic bath in an ice bath and circulating water; step 3, after the exfoliation step in step 2 is completed, centrifuging the dispersion to remove the residual unexfoliated material, and collecting 50% of the top suspension for a second centrifugation; step 4, redissolving the precipitate obtained after separation in 32 mL of methanol and 8 mL of deionized water to obtain a high-concentration MoS2 NS, and then adding 40 mL of deionized water to obtain the final highly concentrated molybdenum disulfide nanosheets.

3. The method for preparing the novel photoreceptor for neural network visual perception according to claim 1, characterized in that: The selected graphene electrode has more grain boundaries and defect-rich double-layer graphene islands, so as to achieve partial oxidation of graphene and molybdenum disulfide when heated; the double-layer graphene islands are dark islands with a diameter of 4-5 microns.

4. The method for preparing a novel photoreceptor for neural network visual perception according to claim 1, wherein: MoS2 nanosheets were deposited onto the graphene electrode at a deposition temperature of 210 degrees Celsius and a total deposition time of 90 minutes.

5. The method for preparing the novel photoreceptor for neural network visual perception according to claim 1, wherein: The heating to induce oxidation of graphene and MoS2 involved heating at 160 degrees Celsius for 15 minutes in air and in vacuum, respectively.

6. A novel photoreceptor for neural network visual perception prepared by the preparation method according to any one of claims 1 to 5.

7. The application of the novel photoreceptor for neural network visual perception according to claim 6, characterized in that: The excitation (Set) process, the novel photoreceptor for neural network visual perception described in claim 6, utilizes the stimulation of light and bias voltage to drift the graphene oxide of the photovoltaic cell to the oxygen vacancies near the interface and then reduce it to graphene, and further oxidize the graphene oxide of the photovoltaic cell on the other side, and finally obtain two opposing photovoltaic cells to generate a larger photocurrent, so as to change the device's conductance and the photocurrent under zero bias; the reset (Reset) process includes applying a reverse bias voltage, and the reduced graphene is replaced by MoS under the action of the bias voltage. 2-x O x The oxygen ions drifting from the middle are oxidized into graphene oxide; the further oxidized graphene on the other side is reduced, so that the photocurrent difference between the two opposing photovoltaic cells is reduced, so that the photocurrent of the memphotoresponder described in claim 6 under zero bias decreases, and the photocurrent reset of the entire device is achieved.

Images