A D 2 NN processor with polarization, distance, and rotation multi-dimensional multiplexing and its implementation method

Abstract

The present invention relates to a D<supgt;2< / supgt>NN processor for polarization, distance, and rotation multi-dimensional multiplexing and its implementation method. The processor includes a spatial structured light generation layer and multiple metasurface layers. By constructing a diffraction neural network to simulate the propagation process of terahertz waves, and combining the gradient descent and backpropagation algorithms to optimize the phase distribution of the multiple metasurface layers, two tasks of joint training of "x polarization - distance - rotation" and "y polarization - distance - rotation" are completed respectively. Using the propagation phase principle, the geometric parameters of the micron-scale silicon pillar units construct independent diffraction channels in the x and y polarization states to achieve polarization multiplexing. At the same time, a spatial displacement and rotation regulation mechanism of the metasurface layer is introduced to construct a multi-dimensional information processing path and improve the information processing ability of the system. The problems existing in D<supgt;2< / supgt>NN, such as limited task carrying capacity, insufficient information security, and low hardware integration, are solved. The processor of the present invention has the advantages of flexible design, compact volume, and high integration.

Description

Technical Field

[0001] The present invention belongs to the cross - field of optical computing and artificial intelligence, and particularly relates to an all - optical diffraction deep neural network (D 2 NN) processor based on a multi - dimensional multiplexing mechanism, and more particularly to a system and method for realizing multi - functional, highly integrated and highly secure information transmission by integrating multi - dimensional parameters of polarization, distance and rotation. Background Art

[0002] All - optical diffraction deep neural network (D 2 NN) has significant advantages in improving data processing speed and reducing power consumption, injecting new impetus into the development of optical computing and artificial intelligence. However, current D 2 NN still faces significant challenges in terms of task processing capacity, system flexibility and integration level. In existing research, most D 2 NN only implement specific task processing with low functional integration, making it difficult to meet the requirements of efficient intelligent processing in complex scenarios. Innovative solutions such as the plug - and - play diffraction neural network optimization method and task recognition device based on metasurface (Publication No.: CN116596050A) achieve the switching of two recognition tasks by switching the plug - and - play components implemented by metasurface in the network, improving the flexibility of the plug - and - play diffraction neural network while reducing computational resource consumption and training time. Another patent, a metasurface device and implementation method for dual - target spatial position and image recognition (Publication No.: CN118484891A), synchronously recognizes the spatial positions and image contents of two targets, enhancing the processing efficiency of multi - target detection tasks; and demonstrates various D 2 NN with function switching using the inherent property of light wavelength, expanding the processing information capacity. Although the above - mentioned work has made breakthroughs in function expansion, there are still the following limitations: low functional integration: the task capacity of existing D 2 NN systems is limited and the structural compactness is insufficient; weak information security: the limited information processing dimension of D 2 NN systems leads to high predictability in the data transmission process, making it easy to be stolen or reconstructed. Lack of redundancy protection and physical - layer encryption mechanisms makes it difficult to meet the requirements of high - security information processing; insufficient hardware compatibility: existing solutions rely on complex optical devices and mechanical replacements, restricting system integration and application expansion. Summary of the Invention

[0003] Aiming at the problems of limited task - carrying capacity, insufficient information security and low hardware integration existing in current D 2 NN, the present invention proposes an all - optical diffraction neural network processor based on multi - dimensional multiplexing of polarization, distance and rotation and its implementation method. This processor integrates three physical degrees of freedom into a unified D 2In the NN architecture, a multi-dimensional information processing framework with an adjustable task path is constructed, significantly enhancing the information capacity of the system and reducing interference between different tasks.

[0004] The technical solution of the present invention is as follows:

[0005] A polarization, distance, and rotation multi-dimensional multiplexing D 2 NN processor, including, arranged in sequence along the incident light propagation direction: a spatial structured light generation layer for amplitude modulation patterning of the incident light to achieve spatial encoding of the incident light field; multiple metasurface layers, each layer including a micron-scale silicon pillar unit array arranged in a preset order for achieving multi-dimensional diffraction control; wherein, the geometric parameters of the micron-scale silicon pillar units are designed based on the propagation phase principle, and independent diffraction channels are constructed in the x and y polarization states respectively to achieve polarization multiplexing; the phase distribution of the multiple metasurface layers is obtained by training a diffraction neural network model, and the training process uses gradient descent and backpropagation algorithms, and a dual-channel polarization decoupling training strategy is adopted to complete the joint training tasks of "x polarization - distance - rotation" and "y polarization - distance - rotation" respectively, generating the optimal diffraction phase distribution D 2 NN X and D 2 NN Y , and by independently or cooperatively controlling the polarization state of light, adjusting the interlayer distance and rotation mode, multiple optical computing task processing under different channels of polarization, distance, and rotation multi-dimensional separation or multiplexing is achieved.

[0006] Further, the micron-scale silicon pillar units are designed with independent phase responses in the x and y polarization states respectively. By scanning different geometric parameters, the corresponding transmission coefficients and phase distributions are obtained, and the phase information is assigned to the length and width dimensions of the micron-scale silicon pillars using the propagation phase principle, thereby achieving function separation and multiplexing under different polarization states; the length and width of the micron-scale silicon pillar units in the micron-scale silicon pillar unit array respectively achieve x the training phase distribution under polarization and y the training phase distribution under polarization, thereby achieving x the function switching of rotation multiplexing under polarization and y the function switching of distance multiplexing under polarization.

[0007] Further, by scanning the length and width parameters of the micron-scale silicon pillar unit structure under the conditions of x-polarized incident light and y-polarized incident light respectively, the transmission spectra and phase responses of the micron-scale silicon pillar unit structures of different sizes are obtained, and the metasurface micron-scale silicon pillar unit structures that meet the requirements are selected from them to achieve phase control for polarization separation.

[0008] Further, the spatial structured light generation layer adopts any one of the following forms:

[0009] Metal imaging plate: a transparent substrate made of silicon dioxide material with a thickness of 500 microns; a patterned metal film deposited on the upper surface of the transparent substrate, which is a gold film with a thickness of 5 microns; wherein, the patterned metal film allows terahertz waves to transmit through the patterned area, and the remaining metal areas block terahertz waves, thereby realizing the spatial modulation of the amplitude of the incident light field and serving as the input signal encoding structure of the diffraction neural network;

[0010] Programmable spatial light modulator SLM: used to dynamically load the input image pattern and modulate the amplitude of the terahertz wave incident light field.

[0011] Furthermore, the metasurface layer is prepared by deep silicon etching process, using silicon wafer material. The overall thickness of the silicon wafer is 1000 microns, the thickness of the deep silicon etching is 500 microns, and the lateral length and width of each metasurface device are 13200 microns; the period of the metasurface micron-scale silicon pillar unit structure obtained after deep silicon etching is 110 microns, and its thickness is 500 microns; it operates at the set working frequency of 0.7 THz, and the number of its arrays is 120*120.

[0012] Furthermore, the distance-rotation combination method includes four schemes: none, distance, rotation, distance-rotation; by combining and adjusting the spatial structured light generation layer, switching the polarization state of the incident light, changing the spatial distance between each metasurface layer and the detection plane, or by rotating the metasurface layer, the optical path of the light field is changed to achieve optical computing tasks with different functions. The optical computing tasks include basic mathematical operations, different types of single / multi-target recognition, image analysis and processing, and signal encoding functions.

[0013] Furthermore, the incident light is terahertz wave.

[0014] A method for implementing a polarization, distance, and rotation multi-dimensional multiplexing D 2 NN processor, used for the implementation of the polarization, distance, and rotation multi-dimensional multiplexing D 2 NN processor, which converts the target input image into an amplitude-modulated pattern of the incident light, and sequentially passes through multiple diffraction layers to obtain the logical calculation result or recognition output on the perception layer; the specific training process is as follows:

[0015] Use the Rayleigh-Sommerfeld diffraction integral formula to simulate the process of the light field propagating from the nth layer to the n+1th layer; wherein, represents the transmission coefficient of the nth layer, where A is the amplitude, φ is the phase, i is the imaginary unit, and a = 1 when only training a pure-phase diffraction neural network; the light field distribution in the form of the Rayleigh-Sommerfeld diffraction theory propagation is expressed as:

[0016] (1)

[0017] Among them, U(r n ) is the transmitted electric field distribution of the nth layer, is the transmitted electric field distribution on the (n + 1)th layer, n is the number of diffraction layers. In the formula, h(r n+1 -r n ) represents the impulse response during the transfer process, and the expression is:

[0018] (2)

[0019] Among them, is the Euclidean distance from the point r n on the nth layer to the point r n+1 on the (n + 1)th layer, λ is the incident wavelength, i = is the imaginary unit, x n+1 , y n+1 , z n+1 are respectively the spatial coordinate components of a certain point on the (n + 1)th layer , , are respectively the spatial coordinate components of the ath neuron on the nth layer; S represents the integration region, which is all points within the entire wavefront area; this integration represents the weighted sum and superposition of the light field distribution of the entire nth layer to calculate the light field distribution of a certain point on the (n + 1)th layer; through this method, the propagation of light waves in media with different layer structures can be predicted and calculated; the number of diffraction layers is represented by n; by setting the propagation distance and rotation method (single layer / multilayer rotation of 180°) between different diffraction layers, a corresponding number of calculation channels are formed, which can correspond to optical calculation tasks with different functions. During a single training for a single polarization, the multi-layer phase distribution under different propagation distance conditions and rotation methods is optimized adaptively at the same time. The gradient is calculated using the backpropagation algorithm, and the phase is iteratively updated in combination with the gradient descent method until the loss function converges to obtain the optimal phase result. This training process ensures that the network can output the corresponding optical calculation task results at the sensing layer, i.e., the detection plane, under different diffraction distances and rotation methods.

[0020] Furthermore, the diffraction neural network model is developed based on the Python language and constructed and trained using the Pytorch deep learning framework; during the training process, the neuron size is set to 110 microns, 8 samples are input in batches, and the phase weights are independently trained for each polarization state. The joint training method is used in each independent training process to ensure that the functions between channels do not interfere with each other.

[0021] Furthermore, the transmission characteristics and phase response of micron-scale silicon pillar units with different structural sizes under x- and y-linearly polarized light irradiation are simulated and analyzed through an electromagnetic full-wave simulation software to support D 2Diffraction Performance Modeling and Functional Verification of NN Processor in Multiple Dimensions.

[0022] The beneficial effects of the present invention are as follows:

[0023] The present invention takes a three-layer micron-scale metasurface layer prepared by deep silicon etching process as a platform, integrates a multi-dimensional all-optical diffraction processor architecture of three physical dimensions of polarization, distance and rotation, and successfully demonstrates various types of terahertz optical computing tasks in multi-channel dimensions, covering basic mathematical operations, different categories of single / multi-target multi-category recognition, image analysis and processing, and signal encoding functions. Further, a high-security information transmission system is constructed by allocating a large amount of input information through multi-dimensional channels. This all-optical diffraction processor integrating the dimensions of polarization, displacement and rotation is expected to achieve important applications in the fields of next-generation optical storage, optical communication and more advanced optical devices due to its versatility advantages. The specific improvements are as follows:

[0024] 1. High functional integration degree. Through the multiplexing of three dimensions of polarization, displacement and rotation, it supports various types of terahertz optical computing tasks, covering basic mathematical operations (such as logical operations in Example 1), different categories of single / multi-target multi-category recognition (such as handwritten digits / fashion items in Example 1), image analysis and processing, signal encoding (such as Morse code in Example 2) functions, etc., and the task capacity is increased by more than 3 times;

[0025] 2. Strong controllability of physical parameters. By adjusting the polarization state of the incident light, the distance between diffraction layers and the rotation angle, the phase distribution can be flexibly changed to realize the dynamic switching of computing functions, and the versatility and reconstruction ability of the system are enhanced;

[0026] 3. Strong manufacturability and trainability. The metasurface structure is based on silicon materials and conventional micro-nano manufacturing processes, and has good processability; the training method is based on the open-source Pytorch platform and combines MNIST-like datasets, and has good transferability and engineering implementation foundation;

[0027] 4. Compact system architecture. Compared with the existing multi-wavelength or multi-layer independent structures, the present invention uses a single terahertz wave source and a compact metasurface to realize complex optical computing tasks, and has significant miniaturization and system integration advantages. Description of the Drawings

[0028] Figure 1 It is a three-dimensional schematic diagram of a D 2 NN processor with multi-dimensional multiplexing of polarization, distance and rotation according to an embodiment of the present invention;

[0029] Figure 2 It is a D 2Schematic diagram of the dynamic phase design of the NN processor; (a) Schematic diagram of polarization multiplexing of the transmission phase; (b) Schematic diagram of the unit structure of the transmission phase; (c) Top view of the unit structure of the transmission phase; (d) x Transmission phase distribution when polarized light is incident; (e) y Transmission phase distribution when polarized light is incident; (f) x Transmittance distribution when polarized light is incident; (g) y Transmittance distribution when polarized light is incident;

[0030] Figure 3 In the embodiment of the present invention y Simulation result diagram of the switching between AND gate and OR gate of logic gates under the distance regulation between the polarization data encoding layer, the first metasurface, the second metasurface, the third metasurface, and the sensing plane; (a) Schematic diagram of the AND operation design of the logic gate; (b) Schematic diagram of the OR operation design of the logic gate; (c) Simulation results of the AND gate and OR gate calculations for the input "0" "0" in the distance dimension between the data encoding layer, the first metasurface, the second metasurface, the third metasurface, and the sensing plane; (d) Simulation results of the AND gate and OR gate calculations for the input "0" "1" in the distance dimension between the data encoding layer, the first metasurface, the second metasurface, the third metasurface, and the sensing plane; (e) Simulation results of the AND gate and OR gate calculations for the input "1" "0" in the distance dimension between the data encoding layer, the first metasurface, the second metasurface, the third metasurface, and the sensing plane; (f) Simulation results of the AND gate and OR gate calculations for the input "1" "1" in the distance dimension between the data encoding layer, the first metasurface, the second metasurface, the third metasurface, and the sensing plane;

[0031] Figure 4 In the embodiment of the present invention x Simulation result diagram of the switching between dual-digit recognition and dual-fashion item recognition under the regulation of the second and third metasurfaces of polarization rotation; (a) Schematic diagram of the dual-target digit recognition design; (b) Schematic diagram of the dual-target fashion item recognition design; (c) Simulation results of the dual-target digit recognition; (d) Simulation results of the dual-target fashion item recognition after rotating the second and third metasurfaces by 180 degrees; (e) Accuracy curve of diffraction neural training; (f) Confusion matrix calculated by the diffraction neural network;

[0032] Figure 5The application process and simulation results of the multi-dimensional task path regulation mechanism based on the joint control of polarization-distance-rotation three degrees of freedom in the physical layer encryption information transmission system in the embodiments of the present invention; (a) Input information and cipher book; (b) Schematic diagram of the secret key set in the system. It consists of three types of physical parameters: polarization state, the distance between diffraction layers, and the rotation angle of some metasurface layers. The three jointly determine the selection of the task path; (c) Schematic diagram of the secret key and simulation result diagram of the physical layer encryption information transmission of the multi-dimensional task path regulation mechanism; (d) Information decryption result diagram under the correct key configuration.

[0033] Attached drawing reference signs:

[0034] 1. Spatial structured light generation layer; 2. Micron-scale silicon pillar unit array structure. Specific implementation manners

[0035] The present invention will be described in detail below with reference to the attached drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and gives detailed implementation manners and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0036] Implementation principle / design concept:

[0037] Due to the principle of dynamic phase, the responses of the metasurface unit structure to the x-polarized and y-polarized incident terahertz waves are completely independent in physical implementation and have natural isolation. Therefore, the present invention considers introducing the polarization state as an independent multiplexing dimension to significantly improve the information processing density, and the polarization control technology is mature. The polarization state of the incident wave can be precisely controlled through a polarizer, etc., and it is easy to implement; secondly, the wavefront reconstruction results at different propagation distances are significantly different, which can be regarded as a "parameter with adjustable optical path". The diffraction field is guided to evolve multiple recognition results on different planes by using the regulation ability of the propagation distance, which is equivalent to increasing the network function dimension; by spatially rotating some diffraction layers, the structural symmetry is broken, and different diffraction responses are introduced. The metasurface layer with rotation has strong operability. Combining the three independent and easily controllable factors of polarization, distance, and rotation to achieve a multi-dimensional multiplexing D 2 NN processor, which has the characteristics of low crosstalk, high multiplexing, and strong scalability, significantly improves the task processing ability and information capacity of the system, and realizes the switchable multi-task processing ability under the same D 2 NN architecture, breaking through the task-bearing bottleneck of the traditional single-dimensional D 2 NN.

[0038] Overall overview:

[0039] The D of the present invention 2The NN processor includes a spatial structured light generation layer and multiple metasurface layers arranged in sequence along the propagation direction of the incident light. The metasurface layer includes an array of microscale silicon pillar units fabricated by deep silicon etching, which can achieve sub-wavelength phase modulation of the incident terahertz wave. The processor is trained using a deep neural network algorithm to complete different task responses under specific polarization states, distances, and rotation configurations, realizing the intelligent terahertz optical computing function of a single device for multi-task switching. The spatial structured light generation layer uses a metal imaging plate or a spatial light modulator to modulate the amplitude of the light field, extract the electric field distribution intensity on the detection plane, and can achieve various types of multi-modal terahertz optical computing tasks. By adjusting the interlayer distance and rotation angle, the system can flexibly switch the computing function and achieve multi-task multiplexing. The training process of the diffraction neural network uses the gradient descent algorithm and the backpropagation algorithm, adopting a dual-channel polarization decoupling training strategy. Under the channel dimension allocation of "x polarization - rotation - distance" and "y polarization - rotation - distance" respectively, various types of multi-modal terahertz optical computing tasks are realized. The "rotation - distance" regulation dimension can be used independently or collaboratively, supporting flexible optimization of the multi-dimensional parameter space through separate or combined regulation. Therefore, for the training under each polarization state, the joint training method is used, and two independent joint trainings are carried out separately to achieve the optimal phase distribution design under different degrees of freedom, and D 2 NN X and D 2 NN Y are obtained respectively, corresponding to realizing the terahertz optical computing tasks of different channels. The trained diffraction structure can dynamically switch different computing task paths according to the polarization state, interlayer distance, or interlayer rotation angle of the input light, thereby realizing the fusion of multiple information processing functions in the same system.

[0040] Embodiment 1:

[0041] The present invention proposes a D 2 NN processor with polarization, distance, and rotation multi-dimensional multiplexing and its implementation method. The structure of the multi-dimensional D 2 NN processor includes a spatial structured light generation layer and three metasurface layers arranged in sequence along the propagation direction of the incident light. The three metasurface layers are arranged in sequence along the propagation direction of the incident light. Each metasurface layer includes a microscale silicon pillar unit array and a dielectric substrate arranged in sequence along the propagation direction of the incident light. Applying the propagation phase, the length and width of the microscale silicon pillar units in the microscale silicon pillar unit array respectively achieve x the training phase distribution under polarization and y the training phase distribution under polarization, thereby realizing x the function switching of rotation multiplexing under polarization and y the function switching of distance multiplexing under polarization, so as to realize the overall polarization, distance, and rotation multi-dimensional D 2NN processor.

[0042] The diffractive neural network is a computational model that combines optical physics and artificial intelligence. Its core idea is to simulate the process of light propagation and diffraction in space and construct an all-optical system analogous to traditional neural networks. In this model, light waves are regarded as carriers of information and are transmitted between neuron layers in the medium. Its computational mechanism is based on the Huygens-Fresnel principle, that is: each point on a wavefront can be regarded as a new secondary wave source. At a position with a propagation distance of r, the electric field generated by the secondary wave source can be expressed as:

[0043]

[0044] E(x, y, z) is the complex electric field vector at point (x, y, z); w(x, y) is the amplitude weighting function of the incident wave on the (x, y) plane; A(x, y, z) is the amplitude modulation factor during the propagation process; φ(x, y, z) is the phase distribution at this point; e j φ(x,y,z) is the phase modulation factor, indicating the influence of the phase on the electric field. j is the imaginary unit. These secondary waves propagate forward in space to form a new wavefront, thus realizing the layer-by-layer transmission of information. These diffractive layers can be of transmissive or reflective structures. Each spatial position unit within the layer is equivalent to an optical "neuron", which modulates the input signal by controlling the phase and / or amplitude of the transmitted or reflected light. During the propagation of light waves, the diffractive patterns formed by different materials and structures can regulate and process the light field information to complete tasks such as image classification, logical operations, and target recognition. The training of the diffractive neural network is carried out on a computer relying on deep learning, including multiple rounds of training and parameter adjustment to optimize the network structure and improve performance.

[0045] A D 2 NN processor with polarization, distance, and rotation multi-dimensional multiplexing and its implementation method can achieve polarization, distance, and rotation multi-dimensional multiplexing, construct multiple different calculation channels in the same physical system, and realize terahertz optical calculations with multiple different functions. The optical calculation tasks include basic mathematical operations (Boolean logic operations for processing logic gate operations, simple mathematical calculations for performing addition, subtraction, multiplication, and division), single / multi-target recognition of different categories (handwritten digits, fashion items, handwritten English letters), image analysis and processing (non-destructive imaging to maintain the original quality of the image, image edge detection to extract the contour features of the image, digital and image mutual conversion to encode digital data into an image or parse digital information from an image), and signal encoding (converting information into a specific format encoding such as Morse code) functions.

[0046] The metasurface of the present invention has a single periodic structure, including a micron-scale silicon pillar unit structure with deep silicon etching for phase regulation. By scanning the length and width parameters of the micron-scale silicon pillar unit structure under the conditions of x-polarized incident light and y-polarized incident light respectively, the transmission spectra and phase responses of the micron-scale silicon pillar unit structures with different sizes are obtained, and the metasurface micron-scale silicon pillar unit structures that meet the requirements (the transmittance under x and y polarization is higher than 0.7 and the phase response meets the set requirements) are selected to achieve polarization-separated phase regulation.

[0047] The metasurface micron-scale silicon pillar unit is made of high-resistivity silicon material, the unit structure period is 110 microns, and its thickness is 500 microns. It operates at a set working frequency of 0.7 THz, and the number of its arrays is 120*120. The dielectric substrate is made of silicon material, its thickness is 500 microns, and the overall length and width dimensions of the structure are 13200 microns.

[0048] One implementation method in the spatial structured light generation layer is to use a metal imaging plate. The metal imaging plate is made of gold material, with a thickness of 5 microns, and its substrate is made of silicon dioxide, with a thickness of 500 microns.

[0049] In this embodiment, the neural network training process is modeled and optimized based on the Pytorch platform, and the target phase distribution is learned through the backpropagation algorithm. The MNIST (a classic machine learning dataset for handwritten digit recognition) and Fashion-MNIST (a fashion item image classification dataset) public datasets are used as inputs to construct image encodings of different categories. During the training process, the neuron size is set to 110 microns, 8 samples are input in batches, and the phase weights are independently trained for each polarization state. The joint training method is used in each independent training process to ensure that the functions between channels do not interfere with each other.

[0050] In addition, this embodiment also uses a full-wave electromagnetic simulation software to simulate and analyze the micron-scale silicon pillar unit structures with different sizes, obtain the transmission coefficient and phase response at a specific working frequency, guide the fine design of the structure size, and ensure excellent optical performance in actual manufacturing.

[0051] As Figure 1 shown, the multi-layer structure specifically implemented in this embodiment is as follows: including a spatial structured light generation layer (amplitude modulation structure), a multi-layer metasurface layer, and an output detection surface arranged in sequence along the incident light propagation direction. The processor realizes a multi-functional optical computing platform integrating logic operation and image recognition through amplitude modulation encoding, phase regulation, and output recognition of terahertz waves, where θ is the rotation angle between the second metasurface layer and the third metasurface layer. The multi-layer metasurface layer is arranged as the first, second, and third metasurface layers in sequence along the incident light propagation direction, △d is the distance that the metasurface layer moves. In y the case of polarized light incidence, D 2 The NN processor is used as a logic arithmetic unit. By adjusting the distances between the spatial structured light generation layer, the multi-layer metasurface layer and the output plane, the switching of logic gate calculations is realized. Among them, the distance between the spatial structured light generation layer and the output plane does not need to be adjusted, and only the positions of 3 metasurface layers are adjusted; in addition, the three diffraction layers are only examples of embodiments. Actually, they can all be multi-layer (at least 1 layer) diffraction layers, and the distances between them are not limited either; in x the case of polarized light incidence, D 2 The NN processor is used as a dual-object classifier. By adjusting the rotation angles of the second and third metasurface layers, the recognition of different categories of dual objects is realized. Among them, the spatial structured light generation layer and the output plane do not need to be rotated, and only the rotation angles of 3 metasurface layers are adjusted; in addition, the three diffraction layers are only examples of embodiments. Actually, they can all be multi-layer diffraction layers (at least 1 layer), and there is no limit to the single rotation of any layer or the combined rotation of multiple layers. However, the rotation angle can only be 180 degrees (because the x, y polarization channel corresponds to the dynamic phase); in short, as long as the network structure is changed, the corresponding functions can be realized.

[0052] As Figure 2 shown, it is the schematic diagram of the dynamic phase design in the embodiment of the present invention: Figure 2 (a) shows the polarization multiplexing mechanism of the transmission phase. By distinguishing x and y the linear polarization states, the complete separation of two functional channels is realized; through the diffraction neural network calculation, two groups of phase distributions are obtained as D 2 NN x and D 2 NN y . The incident light (including x and y linear polarized light) propagates and is regulated in the multi-layer metasurface structure, and a diffraction pattern related to the target is formed on the output surface. The model training process is as follows:

[0053] Use the Rayleigh-Sommerfeld diffraction integral formula to simulate the process of the light field propagating from the nth layer to the n+1th layer. Among them, represents the transmission coefficient of the nth layer, where A is the amplitude, φ is the phase, and i is the imaginary unit. When only training a pure-phase diffraction neural network, a = 1 is set; the light field distribution in the form of the Rayleigh-Sommerfeld diffraction theory propagation is expressed as:

[0054] (1)

[0055] where U(r n ) is the transmission electric field distribution of the nth layer. is the transmitted electric field distribution on the (n + 1)-th layer, where n is the number of diffraction layers. In the formula, h(r n+1 - r n ) represents the impulse response during the transfer process, and the expression is:

[0056] (2)

[0057] Among them, is the Euclidean distance from the point r n on the n-th layer to the point r n+1 on the (n + 1)-th layer. λ is the incident wavelength, i = is the imaginary unit, and x n+1 , y n+1 , z n+1 are the spatial coordinate components of a certain point on the (n + 1)-th layer respectively , , are the spatial coordinate components of the a-th neuron on the n-th layer respectively; S represents the integration region, which is all points within the entire wavefront area; this integration represents the weighted sum and superposition of the light field distribution of the entire n-th layer to calculate the light field distribution of a certain point on the (n + 1)-th layer; through this method, the propagation of light waves in a medium with different layer structures can be predicted and calculated; the number of diffraction layers is represented by n, and it is designed as a three-layer diffraction layer structure; first, by setting the propagation distances between different diffraction layers (corresponding to two calculation tasks under two distance conditions, which is joint training), different logical function outputs can be achieved. During a single training for y polarization, the three-layer phase distributions under different propagation distance conditions (logical AND gate and OR gate) are simultaneously optimized and adapted. The gradient is calculated using the backpropagation algorithm, and the phase is iteratively updated in combination with the gradient descent method until the loss function converges to obtain the optimal phase result: D 2 NN y = {Phase _ Y1, Phase _ Y2, Phase _ Y3}, where Phase _ Y1, Phase _ Y2, and Phase _ Y3 respectively represent the phase distributions of the three diffraction layers obtained during the "polarization - distance" training. This training process ensures that the network can output the corresponding logic gate operation results at the sensing layer (i.e., the detection plane) under different diffraction distances; by setting different rotation methods between different diffraction layers, different categories of dual-target recognition outputs can be achieved. Through the joint training strategy, for xOptimize the three-layer phase distribution for dual-object image recognition of two categories under different rotation conditions during a single polarization training (dual-object digit recognition and dual-object fashion item recognition are actually four image recognitions of different digits (1, 2) and different fashion items (tops, bottoms) after rotation under two different categories (digits and fashion items)). Calculate the gradient using the backpropagation algorithm, and combine the gradient descent method to iteratively update the phase until the loss function converges to obtain the optimal phase result: D 2 NN x ={Phase _ X1, Phase _ X2, Phase _ X3},Phase _ X1, Phase _ X2, and Phase _ X3 represent the phase distributions of the three diffraction layers obtained during the "polarization-rotation" training, respectively, and output the recognition results of dual digits and dual clothing on the perception layer. This training process ensures that the network can output corresponding dual-object recognition results of different categories under different rotation modes; Figure 2 (b) is a three-dimensional view of the micro-scale silicon pillar unit structure. The length and width of the micro-scale silicon pillar unit structure respectively affect x and y the transmission phase in the polarization state; Figure 2 (c) is a top view of the unit structure, showing the periodic arrangement of the micro-scale silicon pillar unit structure; Figure 2 (d) and Figure 2 (e) are x and y the designed transmission phase distribution diagrams under polarized incidence, respectively; Figure 2 (f) and Figure 2 (g) show the transmittance distributions of the micro-scale silicon pillar unit structure under different polarization states, used to verify its high efficiency and low loss characteristics.

[0058] As Figure 3 shown, it is the simulation result of realizing the switching of logic operation functions by regulating the interlayer distance under polarized light irradiation. Set the distance " y 1, d 4" from the spatial structured light generation layer to the first metasurface layer to be 5000 microns, and the distances " d 2, d 3" between the diffraction layers are both 10000 microns, and the output is the "AND gate" logic result, where d 1, d 2, d 3, d 3, d4 respectively represent the distance parameters between the spatial structured light generation layer, the metasurface layer, and the detection plane; adjust the distance between the spatial structured light generation layer and the first metasurface layer to 10,000 micrometers, the diffraction layer spacing to 5,000 micrometers, and the output is the logical result of "OR gate", where d 1’, d 2’, d 3’, d 4’ respectively represent the distance parameters between the spatial structured light generation layer, the metasurface layer, and the detection plane after changing the distance. Without changing the positions of the spatial structured light generation layer and the detection plane, d 1, d 2, d 3, d 4 to d 1’, d 2’, d 3’, d 4’ is that the greater the numerical change, the greater the difference in the wavefront reconstruction effect. As much as possible, reduce the crosstalk between multitasks achieved by changing the distance, which reflects the high programmability and functional flexibility of the system in the spatial dimension.

[0059] Such as Figure 4 shown, for x the simulation results of realizing the switching of dual-target recognition functions by rotating the second and third metasurface structures under the irradiation of polarized light. Input dual-digit patterns of different shapes, after three-layer diffraction processing, recognition light spots are formed on the output detection surface. Rotate the second and third metasurface layers 180 degrees respectively, and the dual-digit recognition can be switched to the dual-fashion item recognition task. Due to the multiplexing of polarization channels, the rotation angle can only be 180 degrees (but one layer can be rotated 180 degrees, or two layers can be rotated 180 degrees as different channels). After three-layer diffraction processing, recognition light spots corresponding to the dual-fashion items are formed on the output detection surface, which reflects the high responsiveness and reconstruction ability of the system to the rotation dimension.

[0060] Example 2:

[0061] The multi-dimensional multiplexing D 2 NN can provide multiple channels for encoding and processing information, and each channel plays a specific role in terms of secure key sharing. By collecting the information of these independent channels, a powerful information encryption platform can be established. Compared with the single-dimensional D 2 NN, the multi-dimensional method provides an additional dimension for information protection and significantly reduces the risk of leakage.

[0062] On this basis, a method for physical layer encrypted information transmission based on a multi-dimensional task path regulation mechanism is further proposed, and an optical secure transmission system integrating three-dimensional joint control is constructed. Specifically, the "polarization-distance-rotation" task is trained twice and Morse code is used as the output target of the training task. Under x-polarization, the distances D1 and D2 and whether to rotate are adjusted respectively to obtain D 2 NN X , to implement a terahertz optical computing task that outputs the corresponding Morse code of "META" when inputting the handwritten numbers "2468" (where D1 represents " d 1、 d 2、 d 3、 d 4" distance situation, D2 represents " d 1’、 d 2’、 d 3’、 d 4’" distance situation), under y-polarization, the distances D1 and D2 and whether to rotate are adjusted respectively to obtain D 2 NN Y , to implement a terahertz optical computing task that outputs the corresponding Morse code of "USST" when inputting the handwritten numbers "3579", D 2 The NN processor can use the optical parameter configuration (polarization state, interlayer distance or interlayer rotation angle) of the training process as the encryption transmission key according to the structural characteristics of the input light, so that the output of different task paths can be decoded only under specific configuration combinations, thereby realizing encryption control based on physical parameters and realizing an information security transmission system that hides and transmits multiple information without being cracked in the same system. By integrating multiple terahertz optical computing tasks on the same D 2 NN platform and using the key control function, an encrypted information transmission mechanism under the same source structure and different configurations is realized, which significantly improves the confidentiality and anti-cracking ability of the optical information processing system and has broad potential for application in scenarios with high security requirements.

[0063] Such as Figure 5As shown, the application process and simulation results of the multi-dimensional task path regulation mechanism based on the joint control of polarization-distance-rotation three degrees of freedom in the physical layer encryption information transmission system. The input information comes from the MNIST (a classic machine learning dataset for handwritten digit recognition) dataset and is modulated in the form of optical patterns. The cipherbook sets the mapping rules between the plaintext and the task path through Morse code, realizing the initial binding of information and path control. Polarization, distance, and rotation serve as keys, and the three jointly determine the selection of the task path. Under different key combinations, the diffraction patterns generated by the system change significantly, and the task path has strong selectivity. It is difficult to obtain the correct information in the output image without a matching key configuration, reflecting the enhanced effect of physical layer security. If the key is correct, the system can successfully reconstruct the input image on the detection plane, realizing the high-fidelity restoration of the original information, and the multi-dimensional encryption mechanism is effective in decoding in real scenarios.

[0064] The distance and rotation can be used alone or in combination; in Example 1, they are used alone, and in Example 2, they are used together.

[0065] In summary, the present invention constructs an all-optical D 2 NN processing platform that integrates three dimensions of polarization, distance, and rotation, and combines metasurface regulation and deep learning optimization to achieve an optical neural network computing architecture with high efficiency, programmability, and flexible task switching, which has important theoretical value and application prospects.

[0066] The above-described embodiments only represent two implementation manners of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. A D of polarization, distance, and rotation multi-dimensional multiplexing 2 NN processor, characterized in that, Including, arranged in sequence along the propagation direction of incident light: a spatial structured light generation layer for performing patterned amplitude modulation on the incident light to achieve spatial encoding of the incident light field; multiple metasurface layers, each layer including a micron-scale silicon pillar unit array arranged in a preset order for achieving multi-dimensional diffraction control; wherein, the geometric parameters of the micron-scale silicon pillar units are designed based on the propagation phase principle, and independent diffraction channels are constructed in the x and y polarization states respectively to achieve polarization multiplexing; the phase distribution of the multiple metasurface layers is obtained by training a diffraction neural network model, and the training process uses gradient descent and backpropagation algorithms, and a two-channel polarization decoupling training strategy is adopted to separately complete the joint training tasks of "x polarization - distance - rotation" and "y polarization - distance - rotation", and generate the optimal diffraction phase distribution D 2 NN X and D 2 NN Y , by independently or synergistically controlling the polarization state of light, adjusting the interlayer distance and rotation mode, multiple optical computing task processing under different channels of polarization, distance, and rotation multi-dimensional separation or multiplexing is achieved.

2. A D NN processor for polarization, distance, and rotation multi-dimensional multiplexing according to claim 1, characterized in that, 2 The micron-scale silicon pillar units are designed with independent phase responses in the x-polarization state and the y-polarization state respectively. By scanning different geometric parameters, the corresponding transmission coefficients and phase distributions are obtained. Using the propagation phase principle, the phase information is imparted to the length and width dimensions of the micron-scale silicon pillars, thereby realizing the function separation and multiplexing in different polarization states; the lengths and widths of the micron-scale silicon pillar units in the micron-scale silicon pillar unit array respectively achieve x the training phase distribution under y polarization and x the function switching of rotational multiplexing under y polarization and the function switching of distance multiplexing under ​ 3. A D 2 NN processor for polarization, distance, and rotation multiplexing, characterized in that By scanning the length and width parameters of the micron-scale silicon pillar unit structure under the conditions of x-polarized incident light and y-polarized incident light respectively, the transmission spectra and phase responses of the micron-scale silicon pillar unit structures with different sizes are obtained, and the metasurface micron-scale silicon pillar unit structures meeting the requirements are selected from them to achieve polarization-separated phase modulation.

4. A D NN processor for polarization, distance, and rotation multi-dimensional multiplexing according to claim 1, characterized in that 2 The spatial structured light generation layer adopts any one of the following forms: ​ Metal imaging plate: a transparent substrate made of 500-micron-thick silica material; a patterned metal film deposited on the upper surface of the transparent substrate, which is a 5-micron-thick gold film; wherein, the patterned metal film allows terahertz waves to transmit through the patterned area, and the remaining metal areas block terahertz waves, thereby realizing the spatial modulation of the amplitude of the incident light field and serving as the input signal encoding structure of the diffraction neural network; Programmable spatial light modulator SLM: used to dynamically load the input image pattern and modulate the amplitude of the terahertz wave incident light field.

5. A D NN processor for polarization, distance, and rotation multi-dimensional multiplexing according to claim 1, characterized in that, 2 The metasurface layer is prepared by deep silicon etching process, using silicon wafer material. The overall thickness of the silicon wafer is 1000 microns, the thickness of deep silicon etching is 500 microns, and the lateral length and width dimensions of each metasurface device are 13200 microns; the period of the metasurface micron-scale silicon pillar unit structure obtained after deep silicon etching is 110 microns, and its thickness is 500 microns; it operates at the set working frequency of 0.7 THz, and its array number is 120*120. ​ 6. A D 2 NN processor for polarization, distance, and rotation multi-dimensional multiplexing, characterized in that The distance-rotation combination method includes four schemes: none, distance, rotation, distance-rotation; by combining and adjusting the spatial structured light generation layer, switching the polarization state of the incident light, changing the spatial distance between each metasurface layer and the detection plane, or by rotating the metasurface layer, the optical field propagation path is changed to achieve optical computing tasks with different functions. The optical computing tasks include basic mathematical operations, different types of single / multi-target recognition, image analysis and processing, and signal encoding functions.

7. A D 2 NN processor for polarization, distance, and rotation multi-dimensional multiplexing, characterized in that The incident light is terahertz wave.

8. A method for implementing a DNN processor with polarization, distance, and rotation multi-dimensional multiplexing, characterized in that, 2 D for polarization, distance, and rotation multi-dimensional multiplexing as described in any one of claims 1-7 2 Implementation of the NN processor, converting the target input image into an amplitude modulation pattern of incident light, passing through multiple diffraction layers in sequence, and obtaining a logical calculation result or recognition output on the sensing layer; the specific training process is as follows: ​ The process of simulating the propagation of the optical field from the n-th layer to the (n + 1)-th layer using the Rayleigh-Sommerfeld diffraction integral formula; among them, represents the transmission coefficient of the n-th layer, where A is the amplitude, φ is the phase, and i is the imaginary unit. When only training a pure-phase diffraction neural network, a = 1 is set; the optical field distribution in the propagation form of the Rayleigh-Sommerfeld diffraction theory is expressed as: （1） where U(r n ) is the transmitted electric field distribution of the n-th layer, is the transmitted electric field distribution on the (n + 1)-th layer, n is the number of diffraction layers, and in the formula h(r n+1 -r n ) represents the impulse response during the transfer process, and the expression is: （2） wherein, is the Euclidean distance between the point r n on the n-th layer and the point r n +1 on the (n + 1)-th layer, λ is the incident wavelength, i = is the imaginary unit, x n+1 , y n+1 , z n+1 are respectively the spatial coordinate components of a certain point on the (n + 1)-th layer , , are respectively the spatial coordinate components of the a-th neuron on the n-th layer; S represents the integration region, which is all points within the entire wavefront area; this integration represents performing weighted sum and superposition on the light field distribution of the entire n-th layer to calculate the light field distribution of a certain point on the (n + 1)-th layer; through this method, the propagation of light waves in a medium with different layer structures can be predicted and calculated; the number of diffraction layers is represented by n; by setting the propagation distances and rotation modes between different diffraction layers, a corresponding number of calculation channels are formed, which can correspond to optical calculation tasks with different functions. During a single training for a single polarization, the multi-layer phase distributions under different propagation distance conditions and rotation modes are simultaneously optimized and adapted. The gradient is calculated using the backpropagation algorithm, and the phase is iteratively updated in combination with the gradient descent method until the loss function converges to obtain the optimal phase result. This training process ensures that the network can output the corresponding optical calculation task results at the sensing layer, i.e., the detection plane, under different diffraction distances and rotation modes.

9. A method for implementing a D 2 NN processor with polarization, distance, and rotation multi-dimensional multiplexing, characterized in that The diffraction neural network model is developed based on the Python language and is constructed and trained using the Pytorch deep learning framework; during the training process, the neuron size is set to 110 microns, 8 samples are input in batches, and the phase weights are independently trained for each polarization state. The joint training method is used in each independent training process to ensure that the functions between channels do not interfere with each other.

10. A method for implementing a DNN processor with polarization, distance, and rotation multi-dimensional multiplexing according to claim 8, characterized in that, 2 The transmission characteristics and phase responses of micron-scale silicon pillar units with different structural dimensions under the illumination of x- and y-linearly polarized light are simulated and analyzed by an electromagnetic full-wave simulation software, which is used to support the diffraction performance modeling and functional verification of the D 2 NN processor in multiple dimensions. ​

Images