Tellurium nanowire optical fusion multiplication and addition processing system and method based on body photovoltaic effect

Abstract

The invention discloses a tellurium nanowire optical fusion multiplication and addition processing system and method based on a body photovoltaic effect, and belongs to the technical field of photon calculation. Through the bulk photovoltaic effect of the spiral tellurium nanowires and the non-centrosymmetric structure, the fusion multiplication and addition operation is directly realized in the optical domain, an additional photoelectric detector (PD), an analog-to-digital converter (ADC) and an integrated circuit are not needed, and the photoelectric conversion delay and the power consumption are reduced by more than 90%; meanwhile, the influence of phase noise, thermal fluctuation and optical coupling loss is eliminated, and the system stability is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of photonic computing technology, specifically to a tellurium nanowire optical fusion multiply-accumulate processing system and method based on the bulk photovoltaic effect. Background Technology

[0002] In recent years, the rapid development of fields such as artificial intelligence, deep learning, and edge computing has placed higher demands on the performance of hardware processing units, especially in efficient, low-power, and scalable matrix operations. These applications typically require processing massive amounts of data and performing complex mathematical calculations. For example, matrix multiplication and addition operations play a central role in scenarios such as image recognition, natural language processing, and real-time decision-making. Traditionally, these tasks have been primarily handled by digital processing units, including central processing units (CPUs), graphics processing units (GPUs), and tensor processing units (TPUs). Although these processors have been continuously optimized in terms of architecture design and manufacturing processes, their basic operating principles still rely on the switching actions of electronic devices and clock synchronization mechanisms. This leads to increasingly prominent power consumption and latency issues when handling large-scale parallel computing, making it difficult to meet the ever-increasing computing demands.

[0003] In the fields of artificial intelligence and high-speed computing, the application of existing digital processing units faces significant challenges. Taking CPUs and GPUs as examples, when performing matrix operations, they need to switch and transmit electronic signals bit by bit. This process not only generates a lot of heat but is also limited by the upper limit of clock frequency, making it difficult to further improve energy efficiency and computing speed. As the scale of neural network models continues to expand, especially the multi-layer matrix operations involved in deep learning, the limitations of traditional electronic processors in terms of power consumption, heat dissipation, and response time become increasingly apparent. Although process miniaturization and architectural optimization (such as multi-core design) can alleviate these problems to some extent, the physical characteristics of electronic devices fundamentally restrict further performance improvements, leading to decreased system efficiency in high-concurrency tasks and making it difficult to support low-latency applications such as real-time edge computing. To address these challenges, optical computing, as an emerging solution, is gradually gaining attention. Optical computing uses photons instead of electrons to perform computational tasks. Its advantages lie in the high speed of optical signal transmission, low power consumption, and natural suitability for parallel processing. Existing optical computing architectures typically employ optical modules such as multi-mode interferometers (MMIs), Mach-Zehnder interferometers (MZIs), waveguide arrays, or microring resonators to implement basic computational functions. For example, multi-mode interferometers can perform linear transformations through the interference effect of light waves, while Mach-Zehnder interferometers are often used to modulate optical signals to simulate multiply-accumulate operations. These techniques attempt to improve computational efficiency through the high parallelism of optical devices and have demonstrated potential in tasks such as neural network inference in laboratory environments.

[0004] However, existing optical computing solutions still have significant shortcomings. These architectures often rely on numerous external components, such as photodetectors, modulators, and electro-optical conversion devices, to complete the interaction between optical and electrical signals, resulting in complex system structures, large sizes, and low integration. Optical modules themselves, such as microring resonators, are highly sensitive to manufacturing processes and environmental changes (e.g., temperature fluctuations), easily introducing computational errors and reducing reliability. Furthermore, current optical computing systems still require auxiliary processing with electronic devices when supporting nonlinear operations (such as activation functions in neural networks), which not only increases power consumption and latency but also limits their application in all-optical computing scenarios. Overall, existing technologies struggle to achieve a balance between computational density, energy efficiency, and scalability, failing to fully unleash the potential of optical computing in the field of artificial intelligence. Summary of the Invention

[0005] The purpose of this invention is to provide a tellurium nanowire optical fusion multiply-accumulate processing system and method based on the bulk photovoltaic effect, in order to solve the problem that the existing processing framework requires additional photoelectric conversion devices to realize fusion multiply-accumulate operations.

[0006] To achieve the above objectives, the present invention employs the following technical solution: This invention discloses a tellurium nanowire optical fusion multiply-accumulate processing system based on bulk photovoltaic effect, including a helical tellurium nanowire computing unit, a spatial light modulator, a laser source, a dual 4f focusing system, and a signal readout module; The helical tellurium nanowire computing unit includes helical tellurium nanowires; the helical tellurium nanowires are grown in suspension on an insulating substrate and have metal electrodes at both ends; the helical tellurium nanowires have a non-centrosymmetric structure. The laser beam is used to act on a spatial light modulator, causing the spatial light modulator to modulate the laser beam and encode the vector input of the image pixels into the optical power intensity of the laser beam and the spatial position of the laser beam on the helical tellurium nanowire. The dual 4f focusing system applies a laser beam with the aforementioned laser beam power intensity and spatial position to the helical tellurium nanowire computing unit, generating a photocurrent signal. The signal reading module is used to acquire the photocurrent signal output from both ends of the helical tellurium nanowire. This photocurrent signal is the result of optical fusion multiplication and addition operations, and image recognition is achieved through convolution operations and affine transformations performed by a convolutional neural network.

[0007] Furthermore, the metal electrodes are the source and drain electrodes, respectively; the length of the spiral tellurium nanowire is 300~600 μm; the metal electrodes are composed of Cr electrodes and Au electrodes; the thicknesses of the Cr electrodes and Au electrodes are 5~10 nm and 30~50 nm, respectively; the horizontal distance between the source and drain electrodes, i.e., the channel length, is 5 μm~80 μm.

[0008] Furthermore, the spatial light modulator is a phase-modulated reflective spatial light modulator; The vector consists of an input vector and a weight vector; the input vector includes image pixel grayscale values ​​and function parameters. The weight vector consists of convolution kernel weights and linear regression coefficients.

[0009] Furthermore, the laser source is a broadband tunable light source; the output laser wavelength of the laser source is 1100nm~2000nm.

[0010] Furthermore, the helical tellurium nanowire computing unit has a customizable linear response region with a length ranging from 10 μm to 600 μm; the system parallelism P is calculated according to the formula P = 2L / S. -1 The calculation is performed, where L is the length of the linear response region and S is the beam scanning step size; the value of S is 10 nm.

[0011] Furthermore, the responsivity of the helical tellurium nanowires is stable between 1.2 A / W and 125.2 A / W in the spectral range of 1100 nm to 2000 nm.

[0012] This invention also discloses an optical fusion multiply-add (OFMA) operation method based on helical tellurium nanowires, which uses the above-mentioned system and includes the following steps: The laser beam emitted from the laser source is applied to the spatial light modulator, and the vector input of the image pixels is encoded as the optical power intensity and the laser beam. The laser beam with the above-mentioned laser beam power intensity and spatial position is applied to the helical tellurium nanowire computing unit through a dual 4f focusing system. Then, under the local asymmetric illumination condition of the helical tellurium nanowire computing unit, a photocurrent signal proportional to the product of the laser beam power intensity and the spatial position is generated. The photocurrent signal output from both ends of the helical tellurium nanowire is collected by the signal acquisition module to realize the multiplication operation.

[0013] Furthermore, the laser beam emitted from the laser source is applied to the spatial light modulator, and the vector input of the image pixels is encoded as the optical power intensity and the laser beam. The laser beam with the aforementioned laser beam power intensity and spatial application position is applied to the helical tellurium nanowire computing unit through a dual 4f focusing system. The photocurrent signal is summed through carrier diffusion superposition, and the photocurrent signal output from both ends of the helical tellurium nanowire is acquired through the signal acquisition module to complete the addition operation.

[0014] The present invention also discloses an image classification method, which uses the above-described system and includes the following steps: S1: The handwritten digital image is divided into sub-image pixels. The laser beam emitted by the laser source acts on the spatial light modulator. The spatial light modulator modulates the incident laser beam and encodes the vector of each sub-image pixel into the laser beam power intensity and the spatial position of the laser beam on the helical tellurium nanowire. S2: The modulated beam is projected onto a designated spatial position on the spiral tellurium nanowire in the spiral tellurium nanowire computing unit through a dual 4f system, and then a photocurrent signal is generated; S3: After acquiring the photocurrent signal, the signal reading module feeds it back to the spatial light modulator to perform ReLU activation operation, uses the traditional convolutional neural network architecture to perform convolution operation and perform affine transformation, and outputs the image classification result.

[0015] Furthermore, in S1, the 15×15 pixel MNIST handwritten digit image is divided into 49 overlapping 3×3 pixel subarrays; In S2, a 3×3 convolution kernel weight is loaded through a spatial light modulator and encoded into 49 sets of beam spatial positions.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes the bulk photovoltaic effect and non-centrosymmetric structure of helical tellurium nanowires to directly achieve fused multiply-accumulate operations in the optical domain, eliminating the need for additional photodetectors (PDs), analog-to-digital converters (ADCs), and integrated circuits. This reduces photoelectric conversion delay and power consumption by more than 90%. At the same time, it eliminates the effects of phase noise, thermal fluctuations, and optical coupling losses, significantly improving system stability.

[0017] Furthermore, the spiral tellurium nanowires used in this invention have customizable lengths, which can be extended N times, achieving a calculated density of 16 TOPS / mm. 2 Far exceeding existing photonic processors (typically below 1 TOPS / mm) 2 This meets the needs of large-scale AI computing.

[0018] Furthermore, the helical tellurium nanowires possess a wide spectral response of 1100-2000 nm, a thermal stability range of room temperature to 140 ℃, and ultra-low noise characteristics. They do not require wavelength screening or thermal control modules and can operate stably in complex environments, reducing system deployment costs.

[0019] Furthermore, the fusion multiply-accumulate system based on helical tellurium nanowires can simultaneously perform matrix-vector multiplication and image classification convolution operations, with an image classification accuracy of up to 90%, providing an integrated photonic computing solution for fields such as artificial intelligence and signal processing. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the optical fusion multiply-add (OFMA) processing system based on helical tellurium nanowires according to the present invention; Figure 2 A scanning electron microscope of the OFMA device with source and drain electrodes of the present invention; Figure 3 The responsivity and detectivity of the optical computing unit of the helical tellurium nanowire under 1100nm-2000nm laser light; Figure 4 The photocurrent of the optical computing unit based on helical tellurium nanowires in this invention ranges from room temperature to 140°C; Figure 5 This is a conceptual diagram of the OFMA-based computing system of the present invention; Figure 6 This is a schematic diagram of the OFMA processing unit; Figure 7 The nanowire channel length versus the calculated density (FLOPs / mm) 2 The dependency on the operation count for each channel; Figure 8A convolutional neural network (CNN) architecture for MNIST image classification; Figure 9 For comparison of classification accuracy. Detailed Implementation

[0021] To enable those skilled in the art to understand the features and effects of the present invention, the terms and expressions used in the specification and claims are explained and defined in general below. Unless otherwise specified, all technical and scientific terms used herein have the ordinary meaning understood by those skilled in the art regarding the present invention, and in case of conflict, the definitions in this specification shall prevail.

[0022] The theories or mechanisms described and disclosed herein, whether right or wrong, should not in any way limit the scope of the invention, that is, the contents of the invention can be implemented without being limited by any particular theory or mechanism.

[0023] In this document, all features defined by numerical ranges or percentage ranges, such as numerical values, quantities, contents, and concentrations, are for the sake of brevity and convenience only. Accordingly, descriptions of numerical ranges or percentage ranges should be considered as covering and specifically disclosing all possible sub-ranges and individual numerical values ​​(including integers and fractions) within those ranges.

[0024] In this article, unless otherwise specified, “contains,” “includes,” “containing,” “has,” or similar terms cover the meanings of “composed of” and “mainly composed of,” for example, “A contains a” covers the meanings of “A contains a and others” and “A contains only a.”

[0025] For the sake of brevity, not all possible combinations of the technical features in each implementation scheme or embodiment are described herein. Therefore, as long as there is no contradiction in the combination of these technical features, the technical features in each implementation scheme or embodiment can be combined arbitrarily, and all possible combinations should be considered within the scope of this specification.

[0026] This invention provides an OFMA processing system based on helical tellurium nanowires, such as... Figure 1 As shown, where and Image pixels are encoded into the optical power intensity and spatial position of a laser beam using a spatial light modulator. All multiplications and accumulations are performed simultaneously within a single OFMA unit.

[0027] The helical tellurium nanowire computing unit includes a helical tellurium nanowire, an insulating substrate, and metal electrodes. The metal electrodes are grown on the surface of the insulating substrate by electron beam evaporation and serve as the source and drain electrodes, respectively. The aforementioned helical tellurium nanowires are grown in suspension on an insulating substrate, which includes silicon oxide and silicon nitride. The aforementioned metal electrodes are composed of Cr+Au, with thicknesses of 5~10 nm and 30~50 nm, respectively; and the horizontal distance between the electrodes, i.e., the channel length, is 5 μm~80 μm. The spiral tellurium nanowires have a non-centrosymmetric structure and a length of 300~600 μm; The laser source is used to generate the photocurrent signal of the helical tellurium nanowire; the helical tellurium nanowire has a non-centrosymmetric structure, and the polarity of the photocurrent can be programmably switched through local illumination and non-local illumination. Image grayscale values ​​and convolution kernel weight vectors are encoded as the optical power intensity of the laser beam and the spatial position of the laser beam on the helical tellurium nanowire; and the dual 4f focusing system is used to focus the laser beam with a certain optical power intensity modulated by the spatial light modulator and project it onto a specified spatial position on the helical tellurium nanowire computing unit to generate a photocurrent signal. The laser beam is used to act on a spatial light modulator, causing the spatial light modulator to modulate the laser beam and encode the vector input of the image pixels into the optical power intensity of the laser beam and the spatial position of the laser beam on the helical tellurium nanowire. The signal reading module is used to acquire the photocurrent signal output from both ends of the helical tellurium nanowire. This photocurrent signal is the result of optical fusion multiplication and addition operations, and image recognition is achieved through convolution operations and affine transformations performed by a convolutional neural network (CNN).

[0028] Furthermore, the spatial light modulator is a phase-modulated reflective spatial light modulator; the vector is an input vector and a weight vector; the input vector includes image pixel gray levels and function parameters; the weight vector is convolution kernel weights and linear regression coefficients.

[0029] Furthermore, the laser source is a broadband tunable source capable of outputting laser light with a wavelength range covering 1100nm~2000nm.

[0030] Furthermore, the helical tellurium nanowire computing unit has a customizable linear response region with a length ranging from 10 μm to 600 μm; the system parallelism P is calculated according to the formula P = 2L / S. -1 The calculation is performed, where L is the length of the linear response region and S is the beam scanning step size; the value of S is 10 nm.

[0031] Furthermore, the helical tellurium nanowires maintain stable thermal emission behavior within a temperature range of room temperature to 140 °C, requiring no additional thermal control device, such as... Figure 4As shown; its current noise power spectral density is as low as tens of femtoamperes / square hertz and its responsivity is stable between 1.2 A / W and 125.2 A / W in the spectral range of 1100 nm to 2000 nm, such as Figure 3 As shown.

[0032] This invention also discloses an optical fusion multiplication-addition method based on helical tellurium nanowires, comprising the following steps: The input vector is encoded as beam intensity using a spatial light modulator, and the weight vector is encoded as the spatial position of the beam on the spiral tellurium nanowire. By using a dual 4f system, the spatial light modulator focuses the encoded light beam onto the spatial position of the spiral tellurium nanowire. Under local asymmetric illumination conditions, a photocurrent proportional to the product of the input vector (i.e., light intensity) and the weight vector (spatial position of the nanowire) is generated, thus realizing the multiplication operation. When multiple laser beams modulated by the spatial light modulator act on the spiral tellurium nanowire simultaneously, the photocurrent is summed through carrier diffusion and superposition, thus completing the addition operation and directly outputting the total photocurrent signal corresponding to the vector dot product.

[0033] This invention also discloses an image classification method, comprising the following steps: S1: Data preprocessing: The 15×15 pixel MNIST handwritten digit image is divided into 49 overlapping 3×3 pixel subarrays, and the gray value of each pixel is encoded into the intensity of laser (1100~1550 nm laser spectrum range) through a spatial light modulator. S2: Weight setting: Load 3×3 convolution kernel weights through a spatial light modulator and encode them into 49 sets of beam spatial positions; S3: Parallel convolution operation: The dual 4f system synchronously projects 49 sets of input subarray beams (encoded as optical power intensity) and weight (spatial position of the spiral tellurium nanowire) beams onto 49 channels of the spiral tellurium nanowire. Each channel outputs the dot product photocurrent of the corresponding subarray and convolution kernel. S4: Activation and Output: After the signal reading module acquires the photocurrent, it feeds it back to the spatial light modulator to perform the ReLU activation operation (i.e., return to zero negative current), and then performs affine transformation through 10 selected channels to output the image classification result; the image classification accuracy of this method is not less than 90%.

[0034] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0035] The fabrication method of the helical tellurium nanowire computing unit is as follows: Step 1: Select silicon dioxide as the insulating substrate and clean it.

[0036] Step 2: The metal electrode (Cr (5nm) + Au (30nm)) is prepared by electron beam evaporation, with a channel length of 10 μm.

[0037] Step 3: Subsequently, the grown helical tellurium nanowires were precisely transferred onto the metal electrode using a dry transfer method, resulting in a helical tellurium nanowire computing unit; the electron microscope image of the helical tellurium nanowire computing unit is shown below. Figure 2 As shown.

[0038] Example 2 A single-channel optical fusion multiplication-accumulation operation method is as follows: Step 1: Using a 1920×1080 resolution phase-modulated reflective spatial light modulator, the image grayscale values ​​and convolution kernel weight vectors are encoded into the optical power intensity of a 1550 nm laser beam and the linear spatial position of the laser beam on the helical tellurium nanowire. This is then precisely focused onto the helical tellurium nanowire using a dual 4f focusing system. Step 2: Illumination mode control: When positive or negative weights are required, a local asymmetric illumination mode is used to make the diffusion effect of photogenerated carriers dominant; when zero weights are required or diffusion components are suppressed, a uniform symmetric illumination mode is used to make the carrier drift effect dominant.

[0039] Step 3: Fusion Multiply-Accumulate Calculation: When multiple beams of light simultaneously illuminate the same helical tellurium nanowire channel, the photogenerated carriers generated by different light spots are linearly superimposed inside the nanowire. The total photocurrent measured at the output terminal satisfies: Iout=Σ( k ·Ai·Bi) in k This is a scaling factor related to the material and device structure.

[0040] Example 3 In this embodiment, a single helical tellurium nanowire is divided into multiple independent computational channels along the axial direction, each channel being 10 μm in length, with multiple channels sharing the same helical tellurium nanowire. Multiple light spot arrays are simultaneously generated at different channel positions using a spatial light modulator. Each channel independently performs the aforementioned fusion multiplication-addition operation once, thereby performing multi-path vector inner product calculations in parallel within a single illumination command. As the length of the helical tellurium nanowire increases, the number of parallelizable channels also increases, significantly improving computational parallelism and computational density. Figure 7 .

[0041] Example 4 The following is an application of handwritten image classification on the MNIST dataset using helical tellurium nanowire computing units: Figure 5 This schematically illustrates how 49 input subarrays interact with the convolution kernel through 49 helical tellurium nanowire computational units, each of which integrates a tellurium nanowire channel. Figure 6 Sub-image pixel values ​​are encoded as light intensity, and kernel weights are encoded as spatial locations. Each helical tellurium nanowire computational unit outputs a photocurrent Y. i =∑x ij ·w ij Corresponding to the dot product between input vectors, the resulting analog photocurrent is digitized and mapped back to the spatial light modulator, where the rectified linear unit (ReLU) activation function is applied by suppressing negative photocurrent values ​​and re-encoding the positive output into a new light intensity for use in the next network layer. As shown in the network architecture ( Figure 8 The first, second, and third layers require 49, 25, and 9 convolution operations, respectively. In our experiments, due to equipment limitations, these convolution operations were performed sequentially using a single helical tellurium nanowire computing unit, and image classification was successfully achieved through 10 affine transformations, such as... Figure 9 As shown, the classification accuracy achieved by the OFMA architecture (90.0%) is very close to that of the CPU (92.7%).

[0042] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A tellurium nanowire optical fusion multiply-accumulate processing system based on bulk photovoltaic effect, characterized in that, It includes a helical tellurium nanowire computing unit, a spatial light modulator, a laser source, a dual 4f focusing system, and a signal readout module; The helical tellurium nanowire computing unit includes helical tellurium nanowires; the helical tellurium nanowires are grown in suspension on an insulating substrate and have metal electrodes at both ends; the helical tellurium nanowires have a non-centrosymmetric structure. The laser beam is used to act on a spatial light modulator, causing the spatial light modulator to modulate the laser beam and encode the vector input of the image pixels into the optical power intensity of the laser beam and the spatial position of the laser beam on the helical tellurium nanowire. The dual 4f focusing system applies a laser beam with the aforementioned laser beam power intensity and spatial position to the helical tellurium nanowire computing unit, generating a photocurrent signal. The signal reading module is used to acquire the photocurrent signal output from both ends of the helical tellurium nanowire. This photocurrent signal is the result of optical fusion multiplication and addition operations, and image recognition is achieved through convolution operations and affine transformations performed by a convolutional neural network.

2. The tellurium nanowire optical fusion multiply-accumulate processing system based on bulk photovoltaic effect according to claim 1, characterized in that, The metal electrodes are the source and drain electrodes, respectively; the length of the spiral tellurium nanowire is 300~600 μm; the metal electrodes are composed of Cr and Au electrodes; the thicknesses of the Cr and Au electrodes are 5~10 nm and 30~50 nm, respectively; the horizontal distance between the source and drain electrodes, i.e., the channel length, is 5 μm~80 μm.

3. The tellurium nanowire optical fusion multiply-accumulate processing system based on bulk photovoltaic effect according to claim 1, characterized in that, The spatial light modulator is a phase-modulated reflective spatial light modulator; The vector consists of an input vector and a weight vector; the input vector includes image pixel grayscale values ​​and function parameters. The weight vector consists of convolution kernel weights and linear regression coefficients.

4. The tellurium nanowire optical fusion multiply-accumulate processing system based on bulk photovoltaic effect according to claim 1, characterized in that, The laser source is a broadband tunable light source; the output laser wavelength of the laser source is 1100nm~2000nm.

5. The tellurium nanowire optical fusion multiply-accumulate processing system based on bulk photovoltaic effect according to claim 1, characterized in that, The helical tellurium nanowire computing unit has a customizable linear response region with a length ranging from 10 μm to 600 μm; the system parallelism P is calculated according to the formula P=2L / S. -1 The calculation is performed, where L is the length of the linear response region and S is the beam scanning step size; the value of S is 10 nm.

6. The tellurium nanowire optical fusion multiply-accumulate processing system based on bulk photovoltaic effect according to claim 1, characterized in that, The responsivity of the spiral tellurium nanowires is stable between 1.2 A / W and 125.2 A / W in the spectral range of 1100 nm to 2000 nm.

7. An optical fusion multiplication and addition method based on helical tellurium nanowires, characterized in that, The process, performed using the system described in any one of claims 1 to 6, includes the following steps: The laser beam emitted from the laser source is applied to the spatial light modulator, and the vector input of the image pixels is encoded as the optical power intensity and the laser beam. The laser beam with the above-mentioned laser beam power intensity and spatial position is applied to the helical tellurium nanowire computing unit through a dual 4f focusing system. Then, under the local asymmetric illumination condition of the helical tellurium nanowire computing unit, a photocurrent signal proportional to the product of the laser beam power intensity and the spatial position is generated. The photocurrent signal output from both ends of the helical tellurium nanowire is collected by the signal acquisition module to realize the multiplication operation.

8. The optical fusion multiplication and addition method based on helical tellurium nanowires according to claim 7, characterized in that, The laser beam emitted from the laser source is applied to the spatial light modulator, and the vector input of the image pixels is encoded as the optical power intensity and the laser beam. The laser beam with the above-mentioned laser beam power intensity and spatial position is applied to the helical tellurium nanowire computing unit through a dual 4f focusing system. The photocurrent signal is summed through carrier diffusion superposition, and the photocurrent signal output from both ends of the helical tellurium nanowire is collected by the signal acquisition module to complete the addition operation.

9. An image classification method, characterized in that, The process, performed using the system described in any one of claims 1 to 6, includes the following steps: S1: The handwritten digital image is divided into sub-image pixels. The laser beam emitted by the laser source acts on the spatial light modulator. The spatial light modulator modulates the incident laser beam and encodes the vector of each sub-image pixel into the laser beam power intensity and the spatial position of the laser beam on the helical tellurium nanowire. S2: The modulated beam is projected onto a designated spatial position on the spiral tellurium nanowire in the spiral tellurium nanowire computing unit through a dual 4f system, and then a photocurrent signal is generated; S3: After acquiring the photocurrent signal, the signal reading module feeds it back to the spatial light modulator to perform ReLU activation operation, uses the traditional convolutional neural network architecture to perform convolution operation and perform affine transformation, and outputs the image classification result.

10. The image classification method according to claim 9, characterized in that, In S1, the 15×15 pixel MNIST handwritten digit image is divided into 49 overlapping 3×3 pixel subarrays; In S2, a 3×3 convolution kernel weight is loaded through a spatial light modulator and encoded into 49 sets of beam spatial positions.

Images