A magnetic-aware deep learning driven intelligent metamaterial on-demand design system and method

Abstract

This invention relates to a magnetic sensing deep learning-driven intelligent metamaterial on-demand design system and method. By setting up a terminal device, a unit cell structure device, a prediction device, and a simulation device, the unit cell structure device can generate a magnetic field modulation spectrum with the same dimension as the target band structure matrix using user-defined target parameters. Furthermore, it uses a magnetic conditional generative adversarial network model algorithm to transform the magnetic field modulation spectrum into multiple candidate unit cell structures. The prediction device can screen out all candidate unit cell structures whose bandgap attribute characteristics deviate from the target bandgap by less than a first threshold. The simulation device can perform finite element simulation on each candidate unit cell structure screened by the prediction device, selecting all candidate unit cell structures that simultaneously meet the bandgap performance threshold and engineering constraints. Ultimately, this achieves rapid and accurate prediction of the band structure of intelligent metamaterials under given excitation field conditions and on-demand design of unit cell structures.

Description

Technical Field

[0001] This invention relates to the field of intelligent metamaterial design and elastic wave manipulation technology, and more specifically, to a magnetic sensing deep learning-driven intelligent metamaterial on-demand design system and method. Background Technology

[0002] Metamaterials, as functional materials with artificially designed structures, can achieve specific wave transmission bandgap characteristics through the periodic arrangement of unit cell structures, and have significant application value in vibration suppression, noise isolation, and waveguide manipulation. However, once traditional metamaterials are fabricated, their bandgap position and bandwidth are essentially fixed, making it difficult to adjust them online according to operating conditions. Magnetorheological elastomers (MREs) are a class of intelligent flexible materials composed of a polymer matrix and soft magnetic particles, whose elastic modulus and damping properties can be significantly and reversibly controlled under the influence of an external magnetic field. Combining MREs with metallic materials to construct intelligent metamaterials enables real-time online control of the bandgap.

[0003] In existing technologies, the design of MRE metamaterials largely relies on empirical structural iterations or finite element simulation optimization, resulting in long design cycles and poor compatibility between bandgap performance and magnetic fields. Furthermore, bandgap control often employs a "passive response" control mode, indirectly adjusting the mechanical properties of the MRE to change the bandgap by altering the external magnetic field, making it difficult to achieve a closed loop of "on-demand design - precise control - performance matching." Bandgap prediction for MRE metamaterials mainly relies on theoretical calculations or finite element simulations, requiring repeated model building under different geometric parameters and magnetic field strengths, applying Bloch periodic boundary conditions, and solving large-scale generalized eigenvalue problems to obtain band structure and bandgap information. These methods are computationally intensive and time-consuming per simulation, making it difficult to support large-scale searches and optimizations in high-dimensional design spaces. In the design of MRE metamaterial unit cell structures, traditional reverse design methods based on topology optimization or heuristic algorithms require repeated physical simulations in each optimization iteration, resulting in low computational efficiency, sensitivity to the initial design, and a tendency to get trapped in local optima, making it difficult to achieve coordinated optimization of structure and magnetic field.

[0004] In recent years, deep learning has been introduced into the field of metamaterial design. Some works have used convolutional neural networks and generative adversarial networks to construct surrogate models between structure and bandgap, achieving rapid prediction of bandgap location or bandwidth, or using generative networks for reverse design of topology. Although deep learning technology has begun to be applied to metamaterial design, it has mostly focused on the mapping relationship between structural parameters and bandgap performance, failing to incorporate the key regulatory factor of excitation field (such as magnetic field) information into the design input dimension. This results in the designed metamaterials failing to maintain the stability of the target bandgap performance under different magnetic field environments. With changes in magnetic field strength, the elastic modulus of the metamaterial exhibits significant nonlinear changes, causing changes in the location, width, and existence of the bandgap in the unit cell structure, making structural design a strongly nonlinear, multiphysics coupled problem. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to overcome the technical defects of existing deep learning technology in metamaterial design, which is difficult to maintain the stability of target bandgap performance and the instability of the position, width and existence of bandgap in unit cell structure. In order to overcome the defects of the prior art, this invention provides a magnetic sensing deep learning driven intelligent metamaterial on-demand design system and method, specifically including a magnetic sensing deep learning driven intelligent metamaterial on-demand design system and a magnetic sensing deep learning driven intelligent metamaterial on-demand design method.

[0006] This invention provides a magnetic sensing deep learning-driven intelligent metamaterial on-demand design system, comprising: [missing information - likely related to a closed loop] Terminal equipment, used to allow users to input target parameters including the target band structure matrix and the target magnetic field strength; The single-cell structure device is configured to generate a magnetic field modulation spectrum with the same dimension as the target band structure matrix using the target parameters according to the Gaussian perturbation function rule, and to transform the magnetic field modulation spectrum into multiple candidate single-cell structures through a model algorithm of magnetic conditional generative adversarial network. A prediction device is used to predict the band structure matrix and / or band gap attribute features of each candidate unit cell structure obtained by the unit cell structure device under the target magnetic field, and to screen out all candidate unit cell structures whose band gap attribute features deviate from the target band gap by less than a first threshold. The simulation equipment is configured to operate as follows: During formal operation, finite element simulation is performed on each candidate unit cell structure screened by the prediction device to obtain the accurate band structure matrix and / or accurate bandgap attribute characteristics of each candidate unit cell structure under the target magnetic field conditions. According to the design requirements of the target engineering system, the engineering indicators of each candidate unit cell structure are checked, and all candidate unit cell structures that simultaneously meet the bandgap performance threshold and engineering constraints are screened out. Then, the terminal device is called to display these candidate unit cell structures. During parameter optimization, finite element simulation is performed on each unit cell structure to obtain the accurate band structure matrix and / or accurate bandgap property characteristics of each unit cell structure under the target magnetic field conditions.

[0007] The magnetic sensing deep learning-driven intelligent metamaterial on-demand design system disclosed in this invention addresses the aforementioned technical deficiencies by setting up a terminal device, a unit cell structure device, a prediction device, and a simulation device. The unit cell structure device generates a magnetic field modulation spectrum of the same dimension as the target band structure matrix using user-defined target parameters according to a Gaussian perturbation function rule. The unit cell structure device further transforms the magnetic field modulation spectrum into multiple candidate unit cell structures through a magnetic conditional generative adversarial network (GAN) model algorithm. This GAN model algorithm can generate multiple sets of candidate unit cell structures simultaneously under given target band structure and target magnetic field strength, while also ensuring that the generated structures meet manufacturability requirements such as bandgap performance and connectivity. Simultaneously, by setting up a prediction device, the system can predict the band structure matrix and / or bandgap attribute characteristics of each candidate unit cell structure acquired by the unit cell structure device under the target magnetic field, and select all candidate unit cell structures whose bandgap attribute characteristics deviate from the target bandgap by less than a first threshold. By setting up simulation equipment to perform finite element simulations on each candidate unit cell structure selected by the prediction equipment, the accurate band structure matrix and / or accurate bandgap attribute characteristics of each candidate unit cell structure under the target magnetic field conditions are obtained. This explicit magnetic field-structure coupling modeling scheme enables rapid and accurate prediction of the band structure of intelligent metamaterials under a given excitation field (magnetic field strength) and on-demand design of unit cell structures. The combination of simulation equipment, prediction equipment, and unit cell structure equipment forms a magnetic field sensing encoding and deep learning generation framework. This generative architecture overcomes the technical shortcomings of existing deep learning technologies applied to metamaterial design, such as difficulty in maintaining the stability of the target bandgap performance and instability in the position, width, and existence of the bandgap in unit cell structures. It significantly improves the efficiency and accuracy of band structure prediction, reducing the bandgap prediction time of a single structure under a given magnetic field from minutes in traditional finite element simulations to seconds or even milliseconds while ensuring a small average relative error, thus significantly improving the efficiency of design iteration. The simulation equipment can also screen out all candidate unit cell structures that simultaneously meet the bandgap performance threshold and engineering constraints, enabling the evaluation of large-scale candidate structures within a reasonable timeframe. This effectively supports the design optimization of engineering systems such as vehicle suspensions and vibration isolation components, providing support for rapid engineering decision-making. The entire system is not only applicable to MRE elastic metamaterials but can also be extended to piezoelectric metamaterials, acoustic metamaterials, and other multiphysics tunable systems. This method provides a data-driven solution for the tunable bandgap design of various smart materials and structures.

[0008] In one possible implementation, the unit cell structure device includes: The modulation module, which communicates with the terminal device, is configured to normalize the target band structure matrix and substitute the target magnetic field strength and the normalized target band structure matrix into a Gaussian perturbation function to generate a magnetic field modulation spectrum with the same dimension as the target band structure matrix. A magnetic conditional generative adversarial network, which communicates with both the modulation module and the prediction device, is configured to convert the magnetic field modulation spectrum into multiple candidate unit cell structures.

[0009] The magnetic field modulation spectrum is constructed by perturbing the Gaussian spectrum, and a magnetic conditional generative adversarial network is used as the generator. The generator contains a Transformer module, which can generate multiple sets of candidate unit cell structures at one time under given target band structure and target magnetic field strength. This realizes the on-demand design of unit cell structures that meet the expected band structure under any magnetic field strength.

[0010] In one possible implementation, the Gaussian perturbation function is calculated as follows: , In the formula, These are the matrix elements of the normalized band structure matrix. For row indicators, List the indicators; It is a magnetic field modulation spectrum. The magnetic field strength, For the maximum magnetic field strength, , and These are learnable magnetic field-related parameters.

[0011] In one possible implementation, the prediction device includes: The weighting module communicates with the magnetic condition generative adversarial network and is used to calculate the magnetic field response weight map of the candidate unit cell structure. The candidate unit cell structure and the magnetic field response weight map are spliced ​​together according to the channel dimension to form a dual-channel input tensor. The generation module, which communicates with both the weighting module and the simulation device, is configured to first call its included magnetic conditional attention mapping prediction network to convert the dual-channel input tensor into the band structure matrix and / or bandgap attribute features of the candidate unit cell structure under the target magnetic field, and then filter out all candidate unit cell structures whose bandgap attribute features deviate from the target bandgap by less than a first threshold.

[0012] The prediction device, equipped with the aforementioned structure and functions, jointly encodes the magnetic field as a spatial weighted graph with the structural topology through a weighting module. Using a magnetic conditional attention mapping prediction network, it concatenates the original structural graph and the magnetic field response weighted graph into a dual-channel input. Introducing a channel-space joint attention mechanism, it adaptively allocates the importance of geometric and magnetic field features within a multi-scale feature space, enabling high-precision and rapid prediction of the bandgap characteristics of two-dimensional MRE metamaterials under various magnetic field conditions. This network can reduce the bandgap prediction time for a single structure under a given magnetic field from minutes in traditional finite element simulations to seconds or milliseconds while maintaining a small average relative error. This provides support for the rapid screening of large batches of candidate structures and different magnetic circuit schemes, significantly improving the coupled modeling capability of magnetic fields and structures, and achieving rapid and accurate prediction of the band diagram and bandgap of a unit cell structure under any given magnetic field.

[0013] In one possible implementation, the weighting module is configured to perform the following steps: A1: The Sobel operator is used to calculate the structural gradient of the candidate unit cell structure to obtain the gradient magnitude map. At the same time, the structural map of the candidate unit cell structure is inverted to obtain the MRE region mask. A2: The gradient magnitude map and the MRE region mask are weighted and summed to obtain the comprehensive sensitivity map; A3: Calculate the point-by-point product of the normalized magnetic field strength and the comprehensive sensitivity map to obtain the magnetic field response weight map; A4: Concatenate the candidate unit cell structure with the magnetic field response weight map along the channel dimension to obtain the dual-channel input tensor.

[0014] By using structural gradient magnetic sensing encoding, the magnetic field is jointly encoded with the structural topology in the form of a spatial weighted graph, which significantly improves the coupling modeling capability of magnetic field and structure, and enables rapid and accurate prediction of band diagram and band gap of unit cell structure under a given arbitrary magnetic field.

[0015] Another technical solution of the present invention is to provide a magnetic sensing deep learning-driven on-demand design method for intelligent metamaterials, the method comprising the following steps: S1: Using seed point-based morphological dilation operations on a computer device, multiple relatively concentrated and continuous unit cell structures of MRE materials are generated; S2: Perform finite element simulation on each unit cell structure using simulation equipment to obtain the accurate band structure matrix and / or accurate bandgap property characteristics of each unit cell structure under magnetic field conditions; S3: Using the precise band structure and / or precise bandgap properties of the unit cell obtained in step S2 as the ideal output, the weighted sum of the mean absolute error and the mean relative error is used as the loss function to optimize the parameters of the prediction device. S4: Using the multiple unit cell structures obtained in step S1, the parameters of the unit cell structure device are optimized by using the weighted sum of the adversarial loss function, topological connectivity loss function and band matching loss function as the total loss function. S5: The user inputs target parameters, including the target band structure matrix and the target magnetic field strength, through the terminal device. Then, by running the unit cell structure device, prediction device and simulation device in sequence, all candidate unit cell structures that simultaneously meet the band gap performance threshold and engineering constraints are selected.

[0016] The method disclosed in this invention first completes explicit magnetic field-structure coupling modeling through steps S1 and S2, enabling rapid and accurate prediction of the band structure of intelligent metamaterials under a given excitation field (magnetic field strength) and on-demand design of unit cell structures. Subsequently, the weighted sum of mean absolute error and mean relative error is used as a loss function to optimize the parameters of the prediction device, ensuring that the generated structure meets manufacturability requirements such as bandgap performance and connectivity. This significantly improves the coupling modeling capability between the magnetic field and the structure, enabling rapid and accurate prediction of the band diagram and bandgap of unit cell structures under any given magnetic field, and also achieving on-demand design of unit cell structures that meet the expected band structure under any magnetic field strength. The prediction network constructed through steps S2, S3, and S4 can reduce the bandgap prediction time of a single structure under a given magnetic field from minutes in traditional finite element simulation to seconds or even milliseconds while maintaining a small mean relative error. This significantly improves the efficiency of design iteration and overcomes the technical shortcomings of existing deep learning techniques applied to metamaterial design, such as difficulty in maintaining the stability of the target bandgap performance and instability in the position, width, and existence of the bandgap in unit cell structures. This technology enables the evaluation of a large number of candidate structures within a reasonable time, thereby effectively supporting the design optimization of engineering systems such as vehicle suspension and vibration isolation components, and providing support for rapid engineering decision-making.

[0017] In one possible implementation, step S1 includes the following steps: S11: Encode the MRE material and the metal material with 0 and 1 respectively, and draw a regular shape on the pixel value matrix where all pixel values ​​are equal to the corresponding pixel values ​​of the metal material. Then, arrange multiple seed points on several sides of the regular shape; wherein, the regular shape is a right triangle, rectangle or rhombus. S12: Starting from the seed point, generate an MRE material region on the pixel value matrix where the proportion of the corresponding pixel value of the MRE material encoding on the pixel value matrix is ​​within a preset range through multi-stage morphological dilation operation. S13: The unit cell structure diagram is generated by performing multiple orthogonal transformations on the MRE material region in the pixel value matrix.

[0018] The above steps ultimately generate a binary distribution topology of MRE and metal (such as tungsten). During the generation process, the magnetic field prior of "MRE magnetic drive variable stiffness characteristics" is considered. Morphological expansion operation based on seed points is used to generate a relatively concentrated and continuous unit cell structure of MRE material, which lays the structural foundation for the magnetic sensing deep learning network to accurately learn the mapping relationship of "magnetic field-MRE variable stiffness-adjustable bandgap".

[0019] In one possible implementation, step S3 includes the following steps: S31: The Sobel operator is used to calculate the structural gradient of each unit cell structure map to obtain their respective gradient magnitude maps, and the inversion operation is performed on each unit cell structure map to obtain their respective MRE region masks. S32: Perform weighted summation on the gradient magnitude map and MRE region mask of each unit cell structure map to obtain their respective comprehensive sensitivity maps; S33: Calculate the point-by-point product of the normalized magnetic field strength and the comprehensive sensitivity map of each unit cell structure map to obtain the respective magnetic field response weight map; S34: Concatenate each unit cell structure diagram with the magnetic field response weight diagram of that unit cell structure diagram according to the channel dimension to obtain their respective dual-channel input tensors; S35: Using the precise band structure matrix and / or precise bandgap attribute features of the unit cell structure obtained in step S2 as the ideal output, and using the weighted sum of the mean absolute error and the mean relative error as the loss function, the parameters of the magnetic conditional attention mapping prediction network in the prediction device are optimized until the prediction error of the magnetic conditional attention mapping prediction network is lower than a preset threshold, at which point the parameter optimization of the prediction device is considered complete.

[0020] The above steps, through structural gradient magnetic sensing encoding, fuse the boundary gradient information of the unit cell with the MRE region mask according to weights, forming a comprehensive structural feature sensitive to the bandgap. This feature is then multiplied point-by-point with the normalized magnetic field strength to generate a magnetic field response weight map. This effectively suppresses the magnetic field contribution of non-MRE regions at the feature level, ensuring that the magnetic field primarily affects the characterization of the MRE region. This step, on the one hand, highlights the sensitivity of the MRE region to magnetic field changes, reducing the interference of the magnetic field distribution in metallic material regions on bandgap prediction; on the other hand, it improves the modeling ability of the subsequent magnetic condition attention mapping network for the "magnetic field-structure" coupling characteristics, which is beneficial for improving magnetic field utilization and bandgap control accuracy under given magnetic conditions.

[0021] Furthermore, by constructing and training a magnetic condition attention prediction network, the original structure diagram and the magnetic field response weight diagram are concatenated into a dual-channel input. A channel-space joint attention mechanism is introduced to adaptively allocate the importance of geometric and magnetic field features within a multi-scale feature space, enabling high-precision and rapid prediction of the bandgap properties of two-dimensional MRE metamaterials under various magnetic field conditions. This network can reduce the bandgap prediction time for a single structure under a given magnetic field from minutes in traditional finite element simulations to seconds or milliseconds while maintaining a small average relative error, providing support for the rapid screening of large batches of candidate structures and different magnetic circuit schemes.

[0022] In one possible implementation, step S4 includes the following steps; S41: Normalize the exact band structure matrix of each unit cell structure under magnetic field conditions, and substitute the magnetic field strength and normalized band structure matrix set in the magnetic field conditions into the Gaussian perturbation function to generate the magnetic field modulation spectrum of each structure with the same dimension as the band structure matrix. S42: Input the magnetic field modulation spectrum of each unit cell structure into the magnetic condition generative adversarial network in the unit cell structure device to generate their respective predicted unit cell structure diagrams. S43: Input each unit cell structure and its predicted unit cell structure map into the PatchGAN discriminator network based on spectral normalization to obtain their respective local structure authenticity score matrices. S44: Using the weighted sum of the adversarial loss function, topological connectivity loss function, and bandgap matching loss function as the total loss function, substitute the local structure authenticity score matrix of each unit cell structure into the total loss function, and train the PatchGAN discriminator network and the magnetic conditional generative adversarial network using an alternating optimization method until the connectivity and bandgap matching accuracy of the predicted unit cell structure graph output by the magnetic conditional generative adversarial network meet the preset requirements. This is considered as completing the parameter optimization of the unit cell structure device.

[0023] The above steps construct a joint conditional tensor of band frequency and magnetic field based on Gaussian spectral perturbation, and optimize the model parameters of the magnetic conditional generative adversarial network (GAN) to generate multiple sets of candidate unit cell structures at once, given a target band structure and target magnetic field strength. By utilizing joint constraints of adversarial loss, topological connectivity loss, and dispersion matching loss, the generated structures satisfy manufacturability requirements such as bandgap performance and connectivity.

[0024] In one possible implementation, step S5 includes the following steps: S51: The user inputs target parameters, including the target band structure matrix and the target magnetic field strength, on the terminal device; S52: Using a single-cell structure device according to the Gaussian perturbation function rule, a magnetic field modulation spectrum with the same dimension as the target band structure matrix is ​​generated using the target parameters. The magnetic field modulation spectrum is then transformed into multiple candidate single-cell structures using a model algorithm of a magnetic conditional generative adversarial network. S53: The band structure matrix and / or band gap attribute characteristics of each candidate unit cell structure obtained by the unit cell structure device under the target magnetic field are predicted by the prediction device, and all candidate unit cell structures whose band gap attribute characteristics deviate from the target band gap by less than a first threshold are selected. S54: The simulation equipment performs finite element simulation verification on each candidate unit cell structure selected by the prediction equipment to obtain the accurate band structure matrix and / or accurate bandgap attribute characteristics of each candidate unit cell structure under the target magnetic field conditions. According to the design requirements of the target engineering system, the engineering indicators of each candidate unit cell structure are checked, and all candidate unit cell structures that simultaneously meet the bandgap performance threshold and engineering constraints are selected. Then, the terminal device is called to display these candidate unit cell structures. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall structure of the magnetic sensing deep learning-driven intelligent metamaterial on-demand design system disclosed in the embodiments of this application. Figure 2 This is a schematic diagram of the band structure prediction network and magnetic conditional attention mapping prediction network disclosed in the embodiments of this application. In the figure, a represents the magnetic conditional attention mapping prediction network and b represents the band structure prediction network. Figure 3 This is the structural gradient magnetosensory map disclosed in the embodiments of this application; Figure 4 This is a schematic diagram of the optimization of magnetic conditional attention mapping prediction network parameters disclosed in the embodiments of this application; Figure 5 This is a schematic diagram of the magnetic conditional generative adversarial network parameter optimization disclosed in the embodiments of this application; Figure 6 This is a diagram illustrating the modulatory effect of the square magnetic field on the band structure as disclosed in the embodiments of this application; Figure 7 This is a training error curve diagram disclosed in the embodiments of this application; Figure 8 The diagram in the middle shows the band structure of the unit cell structure designed on demand under different magnetic field strengths disclosed in the embodiments of this application, as well as the band structure diagram of its network prediction and simulation. Detailed Implementation

[0026] First, those skilled in the art should understand that these embodiments are merely used to explain the technical principles of the embodiments of this application and are not intended to limit the scope of protection of the embodiments of this application. Those skilled in the art can make adjustments as needed to adapt to specific application scenarios.

[0027] In the embodiments of this application, unless otherwise explicitly specified and limited, communication or communication connection between the first feature and the second feature means that there is information transmission between the first feature and the second feature. This information transmission can be unidirectional or bidirectional, and the communication connection can be achieved through electrical connection of wires, wireless communication, communication through electromagnetic media (such as semiconductors), communication through channels, etc. Furthermore, unless otherwise specified, the base of the logarithmic function used is 2.

[0028] In the embodiments of this application, unless otherwise explicitly specified and limited, "above," "below," "in front of," or "behind" the second feature can mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Furthermore, "above," "on top of," and "over" the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature. "Before," "in front of," and "in front of" the second feature can mean that the first feature is directly in front of or diagonally in front of the second feature, or simply indicates that the first feature precedes the second feature in sequence. "After," "behind," and "behind" the second feature can mean that the first feature is directly behind or diagonally behind the second feature, or simply indicates that the first feature is after the second feature in sequence.

[0029] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] See Figures 1 to 8 This embodiment discloses a magnetic sensing deep learning-driven intelligent metamaterial on-demand design system. Figure 1 A schematic diagram of the overall structure of the system, such as Figure 1 As shown, the system includes a terminal device, a unit cell structure device, a prediction device, and a simulation device. The unit cell structure device communicates with the terminal device, the prediction device communicates with the unit cell structure device, and the simulation device communicates with both the terminal device and the prediction device. The terminal device includes a display and an editor on the display, used for user input of target parameters including the target band structure matrix and the target magnetic field strength.

[0031] See Figure 1In this system, the unit cell structure device is configured to generate a magnetic field modulation spectrum with the same dimension as the target band structure matrix using the target parameters according to the Gaussian perturbation function rule. The magnetic field modulation spectrum is then transformed into multiple candidate unit cell structures using a model algorithm based on magnetic conditional generative adversarial networks. For example... Figure 1 As shown, in this embodiment, the unit cell structure device includes a modulation module and a magnetic conditional generative adversarial network, wherein the modulation module communicates with the terminal device, and the magnetic conditional generative adversarial network communicates with both the modulation module and the prediction device.

[0032] See Figure 2 and Figure 5 In the unit cell structure device, the modulation module is configured to normalize the target band structure matrix and substitute the target magnetic field strength and the normalized target band structure matrix into a Gaussian perturbation function to generate a magnetic field modulation spectrum with the same dimension as the target band structure matrix. In this embodiment, the Gaussian perturbation function is calculated as follows: , In the formula, These are the matrix elements of the normalized band structure matrix. For row indicators, List the indicators; It is a magnetic field modulation spectrum. The magnetic field strength, In this embodiment, the maximum magnetic field strength is... Take 150 mT, , and These are learnable magnetic field-related parameters.

[0033] See Figure 2 and Figure 5 In a single-cell device, a magnetic conditional generative adversarial network is configured to transform the magnetic field modulation spectrum into multiple candidate single-cell structures. For example... Figure 5 As shown, in this embodiment, the magnetic conditional generative adversarial network adopts a U-Net structure. A Transformer module is introduced at the bottleneck layer to capture long-range dependencies between different energy bands. The generator outputs an N×N unit cell structure diagram that has been activated by a sigmoid function and binarized with a 0.5 threshold. Specifically, in this embodiment, N is set to 50. See [link to documentation]. Figure 5The encoding end of the magnetic conditional generative adversarial network consists of multiple layers of convolution and pooling, with the pooling operation being a downsampling operation. After downsampling to a smaller spatial resolution, a multi-head self-attention Transformer module (referred to as the Transformer module in the figure) is embedded in the bottleneck layer. The decoding end restores the spatial resolution through upsampling and skip connections, outputting a 50×50 structure probability map activated by Sigmoid, thereby achieving 0.5 threshold binarization of the output probability map to obtain the final unit cell structure map.

[0034] See Figures 1-5 In this system, the prediction device is used to predict the band structure matrix and / or bandgap attribute characteristics of each candidate unit cell structure acquired by the unit cell structure device under the target magnetic field, and to screen out all candidate unit cell structures whose bandgap attribute characteristics deviate from the target bandgap by less than a first threshold. For example... Figure 1 As shown, in this embodiment, the prediction device includes a weighting module and a generation module. The weighting module communicates with the magnetic condition generative adversarial network, and the generation module communicates with both the weighting module and the simulation device.

[0035] See Figure 2 , Figure 3 and Figure 4 In the prediction device, the weighting module is used to calculate the magnetic field response weight map of the candidate unit cell structure. The candidate unit cell structure and the magnetic field response weight map are then concatenated along the channel dimension to form a dual-channel input tensor. For example... Figure 3 As shown, in this embodiment, the weighting module is configured to perform the following steps: A1: The Sobel operator is used to calculate the structure gradient of the candidate unit cell structure, obtaining the gradient magnitude map. Simultaneously, the structure map of the candidate unit cell structure is inverted to obtain the MRE region mask. For details, see [link to documentation]. Figure 3 For each unit cell structure diagram (i.e., the original structure), the Sobel operator is used to calculate the structure gradient, resulting in the following: Figure 3 The gradient magnitude diagram shown Inverting the structure diagram yields... Figure 3 The MRE region mask shown .

[0036] A2: The gradient magnitude map and the MRE region mask are weighted and summed to obtain the comprehensive sensitivity map. Specifically, the constructed comprehensive sensitivity map... for: , in, and ( + =1, taken in the example =0.6、 =0.4) is the weighting coefficient.

[0037] A3: Calculate the point-by-point product of the normalized magnetic field strength and the comprehensive sensitivity map to obtain the magnetic field response weight map. For details, see [link to documentation]. Figure 3 Using normalized magnetic field strength and comprehensive sensitivity map Calculate the magnetic field response weighted diagram for: , Accordingly, the constitutive relation of "selective modulation of MRE materials by magnetic field" is transformed into a trainable spatial prior, so that the network weights automatically satisfy ∇w· during backpropagation. B The implicit physical constraint ≡0 (the magnetic field gradient in the non-MRE material region is zero) reduces the traditional material-magnetic field coupling problem, which requires repeated finite element analysis, to a one-time feature masking operation. This operation is equivalent to suppressing or approximately zeroing the magnetic field contribution of the non-MRE material region at the feature level, so that the magnetic field mainly acts on the characterization of the MRE material region. On the one hand, it highlights the sensitivity of the MRE material region to magnetic field changes and reduces the interference of the magnetic field distribution in the metallic material region on bandgap prediction; on the other hand, it improves the ability of the subsequent magnetic condition attention mapping network to model the "magnetic field-structure" coupling characteristics, which is beneficial to improving the magnetic field utilization and bandgap control accuracy under given magnetic conditions.

[0038] A4: Concatenate the candidate unit cell structure with the magnetic field response weight map along the channel dimension to obtain a two-channel input tensor. Specifically, the unit cell structure map... With magnetic field response weighting diagram The input tensors are concatenated along the channel dimension to form a dual-channel input tensor of size 2×N×N, which serves as the input to the magnetic conditional attention mapping prediction network (referred to as the prediction network in the figure) to achieve explicit joint encoding of the magnetic field and structure.

[0039] See Figure 2 , Figure 4 and Figure 6 In the prediction device, the weight generation module is configured to first call its included magnetic conditional attention mapping prediction network (hereinafter referred to as the prediction network) to transform the dual-channel input tensor into the band structure matrix and / or band gap attribute features of the candidate unit cell structure under the target magnetic field. In this embodiment, the predicted value is the band structure matrix, such as... Figure 6 As shown; subsequently, all candidate unit cell structures whose bandgap attribute characteristics deviate from the target bandgap by less than a first threshold are selected. For example... Figure 2 and Figure 4As shown, the encoding end of the prediction network employs multi-layer convolution and pooling to extract multi-scale structure-magnetic field joint features; the intermediate feature layer introduces the CBAM attention mechanism. Its channel attention weight matrix is ​​hard-masked by the MRE mask before training, ensuring that the magnetic field channel weights are always zero at non-MRE pixel locations; spatial attention pre-enhances boundary pixels through the structure gradient map; the decoding end outputs predicted values ​​with the same dimension as the band structure matrix through a fully connected layer, thereby extracting bandgap attribute features such as the upper and lower bandgap limits.

[0040] In this embodiment, the prediction network comprises a 4-layer convolutional-pooling encoder, an intermediate feature module with channel-spatial attention, and a 2-layer fully connected decoder. The convolutional kernel size is 3×3 with a stride of 1. The network input is a 2×50×50 tensor, and the output is a 10×11 matrix with the same dimensions as the band structure matrix.

[0041] In this system, the simulation equipment is configured to operate as follows: During formal operation, finite element simulation is performed on each candidate unit cell structure selected by the prediction equipment to obtain the accurate band structure matrix and / or accurate bandgap attribute characteristics of each candidate unit cell structure under the target magnetic field conditions. Based on the design requirements of the target engineering system, the engineering indicators of each candidate unit cell structure are checked, and all candidate unit cell structures that simultaneously meet the bandgap performance threshold and engineering constraints are selected. Subsequently, the terminal device is invoked to display these candidate unit cell structures. During parameter optimization, finite element simulation is performed on each unit cell structure to obtain the accurate band structure matrix and / or accurate bandgap attribute characteristics of each unit cell structure under the target magnetic field conditions.

[0042] Throughout both the parameter optimization and formal operation phases, the simulation equipment performs finite element simulations on each unit cell structure to obtain the accurate band structure matrix and / or accurate bandgap properties of each unit cell structure under the target magnetic field conditions. The specific process is as follows: First, the generated unit cell structure is imported into the finite element simulation software COMSOL, and the corresponding material parameters such as density, elastic modulus, and Poisson's ratio are set. Among these, the elastic modulus of the MRE material... Determined by the following formula: , in, (Unit: mT) represents the applied magnetic field strength (in the example). Take 150mT), These are the coefficients obtained from the experimental fitting.

[0043] Subsequently, Bloch periodic boundary conditions were set, and several wave vector sampling points (11 in this example) were selected along the irreducible Brillouin zone path. Using the characteristic frequency interface of the solid mechanics module, the first few orders (10 in this example) of eigenfrequency were solved, i.e., the precise bandgap property characteristics, ultimately yielding... F × K The band structure matrix R is 10×11 dimensional.

[0044] Finally, the intrinsic frequencies are normalized, and the lower limit frequency, upper limit frequency, and bandwidth of the bandgap (BG) for each frequency band are extracted. Based on this, a bandgap target dataset for training can be constructed, yielding sample pairs {unit cell structure S, magnetic field}. B Let R be the band structure matrix and BG be the band gap.

[0045] See Figure 6 During the parameter optimization stage, the unit cell structure was imported into the finite element software COMSOL, and the density and Poisson's ratio of the MRE material were set to 3492.26 kg / m³. 3 And 0.47. Its elastic modulus Expressed as magnetic field strength B The cubic polynomial function has the following polynomial coefficients: [6.982 × 10⁻⁶]. -7 -1.092×10 -4 6.461×10 -3 , 2.765]. Magnetic field strength B The temperature was varied within the range of 0-150 mT, with a step size of 10 mT. The elastic modulus, density, and Poisson's ratio of the tungsten metallic material were set to 344 GPa, 19100 kg / m³, and 19100 kg / m³, respectively. 3 And 0.28. Set Bloch periodic boundary conditions, select 11 wave vector sampling points on the M–Γ–X–M path, solve for the first 10 eigenfrequency, obtain a 10×11 dimensional band structure matrix, and normalize the frequencies. Figure 6 (a) in the diagram is a unit cell structure diagram under magnetic field excitation. Figure 6 (b) in the figure represents the band structure matrix under different excitation magnetic fields.

[0046] See Figure 2 , Figure 4 , Figure 5 , Figure 6 , Figure 7 and Figure 8 The following will further disclose the usage method of the magnetic sensing deep learning-driven intelligent metamaterial on-demand design system described in this embodiment, namely, a magnetic sensing deep learning-driven intelligent metamaterial on-demand design method, which includes the following steps: S1: Unit cell structure construction: Using seed point-based morphological dilation operations on a computer device, multiple relatively concentrated and continuous unit cell structures of MRE materials are generated.

[0047] Specifically, the purpose of this step is to generate a binary distribution topology of MRE and metals (such as tungsten). Considering the magnetic field prior of "MRE magnetic drive variable stiffness characteristics", a morphological expansion operation based on seed points is used to generate a relatively concentrated and continuous unit cell structure of MRE material, laying a structural foundation for the magnetic sensing deep learning network to accurately learn the mapping relationship of "magnetic field-MRE variable stiffness-tunable bandgap".

[0048] In this embodiment, step S1 includes the following steps: S11: Seed point layout and initial topology generation: Encode the MRE material and the metal material with 0 and 1 respectively, and draw a regular shape on the pixel value matrix where all pixel values ​​are equal to the corresponding pixel values ​​of the metal material. Then, arrange multiple seed points on several sides of the regular shape; wherein, the regular shape is a right triangle, rectangle or rhombus.

[0049] Specifically, the MRE and metallic materials are first encoded with 0 and 1 respectively, and an N×N matrix (N is 50 in this example) is defined as the design domain, with 1 / 8 of its right-angled triangular subdomain taken as the actual design domain. Considering the prior magnetic field, a binary distribution unit cell initial topology is generated based on the symmetry axis anchoring 0-value seed points. Specifically, starting from the all-1-valued actual design domain, m values ​​are arranged on the hypotenuse and horizontal right-angled sides of the right-angled triangular subdomain. [3,7] seed points (m is 5 in the example).

[0050] S12: Starting from the seed point, generate an MRE material region on the pixel value matrix where the percentage of the encoded pixel value of the MRE material is within a preset range through multi-stage morphological dilation operations. In this embodiment, the preset range is 40%-70%.

[0051] S13: The unit cell structure diagram is generated by performing multiple orthogonal transformations on the MRE material region in the pixel value matrix. Specifically, the 1 / 8 actual design domain is rotated and mirrored using the P4m symmetry group to obtain the complete N×N unit cell structure diagram. See Figure 6 (a) in the middle.

[0052] S2: Finite element simulation is performed on each unit cell structure using simulation equipment to obtain the accurate band structure matrix and / or accurate bandgap properties of each unit cell structure under magnetic field conditions. In this embodiment, the band structure matrix obtained is under different excitation magnetic fields, such as... Figure 6 As shown in (b) of the diagram.

[0053] S3: Using the precise band structure and / or precise bandgap properties of the unit cell obtained in step S2 as the ideal output, the weighted sum of the mean absolute error and the mean relative error is used as the loss function to optimize the parameters of the prediction device.

[0054] See Figure 3 In this embodiment, step S3 includes the following steps: S31: The Sobel operator is used to calculate the structural gradient of each unit cell structure map to obtain their respective gradient magnitude maps, and the inversion operation is performed on each unit cell structure map to obtain their respective MRE region masks.

[0055] S32: Perform weighted summation of the gradient magnitude map and MRE region mask for each unit cell structure map to obtain their respective comprehensive sensitivity maps.

[0056] S33: Calculate the point-by-point product of the normalized magnetic field strength and the comprehensive sensitivity map of each unit cell structure map to obtain the respective magnetic field response weight map.

[0057] S34: Concatenate each unit cell structure diagram with its magnetic field response weight diagram according to the channel dimension to obtain their respective dual-channel input tensors.

[0058] S35: See also Figure 4 Using the precise band structure matrix and / or precise bandgap attribute features of the unit cell obtained in step S2 as the ideal output, the weighted sum of the mean absolute error and the mean relative error is used as the loss function, and the Adam optimizer is used to optimize the parameters of the magnetic conditional attention mapping prediction network in the prediction device. Figure 7 The prediction error curve of the magnetic conditional attention mapping prediction network is disclosed. The parameter optimization of the prediction device is considered to be completed when the prediction error of the magnetic conditional attention mapping prediction network is lower than a preset threshold. Figure 8 This paper demonstrates the band structure of on-demand designed unit cell structures under different magnetic field strengths, along with their magnetic condition attention mapping prediction network prediction and simulation. Figure 8 It is evident that the magnetic conditional attention mapping prediction network has a relatively accurate prediction effect.

[0059] S4: See also Figure 5 Using the multiple unit cell structures obtained in step S1, the parameters of the unit cell structure device are optimized by using the weighted sum of the adversarial loss function, topological connectivity loss function, and band matching loss function as the total loss function.

[0060] See Figure 5 In this embodiment, step S4 includes the following steps; S41: Normalize the exact band structure matrix of each unit cell under magnetic field conditions, and substitute the magnetic field strength and the normalized band structure matrix set in the magnetic field conditions into the Gaussian perturbation function to generate the magnetic field modulation spectrum of each unit cell with the same dimension as the band structure matrix.

[0061] S42: Input the magnetic field modulation spectrum of each unit cell structure into the magnetic condition generative adversarial network in the unit cell structure device to generate their respective predicted unit cell structure diagrams.

[0062] S43: Input each unit cell structure and its predicted unit cell structure map into the PatchGAN discriminator network based on spectral normalization to obtain their respective local structure authenticity score matrices.

[0063] S44: Using the weighted sum of the adversarial loss function, topological connectivity loss function, and bandgap matching loss function as the total loss function, substitute the local structure authenticity score matrix of each unit cell structure into the total loss function, and train the PatchGAN discriminator network and the magnetic conditional generative adversarial network using an alternating optimization method until the connectivity and bandgap matching accuracy of the predicted unit cell structure graph output by the magnetic conditional generative adversarial network meet the preset requirements. This is considered as completing the parameter optimization of the unit cell structure device.

[0064] In the parameter optimization process, this embodiment constructs a magnetic conditional generative adversarial network as the generator and a PatchGAN discriminator network as the discriminator. Figure 2 The band structure prediction network is shown in b. Subsequently, the total generator loss function is the weighted sum of the adversarial loss function, topological connectivity loss function, and band matching loss function: , in, , and The weighting coefficients for the three types of losses are as follows (in the example). ); Adversarial loss function The generated structure distribution is encouraged to approximate the true structure distribution; topological connectivity loss function Used to penalize structures that are not 4-connected or contain isolated pixels; band matching loss function The generated structure's bandgap is quickly predicted by calling a pre-trained prediction network and then compared with the target bandgap.

[0065] S5: The user inputs target parameters, including the target band structure matrix and the target magnetic field strength, through the terminal device. Then, by running the unit cell structure device, prediction device and simulation device in sequence, all candidate unit cell structures that simultaneously meet the band gap performance threshold and engineering constraints are selected.

[0066] Specifically, step S5 includes the following steps: S51: The user inputs target parameters, including the target band structure matrix and the target magnetic field strength, on the terminal device.

[0067] S52: Using a single-cell structure device and following the Gaussian perturbation function rule, a magnetic field modulation spectrum with the same dimension as the target band structure matrix is ​​generated using the target parameters. The magnetic field modulation spectrum is then transformed into multiple candidate single-cell structures using a model algorithm of a magnetic conditional generative adversarial network.

[0068] S53: The band structure matrix and / or band gap attribute features of each candidate unit cell structure obtained by the unit cell structure device under the target magnetic field are predicted by the prediction device, and all candidate unit cell structures whose band gap attribute features deviate from the target band gap by less than a first threshold are selected.

[0069] S54: The simulation equipment performs finite element simulation verification on each candidate unit cell structure selected by the prediction equipment to obtain the accurate band structure matrix and / or accurate bandgap attribute characteristics of each candidate unit cell structure under the target magnetic field conditions. According to the design requirements of the target engineering system, the engineering indicators of each candidate unit cell structure are checked, and all candidate unit cell structures that simultaneously meet the bandgap performance threshold and engineering constraints are selected. Then, the terminal equipment is called to display these candidate unit cell structures.

[0070] Specifically, finite element simulation is performed on the candidate structures selected through the bandgap deviation screening in step S53 to verify their accurate band structure and bandgap properties under the target magnetic field conditions, and the deviation between them and the target bandgap is calculated. Simultaneously, based on the design requirements of the target engineering system (e.g., vehicle suspension or panel), engineering indicators such as equivalent stiffness, equivalent damping, overall thickness, mounting interface dimensions, and reserved magnetic circuit layout space of the candidate unit cell structures are checked to select those that simultaneously meet the bandgap performance threshold and engineering constraints. When the deviation between the simulated bandgap and the target bandgap is less than the threshold and the corresponding structure meets the aforementioned connectivity, manufacturability, and engineering constraints, the structure is output as the final design result. If further optimization is needed, the final design result can be added to the dataset as a new sample to update or incrementally train the prediction network and the on-demand design network, achieving a closed-loop iteration of design-verification-retraining for practical engineering scenarios.

[0071] A metamaterial bandgap adaptive control system with real-time vibration sensing. This system deploys an accelerometer at the vibration source and acquires vibration signals online using a data acquisition card. The controller performs STFT analysis to extract the dominant vibration frequency. Based on a preset "frequency-optimal current" mapping strategy, it determines the optimal driving current for vibration damping under the current operating conditions. This current is then output through a programmable current source or power amplifier module to drive an external electromagnetic coil to generate a corresponding magnetic field. This, in turn, adjusts the stiffness of the metamaterial's metamaterial MRE in real time, causing its bandgap to adaptively shift to near the dominant frequency, thus achieving intelligent vibration reduction for time-varying vibrations under dynamic environments.

[0072] The system disclosed in this embodiment achieves rapid and accurate prediction of band structure and on-demand design of unit cell structures for smart metamaterials under given excitation field (magnetic field strength) conditions through explicit magnetic field-structure coupling modeling. By using structural gradient magnetic sensing encoding, the magnetic field is jointly encoded with the structural topology in the form of a spatial weighted graph, significantly improving the coupling modeling capability between the magnetic field and structure, and enabling rapid and accurate prediction of the band diagram and band gap of the unit cell structure under any given magnetic field. By constructing a frequency-magnetic field joint conditional tensor through Gaussian spectrum perturbation, and introducing a Transformer module into the generator to form a magnetic conditional generative adversarial network, on-demand design of unit cell structures that satisfy the expected band structure under any magnetic field strength is achieved. Furthermore, this magnetic field sensing encoding and deep learning generative framework is not only applicable to MRE elastic metamaterials, but can also be extended to piezoelectric metamaterials, acoustic metamaterials, and other multiphysics tunable systems. This method provides a data-driven solution for the tunable band gap design of various smart materials and structures.

[0073] Meanwhile, referring to 7 and 8, the prediction network constructed in this embodiment can reduce the bandgap prediction time of a single structure under a given magnetic field from minutes in traditional finite element simulation to seconds or even milliseconds while ensuring a small average relative error (e.g., 2%), significantly improving the efficiency of design iteration. This technology can support the evaluation of a large number of candidate structures within a reasonable time, thereby effectively supporting the design optimization of engineering systems such as vehicle suspensions and vibration isolation components, and providing support for rapid engineering decision-making.

[0074] In the description of this application, the references to terms such as "an embodiment," "some embodiments," "in this embodiment," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0075] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A magnetic sensing deep learning-driven intelligent metamaterial on-demand design system, characterized in that, Including those connected in series to form a closed loop: Terminal equipment, used to allow users to input target parameters including the target band structure matrix and the target magnetic field strength; The single-cell structure device is configured to generate a magnetic field modulation spectrum with the same dimension as the target band structure matrix using the target parameters according to the Gaussian perturbation function rule, and to transform the magnetic field modulation spectrum into multiple candidate single-cell structures through a model algorithm of magnetic conditional generative adversarial network. A prediction device is used to predict the band structure matrix and / or band gap attribute features of each candidate unit cell structure obtained by the unit cell structure device under the target magnetic field, and to screen out all candidate unit cell structures whose band gap attribute features deviate from the target band gap by less than a first threshold. The simulation equipment is configured to operate as follows: During formal operation, finite element simulation is performed on each candidate unit cell structure screened by the prediction device to obtain the accurate band structure matrix and / or accurate bandgap attribute characteristics of each candidate unit cell structure under the target magnetic field conditions. According to the design requirements of the target engineering system, the engineering indicators of each candidate unit cell structure are checked, and all candidate unit cell structures that simultaneously meet the bandgap performance threshold and engineering constraints are screened out. Then, the terminal device is called to display these candidate unit cell structures. During parameter optimization, finite element simulation is performed on each unit cell structure to obtain the accurate band structure matrix and / or accurate bandgap property characteristics of each unit cell structure under the target magnetic field conditions.

2. The magnetic sensing deep learning-driven intelligent metamaterial on-demand design system according to claim 1, characterized in that, The unit cell structure device includes: The modulation module, which communicates with the terminal device, is configured to normalize the target band structure matrix and substitute the target magnetic field strength and the normalized target band structure matrix into a Gaussian perturbation function to generate a magnetic field modulation spectrum with the same dimension as the target band structure matrix. A magnetic conditional generative adversarial network, which communicates with both the modulation module and the prediction device, is configured to convert the magnetic field modulation spectrum into multiple candidate unit cell structures.

3. The magnetic sensing deep learning-driven intelligent metamaterial on-demand design system according to claim 2, characterized in that, The Gaussian perturbation function is calculated as follows: ， In the formula, These are the matrix elements of the normalized band structure matrix. For row indicators, List the indicators; It is a magnetic field modulation spectrum. The magnetic field strength, For the maximum magnetic field strength, , and These are learnable magnetic field-related parameters.

4. The magnetic sensing deep learning-driven intelligent metamaterial on-demand design system according to claim 2 or 3, characterized in that, The prediction device includes: The weighting module communicates with the magnetic condition generative adversarial network and is used to calculate the magnetic field response weight map of the candidate unit cell structure. The candidate unit cell structure and the magnetic field response weight map are spliced ​​together according to the channel dimension to form a dual-channel input tensor. The generation module, which communicates with both the weighting module and the simulation device, is configured to first call its included magnetic conditional attention mapping prediction network to convert the dual-channel input tensor into the band structure matrix and / or bandgap attribute features of the candidate unit cell structure under the target magnetic field, and then filter out all candidate unit cell structures whose bandgap attribute features deviate from the target bandgap by less than a first threshold.

5. The magnetic sensing deep learning-driven intelligent metamaterial on-demand design system according to claim 4, characterized in that, The weighting module is configured to perform the following steps: A1: The Sobel operator is used to calculate the structural gradient of the candidate unit cell structure to obtain the gradient magnitude map. At the same time, the structural map of the candidate unit cell structure is inverted to obtain the MRE region mask. A2: The gradient magnitude map and the MRE region mask are weighted and summed to obtain the comprehensive sensitivity map; A3: Calculate the point-by-point product of the normalized magnetic field strength and the comprehensive sensitivity map to obtain the magnetic field response weight map; A4: Concatenate the candidate unit cell structure with the magnetic field response weight map along the channel dimension to obtain the dual-channel input tensor.

6. A magnetic sensing deep learning-driven intelligent metamaterial on-demand design method, characterized in that, The magnetic sensing deep learning-driven intelligent metamaterial on-demand design system according to any one of claims 1-5 includes the following steps: S1: Using seed point-based morphological dilation operations on a computer device, multiple relatively concentrated and continuous unit cell structures of MRE materials are generated; S2: Perform finite element simulation on each unit cell structure using simulation equipment to obtain the accurate band structure matrix and / or accurate bandgap property characteristics of each unit cell structure under magnetic field conditions; S3: Using the precise band structure and / or precise bandgap properties of the unit cell obtained in step S2 as the ideal output, the weighted sum of the mean absolute error and the mean relative error is used as the loss function to optimize the parameters of the prediction device. S4: Using the multiple unit cell structures obtained in step S1, the parameters of the unit cell structure device are optimized by using the weighted sum of the adversarial loss function, topological connectivity loss function and band matching loss function as the total loss function. S5: The user inputs target parameters, including the target band structure matrix and the target magnetic field strength, through the terminal device. Then, by running the unit cell structure device, prediction device and simulation device in sequence, all candidate unit cell structures that simultaneously meet the band gap performance threshold and engineering constraints are selected.

7. The magnetic sensing deep learning-driven intelligent metamaterial on-demand design method according to claim 6, characterized in that, Step S1 includes the following steps: S11: Encode the MRE material and the metal material with 0 and 1 respectively, and draw a regular shape on the pixel value matrix where all pixel values ​​are equal to the corresponding pixel values ​​of the metal material. Then, arrange multiple seed points on several sides of the regular shape; wherein, the regular shape is a right triangle, rectangle or rhombus. S12: Starting from the seed point, generate an MRE material region on the pixel value matrix where the proportion of the corresponding pixel value of the MRE material encoding on the pixel value matrix is ​​within a preset range through multi-stage morphological dilation operation. S13: The unit cell structure diagram is generated by performing multiple orthogonal transformations on the MRE material region in the pixel value matrix.

8. The magnetic sensing deep learning-driven intelligent metamaterial on-demand design method according to claim 6 or 7, characterized in that, Step S3 includes the following steps: S31: The Sobel operator is used to calculate the structural gradient of each unit cell structure map to obtain their respective gradient magnitude maps, and the inversion operation is performed on each unit cell structure map to obtain their respective MRE region masks. S32: Perform weighted summation on the gradient magnitude map and MRE region mask of each unit cell structure map to obtain their respective comprehensive sensitivity maps; S33: Calculate the point-by-point product of the normalized magnetic field strength and the comprehensive sensitivity map of each unit cell structure map to obtain the respective magnetic field response weight map; S34: Concatenate each unit cell structure diagram with the magnetic field response weight diagram of that unit cell structure diagram according to the channel dimension to obtain their respective dual-channel input tensors; S35: Using the precise band structure matrix and / or precise bandgap attribute features of the unit cell structure obtained in step S2 as the ideal output, and using the weighted sum of the mean absolute error and the mean relative error as the loss function, the parameters of the magnetic conditional attention mapping prediction network in the prediction device are optimized until the prediction error of the magnetic conditional attention mapping prediction network is lower than a preset threshold, at which point the parameter optimization of the prediction device is considered complete.

9. The magnetic sensing deep learning-driven intelligent metamaterial on-demand design method according to claim 8, characterized in that, Step S4 includes the following steps; S41: Normalize the exact band structure matrix of each unit cell structure under magnetic field conditions, and substitute the magnetic field strength and normalized band structure matrix set in the magnetic field conditions into the Gaussian perturbation function to generate the magnetic field modulation spectrum of each structure with the same dimension as the band structure matrix. S42: Input the magnetic field modulation spectrum of each unit cell structure into the magnetic condition generative adversarial network in the unit cell structure device to generate their respective predicted unit cell structure diagrams. S43: Input each unit cell structure and its predicted unit cell structure map into the PatchGAN discriminator network based on spectral normalization to obtain their respective local structure authenticity score matrices. S44: Using the weighted sum of the adversarial loss function, topological connectivity loss function, and bandgap matching loss function as the total loss function, substitute the local structure authenticity score matrix of each unit cell structure into the total loss function, and train the PatchGAN discriminator network and the magnetic conditional generative adversarial network using an alternating optimization method until the connectivity and bandgap matching accuracy of the predicted unit cell structure graph output by the magnetic conditional generative adversarial network meet the preset requirements. This is considered as completing the parameter optimization of the unit cell structure device.

10. The magnetic sensing deep learning-driven intelligent metamaterial on-demand design method according to claim 9, characterized in that, Step S5 includes the following steps: S51: The user inputs target parameters, including the target band structure matrix and the target magnetic field strength, on the terminal device; S52: Using a single-cell structure device according to the Gaussian perturbation function rule, a magnetic field modulation spectrum with the same dimension as the target band structure matrix is ​​generated using the target parameters. The magnetic field modulation spectrum is then transformed into multiple candidate single-cell structures using a model algorithm of a magnetic conditional generative adversarial network. S53: The band structure matrix and / or band gap attribute characteristics of each candidate unit cell structure obtained by the unit cell structure device under the target magnetic field are predicted by the prediction device, and all candidate unit cell structures whose band gap attribute characteristics deviate from the target band gap by less than a first threshold are selected. S54: The simulation equipment performs finite element simulation verification on each candidate unit cell structure selected by the prediction equipment to obtain the accurate band structure matrix and / or accurate bandgap attribute characteristics of each candidate unit cell structure under the target magnetic field conditions. According to the design requirements of the target engineering system, the engineering indicators of each candidate unit cell structure are checked, and all candidate unit cell structures that simultaneously meet the bandgap performance threshold and engineering constraints are selected. Then, the terminal device is called to display these candidate unit cell structures.

Images