A metasurface dynamic inverse generation method based on a liquid generation architecture

Abstract

The application discloses a metasurface dynamic inverse generation method based on a liquid generation architecture, first obtains target electromagnetic response data to be designed and pre-processes, obtains target response condition input, and obtains basic condition features through a convolutional neural network condition encoder; the basic condition features are input into a continuous time dynamic liquid network to generate residual compensation features; the basic condition features and the residual compensation features are fused to obtain enhanced condition features; the enhanced condition features are input into a condition diffusion model to generate a candidate metasurface structure; finally, the predicted electromagnetic response corresponding to the candidate metasurface structure is calculated, and the inverse generated metasurface structure is output according to the error evaluation result between the predicted electromagnetic response and the target electromagnetic response. The method performs state-dependent dynamic residual compensation on the condition representation of the target electromagnetic response, and generates a candidate metasurface coding structure through a condition generation network, thereby improving the response matching precision of the global inverse design of the metasurface.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and electromagnetic metasurface technology, specifically a dynamic reverse generation method for metasurfaces based on a liquid generation architecture. Background Technology

[0002] Metasurfaces can flexibly control the response of electromagnetic waves, such as amplitude, phase, polarization, scattering direction, and far-field pattern, by adjusting the geometric parameters, material parameters, arrangement, or encoding state of the unit structure. They play an important role in applications such as beamforming, intelligent reflector communication, radar cross section reduction, electromagnetic stealth, and terahertz devices.

[0003] Traditional metasurface inverse design typically relies on parameter scanning, full-wave simulation, and iterative manual experience. When the design target is a complex far-field pattern, spatial energy distribution, or multi-point scattering response, the spatial dimension of the structural parameters is high, and there is a strong nonlinear mapping relationship between the target response and the structural parameters, resulting in high computational cost and slow convergence speed for conventional optimization methods.

[0004] In recent years, deep learning technology has been gradually introduced into the reverse engineering of metasurfaces. Existing solutions typically employ architectures such as convolutional neural networks (CNNs), multilayer perceptrons (MLPs), autoencoders (AEs), conditional diffusion models, and flow-based generative models to learn the nonlinear mapping relationship between the target electromagnetic response and structural parameters, and generate candidate metasurface structures under the constraints of the target response. These methods have improved design efficiency to some extent.

[0005] However, existing deep learning-based metasurface reverse design methods still generally suffer from the following technical problems:

[0006] Conditional encoding mechanisms lack dynamic state modeling capabilities. Existing models typically employ CNNs or MLPs to perform a one-time static feature mapping on the target's electromagnetic response. This means that once conditional encoding is complete, the extracted conditional features remain unchanged throughout the generation process. Because the target's far-field response itself contains information with complex spatial distributions, such as main lobes, side lobes, nulls, and energy decay, the state dependencies and dynamic correlations between different target responses and between different spatial locations of the same response cannot be fully expressed by static encoding. This results in insufficient descriptive power of conditional representations for the target response.

[0007] Both approaches fall short in representing the global correlation features in the target response. CNN encoders can extract local spatial texture features, but their ability to model the global dependencies between the main lobe and side lobes, and between side lobes, as well as the spatial energy variation trend in the far-field distribution is limited. While MLP encoders can establish fully connected nonlinear mappings, they flatten the two-dimensional spatial input into a one-dimensional vector, ignoring the spatial structural correlation of the target response. Neither approach can fully represent the global state information in complex target responses, thus limiting the matching accuracy between the generated structure and the target response.

[0008] The lack of an adaptive compensation mechanism to adapt to differences in target response states is a significant issue. Existing models use the same set of static conditional coding parameters to process different types of electromagnetic response targets (such as single-beam, multi-beam, wide-beam, or specific null distributions), lacking a mechanism for adaptive dynamic compensation based on differences in input response states. This leads to a substantial deviation between the forward physical response and the design target when generating candidate structures under complex and varied response conditions, a problem particularly pronounced in applications requiring high-precision beam pointing and sidelobe suppression.

[0009] In summary, how to introduce an adaptive compensation mechanism with dynamic state modeling capabilities into the conditional coding stage without changing the existing generative network backbone structure, so as to improve the physical response matching accuracy of metasurface reverse design results, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and propose a dynamic reverse generation method for metasurfaces based on a liquid generation architecture. This method utilizes an NCP-CfC network to perform state-dependent dynamic residual compensation for the conditional characterization of the target electromagnetic response, and generates candidate metasurface encoding structures through a conditional generation network, thereby improving the response matching accuracy of global reverse design of metasurfaces.

[0011] To achieve the above objectives, the technical solution specifically adopted by the present invention is as follows:

[0012] A dynamic inverse generation method for metasurfaces based on a liquid generation architecture includes the following steps:

[0013] Step 1: Obtain the target electromagnetic response data to be designed, and preprocess the target electromagnetic response data to obtain the target response condition input;

[0014] Step 2: Using the target response conditions as input, extract spatial features of the target electromagnetic response through a convolutional neural network conditional encoder, and obtain basic conditional features through pooling.

[0015] Step 3: Input the basic condition features into the continuous-time dynamic liquid network. The continuous-time dynamic liquid network updates its state based on the current input features and the previous hidden state to obtain the current liquid hidden state, and generates residual compensation features based on the current liquid hidden state.

[0016] Step 4: Fuse the basic condition features with the residual compensation features to obtain the enhanced condition features;

[0017] Step 5: Input the enhanced conditional features into the conditional diffusion model to generate candidate metasurface structures;

[0018] Step 6: Calculate the predicted electromagnetic response corresponding to the candidate metasurface structure, and output the reverse-generated metasurface structure based on the error evaluation result between the predicted electromagnetic response and the target electromagnetic response.

[0019] Preferably, the continuous-time dynamic liquid network includes an NCP sparse connection network and a CfC liquid dynamic network; the connection relationship within the CfC liquid dynamic network is constrained by the NCP sparse connection network, and the current liquid hidden state is generated based on the current input features and the previous hidden state through the CfC liquid dynamic network.

[0020] Preferably, the CfC liquid dynamic network determines two candidate states based on the current input features and the previous hidden state, and then performs weighted fusion of the two candidate states using a dynamic gating coefficient related to the state update interval to obtain the current liquid hidden state.

[0021] Preferably, when fusing the basic condition features and residual compensation features, a gated residual method is used for fusion: the residual compensation features are superimposed onto the basic condition features through learnable residual gate coefficients to obtain the enhanced condition features.

[0022] Preferably, the conditional diffusion model uses U-Net as the denoising backbone network, and under the constraint of enhanced conditional features, it gradually recovers and generates candidate metasurface structures from structural states with added noise of different intensities.

[0023] Preferably, in step 2, the main lobe position, side lobe distribution, spatial energy change and local texture features of the target electromagnetic response are extracted by the convolutional neural network conditional encoder, and the extracted features are mapped to the conditional feature dimension that matches the subsequent conditional diffusion model.

[0024] Preferably, the preprocessing method involves sequentially performing logarithmic compression, percentile truncation, and normalization on the target electromagnetic response data.

[0025] Preferably, the predicted electromagnetic response is obtained through physical analytical calculation, fast Fourier transform approximation, forward prediction network, or full-wave simulation interface, and the error evaluation includes root mean square error, correlation coefficient, and relative error.

[0026] Preferably, the metasurface structure includes a metasurface, a reflective array, a smart reflective surface, a frequency selective surface, and an electromagnetic functional structure in an antenna array. The target electromagnetic response includes a far-field pattern, a scattering response, a beamforming response, an amplitude distribution, and a phase distribution.

[0027] Preferably, the metasurface structure is a binary encoded metasurface, the target electromagnetic response is the target far-field amplitude distribution, and the candidate metasurface structure is binarized to form a two-dimensional binary encoded matrix.

[0028] Preferably, the NCP sparse connection network and the CfC liquid dynamics network constitute an NCP-CfC liquid dynamics unit. The NCP-CfC liquid dynamics unit has 256 units and is equipped with a hybrid memory module. The hybrid memory module is used to retain historical condition information and global correlation information in the far-field feature sequence. When the hybrid memory module is enabled, the memory state of the hybrid memory module is initialized to a zero vector in the first feature processing step.

[0029] Preferably, before performing reverse generation, a training sample pair is formed using the target electromagnetic response data and the corresponding metasurface structure data. The conditional diffusion model is trained by adding noise of different intensities to the real metasurface structure, and the basic conditional coding module and the continuous-time dynamic liquid network are trained together.

[0030] This invention also provides a dynamic global inverse generation system for metasurfaces based on a liquid generation architecture, comprising:

[0031] The basic condition encoding module is used to receive target response condition input, extract target response features, and output basic condition features.

[0032] A continuous-time dynamic liquid network, connected to the basic condition encoding module, is used to receive the target response features, perform state dependency updates on the input features and hidden states through the liquid dynamic unit, and output residual compensation features.

[0033] The feature fusion module is connected to the basic condition encoding module and the continuous-time dynamic liquid network respectively, and is used to fuse the basic condition features with the residual compensation features to output enhanced condition features;

[0034] A conditional diffusion model, connected to the feature fusion module, is used to generate candidate metasurface structures under the constraints of the enhanced conditional features.

[0035] The post-processing and screening module, connected to the conditional diffusion model, is used to discretize or binarize the candidate metasurface structures, calculate the predicted electromagnetic response corresponding to each candidate metasurface structure, and screen and output the final reverse design results based on the error evaluation results between each predicted electromagnetic response and the target electromagnetic response.

[0036] Preferably, the NCP-CfC liquid dynamic unit is composed of an NCP sparse connection network and a CfC liquid dynamic network. The number of units in the NCP-CfC liquid dynamic unit is 256, and a hybrid memory module is set. The hybrid memory module is used to retain historical condition information and global correlation information in the far-field feature sequence. When the hybrid memory module is enabled, the memory state of the hybrid memory module is initialized to a zero vector in the first feature processing step.

[0037] Preferably, the feature fusion module uses a gated residual method for feature fusion: the residual compensation features are weighted by learnable residual gate coefficients, and the weighted residual compensation features are added to the basic condition features to output the enhanced condition features.

[0038] Preferably, the conditional diffusion model uses U-Net as the denoising backbone network, and under the constraint of the enhanced conditional features, it gradually recovers and generates candidate metasurface structures from structural states with added noise of different intensities.

[0039] Preferably, the basic conditional encoding module is a convolutional neural network conditional encoder, which is used to extract the main lobe position, side lobe distribution, spatial energy change and local texture features of the target electromagnetic response, and map the extracted features to a conditional feature dimension that matches the conditional diffusion model.

[0040] This invention has the following characteristics and beneficial effects:

[0041] This invention adds a dynamic liquid network residual compensation module after the basic conditional encoding module. This module uses a continuous-time liquid network to perform state-dependent updates on the input features and hidden states. Because the liquid network has the ability to dynamically adjust its internal response characteristics according to changes in the input state, its output residual compensation features can reflect the state differences between different target electromagnetic responses, thus providing additional state-sensitive descriptive information on top of the basic static encoding. By fusing the basic conditional features with this state-dependent residual compensation feature, the resulting enhanced conditional features more fully express information such as the main lobe position, sidelobe distribution, null depth, and spatial energy variation trend of the target response, providing more accurate conditional constraints for subsequent network generation.

[0042] In dynamic liquid networks, liquid units update their hidden states sequentially according to the feature sequence. The processing result at the current feature position is influenced by the information of the previous feature positions, thus forming a memory and transmission mechanism for the correlation between features in the temporal (or spatial sequence) dimension. When this mechanism is applied to far-field response feature sequences, the model can further capture the global dependencies between the main lobe and side lobes, and between sampling points at different angles, while preserving local spatial features. This has direct technical significance for design scenarios that require global coordinated control, such as multi-beam, wide-beam, or complex null distributions in the target response, overcoming the problem of loss of global correlation information caused by conventional CNN encoders that can only extract local receptive field features.

[0043] This invention integrates a dynamic liquid network as a residual compensation module between the basic condition encoding network and the condition generation network, and adaptively fuses the basic condition features and dynamic compensation features through a gated residual method. This structural design has the following advantages:

[0044] The dynamic liquid network only makes supplementary corrections to the basic coding results, rather than replacing the original coding path. Therefore, even if the output of the liquid network module is poor in some edge cases, the basic condition features can still independently support the generation process and ensure the basic performance lower limit of the system.

[0045] The basic condition coding module and the condition generation network backbone can use existing mature architectures. In actual deployment, only a dynamic liquid network residual compensation module needs to be embedded between the two. There is no need to redesign the entire generation network, which facilitates the upgrading and modular replacement of existing systems.

[0046] The dynamic liquid network introduces state-dependent time mapping parameters and a continuous-time dynamic gating mechanism when updating hidden states, which has a certain filtering and inertial smoothing effect on changes in input features. When there is measurement noise, numerical discretization error, or small perturbations introduced by the preprocessing stage in the target electromagnetic response data, the dynamic evolution process of the liquid network can, to a certain extent, suppress the influence of sudden input fluctuations on the expression of conditional features, making the final generated enhanced conditional features more stable, thereby improving the output consistency of the subsequent generated structure under different data quality conditions.

[0047] This invention, after generating candidate metasurface structures, further discretizes them, calculates their forward responses, and then selects the best output based on the error evaluation results between the predicted and target responses. This mechanism ensures that the final output reverse design results are directly verified at the forward physical response level, rather than relying solely on implicit constraints within the generator network. By generating multiple candidate structures for the same target response conditions and selecting the best one, the risk of a single generation result deviating from the design objective can be effectively reduced, ensuring that the output metasurface encoded structure has a quantifiable response matching guarantee.

[0048] Because the dynamic liquid network residual compensation module can generate differentiated compensation features for different target response input states, the method of this invention is adaptable to various design target types, including single-beam, multi-beam, wide-beam, far-field distributions with specific null constraints, and scattering energy modulation. Compared to generative models that rely solely on static conditional encoding, this invention can automatically adjust the conditional feature expression through dynamic state-dependent updates when facing design tasks with a large range of response morphology variations and complex spatial distributions, thereby broadening the applicability of deep learning inverse design methods in metasurface engineering applications. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a dynamic global reverse design method for metasurfaces based on a liquid generation architecture, according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the overall model structure in this embodiment of the invention, consisting of CNN basic conditional encoding, NCP-CfC dynamic liquid network residual compensation, gated residual fusion, and conditional diffusion U-Net.

[0051] Figure 3 This is a schematic diagram of the structure of the dynamic liquid network residual compensation module, which consists of NCP sparse connection constraints, CfC dynamic gating state updates, and gating residual fusion, in an embodiment of the present invention.

[0052] Figure 4 This is a main comparison chart of the multilayer perceptron conditional generation model, convolutional conditional generation model, and liquid generation architecture model in the reverse engineering performance evaluation metrics of embodiments of the present invention. Detailed Implementation

[0053] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0054] Example 1

[0055] This embodiment provides a dynamic reverse generation method for metasurfaces based on a liquid generation architecture, such as... Figure 1 As shown, it includes the following steps:

[0056] Step S1: Obtain the target electromagnetic response data to be designed, and normalize the target electromagnetic response data to obtain the target response condition input.

[0057] Specifically, in this embodiment, the metasurface structure is a binary encoded metasurface, the target electromagnetic response is the target far-field amplitude distribution, and the output structure is a two-dimensional binary encoded matrix. To improve the numerical stability of the training process, the target far-field response is sequentially subjected to logarithmic compression, percentile truncation, and normalization. Let the amplitude at the i-th sampling point of the target far-field response be... Its preprocessing process can be expressed as:

[0058]

[0059]

[0060]

[0061] in, This represents the logarithmically compressed response value. This represents the response value after percentile truncation. and These represent the 2nd and 98th percentiles of the logarithmic value of the target's far-field response in the training set, respectively. For numerically stable terms, This represents the normalized target far-field condition input.

[0062] Step S2: Input the target response conditions into the convolutional neural network conditional encoder to extract basic conditional features.

[0063] Specifically, the convolutional neural network conditional encoder extracts the main lobe position, side lobe distribution, spatial energy variation, and local texture features of the target in the far field, and maps the extracted features to a conditional feature dimension that matches the conditional diffusion U-Net backbone.

[0064] Step S3: Input the basic condition features into the continuous-time dynamic liquid network, update the state based on the current input features and the previous hidden state to obtain the current liquid hidden state, and generate residual compensation features based on the current liquid hidden state.

[0065] Specifically, such as Figure 2 and Figure 3 As shown, the continuous-time dynamic liquid network includes an input mapping layer, an NCP-CfC liquid dynamic unit, a gated residual mapping layer, and an output mapping layer.

[0066] In this embodiment, the NCP-CfC liquid dynamic unit is composed of an NCP sparse connection network and a CfC liquid dynamic network. The number of units in the NCP-CfC liquid dynamic unit is 256, and a hybrid memory module is set. The hybrid memory module is used to retain historical condition information and global correlation information in the far-field feature sequence. When the hybrid memory module is enabled, the memory state of the hybrid memory module is initialized to a zero vector in the first feature processing step.

[0067] This process can be represented as:

[0068] In this embodiment, the basic conditional features are obtained by a convolutional neural network conditional encoder extracting features from the target response conditional input. The convolutional neural network conditional encoder processes the basic conditional features step by step; in the first feature processing step, the previous hidden state is initialized as a zero vector, and in subsequent feature processing steps, the previous hidden state is the liquid hidden state obtained in the previous feature processing step.

[0069]

[0070]

[0071]

[0072] in, This represents the dynamic gating coefficient of the CfC liquid dynamic network at the t-th feature position; and This represents the time mapping parameters determined by the current input features and the previous hidden state; Indicates the update interval between adjacent feature states; and This represents two candidate states generated under the constraints of the NCP sparse connection network. This indicates the updated hidden state of the liquid; This represents the residual compensation feature obtained from the liquid hidden state mapping; Represents the learnable residual gating coefficients; and These represent the basic conditional features and the enhanced conditional features of the CNN, respectively. In this embodiment, t represents the position index in the far-field feature sequence, rather than the electromagnetic wave physical time or the diffusion sampling time step. In the formula, ⊙ represents element-wise multiplication, i.e., the Hadamard product.

[0073] Step S4: Fuse the basic condition features with the residual compensation features to obtain the enhanced condition features.

[0074] Specifically, this embodiment uses a gated residual method to fuse the basic condition features of CNN and the residual compensation features output by the NCP-CfC dynamic liquid network, thereby introducing state dependency compensation information of the liquid dynamic unit while retaining the spatial coding capability of CNN.

[0075] Step S5: Input the enhanced conditional features into the conditional diffusion model to generate one or more candidate metasurface structures.

[0076] Specifically, in this embodiment, the conditional diffusion model uses a U-Net denoising backbone. Referring to the training method of the conditional diffusion model, Gaussian noise of varying intensities is added to real electromagnetic functional structure samples, and the network is trained to recover the real structure under enhanced conditional feature constraints. This process can be represented as:

[0077]

[0078]

[0079] in, This represents a sample of a real electromagnetic functional structure. This represents the structural state after noise is added; n represents Gaussian noise that follows a standard normal distribution. Indicates noise intensity; This represents a conditional diffusion U-Net denoising network with parameter θ. This indicates the characteristics of liquid residual enhancement conditions; Represents the loss weight related to noise intensity; This represents the training loss of the conditional diffusion model.

[0080] The training of the convolutional neural network conditional encoder, continuous-time dynamic liquid network, and conditional diffusion model described above was completed using paired training samples. Each training sample includes the target electromagnetic response and its corresponding real electromagnetic functional structure. During training, Gaussian noise of different intensities was added to the real structure, and the denoising process was constrained by the enhanced conditional features of the target electromagnetic response. The model parameters were updated by minimizing the difference between the denoised structure and the real structure. After the model training is completed, the actual reverse engineering stage only requires the input of the target electromagnetic response to be designed, without prior knowledge of its corresponding real structure.

[0081] Step S6: Discretize or binarize one or more candidate metasurface structures, calculate the predicted electromagnetic response corresponding to each candidate metasurface structure, and filter and output the final reverse generation result based on the error evaluation results between the predicted electromagnetic response and the target electromagnetic response.

[0082] Specifically, the continuous value matrix output by the conditional diffusion generation module is converted into a binary encoding matrix, and the forward far-field calculation is performed on the binary encoding matrix to obtain the predicted electromagnetic response corresponding to the candidate metasurface structure.

[0083] In this embodiment, the continuous value matrix is ​​converted into a binary encoded matrix in the following manner:

[0084]

[0085] in, This represents the continuous value in the m-th row and n-th column of the output matrix of the conditional diffusion generation module. This represents the binary encoding result at the corresponding position.

[0086] The binary encoding results are mapped to the metasurface element response, and the predicted electromagnetic response is obtained based on the array far-field calculation:

[0087]

[0088] in, This represents the element response of the metasurface element in the m-th row and n-th column, determined by the binary encoded state. and This represents the position coordinates of the corresponding metasurface unit, where k represents the wavenumber. and The observed direction angle is indicated; the predicted electromagnetic response is normalized according to the maximum amplitude.

[0089] Multiple candidate metasurface structures are screened based on the error evaluation results between the predicted and target electromagnetic responses, and the final reverse generation result is output. This embodiment uses root mean square error (RMSE), correlation coefficient (Corr), relative error (RelErr), intersection-over-union ratio (IoU), and accuracy (Acc) to evaluate the reverse generation result. Specifically, RMSE measures the magnitude error between the predicted and target responses; a smaller RMSE value indicates higher matching accuracy. Corr measures the linear correlation between the two response distributions; a value closer to 1 indicates more consistent distribution shapes. Relative error (RelErr) measures the overall L2 error of the predicted response relative to the target response; a smaller RelErr value indicates lower relative deviation. IoU measures the overlap between the generated structure and the reference structure in the effective cell region; a larger IoU value indicates higher structural region consistency. Accuracy (Acc) measures the proportion of coding units in the generated structure that match the corresponding coding units in the reference structure; a larger IoU value indicates a higher degree of coding matching. Among multiple candidate metasurface structures, the candidate metasurface structure with smaller root mean square error and higher correlation of response distribution is preferentially selected as the final reverse generation result.

[0090] Furthermore, it should be noted that before performing reverse generation, the target electromagnetic response data and the corresponding metasurface structure data are used to form a training sample pair. The conditional diffusion model is trained by adding noise of different intensities to the real metasurface structure, and the basic conditional encoding module and the continuous-time dynamic liquid network are trained together.

[0091] In this embodiment, the liquid generation architecture model is formed by combining CNN conditional coding, NCP-CfC dynamic liquid network, gated residual fusion, and hybrid memory module. The NCP-CfC dynamic liquid network has 256 liquid units, and the hybrid memory module is enabled. This structure introduces NCP-CfC dynamic liquid network residual compensation on top of convolutional conditional coding to enhance the state-dependent representation capability of the target electromagnetic response conditions.

[0092] Specifically, the hybrid memory module can be implemented using a memory-gated unit. For the t-th position in the far-field feature sequence, the memory-gated unit generates an updated memory state based on the current input features and the memory state of the previous position, and then sends this memory state to the CfC dynamic liquid unit to participate in the hidden state update. In this way, the model not only utilizes the current local features, but also retains the conditional information in the preceding feature positions, thereby enhancing its ability to express the global correlation of complex far-field responses.

[0093] To verify the effectiveness of this invention, this embodiment sets up a main comparison of three models: Multilayer Perceptron Conditional Generator (MLPCondural Generator, MCG), Convolutional Conditional Generator (CNN Conditional Generator, CCG), and Liquid Generative Architecture (LGA). All three models use the same training set, validation set, test set, and evaluation metrics. Under the condition of CFG=3.0 and 100 test samples, 8 candidate structures are generated for each objective. For RMSE, Corr, and RelErr, the candidate structure with the smallest RMSE for each objective is selected, and the corresponding metrics are calculated. For IoU and Acc, the candidate structure with the largest IoU for each objective is selected, and the corresponding metrics are calculated. Then, the arithmetic mean is calculated for each of the 100 objectives. The results are shown in Table 1.

[0094] Table 1 shows the evaluation comparison data of the three types of models.

[0095]

[0096] As shown in Table 1, the multilayer perceptron conditional generation model suffers from high far-field matching error due to its lack of two-dimensional spatial feature modeling capability; the convolutional conditional generation model significantly improves reverse engineering performance through convolutional spatial encoding; the liquid generation architecture model of this invention introduces NCP-CfC dynamic liquid network residual compensation on the basis of convolutional spatial encoding, further achieving the lowest RMSE, highest Corr, lowest RelErr, and higher IoU and Acc. Specifically, as... Figure 4 As shown, compared to the convolutional conditional generative model (CCG), the liquid generative architecture model (LGA) reduced RMSE from 0.0749 to 0.0590, a reduction of approximately 21.2%; Corr increased from 0.9431 to 0.9643, an improvement of approximately 2.2%; RelErr decreased from 0.1640 to 0.1294, a reduction of approximately 21.1%; IoU increased from 0.6610 to 0.6879, an improvement of approximately 4.1%; and Acc increased from 0.7922 to 0.8092, an improvement of approximately 2.1%. Compared to the multilayer perceptron conditional generative model (MCG), the liquid generative architecture model (LGA) reduced RMSE by approximately 40.3%, RelErr by approximately 40.1%, Corr by approximately 7.2%, IoU by approximately 11.1%, and Acc by approximately 6.7%. This demonstrates that the present invention can effectively improve the accuracy of metasurface reverse engineering.

[0097] Example 2

[0098] This embodiment presents a dynamic global inverse design system for metasurfaces based on a liquid generation architecture, comprising:

[0099] The basic conditional encoding module is used to receive target response conditional input, extract target response features, and output basic conditional features. The basic conditional encoding module is a convolutional neural network conditional encoder, which is used to extract the main lobe position, side lobe distribution, spatial energy change and local texture features of the target electromagnetic response, and map the extracted features to the conditional feature dimension that matches the electromagnetic functional structure generation network.

[0100] A continuous-time dynamic liquid network, connected to the basic condition coding module, is used to receive the target response features, perform state-dependent updates on the input features and hidden states through the NCP-CfC liquid dynamic unit, and output residual compensation features.

[0101] The NCP sparse connection network and the CfC liquid dynamics network constitute the NCP-CfC liquid dynamics unit. The NCP-CfC liquid dynamics unit has 256 units and is equipped with a hybrid memory module. The hybrid memory module is used to retain historical condition information and global correlation information in the far-field feature sequence. When the hybrid memory module is enabled, the memory state of the hybrid memory module is initialized to a zero vector in the first feature processing step.

[0102] The feature fusion module is connected to the basic condition encoding module and the continuous-time dynamic liquid network respectively, and is used to fuse the basic condition features with the residual compensation features to output enhanced condition features;

[0103] The feature fusion module uses a gated residual method for feature fusion: the residual compensation features are weighted by learnable residual gate coefficients, and the weighted residual compensation features are added to the basic condition features to output the enhanced condition features.

[0104] A conditional diffusion model, connected to the feature fusion module, is used to generate candidate metasurface structures under the constraints of the enhanced conditional features.

[0105] The conditional diffusion model uses U-Net as the denoising backbone network and gradually recovers candidate supersurface structures from structural states with added noise of different intensities under the constraints of the enhanced conditional features.

[0106] The post-processing and screening module, connected to the conditional diffusion generation module, is used to discretize or binarize the candidate metasurface structures, calculate the predicted electromagnetic response corresponding to each candidate metasurface structure, and screen and output the final reverse design results based on the error evaluation results between each predicted electromagnetic response and the target electromagnetic response.

[0107] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for dynamic reverse generation of metasurfaces based on a liquid-generating architecture, characterized in that, Includes the following steps: Step 1: Obtain the target electromagnetic response data to be designed, and preprocess the target electromagnetic response data to obtain the target response condition input; Step 2: Using the target response conditions as input, extract spatial features of the target electromagnetic response through a convolutional neural network conditional encoder, and obtain basic conditional features through pooling. Step 3: Input the basic condition features into the continuous-time dynamic liquid network. The continuous-time dynamic liquid network updates its state based on the current input features and the previous hidden state to obtain the current liquid hidden state, and generates residual compensation features based on the current liquid hidden state. Step 4: Fuse the basic condition features with the residual compensation features to obtain the enhanced condition features; Step 5: Input the enhanced conditional features into the conditional diffusion model to generate candidate metasurface structures; Step 6: Calculate the predicted electromagnetic response corresponding to the candidate metasurface structure, and output the reverse-generated metasurface structure based on the error evaluation result between the predicted electromagnetic response and the target electromagnetic response.

2. The method according to claim 1, characterized in that, The continuous-time dynamic liquid network includes an NCP sparse connection network and a CfC liquid dynamic network. The connection relationships within the CfC liquid dynamic network are constrained by the NCP sparse connection network, and the current liquid hidden state is generated based on the current input features and the previous hidden state through the CfC liquid dynamic network.

3. The method according to claim 1, characterized in that, The CfC liquid dynamic network determines two candidate states based on the current input features and the previous hidden state, and then performs weighted fusion of the two candidate states using a dynamic gating coefficient related to the state update interval to obtain the current liquid hidden state.

4. The method according to claim 1, characterized in that, When fusing the basic condition features and residual compensation features, a gated residual method is used for fusion: the residual compensation features are superimposed on the basic condition features through learnable residual gate coefficients to obtain the enhanced condition features.

5. The method according to claim 1, characterized in that, The conditional diffusion model uses U-Net as the denoising backbone network and, under the constraint of enhanced conditional features, gradually recovers and generates candidate metasurface structures from structural states with added noise of different intensities.

6. The method according to claim 1, characterized in that, In step 2, the main lobe position, side lobe distribution, spatial energy change and local texture features of the target electromagnetic response are extracted by the convolutional neural network conditional encoder, and the extracted features are mapped to the conditional feature dimension that matches the subsequent conditional diffusion model.

7. The method according to any one of claims 1 to 6, characterized in that, The preprocessing method involves sequentially performing logarithmic compression, percentile truncation, and normalization on the target electromagnetic response data.

8. The method according to claim 1, characterized in that, The predicted electromagnetic response is obtained through physical analytical calculation, fast Fourier transform approximation, forward prediction network or full-wave simulation interface, and the error evaluation includes root mean square error, correlation coefficient and relative error.

9. The method according to claim 1, characterized in that, The metasurface structure includes a metasurface, a reflective array, a smart reflective surface, a frequency selective surface, and electromagnetic functional structures in an antenna array. The target electromagnetic response includes a far-field pattern, a scattering response, a beamforming response, an amplitude distribution, and a phase distribution.

10. The method according to claim 1, characterized in that, The metasurface structure is a binary encoded metasurface, the target electromagnetic response is the target far-field amplitude distribution, and the candidate metasurface structure is binarized to form a two-dimensional binary encoded matrix.

11. The method according to claim 1, characterized in that, The NCP sparse connection network and the CfC liquid dynamics network constitute the NCP-CfC liquid dynamics unit. The NCP-CfC liquid dynamics unit has 256 units and is equipped with a hybrid memory module. The hybrid memory module is used to retain historical condition information and global correlation information in the far-field feature sequence. When the hybrid memory module is enabled, the memory state of the hybrid memory module is initialized to a zero vector in the first feature processing step.

12. The method according to claim 1, characterized in that, Before performing reverse generation, the target electromagnetic response data and the corresponding metasurface structure data are used to form training sample pairs. The conditional diffusion model is trained by adding noise of different intensities to the real metasurface structure, and the basic conditional coding module and the continuous-time dynamic liquid network are trained together.

13. A dynamic global reverse design system for metasurfaces based on a liquid-generating architecture, characterized in that, include: The basic condition encoding module is used to receive target response condition input, extract target response features, and output basic condition features. A continuous-time dynamic liquid network, connected to the basic condition coding module, is used to receive the target response features, perform state-dependent updates on the input features and hidden states through the NCP-CfC liquid dynamic unit, and output residual compensation features. The feature fusion module is connected to the basic condition encoding module and the continuous-time dynamic liquid network respectively, and is used to fuse the basic condition features with the residual compensation features to output enhanced condition features; A conditional diffusion model, connected to the feature fusion module, is used to generate candidate metasurface structures under the constraints of the enhanced conditional features. The post-processing and screening module, connected to the conditional diffusion model, is used to discretize or binarize the candidate metasurface structures, calculate the predicted electromagnetic response corresponding to each candidate metasurface structure, and screen and output the final reverse design results based on the error evaluation results between each predicted electromagnetic response and the target electromagnetic response.

14. The system according to claim 13, characterized in that, The NCP-CfC liquid dynamic unit is composed of an NCP sparse connection network and a CfC liquid dynamic network. The number of units in the NCP-CfC liquid dynamic unit is 256, and a hybrid memory module is set up. The hybrid memory module is used to retain historical condition information and global correlation information in the far-field feature sequence. When the hybrid memory module is enabled, the memory state of the hybrid memory module is initialized to a zero vector in the first feature processing step.

15. The system according to claim 13, characterized in that, The feature fusion module uses a gated residual method for feature fusion: the residual compensation features are weighted by learnable residual gate coefficients, and the weighted residual compensation features are added to the basic condition features to output the enhanced condition features.

16. The system according to claim 13, characterized in that, The conditional diffusion model uses U-Net as the denoising backbone network and, under the constraint of the enhanced conditional features, gradually recovers and generates candidate metasurface structures from structural states with added noise of different intensities.

17. The system according to claim 13, characterized in that, The basic conditional encoding module is a convolutional neural network conditional encoder, which is used to extract the main lobe position, side lobe distribution, spatial energy change and local texture features of the target electromagnetic response, and map the extracted features to a conditional feature dimension that matches the conditional diffusion model.

Images