A dynamic multifunctional optical metasurface design method and system

Abstract

This application provides a dynamic multifunctional optical metasurface design method and system. The method includes: acquiring functional requirements under different application scenarios and setting the target reflection spectrum accordingly; inputting the target reflection spectrum into a reverse retrieval network, which is then compressed by an encoder and predicted by a decoder parameter to obtain the predicted design parameters of the meta-unit structure; inputting the predicted design parameters of the meta-unit structure into a forward prediction network, which is then calculated via forward propagation to obtain the predicted reflection spectrum; based on the degree of matching between the predicted reflection spectrum and the target reflection spectrum, a phase retrieval algorithm is used to establish a bidirectional connection between the design domain and the physical domain, and by continuously adjusting the predicted design parameters of the meta-unit structure, the predicted reflection spectrum gradually approaches the target reflection spectrum to obtain the ideal design parameters of the meta-unit structure; and arranging the meta-units in the operating space according to the ideal design parameters of the meta-unit structure to form a metasurface structure that can be dynamically and multifunctionally controlled.

Description

Technical Field

[0001] This application relates to the field of optical metasurface technology, and more specifically, to a dynamic multifunctional optical metasurface design method and system. Background Technology

[0002] Optical metasurfaces are two-dimensional planar structures composed of subwavelength units. Due to their flexible control over the wavefront of light, they have shown great application potential in fields such as high-density optical storage, holographic imaging, and nanoprinting. However, once traditional optical metasurfaces are fabricated, their functions are fixed and lack dynamic adjustment capabilities, which greatly limits their further application in terms of information storage density, security, and multifunctionality.

[0003] To address this issue, researchers have proposed embedding phase change materials (PCMs) into optical metasurfaces. By leveraging the significant changes in optical refractive index between crystalline and amorphous states of PCMs, dynamic functional modulation of the optical metasurface can be achieved. However, existing deep learning-based design methods mostly focus on forward design, i.e., predicting optical responses from the metacell structure. While this method can provide some reference for the design of optical metasurfaces, it still suffers from limitations in algorithm models and training data, including limited design capabilities, lack of global optimization, and reliance on prior knowledge, making it difficult to meet the design requirements of complex, multifunctional optical metasurfaces.

[0004] Therefore, in order to overcome the bottleneck of existing deep learning forward design methods in the design of complex multifunctional optical metasurfaces, it is necessary to explore a novel deep learning-based reverse design method to achieve efficient, accurate design of complex multifunctional optical metasurfaces with global optimization capabilities. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a dynamic multifunctional optical metasurface design method and system to address the shortcomings of the prior art.

[0006] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A dynamic multifunctional optical metasurface design method, comprising the following steps:

[0007] S1. Obtain the functional requirements under different application scenarios and set the target reflectance spectrum accordingly;

[0008] S2. Input the target reflectance spectrum into the reverse retrieval network, and obtain the supercell structure prediction design parameters through encoder compression and decoder parameter prediction.

[0009] S3. Input the supercell structure prediction design parameters into the forward prediction network, and obtain the predicted reflection spectrum through forward propagation calculation of the network;

[0010] S4. Based on the degree of matching between the predicted reflection spectrum and the target reflection spectrum, a two-way connection is established between the design domain and the physical domain using a phase retrieval algorithm. By continuously adjusting the predicted design parameters of the supercell structure, the predicted reflection spectrum gradually approaches the target reflection spectrum, thereby obtaining the ideal design parameters of the supercell structure.

[0011] S5. Arrange the super-units in the operating space according to the ideal design parameters of the super-unit structure to form a metasurface structure that can achieve dynamic multi-functional control.

[0012] Furthermore, in step S1, obtaining the functional requirements under different application scenarios and setting the target reflectance spectrum accordingly includes:

[0013] S11. When the functional requirement is determined to be image edge extraction, the target reflectance spectrum is set to have strong reflection in the high-frequency region and weak reflection in the low-frequency region in order to emphasize spatial high-frequency information.

[0014] S12. When the functional requirement is determined to be image recognition, the target reflectance spectrum is set to have a single-peak structure, and the peak position corresponds to the wavelength of the recognition channel.

[0015] S13. When the functional requirement is determined to be image encryption display, the target reflectance spectrum is set to have a uniform reflectance intensity distribution in the visible light range in order to suppress information leakage.

[0016] Furthermore, in step S2, the reverse retrieval network adopts an encoder-decoder structure, wherein the encoder is used to extract features and compress the latent space of the input target reflection spectrum, and the decoder is used to predict parameters based on the compressed latent space features and in combination with preset optical requirements.

[0017] Furthermore, the loss function of the reverse retrieval network is shown in the following formula:

[0018] ;

[0019] in, This is the mean square error loss term. For the number of parameters, For prediction parameters, For actual parameters, This is a penalty for light transmittance. To preset the minimum light transmittance, To predict light transmittance, This is a reflectivity penalty term. To preset the maximum reflectivity, To predict reflectivity, Here, represents the dynamic weight of the transmittance penalty term, e represents the current training epoch, and E represents the total number of training epochs. The maximum weight of the transmittance penalty term. The dynamic weights of the reflectivity penalty term. The maximum weight of the reflectivity penalty term. This represents the minimum weight of the reflectivity penalty term.

[0020] Furthermore, in step S3, a multimodal fusion input method is adopted to fuse the input numerical parameters with the superunit structure image and the superunit structure text description information. The resulting fused data is input into the forward prediction network. The superunit structure image is subjected to convolutional pooling operation by the convolutional neural network branch of the network to extract the spatial structure features in the image. The superunit structure text description information is subjected to word vector encoding and semantic analysis by the natural language processing branch of the network to mine the functional semantic features contained in the text. The numerical parameters are transformed and mapped by the fully connected neural network branch of the network to map the numerical parameters to a feature space suitable for fusion with image and text features. Finally, the multimodal fusion branch of the network integrates the various features to form a comprehensive feature representation.

[0021] Furthermore, in step S4, the step of establishing a bidirectional connection between the design domain and the physical domain using the phase retrieval algorithm, and continuously adjusting the predicted design parameters of the supercell structure to gradually approximate the target reflection spectrum, thereby obtaining the ideal design parameters of the supercell structure, includes:

[0022] S41. Calculate the error between the predicted light field and the ideal light field in the physical domain;

[0023] S42. Transform the error inversely to the design domain and update the phase distribution;

[0024] S43. Remap the updated phase distribution to the supercell structure parameter space to obtain new geometric parameters;

[0025] S44. Repeat the above steps until the spectral and phase errors are below the set threshold.

[0026] Secondly, this application discloses a dynamic multifunctional optical metasurface design system, the system comprising a scene analysis module, a target spectrum inverse analysis module, a parameter forward prediction module, a parameter adjustment module, and a metacell arrangement module, wherein:

[0027] The scenario analysis module is used to obtain the functional requirements under different application scenarios and set the target reflectance spectrum accordingly.

[0028] The target spectrum reverse analysis module is used to input the target reflectance spectrum into the reverse retrieval network, and obtain the super unit structure prediction design parameters after encoder compression and decoder parameter prediction.

[0029] The parameter forward prediction module is used to input the supercell structure prediction design parameters into the forward prediction network, and obtain the predicted reflectance spectrum through forward propagation calculation of the network;

[0030] The parameter adjustment module is used to establish a two-way connection between the design domain and the physical domain based on the degree of matching between the predicted reflection spectrum and the target reflection spectrum, using a phase retrieval algorithm. By continuously adjusting the predicted design parameters of the supercell structure, the predicted reflection spectrum gradually approaches the target reflection spectrum, thereby obtaining the ideal design parameters of the supercell structure.

[0031] The super-unit arrangement module is used to arrange super-units in the operating space according to the ideal design parameters of the super-unit structure, forming a metasurface structure that can achieve dynamic and multifunctional control.

[0032] Thirdly, this application discloses a readable storage medium, which includes a dynamic multifunctional optical metasurface design method program. When the dynamic multifunctional optical metasurface design method program is executed by a processor, it implements the steps of the method described in any of the preceding claims.

[0033] The beneficial effects of this invention are: by combining the reverse retrieval network and the forward prediction network, the target spectrum is first set according to the functional requirements and the predicted design parameters are quickly obtained. Then, through forward prediction verification and spectral matching optimization, the ideal design parameters are accurately obtained. Finally, the super units are arranged to form a metasurface structure, realizing the efficient transformation from functional requirements to structural design. It can accurately and quickly design a metasurface structure that can be dynamically and multifunctionally controlled and meets the needs of different application scenarios. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a dynamic multifunctional optical metasurface design method disclosed in this invention.

[0035] Figure 2 This is a schematic diagram of the structure of a dynamic multifunctional optical metasurface design system disclosed in this invention;

[0036] Figure 3 This is a schematic diagram of the structure of a readable storage medium disclosed in this invention. Detailed Implementation

[0037] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0038] like Figure 1 As shown, this application discloses a dynamic multifunctional optical metasurface design method, which includes the following steps:

[0039] Step S1: Obtain the functional requirements under different application scenarios and set the target reflectance spectrum accordingly.

[0040] Specifically, this application selects phase change materials. As a core material for dynamic regulation. It should be noted that... It exhibits growth-driven crystallization and low melting point characteristics, compared to traditional phase change materials. Its phase transition period is faster and its energy consumption is lower. Between the crystalline and amorphous states, The optical refractive index changes significantly, providing an excellent material basis for the dynamic manipulation of optical metasurfaces. When In its amorphous state, optical metasurfaces can achieve image edge extraction based on superunit amplitude modulation and far-field image enhancement based on phase modulation; while when... When transformed into a crystalline state, image recognition can be achieved using the correct polarization state. Furthermore, even if the correct polarization state is input, the encrypted information remains invisible, achieving double encryption of the information and greatly enhancing its security.

[0041] Step S2: The target reflectance spectrum is input into the reverse retrieval network, and the supercell structure prediction design parameters are obtained after encoder compression and decoder parameter prediction.

[0042] It should be noted that each metacell, as the basic working unit of the optical metasurface, can achieve amplitude and phase modulation of reflected light by carefully designing its structural design parameters such as size, rotation angle, and working wavelength, thereby meeting the needs of different optical functions.

[0043] Specifically, each supercell employs a metal-insulator-metal (MIM) plasmon resonator structure. This structure enhances optical response and enables precise phase control. Specifically, the length and width of each supercell are designed to vary in 5-nm increments between 20nm and 195nm, while the rotation angle varies in 10-degree increments between 0° and 170°. The operating wavelength is selected as 520nm. By adjusting these parameters, the supercell can achieve amplitude modulation in the near field by controlling the rotation angle according to Malus's law, and phase modulation in the far field by changing its size and rotation angle, thus meeting the requirements of different optical functions.

[0044] Step S3: Input the supercell structure prediction design parameters into the forward prediction network, and obtain the predicted reflection spectrum through forward propagation calculation.

[0045] Specifically, regarding the numerical parameters input to the forward prediction network, it should be noted that these parameters consist of supercell structure-related parameters and optical property parameters. The supercell structure-related parameters include the supercell's geometric dimensions (including length L, width W, height H, and rotation angle θ) and the phase change material. Phase identifier The optical characteristic parameters include the reflectance spectrum. Reflection phase and polarization-related parameters .

[0046] Step S4: Based on the degree of matching between the predicted reflection spectrum and the target reflection spectrum, a two-way connection is established between the design domain and the physical domain using a phase retrieval algorithm. By continuously adjusting the predicted design parameters of the supercell structure, the predicted reflection spectrum gradually approaches the target reflection spectrum, thereby obtaining the ideal design parameters of the supercell structure.

[0047] Step S5: Arrange the super-units in the operating space according to the ideal design parameters of the super-unit structure to form a metasurface structure that can achieve dynamic multi-functional control.

[0048] Specifically, by arranging these optimized supercells in space to form a complete metasurface, the following dynamic functions can be achieved:

[0049] (1) Amorphous state : Perform image edge extraction and enhancement (amplitude / phase dual modulation).

[0050] (2) Crystalline state Perform polarization-selective image recognition and information encryption.

[0051] Ultimately, the constructed optical metasurface can achieve triple reconfigurable functions of image recognition, enhancement, and encryption under phase transition drive, demonstrating the characteristics of low energy consumption, high integration, and strong dynamism.

[0052] As can be seen from the above, the dynamic multifunctional optical metasurface design method disclosed in this application combines a reverse retrieval network and a forward prediction network. First, the target spectrum is set according to the functional requirements and the predicted design parameters are quickly obtained. Then, the ideal design parameters are accurately obtained through forward prediction verification and spectral matching optimization. Finally, the metacells are arranged to form a metasurface structure, realizing the efficient transformation from functional requirements to structural design. It can accurately and quickly design a metasurface structure that can be dynamically and multifunctionally controlled and meets the needs of different application scenarios.

[0053] In one embodiment, step S1, which involves acquiring functional requirements for different application scenarios and setting the target reflectance spectrum accordingly, includes:

[0054] Step S11: When the functional requirement is determined to be image edge extraction, the target reflectance spectrum is set to have strong reflection in the high-frequency region and weak reflection in the low-frequency region in order to emphasize spatial high-frequency information.

[0055] Specifically, image edges correspond to regions in an image where grayscale or color changes drastically; these changes represent high-frequency information in the frequency domain. This application sets the target's reflectance spectrum to have strong reflection in the high-frequency region and weak reflection in the low-frequency region, utilizing the correspondence between optical properties and frequency domain information. Specifically, when light shines on a structure with these target reflectance spectral characteristics, more of the high-frequency components are reflected, while relatively less of the low-frequency components are reflected. Thus, in subsequent image acquisition and processing, by detecting the characteristics of the reflected light, edge information in the image can be highlighted, thereby achieving more effective image edge extraction.

[0056] Step S12: When the functional requirement is determined to be image recognition, the target reflectance spectrum is set to have a single-peak structure, and the peak position corresponds to the wavelength of the recognition channel.

[0057] Specifically, image recognition typically relies on specific optical features to distinguish different target objects, and spectral information is one of the key features. Setting the target's reflectance spectrum to have a single-peak structure, with the peak position corresponding to the wavelength of the recognition channel, is based on the absorption, reflection, or scattering characteristics of materials for specific wavelengths of light. In image recognition scenarios, different categories of objects may have different optical properties. Setting the peak position of the target's reflectance spectrum at the wavelength of the recognition channel can highlight these differences to the greatest extent, enabling accurate differentiation of different image targets.

[0058] Step S13: When the functional requirement is determined to be image encryption display, the target reflectance spectrum is set to have a uniform reflectance intensity distribution in the visible light range in order to suppress information leakage.

[0059] Specifically, the core of image encryption lies in preventing unauthorized access to valid image information. Within the visible light range, the human eye and common optical detection devices primarily perceive image content based on the intensity and distribution of reflected light from objects. When the target's reflectance spectrum is set to have a uniform intensity distribution within the visible light range, it means that the intensity of reflected light across the entire display area is almost uniform, without significant variations in brightness or characteristic patterns. This makes it difficult to extract useful information related to the original image from the reflected light, effectively suppressing information leakage and ensuring the security of the image content.

[0060] In one embodiment, in step S2, the reverse retrieval network adopts an encoder-decoder structure, wherein the encoder is used to extract features and compress the latent space of the input target reflectance spectrum, and the decoder is used to predict parameters based on the compressed latent space features and in combination with preset optical requirements.

[0061] Specifically, for functions such as image edge extraction, recognition, and encrypted display, the encoder, through feature extraction and latent space compression of the target reflectance spectrum, can efficiently capture key information of the target reflectance spectrum, remove noise and redundant data, and make subsequent parameter prediction more accurate and stable. The decoder, combined with preset optical requirements, performs parameter prediction, ensuring that the generated supercell structure design parameters not only meet the characteristics of the target reflectance spectrum but also conform to the performance requirements of practical optical applications. Compared with traditional design methods, this inverse retrieval network has a higher degree of automation and design efficiency. Through extensive data training, it can automatically learn the complex mapping relationship between the target reflectance spectrum and structural parameters, quickly generating the optimal design scheme.

[0062] In one embodiment, the loss function of the reverse retrieval network is shown in the following formula:

[0063] ;

[0064] in, This is the mean square error loss term. For the number of parameters, For prediction parameters, For actual parameters, This is a penalty for light transmittance. To preset the minimum light transmittance, To predict light transmittance, This is a reflectivity penalty term. To preset the maximum reflectivity, To predict reflectivity, Here, represents the dynamic weight of the transmittance penalty term, e represents the current training epoch, and E represents the total number of training epochs. The maximum weight of the transmittance penalty term. The dynamic weights of the reflectivity penalty term. The maximum weight of the reflectivity penalty term. This represents the minimum weight of the reflectivity penalty term.

[0065] It should be noted that the loss function formula above introduces penalty terms related to optical performance, such as transmittance and reflectance. When the supercell structure generated by the predicted parameters does not meet these optical performance indicators, the corresponding penalty terms increase the loss value, thereby guiding the network to optimize in a direction that meets the optical requirements during training.

[0066] Specifically, in this inverse retrieval network, the mean squared error loss term primarily measures the difference between the predicted and actual parameters. The transmittance penalty term aims to ensure that the generated supercell structure meets the preset transmittance requirements. When the predicted transmittance... Below the preset minimum transmittance hour, It is a positive number. This will generate a positive value as a penalty for the network failing to meet the transmittance requirement. This penalty term will be added to the total loss function, prompting the network to adjust its parameters during training to improve the predicted transmittance. as high as possible This is to meet the lower limit requirement of transmittance in practical applications. Similar to the transmittance penalty term, the reflectance penalty term is used to ensure that the generated supercell structure meets the preset reflectance limit. If the predicted reflectance... Exceeded the preset maximum reflectivity , A positive value is calculated as a penalty for the network failing to meet the reflectivity requirement. This penalty term prompts the network to optimize its parameters during training, improving its ability to predict reflectivity. as low as possible This is to meet the upper limit requirements for reflectivity in practical applications.

[0067] It should be further noted that as the number of training rounds e increases, The transmittance penalty will gradually increase. This means that in the early stages of training, the network pays relatively little attention to transmittance and focuses mainly on learning the basic pattern of parameter prediction; as training progresses, the network will gradually increase its attention to transmittance to ensure that the final generated supercell structure can meet the transmittance requirements.

[0068] It should be further noted that in the early stages of training, Approaching The reflectivity penalty term is given a large weight to make the network quickly focus on the reflectivity limitation; as training progresses, the weight is gradually reduced to avoid over-limiting the network's learning of other aspects, while ensuring that the final result meets the reflectivity requirement.

[0069] In one embodiment, in step S3, a multimodal fusion input method is adopted to fuse the input numerical parameters with the superunit structure image and the superunit structure text description information. The resulting fused data is input to the forward prediction network. The superunit structure image is subjected to convolutional pooling operation by the convolutional neural network branch of the network to extract the spatial structure features in the image. The superunit structure text description information is subjected to word vector encoding and semantic analysis by the natural language processing branch of the network to mine the functional semantic features contained in the text. The numerical parameters are transformed and mapped by the fully connected neural network branch of the network to map the numerical parameters to a feature space suitable for fusion with image and text features. Finally, the multimodal fusion branch of the network integrates the various features to form a comprehensive feature representation.

[0070] It should be noted that this multimodal fusion input and processing method fully leverages the advantages of different modalities of data. Images can intuitively present the geometry and spatial layout of the supercell, while text descriptions accurately convey the functional requirements and objectives of the design, and numerical parameters provide specific quantitative indicators. Through the targeted processing of different modalities by each branch network and the integration of the multimodal fusion branches, the characteristic information of the supercell structure can be captured comprehensively and deeply, effectively compensating for the shortcomings of single-modal data. This significantly improves the prediction accuracy and reliability of the forward prediction network for the optical properties of the supercell structure, laying a solid foundation for subsequent implementation of complex functions such as image recognition, enhancement, and encryption.

[0071] In one embodiment, step S4, which involves establishing a bidirectional connection between the design domain and the physical domain using a phase retrieval algorithm and continuously adjusting the supercell structure's predicted design parameters to gradually approximate the target reflection spectrum, thereby obtaining the ideal design parameters for the supercell structure, includes:

[0072] Step S41: Calculate the error between the predicted light field and the ideal light field in the physical domain.

[0073] Step S42: Transform the error inversely to the design domain and update the phase distribution.

[0074] Step S43: Remap the updated phase distribution to the supercell structure parameter space to obtain new geometric parameters.

[0075] Step S44: Repeat the above steps until the spectral and phase errors are below the set threshold.

[0076] Specifically, this process utilizes the iterative approach of phase retrieval algorithms to establish a closed-loop feedback optimization between the "design domain" (parameter space of the supercell) representing the physical structure and the "physical domain" (spectrum and phase of the far field) representing the optical response. Its core objective is to deduce the phase distribution of the metasurface that can produce the target's reflection spectrum (i.e., the light field amplitude), given the target's reflection spectrum, and ultimately determine the corresponding supercell geometry. The principle of this iterative optimization process can be broken down as follows:

[0077] Regarding step S41, it should be noted that this step applies constraints in the "physical domain." First, based on the supercell structure parameters obtained from the previous iteration or the initial prediction, this application calculates the resulting optical response, i.e., the "predicted light field," through a forward prediction network (or electromagnetic simulation). This light field contains amplitude (predicted reflection spectrum) and phase information. Simultaneously, we have a design target, the "ideal light field," whose amplitude is determined by a preset "target reflection spectrum," but whose phase is unknown.

[0078] The core operation of "calculation error" is to replace the amplitude of the "predicted light field" with the amplitude of the "ideal light field" (i.e., the target reflection spectrum), while retaining the phase information of the "predicted light field." This creates a new light field that combines the target amplitude and the predicted phase. The physical meaning of this operation is to force the light field to meet the preset design requirements in terms of amplitude, and then observe how the phase needs to be adjusted to achieve this goal.

[0079] Regarding step S42, it's important to note that this step involves feeding back the constraint information from the physical domain to the "design domain." Here, this application propagates the new light field (target amplitude + predicted phase) formed in the previous step from the physical domain (far field) back to the design domain (metasurface plane) using mathematical methods such as inverse Fourier transform. After the inverse transform, a new complex amplitude distribution is obtained on the metasurface plane. The phase component of this new distribution is the "updated phase distribution." This new phase is a correction made by the algorithm to the original phase distribution based on the requirements of the target spectrum, making it closer to the phase distribution that produces the ideal result.

[0080] Regarding step S43, it should be noted that this step is crucial in connecting the abstract design with the concrete implementation. In step S42, this application obtained an ideal, continuous phase distribution pattern. However, in actual manufacturing, the phase cannot be directly "printed"; instead, it needs to be achieved by arranging "supercells" with specific geometries.

[0081] Therefore, a pre-established "phase-structure" database or mapping relationship is needed (this relationship can be obtained through previous forward network training or extensive simulations). This database or mapping relationship records the phase response that supercells with different geometric parameters (such as length L, width W, rotation angle θ, etc.) can produce. The "remapping" process involves traversing every point (or region) on the "updated phase distribution map" and, based on the required phase value, querying or matching the geometric parameters of the supercells that can achieve that phase. In this way, the abstract phase requirement is materialized into a completely new and specific set of "supercell structure design parameters".

[0082] Regarding step S44, it should be noted that this step serves as the loop and termination condition for the entire optimization process. Specifically, this application uses the new geometric parameters obtained in step S43 as the starting point for a new round of iterations and re-executes steps S41 to S43. In each loop, due to the application of dual constraints in the physical domain (amplitude) and the design domain (structure-phase relationship), the predicted reflection spectrum gradually approaches the target reflection spectrum. Specifically, by calculating the mean square error and other deviation indices between the predicted spectrum and the target spectrum in each round, when this error is determined to be less than a pre-set sufficiently small threshold, the optimization process can be considered to have converged. The resulting superunit structure geometric parameters are the "ideal design parameters" that meet the design requirements.

[0083] In summary, the iterative process from steps S41 to S44 is essentially a feedback loop that continuously searches for the optimal solution in the solution space. It cleverly combines physical constraints (target spectrum) and structural feasibility (phase-structure mapping), and by transferring and correcting information back and forth between the physical domain and the design domain, it ultimately converges efficiently and accurately to an optical metasurface physical structure design scheme that meets the complex functional requirements.

[0084] Please refer to Figure 2 This application discloses a dynamic multifunctional optical metasurface design system, which includes a scene analysis module, a target spectrum inverse analysis module, a parameter forward prediction module, a parameter adjustment module, and a metacell arrangement module, wherein:

[0085] The scenario analysis module is used to obtain the functional requirements under different application scenarios and set the target reflectance spectrum accordingly.

[0086] The target spectrum inverse analysis module is used to input the target reflectance spectrum into the inverse retrieval network, and obtain the supercell structure prediction design parameters through encoder compression and decoder parameter prediction.

[0087] The parameter forward prediction module is used to input the supercell structure prediction design parameters into the forward prediction network, and obtain the predicted reflection spectrum through forward propagation calculation of the network.

[0088] The parameter adjustment module is used to establish a two-way connection between the design domain and the physical domain based on the degree of matching between the predicted reflection spectrum and the target reflection spectrum, using a phase retrieval algorithm. By continuously adjusting the predicted design parameters of the supercell structure, the predicted reflection spectrum gradually approaches the target reflection spectrum, thereby obtaining the ideal design parameters of the supercell structure.

[0089] The super-unit arrangement module is used to arrange super-units in the operating space according to the ideal design parameters of the super-unit structure, forming a metasurface structure that can achieve dynamic and multifunctional control.

[0090] In one embodiment, the above modules are also used to implement a dynamic multifunctional optical metasurface design method as described in any of the foregoing method embodiments, and this application does not limit this.

[0091] As can be seen from the above, the dynamic multifunctional optical metasurface design system disclosed in this application combines a reverse retrieval network and a forward prediction network. First, the target spectrum is set according to the functional requirements and the predicted design parameters are quickly obtained. Then, the ideal design parameters are accurately obtained through forward prediction verification and spectral matching optimization. Finally, the metaunits are arranged to form a metasurface structure, realizing the efficient transformation from functional requirements to structural design. It can accurately and quickly design a metasurface structure that can be dynamically and multifunctionally controlled and meets the needs of different application scenarios.

[0092] Please refer to Figure 3 This application discloses a readable storage medium, which includes a dynamic multifunctional optical metasurface design method program. When the dynamic multifunctional optical metasurface design method program is executed by a processor, it implements the steps of the method described in any of the preceding claims.

[0093] As can be seen from the above, the readable storage medium disclosed in this application combines a reverse retrieval network and a forward prediction network. First, the target spectrum is set according to the functional requirements and the predicted design parameters are quickly obtained. Then, through forward prediction verification and spectral matching optimization, the ideal design parameters are accurately obtained. Finally, the meta-units are arranged to form a metasurface structure, realizing the efficient transformation from functional requirements to structural design. It can accurately and quickly design a metasurface structure that can be dynamically and multifunctionally controlled and meets the needs of different application scenarios.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dynamic multi-functional optical metasurface design method, characterized in that, The method comprises the following steps: S1, obtaining functional requirements in different application scenarios, and setting target reflection spectrum accordingly; S2, inputting the target reflection spectrum into the reverse retrieval network, compressing through the encoder and predicting the parameters through the decoder to obtain the super cell structure prediction design parameters; S3, inputting the super cell structure prediction design parameters into the forward prediction network, calculating through network forward propagation to obtain the predicted reflection spectrum; S4, based on the matching degree between the predicted reflection spectrum and the target reflection spectrum, using a phase recovery algorithm to establish a two-way relationship between the design domain and the physical domain, and by continuously adjusting the super cell structure prediction design parameters, the predicted reflection spectrum gradually approaches the target reflection spectrum, and the ideal super cell structure design parameters are obtained; S5, arranging the super cell in the operation space according to the ideal super cell structure design parameters to form a super surface structure that can realize dynamic multi-functional regulation and control; The loss function of the reverse retrieval network is as follows: ； wherein, is a mean squared error loss term, is a number of parameters, is a predicted parameter, is a true parameter, is a transmittance penalty term, is a preset minimum transmittance, is a predicted transmittance, is a reflectance penalty term, is a preset maximum reflectance, is a predicted reflectance, is a dynamic weight for the transmittance penalty term, e is a current training epoch, E is a total training epoch, is a maximum weight for the transmittance penalty term, is a dynamic weight for the reflectance penalty term, is a maximum weight for the reflectance penalty term, is a minimum weight for the reflectance penalty term.

2. The method of claim 1, wherein, In step S1, the functional requirements in different application scenarios are obtained, and the target reflection spectrum is set accordingly, which comprises: S11, when the functional requirement is image edge extraction, the target reflection spectrum is set to reflect strongly in the high frequency region and weakly in the low frequency region to emphasize the spatial high frequency information; S12, when the functional requirement is image recognition, the target reflection spectrum is set to be unimodal, and the peak position corresponds to the recognition channel wavelength; S13, when the functional requirement is image encryption display, the target reflection spectrum is set to have uniform reflection intensity distribution in the visible light range to suppress information leakage.

3. The method of claim 1, wherein, In step S2, the reverse retrieval network adopts an encoder-decoder structure, wherein the encoder is used for feature extraction and latent space compression processing of the input target reflection spectrum, and the decoder is used for parameter prediction based on the compressed latent space features and combined with the preset optical requirements.

4. The method of claim 1, wherein, In step S3, a multi-modal fusion input method is adopted to fuse the input numerical parameters, super cell structure images and super cell structure text description information, wherein the fusion data formed is input into the forward prediction network, the spatial structure features in the image are extracted through convolution and pooling operations of the convolutional neural network branch in the network, the functional semantic features contained in the text are mined through word vector encoding and semantic analysis processing of the natural language processing branch in the network, and the numerical parameters are mapped to a feature space suitable for fusion with image and text features through feature transformation and mapping of the fully connected neural network branch in the network. Finally, the features are integrated through the multi-modal fusion branch in the network to form a comprehensive feature representation.

5. The method of claim 1, wherein, In step S4, the phase recovery algorithm is used to establish a two-way relationship between the design domain and the physical domain, and the ideal super cell structure design parameters are obtained by continuously adjusting the super cell structure prediction design parameters so that the predicted reflection spectrum gradually approaches the target reflection spectrum, comprising: S41, calculating the error between the predicted light field and the ideal light field in the physical domain; S42, inverse transform the error to the design domain, and update the phase distribution; S43, remap the updated phase distribution to the supercell structure parameter space to obtain new geometric parameters; S44, repeat the above steps until the spectrum and phase error are below the set threshold.

6. A dynamic multi-functional optical metasurface design system, characterized in that, The system comprises a scene analysis module, a target spectrum inverse analysis module, a parameter forward prediction module, a parameter adjustment module, and a supercell arrangement module, wherein: The scene analysis module is configured to obtain functional requirements in different application scenarios and set target reflectance spectra accordingly; The target spectrum inverse analysis module is configured to input the target reflectance spectrum into an inverse retrieval network, compress it through an encoder, and predict the parameters through a decoder to obtain supercell structure prediction design parameters; The parameter forward prediction module is configured to input the supercell structure prediction design parameters into a forward prediction network, calculate the predicted reflectance spectrum through network forward propagation; The parameter adjustment module is configured to establish a two-way link between the design domain and the physical domain based on the matching degree between the predicted reflectance spectrum and the target reflectance spectrum using a phase recovery algorithm, continuously adjust the supercell structure prediction design parameters, make the predicted reflectance spectrum gradually approach the target reflectance spectrum, and obtain ideal supercell structure design parameters; The supercell arrangement module is configured to arrange the supercells in the operation space according to the ideal supercell structure design parameters to form a dynamic multifunctional optical super surface structure; The loss function of the inverse retrieval network is shown in the following formula: ； wherein, is a mean squared error loss term, is a number of parameters, is a predicted parameter, is a true parameter, is a transmittance penalty term, is a preset minimum transmittance, is a predicted transmittance, is a reflectance penalty term, is a preset maximum reflectance, is a predicted reflectance, is a dynamic weight for the transmittance penalty term, e is a current training epoch, E is a total training epoch, is a maximum weight for the transmittance penalty term, is a dynamic weight for the reflectance penalty term, is a maximum weight for the reflectance penalty term, is a minimum weight for the reflectance penalty term.

7. A readable storage medium characterized by, The readable storage medium comprises a dynamic multifunctional optical super surface design method program, and the dynamic multifunctional optical super surface design method program is executed by the processor to realize the steps of the method in any one of claims 1 to 5.

Images