Non-interleaved jones matrix eight-channel regulation method and regulation device thereof

Abstract

The application discloses a non-interleaved Jones matrix eight-channel regulation method and a regulation device thereof.The regulation method comprises the following steps: according to the prior experience in the communication field, converting the Jones matrix of an optimized super unit into the Jones matrix of an optimized entire super device to utilize the superposition and optimization of the polarized electromagnetic wave in the space propagation process; adopting two-layer unitary symmetric Jones matrix super surfaces to be cascaded to realize eight-channel regulation of the parameters in the Jones matrix; and the eight-channel regulation of the parameters in the Jones matrix is specifically as follows: a matrix diffraction deep neural network framework is used to pass through the space cascade of multiple layers of Jones matrix super surfaces, the acquired polarization information is brought into an all-optical network, and diversified inference analysis is performed. The application solves the problems of low efficiency, high complexity and insufficient expansibility in various frequency bands in the prior art, and converts the existing optimized super unit Jones matrix into the optimized entire super device Jones matrix.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a non-interleaved Jones matrix eight-channel control method and its control device. Background Technology

[0002] Artificial neural networks (ANNs) are composed of many artificial neurons connected like those in the human brain. They are a promising research topic, commonly used for inference tasks such as classification, regression, or prediction. Deep neural networks (DNNs), as a further extension of ANNs, construct multi-layered networks to perform hierarchical feature extraction and data representation, thereby enabling more complex pattern recognition and classification tasks. DNNs have been successfully applied to the reverse design of various optical manipulation meta-devices. All-optical diffractive deep neural networks (DNNs) are another example. 2 As an emerging all-hardware DNN, DNNs construct physical networks through spatially cascaded multilayer diffractive surfaces and achieve interlayer connections via electromagnetic (EM) wave propagation. Compared to electronic neural networks, DNNs... 2 The all-optical processing characteristics of neural networks (NNs) offer inherent advantages such as high computational speed, high parallelism, and low resource consumption. 2 The combination of neural network (NN) frameworks and metasurfaces has achieved great success in embedding functional meta-devices such as image classification, optical logic operations, optical linear transformations, and pulse shaping systems. These designs mostly employ multi-layered, spatially cascaded metasurfaces with a large number of neurons to obtain the dimension of the solution space for complex-valued transformations. However, due to limitations in the input and output fields of view, deeper diffraction networks and more neurons cannot further increase the dimension of the solution space. Therefore, under the same conditions, such as field of view, number of layers, and number of neurons, diffraction networks possess stronger inference performance, higher parallelism, and greater channel multiplexing capabilities, making them crucial in line with current trends in artificial intelligence development.

[0003] Existing improvements D 2 Methods for improving the inference accuracy of neural networks mainly include differential detection, optoelectronic network combination, nonlinear activation, and residual D. 2 NNs (Res-D) 2 NNs) and Fourier space D 2 NNs (FD) 2 (NNs). The generalization ability of these frameworks is mainly enhanced by increasing the number of layers or optical elements in the network, or by performing certain transformations on the output field of view. Furthermore, several spectral multiplexing D... 2NNs framework. The spectral multiplexing D2NNs framework utilizes the properties of materials at different frequency points to encode multiple tasks into multiple wavelength channels during joint training, enabling functions such as multi-task synchronous or time-division execution, spatial multiplexing filters, and multispectral imaging.

[0004] Polarization, as a key modulation dimension of electromagnetic waves, has also been introduced into all-optical networks. A dual-channel all-optical classifier has been reported. This framework is based on D... 2 NNs train polarization-insensitive neurons separately for two different tasks, thereby achieving polarization multitasking through polarization-multiplexed meta-atoms. Another approach to introducing polarization into all-optical networks is to insert additional untrained polarizer arrays between diffraction layers. By jointly training polarization-insensitive neurons with multitasking encoded by different combinations of input and output polarization states, a two-channel linear transformation is achieved when one polarizer array is inserted, and a four-channel linear transformation is achieved when two polarizer arrays are inserted. Achieving all-optical networks using available polarization multiplexing frameworks is not straightforward, as it relies on the polarization multiplexing capability of the superunit or additional polarization modulation components. It may involve complex designs of multi-channel polarization-multiplexed superunits or the unnecessarily introduction of more polarization modulation components. Therefore, to achieve all-optical networks with stronger polarization multiplexing capabilities and more compact configurations, it is necessary to directly introduce trainable polarization-modulated neurons into the D... 2 In NNs.

[0005] Currently, research on the limitations of polarization manipulation of Jones matrix metasurfaces mainly focuses on optimizing the metacells. This primarily requires staggered and multi-layered stacking designs, leading to low efficiency, high complexity, and insufficient scalability across various frequency bands.

[0006] By interleaving four birefringent metamaterials with specific position coordinates and rotation angles, the six parameters of the Jones matrix can be controlled to reach the upper limit of a single-layer metamaterial surface.

[0007] The full eight-parameter control of the Jones matrix is ​​achieved through a spatially cascaded bilayer metasurface, which consists of staggered metaunits that also have six modulation degrees of freedom in the Jones matrix.

[0008] However, staggered metasurfaces with a unit cell period of approximately one working wavelength operate in higher-order diffraction modes, leading to reduced spatial resolution, complex detection configurations, and low holographic efficiency.

[0009] In this work, the present invention proposes matrix diffractive deep neural networks (MDNs). 2Using the NN framework to design multilayer Jones matrix metasurfaces can effectively facilitate polarization manipulation in metadevices. To make this method easily scalable to different frequencies, we employ a unitary symmetric matrix platform (birefringent metasurface) to realize MD with polarization coding capabilities. 2 NNs. Based on the proposed MD 2 In NNs, we implemented a non-interleaved, high-efficiency Jones matrix eight-channel modulation function. Experimental results show that MD 2 The effectiveness and superiority of NNs (Neural Networks). Summary of the Invention

[0010] This invention provides a non-interleaved Jones matrix eight-channel control method and control device to solve the problems of low efficiency, high complexity and insufficient scalability in various frequency bands in the prior art, and transforms the existing Jones matrix of optimized meta-units into a Jones matrix of optimized entire meta-device.

[0011] This invention is achieved through the following technical solution: A non-interleaved Jones matrix eight-channel control method, the control method includes transforming the Jones matrix of the optimized meta-unit into the Jones matrix of the entire meta-device based on the prior experience in the field of communications, so as to utilize the superposition and optimization of polarized electromagnetic waves during space propagation. Two-layer unitary symmetric Jones matrix metasurfaces are cascaded to achieve eight-channel control of parameters in the Jones matrix. Specifically, the eight-channel control of parameters in the Jones matrix is ​​achieved by a matrix diffraction deep neural network framework that incorporates the acquired polarization information into an all-optical network through spatial cascading of multiple Jones matrix metasurfaces, thereby performing diverse inference analysis.

[0012] Furthermore, the matrix diffraction deep neural network MD 2 A neural network consists of an input layer, hidden layers, and an output layer. The matrix diffraction deep neural network MD 2 The input layer of the NN is a polarized plane electromagnetic wave, and the matrix diffraction deep neural network MD... 2 The hidden layers of the neural network are spatially cascaded Jones matrix hypersurfaces.

[0013] Furthermore, the matrix diffraction deep neural network MD 2 The multipolar-specific coding task of the input layer of the neural network is fed into the hidden layer for processing; The matrix diffraction deep neural network MD 2 The output layer of a neural network can be customized based on pre-executed tasks.

[0014] Furthermore, the matrix diffraction deep neural network MD 2The input layer of the NN consists of N×N Jones vectors, which are electromagnetic wave inputs of N×N pixels with preset polarization codes; The matrix diffraction deep neural network MD 2 The hidden layers of a neural network consist of subwavelength units.

[0015] Furthermore, due to the discretization design of the metasurface, the periodicity of the input and output layers is the same as that of a metaunit in the hidden layer, which facilitates computation; Located at discrete locations ( x d , y d , z d The meta-atoms of the m-th layer can be represented by a Jones matrix: (1) in, T and φ These represent the transmitted amplitude and phase of co-polarized or orthogonally polarized electromagnetic waves, respectively. Furthermore, the Jones matrix has eight trainable variables. For a two-layer cascaded hypersurface, the eight degrees of freedom in the Jones matrix can be controlled and manipulated; A unitary symmetric matrix platform can be represented as: (2) in, φ x and φ y They are respectively x polarization and y polarization phase, R(θ) For the angle is θ The rotation matrix. Furthermore, the unitary symmetric matrix platform has three trainable parameters, derived from the input layer or located at (…). x d , y d , z d The complex amplitude electric field of any polarized electromagnetic wave propagating between layers can be described by the Jones vector: (3) in, Ex and Ey It is along x and y Complex amplitude electric field in the direction of vibration, A and φ It refers to amplitude and phase.

[0016] Furthermore, using Rayleigh-Sommerfeld diffraction theory, by connecting the layers through electromagnetic waves propagating in free space and treating neurons as secondary wave sources, it can be modeled as follows: (4) Therefore, located at ( x, y, z The propagating electric field vector at point () E M Modulated by the Mth layer Jones matrix, it can be expressed as: (5) in, A This represents the entire discrete pixel on the (M-1)th layer. After the polarized electromagnetic wave is modulated in multiple diffraction layers, we use custom operations in the output layer to implement a meta-device with specific functions.

[0017] A non-interleaved Jones matrix eight-channel control device, wherein the Jones matrix eight-channel control device applies the non-interleaved Jones matrix eight-channel control method described above. For the Jones matrix eight-channel control device, the input layer is encoded by a specific polarization plane electromagnetic wave; The output layer is defined as the intensity distribution of electromagnetic waves received under a specific polarization; The mean squared error loss function is used as an error assessment to converge the intensity distribution of a specific polarization in the output layer to a preset value; Create MD 2 Forward propagation model of neural networks; MD 2 After the forward propagation model of the neural network is trained, the required functions are implemented using the gradient descent algorithm and the backpropagation algorithm. In MD 2 During the training of the forward propagation model of neural networks, different polarization channels for different tasks share the same Jones matrix parameters for each diffraction layer.

[0018] A multi-channel transmission method in communication, wherein the transmission method applies the non-interleaved Jones matrix eight-channel control device as described above, and adopts two layers of Jones matrix metasurface cascade to realize eight-channel control of parameters in the Jones matrix.

[0019] The beneficial effects of this invention are: This invention achieves non-interleaved, high-efficiency eight-channel Jones matrix modulation, demonstrating the effectiveness of a novel strategy that introduces polarization into electromagnetic field modulation of metasurfaces via Jones matrix metasurfaces. This could potentially drive the development of all-optical networks toward multi-task integration and more advanced functional devices.

[0020] The MD proposed in this invention 2With the assistance of NNs, non-interlaced birefringent meta-units with relatively small modulation degrees of freedom can be used to realize eight-channel modulation of the Jones matrix.

[0021] The matrix diffraction deep neural network MD of the present invention 2 NN deploys polarization-sensitive Jones matrix metasurfaces onto all-optical polarization multiplexing networks to perform complex inference tasks and reverse design of metadevices. Attached Figure Description

[0022] Figure 1 This is the MD of the present invention. 2 The schematic diagram and framework diagram of NN.

[0023] Figure 2 The present invention relates to a Jones matrix eight-channel modulated meta-device, wherein (a) four sets of printed and holographic images are integrated with a single meta-device by combining different input and output polarizations. z 1=0mm, z 2=40mm, z 3=80mm, (b) Simulation results of printing and holographic images correspond to the Jones matrix. J 11 , J 12 , J 21 and J 22 The experimental results for components, (c) printed and holographic images correspond to the Jones matrix. J 11 , J 12 , J 21 and J 22 Quantity. Detailed Implementation

[0024] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0025] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0026] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0027] The following is in conjunction with the appendix to this application specification. Figure 1-2 The technical solutions in the embodiments of this application are clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0028] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0029] Example 1 like Figure 1 As shown, a non-interleaved Jones matrix eight-channel control method is described. The control method includes transforming the Jones matrix of the optimized meta-unit into the Jones matrix of the entire meta-device based on the prior experience in the field of communications, so as to utilize the superposition and optimization of polarized electromagnetic waves during space propagation. Two-layer unitary symmetric Jones matrix metasurfaces are cascaded to achieve eight-channel control of parameters in the Jones matrix. Specifically, the eight-channel control of parameters in the Jones matrix is ​​achieved by a matrix diffraction deep neural network framework that incorporates the acquired polarization information into an all-optical network through spatial cascading of multiple Jones matrix metasurfaces, thereby performing diverse inference analysis.

[0030] A non-interleaved Jones matrix eight-channel control method, wherein the matrix diffraction deep neural network MD 2 NN (matrix diffractive deep neural networks, MD 2 A neural network (NN) consists of an input layer, hidden layers, and an output layer. Both the input layer and the hidden layer are represented using Jones vectors and Jones matrices.

[0031] The matrix diffraction deep neural network MD 2 The input layer of the NN is a polarized plane electromagnetic wave, and the matrix diffraction deep neural network MD... 2 The hidden layers of the neural network are spatially cascaded Jones matrix hypersurfaces.

[0032] A non-interleaved Jones matrix eight-channel control method, wherein the matrix diffraction deep neural network MD 2 The multipolar-specific coding task of the input layer of the neural network is fed into the hidden layer for processing; The matrix diffraction deep neural network MD 2 The output layer of a neural network can be customized based on pre-executed tasks.

[0033] A non-interleaved Jones matrix eight-channel control method, wherein the matrix diffraction deep neural network MD 2 The input layer of the NN consists of N×N Jones vectors, which are electromagnetic wave inputs of N×N pixels with preset polarization codes; The matrix diffraction deep neural network MD 2 The hidden layers of a neural network consist of various subwavelength units with polarization modulation capabilities.

[0034] A non-interleaved Jones matrix eight-channel control method, due to the discretization design of the metasurface, the periodicity of the input and output layers is the same as that of a metasurface unit in the hidden layer, which facilitates calculation; Located at discrete locations ( x d , y d , z d The meta-atoms of the m-th layer can be represented by a Jones matrix: (1) in, T and φ These represent the transmitted amplitude and phase of co-polarized or orthogonally polarized electromagnetic waves, respectively. A non-interleaved Jones matrix eight-channel control method is disclosed, wherein the Jones matrix has eight trainable variables. For a two-layer spatial cascaded hypersurface, the eight degrees of freedom in the Jones matrix can be controlled and manipulated. Therefore, the more variables expected in the Jones matrix, the more challenging it is to physically realize the superstructure. Thus, we chose a unitary symmetric matrix platform, which can be represented as: (2) in, φ x and φ y They are respectively x polarization and y polarization phase, R(θ) For the angle is θ The rotation matrix. A non-interleaved Jones matrix eight-channel control method, wherein the unitary symmetric matrix platform has three trainable parameters, physically indicating that the phase of any orthogonally polarized electromagnetic wave can be independently manipulated; for input from the input layer or located at ( x d , y d , z d The complex amplitude electric field of any polarized electromagnetic wave propagating between layers can be described by the Jones vector: (3) in, Ex and Ey It is along x and y Complex amplitude electric field in the direction of vibration, A and φ It refers to amplitude and phase.

[0035] A non-interleaved Jones matrix eight-channel modulation method utilizes Rayleigh-Sommerfeld diffraction theory to connect the layers through electromagnetic waves propagating in free space, and treats neurons as secondary wave sources, which can be modeled as follows: (4) Therefore, located at ( x, y, z The propagating electric field vector at point () E M Modulated by the Mth layer Jones matrix, it can be expressed as: (5) in, A This represents the entire discrete pixel on the (M-1)th layer. After the polarized electromagnetic wave is modulated in multiple diffraction layers, we use custom operations in the output layer to implement a meta-device with specific functions.

[0036] Specifically, the all-optical diffractive deep neural network (D2NN) framework, as a hardware platform, has proven capable of realizing various advanced functional metadevices, characterized by high parallelism and high processing speed. However, the design approach of incorporating trainable polarization-modulated neurons into D2NNs has not been fully explored, potentially leading to higher integration and greater task-load capabilities. Here, we propose matrix diffractive deep neural networks (M-D2NN), deploying polarization-sensitive Jones matrix metasurfaces onto an all-optical polarization multiplexing network to perform complex inference tasks and reverse design of metadevices. Through M-D2NN, we achieve non-interleaved, efficient eight-channel Jones matrix modulation, demonstrating the effectiveness of a novel strategy that introduces polarization into electromagnetic field modulation of metasurfaces via Jones matrix metasurfaces. This may drive the development of all-optical networks towards multi-task integration and more advanced functional devices.

[0037] Example 2 A non-interleaved Jones matrix eight-channel control device, wherein the Jones matrix eight-channel control device applies the non-interleaved Jones matrix eight-channel control method described above. For the Jones matrix eight-channel control device, the input layer is encoded by a specific polarization plane electromagnetic wave; The output layer is defined as the intensity distribution of electromagnetic waves received under a specific polarization; The mean squared error loss function is used as an error assessment to converge the intensity distribution of a specific polarization in the output layer to a preset value; Create MD 2 Forward propagation model of neural networks; MD 2 After the forward propagation model of the neural network is trained, the required functions are implemented using the gradient descent algorithm and the backpropagation algorithm. It is worth noting that in MD 2 During the training of the forward propagation model of neural networks, different polarization channels for different tasks share the same Jones matrix parameters for each diffraction layer.

[0038] Specifically, a non-interleaved Jones matrix eight-channel control method is used to realize the function of the device and control the multi-channel parameters in communication.

[0039] Example 3 A multi-channel transmission method in communication, wherein the transmission method applies the non-interleaved Jones matrix eight-channel control method as described in Embodiment 2, and adopts two layers of Jones matrix supersurface cascade to realize eight-channel control of parameters in the Jones matrix.

[0040] Specifically, a non-interleaved Jones matrix eight-channel control method is used to realize the function of the device and control the multi-channel parameters in communication.

[0041] Example 4 like Figure 2 As shown, to verify feasibility, four digital images with the strongest background intensity—"2", "0", "2", and "3"—were selected as printed images to characterize amplitude control. Additionally, the letter images "M", "E", "T", and "A" were used as holographic images to characterize phase control. The Jones matrix of the meta-device is encoded by the aforementioned four sets of printed and holographic images, as shown below. Figure 2 As shown in (a), and optimized using M-D2NNs, the distance between the two birefringent metasurfaces is 40 mm, and the imaging plane is 40 mm from the second metasurface. Each metasurface consists of 30 × 30 birefringent metaunits. Simulation results of the meta-device used to implement eight-channel Jones matrix operations are shown in (a). Figure 2 As shown in (b), the simulation results are consistent with the preset effects. Each Jones matrix channel exhibits flexible amplitude manipulation in the printed image and phase manipulation in the holographic image, with no obvious crosstalk between channels. The simulation efficiencies for the holographic images of “M”, “E”, “T”, and “A” are 25.6%, 24.3%, 20.1%, and 21.7%, respectively. This meta-device is manufactured using printed circuit board technology. The experimental setup was conducted in a microwave chamber. By adjusting the input and receive polarization states, four sets of printed and holographic images were captured, and the results are consistent with the simulation results, as shown in (b). Figure 2 As shown in (c), the experimental efficiencies for the holographic images of “M”, “E”, “T”, and “A” are 20.4%, 19.3%, 16.1%, and 17.3%, respectively. The slight discrepancy between the simulation and experimental results may stem from manufacturing errors and positioning accuracy during the measurement process. These experimental results demonstrate that the Jones matrix eight-channel modulated meta-device is successfully demonstrated based on M-D2NNs.

[0042] The proposed eight-channel modulated metadevice for the Jones matrix utilizes the most conventional metaunits, namely non-interlaced birefringent metaunits, which facilitates expansion to all frequency bands. The period of the metaunits used is approximately 0.35 times the operating wavelength, unlike previous works where the period was approximately equal to the operating wavelength. The advantage of a smaller period relative to the wavelength is that it allows for higher spatial resolution without sacrificing the metadevice's subwavelength characteristics and smaller relative aperture. Furthermore, the transmitted energy of this unit is primarily concentrated in the 0th-order channel, and higher-order diffraction is virtually nonexistent. It eliminates the need for an oblique-incident tilt detection configuration for image enhancement, resulting in a more compact and integrated metadevice. Moreover, it leads to more efficient holograms; our metadevice is approximately 40% more efficient than previously reported schemes. In summary, the non-interlaced birefringent platform for eight-channel modulated Jones matrix offers advantages for polarization-multichannel metadevices, including easy expansion to all frequency bands, more efficient hologram effects, and a more compact configuration.

Claims

1. A non-interleaved Jones matrix eight-channel control method, characterized in that, The control method includes, based on the prior experience in the field of communications, transforming the Jones matrix of the optimized meta-unit into the Jones matrix of the entire meta-device, in order to utilize the superposition and optimization of polarized electromagnetic waves during space propagation. Two-layer unitary symmetric Jones matrix metasurfaces are cascaded to achieve eight-channel control of parameters in the Jones matrix. Specifically, the eight-channel control of parameters in the Jones matrix is ​​achieved by a matrix diffraction deep neural network framework that incorporates the acquired polarization information into an all-optical network through spatial cascading of multiple Jones matrix metasurfaces, thereby performing diverse inference analysis. The matrix diffraction deep neural network MD 2 A neural network consists of an input layer, hidden layers, and an output layer. The matrix diffraction deep neural network MD 2 The input layer of the NN is a polarized plane electromagnetic wave, and the matrix diffraction deep neural network MD... 2 The hidden layers of the neural network are spatially cascaded Jones matrix hypersurfaces; The matrix diffraction deep neural network MD 2 The multipolar-specific coding task of the input layer of the neural network is fed into the hidden layer for processing; The matrix diffraction deep neural network MD 2 The output layer of a neural network is customized based on pre-executed tasks; The matrix diffraction deep neural network MD 2 The input layer of the NN consists of N×N Jones vectors, which are electromagnetic wave inputs of N×N pixels with preset polarization codes; The matrix diffraction deep neural network MD 2 The hidden layers of a neural network consist of subwavelength units; Due to the discretization design of the metasurface, the periodicity of the input and output layers is the same as that of a metaunit in the hidden layer, which facilitates computation. Located at discrete locations ( x d , y d , z d The meta-atoms of the m-th layer are represented by the Jones matrix: （1） in, T and φ These represent the transmitted amplitude and phase of co-polarized or orthogonally polarized electromagnetic waves, respectively. The Jones matrix has eight trainable variables; for a two-layer cascaded hypersurface, the eight degrees of freedom in the Jones matrix are controllable and manipulable. A unitary symmetric matrix platform is represented as: （2） in, φ x and φ y They are respectively x polarization and y polarization phase, R(θ) For the angle is θ The rotation matrix.

2. The non-interleaved Jones matrix eight-channel control method according to claim 1, characterized in that, The unitary symmetric matrix platform has three trainable parameters, which are derived from the input layer or located at (…). x d , y d , z d The complex amplitude electric field of any polarized electromagnetic wave propagating between layers is described by the Jones vector: （3） in, Ex and Ey It is along x and y Complex amplitude electric field in the direction of vibration, A and φ It refers to amplitude and phase.

3. The non-interleaved Jones matrix eight-channel control method according to claim 2, characterized in that, Using Rayleigh-Sommerfeld diffraction theory, the layers are connected by electromagnetic waves propagating in free space, and neurons are treated as secondary wave sources, modeled as follows: （4） Therefore, located at ( x,y,z The propagating electric field vector at point () E M Modulated by the Mth layer Jones matrix, it is expressed as: （5） in, A This represents the entire discrete pixel on the (M-1)th layer; polarized electromagnetic waves are modulated in multiple diffraction layers, and then custom operations are used in the output layer to realize a meta-device with specific functions.

4. A non-interleaved Jones matrix eight-channel control device, characterized in that, The Jones matrix eight-channel control device uses the non-interleaved Jones matrix eight-channel control method as described in any one of claims 1-3; For the Jones matrix eight-channel control device, the input layer is encoded by a specific polarization plane electromagnetic wave; The output layer is defined as the intensity distribution of electromagnetic waves received under a specific polarization; The mean squared error loss function is used as an error assessment to converge the intensity distribution of a specific polarization in the output layer to a preset value; Create MD 2 Forward propagation model of neural networks; MD 2 After the forward propagation model of the neural network is trained, the required functions are implemented using the gradient descent algorithm and the backpropagation algorithm. In MD 2 During the training of the forward propagation model of neural networks, different polarization channels for different tasks share the same Jones matrix parameters for each diffraction layer.

5. A multi-channel transmission method in communication, characterized in that, The transmission method uses the non-interlaced Jones matrix eight-channel control device as described in claim 4, employing a cascaded two-layer Jones matrix metasurface to achieve eight-channel control of parameters in the Jones matrix.

Images