Light beam demodulation detector based on multi-layer metasurface and construction method thereof

Abstract

The invention provides a light beam demodulation detector based on a multi-layer metasurface and a construction method of the light beam demodulation detector based on the multi-layer metasurface, and the light beam demodulation detector comprises an image processing chip and the multi-layer metasurface. The structured light beam to be identified generates a plurality of diffraction beams after passing through the diffraction neural network formed by the multi-layer metasurface; the diffracted light beams generate different light intensity distributions on the upper surface of the image processing chip, so that the image processing chip determines the type of the structured light beams to be recognized based on the light intensity distributions, and demodulation recognition of the light beams is achieved. According to the light beam demodulation detector based on the multi-layer metasurface and the construction method of the light beam demodulation detector, a demodulation component is constructed by utilizing diffraction neural network simulation and the multi-layer metasurface so as to realize demodulation transmission of structured light beams, and recognition of the types of the light beams is realized in combination with an image processing chip; and a compact light beam demodulation detection device with a miniaturized structure is optimally designed.

Description

Technical Field

[0001] The present invention relates to the technical field of optoelectronic detectors and neural network-optimized optical technologies, and particularly relates to a beam demodulation detector based on a multi-layer metasurface and a construction method thereof. Background Art

[0002] [[ID=⑧]]Modern communication systems mainly rely on electrical signals for information transmission. However, with the rapid development of the Internet and AI large models, problems faced by traditional electrical communications, such as limited spectrum resources, electromagnetic interference, and signal attenuation, have gradually emerged. Especially in high-frequency signal transmission, the above problems are more prominent, restricting the performance and efficiency of the system and making it difficult to meet the surging data demand. In contrast, optical communication technology uses light as an information carrier and has significant advantages such as wider bandwidth, low attenuation, strong anti-interference ability, and high security. It is expected to become a key technology to break through the development bottleneck of traditional communications, especially suitable for future high-throughput, high-data-rate, long-distance transmission, and anti-interference application scenarios.

[0003] Free Space Optical Communications (FSO) is an important optical communication technology that directly transmits optical signals through air or other transparent media. While having excellent optical communication characteristics, it effectively avoids the wiring limitations of existing fiber optic communication technologies that require laying expensive and complex fiber optic equipment, showing the advantages of flexible deployment and low cost. Currently, the research on FSO systems mainly focuses on solving two core problems: optical signal multiplexing and demodulation. In terms of multiplexing, it involves signal coding methods (such as mode division multiplexing, wavelength division multiplexing, etc.) and the optimization of multiplexing strategies to achieve higher data throughput and lower crosstalk. Vortex light with Orbital Angular Momentum (OAM) can support high-density, multi-channel, and low-crosstalk multiplexing due to its inherent mode orthogonality and sufficient number of modes, making the OAM-based mode division multiplexing technology highly concerned in the field. However, limiting the research to only OAM modes may be difficult to achieve higher-dimensional information coding and multiplexing, restricting the application boundaries of optical communications. Other types of structured light, such as Hermite-Gaussian beams, Mathieu beams, etc., also have excellent application potential due to their unique polarization states, complex amplitude distributions, and richer mode types.

[0004] In terms of demodulation, existing structured light demodulation devices such as Dammann gratings, multi-plane light converters, and Mach-Zehnder interferometers face limitations in grating diffraction efficiency, large volume, high system complexity, and the problem that losses increase with the number of modes, making it difficult to meet the industrial requirements for miniaturization and integration. A highly promising alternative is the mode demodulation device based on metasurfaces. The two-dimensional planar structure can significantly reduce the device volume, and by designing nanostructures, sub-wavelength scale manipulation of the optical field can be achieved, with extremely high mode selectivity and efficiency. Although they exhibit excellent vortex light mode recognition capabilities, most are integrated on optical fibers rather than modulating free-space light. Existing research has demonstrated the combination of metasurfaces and diffractive neural networks (D 2 NN, Diffractive Deep Neural Network). Based on deep learning design, it can achieve highly flexible regulation of optical signal modes, has end-to-end optimization capabilities, and can further enhance the accuracy of mode selection and recognition. However, due to the high requirements for the interlayer alignment accuracy of multi-layer diffractive devices in D 2 NN, the implementation of multi-layer designs still faces huge challenges in practical applications. To reduce the impact of alignment problems, current research mainly focuses on structures with a small number of layers (single or double layers) or long wavelength bands (such as terahertz, etc.), which to a certain extent limits the performance and application scope of demodulation devices.

[0005] Moreover, all of the above devices or systems face a common problem, that is, almost all only focus on the demultiplexing of the orbital angular momentum modes of vortex light, while ignoring the potential of other types of structured light. However, the types and modes of structured light are diverse, each with unique advantages, and can be used for different communication requirements and application scenarios. Existing research has restricted the diversity and flexibility of optical communication systems and also hindered their practical promotion in a wider range of fields. Therefore, it is crucial to improve the range of structured light recognition types of existing structured light demodulation devices, achieve efficient classification of multiple types of structured light, achieve high accuracy and low energy consumption, and overcome the deficiencies of existing devices in miniaturization and integration. Summary of the Invention

[0006] The present invention aims to provide a beam demodulation detector based on multi-layer metasurfaces and its construction method to solve the above technical problems, achieve the demodulation and transmission of structured light beams, and optimize and design a compact and miniaturized structured light beam demodulation detection device.

[0007] To solve the above technical problems, the present invention provides a beam demodulation detector based on multi-layer metasurfaces, including an image processing chip and multi-layer metasurfaces, wherein:

[0008] The upper surface of the image processing chip is fixedly connected to the bottom of the multi-layer metasurfaces;

[0009] The multi-layer metasurface includes at least four layers of metasurfaces, and different metasurfaces are stacked and fixedly connected; the metasurface is composed of N*N micro-nano units. During the processing of the multi-layer metasurface, based on D 2 NN and the backpropagation algorithm, the multi-layer phase distribution of the micro-nano units is obtained, so as to determine the preparation parameters of the metasurface based on the multi-layer phase distribution, and the metasurfaces are aligned layer by layer and stacked and fixed according to a preset sequence based on the preparation parameters to obtain a multi-layer metasurface for beam demodulation;

[0010] When the structure light beam to be recognized passes through the beam demodulation detector for demodulation and recognition, after the structure light beam to be recognized passes through the diffraction neural network composed of the multi-layer metasurface, several diffracted beams are generated; the several diffracted beams generate different light intensity distributions on the upper surface of the image processing chip, so that the image processing chip determines the type of the structure light beam to be recognized based on the light intensity distribution, thereby realizing the demodulation and recognition of the beam.

[0011] The above solution simulates and constructs a deep diffraction neural network by stacking and combining multi-layer metasurfaces. When the structure light beam to be recognized passes through the multi-layer vertically combined metasurfaces, the simulated and constructed deep diffraction neural network demodulates and transmits the structure light beam, realizing the transmission of the structure light beam to be recognized to the detection area at the corresponding position of the image processing chip. Furthermore, the image processing chip completes the recognition of the type of the structure light beam to be recognized through simple image recognition and analysis; at the same time, due to the miniaturization characteristics of the metasurface, and the multi-layer metasurfaces are stacked and fixedly connected and then integrated on the image processing chip, realizing high-density optical information processing, so that the beam demodulation detector composed of the multi-layer metasurface and the image processing chip has the characteristics of integration and compact structure; at the same time, adopting the beam demodulation detector structure of the present application, directly performing physical demodulation and beam splitting on the optical signal in free space, migrating the entire demodulation process to the optical part, without the need for an analog-to-digital conversion process, thereby effectively sharing the processing pressure of electronic components, greatly improving the processing speed, reducing energy consumption, overcoming the limitations of existing mode demultiplexing devices in terms of complexity, loss and integration, and providing a technical basis for the next-generation high-throughput free-space optical communication system.

[0012] Further, a transparent filling layer is provided between different metasurfaces, and the transparent filling layer is used to realize equidistant fixed connection between different metasurfaces.

[0013] Further, the transparent filling layer is also used to support different metasurfaces.

[0014] In the above solution, different metasurfaces are connected and fixed by a transparent filling layer to form a multi-layer metasurface, and the layer spacing of the multi-layer metasurface is determined by controlling the thickness of the transparent filling layer, ensuring the compact structure of the multi-layer metasurface.

[0015] Further, the beam demodulation detector further includes a transparent substrate, the upper surface of the transparent substrate is used to carry a multi-layer metasurface, and the lower surface of the transparent substrate is fixedly connected to the upper surface of the image processing chip.

[0016] The above solution realizes the integrated combination of the multi-layer metasurface and the image processing chip through the transparent substrate at the bottom of the multi-layer metasurface.

[0017] Further, the beam demodulation detector further includes alignment marks fixedly connected to the upper surface of the transparent substrate; during the process of stacking and fixing the metasurfaces to form the multi-layer metasurface, the micro-nano unit array of the metasurface is aligned and exposed layer by layer according to the alignment marks.

[0018] Further, during the process of stacking and fixing the metasurfaces to form the multi-layer metasurface, the micro-nano unit array of the metasurface is aligned and exposed layer by layer according to the orientation of the alignment marks.

[0019] In the above solution, by setting alignment marks on the transparent substrate, during the processing of the multi-layer metasurface, a high-performance electron beam lithography system (EBPG) can be used to expose the structures of the upper and lower layer metasurfaces based on this as a reference, realizing precise alignment and stacking, and ensuring the processing accuracy of the structural units of the multi-layer metasurface.

[0020] Further, the upper surface of the image processing chip is divided into several detection zones, and the several detection zones are provided with labels corresponding to the types of structured light beams; during the process of demodulating and identifying the to-be-identified structured light beam through the beam demodulation detector, after the to-be-identified structured light beam passes through the physical diffraction neural network composed of the multi-layer metasurface and generates several diffracted beams, the several diffracted beams are respectively focused within the corresponding detection zones, generating a specific light intensity distribution, so that the image processing chip determines the label of the corresponding detection zone based on the light intensity distribution, and further determines the type of the to-be-identified structured light beam, thereby realizing the demodulation and identification of the beam.

[0021] In the above solution, by dividing several detection zones on the upper surface of the image processing chip and corresponding the labels of the detection zones to the types of structured light, the identification process after beam demodulation is made clearer and simpler, without a complex image processing process.

[0022] Further, the micro-nano unit adopts a cylindrical structure.

[0023] In the above solution, by adopting a cylindrical micro-nano unit structure, due to the rotational symmetry of the micro-nano unit with a cylindrical structure, the multi-layer metasurface is insensitive to the polarization direction of incident light, making it suitable for processing unpolarized light or randomly polarized light; moreover, the cylindrical structure has fewer parameters, and continuous phase modulation can be achieved in a wide wavelength range by only adjusting the diameter and height of the cylinder, reducing the design difficulty and processing complexity.

[0024] A beam demodulation detector based on a multi-layer metasurface provided by the present invention simulates and constructs a deep diffraction neural network by stacking and combining multi-layer metasurfaces. When the structure light beam to be recognized passes through the multi-layer vertically combined metasurfaces, the simulated deep diffraction neural network demodulates and transmits the structure light beam, so as to transmit the structure light beam to be recognized to the corresponding detection area of the image processing chip. Then, the image processing chip completes the recognition of the type of the structure light beam to be recognized through simple image recognition and analysis; a transparent filling layer is used to connect multi-layer different metasurfaces into an integrated whole, significantly reducing the overall volume of the multi-layer metasurface component to the micron level, and the single-chip area is less than 0.01 mm 2 , and it can be integrated on the image processing chip to realize high-density optical information processing. Furthermore, the beam demodulation detector composed of the multi-layer metasurface and the image processing chip has the characteristics of miniaturization and compact structure; alignment marks are added to optimize the multi-layer alignment exposure process, greatly reducing the influence of the alignment problem in the actual device preparation process of the multi-layer diffraction neural network system; at the same time, by adopting the structure of the beam demodulation detector of the present application, the optical signal is directly physically demodulated and split in free space, and the entire demodulation process is migrated to the optical part without the need for an analog-to-digital conversion process, thus effectively sharing the processing pressure of electronic components, greatly improving the processing speed, reducing energy consumption, overcoming the limitations of the existing mode demultiplexing devices in terms of complexity, loss and integration, and having the ability of efficient splitting and accurate recognition of superimposed structured light, which helps to solve the high-dimensional information transmission requirements in multi-modal free space optical communication, provides more coding and decoding methods for optical communication, and provides a technical basis for the next-generation high-throughput free space optical communication system.

[0025] The present invention also provides a construction method for a beam demodulation detector based on a multi-layer metasurface, which is used to construct a beam demodulation detector based on a multi-layer metasurface as described above. The construction method includes:

[0026] Construct an initial diffraction neural network based on the Rayleigh-Sommerfeld diffraction theory and the metasurface modulation principle;

[0027] Define and obtain a single-class recognition loss function and a multi-class demodulation loss function based on Maxwell's equations and the artificial neural network principle;

[0028] Based on D 2The NN model, single-class recognition loss function, and multi-class demodulation loss function are used to optimize and train the initial diffraction neural network through step-by-step joint learning to obtain a multi-layer phase distribution;

[0029] Based on the multi-layer phase distribution, micro-nano processing is carried out to obtain a multi-layer metasurface, and then the multi-layer metasurface and the image processing chip are fixedly connected to construct a beam demodulation detector.

[0030] A construction method proposed in the above solution is as follows. First, based on the corresponding principles and objectives, a D 2 NN model is designed, and a demodulation and recognition loss function related to structured light is constructed. Then, an initial diffraction neural network is defined, and based on this initial diffraction neural network, the corresponding loss function, optimizer, and structured light dataset, etc., iterative optimization is carried out for the demodulation and recognition tasks to obtain the required neural network, so as to obtain the corresponding phase distribution parameters. Finally, the required multi-layer phase distribution is extracted according to the optimized neural network, and based on this, micro-nano processing is carried out to obtain a multi-layer metasurface simulating a physical neural network, and then an image processing chip is integrated to form a beam demodulation detector. The device constituted by this method can realize the demodulation transmission of structured light and the recognition of different types.

[0031] Furthermore, based on Maxwell's equations and the principle of artificial neural networks, a single-class recognition loss function and a multi-class demodulation loss function are defined, including:

[0032] Based on the structured light beam to be recognized and the Rayleigh-Sommerfeld diffraction theory, the complex amplitude of the output beam is obtained, and the calculation process satisfies the following formula:

[0033]

[0034] In the formula: U o (x r , y s ) represents the complex amplitude of the output beam, U i (x m , y n ) represents the complex amplitude of the structured light beam to be recognized, (x r , y s ) represents the observation point plane coordinates, (x m , y n ) represents the light source plane coordinates of the structured light beam to be recognized, and r represents the propagation distance between the observation point and the light source;

[0035] Based on the output beam and the metasurface modulation principle, the complex amplitude of the modulated beam is obtained, and the hidden layer in the artificial neural network is simulated to modulate the phase of the output beam, and the calculation process satisfies the following formula:

[0036]

[0037] Wherein, U m (x r , y s ) represents the complex amplitude of the modulated light beam, and U o (x r , y s ) represents the complex amplitude of the output light beam. represents the phase modulation amount;

[0038] Based on the complex amplitude of the modulated light beam, the target light intensity is obtained, and based on the target light intensity, the recognition result of the structured light beam to be recognized is obtained, so as to realize the construction of the initial diffraction neural network.

[0039] In the above solution, the complex amplitude of the output light beam is obtained through the structured light beam to be recognized and the Rayleigh-Sommerfeld diffraction theory, simulating the weight connection in the artificial neural network and undertaking the function of the forward propagation algorithm; the complex amplitude of the modulated light beam is obtained based on the output light beam and the metasurface modulation principle, simulating the hidden layer in the artificial neural network to modulate the phase of the output light beam, so as to finally realize the construction of the initial diffraction neural network.

[0040] Furthermore, based on the Maxwell's equations and the principle of artificial neural network, the single-class recognition loss function and the multi-class demodulation loss function are defined, including:

[0041] Based on the Maxwell's equations and the principle of artificial neural network, the single-class recognition loss function is defined, and the calculation process is shown as follows:

[0042]

[0043] Where: L single represents the loss function, L MSE,weighted represents the weighted mean square error between the output light beam positioning and the current detection partition, L residual represents the physical residual loss related to the Maxwell's equations, L eff represents the energy efficiency penalty term, y i represents the actual light intensity distribution, represents the target light intensity distribution, I in represents the incident light intensity of the structured light beam to be recognized, represents the Fourier transform of the output electric field, e target represents the target energy efficiency, w i represents the weight corresponding to the loss of the i-th class, K represents the number of structured light patterns to be recognized, α, β, and γ respectively represent the linear superposition weights of the three loss functions, and i represents the number of the current detection partition. represents the target light intensity within the current detection partition;

[0044] Define and obtain various types of demodulation loss functions based on Maxwell's equations and the principle of artificial neural networks. The calculation process is shown in the following formula:

[0045]

[0046] In the formula: L mixed represents various types of demodulation loss functions, L MSE represents the weighted mean square error between the output beam positioning and the current detection partition, L residual represents the physical residual loss related to Maxwell's equations, L eff represents the energy efficiency penalty term, y i represents the actual light intensity distribution, represents the target light intensity distribution, I in represents the incident light intensity of the structured light beam to be recognized, represents the Fourier transform of the output electric field, e target represents the target energy efficiency, K represents the number of structured light patterns to be recognized, α, β, and γ respectively represent the linear superposition weights of the three loss functions, and i represents the number of the current detection partition, represents the target light intensity within the current detection partition.

[0047] In the above solution, a single-type recognition loss function and various types of demodulation loss functions are defined and obtained based on Maxwell's equations and the principle of artificial neural networks, and the linear superposition weights of various loss functions are introduced to achieve the trade-off of multi-task priorities, providing accurate algorithm data support for the subsequent optimization process of the neural network.

[0048] Furthermore, based on the D 2 NN model, the single-type recognition loss function and various types of demodulation loss functions are used for step-by-step joint learning to optimize and train the initial diffraction neural network, obtaining a multi-layer phase distribution, including:

[0049] Iteratively optimize the initial diffraction neural network based on the preset single-structured light training set and the single-type recognition loss function, perform single-structured light optimization using the gradient descent method based on the initial parameters of the initial diffraction neural network, and obtain the value of the single-type recognition loss function in real time. When the value of the single-type recognition loss function meets the preset stability condition, end the iteration to obtain the single-structured light optimization model, and obtain the optimized phase distribution based on the single-structured light optimization model;

[0050] Based on the single-structured light optimization model, the preset aliased light training set, and various types of demodulation loss functions, perform iterative training. During the training process, calculate the gradient of various types of demodulation loss functions during the aliased light demodulation optimization process, perform aliased light demodulation optimization on the optimized phase distribution based on the optimizer according to the gradient, and update the convergence situation of various types of demodulation loss functions and the aliased light demodulation situation in real time;

[0051] Obtain the recognition accuracy value during the optimization process of aliased light demodulation based on the aliased light demodulation situation;

[0052] When the convergence situation of the function meets the preset convergence condition and the recognition accuracy value is stable, terminate the optimization process and obtain the multi-layer phase distribution based on the current aliased light optimization model.

[0053] In the above solution, to simultaneously achieve single-type recognition of structured light and demodulation and recognition of superimposed light, a step-by-step joint learning is adopted for different subtasks, and training is carried out separately according to their respective loss functions and task objectives, but the model shares phase parameters; at the same time, since the loss function of the single structured light type recognition task converges relatively fast, the optimization of subsequent tasks has little impact on its recognition accuracy, and the aliased structured light recognition task will fall into a local optimal solution if directly trained without proper parameter initialization. Therefore, during the entire training process, first iterate for the single structured light type recognition task, and end the optimization when its loss function is stable; based on the optimized model, further train for the aliased structured light recognition task and calculate the gradient of the loss function, and optimize the phase parameters of the metasurface according to the gradient through the optimizer. When the loss function converges and the accuracy is stable, end the optimization. The finally obtained multi-layer phase distribution provides a basis for the subsequent design of the metasurface structure.

[0054] Furthermore, based on the multi-layer phase distribution, micro-nano processing is carried out to prepare a multi-layer metasurface, and then the multi-layer metasurface and the image processing chip are fixedly connected to construct a beam demodulation detector, including:

[0055] Carry out simulation based on the multi-layer phase distribution to obtain the processing layout of the structure distribution of each layer of the metasurface;

[0056] Carry out layer-by-layer preparation based on the processing layout to obtain a multi-layer metasurface, realizing the physical construction of the diffraction neural network, and then fixedly connect the multi-layer metasurface and the image processing chip to construct a beam demodulation detector.

[0057] Furthermore, the process of carrying out layer-by-layer preparation based on the processing layout to obtain a multi-layer metasurface includes: carrying out layer-by-layer coating, alignment exposure, etching and filling treatment on a transparent substrate based on the processing layout to obtain a multi-layer metasurface.

[0058] A method for constructing a beam demodulation detector based on a multi-layer metasurface provided by the present invention, by designing D based on the corresponding principle and objective 2NN model, construct a demodulation and recognition loss function related to structured light, define an initial diffraction neural network, and perform iterative optimization for demodulation and recognition tasks based on the initial diffraction neural network, corresponding loss function, optimizer, and structured light dataset to obtain the required neural network. Finally, extract the required multi-layer phase distribution based on the optimized neural network, and perform micro-nano processing based on this to obtain a multi-layer metasurface that simulates a physical neural network, and then integrate an image processing chip to form a beam demodulation detector. The device formed by this method can achieve the demodulation and transmission of structured light and then realize the recognition of different types of structured light. At the same time, it proves the implementation process and feasibility of directly performing physical demodulation and beam splitting on optical signals in free space, transfers the entire demodulation process to the optical part, eliminates the need for analog-to-digital signal conversion, effectively shares the processing pressure of electronic components in the device constructed by this method, greatly improves the processing speed, and reduces energy consumption. Brief Description of the Drawings

[0059] Figure 1 Schematic diagram of a beam demodulation detector based on a multi-layer metasurface provided by an embodiment of the present invention;

[0060] Figure 2 Schematic diagram of a beam demodulation detector based on a four-layer metasurface provided by an embodiment of the present invention;

[0061] Figure 3 Schematic diagram of the beam demodulation process of a beam demodulation detector based on a four-layer metasurface provided by an embodiment of the present invention;

[0062] Figure 4 Schematic diagram of the construction method of a beam demodulation detector based on a four-layer metasurface provided by an embodiment of the present invention;

[0063] Figure 5 Schematic diagram of the cylindrical micro-nano unit structure in the construction method of a beam demodulation detector based on a four-layer metasurface provided by an embodiment of the present invention;

[0064] Figure 6 Schematic diagram of beam recognition in the construction method of a beam demodulation detector based on a four-layer metasurface provided by an embodiment of the present invention;

[0065] In the figure: 01, multi-layer metasurface; 02, image processing chip; 03, transparent filling layer; 04, transparent substrate; 05, alignment mark; 06, detection zone; 61, multi-target diffraction result; 62, single-target diffraction result; 07, single-type structured light recognition; 08, superimposed beam; 81, superposition of multiple single-type structured lights; 09, structured light beam to be recognized; 90, input structured light, 91, output light intensity distribution; 10, two-dimensional array of dielectric material nanocolumns; 11, four-layer metasurface. Detailed implementation manners

[0066] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0067] Embodiment 1:

[0068] This embodiment provides a beam demodulation detector based on a multi-layer metasurface, as Figure 1 shown, including a multi-layer metasurface 01 and an image processing chip 02, wherein:

[0069] The upper surface of the image processing chip 02 is fixedly connected to the bottom of the multi-layer metasurface 01;

[0070] The multi-layer metasurface 01 includes at least four layers of metasurfaces, and different metasurfaces are stacked and fixedly connected;

[0071] The metasurface is composed of N*N micro-nano units. During the processing of the multi-layer metasurface 01, the multi-layer phase distribution of the micro-nano units is obtained by the D 2 NN algorithm based on deep learning, the preparation parameters of the metasurface are determined based on the multi-layer phase distribution, and the metasurfaces are processed layer by layer and stacked and fixed in a preset sequence based on the preparation parameters to obtain the multi-layer metasurface 01 for beam demodulation;

[0072] When the structure light beam to be recognized passes through the beam demodulation detector for demodulation and recognition, the structure light beam to be recognized generates a number of diffracted beams after passing through the diffraction neural network composed of the multi-layer metasurface 01; different light intensity distributions are generated on the upper surface of the image processing chip 02 by the number of diffracted beams, so that the image processing chip 02 determines the type of the structure light beam to be recognized based on the light intensity distribution, thereby realizing the demodulation and recognition of the beam.

[0073] The above solution constructs a deep diffraction neural network by stacking and combining multiple layers of metasurfaces 01. When the structured light beam to be recognized passes through the vertically combined multiple layers of metasurfaces, the simulated deep diffraction neural network demodulates and transmits the structured light beam, realizing the transmission of the structured light beam to be recognized to the detection area at the corresponding position of the image processing chip 02. Then, the image processing chip 02 completes the recognition of the type of the structured light beam to be recognized through simple image recognition and analysis. At the same time, due to the miniaturization characteristics of the metasurface, and the multiple layers of metasurfaces 01 are stacked and fixedly connected and integrated on the image processing chip 02 to achieve high-density optical information processing, the beam demodulation detector composed of the multiple layers of metasurfaces 01 and the image processing chip 02 has the characteristics of integration and compact structure. At the same time, adopting the beam demodulation detector structure of the present application directly performs physical demodulation and beam splitting on the optical signal in free space, migrating the entire demodulation process to the optical part without the need for an analog-to-digital conversion process, thereby effectively sharing the processing pressure of electronic components, greatly improving the processing speed, reducing energy consumption, overcoming the limitations of existing mode demultiplexing devices in terms of complexity, loss, and integration, and providing a technical basis for the next-generation high-throughput free space optical communication system.

[0074] In the specific implementation process, the metasurface is constructed by a two-dimensional array of micro-nano units, i.e., dielectric material nano-columns with a quantity of N*N (N>240), based on the D 2 NN algorithm. The loss is calculated through the designed loss function, and the parameters of each "neuron", i.e., the phase value regulated by each micro-nano unit in the metasurface, are updated by backpropagation iterative optimization.

[0075] In the specific implementation process, preferably, the dielectric material used as the micro-nano unit is silicon nitride (Si3N4) or titanium dioxide (TiO2) with excellent optical properties in the visible light range; the image processing chip 02 preferably adopts a CMOS image processing chip, and the image processing chip 02 and the multiple layers of metasurfaces 01 can be integrated by pasting with a special optical adhesive or other methods.

[0076] Optionally, a transparent filling layer 03 is provided between different metasurfaces, and the transparent filling layer 03 is used to achieve the support and equidistant fixed connection between different metasurfaces.

[0077] In the specific implementation process, the transparent filling layer 03 can be selected as an SU-8 photoresist filling layer. Except for the topmost metasurface, each layer of the metasurface is coated and leveled with its micro-nano structure by spin-coating SU-8 photoresist. The connection between two layers of metasurfaces is realized through the SU-8 photoresist filling layer 3, and the layer spacing is determined by the spin-coating thickness of the SU-8 photoresist filling layer 3.

[0078] Optionally, the beam demodulation detector further includes a transparent substrate 04. The upper surface of the transparent substrate 04 is used to carry the multi-layer metasurface 01, and the lower surface of the transparent substrate 04 is fixedly connected to the upper surface of the image processing chip 02.

[0079] In the specific implementation process, the integration combination of the multi-layer metasurface 01 and the image processing chip 02 is realized through the transparent substrate 04 at the bottom of the multi-layer metasurface 01.

[0080] Optionally, the beam demodulation detector further includes an alignment mark 05, and the alignment mark 05 is fixedly connected to the upper surface of the transparent substrate 04; during the process of stacking and fixing the metasurfaces to form the multi-layer metasurface 01, the micro-nano unit array of the metasurface is aligned and exposed layer by layer according to the orientation of the alignment mark 05.

[0081] In the specific implementation process, the alignment mark 05 is used for alignment during the multi-layer lithography process. During this process, the alignment mark 05 is always exposed and not covered by the transparent filling layer 03. At the same time, to ensure the processing accuracy and interlayer alignment degree of the multi-layer metasurface structural units, in the device preparation, the structure of the metasurface is exposed by using the alignment exposure technology through a high-performance electron beam lithography system (EBPG).

[0082] Optionally, the upper surface of the image processing chip 02 is divided into several detection zones, and the several detection zones are provided with labels corresponding to the types of structured light beams; during the process of demodulating and identifying the to-be-identified structured light beam through the beam demodulation detector, after the to-be-identified structured light beam passes through the physical diffraction neural network formed by the multi-layer metasurface 01 and generates several diffracted beams, the several diffracted beams are respectively focused within the corresponding detection zones, generating a specific light intensity distribution, so that the image processing chip 02 determines the label of the corresponding detection zone based on the light intensity distribution, and further determines the type of the to-be-identified structured light beam, thereby realizing the demodulation and identification of the beam.

[0083] In the specific implementation process, by dividing several detection zones on the upper surface of the image processing chip 02 and corresponding the detection zone labels to the structured light types, the identification process after beam demodulation becomes clearer and simpler, without a complex image processing process.

[0084] Optionally, the micro-nano unit adopts a cylindrical structure.

[0085] In the specific implementation process, a transmissive metasurface is selected. The correspondence between the phase value and the micro-nano unit structure is preferably determined by the transmission phase theory. The preferred structure is cylindrical. Due to the rotational symmetry of the micro-nano units with a cylindrical structure, the multi-layer metasurface 01 is insensitive to the polarization direction of the incident light, making it suitable for processing unpolarized light or randomly polarized light. Moreover, the cylindrical structure has fewer parameters, and continuous phase modulation can be achieved in a wide wavelength range by only adjusting the diameter and height of the cylinder, reducing the design difficulty and processing complexity.

[0086] A beam demodulation detector based on a multi-layer metasurface provided in this embodiment broadens the range of available structured light types in FSO communication and realizes high-performance demodulation and classification of multiple types of structured light: Existing mode demodulation devices mainly focus on the demultiplexing of the orbital angular momentum mode of vortex light, ignoring the potential application value of other types of structured light (such as Mathieu beams, Airy beams, cosine-Gaussian beams, etc.). The purpose of this embodiment is to develop a device that can efficiently demodulate multiple types of structured light, including both the demodulation and recognition of the existing orbital angular momentum mode of vortex light, and at the same time can achieve the demodulation and recognition of different types of structured light, thereby expanding the applicable range of the optical communication system and providing more encoding and decoding methods for optical communication. At the same time, in the specific implementation process, the beam demodulation detector of this embodiment shows excellent demodulation and recognition capabilities. The recognition accuracy for a single type of beam exceeds 99.00%, the beam splitting and recognition accuracy in the case of the overlap of two structured beams exceeds 87.50%, and for the blind test (not encountered during the training process) of the overlapping beam of three OAM modes, the obtained accuracy is higher than 76%, realizing the accurate demodulation and recognition of different beam types. Secondly, it realizes all-optical demodulation processing, simplifies the digital-to-analog conversion process, and reduces energy consumption: Communication technologies based on electrical signals usually need to convert digital signals into analog signals for transmission, and at the same time, the receiving end needs to process the signal into digital information. The digital-to-analog conversion process increases the system complexity and energy consumption. The beam demodulation detector proposed in this embodiment has an energy consumption of 0.36W during the training stage. After the training is completed and the parameters are fixed, the energy consumption in the inference stage is only 0.15W. It directly performs physical demodulation and beam splitting on the optical signal in free space, that is, migrates the entire demodulation process to the optical part, without the need for analog-to-digital signal conversion, effectively sharing the processing pressure of electronic components, greatly improving the processing speed (close to the speed of light) and reducing energy consumption. Further, it has the advantages of miniaturization and easy integration. The single-chip area is less than 0.01mm 2 , significantly reducing the overall volume of the device to the micron level (96×96×400μm 3) It can be integrated onto an image processing chip to achieve high-density optical information processing, solving the problems of large size and complex system commonly found in current structured light demodulation devices. In addition, the multi-layer diffraction neural network system faces alignment problems during the actual device fabrication process and needs to balance the alignment degree and the number of physical network layers. Based on the requirements of modern communication for high demodulation capabilities of devices, a beam demodulation detection device based on a highly aligned multi-layer metasurface model with a compact structure for different types of structured light and orbital angular momentum mode recognition is finally designed.

[0087] Embodiment 2:

[0088] This embodiment provides a beam demodulation detector based on a four-layer metasurface, as Figure 2 shown, including: a two-dimensional array of dielectric material nanocolumns 10, an image processing chip 02, a transparent filling layer 03, a transparent substrate 04, and alignment marks 05, where:

[0089] A four-layer metasurface is formed by the four-layer two-dimensional array of dielectric material nanocolumns 10, the transparent filling layer 03, the transparent substrate 04, and the alignment marks 05: The four-layer metasurfaces are stacked and fixed through the transparent filling layer 03;

[0090] The lower surface of the bottom metasurface 13 is fixedly connected to the upper surface of the transparent substrate 04;

[0091] The lower surface of the transparent substrate 04 is fixedly connected to the upper surface of the image processing chip 02;

[0092] The bottom of the alignment marks 05 is disposed on the upper surface of the transparent substrate 04;

[0093] Specifically, the four-layer metasurface includes an upper metasurface, two middle metasurfaces, and a bottom metasurface: The upper metasurface is composed of an upper transparent filling layer and an upper two-dimensional array of dielectric material nanocolumns, and the bottom of the upper two-dimensional array of dielectric material nanocolumns is vertically fixed on the upper surface of the upper transparent filling layer; The middle metasurface is composed of a middle transparent filling layer and a middle two-dimensional array of dielectric material nanocolumns, and the bottom of the middle two-dimensional array of dielectric material nanocolumns is vertically fixed on the upper surface of the middle transparent filling layer and is covered and filled by the upper transparent filling layer; The bottom metasurface is composed of the transparent substrate 04 and a bottom two-dimensional array of dielectric material nanocolumns, and the bottom of the bottom two-dimensional array of dielectric material nanocolumns is vertically fixed on the upper surface of the transparent substrate 04 and is covered and filled by the middle transparent filling layer;

[0094] In the specific implementation process, the middle metasurface and the bottom metasurface are covered and filled with the micro-nano structure by spin-coating an SU-8 photoresist filling layer; During the processing of the four-layer metasurface, the alignment marks 05 are always exposed to facilitate the precise alignment of the four-layer metasurface in the vertical direction during exposure.

[0095] When the structured light beam to be recognized is demodulated and recognized by the beam demodulation detector, after the structured light beam to be recognized passes through the diffraction neural network composed of multiple metasurfaces, several diffracted beams are generated; the several diffracted beams generate different light intensity distributions on the upper surface of the image processing chip 02, so that the image processing chip 02 determines the type of the structured light beam to be recognized based on the light intensity distribution, thereby realizing the demodulation and recognition of the beam;

[0096] Among them, several detection partitions 06 are defined on the upper surface of the image processing chip 02, and the several detection partitions 06 are provided with labels corresponding to the types of structured light; in the process of demodulating and recognizing the structured light beam to be recognized by the beam demodulation detector, after the structured light beam to be recognized passes through the physical diffraction neural network composed of four layers of metasurfaces 7 and several diffracted beams are generated, the several diffracted beams are respectively focused within the corresponding detection partitions 06 to generate a specific light intensity distribution, so that the image processing chip 02 determines the label of the corresponding detection partition 06 based on the light intensity distribution, and further determines the type of the structured light beam to be recognized.

[0097] In the specific implementation process, as Figure 3 shown, in the process of demodulating and recognizing a single type of structured light 07 (such as Bessel-Gaussian beam, Laguerre-Gaussian beam, etc.) or a superimposed beam 08 by using the above-mentioned beam demodulator, where 81 represents the process of superimposing multiple single types of structured light, the incident structured light beam 09 to be recognized passes through the physical diffraction neural network composed of four layers of metasurfaces 11 for diffraction propagation, so that a certain type of structured light is transmitted into the detection partition 06 corresponding to its label, as Figure 3 shown by the multi-target diffraction result 61 and the single-target diffraction result 62 in.

[0098] Embodiment 3:

[0099] This embodiment provides a construction method of a beam demodulation detector based on a multi-layer metasurface for constructing the above-mentioned beam demodulation detector based on a multi-layer metasurface, as Figure 4 shown, the construction method includes the steps:

[0100] S1: Construct an initial diffraction neural network model based on the Rayleigh-Sommerfeld diffraction theory and the metasurface modulation principle;

[0101] S2: Define and obtain a single-type recognition loss function and a multi-type demodulation loss function based on Maxwell's equations and the principle of artificial neural networks;

[0102] S3: Optimize and train the initial diffraction neural network by using a distributed joint learning based on the D 2 NN model, the single-type recognition loss function and the multi-type demodulation loss function to obtain a multi-layer phase distribution;

[0103] S4: Based on the multi-layer phase distribution, perform micro-nano processing to obtain a multi-layer metasurface, and then fixedly connect the multi-layer metasurface and the image processing chip to construct a beam demodulation detector.

[0104] A construction method proposed by the above solution first designs a D 2 NN model based on the corresponding principle and target, constructs a demodulation and recognition loss function related to structured light, then defines an initial diffraction neural network, and performs iterative optimization for the demodulation and recognition tasks based on the initial diffraction neural network, the corresponding loss function, the optimizer, and the structured light data set, etc. to obtain the required neural network, so as to obtain the corresponding phase distribution parameters. Finally, extract the required multi-layer phase distribution according to the optimized neural network, and perform micro-nano processing based on this to obtain a multi-layer metasurface simulating a physical neural network, and then integrate an image processing chip to form a beam demodulation detector. The device formed by this method can realize the demodulation transmission of structured light and the recognition of different types.

[0105] Optionally, step S1 includes:

[0106] Obtain the complex amplitude of the output beam based on the structured light beam to be recognized and the Rayleigh-Sommerfeld diffraction theory, simulate the weight connection in the artificial neural network, undertake the function of the forward propagation algorithm, and the calculation process satisfies the following formula:

[0107]

[0108] In the formula: U o (x r , y s ) represents the complex amplitude of the output beam, U i (x m , y n ) represents the complex amplitude of the structured light beam to be recognized, (x r , y s ) represents the coordinate of the observation point plane, (x m , y n ) represents the coordinate of the light source plane of the structured light beam to be recognized, and r represents the propagation distance between the observation point and the light source;

[0109] Obtain the complex amplitude of the modulated beam based on the output beam and the metasurface modulation principle, simulate the hidden layer in the artificial neural network to modulate the phase of the output beam, and the calculation process satisfies the following formula:

[0110]

[0111] In the formula, U m (x r , y s ) represents the complex amplitude of the modulated beam, Uo (x r ,y s ) represents the complex amplitude of the output light beam, Indicates the phase modulation amount;

[0112] The target light intensity is obtained based on the complex amplitude of the modulated light beam, and the recognition result of the structured light beam to be recognized is obtained based on the target light intensity, thereby realizing the construction of the initial diffraction neural network.

[0113] In the specific implementation process, the six types of structured light (including Laguerre-Gaussian, Bessel-Gaussian, Hermite-Gaussian, Mattheus, cosine-Gaussian and Airy beams) with a wavelength of 532nm and five modes of vortex beams carrying OAM are demodulated and identified. As an example, this embodiment is further described in detail: the forward propagation process of structured light of different modes in the diffraction neural network is based on the Rayleigh-Sommerfeld diffraction theory, assuming that the complex amplitudes of the input light field and the output light field are U i and U 0 , it can be expressed as:

[0114]

[0115] in, is the distance between the light source and the observation point;

[0116] Taking into account the modulation of the light field by the metasurface, the expression of the output light field is modified as follows:

[0117]

[0118] After the structured light beam to be identified has passed through the forward propagation process, it will generate a light intensity distribution on the acquisition plane of the CMOS image processing chip. The light intensity of different detection partitions 06 can be expressed as i is the number of different detection areas, marked as numbers "0" to "9" respectively, corresponding to the label value of the structured light type. The category corresponding to the number of the area with the largest light intensity value is selected as the output result.

[0119] Step S2 includes: defining and obtaining a single-category recognition loss function based on Maxwell's equations and the principle of artificial neural networks. The calculation process is shown in the following formula:

[0120]

[0121] Where: L single represents the loss function, L MSE,weighted represents the weighted mean square error between the output beam positioning and the current detection partition, L residual represents the physical residual loss associated with Maxwell's equations, L eff represents the energy efficiency penalty term, yi represents the actual light intensity distribution represents the target light intensity distribution, I in represents the incident light intensity of the structured light beam to be recognized represents the Fourier transform of the output electric field, e target represents the target energy efficiency, w i represents the weight corresponding to the loss of the i-th category, K represents the number of structured light patterns to be recognized, α, β, and γ respectively represent the linear superposition weights of the three loss functions, and i represents the number of the current detection partition represents the target light intensity within the current detection partition

[0122] Based on the Maxwell's equations and the principle of artificial neural network, a multi-category demodulation loss function is defined, and the calculation process is shown as follows

[0123]

[0124] In the formula, where: L mixed represents the multi-category demodulation loss function, L MSE represents the weighted mean square error between the output beam positioning and the current detection partition, L residual represents the physical residual loss related to the Maxwell's equations, L eff represents the energy efficiency penalty term, y i represents the actual light intensity distribution represents the target light intensity distribution, I in represents the incident light intensity of the structured light beam to be recognized represents the Fourier transform of the electric field, e target represents the target energy efficiency, K represents the number of structured light patterns to be recognized, α, β, and γ respectively represent the linear superposition weights of the three loss functions, and i represents the number of the current detection partition represents the target light intensity within the current detection partition

[0125] In the specific implementation process, to simultaneously achieve the single-category recognition of structured light and the demodulation and recognition of superimposed light, a step-by-step joint learning is adopted for different subtasks, and they are trained respectively according to their own loss functions and task objectives, but the model shares the phase parameters. Define the loss function L of the single-category recognition target of structured light single as the weighted mean square error L between the positioning of the output beam of the last layer of metasurface and the target detection area MSE,weighted , the physical residual loss L related to the Maxwell's equations residual and the linear combination of the energy efficiency penalty term L eff :

[0126]

[0127] Meanwhile, according to the above definition of the loss function, the loss function for the demodulation and recognition tasks of the superimposed beam is the mean square error \(L\). MSE and the physical residual loss \(L\) related to the Maxwell's equations residual and the linear sum of the energy efficiency penalty term \(L\). eff That is:

[0128]

[0129] Optionally, step S3 includes:

[0130] Iteratively optimizing the initial diffraction neural network based on a preset single structured light training set and a single-class recognition loss function, performing single structured light optimization using the gradient descent method based on the initial parameters of the initial diffraction neural network, and obtaining the single-class recognition loss function value in real time. When the single-class recognition loss function value meets the preset stability condition, end the iteration to obtain a single structured light optimization model, and obtain an optimized phase distribution based on the single structured light optimization model;

[0131] Performing iterative training based on the single structured light optimization model, a preset aliased light training set, and a multi-class demodulation loss function. During the training process, calculate the gradient of the multi-class demodulation loss function during the aliased light demodulation optimization process, perform aliased light demodulation optimization on the optimized phase distribution based on the optimizer according to the gradient, and update the convergence situation of the multi-class demodulation loss function and the aliased light demodulation situation in real time;

[0132] Obtain the recognition accuracy value during the aliased light demodulation optimization process based on the aliased light demodulation situation;

[0133] When the function convergence situation meets the preset convergence condition and the recognition accuracy value is stable, terminate the optimization process and obtain the multi-layer phase distribution based on the current aliased light optimization model.

[0134] In the specific implementation process, since the loss function of the single structured light type recognition task converges relatively quickly, the optimization of subsequent tasks has little impact on its recognition accuracy, and the aliased structured light recognition task will fall into a local optimal solution if directly trained without proper parameter initialization. Therefore, during the entire training process, first perform iteration for the single structured light type recognition task and end the optimization when its loss function is stable; further train the aliased structured light recognition task based on the optimized model and calculate the gradient of the loss function, and use the Adam optimizer to optimize the phase parameters of the metasurface according to the gradient. End the optimization when the loss function converges and the accuracy is stable. The finally obtained phase distribution provides a basis for the design of subsequent metasurface structures.

[0135] Optionally, step S4 includes: performing simulation based on the multi-layer phase distribution to obtain the processing layout of the structural distribution of each layer of the metasurface; performing layer-by-layer coating, alignment exposure, etching, and filling on the transparent substrate based on the processing layout to obtain the multi-layer metasurface 01, realizing the physical construction of the diffraction neural network, and then fixedly connecting the multi-layer metasurface 01 and the image processing chip 02 to construct the beam demodulation detector.

[0136] In the specific implementation process, a transmissive scheme is adopted to design the metasurface, and the computer software COMSOL Multiphysics is used to simulate the metasurface based on the transmission phase. Table 1 below shows the processing parameters of the metasurface based on Si3N4 cylindrical micro-nano units of different layers to achieve a phase coverage of 0-2π at a light wavelength of 532 nm. Among them, the bottom-layer metasurface selects K9 glass as the transparent substrate, the middle-layer metasurface and the upper-layer metasurface select SU-8 photoresist as the substrate, and the filling layer selects SU-8 photoresist. There is no filling layer above the upper-layer metasurface. Based on the above relationship between the diameter of the micro-nano unit and the phase change, the nano-columns are arranged with reference to the simulated phase distribution map to obtain the GDS layout for metasurface processing. As Figure 5 shown, at a wavelength of 532 nm, the phase change of the Si3N4 cylindrical micro-nano unit can cover 0-2π when the period p is 400 nm and the diameter d is in the range of 60 nm - 370 nm.

[0137] Table 1 Processing parameters of the metasurface of different layers

[0138] Item Substrate Filling layer Period p / nm Diameter d / nm Height h / nm Bottom layer K9 glass SU-8 photoresist 400 60-370 1600 Intermediate layer SU-8 photoresist SU-8 photoresist 400 60-370 1600 Upper layer SU-8 photoresist Air 400 60-370 650

[0139] In the specific implementation process, as Figure 6 shown, Figure 6 (a) represents the recognition of single-type structured light, Figure 6(b) represents the demodulation and recognition of aliased light. Among them, 90 represents the input structured light, and 91 represents the output light intensity distribution. In the simulation of single-type recognition of structured light, different types of structured light are respectively marked as "0" and "6"-"9", and different OAM light modes are respectively marked as "1"-"5". Ten detection areas with equal areas are set on the detection plane of the image processing chip, corresponding to the labels of "0"-"9" respectively. When the Mathieu beam (labeled as "7") passes through the multi-layer metasurface neural network, the modulated diffracted light will be focused on the detection area corresponding to the label "7", realizing the recognition task of a single structured light. In the actual task, after 200 rounds of iterative training, the simulation accuracy reaches 99.03%; in the demodulation and recognition task of superimposed light, the optical fields of any two modes of structured light are superimposed, and the generated superimposed light records the labels of the two initial optical fields at the same time. When the new aliased light formed by superimposing the Bessel-Gaussian beam (labeled as "1") and the Airy beam (labeled as "9") passes through the modulation of the four-layer metasurface, the diffracted light will be divided into two beams and respectively focused on the detection areas labeled as "1" and "9". The accuracy of the model obtained through the above-mentioned 200 rounds of iterative training reaches 87.86%, effectively improving the accuracy of beam demodulation and recognition.

[0140] Table 2 Corresponding relationship between different types of structured light, OAM modes and labels

[0141]

[0142] In the specific implementation process, as shown in Table 3, the number of layers of the metasurface is an important factor affecting the performance. To a certain extent, increasing the number of metasurfaces can be regarded as increasing the depth of the diffractive neural network, which improves the accuracy of both single-type structured light and aliased structured light, reaching 99.65% and 91.27% respectively when the total number of layers is 8.

[0143] Table 3 Influence of the number of layers of multi-layer metasurfaces on the performance of the beam demodulator

[0144] Number of layers of multi-layer metasurface Accuracy of single-structure light recognition Accuracy of aliased structure light recognition 4 99.03％ 87.86％ 6 99.65％ 90.45％ 8 99.65％ 91.27％ 10 99.65％ 89.80％

[0145] A construction method of a beam demodulation detector based on a multi-layer metasurface provided in this embodiment designs D based on the corresponding principle and target 2NN model, construct a demodulation and recognition loss function related to structured light, and define an initial diffraction neural network. Based on this initial diffraction neural network, the corresponding loss function, optimizer, and structured light dataset, etc., perform iterative optimization for the demodulation and recognition tasks to obtain the required neural network, so as to obtain the corresponding phase distribution parameters. Finally, extract the required multi-layer phase distribution according to the optimized neural network, and based on this, perform micro-nano processing to obtain a multi-layer metasurface that simulates a physical neural network, and then integrate an image processing chip to form a beam demodulation detector. The device constituted by this method can realize the demodulation and transmission of structured light and then realize the recognition of different types of structured light; at the same time, it proves the implementation process and feasibility of directly performing physical demodulation and beam splitting on optical signals in free space, migrates the entire demodulation process to the optical part, without the need for analog-to-digital signal conversion, effectively shares the processing pressure of the electronic components in the device constructed by this method, greatly improves the processing speed (close to the speed of light) and reduces energy consumption; provides a construction method of a beam demodulator that can simultaneously identify and classify different types of structured light (such as Laguerre-Gaussian beams, Bessel-Gaussian beams, cosine-Gaussian beams, Mathieu beams, and Airy beams, etc.) and the orbital angular momentum (OAM) mode of vortex light.

[0146] The above is the preferred implementation manner of the present invention. It should be noted that for those of ordinary skill in the art of the present technology, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A beam demodulation detector based on a multi-layer metasurface, characterized in that, It includes an image processing chip and a multi-layer metasurface, where: The upper surface of the image processing chip is fixedly connected to the bottom of the multi-layer metasurface; The multi-layer metasurface includes at least four layers of metasurfaces, and different metasurfaces are stacked and fixedly connected; The metasurface is composed of micro-nano units arranged in an N*N array. During the processing of the multi-layer metasurface, based on D 2 NN algorithm, the multi-layer phase distribution of the micro-nano units is obtained, and the preparation parameters of the metasurface are determined based on the multi-layer phase distribution. Based on the preparation parameters, the metasurface is processed layer by layer and stacked and fixed in a preset sequence to obtain a multi-layer metasurface for beam demodulation; When the structure light beam to be recognized is demodulated and recognized by the beam demodulation detector, after the structure light beam to be recognized passes through the diffraction neural network formed by the multi-layer metasurface, several diffraction beams are generated; the several diffraction beams generate different light intensity distributions on the upper surface of the image processing chip, so that the image processing chip determines the type of the structure light beam to be recognized based on the light intensity distribution, thereby realizing the demodulation and recognition of the beam.

2. The beam demodulation detector based on a multi-layer metasurface according to claim 1, wherein A transparent filling layer is provided between different metasurfaces, and the transparent filling layer is used to realize the equidistant fixed connection between different metasurfaces.

3. The beam demodulation detector based on a multi-layer metasurface according to claim 1, characterized in that It further includes a transparent substrate, the upper surface of the transparent substrate is used to carry the multi-layer metasurface, and the lower surface of the transparent substrate is fixedly connected to the upper surface of the image processing chip.

4. The beam demodulation detector based on a multi-layer metasurface according to claim 3, characterized in that, It further includes an alignment mark, and the alignment mark is fixedly connected to the upper surface of the transparent substrate; during the process of stacking and fixing the metasurfaces to form the multi-layer metasurface, the micro-nano unit array of the metasurface is aligned and exposed layer by layer according to the alignment mark.

5. The beam demodulation detector based on a multi-layer metasurface according to claim 1, characterized in that The upper surface of the image processing chip is divided into several detection zones, and several of the detection zones are provided with labels corresponding to the types of structure light beams; During the process of demodulating and recognizing the structure light beam to be recognized by the beam demodulation detector, after the structure light beam to be recognized passes through the physical diffraction neural network formed by the multi-layer metasurface and several diffraction beams are generated, the several diffraction beams are respectively focused within the corresponding detection zones to generate a specific light intensity distribution, so that the image processing chip determines the label of the corresponding detection zone based on the light intensity distribution, and further determines the type of the structure light beam to be recognized, thereby realizing the demodulation and recognition of the beam.

6. The beam demodulation detector based on a multi-layer metasurface according to claim 1, wherein The micro-nano unit adopts a cylindrical structure.

7. A construction method of a beam demodulation detector based on a multi-layer metasurface, characterized in that, A method for constructing a beam demodulation detector based on a multi-layer metasurface according to any one of claims 1 to 6, the construction method includes: Constructing an initial diffraction neural network based on the Rayleigh-Sommerfeld diffraction theory and the metasurface modulation principle; Defining and obtaining a single-type recognition loss function and a multi-type demodulation loss function based on Maxwell's equations and the artificial neural network principle; Based on D 2 The NN model, single-class recognition loss function, and multi-class demodulation loss function are used to optimize and train the initial diffraction neural network through step-by-step joint learning to obtain a multi-layer phase distribution; Performing micro-nano processing and preparation based on the multi-layer phase distribution to obtain a multi-layer metasurface, and then fixedly connecting the multi-layer metasurface and the image processing chip to construct a beam demodulation detector.

8. The construction method of a beam demodulation detector based on a multi-layer metasurface according to claim 7, characterized in that, The constructing of the initial diffraction neural network based on the Rayleigh-Sommerfeld diffraction theory and the metasurface modulation principle includes: Obtaining the complex amplitude of the output beam based on the structure light beam to be recognized and the Rayleigh-Sommerfeld diffraction theory, and the calculation process satisfies the following formula: Where: U o (x r , y s ) represents the complex amplitude of the output light beam, U i (x m , y n ) represents the complex amplitude of the structured light beam to be recognized, (x r , y s ) represents the coordinates of the observation point plane, (x m , y n ) represents the coordinates of the light source plane of the structured light beam to be recognized, and r represents the propagation distance between the observation point and the light source; Obtaining the complex amplitude of the modulated beam based on the output beam and the metasurface modulation principle, and simulating the hidden layer in the artificial neural network to modulate the phase of the output beam, and the calculation process satisfies the following formula: where U m (x r , y s ) represents the complex amplitude of the modulated light beam, and U o (x r , y s ) represents the complex amplitude of the output light beam, represents the phase modulation amount; Obtain the target light intensity based on the complex amplitude of the modulated light beam, and obtain the recognition result of the structured light beam to be recognized based on the target light intensity, so as to realize the construction of the initial diffraction neural network.

9. The construction method of a beam demodulation detector based on a multi-layer metasurface according to claim 7, characterized in that, Define and obtain the single-class recognition loss function and the multi-class demodulation loss function based on Maxwell's equations and the principle of artificial neural networks, including: Define and obtain the single-class recognition loss function based on Maxwell's equations and the principle of artificial neural networks. The calculation process is shown in the following formula: Where: L single represents the single-class recognition loss function, L MSE,weighted represents the weighted mean square error between the output beam localization and the current detection partition, L residual represents the physical residual loss related to the Maxwell equations, L eff represents the energy efficiency penalty term, y i represents the actual light intensity distribution, represents the target light intensity distribution, I in represents the incident light intensity of the structured light beam to be recognized, represents the Fourier transform of the output electric field, e target represents the target energy efficiency, w i represents the weight corresponding to the loss of the i-th class, K represents the number of structured light patterns to be recognized, α, β, and γ respectively represent the linear superposition weights of the three loss functions, and i represents the number of the current detection partition, represents the target light intensity within the current detection partition; Define and obtain the multi-class demodulation loss function based on Maxwell's equations and the principle of artificial neural networks. The calculation process is shown in the following formula: Wherein, in the formula: L mixed represents a variety of demodulation loss functions, L MSE represents the weighted mean square error between the output beam positioning and the current detection partition, L residual represents the physical residual loss related to the Maxwell equations, L eff represents the energy efficiency penalty term, y i represents the actual light intensity distribution, represents the target light intensity distribution, I in represents the incident light intensity of the structured light beam to be recognized, represents the Fourier transform of the output electric field, e target represents the target energy efficiency, K represents the number of structured light patterns to be recognized, α, β, and γ respectively represent the linear superposition weights of the three loss functions, and i represents the number of the current detection partition, represents the target light intensity within the current detection partition.

10. The construction method of a beam demodulation detector based on a multi-layer metasurface according to claim 7, characterized in that, The above-mentioned based on D 2 The NN model, the single-class recognition loss function, and the multi-class demodulation loss function are used for step-by-step joint learning to optimize and train the initial diffraction neural network to obtain the multi-layer phase distribution, including: Iteratively optimize the initial diffraction neural network based on the preset single-structured light training set and the single-class recognition loss function. Perform single-structured light optimization on the initial parameters of the initial diffraction neural network by the gradient descent method, and obtain the single-class recognition loss function value in real time. When the single-class recognition loss function value meets the preset stability condition, end the iteration to obtain the single-structured light optimization model, and obtain the optimized phase distribution based on the single-structured light optimization model. Perform iterative training based on the single-structured light optimization model, the preset aliased light training set, and the multi-class demodulation loss function. Calculate the gradient of the multi-class demodulation loss function during the aliased light demodulation optimization process during the training process. Based on the optimizer, perform aliased light demodulation optimization on the optimized phase distribution according to the gradient, and update the convergence situation of the multi-class demodulation loss function and the aliased light demodulation situation in real time. Obtain the recognition accuracy value during the aliased light demodulation optimization process based on the aliased light demodulation situation. When the function convergence situation meets the preset convergence condition and the recognition accuracy value is stable, terminate the optimization process and obtain the multi-layer phase distribution based on the current aliased light optimization model.

11. A method for constructing a beam demodulation detector based on a multi-layer metasurface according to claim 7, characterized in that Perform micro-nano processing and preparation based on the multi-layer phase distribution to obtain a multi-layer metasurface, and then fixedly connect the multi-layer metasurface and the image processing chip to construct a beam demodulation detector, including: Perform simulation based on the multi-layer phase distribution to obtain the processing layout of the structure distribution of each layer of the metasurface. Perform layer-by-layer preparation based on the processing layout to obtain a multi-layer metasurface, realize the physical construction of the diffraction neural network, and then fixedly connect the multi-layer metasurface and the image processing chip to construct a beam demodulation detector.

Images