Neuromorphic sensor based on programmable photodiode and preparation method

Abstract

The invention discloses a neuromorphic sensor based on a programmable photodiode and a preparation method, and relates to the technical field of photoelectric sensors, and the programmable sensor is equipped with adjustable positive and negative weights and forms an ANN for advanced cognitive calculation. Comprising a substrate, a photosensitive material formed on the substrate, and an electrode covering the photosensitive material. The neuromorphic image sensor comprises n * n pixel points, each pixel point comprises n sub-pixel points, and each sub-pixel point is composed of a programmable photodiode; n > = 3. The substrate is a glass substrate. The pattern of the photosensitive material on the substrate is a line array with fixed line width and line spacing. The electrode is a gold electrode, the length of the electrode is 300-500 microns, the width of the electrode is 100-300 microns, the thickness of the electrode is 40-60 nanometers, and the width of a channel between the electrodes of each photodiode is 50-150 microns. The programmable neuromorphic sensor can form an ANN to carry out advanced cognitive calculation, the hardware processing difficulty is simplified, and the power consumption is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of optoelectronic sensors, and particularly to a neuromorphic sensor based on a programmable photodiode and a preparation method thereof. Background Art

[0002] Traditional vision architectures adopt a design with physically separated sensors, memories, and processing units. This design requires the transmission of a large amount of raw data throughout the signal chain, resulting in significant energy consumption, latency issues, and increased hardware costs. With the continuous development of latency-sensitive applications and sensor-intensive platforms, efforts have been made to eliminate the interfaces between sensors, memories, and processors through new architectures, thereby improving the efficiency and performance of the system. With the emergence of memory materials, in-memory computing technology that combines storage and computing modules has been widely developed and promoted. In recent years, inspired by multifunctional image sensors, in-sensor computing technology that combines sensors and computing modules has also been widely studied and explored. However, a memory-aware and computing architecture that combines the three major modules of sensing, storage, and computing has not been developed yet.

[0003] So far, various brain-inspired vision sensors have been designed to simulate specific functions of the human retina and achieve in-situ visual preprocessing. The optimization of raw data can significantly accelerate the subsequent calculation process. However, the "calculation" here only refers to the preprocessing process of signals, and the recognition of images still depends on external post-processing. More precisely, this should be called a "preprocessing" process because its output is still a representation of the input signal. To achieve image perception and recognition, advanced in-sensor computing based on matrix-vector multiplication operations should be performed, where positive and negative weights are preconditions. Currently, this issue is being widely studied. For example, Fudan University utilized the polarization characteristics of ferroelectric materials to program the photon memory intensity by modulating the energy barrier and developed a programmable ferroelectric bionic vision hardware with selective attention. However, this device introduces a gate, the circuit is complex, so the power consumption is high and it is difficult to integrate on a large scale; Nanjing University developed a vision sensor with programmable positive and negative light responses by adjusting the gate voltage of each pixel for simultaneous image sensing and processing. However, this method does not achieve a true negative response and requires additional circuit assistance during application, thus introducing problems of high power consumption and difficult hardware expansion. In addition, another problem with advanced in-sensor computing is the weights stored in external memories, which inevitably leads to additional power consumption and difficulties in large-scale integration.

[0004] Most neuromorphic vision sensors use MoS2, MoOx, In2O3, ZnO, and other metal chalcogenides, metal oxides, or carbon-based materials (such as graphene) as photosensitive materials. These materials have high carrier mobilities and appropriate bandgaps, but most of them are prepared by high-temperature chemical vapor deposition or physical vapor deposition. The harsh preparation conditions bring high costs and are difficult to integrate on a large scale.

[0005] Variable-sensitivity photodetectors, dual-gate diodes, two-terminal optical memory devices, and gate-controlled vision sensors often face problems such as bias-dependent dark currents, high power consumption, variable photocurrents, complex fabrication processes, and low light responsivity. The pseudo-negative weights obtained by source-drain voltage or by subtraction are used as artificial neural network weights. However, this method has problems of high power consumption and difficulty in expanding the hardware area. In addition, the weights stored in external devices result in additional power consumption and are difficult to integrate on a large scale.

[0006] Therefore, choosing appropriate photosensitive semiconductor materials and developing a new type of programmable neuromorphic vision sensor that can integrate high-performance perception, weight storage, and advanced computing simultaneously has important theoretical significance and broad application prospects. Summary of the Invention

[0007] In view of this, the present invention provides a neuromorphic sensor based on a programmable photodiode and a preparation method, which can solve the deficiencies in the prior art and provide a neuromorphic sensor with excellent weight programmable performance through innovative material selection and device design.

[0008] To achieve the above object, the technical solution of the present invention is: A neuromorphic sensor based on a programmable photodiode includes a substrate, a photosensitive material formed on the substrate, and an electrode covering the photosensitive material.

[0009] The neuromorphic image sensor includes n×n pixel points, each pixel point contains n sub-pixel points, and each sub-pixel point is composed of a programmable photodiode; n≥3.

[0010] The substrate uses a glass substrate.

[0011] The pattern of the photosensitive material on the substrate is: a line array with a fixed line width and line spacing.

[0012] The electrode is a gold electrode. The electrode length is 300 - 500 microns, the width is 100 - 300 microns, and the thickness is 40 - 60 nanometers. The channel width between the electrodes of each photodiode is 50 - 150 microns.

[0013] Furthermore, the photosensitive material selects AgBiS2 material; the photosensitive material is patterned and grown on the substrate by a seed layer stimulation method.

[0014] Another embodiment of the present invention further provides a method for preparing a neuromorphic sensor based on a programmable photodiode, comprising the following steps:

[0015] S1: Provide a glass substrate.

[0016] S2: Clean the glass substrate; cleaning the glass substrate includes: ultrasonically cleaning the glass substrate with acetone, ethanol, and water for 20 - 30 minutes respectively.

[0017] S3: Prepare a pixel array pattern on the glass substrate through a lithography process; the pixel array pattern is: an n×n pixel array pattern, each pixel is composed of n sub - pixels, each sub - pixel is a photosensitive material part of a programmable photodiode, and the pattern of the photosensitive material part is a micron - wire array with a fixed line width and line spacing, where n≥3.

[0018] S4: Place the glass substrate with the lithographed pixel array pattern into a thermal evaporation coating machine to evaporate silver, and obtain a silver material pixel array pattern after stripping.

[0019] S5: Place the glass substrate with the silver material pixel array pattern into a plasma cleaner to bombard and oxidize it into a silver oxide seed layer.

[0020] S6: Spin - coat the AgBiS2 precursor solution onto the silver oxide seed layer through a spin coater.

[0021] S7: Place the glass substrate spin - coated with the AgBiS2 precursor solution in a glove box and heat it to form the AgBiS2 photosensitive material.

[0022] S8: Lithograph a pixel array electrode pattern on the photosensitive material.

[0023] S9: Place the substrate lithographed with the pixel array electrode pattern into a thermal evaporation coating machine to evaporate a gold electrode, and a neuromorphic sensor based on a programmable photodiode can be obtained after stripping.

[0024] Further, the thickness of the evaporated silver material in S4 is 5 - 30 nanometers. The purpose of this step is to pattern and arrange a silver seed layer on the glass substrate, and the subsequently grown AgBiS2 photosensitive material only grows in a patterned manner where there is a silver seed layer.

[0025] The stripping process in S4 includes: soaking in acetone for 5 - 20 minutes.

[0026] Further, the oxidation parameters of the plasma cleaner in S5 are: bombarding the silver material for 10 - 30 minutes under an oxygen flow rate of 2 - 10 sccm to oxidize the evaporated silver material into silver oxide.

[0027] Further, the method for preparing the AgBiS2 precursor solution in S6 is as follows: Weigh AgCl, BiCl3, and thiourea and dissolve them in an N,N-dimethylformamide solution, with the molar ratio of AgCl to BiCl3 being 0-2:1, and stir to mix them evenly.

[0028] The spin coating process in S6 is as follows: Place the substrate in a spin coater and rotate it at a speed of no less than 2000 revolutions per minute for 10-60 seconds, and drop 15-30 microliters of the evenly mixed precursor solution during the rotation.

[0029] Further, the heating parameters in the glove box in S7 are as follows: Place the substrate on a hot plate at 150-200 degrees Celsius and heat it for 3-10 minutes.

[0030] Further, steps S6-S7 need to be repeated 1-5 times during the preparation of the sensor pixel array to form a uniform and dense AgBiS2 thin film.

[0031] Further, the pixel array electrode pattern in S8 is an n×n pixel array pattern, each pixel is composed of n sub-pixels, and each sub-pixel is composed of a programmable photodiode. The dimensions of its electrode part are: length 300-500 microns, width 100-300 microns, and thickness 40-60 nanometers. The channel between the electrodes of each photodiode is 50-150 microns, and n≥3.

[0032] Further, the thickness of the evaporated gold electrode in S9 is 40-60 nanometers.

[0033] The lift-off process in S9 is as follows: Immerse it in acetone for 10-30 minutes.

[0034] The patterned preparation method of the AgBiS2 material is realized by the method of stimulating with a silver seed layer, that is, the final AgBiS2 material only selectively grows on the places where the silver material is evaporated on the substrate in step S4, and does not grow on the places where the silver material is not evaporated, so as to form a 3×3×3 pixel array pattern on the substrate.

[0035] Beneficial effects:

[0036] 1. The neuromorphic sensor based on a programmable photodiode provided by the present invention is equipped with adjustable positive and negative weights. The principle is as follows: The sensor is a symmetric back-to-back double Schottky diode structure, and a series of positive or negative voltage pulses can be used to induce local electrochemical reduction to form silver filaments, thereby breaking the symmetry and forming an asymmetric single Schottky diode, generating a positive or negative photocurrent, and realizing the function of local weight storage.

[0037] 2. The neuromorphic sensor based on programmable photodiodes provided by the present invention can perform image recognition without any external memory and computing unit. The integrated sensing, memory, and computing mode integrates advanced computing, weight memory, and high-performance sensors, paving the way for a computing architecture with low energy consumption, low latency, and reduced hardware costs.

[0038] 3. The programmable neuromorphic sensor provided by the present invention is prepared based on a low-temperature high-resolution patterning method of AgBiS2 material. It has low processing difficulty and can realize the preparation of large-area pixel image sensors, making it a potential material for realizing neuromorphic hardware.

[0039] 4. The present invention provides a preparation method for a neuromorphic sensor based on programmable photodiodes. Its material process is simple and the performance is excellent. The present invention selects AgBiS2 material that can be patterned at low temperature as the photosensitive material. The preparation process is simple and convenient, and only requires a simple "spin coating - heating" process. At the same time, AgBiS2 material is an environmentally friendly I-V-VI ternary semiconductor material with a wide-band spectral response range (300 - 1600 nm) and an extremely high light absorption coefficient, making it a potential material for facilitating the realization of neuromorphic hardware.

[0040] 5. The present invention provides a preparation method for a neuromorphic sensor based on programmable photodiodes. The programmable sensor is equipped with adjustable positive and negative weights and constitutes an ANN by itself for advanced cognitive computing. The programmable neuromorphic sensor provided by the present invention is a symmetric back-to-back double Schottky diode structure in the initial state. Through a series of positive and negative voltage pulses to induce local electrochemical reduction to form silver filaments, the symmetry is broken, forming an asymmetric single Schottky diode, which is equivalent to a photovoltaic cell without an external bias, generating a positive or negative photocurrent and having a stable and reliable working mechanism. Therefore, the programmable neuromorphic sensor can constitute an ANN by itself for advanced cognitive computing, simplifying the hardware processing difficulty and reducing power consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic structural diagram of the neuromorphic sensor based on programmable photodiodes in the embodiment of the present invention.

[0042] Figure 2 In it, A is a schematic structural diagram and equivalent circuit of the neuromorphic sensor in the initial symmetric state in the embodiment of the present invention.

[0043] Figure 2 In it, B is a schematic structural diagram and equivalent circuit of the neuromorphic sensor after being programmed with negative voltage pulses in the embodiment of the present invention.

[0044] Figure 2C in this figure is the schematic diagram of the device structure and the equivalent circuit of the neuromorphic sensor after positive voltage pulse programming in the embodiment of the present invention.

[0045] Figure 2 D in this figure is the simulated photocurrent of the neuromorphic sensor in the initial symmetric state in the embodiment of the present invention.

[0046] Figure 2 E in this figure is the simulated photocurrent of the neuromorphic sensor after negative voltage pulse programming in the embodiment of the present invention.

[0047] Figure 2 F in this figure is the simulated photocurrent of the neuromorphic sensor after positive voltage pulse programming in the embodiment of the present invention.

[0048] Figure 3 A in this figure is the current-voltage characteristic curve of the neuromorphic sensor in the initial symmetric state in the embodiment of the present invention.

[0049] Figure 3 B in this figure is a complete current-voltage test cycle of the neuromorphic sensor in the embodiment of the present invention.

[0050] Figure 3 C in this figure is the current-voltage characteristic curve of the neuromorphic sensor after programming with positive voltage pulses in the embodiment of the present invention.

[0051] Figure 3 D in this figure is the current-voltage characteristic curve of the neuromorphic sensor after programming with negative voltage pulses in the embodiment of the present invention.

[0052] Figure 4 A in this figure is the SEM-EDS elemental analysis diagram of the neuromorphic sensor after silver precipitation by voltage pulse programming in the embodiment of the present invention.

[0053] Figure 4 B in this figure is the high-resolution TEM image and the corresponding TEM-EDS elemental analysis diagram of the neuromorphic sensor after silver precipitation by voltage pulse programming in the embodiment of the present invention.

[0054] Figure 5 These are the in-situ SEM and TEM images of silver precipitation and retraction during the programming modulation of positive and negative voltage pulses of the neuromorphic sensor in the embodiment of the present invention.

[0055] Figure 6 A in this figure is the current-voltage characteristic curve of the neuromorphic sensor under 808-nm light irradiation after 2 positive voltage pulses and 23 identical positive voltage pulses programming in the embodiment of the present invention.

[0056] Figure 6 B in this figure is in the embodiment of the present invention Figure 6Based on A modulation, the current-voltage characteristic curves after applying 10 negative voltage pulses and 25 identical negative voltage pulses are respectively shown.

[0057] Figure 7 This is the light responsivity LTP and LTD characteristics of the neuromorphic sensor in three cycles in the embodiments of the present invention.

[0058] Figure 8 In the embodiments of the present invention, for the neuromorphic sensor under 808-nm light illumination, the holding time in the positive light response state is greater than 2000 seconds.

[0059] Figure 9 A in it is the original grayscale image used for convolution processing in the embodiments of the present invention.

[0060] Figure 9 B in it are three convolution kernels configured with the actual light responsivity of the neuromorphic sensor in the embodiments of the present invention, respectively representing vertical, horizontal, and isolated point edge extraction.

[0061] Figure 10 A in it is used in the embodiments of the present invention Figure 9 The result of extracting image features using the convolution kernel configured with the actual light responsivity in B.

[0062] Figure 10 B in it is the simulation result based on all-floating weights using software simulation in the embodiments of the present invention.

[0063] Figure 11 A in it is the weight distribution of all 27 independent devices in the sensor array in the embodiments of the present invention.

[0064] Figure 11 B in it are the real-time output photocurrents of three kinds of pictures in the embodiments of the present invention, and the photocurrent of the target picture is well separated from that of other pictures.

[0065] Figure 12 This is the simulation calculation demonstration of a four-layer CNN based on a neuromorphic sensor array in the embodiments of the present invention, used for feature extraction and prediction, and the system block diagram for driving a drone to follow a robotic dog.

[0066] Figure 13 This is the real-time view of an eye movement-driven drone following a robotic dog in the embodiments of the present invention. A and C are the perspectives and panoramic views of the drone moving to the right; B and D are the perspectives and panoramic views of the drone moving to the upper right.

[0067] Figure 14 This is the accuracy of all-floating weights, all-electric weights, and optoelectronic weights for eye movement detection in the embodiments of the present invention.

[0068] Figure 15This is the confusion matrix for eye movement detection in the real-time example of the present invention. Detailed implementation mode

[0069] The present invention will be described in detail below in conjunction with the accompanying drawings and by way of examples.

[0070] The present invention provides a neuromorphic sensor based on a programmable photodiode and a preparation method thereof. The sensor can program the light response state through a voltage pulse to realize the integrated neuromorphic photoelectric sensing functions of perception, storage, and calculation.

[0071] This neuromorphic sensor, as Figure 1 shown, includes: a substrate 1, a photosensitive material 2 formed on the substrate, and an electrode 3 covering the photosensitive material. Among them:

[0072] The neuromorphic image sensor includes n×n pixel points, each pixel point contains n sub-pixel points, and each sub-pixel point is composed of a programmable photodiode, where n≥3.

[0073] The neuromorphic sensor uses a glass substrate.

[0074] The photosensitive material is selected as AgBiS2 material.

[0075] The photosensitive material is patterned and grown on the substrate by the seed layer stimulation method.

[0076] The pattern of the photosensitive material on the substrate is: a line array with a fixed line width and line spacing.

[0077] The electrode is a gold electrode, with an electrode length of 300 - 500 microns, a width of 100 - 300 microns, and a thickness of 40 - 60 nanometers. The channel width between the electrodes of each photodiode is 50 - 150 microns.

[0078] The present invention also provides a preparation method for a neuromorphic sensor based on a programmable photodiode. The steps of this method include:

[0079] S1: Provide a glass substrate;

[0080] S2: Clean the glass substrate; in the embodiment of the present invention, cleaning the glass substrate includes: ultrasonically cleaning the glass substrate with acetone, ethanol, and water for 20 - 30 minutes respectively.

[0081] S3: Prepare a pixel array pattern on the glass substrate through a lithography process; in the embodiment of the present invention, the lithography process includes: spin coating, pre-baking, exposure, and development. The pixel array pattern is: an n×n pixel array pattern, each pixel point is composed of n sub-pixel points, each sub-pixel point is the photosensitive material part of a programmable photodiode, and the pattern of the photosensitive material part is a micron line array with a fixed line width and line spacing, where n≥3.

[0082] S4: Place the glass substrate lithographed with the pixel array pattern into a thermal evaporation coating machine to evaporate silver, and after stripping, obtain the silver material pixel array pattern; in the embodiment of the present invention, the thickness of the evaporated silver material is 5 - 30 nanometers. The purpose of this step is to pattern and arrange a silver seed layer on the glass substrate, and the subsequently grown AgBiS2 photosensitive material only grows in a patterned manner where there is a silver seed layer.

[0083] In the embodiment of the present invention, the stripping process includes: soaking in acetone for 5 - 20 minutes.

[0084] S5: Place the glass substrate with the silver material pixel array pattern into a plasma cleaner to bombard and oxidize it into a silver oxide seed layer; in the embodiment of the present invention, the oxidation parameters of the plasma cleaner are: bombarding the silver material for 10 - 30 minutes under an oxygen flow rate of 2 - 10 sccm to oxidize the evaporated silver material into silver oxide.

[0085] S6: Spin - coat the AgBiS2 precursor solution onto the silver oxide seed layer through a spin coater. In the embodiment of the present invention, the preparation method of the AgBiS2 precursor solution is: weigh AgCl, BiCl3, and thiourea and dissolve them in N,N - dimethylformamide solution, so that the molar ratio of AgCl to BiCl3 is 0 - 2:1, and stir to mix them evenly.

[0086] In the embodiment of the present invention, the spin - coating process is: place the substrate into a spin coater and rotate it at a speed of not less than 2000 revolutions per minute for 10 - 60 seconds, and drop 15 - 30 microliters of the evenly mixed precursor solution during the rotation process.

[0087] S7: Place the glass substrate spin - coated with the AgBiS2 precursor solution in a glove box and heat it to form the AgBiS2 photosensitive material; in the embodiment of the present invention, the heating parameters in the glove box are: place the substrate on a hot plate at 150 - 200 degrees Celsius and heat it for 3 - 10 minutes.

[0088] In the embodiment of the present invention, during the preparation of the sensor pixel array, steps S6 - S7 (i.e., spin - coating - heating) need to be repeated 1 - 5 times to form a uniform and dense AgBiS2 thin film.

[0089] S8: Lithograph the pixel array electrode pattern on the photosensitive material; in the embodiment of the present invention, the pixel array electrode pattern is an n×n pixel array pattern, each pixel is composed of n sub - pixels, and each sub - pixel is composed of a programmable photodiode. The size of its electrode part is: length 300 - 500 microns, width 100 - 300 microns, thickness 40 - 60 nanometers. The channel between each photodiode electrode is 50 - 150 microns, and n≥3.

[0090] S9: Place the substrate lithographed with the pixel array electrode pattern in a thermal evaporation coating machine to evaporate the gold electrode, and after stripping, a neuromorphic sensor of the programmable photodiode can be obtained.

[0091] In the embodiment of the present invention, the thickness of the evaporated gold electrode is 40 - 60 nanometers. The stripping process is: soak in acetone for 10 - 30 minutes.

[0092] In the embodiment of the present invention, the patterning preparation method of the AgBiS2 material mentioned in this step is realized by the method of stimulating with a silver seed layer, that is, the final AgBiS2 material only grows selectively on the places where the silver material is evaporated on the substrate in step S4, and does not grow where the silver material is not evaporated, so as to form a 3×3×3 pixel array pattern on the substrate.

[0093] The working principle of the programmable photodiode is as follows: There is a unique cation migration characteristic in the AgBiS2 semiconductor. The reversible formation of Ag filaments is used to break and reconfigure the photovoltaic polarity in the device, thereby establishing a reliable non-volatile model for the positive and negative weight bipolar adjustable light response of the neural network.

[0094] Specifically, in the initial state, a double Schottky photodiode is designed using a metal-semiconductor-metal structure, and an AgBiS2 semiconductor layer is patterned and deposited on the contact surfaces of the two gold electrodes, as shown in Figure 2 A in. Under illumination, the built-in electric fields of the two photodiodes decrease by the same amplitude (i.e., the open-circuit voltage), generating a zero net potential difference across the entire pixel and no photocurrent at 0V, as shown in Figure 2 D in. Then, a series of negative voltage pulses are applied to one electrode to induce local electrochemical reduction to form silver filaments, which will reduce the work function of the local contact and ultimately eliminate the initial Schottky barrier. Therefore, the symmetry is broken, and only one diode remains on the other electrode ( Figure 2 B in), equivalent to a photovoltaic cell without an external bias, generating a positive photocurrent ( Figure 2 E in). Subsequently, if a series of positive voltage pulses are applied, it will promote the oxidation / ionization of metallic silver, thus dissolving back into the chalcogenide solid solution, and at the same time forming new silver filaments at the opposite electrode ( Figure 2 C in), generating a negative photocurrent ( Figure 2 F in).

[0095] Figure 3 A in shows the current-voltage characteristic curve of the symmetric device in the initial state. When a small voltage is applied to either polarity of the symmetric diode, one diode is reverse-biased to block the current. However, a higher voltage can trigger the formation of silver, thereby reducing the barrier height of the diode and promoting an exponential increase in current. Cyclic voltammetry further reveals this tunable polarity transformation process, as shown inFigure 3 As shown by B in , a complete test cycle is presented, showing a combination of two typical diode characteristics with opposite polarities. When the symmetry is broken at high voltages, the elimination of the Schottky barrier results in an asymmetric single-diode configuration. The critical switching voltage is around ±10 V, after which the diode polarity switches immediately, as Figure 2 shown by the equivalent circuit in B of . The device can be programmed reversibly through a sequence of positive or negative voltage pulses, as Figure 3 shown by B and C in , showing single-diode characteristics.

[0096] Figure 4 As characterized by SEM-EDS for the formation of silver in A of , it can be seen that Bi and S coexisting with Ag are uniformly distributed throughout the material, while elemental silver accumulates in specific regions due to the electrochemical reaction triggered by voltage pulses. The same conclusion can be obtained from Figure 4 the cross-sectional TEM-EDS in B of . Figure 5 In-situ SEM and cross-sectional TEM images at the initial state, silver precipitation state, and silver retraction state under voltages of 0 V, 10 V, and -10 V are shown respectively.

[0097] As Figure 6 shown by A in , when the programming voltage is +10 V, the unidirectional photodiode first generates a negative short-circuit photocurrent, which increases with the increase in the number of programming pulses; when the programming voltage is reversed, as Figure 6 shown by B in , the photocurrent reverses back to 0 and further increases in the positive direction with the increase in the number of programming pulses, indicating the bipolar tunability of the photovoltaic cell.

[0098] Figure 7 The "set-reset" process of regulating the photocurrent through the number of programming voltage pulses within three cycles is given. Each "set-reset" process consists of LTP and LTD processes of photocurrents in 52 different states respectively. Figure 8 The holding time of the programmable device is shown. The amplitude of its light response can last for more than 2000 s, demonstrating its good non-volatility, which meets the prerequisite for moving towards array-level hardware for image information processing within sensors.

[0099] First, by constructing an optoelectronic convolution kernel to simultaneously achieve image sensing and processing, its potential application at the array level is demonstrated. Figure 9 A in is the original grayscale image for convolution processing, Figure 9 and B in are three convolution kernels configured with actual light responsivities for edge detection and sharpening. The corresponding four image feature extraction results after convolution operation are as Figure 10 shown by A in , which is almost consistent with the simulation results based on fully floating weights in B of . Figure 10 in .

[0100] In addition to the image preprocessing function, its inherent parallel vector matrix multiplication and accumulation (MAC) capabilities enable the neuromorphic photodiode array to perform real-time analog computing. For example, folding the three layers of neurons in the 3×3×3 programmable photodiode array shown in Figure 1 into one layer to form an artificial neural network (ANN) within the sensor for MAC-based picture classification. The weight distribution of all 27 independent devices is as shown in Figure 11 A, and they are programmed in the sensing hardware through voltage pulses. When the three patterns of "├", "T", and "┤" are sequentially projected onto the sensor array through the mask template, the sensor generates the highest output currents of I1, I2, and I3 respectively to classify pictures, as shown in Figure 11 B. Since the bipolar programmable photodiode produces symmetric positive and negative light responses, the output current of the pseudo neuron is almost zero, indicating high accuracy when performing the ANN classification task.

[0101] Next, the potential of neuromorphic programmable devices to perform actual complex tasks was further explored. Figure 12 An example of a CNN hybrid network composed of a photoelectric convolutional kernel and a pure electric convolutional kernel / full connection layer is shown, demonstrating the prospect of constructing a hardware-implementable deep learning network. In this configuration, the pure electric CNN can also be implemented by this symmetric programmable photodiode because the current rectification polarity can also be switched during programming to generate a coordinated conductivity, thereby realizing the MAC operation through Ohm's law. Through the exposure of the optical eye image, MAC operations are performed in the CNN for feature extraction and dimensionality reduction, full connection information classification, converting the eye movement information into position and direction, and finally applying it to drive the drone actuator. Figure 13 It shows that the drone is controlled by the eye movement position monitored in real time and closely follows the mechanical dog moving to the right and the upper right direction. The recognition accuracy of this neural network shows an exponential upward trend during different training periods and finally approaches 100% ( Figure 14 and Figure 15 ). Through the camera equipped on the drone and its video feedback, personnel can perform precise control through the first-person perspective, providing insights for the next generation of eye-driven immersive and diverse human-computer interaction applications.

[0102] In summary, the above is only the preferred embodiment of the present invention and is not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A neuromorphic sensor based on a programmable photodiode, characterized in that: include: A substrate, a photosensitive material formed on the substrate, and an electrode covering the photosensitive material; The neuromorphic image sensor includes n×n pixels, each pixel includes n sub-pixels, and each sub-pixel is composed of a programmable photodiode; n≥3; The substrate is a glass substrate; The pattern of the photosensitive material on the substrate is: a line array with fixed line width and line spacing; The electrode is a gold electrode with a length of 300-500 microns, a width of 100-300 microns, and a thickness of 40-60 nanometers. The width of the channel between electrodes of each photodiode is 50-150 microns.

2. A neuromorphic sensor based on a programmable photodiode as claimed in claim 1, characterized in that: The photosensitive material is AgBiS2 material; the photosensitive material is patterned and grown on the substrate through a seed layer stimulation method.

3. A method for preparing a neuromorphic sensor based on a programmable photodiode, characterized in that: The steps include: S1: providing a glass substrate; S2: cleaning the glass substrate; the cleaning of the glass substrate comprises: using acetone, ethanol, and water to ultrasonically clean the glass substrate for 20-30 minutes respectively; S3: preparing a pixel array pattern on a glass substrate by photolithography; the pixel array pattern is: an n×n pixel array pattern, each pixel point is composed of n sub-pixels, each sub-pixel point is a photosensitive material part of a programmable photodiode, and the pattern of the photosensitive material part is a micrometer line array with a fixed line width and line spacing, n≥3; S4: placing the glass substrate with the pixel array pattern photoetched thereon into a thermal evaporation coating machine to evaporate silver, and obtaining a silver material pixel array pattern after peeling; S5: placing the glass substrate with the silver material pixel array pattern into a plasma cleaning machine for bombardment oxidation to form a silver oxide seed layer; S6: Spin-coat the AgBiS2 precursor solution onto the silver oxide seed layer using a spin coater. S7: placing the glass substrate spin-coated with the AgBiS2 precursor solution in a glove image and heating it to form an AgBiS2 photosensitive material; S8: photolithography a pixel array electrode pattern on the photosensitive material; S9: Place the substrate with the pixel array electrode pattern photolithographically formed in a thermal evaporation coating machine to evaporate the gold electrode, and after peeling off, a neuromorphic sensor with a programmable photodiode can be obtained.

4. A method for preparing a neuromorphic sensor based on a programmable photodiode as claimed in claim 3, characterized in that: The thickness of the evaporated silver material in S4 is 5-30 nanometers. The purpose of this step is to pattern the silver seed layer on the glass substrate. The subsequently grown AgBiS2 photosensitive material is only patterned and grown where the silver seed layer is present. The stripping process in S4 includes: soaking in acetone for 5-20 minutes.

5. The method for preparing a neuromorphic sensor based on a programmable photodiode according to claim 3, characterized in that: The oxidation parameters of the plasma cleaning machine in S5 are: bombarding the silver material for 10-30 minutes at an oxygen flow rate of 2-10 sccm to oxidize the evaporated silver material into silver oxide.

6. A method for preparing a neuromorphic sensor based on a programmable photodiode as claimed in claim 3, characterized in that: The preparation method of the AgBiS2 precursor solution in S6 is as follows: weigh AgCl, BiCl3, and thiourea, dissolve them in N,N-dimethylformamide solution, so that the molar ratio of AgCl to BiCl3 is 0-2:1, and stir to mix them evenly; The spin coating process in S6 is as follows: the substrate is placed in a coating machine and rotated at a speed of not less than 2000 revolutions per minute for 10-60 seconds, and 15-30 microliters of a well-mixed precursor solution is added dropwise during the rotation process.

7. The method for preparing a neuromorphic sensor based on a programmable photodiode according to claim 3, characterized in that: The heating parameters in the glove box in S7 are: placing the substrate on a hot plate at 150-200 degrees Celsius and heating it for 3-10 minutes.

8. The method for preparing a neuromorphic sensor based on a programmable photodiode according to claim 3, characterized in that: During the preparation of the sensor pixel array, steps S6-S7 need to be repeated 1-5 times to form a uniform and dense AgBiS2 film.

9. A method for preparing a neuromorphic sensor based on a programmable photodiode as claimed in claim 8, characterized in that: The pixel array electrode pattern in S8 is an n×n pixel array pattern, each pixel is composed of n sub-pixels, each sub-pixel is composed of a programmable photodiode, and the electrode portion dimensions are: length 300-500 microns, width 100-300 microns, thickness 40-60 nanometers. The channel between each photodiode electrode is 50-150 microns, n≥3.

10. The method for preparing a neuromorphic sensor based on a programmable photodiode according to claim 8, characterized in that: The thickness of the evaporated gold electrode in S9 is 40-60 nm; The stripping process in S9 is: soaking in acetone for 10-30 minutes; The patterned preparation method of the AgBiS2 material is achieved by using a silver seed layer stimulation method, that is, the final AgBiS2 material selectively grows only in the places where the silver material is evaporated on the substrate in step S4, and will not grow in the places where the silver material is not evaporated, thereby forming a 3×3×3 pixel array pattern on the substrate.

Images