High-temperature superconducting band-pass filter design method, system, storage medium and device

Abstract

The application discloses a high-temperature superconducting band-pass filter design method, system, storage medium and equipment. The method first determines design indexes including a center frequency, a relative bandwidth and the like, and an extensible structure framework formed by cascading multiple coupling lines and one more branch lines; then, equivalent circuit models of differential mode and common mode of the filter are respectively established according to a transmission line theory; subsequently, a multilayer perceptron type circuit level physical information neural network model with residual connection is constructed, the design indexes are input into the model, and equivalent circuit matrices and a target differential mode transmission function are embedded into a loss function to complete training; the coupling line even mode, odd mode characteristic impedance and the branch line characteristic impedance are output by the trained model, and a filter physical layout is generated according to the characteristic impedances. The application fuses a physical information neural network, can quickly generate a superconducting filter layout, and effectively improves design efficiency and prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of microwave devices and circuit technology, specifically to a design method, system, storage medium, and device for a high-temperature superconducting bandpass filter. Background Technology

[0002] Balanced bandpass filters are widely used in communication, radar, and satellite systems because they can suppress common-mode interference and improve system anti-interference capabilities. As systems evolve towards wider bandwidth and higher sensitivity, broadband balanced bandpass filters, while achieving good differential-mode passband response, also need to achieve high levels of common-mode rejection over a wide frequency range.

[0003] To reduce insertion loss and improve out-of-band rejection, high-temperature superconducting materials have been introduced into filter design. Broadband balanced high-temperature superconducting bandpass filters typically employ complex transmission line networks composed of multi-stage coupling lines and branch lines. As the order increases, the number of structural parameters increases significantly, making the design extremely dependent on full-wave electromagnetic simulation and manual parameter tuning. This results in long design cycles and makes rapid engineering difficult.

[0004] In existing technologies, commonly used data-driven surrogate models include artificial neural networks, support vector machines, and Gaussian process regression, which are used to replace some electromagnetic simulations. However, such models usually do not show the physical laws of introducing transmission lines and electromagnetic fields, and belong to "black box" regression. They have limited generalization ability and are prone to producing non-physical or distorted predictions when training data is insufficient or design parameters exceed the training range. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a design method, system, storage medium and device for high-temperature superconducting bandpass filters, and to solve the above-mentioned problems described in the prior art.

[0006] A first aspect of the present invention is to provide a design method for a high-temperature superconducting bandpass filter, the method comprising:

[0007] The design specifications and scalable structural framework of the filter are determined. The design specifications include center frequency, relative bandwidth, filter order, differential mode return loss and common mode rejection. The structural framework consists of multiple cascaded coupling lines and one additional branch line.

[0008] Based on transmission line theory, differential-mode equivalent circuit models and common-mode equivalent circuit models of the filter structure are constructed respectively.

[0009] Construct a circuit-level physical information neural network model using a multilayer perceptron architecture, including setting multiple hidden layers and configuring a corresponding number of neurons, while adding residual connections;

[0010] The design metrics are input into the circuit-level physical information neural network model, and the matrix derived from the equivalent circuit and the target differential mode transfer function are embedded into the model's loss function to complete model training.

[0011] Using the trained circuit-level physical information neural network model, the even-mode characteristic impedance, odd-mode characteristic impedance, and characteristic impedance of the branch lines of the coupled line are output.

[0012] The physical layout of the filter is generated based on the various characteristic impedances of the output.

[0013] According to one aspect of the above technical solution, the target differential mode transfer function adopts an equirippled Chebyshev form, specifically implemented as follows:

[0014] Determine the passband ripple level and cutoff frequency of the Chebyshev response, and calculate the corresponding electrical length;

[0015] Construct a Chebyshev polynomial function of the first kind of degree;

[0016] The target differential transfer function is generated based on the Chebyshev polynomial function, and the loss function of the circuit-level physical information neural network model is embedded in it.

[0017] According to one aspect of the above technical solution, the loss function of the circuit-level physical information neural network model includes a frequency-weighted loss term, and the specific implementation of the frequency-weighted loss term includes:

[0018] Identify the passband frequency range and out-of-band frequency range of the filter;

[0019] A high weighting coefficient is set for frequency points within the passband frequency range, and a low weighting coefficient is set for frequency points within the out-of-band frequency range.

[0020] The corresponding weighting coefficients are multiplied by the scattering parameter prediction error to obtain the frequency-weighted loss term, which is then embedded into the total loss function of the circuit-level physical information neural network model.

[0021] According to one aspect of the above technical solution, the circuit-level physical information neural network model satisfies the passive boundary condition, specifically implemented as follows:

[0022] The condition information of the passive boundary condition is invoked, and the amplitude constraint that the scattering parameter must satisfy is determined based on the condition information;

[0023] The passive constraint is transformed into a penalty term and embedded into the loss function of the circuit-level physical information neural network model;

[0024] During the training process of the circuit-level physical information neural network model, a penalty is imposed on the prediction results that violate the passivity constraint in order to achieve boundary condition constraints.

[0025] According to one aspect of the above technical solution, the method further includes:

[0026] Based on the physical layout, a layout-level physical information neural network model is constructed, using a one-dimensional convolutional autoencoder model, with the encoder being a convolutional neural network and the decoder being a multi-layer fully connected network.

[0027] A hybrid loss function is constructed for the layout-level physical information neural network model. The hybrid loss function includes the mean square error terms of the actual geometric parameters and the predicted geometric parameters, the mean square error terms of the actual scattering parameters and the reconstructed scattering parameters under frequency weighting, and physical constraint terms of passivity and port matching are added.

[0028] Obtain an initial dataset, train the layout-level physical information neural network model, and establish a mapping relationship between layout geometric parameters and differential and common-mode scattering parameters through the layout-level physical information neural network model;

[0029] Using the decoder of the aforementioned map-level physical information neural network model as the electromagnetic proxy model, the number of search proxies and the maximum number of iterations of the artificial lemming optimization algorithm are set to complete the algorithm initialization;

[0030] The artificial lemming optimization algorithm is used to perform global search and local refinement of the layout geometry parameters, and combined with an automatic data incremental training strategy to continuously optimize the layout-level physical information neural network model, outputting the optimal layout geometry parameters that meet the design specifications.

[0031] According to one aspect of the above technical solution, the steps of using the artificial lemming optimization algorithm to perform global search and local refinement of the layout geometric parameters, and continuously optimizing the layout-level physical information neural network model in conjunction with an automatic data incremental training strategy, to output the optimal layout geometric parameters that meet the design specifications, include:

[0032] The loss function of the artificial lemming optimization algorithm is constructed based on the return loss level, insertion loss level of the differential mode response, and insertion loss level of the common mode response.

[0033] Set the number of search agents and the maximum number of iterations for the artificial lemming optimization algorithm, and initialize the position parameters of each agent;

[0034] Using the decoder of the layout-level physical information neural network model, the scattering parameters corresponding to each agent position are predicted, and the corresponding loss function values ​​are calculated.

[0035] Update the agent position based on the loss function value, perform a global search, refine the parameters near the optimal solution locally, and output the optimal layout geometric parameters.

[0036] According to one aspect of the above technical solution, the step of predicting the scattering parameters corresponding to each agent position and calculating the corresponding loss function value using the decoder of the layout-level physical information neural network model includes:

[0037] The layout geometry parameters corresponding to each agent position in the artificial lemming optimization algorithm are input into the encoder of the trained layout-level physical information neural network model to extract the deep feature vector of the layout geometry parameters.

[0038] The deep feature vector is input into the decoder, and the corresponding differential mode scattering parameters and common mode scattering parameters are reconstructed through upsampling and dimensionality transformation of a multi-layer fully connected network.

[0039] Extract the return loss level and insertion loss level corresponding to the differential mode scattering parameters, and the insertion loss level corresponding to the common mode scattering parameters to obtain the level parameters;

[0040] The level parameters are substituted into the pre-constructed loss function of the artificial lemming optimization algorithm to calculate the loss function value corresponding to each agent position.

[0041] A second aspect of the present invention is to provide a high-temperature superconducting bandpass filter design system, applied to the method described in the above-mentioned technical solution, the system comprising:

[0042] The device configuration module is used to determine the design specifications and scalable structural framework of the filter. The design specifications include center frequency, relative bandwidth, filter order, differential mode return loss and common mode rejection. The structural framework consists of multiple cascaded coupling lines and one additional branch line.

[0043] The first construction module is used to construct the differential-mode equivalent circuit model and the common-mode equivalent circuit model of the filter structure based on transmission line theory.

[0044] The second building module is used to build a circuit-level physical information neural network model with a multilayer perceptron architecture, including setting multiple hidden layers and configuring a corresponding number of neurons, while adding residual connections.

[0045] The model training module is used to input the design metrics into the circuit-level physical information neural network model, embed the matrix derived from the equivalent circuit and the target differential mode transfer function into the model's loss function, and complete the model training.

[0046] The impedance classification module is used to output the even-mode characteristic impedance, odd-mode characteristic impedance, and characteristic impedance of the branch lines of the coupled lines using the trained circuit-level physical information neural network model.

[0047] The layout generation module is used to generate the physical layout of the filter based on the various characteristic impedances of the output.

[0048] A third aspect of the present invention is to provide a readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method described in the above-described technical solution.

[0049] A fourth aspect of the present invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described in the above technical solutions.

[0050] Compared with existing technologies, the advantages of using the high-temperature superconducting bandpass filter design method, system, storage medium, and device shown in this invention are as follows:

[0051] This invention effectively addresses the technical pain points of existing broadband balanced high-temperature superconducting bandpass filters, such as long design cycles, reliance on manual parameter tuning, high engineering difficulty, and weak generalization ability and susceptibility to non-physical predictions by traditional data-driven surrogate models. By defining a scalable structural framework and constructing differential-mode and common-mode equivalent circuit models based on transmission line theory, physical laws are embedded into a circuit-level physical information neural network. This breaks through the limitations of traditional "black-box" regression, significantly improving the model's generalization ability. Even when training data is insufficient or design parameters exceed the training range, it can still output prediction results that conform to physical laws. Simultaneously, the model can directly output the key characteristic impedances of coupling lines and branch lines, quickly generating the filter's physical layout. This significantly reduces the workload of full-wave electromagnetic simulation and manual parameter tuning, shortens the design cycle, reduces engineering difficulty, and balances the filter's core performance characteristics such as differential-mode passband response, common-mode rejection, and low insertion loss, adapting to the broadband and high-sensitivity development needs of communication, radar, and satellite systems. Attached Figure Description

[0052] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0053] Figure 1 This is a flowchart illustrating the high-temperature superconducting bandpass filter design method proposed in this embodiment of the invention.

[0054] Figure 2 This is the differential-mode equivalent circuit of the broadband balanced high-temperature superconducting bandpass filter proposed in the embodiments of the present invention;

[0055] Figure 3This is the common-mode equivalent circuit of the broadband balanced high-temperature superconducting bandpass filter proposed in the embodiments of the present invention;

[0056] Figure 4 This is the circuit-level physical information neural network (PINN-1) architecture proposed in the embodiments of the present invention;

[0057] Figure 5 This is the layout-level physical information neural network (PINN-2) architecture proposed in this embodiment of the invention;

[0058] Figure 6 This is a flowchart of the integrated design method of the broadband balanced high-temperature superconducting bandpass filter proposed in this embodiment of the invention, which combines artificial lemming algorithm optimization and a two-stage PINN model.

[0059] Figure 7 This is a diagram of the topology of a second-order filter in an embodiment of the present invention;

[0060] Figure 8 This is a three-dimensional topology diagram of a second-order filter in an embodiment of the present invention;

[0061] Figure 9 This is a topology diagram of the second-order filter in an embodiment of the present invention;

[0062] Figure 10 These are the PINN-1 predicted response and ideal Chebyshev response results of the second-order filter in this embodiment of the invention;

[0063] Figure 11 These are the prediction and simulation diagrams of the second-order filter in this embodiment of the invention;

[0064] Figure 12 This is a structural block diagram of the high-temperature superconducting bandpass filter design system proposed in an embodiment of the present invention. Detailed Implementation

[0065] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0066] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0067] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0068] Example 1

[0069] Please see Figure 1 The first embodiment of the present invention provides a design method for a high-temperature superconducting bandpass filter, the method comprising steps S10-S60:

[0070] Step S10: Determine the design specifications and scalable structural framework of the filter. The design specifications include center frequency, relative bandwidth, filter order, differential mode return loss and common mode rejection. The structural framework consists of multiple cascaded coupling lines and one additional branch line.

[0071] It should be noted that the design specifications clearly define the core performance requirements of the filter. The center frequency determines the operating frequency band of the filter, the relative bandwidth relates to the passband range, the filter order affects the out-of-band suppression effect, the differential mode return loss measures the degree of signal reflection, and the common mode rejection capability determines its anti-interference level.

[0072] In this embodiment, the scalable structural framework adopts a cascaded form of multiple coupling lines and one additional branch line. This structure has the characteristics of flexible adjustment and can adapt to different design performance requirements without changing the overall topology, providing a unified structural foundation for subsequent equivalent modeling and neural network training.

[0073] Step S20: Based on transmission line theory, construct the differential-mode equivalent circuit model and the common-mode equivalent circuit model of the filter structure, respectively.

[0074] It should be noted that the differential-mode signal is the useful signal that the filter needs to transmit effectively, while the common-mode signal is the interference signal that needs to be suppressed. Constructing equivalent circuit models for the two modes of signals respectively can accurately describe the transmission characteristics of different signals in the filter structure.

[0075] In this embodiment, the establishment of two types of equivalent circuit models, the differential-mode equivalent circuit model and the common-mode equivalent circuit model, can transform complex electromagnetic problems into circuit analysis problems, providing a physical basis for the loss function of the subsequent circuit-level physical information neural network model.

[0076] Step S30: Construct a circuit-level physical information neural network model using a multilayer perceptron architecture, including setting multiple hidden layers and configuring a corresponding number of neurons, while adding residual connections.

[0077] Among them, the multilayer perceptron architecture possesses powerful nonlinear fitting capabilities. By setting multiple hidden layers and configuring a corresponding number of neurons, it can enhance the model's ability to learn complex mapping relationships. Adding residual connections can effectively solve the gradient vanishing problem during the training process of deep neural networks, ensuring the training efficiency and accuracy of the model, and enabling the model to stably learn the mapping law from design indicators to characteristic impedance.

[0078] Step S40: Input the design parameters into the circuit-level physical information neural network model, embed the matrix derived from the equivalent circuit and the target differential mode transfer function into the model's loss function, and complete the model training.

[0079] In this embodiment, the matrix derived from the equivalent circuit contains information on the transmission and reflection characteristics of the filter, and the target differential mode transfer function clarifies the ideal performance target of the filter. Embedding both into the loss function allows the neural network model to follow the laws of electromagnetic physics during training, avoiding the "black box" defect of traditional data-driven models, and making the characteristic impedance of the model output physically reasonable.

[0080] Step S50: Using the trained circuit-level physical information neural network model, output the even-mode characteristic impedance, odd-mode characteristic impedance, and characteristic impedance of the branch lines of the coupled lines.

[0081] In this embodiment, even-mode and odd-mode characteristic impedances are key electrical parameters of the coupled line, directly affecting its coupling degree, while the characteristic impedance of the branch line determines its signal transmission capability. The aforementioned circuit-level physical information neural network model directly outputs characteristic impedance parameters that meet performance requirements based on the input design specifications, eliminating the tedious analytical calculations and manual parameter tuning processes of traditional methods, thus significantly improving design efficiency.

[0082] Step S60: Based on the various characteristic impedances of the output, generate the physical layout of the filter.

[0083] It should be noted that there is a clear correspondence between the characteristic impedance and the physical structure dimensions of the filter. Based on the output characteristic impedance parameters and the transmission line impedance calculation formula, the physical dimensions such as the width and spacing of the coupling lines and the width of the branch lines can be directly determined, thereby generating the physical layout of the filter.

[0084] As a preferred example of the present invention:

[0085] like Figure 2 The figure shows the differential-mode equivalent circuit of the broadband balanced high-temperature superconducting bandpass filter proposed in this invention. According to transmission line theory, the first... i Branch lines and coupling lines ABCD A matrix can be represented as:

[0086]

[0087]

[0088] in

[0089]

[0090]

[0091]

[0092]

[0093] total ABCD The matrix can be calculated by multiplying it sequentially by the submatrices of all branch lines and coupling lines:

[0094]

[0095] in, f Da ( q） , f Db ( q） , f Dc ( q） , f Dd ( q） The number of times n A polynomial function. When N When it is an odd number:

[0096]

[0097] When N is even:

[0098]

[0099] in, a Dn , b Dn , c Dn , d Dn All are constants, and are respectively determined by the branches ( Z b1 ,…, Z bN+1 ) and coupling line ( Z e1 , Z o1 ,…, Z eN, Z oN） The characteristic impedance is expressed as follows. Therefore, the transfer function of the DM equivalent circuit can be expressed as:

[0100]

[0101] in, F DM The filter function of the DM equivalent circuit can be derived as follows:

[0102]

[0103] in, X n is F DM The coefficients of the filter function are determined by the characteristic impedances of the branch lines and the coupled lines. When N When the number is odd, the filter function F DM It only includes the coefficients and cosine values ​​of even-numbered terms. Similarly, when N When the number is even, it contains only the coefficients and cosine values ​​of the odd-numbered terms.

[0104] To achieve the characteristics of an equiripple filter, this invention selects an ideal Chebyshev response transfer function as the objective function. By ensuring that the filtering function of the bandpass filter and the ideal Chebyshev objective function have the same response, the characteristic impedance values ​​of each transmission line required in the bandpass filter can be calculated. The Chebyshev transfer function is:

[0105]

[0106] in, , L A It is a fixed ripple level within the passband. F Obj The Chebyshev filter function can be expressed as:

[0107]

[0108] and

[0109]

[0110] in, θ c It is a lower cutoff frequency f c The electrical length below. T 3N+1 and T 3N They are 3N+1 times and 3 N Chebyshev polynomials of the first kind, of degree 1.

[0111] To implement the DM transfer function S DM 21 and the objective function S The matching of Obj 21 is evaluated using the following optimization function:

[0112]

[0113] in

[0114]

[0115] and, f i ( i =1,2,…, n） is the sampling frequency of the passband of the DM equivalent circuit, and n is the number of sampling points in the passband.

[0116] like Figure 3 The diagram shows the common-mode equivalent circuit of the broadband balanced high-temperature superconducting bandpass filter proposed in this invention. Similarly, based on the scattering matrix and ABCD The transformation relationships between matrices can be used to derive the common-mode equivalent circuit. S parameter:

[0117]

[0118] The evaluation function for the common-mode response can be defined as:

[0119]

[0120] in, G CM This is the common-mode rejection level, typically set to 20 dB.

[0121] like Figures 4-5 The figure shows the two-stage PINN model architecture of the broadband balanced high-temperature superconducting bandpass filter proposed in this invention.

[0122] Circuit-level physical information neural network (PINN-1) architecture as follows Figure 4 As shown. Its network architecture includes: a multilayer perceptron (MLP) architecture with three hidden layers, each containing 128 neurons, and residual connections added to improve learning efficiency. The input is the filter design metric (center frequency). f 0. Relative bandwidth FBW, order N Differential mode return loss RL DM Common-mode suppression G CM The output is the characteristic impedance of the coupled line and the branch line.

[0123] Its loss function model includes: the network is trained using a composite loss function, which is also enforced. The loss function is defined as calculated using the predicted impedance through an equivalent circuit model. S Parameters and target Chebyshev response S The mean square error (MSE) between the parameters is represented by the evaluation functions for CM and DM described above:

[0124]

[0125] Furthermore, a physical residual term is added to the loss function. The physical constraint term consists of the residuals of the ABCD matrix, and is expressed as follows:

[0126]

[0127] Therefore, the total loss function of PINN-1 can be expressed as:

[0128]

[0129] in, These are the weighting coefficients.

[0130] Its training strategy includes: creating a large-scale, noise-free training dataset based on Chebyshev response theory of bandpass filters; uniformly and randomly sampling the design parameters within a reasonable range; calculating the target label of each sample in real time using an analytical Chebyshev polynomial approximation formula; and setting the output length to 3×N+3 (N is set to a maximum of 5).

[0131] PINN-2 architecture as Figure 5 As shown, this architecture employs a 1D convolutional autoencoder (1D-CAE) network. The initial layer uses 3×1 convolutional kernels to expand the input channels to 64 dimensions, and combines batch normalization, LeakyReLU, and Dropout mechanisms to gradually increase the channel dimension to 128. After each module, a max-pooling operation halves the sequence length, followed by adaptive max-pooling to compress the features into a uniform 8-dimensional representation. Finally, a flattening layer and a fully connected layer reshape the high-dimensional data into a 1D vector with dimensions corresponding to the input dimension.

[0132] The model extracts hierarchical features to generate deeper geometric parameter representations. These geometric parameters, derived from the encoder, are then fed into the decoder, which utilizes fully connected layers for upsampling and dimensionality expansion to reconstruct the original representation. S Parameters. Ultimately, the decoder model is used as a surrogate model, obtained from the filter's input geometric parameters. S Parameters. Decoder's S The parameters need to approximate the filter's... S Parameters. The reconstruction loss is defined as follows:

[0133]

[0134] Where x and x rec They are actual and predicted, respectively. S Parameters. K is the number of data samples. Furthermore, the encoder's geometric parameters need to approximate the filter's true geometric parameters; the prediction loss is defined as:

[0135]

[0136] Where μ and μ pre These are the actual and predicted geometric parameters, respectively. The objective function to be minimized is... L rec and L pre sum:

[0137]

[0138] All these loss values, and the loss value calculated according to this formula, are used in the backpropagation algorithm to train the entire model. Furthermore, the frequency-weighted loss emphasizes the region near the passband (in which... S The parameter variations are highly nonlinear, and the impact of out-of-band regions is reduced. A frequency-dependent physical constraint regularization term is proposed. L freq Defined as prediction frequency weighted S The sum of the mean square errors between the parameters and the theoretical expected values:

[0139]

[0140]

[0141] in, S pre Indicates prediction S parameter, S true Design representation S parameter, f i For operating frequency, f c It is the center frequency of the filter. p Control the attenuation rate of the weights. Boundary constraints ensure port matching at the center frequency and passivity throughout the operating band.

[0142] The proposed framework incorporates two key electromagnetic constraints through a dedicated boundary loss term: the port impedance constraint at the center frequency and the passivity of the filter. This can be achieved as follows:

[0143]

[0144] To address the issue of unstable accuracy in pure data fitting surrogate models under boundary conditions and sparse data conditions, a hybrid loss function combining physical constraints and data-driven approaches is constructed. Physical constraints are incorporated into the loss function of the one-dimensional autoencoder network, and the network's total loss function is expressed as follows:

[0145]

[0146] like Figure 6 The diagram shows the flowchart of the fast synthesis method for broadband balanced high-temperature superconducting bandpass filters proposed in this invention, which combines artificial lemming algorithm optimization and a two-stage PINN model. The loss function of the ALA is defined based on key parameters as follows:

[0147]

[0148] in, R DM and I DM These are the RL and IL levels in the DM response, respectively. I CM It is the IL level in the CM response. ω 1 and ω 2 represents the passband bandwidth and the stopband bandwidth, respectively.

[0149] A wideband HTS balanced bandpass filter is designed through the synergistic effect of two-stage PINN and ALA algorithms. In the PINN-1 stage, transmission line modeling is performed to obtain the optimal characteristic impedance value that meets design requirements. Subsequently, in the PINN-2 stage, the filter layout is constructed based on the transmission line model, and the optimal physical dimensions are jointly optimized using the ALA algorithm.

[0150] The following experiment verifies this embodiment:

[0151] Using the above comprehensive method, a second-order high-temperature superconducting balanced bandpass filter was designed. For example... Figure 7 and Figure 8 The diagram shows the topology of the second-order filter and its three-dimensional topology. The designed second-order high-temperature superconducting balanced bandpass filter has a center frequency of 4 GHz, a relative bandwidth of 112%, and a return loss of 30 dB within the passband. In the common-mode response, the insertion loss across the entire frequency band is 17 dB. The obtained optimal characteristic impedance is: Z b1 =70.16Ω, Z b2 =26.95Ω, Z e1 =136.52Ω, Z o1 =44.51Ω.

[0152] likeFigure 9 The diagram shows the layout of a second-order filter in an embodiment of the present invention. All dimensions in the diagram are in millimeters (mm): W1=0.09, W2=0.12, W3=1.73, W4=1.3, W5=0.8, W6=1.51, G1=0.05, L1=3.79, L2=8.24, L3=7.35, L4=14.60, L5=2.89, L6=2.91, L7=6.8.

[0153] like Figure 10 The figure shows the PINN-1 predicted response and the ideal Chebyshev response of the second-order filter in this embodiment of the invention. The results show that PINN-1 can effectively predict the characteristic impedance of the filter, making the response close to that of an ideal Chebyshev filter. Training lasted one hour, and after training, prediction only took a few seconds.

[0154] The geometric parameters are initialized based on the characteristic impedance values. Then, upper and lower limits for each parameter can be defined. Afterward, a dataset can be constructed to train the one-dimensional autoencoder (1D-CAE) model. After training convergence, the developed 1D-CAE model achieved an average error of 3.6e on both the training and test sets. -4 and 3.4e -3 This is because some data was magnified by 10. 3 The error value is amplified accordingly. Then, the decoder of the 1D-CAE model is used instead of the electromagnetic model. A loss function is defined, 30 search agents are set during ALA algorithm initialization, and the maximum number of iterations is set to 300. After ALA optimization, all values ​​of the filter's geometric parameters can be obtained.

[0155] like Figure 11 The diagram shows the prediction and simulation results of the second-order filter in this embodiment of the invention. It illustrates the differential-mode and common-mode response results of the PINN-2 model decoder under optimal geometric parameters, as well as the electromagnetic simulation results. Based on the DM and CM response results from the physical model, it can be seen that the DM passband is 1.69-6.32 GHz, and the DM stopband is 0-1.5 GHz and 6.5-8 GHz. Within the DM passband, the reflection loss is better than 27 dB. The full-band common-mode rejection ratio is greater than 10 dB. Compared with the calculated results, the two agree well, indicating that the surrogate model can perfectly match the response results.

[0156] The above practical design examples of second-order broadband balanced high-temperature superconducting filters demonstrate the significant advantages of the proposed method in shortening the design cycle, improving model generalization ability, and ensuring circuit performance. The experimental results and simulation results are in high agreement, confirming that the proposed method has strong engineering applicability and high design reliability.

[0157] Example 2

[0158] The second embodiment of the present invention also provides a design method for a high-temperature superconducting bandpass filter. The method shown in this embodiment is basically similar to that in the first embodiment, except that:

[0159] In this embodiment, the target differential mode transfer function adopts the equirippled Chebyshev form, and its specific implementation includes:

[0160] Determine the passband ripple level and cutoff frequency of the Chebyshev response, and calculate the corresponding electrical length;

[0161] Construct a Chebyshev polynomial function of the first kind of degree;

[0162] The target differential transfer function is generated based on the Chebyshev polynomial function, and the loss function of the circuit-level physical information neural network model is embedded in it.

[0163] It should be noted that the passband ripple level determines the range of signal amplitude fluctuation within the passband, the cutoff frequency defines the boundary of the passband, and the electrical length is a key electrical parameter of the transmission line, which is related to the physical length and operating frequency. Accurate calculation of the electrical length can ensure the matching degree between the transfer function and the actual filter structure.

[0164] Chebyshev polynomials exhibit equiripple characteristics, and their degree is related to the filter's order; higher orders result in better out-of-band suppression. By constructing Chebyshev polynomial functions of corresponding degrees, the ideal differential-mode transmission characteristics of the filter can be accurately described.

[0165] Specifically, by embedding the transfer function into the loss function, the neural network model will use the ideal transfer function as the target during training to optimize the characteristic impedance parameters of the output, so that the passband characteristics of the generated filter are close to the ideal state of equal ripple, thereby improving the passband performance of the filter.

[0166] In this embodiment, the loss function of the circuit-level physical information neural network model includes a frequency-weighted loss term, and the specific implementation of the frequency-weighted loss term includes:

[0167] Identify the passband frequency range and out-of-band frequency range of the filter;

[0168] A high weighting coefficient is set for frequency points within the passband frequency range, and a low weighting coefficient is set for frequency points within the out-of-band frequency range.

[0169] The corresponding weighting coefficients are multiplied by the scattering parameter prediction error to obtain the frequency-weighted loss term, which is then embedded into the total loss function of the circuit-level physical information neural network model.

[0170] It should be noted that the passband is the frequency band in which the filter effectively transmits signals; it is the core operating region of the filter. The out-of-band is the frequency band in which signal transmission needs to be suppressed. Clearly defining the frequency ranges of these two bands provides a basis for setting differentiated weighting coefficients for each band.

[0171] Specifically, a high weighting coefficient is set for frequency points within the passband frequency range, while a low weighting coefficient is set for frequency points within the out-of-band frequency range. This setting allows the model to place greater emphasis on the prediction error in the passband region when calculating the loss value, thus shifting the focus of model training towards passband performance. This ensures that the transmission characteristics within the passband are closer to the ideal target, while appropriately relaxing the requirements for the out-of-band region, thereby improving the training efficiency of the model while ensuring core performance.

[0172] Specifically, scattering parameters directly reflect the transmission and reflection characteristics of the filter. Multiplying the weighting coefficients by the scattering parameter prediction error amplifies the error in the passband region and reduces the error in the out-of-band region. Embedding this weighted loss term into the total loss function can guide the model to prioritize optimizing passband performance during training, enabling the generated filter to have superior performance in the core operating frequency band.

[0173] In this embodiment, the circuit-level physical information neural network model satisfies the passivity boundary condition, specifically implemented as follows:

[0174] The condition information of the passive boundary condition is invoked, and the amplitude constraint that the scattering parameter must satisfy is determined based on the condition information;

[0175] The passive constraint is transformed into a penalty term and embedded into the loss function of the circuit-level physical information neural network model;

[0176] During the training process of the circuit-level physical information neural network model, a penalty is imposed on the prediction results that violate the passivity constraint in order to achieve boundary condition constraints.

[0177] It should be noted that passivity is a fundamental characteristic of passive microwave devices. Passive filters do not generate energy themselves, and their scattering parameters must meet specific amplitude constraints. Clarifying these constraints provides a basis for subsequently converting them into penalty terms for model training.

[0178] By constructing a penalty term, a large loss value will be generated when the scattering parameter corresponding to the characteristic impedance of the model output violates the passivity constraint, forcing the model to adjust the parameters during training so that the output result meets the passivity requirement.

[0179] Specifically, during the training iteration process, the model continuously optimizes the parameters to reduce the loss value. When a prediction result that violates the passive constraint occurs, the penalty term increases the loss value, guiding the model to adjust in the direction that satisfies the passive constraint. Finally, the filter structure corresponding to the characteristic impedance of the output has passive characteristics and can work stably.

[0180] In this embodiment, the method further includes:

[0181] Based on the physical layout, a layout-level physical information neural network model is constructed, using a one-dimensional convolutional autoencoder model, with the encoder being a convolutional neural network and the decoder being a multi-layer fully connected network.

[0182] A hybrid loss function is constructed for the layout-level physical information neural network model. The hybrid loss function includes the mean square error terms of the actual geometric parameters and the predicted geometric parameters, the mean square error terms of the actual scattering parameters and the reconstructed scattering parameters under frequency weighting, and physical constraint terms of passivity and port matching are added.

[0183] Obtain an initial dataset, train the layout-level physical information neural network model, and establish a mapping relationship between layout geometric parameters and differential and common-mode scattering parameters through the layout-level physical information neural network model;

[0184] Using the decoder of the aforementioned map-level physical information neural network model as the electromagnetic proxy model, the number of search proxies and the maximum number of iterations of the artificial lemming optimization algorithm are set to complete the algorithm initialization;

[0185] The artificial lemming optimization algorithm is used to perform global search and local refinement of the layout geometry parameters, and combined with an automatic data incremental training strategy to continuously optimize the layout-level physical information neural network model, outputting the optimal layout geometry parameters that meet the design specifications.

[0186] It should be noted that the one-dimensional convolutional autoencoder can effectively extract deep features of layout geometry parameters. The encoder uses a convolutional neural network, which has a powerful feature extraction capability and can transform high-dimensional layout geometry parameters into low-dimensional feature vectors. The decoder uses a multi-layer fully connected network, which can reconstruct the low-dimensional feature vectors into the corresponding differential-mode and common-mode scattering parameters, and establish a precise mapping relationship between layout geometry parameters and electromagnetic response.

[0187] Among them, the mean square error term of the actual geometric parameters and the predicted geometric parameters included in the hybrid loss function can ensure the model's fitting accuracy to the layout geometric parameters; the mean square error term of the actual scattering parameters and the reconstructed scattering parameters under frequency weighting can improve the model's prediction accuracy of the electromagnetic response; the physical constraint terms of passivity and port matching can ensure that the filter structure corresponding to the layout parameters output by the model has physical realizability and avoid non-physical design results.

[0188] The initial dataset contains a large number of layout geometric parameters and their corresponding electromagnetic response data. By training the model, the model can learn the intrinsic relationship between the two, enabling it to predict the electromagnetic response based on the input layout geometric parameters, thus providing an accurate surrogate model for subsequent optimization.

[0189] The decoder can quickly predict the electromagnetic response corresponding to the geometric parameters of the layout, replacing time-consuming full-wave electromagnetic simulation and significantly improving optimization efficiency. Setting the number of search agents and the maximum number of iterations for the artificial lemming optimization algorithm can ensure the search range and convergence speed of the algorithm, laying the foundation for subsequent global search and local refinement.

[0190] Specifically, the artificial lemming optimization algorithm possesses powerful global search capabilities, enabling it to find the optimal solution within a broad parameter space. Simultaneously, it refines the parameters near the optimal solution locally, further enhancing optimization accuracy. The automatic incremental data training strategy continuously supplements the optimization process with new datasets, continuously optimizing the layout-level physical information neural network model, resulting in increasingly higher prediction accuracy and ultimately outputting the optimal layout geometric parameters that meet the design specifications.

[0191] In this embodiment, the steps of using the artificial lemming optimization algorithm to perform global search and local refinement of the layout geometric parameters, and continuously optimizing the layout-level physical information neural network model in conjunction with an automatic data incremental training strategy, to output the optimal layout geometric parameters that meet the design specifications, include:

[0192] The loss function of the artificial lemming optimization algorithm is constructed based on the return loss level, insertion loss level of the differential mode response, and insertion loss level of the common mode response.

[0193] Set the number of search agents and the maximum number of iterations for the artificial lemming optimization algorithm, and initialize the position parameters of each agent;

[0194] Using the decoder of the layout-level physical information neural network model, the scattering parameters corresponding to each agent position are predicted, and the corresponding loss function values ​​are calculated.

[0195] Update the agent position based on the loss function value, perform a global search, refine the parameters near the optimal solution locally, and output the optimal layout geometric parameters.

[0196] It should be noted that the return loss level reflects the signal reflection degree of the filter, the insertion loss level reflects the signal transmission efficiency of the filter, and the insertion loss level of the common-mode response reflects the anti-interference capability of the filter. Incorporating these three key indicators into the loss function can ensure that the optimized filter meets the design requirements in terms of core performance.

[0197] The number of search agents determines the algorithm's parallel search capability; a larger number allows for a wider search range, but also increases computational cost. The maximum number of iterations determines the algorithm's convergence time; setting this parameter appropriately ensures that the algorithm converges to the optimal solution within a finite time. Initializing the position parameters of each agent provides the algorithm with an initial search starting point.

[0198] The decoder can quickly and accurately predict the electromagnetic response corresponding to each agent position. After obtaining the scattering parameters, it calculates the return loss level and insertion loss level based on the scattering parameters, and then substitutes them into the loss function to calculate the loss value, providing an evaluation basis for the algorithm iteration.

[0199] Specifically, the algorithm adjusts the agent's position based on the magnitude of the loss function value. The smaller the loss value, the higher the probability of the agent moving. Through continuous iteration, it achieves the search for the optimal solution globally. After finding the optimal solution, it refines the parameters in its vicinity to further improve the accuracy of the parameters, and finally outputs the optimal layout geometry parameters that meet the design specifications.

[0200] In this embodiment, the steps of predicting the scattering parameters corresponding to each agent location and calculating the corresponding loss function value using the decoder of the layout-level physical information neural network model include:

[0201] The layout geometry parameters corresponding to each agent position in the artificial lemming optimization algorithm are input into the encoder of the trained layout-level physical information neural network model to extract the deep feature vector of the layout geometry parameters.

[0202] The deep feature vector is input into the decoder, and the corresponding differential mode scattering parameters and common mode scattering parameters are reconstructed through upsampling and dimensionality transformation of a multi-layer fully connected network.

[0203] Extract the return loss level and insertion loss level corresponding to the differential mode scattering parameters, and the insertion loss level corresponding to the common mode scattering parameters to obtain the level parameters;

[0204] The level parameters are substituted into the pre-constructed loss function of the artificial lemming optimization algorithm to calculate the loss function value corresponding to each agent position.

[0205] It should be noted that the encoder can extract features from the input layout geometry parameters, transforming high-dimensional parameter information into low-dimensional feature vectors. These feature vectors contain the core information of the layout geometry parameters, providing a foundation for subsequent scattering parameter reconstruction.

[0206] The decoder's multi-layer fully connected network possesses powerful dimensionality transformation capabilities, enabling it to convert low-dimensional feature vectors into high-dimensional scattering parameters. This achieves a precise mapping from layout geometry parameters to electromagnetic response, replacing time-consuming full-wave electromagnetic simulation. Scattering parameters directly reflect the electromagnetic response; through specific calculation formulas, key performance indicators such as return loss level and insertion loss level can be extracted from these parameters. These indicators are the direct basis for calculating the loss function value.

[0207] Specifically, the loss function can quantitatively evaluate the performance of each agent position. The smaller the loss value, the better the geometric parameters of the layout corresponding to that position. The calculated loss function value will provide a key basis for the agent position update of the artificial lemming optimization algorithm and guide the algorithm to iterate towards the optimal solution.

[0208] Example 3

[0209] Please see Figure 12 The third embodiment of the present invention provides a high-temperature superconducting bandpass filter design system, applied to the method described in any of the above embodiments, the system comprising:

[0210] The device configuration module 10 is used to determine the design specifications and scalable structural framework of the filter. The design specifications include center frequency, relative bandwidth, filter order, differential mode return loss and common mode rejection. The structural framework is composed of multiple coupling lines cascaded with one more branch line.

[0211] The first construction module 20 is used to construct the differential-mode equivalent circuit model and the common-mode equivalent circuit model of the filter structure based on transmission line theory.

[0212] The second building module 30 is used to build a circuit-level physical information neural network model with a multilayer perceptron architecture, including setting multiple hidden layers and configuring a corresponding number of neurons, while adding residual connections.

[0213] The model training module 40 is used to input the design indicators into the circuit-level physical information neural network model, embed the matrix derived from the equivalent circuit and the target differential mode transfer function into the model's loss function, and complete the model training.

[0214] Impedance classification module 50 is used to output the even-mode characteristic impedance, odd-mode characteristic impedance and characteristic impedance of the branch line of the coupled line using the trained circuit-level physical information neural network model.

[0215] The layout generation module 60 is used to generate the physical layout of the filter based on the various characteristic impedances of the output.

[0216] Example 4

[0217] A fourth embodiment of the present invention provides a readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method described in any of the above embodiments.

[0218] Example 5

[0219] A fifth embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described in any of the above embodiments.

[0220] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0221] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. 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 any suitable manner in one or more embodiments or examples.

[0222] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A design method for a high-temperature superconducting bandpass filter, characterized in that, The method includes: The design specifications and scalable structural framework of the filter are determined. The design specifications include center frequency, relative bandwidth, filter order, differential mode return loss and common mode rejection. The scalable structural framework is composed of multiple cascaded coupling lines and one additional branch line. Based on transmission line theory, differential-mode equivalent circuit models and common-mode equivalent circuit models of filter structures are constructed respectively. Construct a circuit-level physical information neural network model using a multilayer perceptron architecture, including setting multiple hidden layers and configuring a corresponding number of neurons, while adding residual connections; The design metrics are input into the circuit-level physical information neural network model. The matrix derived from the equivalent circuit, the target differential mode transfer function, the frequency-weighted loss terms for the passband and out-of-band frequency ranges, and the passive constraints are embedded into the model's loss function to complete model training. Using the trained circuit-level physical information neural network model, the even-mode characteristic impedance, odd-mode characteristic impedance, and characteristic impedance of the branch lines of the coupled line are output. The physical layout of the filter is generated based on the various characteristic impedances of the output.

2. The high-temperature superconducting bandpass filter design method according to claim 1, characterized in that, The target differential mode transfer function adopts the equirippled Chebyshev form, and its specific implementation includes: Determine the passband ripple level and cutoff frequency of the Chebyshev response, and calculate the corresponding electrical length; Construct a Chebyshev polynomial function of the first kind of degree; The target differential transfer function is generated based on the Chebyshev polynomial function, and the loss function of the circuit-level physical information neural network model is embedded in it.

3. The high-temperature superconducting bandpass filter design method according to claim 2, characterized in that, The specific implementation of the frequency-weighted loss term includes: Identify the passband frequency range and out-of-band frequency range of the filter; A high weighting coefficient is set for frequency points within the passband frequency range, and a low weighting coefficient is set for frequency points within the out-of-band frequency range. Multiply the corresponding weighting coefficients by the scattering parameter prediction error to obtain the frequency-weighted loss term.

4. The high-temperature superconducting bandpass filter design method according to claim 2, characterized in that, The circuit-level physical information neural network model satisfies the passive boundary condition, and its specific implementation includes: The condition information of the passive boundary condition is invoked, and the amplitude constraint that the scattering parameter must satisfy is determined based on the condition information; The passive constraint is transformed into a penalty term and embedded into the loss function of the circuit-level physical information neural network model; During the training process of the circuit-level physical information neural network model, a penalty is imposed on the prediction results that violate the passivity constraint in order to achieve boundary condition constraints.

5. The high-temperature superconducting bandpass filter design method according to any one of claims 1-4, characterized in that, The method further includes: Based on the physical layout, a layout-level physical information neural network model is constructed, using a one-dimensional convolutional autoencoder model, with the encoder being a convolutional neural network and the decoder being a multi-layer fully connected network. A hybrid loss function is constructed for the layout-level physical information neural network model. The hybrid loss function includes the mean square error terms of the actual geometric parameters and the predicted geometric parameters, the mean square error terms of the actual scattering parameters and the reconstructed scattering parameters under frequency weighting, and physical constraint terms of passivity and port matching are added. Obtain an initial dataset, train the layout-level physical information neural network model, and establish a mapping relationship between layout geometric parameters and differential and common-mode scattering parameters through the layout-level physical information neural network model; Using the decoder of the aforementioned map-level physical information neural network model as the electromagnetic proxy model, the number of search proxies and the maximum number of iterations of the artificial lemming optimization algorithm are set to complete the algorithm initialization; The artificial lemming optimization algorithm is used to perform global search and local refinement of the layout geometry parameters, and combined with an automatic data incremental training strategy to continuously optimize the layout-level physical information neural network model, outputting the optimal layout geometry parameters that meet the design specifications.

6. The high-temperature superconducting bandpass filter design method according to claim 5, characterized in that, The steps of using the artificial lemming optimization algorithm to perform global search and local refinement of the layout geometric parameters, and continuously optimizing the layout-level physical information neural network model by combining an automatic data incremental training strategy, to output the optimal layout geometric parameters that meet the design specifications, include: The loss function of the artificial lemming optimization algorithm is constructed based on the return loss level, insertion loss level of the differential mode response, and insertion loss level of the common mode response. Set the number of search agents and the maximum number of iterations for the artificial lemming optimization algorithm, and initialize the position parameters of each agent; Using the decoder of the layout-level physical information neural network model, the scattering parameters corresponding to each agent position are predicted, and the corresponding loss function values ​​are calculated. Update the agent position based on the loss function value, perform a global search, refine the parameters near the optimal solution locally, and output the optimal layout geometric parameters.

7. The high-temperature superconducting bandpass filter design method according to claim 6, characterized in that, The steps of predicting the scattering parameters corresponding to each agent location and calculating the corresponding loss function value using the decoder of the layout-level physical information neural network model include: The layout geometry parameters corresponding to each agent position in the artificial lemming optimization algorithm are input into the encoder of the trained layout-level physical information neural network model to extract the deep feature vector of the layout geometry parameters. The deep feature vector is input into the decoder, and the corresponding differential mode scattering parameters and common mode scattering parameters are reconstructed through upsampling and dimensionality transformation of a multi-layer fully connected network. Extract the return loss level and insertion loss level corresponding to the differential mode scattering parameters, and the insertion loss level corresponding to the common mode scattering parameters to obtain the level parameters; The level parameters are substituted into the pre-constructed loss function of the artificial lemming optimization algorithm to calculate the loss function value corresponding to each agent position.

8. A design system for a high-temperature superconducting bandpass filter, characterized in that, The system, applicable to the method of any one of claims 1-7, comprises: The device configuration module is used to determine the design specifications and scalable structural framework of the filter. The design specifications include center frequency, relative bandwidth, filter order, differential mode return loss and common mode rejection. The scalable structural framework is composed of multiple cascaded coupling lines and one additional branch line. The first construction module is used to construct the differential-mode equivalent circuit model and the common-mode equivalent circuit model of the filter structure based on transmission line theory. The second building module is used to build a circuit-level physical information neural network model with a multilayer perceptron architecture, including setting multiple hidden layers and configuring a corresponding number of neurons, while adding residual connections. The model training module is used to input the design indicators into the circuit-level physical information neural network model, and to embed the matrix derived from the equivalent circuit, the target differential mode transfer function, the frequency weighted loss terms of the passband frequency range and the out-of-band frequency range, and the passive constraint conditions into the loss function of the model to complete the model training. The impedance classification module is used to output the even-mode characteristic impedance, odd-mode characteristic impedance, and characteristic impedance of the branch lines of the coupled lines using the trained circuit-level physical information neural network model. The layout generation module is used to generate the physical layout of the filter based on the various characteristic impedances of the output.

9. A readable storage medium having computer instructions stored thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-7.

Images