Passive device synthesis method and system based on frequency response performance segmentation and reconstruction

Abstract

The invention discloses a passive device synthesis method and system based on frequency response performance segmentation and reconstruction, and belongs to the crossing field of microwave technology and artificial intelligence, and the method comprises the following steps: obtaining component parameters of various passive device topological structures for electromagnetic simulation; inputting the component parameters into a corresponding pre-trained passive device frequency response sub-neural network to obtain predicted S parameters of each passive device topological structure, splicing the predicted S parameters and then inputting the spliced predicted S parameters into a pre-trained passive device frequency response reconstruction neural network to obtain predicted S parameters of a cascade structure; and performing global optimization on the prediction S parameter of the cascade structure to obtain an optimal component parameter combination corresponding to the optimal prediction S parameter. The method can solve the problem that in the prior art, structural parameter analysis is only carried out on a single fixed passive device topology, and structural parameter optimization of various passive device topologies cannot be supported.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross field of microwave technology and artificial intelligence, and particularly relates to a passive device synthesis method and system based on frequency response performance segmentation reconstruction. BACKGROUND

[0002] A microwave low-pass filter is a key passive device in a modern wireless communication system, and its performance directly affects the signal quality and overall stability of the system. With the expansion of the communication frequency band to the millimeter wave, more stringent requirements are put forward for the performance indicators of the filter. The traditional design method of the microwave low-pass filter mainly relies on the calculation of a normalized prototype (such as Butterworth and Chebyshev prototype) based on an empirical formula and inverse normalization calculation, which needs manual trial and error optimization. The designer determines the initial parameters according to the target indicators by referring to the design manual, and then adjusts and optimizes repeatedly through electromagnetic simulation software. The optimization iteration time is too long, and in addition, the high-frequency band parasitic effect (such as the parasitic parameters of inductance and capacitance) is difficult to model through the empirical formula, resulting in a significant deviation between the simulation results and the theoretical analysis. This method has three significant defects: first, when the working frequency changes, the device parameters need to be re-optimized, which consumes a large amount of time and computing resources, and the design efficiency is extremely low; second, multi-parameter optimization is easy to fall into a local optimal solution, resulting in errors in the prediction of S parameters, which is difficult to meet the high performance requirements; third, for a complex cascaded network, the traditional ABCD matrix (ABCD matrix is a commonly used two-port network parameter matrix in microwave and radio frequency engineering) manual calculation method is complex to model, is easy to have calculation errors, and is difficult to accurately analyze the nonlinear coupling relationship between the sub-circuit and the system level performance, and cannot efficiently match the system level performance requirements.

[0003] In recent years, deep learning technology has been applied to the design of passive devices to simplify the design process. Chinese patent application number 202410663095.9 discloses a deep learning-based inverse design method for passive integrated devices in radio frequency integrated circuits. This scheme uses a Conditional Generative Adversarial Network (CGAN) to directly generate the structural parameters of the device using the target S21 parameters as input, achieving an inverse mapping between performance indicators and structural parameters. However, this scheme has two major limitations: First, the topology is rigidly bound, lacking universal applicability. The model training is deeply tied to the specific component topology, and the trained CGAN model is only effective for a single fixed topology. When the topology (such as Π-type, T-type, or their cascaded configuration) changes, the model completely fails. In this case, the dataset must be regenerated, and a completely new network must be designed and trained from scratch, resulting in high time and space costs and lacking scalability. Second, it lacks multi-objective global optimization, resulting in a single design dimension. This scheme only provides a direct mapping from S21 parameters to structural parameters but does not incorporate a global optimization algorithm. This makes it impossible for it to efficiently balance and optimize multiple conflicting performance metrics (such as passband insertion loss, stopband rejection, and transition band slope), making it difficult to meet complex engineering design requirements. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for the synthesis of passive devices based on frequency response performance segmentation and reconstruction, which can solve the problem that the existing technology can only perform structural parameter analysis on a single fixed passive device topology and cannot support the structural parameter optimization of multiple passive device topologies.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a passive device synthesis method based on frequency response performance segmentation and reconstruction includes the following steps: Component parameters for electromagnetic simulation of various passive device topologies; The component parameters are input into a pre-trained passive device frequency response sub-neural network to obtain the predicted S-parameters of each passive device topology. The predicted S-parameters of each passive device topology are then concatenated and input into a pre-trained passive device frequency response reconstruction neural network to obtain the predicted S-parameters of the cascaded structure. The predicted S-parameters of the cascaded structure are globally optimized to obtain the optimal combination of component parameters corresponding to the optimal predicted S-parameters.

[0006] In some embodiments, both the pre-trained passive device frequency response sub-neural network and the pre-trained passive device frequency response reconstruction neural network include: an input layer, a hidden layer and an output layer connected in sequence, wherein the hidden layer includes a shared feature layer and the shared feature layer adopts a fully connected network.

[0007] In some implementations, the component parameters are input into a pre-trained passive device frequency response sub-neural network to obtain the predicted S-parameters of each passive device topology, specifically including: The component parameters are transmitted to the shared feature layer through the input layer to extract the topological common features of the component parameters; The common topological features are linearly transformed through the output layer to obtain the predicted S-parameters corresponding to different passive device topologies.

[0008] In some implementations, the step of concatenating the predicted S-parameters of each passive device topology and inputting them into a pre-trained passive device frequency response reconstruction neural network to obtain the predicted S-parameters of the cascaded structure specifically includes: The predicted S-parameters corresponding to the topology of each passive device are concatenated to obtain a cascaded feature vector. The cascaded feature vector is transmitted to the shared feature layer through the input layer to extract the cascaded coupling features of the cascaded feature vector; The cascaded coupling features are linearly transformed through the output layer to obtain the predicted S-parameters of the cascaded structure.

[0009] In some implementations, the step of globally optimizing the predicted S-parameters of the cascaded structure to obtain the optimal combination of component parameters corresponding to the optimal predicted S-parameters specifically includes: An objective function is constructed with the goal of minimizing the error between the predicted S-parameters and the target S-parameters of the cascaded structure. The objective function is the amplitude error between the predicted S-parameters and the preset fixed threshold in the passband and stopband, respectively, and the amplitude error between the predicted S-parameters and the target S-parameters in the transition band, which is constructed by weighted sum of squares. The upper and lower limits of preset component parameters are used as constraints. Based on the particle swarm optimization algorithm, the objective function is minimized and iteratively optimized to obtain the optimal combination of component parameters corresponding to the optimal predicted S-parameters.

[0010] In some embodiments, the pre-training of the passive device frequency response sub-neural network includes the following steps: Sampling is performed within the component parameter range of each preset passive device topology to generate multiple sets of component parameter combinations. The true S-parameters of each set of component parameter combinations in the target frequency band are obtained through electromagnetic simulation. A training dataset is constructed by mapping the component parameter combinations and the true S-parameters. The true S-parameters include the amplitude and phase information of S11, S21 and S22 parameters. During training, the combination of component parameters is used as the input of the passive device frequency response sub-neural network, the corresponding real S-parameters are used as the training target, and the predicted S-parameters of the passive device topology are used as the output. The weights of the passive device frequency response sub-neural network are optimized by minimizing the corresponding complex loss function, and the training is completed until convergence to obtain the pre-trained passive device frequency response sub-neural network. The complex loss function of the passive device frequency response sub-neural network consists of the amplitude error and phase error between the actual S-parameters and the predicted S-parameters of the passive device topology.

[0011] In some embodiments, the pre-training of the passive device frequency response reconstruction neural network includes the following steps: The total S-parameters of the cascaded structure are obtained by combining the parameters of each group of components in the target frequency band and using electromagnetic simulation. During training, the actual S-parameters are used as the input to the passive device frequency response reconstruction neural network, the total S-parameters of the corresponding cascaded structure are used as the training target, and the predicted S-parameters of the cascaded structure are used as the output. The weights of the passive device frequency response reconstruction neural network are optimized by minimizing the corresponding complex loss function, and the training is completed until convergence to obtain the pre-trained passive device frequency response reconstruction neural network. The complex loss function of the passive device frequency response reconstruction neural network consists of the amplitude error and phase error between the total S-parameters of the cascaded structure and the predicted S-parameters of the cascaded structure.

[0012] Secondly, a passive device synthesis system based on frequency response performance segmentation and reconstruction includes: The component parameter acquisition module is used to acquire component parameters for electromagnetic simulation of various passive component topologies; The segmentation and reconstruction module is used to input the component parameters into the corresponding pre-trained passive device frequency response sub-neural network to obtain the predicted S-parameters of each passive device topology, and to concatenate the predicted S-parameters of each passive device topology and input them into the pre-trained passive device frequency response reconstruction neural network to obtain the predicted S-parameters of the cascaded structure. The global optimization module is used to perform global optimization on the predicted S-parameters of the cascaded structure to obtain the optimal combination of component parameters corresponding to the optimal predicted S-parameters.

[0013] Thirdly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the passive device synthesis method based on frequency response performance segmentation and reconstruction.

[0014] Fourthly, a computer program product comprising a computer program that, when executed by a processor, implements the steps of the passive device synthesis method based on frequency response performance segmentation and reconstruction.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a passive device synthesis method based on frequency response performance segmentation and reconstruction. The method predicts the S-parameters of the topology of a single passive device through a passive device frequency response sub-neural network. The passive device frequency response reconstruction neural network learns the cascade coupling characteristics based on the predicted S-parameters of the spliced ​​passive device topologies and finally outputs the predicted S-parameters of the cascaded structure. This design allows for flexible combination of different passive device topologies. When the sub-neural network topology changes, only the sub-neural network needs to be retrained, while the reconstruction neural network does not need to be retrained, thus improving the efficiency of reverse design.

[0016] Furthermore, the hidden layers include a shared feature layer, which extracts common information from the input features, reducing model complexity and computational resource consumption while improving the generalization ability of the feature representation. This structure improves the overall accuracy of S-parameter prediction while maintaining model scalability and supporting various topology adaptations, thereby enhancing design reliability and efficiency.

[0017] Furthermore, in the passive device frequency response sub-neural network, the shared feature layer extracts the topological common features of the component parameters, reduces redundant calculations, ensures high accuracy in S-parameter prediction for different frequency bands, and reduces model training costs and improves design efficiency through feature reuse.

[0018] Furthermore, in the passive device frequency response reconstruction neural network, the predicted S-parameters corresponding to different passive device topologies are used to learn the nonlinear coupling relationships in the cascaded circuit. A shared feature layer extracts cascade coupling features, effectively solving the problems of complex cascade modeling and large calculation errors in traditional methods. This enables rapid and accurate prediction of cascaded circuit performance without the need for manual calculation of the ABCD matrix, improving overall design accuracy and automation.

[0019] Furthermore, the objective function of the particle swarm optimization algorithm is the amplitude error between the predicted S-parameters and the preset fixed threshold in the passband and stopband, respectively, and the amplitude error between the predicted S-parameters and the target S-parameters in the transition band. This is calculated using a weighted sum of squares, comprehensively considering key performance indicators such as passband insertion loss, stopband suppression, and transition band slope to ensure that the optimized component parameters meet engineering requirements. The PSO algorithm avoids the problem of traditional optimization methods easily getting trapped in local optima, efficiently finding the optimal solution through global search and shortening the design cycle. This directly solves the shortcomings of low efficiency and insufficient accuracy in multi-parameter optimization in existing technologies.

[0020] Furthermore, in the pre-training of the passive device frequency response sub-neural network, the weights of the passive device frequency response sub-neural network are optimized by minimizing the corresponding complex loss function. The complex loss function of the passive device frequency response sub-neural network consists of the amplitude error and phase error between the true S-parameters and the predicted S-parameters. Simultaneously optimizing the amplitude error and phase error significantly improves the prediction accuracy in the high-frequency band and overcomes the problem of inaccurate modeling in the high-frequency band by traditional empirical formulas.

[0021] Furthermore, in the pre-training of the passive device frequency response reconstruction neural network, the complex loss function of the passive device frequency response reconstruction neural network consists of the amplitude error and phase error between the total S-parameters of the cascaded structure and the predicted S-parameters of the cascaded structure. This enables the reconstruction neural network to accurately learn the coupling characteristics of the cascaded circuit. The complex loss function ensures that both amplitude and phase are optimized, improving the prediction accuracy of the cascaded structure S-parameters. This avoids the complexity and errors of traditional manual calculation of cascaded S-parameters. Moreover, when the topology changes, retraining is not required; only the S-parameters need to be input to support the automated design of multiple topology combinations. Attached Figure Description

[0022] Figure 1 A flowchart of a passive device synthesis method based on frequency response performance segmentation and reconstruction provided in an embodiment of the present invention; Figure 2 This is a structural diagram of a passive device synthesis system based on frequency response performance segmentation and reconstruction provided in an embodiment of the present invention. Figure 3 A technical framework diagram of the passive device synthesis method based on frequency response performance segmentation and reconstruction provided in the embodiments of the present invention; Figure 4 A flowchart of ADS-MATLAB co-simulation provided in this embodiment of the invention; Figure 5 A framework diagram of a passive device frequency response sub-neural network provided in an embodiment of the present invention; Figure 6 A framework diagram of the step-by-step neural network architecture provided in an embodiment of the present invention; Figure 7 A flowchart of PSO prediction provided for embodiments of the present invention; Figure 8 The fitting curves of the step-by-step neural network architecture provided in this embodiment of the invention are shown in (a) for the S11 parameter, (b) for the S21 parameter, and (c) for the S22 parameter. Figure 9 The simulation results of a low-pass filter with a cutoff frequency of 4 GHz provided in the embodiments of the present invention; Figure 10 The simulation results of a low-pass filter with a cutoff frequency of 2GHz are provided for embodiments of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. The content described herein is for explanation rather than limitation of the present invention.

[0024] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification of this invention are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, systems, products or devices.

[0025] The following is an explanation of the relevant technical terms: S-parameters are scattering parameters, including S11, S21, and S22. S11 is the input reflection coefficient, S21 is the forward transmission coefficient, and S22 is the output reflection coefficient. ADS is an electromagnetic simulation software, short for Advanced Design System. Particle Swarm Optimization (PSO) algorithm; Genetic Algorithm (GA); Simulated Annealing (SA) algorithm; Convolutional Neural Network (CNN); One activation function is ReLU (Rectified Linear Unit), another is GELU (Gaussian Error Linear Unit), and yet another is Leaky ReLU (Leaky Rectified Linear Unit). One optimizer is Adam (Adaptive Moment Estimation). Application Programming Interface (API); High Frequency Structure Simulator (HFSS); Computer Simulation Technology (CST); Two-dimensional (2D); One advanced technical computing environment is MATLAB (Matrix Laboratory).

[0026] like Figure 3The diagram illustrates the technical framework of an embodiment of the present invention. Target parameter settings are the starting point for the design, and users set the performance indicators of passive devices according to engineering requirements. These indicators typically cover key performance aspects of the passband, stopband, and transition band, such as the maximum allowable insertion loss in the passband, the minimum required rejection depth in the stopband, and the required roll-off slope in the transition band. The PSO optimization algorithm, as the global optimization engine, drives the entire design process. It receives performance indicators from the target parameter settings and prediction results from the neural network model, iteratively generating and evaluating different combinations of component parameters. Its core objective is to minimize the error between the overall S-parameters predicted by the neural network and the target S-parameters. The passive device frequency response performance neural network model is the core computational module of the present invention, employing a "segmentation and reconstruction" architecture for quickly and accurately predicting circuit performance. This model contains three cooperating sub-modules. The passive device frequency response characteristic clustering sub-neural network 1 is specifically designed to learn the electromagnetic characteristics of a Π-type circuit. It takes the component parameters C1, L1, and C2 of this topology as input and outputs its independent S-parameter S1. Correspondingly, the passive device frequency response characteristic clustering sub-neural network 2 is specifically designed to learn the electromagnetic characteristics of a T-type circuit. It takes the component parameters L2, L3, and C3 of the topology as input and outputs their independent S-parameter S2. The passive component frequency response performance reconstruction neural network is responsible for learning the overall effect after cascading. It takes the S-parameters S1 and S2 predicted by the two sub-networks as input and predicts the total S-parameter S_total of the entire cascaded structure by capturing the coupling characteristics between circuits. Finally, the system output is the optimized combination of component parameters C1, L1, C2, L2, L3, and C3. When the PSO algorithm converges iteratively, this set of parameters enables the performance of the cascaded circuit to be closest to the preset target index.

[0027] like Figure 4As shown, the start marks the beginning of the entire data generation process. MATLAB parameter initialization is the first step. In this step, the component parameter scan range for passive devices (such as Π-type and T-type circuits) is preset in the MATLAB environment, setting boundaries for subsequent automated scanning. The parameter combination list is generated based on the initialized parameter range, using a sampling method to generate a large number of component parameter combinations to be simulated. These combinations constitute the basic samples of the dataset. The process then enters a loop that continuously iterates through the parameter combination list until the simulation of all combinations is completed. Within the loop, ADS is first called to modify the parameters, i.e., the component parameter values ​​of the current group are written to the ADS circuit design file via the script interface. Next, frequency domain simulation is performed, with the ADS solver performing electromagnetic simulation within the set frequency band to calculate the S-parameters of the circuit. After the simulation is complete, the system extracts and saves the S-parameters, exporting the real and imaginary parts, or key data such as amplitude and phase, from the ADS results and storing them in the specified data file. After completing the simulation and saving of a set of parameters, the process returns to the loop checkpoint. The system checks whether the parameter list has been completely traversed. If the simulation is not complete, it continues looping to simulate the next set of parameters; if it is complete, the loop exits. Finally, after the loop ends, the process generates a dataset. This step summarizes, organizes, and formats all individual simulation results into a normalized dataset that can be directly used for training neural networks. The end indicates that the entire automated data generation task has been successfully completed.

[0028] exist Figure 3 The technical approach shown and such Figure 4 Based on the simulation steps shown, such as Figure 1 The embodiment shown provides a passive device synthesis method based on frequency response performance segmentation and reconstruction, including the following steps: S1, obtain the component parameters for electromagnetic simulation of various passive device topologies; The passive device topologies employ Π-type and T-type structures. The process of acquiring component parameters is achieved through ADS-MATLAB co-simulation, generating high-precision data including parasitic effects. This data is then combined with theoretical formulas to generate rapid pre-training data, resulting in a multidimensional dataset of 21,000 sets (7,000 sets each for Π-type, T-type, and cascaded circuits). The co-simulation process is as follows: Figure 4 As shown, it specifically includes: Initialize parameter range: Define the component parameter scan range for Π type (C1: 0.7-1.4pF, L1: 3-10nH, C2: 0.7-1.4pF) and T type (L2: 1-8nH, L3: 1-8nH, C3: 1-8pF) in MATLAB.

[0029] Automated simulation: ADS is called via the TADSInterface.m function to iteratively modify parameters and execute a 1-10 GHz frequency domain simulation (step size 90.90909 MHz), extracting S... 11 S 21 S 22 The S-parameters are scattering parameters used to describe the energy reflection and transmission characteristics of multiport networks. S11 is the input reflection coefficient, which measures the impedance matching of the ports; the smaller the better. S21 is the forward transmission coefficient, which is the insertion loss for filters; the larger the better in the passband and the smaller the better in the stopband. S22 is the output reflection coefficient, which measures the impedance matching of the output ports; the smaller the better. S12 is the reverse transmission coefficient, usually the same as S21. Each S-parameter contains amplitude and phase information. The values ​​of S-parameters are strongly frequency-dependent; the curves showing the amplitude or phase change with frequency are called S-parameter curves.

[0030] Cascaded structure data generation: The independent S-parameters of the Π-type and T-type circuits are used as inputs, and the total S-parameters of the cascaded circuit are simulated using ADS as the output to construct a cascaded dataset.

[0031] S2, input the component parameters into the corresponding pre-trained passive device frequency response sub-neural network to obtain the predicted S-parameters of each passive device topology, and then input the predicted S-parameters of each passive device topology into the pre-trained passive device frequency response reconstruction neural network to obtain the predicted S-parameters of the cascaded structure. In S2, both the passive device frequency response sub-neural network and the passive device frequency response reconstruction neural network include: an input layer, a hidden layer, and an output layer connected in sequence. The hidden layer includes a shared feature layer, which is a fully connected network. The fully connected network in this embodiment can be replaced by a Convolutional Neural Network (CNN). In specific implementations, the component parameters can be rearranged or mapped to a 2D topological matrix (e.g., a 10×10 matrix) as input to the CNN. By automatically extracting the spatial and topological features between parameters through convolutional layers, the modeling accuracy of complex coupling relationships in cascaded circuits can be improved. The distributed neural network architecture composed of the core passive device frequency response sub-neural network and the passive device frequency response reconstruction neural network remains unchanged, still achieving the advantages of high design efficiency and strong applicability.

[0032] like Figure 5 As shown, the detailed design of the passive device frequency response sub-neural network is as follows: Input layer: Π-type input is 3-dimensional (C1, L1, C2), where C1 and C2 are connected in parallel between the input / output terminals and ground, and L1 is connected in series between the input and output. The whole structure looks like the Greek letter Π. T-type input is 3-dimensional (L2, L3, C3), where L2 and L3 are connected in series on the path between the input and output terminals, and C3 is connected in parallel between the intermediate node of the path between the input and output terminals and ground. The whole structure looks like the English letter T.

[0033] Hidden layer: Employs a shared feature layer. Shared feature layer: 3-layer fully connected network (256→128→64 neurons) to extract common topological features.

[0034] Output layer: 6N f Dimension (6N) f The number of frequency points is 100, ranging from 1 to 10 GHz, with output S. 11 S 21 S 22 (The values ​​of the real and imaginary parts).

[0035] Figure 5 In this model, the layers are: Input Layer -> Hidden Layer -> Output Layer. ReLU is the activation function. X1, X2, and X3 represent the characteristics of the input data, that is, the parameter values ​​of the three components in the Π-type / T-type circuit.

[0036] y1, y2, y3, y4, y5, y6 represent the output values ​​of the network, that is, the real and imaginary parts of the S11, S21, and S22 parameters of the Π-type / T-type circuit, which correspond one-to-one with the real part of S11, the imaginary part of S11, ..., the imaginary part of S21.

[0037] Input Layer is the input layer, Hidden Layer is the hidden layer, and Output Layer is the output layer; n[0]=3 indicates that the dimension of the input layer is 3; n[1]=512 indicates that the number of neurons in the first hidden layer is 512, n[2]=1024 indicates that the number of neurons in the second hidden layer is 1024, n[3]=2048 indicates that the number of neurons in the third hidden layer is 2048, n[4]=1024 indicates that the number of neurons in the fourth hidden layer is 1024, and n[5]=512 indicates that the number of neurons in the fifth hidden layer is 512. n[6]=6 indicates that the dimension of the output layer is 6.

[0038] In the passive device frequency response sub-neural network, the device parameters are transmitted to the shared feature layer through the input layer to extract the topological common features of the device parameters; the topological common features are linearly transformed through the output layer to obtain the predicted S-parameters corresponding to different passive device topologies.

[0039] In the passive device frequency response reconstruction neural network, the predicted S-parameters corresponding to different passive device topologies are concatenated to obtain a cascaded feature vector; the cascaded feature vector is transmitted to the shared feature layer through the input layer to extract the cascaded coupling features of the cascaded feature vector; the cascaded coupling features are linearly transformed through the output layer to obtain the predicted S-parameters of the cascaded structure.

[0040] The passive device frequency response sub-neural network and the passive device frequency response reconstruction neural network together constitute a distributed neural network architecture. A Transformer network with a frequency domain attention mechanism can be used to replace the distributed neural network architecture, which can assign different attention weights to the S-parameter features at different frequency points, thereby enhancing the learning of key features in the high-frequency band (>6GHz). This is expected to further reduce the high-frequency prediction error to within 0.3dB. The model hyperparameters of the distributed neural network architecture can be flexibly adjusted within a certain range without affecting the core fitting ability. For example, the number of convolutional layers can be selected between 8 and 16, the number of fully connected layers can be adjusted between 3 and 6, and the activation function can be replaced by GELU, etc. like Figure 6 As shown, the detailed design of the passive device frequency response reconstruction neural network is as follows, based on the passive device frequency response sub-neural network: Input layer: 12N f S-parameter splicing of π-type and T-type subnetworks, N f (This refers to the number of frequency points, 100 frequency points in total from 1-10GHz). Hidden layer: Shared feature layer structure, enhancing cascaded coupled feature learning. Output layer: 6N f Dimension (total S-parameters of the cascaded circuit); Figure 6In the diagram, Re(S11) is the real part of the S11 parameter, Im(S11) is the imaginary part of the S11 parameter, Re(S21) is the real part of the S21 parameter, Im(S21) is the imaginary part of the S21 parameter, Re(S22) is the real part of the S22 parameter, and Im(S22) is the imaginary part of the S22 parameter. The X values ​​in the two passive device frequency response sub-neural networks on the left represent the three components in the Π-type and T-type circuits, respectively. Y represents the real and imaginary parts of the S11, S21, and S22 parameters of the Π-type and T-type circuits. The passive device frequency response reconstruction neural network on the right takes the outputs of the two passive device frequency response sub-neural networks as input, i.e., it has 12 feature inputs (X1, X2, X3….X11, X12), and the output is the real and imaginary parts of the S11, S21, and S22 parameters of the cascaded circuit after the Π-type and T-type circuits are cascaded.

[0041] In S2, the pre-training of the passive device frequency response sub-neural network includes the following steps: sampling within the component parameter range of each preset passive device topology to generate multiple sets of component parameter combinations; obtaining the real S-parameters of each set of component parameter combinations in the target frequency band through electromagnetic simulation; and constructing a training dataset by mapping the component parameter combinations and the real S-parameters. The real S-parameters include the amplitude and phase information of parameters S11, S21, and S22. During training, the component parameter combination is used as the input to the passive device frequency response sub-neural network, the corresponding real S-parameters are used as the training target, and the predicted S-parameters are used as the output. The weights of the passive device frequency response sub-neural network are optimized by minimizing the corresponding complex loss function. The training continues until convergence to obtain the pre-trained passive device frequency response sub-neural network. The complex loss function of the passive device frequency response sub-neural network consists of the amplitude error and phase error between the real S-parameters and the predicted S-parameters.

[0042] In S2, the pre-training of the passive device frequency response reconstruction neural network includes the following steps: The total S-parameters of the cascaded structure are obtained by combining the parameters of each group of components in the target frequency band and using electromagnetic simulation. During training, the actual S-parameters are used as the input to the passive device frequency response reconstruction neural network, the total S-parameters of the corresponding cascaded structure are used as the training target, and the predicted S-parameters of the cascaded structure are used as the output. The weights of the passive device frequency response reconstruction neural network are optimized by minimizing the corresponding complex loss function, and the training is completed until convergence to obtain the pre-trained passive device frequency response reconstruction neural network. The complex loss function of the passive device frequency response reconstruction neural network consists of the amplitude error and phase error between the total S-parameters of the cascaded structure and the predicted S-parameters of the cascaded structure.

[0043] The complex loss function formulas for the passive device frequency response sub-neural network and the passive device frequency response reconstruction neural network are as follows:

[0044] in, This is the loss value. The number of samples; In the passive device frequency response sub-neural network Let S be the predicted S-parameters for the i-th sample. The true S-parameters of the i-th sample; In passive device frequency response reconstruction neural networks Let S be the predicted S-parameters of the cascade structure of the i-th sample. The total S-parameters of the cascaded structure; It is the square of the amplitude error. The square of the phase error. This is the weighting coefficient for the phase error.

[0045] The optimizer chosen is the Adam optimizer, which incorporates momentum ( With adaptive learning rate characteristics, the initial learning rate is set to 0.001, and weight decay is introduced. To suppress overfitting, regularization measures include early stopping (terminating training after 10 consecutive epochs without improvement) and gradient clipping (norm threshold of 1.0) to prevent overfitting and gradient explosion.

[0046] like Figure 8 As shown, the fitting curve of the step-by-step neural network architecture is displayed. By comparing the actual S-parameters obtained from electromagnetic simulation with the S-parameters predicted by the neural network, this graph intuitively verifies the accuracy and reliability of the neural network model proposed in this invention in the performance modeling of passive devices. The horizontal axis, Frequency, represents the frequency in dB; True indicates the actual S-parameters, and Predicted indicates the predicted S-parameters. The graph includes a comparison of the amplitudes of three key S-parameters, all with frequency (GHz) on the horizontal axis and amplitude (dB) on the vertical axis. Figure 8 In (a), the S11 Magnitude Comparison illustrates the fitting of the input reflection coefficient. The high degree of overlap between the "True" and "Predicted" curves in the figure indicates that the neural network can accurately learn and predict the input port matching characteristics of the circuit at different frequencies. Figure 8In (b), the S21 Magnitude Comparison demonstrates the fit of the forward transmission coefficients. This is a core metric for evaluating filter performance. The two curves fit closely in the passband, transition band, and stopband, proving that the neural network can accurately reproduce the filter's core frequency response, including key characteristics such as insertion loss and stopband rejection. Figure 8 In (c), S22Magnitude Comparison demonstrates the fit of the output reflection coefficient. The high consistency between its predicted curve and the true curve further validates the predictive ability of the neural network model for output port matching behavior.

[0047] S3, perform global optimization on the predicted S-parameters of the cascaded structure to obtain the optimal combination of component parameters corresponding to the optimal predicted S-parameters.

[0048] Global optimization employs the particle swarm optimization algorithm, the process of which is as follows: Figure 7 As shown, the PSO algorithm continuously generates and adjusts component parameter combinations, performs sub-network prediction, obtains S-parameters, and calculates the fitness of the total response obtained through the cascaded network, thereby optimizing the objective. Specifically, it includes the following steps: Optimization problem modeling: (1) The objective function is as follows:

[0049] in For the component parameters, α, β, and γ are different weighting coefficients. Passband_loss: passband insertion loss, the amplitude value of S21 in the passband; stopband_rejection: stopband rejection, the amplitude value of S21 in the stopband; transition_slope: transition band slope, the steepness of the S21 curve from the passband to the stopband.

[0050] (2) The upper and lower limits of the preset component parameters are used as constraints, as shown in the following formula:

[0051] in, Let be the capacitance parameters of the i-th capacitor. This is the lower limit of the capacitance parameter. This represents the upper limit of the capacitor parameter. For inductance parameters, This represents the upper limit of inductance parameters. This represents the lower limit of the inductance parameter.

[0052] The steps of the PSO algorithm are as follows: An objective function is constructed with the goal of minimizing the error between the predicted S-parameters and the target S-parameters of the cascaded structure. Specifically, this objective function is achieved by weighting the amplitude errors between the predicted S-parameters and the preset fixed threshold in the passband and stopband, respectively, and the amplitude errors between the predicted S-parameters and the target S-parameters in the transition band. The specific formula is as follows:

[0053] in, E rror is the total error. f For frequency, Passband For passband, Stopband For the stopband, Transition Transition zone S 21 To predict the magnitude of the S-parameter, Target For the magnitude of the target S-parameter, a fixed threshold of 3dB is preset in the passband and a fixed threshold of 20dB is preset in the stopband.

[0054] Population initialization: 4096 6-dimensional parameter vectors (particles) are randomly generated, representing potential component parameter combinations.

[0055] Proxy model prediction: The neural network is reconstructed by calling the frequency response of the trained passive device to predict the S-parameters corresponding to each particle.

[0056] Fitness calculation: The error is calculated using the formula that specifically implements the above objective function, and this error is used as the fitness value.

[0057] Particle Update: Update particle velocity and position according to the formula:

[0058] in, For inertial weights, As a learning factor, and It is a random number. For the new velocity of the particles, The particle's current velocity. For the optimal position of an individual, This represents the particle's current position. This is the globally optimal position.

[0059] Iteration Termination: After 800-1200 iterations, the particle with the best fitness is output, which is the optimal combination of component parameters.

[0060] The PSO algorithm is invoked in the Terminal, and the execution command is: / opt / DP / Pytorch / Mypytorch / bin / python3.11 / home / HDD / XDIC09 / Projects / Pytorch / CNNClassifier / 1.py.

[0061] The user then enters the cutoff frequency in the Terminal, and can specify the target frequency via the command line. In this embodiment, the Particle Swarm Optimization (PSO) algorithm can be replaced by a Genetic Algorithm (GA). In this approach, component parameters are encoded as gene sequences, and the population is iteratively evolved by simulating selection, crossover, and mutation operations in natural selection, ultimately searching for the optimal parameter combination. This approach is particularly suitable for optimization scenarios with discrete or mixed parameters. Simulated Annealing (SA) can also be used for global optimization. SA simulates the physical annealing process, and by introducing temperature parameters and strategies to control its decay, it effectively avoids premature convergence of the algorithm and may exhibit better robustness when dealing with high-dimensional parameter spaces (e.g., dimension > 8). Depending on the specific optimization efficiency and accuracy requirements, Imperial Competition, hybrid algorithms of PSO and gradient descent, etc., can all be used as alternatives to PSO. The core idea of ​​all these alternatives is to minimize the constructed multi-objective cost function through a global search strategy.

[0062] Finally, a simulation verification process is performed: the optimal combination of component parameters output by the PSO algorithm is substituted into ADS, a Π-type, T-type, or cascaded circuit is built, and a 1-10GHz frequency domain simulation is executed to verify whether the S-parameters meet the target. Figure 9 and Figure 10 As shown, the horizontal axis freq represents frequency in GHz, the vertical axis dB(S(2,1)) represents the amplitude of parameter S21, and the vertical axis dB(S(1,1)) represents the amplitude of parameter S11. Simulation results show that when the cutoff frequency of the reverse-designed low-pass filter is set to 4 GHz, the insertion loss is less than 0.736 dB; when the cutoff frequency is set to 2 GHz, the cutoff frequency condition is met. Electromagnetic simulation results verify the feasibility and effectiveness of the method in this embodiment for designing high-performance passive devices.

[0063] The ADS simulation software used in this embodiment can be replaced with other mainstream electromagnetic simulation tools, such as HFSS (High Frequency Structure Simulator) or CST (Computer Simulation Technology). By calling the API interfaces provided by these software programs and performing co-simulation with MATLAB, the same data generation accuracy and process automation can be maintained. This invention is not limited to lumped parameters such as capacitors and inductors. The method is also applicable to the integrated design of distributed parameter passive devices (such as microstrip lines and couplers). In this alternative, the input parameters of the sub-network need to be adjusted accordingly to physical dimension parameters such as microstrip linewidth, length, and dielectric layer thickness. Its core model and methodology, namely predicting performance through a distributed neural network and optimizing through a global algorithm, are still applicable and can achieve effective optimization design.

[0064] This embodiment has the following advantages: (1) It achieves multiple topology combinations through passive device frequency response sub-neural network and passive device frequency response performance reconstruction neural network. The dimensionality reduction of the main network and sub-network results in low space and time resource consumption and high design efficiency. (2) Strong topology flexibility and applicability: Existing technologies only support a single fixed topology. This invention achieves multiple topology combinations through passive device frequency response sub-neural network and passive device frequency response performance reconstruction neural network. When the sub-neural network topology changes, only the sub-neural network needs to be retrained. The reconstruction neural network does not need to be retrained, which greatly increases the usability of this invention and improves the reverse design efficiency. (3) In the particle swarm algorithm, a highly specialized and composite cost function (objective_function) is designed, which comprehensively considers multiple key filter performance indicators such as in-band performance (passband ripple, passband gain deviation), transition band performance (error at cutoff frequency), and overall fitting degree (full-band amplitude error, phase error).

[0065] like Figure 2 As shown, this embodiment also provides a passive device synthesis system based on frequency response performance segmentation and reconstruction, including: The component parameter acquisition module is used to acquire component parameters for electromagnetic simulation of various passive component topologies; The segmentation and reconstruction module is used to input the component parameters into the corresponding pre-trained passive device frequency response sub-neural network to obtain the predicted S-parameters of each passive device topology, and to concatenate the predicted S-parameters of each passive device topology and input them into the pre-trained passive device frequency response reconstruction neural network to obtain the predicted S-parameters of the cascaded structure. The global optimization module is used to perform global optimization on the predicted S-parameters of the cascaded structure to obtain the optimal combination of component parameters corresponding to the optimal predicted S-parameters.

[0066] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0067] This embodiment also provides a computer device, which includes a processor and a memory. The memory is used to store a computer program (in this embodiment, the computer program includes computational components and iterative components, capable of model calculation and model updating). The computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to realize the corresponding method flow or corresponding function. The processor described in this embodiment can be used in the operation of a passive device synthesis method based on frequency response performance segmentation and reconstruction.

[0068] This embodiment also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the passive device synthesis method based on frequency response performance segmentation and reconstruction in the above embodiment.

[0069] This embodiment also provides a computer program product, which includes a computer program that, when executed by a processor, implements the corresponding steps of the passive device synthesis method based on frequency response performance segmentation and reconstruction in the above embodiment.

[0070] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0071] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0072] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0073] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A passive device synthesis method based on frequency response performance segmentation and reconstruction, characterized in that, Includes the following steps: Component parameters for electromagnetic simulation of various passive device topologies; The component parameters are input into a pre-trained passive device frequency response sub-neural network to obtain the predicted S-parameters of each passive device topology. The predicted S-parameters of each passive device topology are then concatenated and input into a pre-trained passive device frequency response reconstruction neural network to obtain the predicted S-parameters of the cascaded structure. The predicted S-parameters of the cascaded structure are globally optimized to obtain the optimal combination of component parameters corresponding to the optimal predicted S-parameters.

2. The passive device synthesis method based on frequency response performance segmentation and reconstruction according to claim 1, characterized in that, Both the pre-trained passive device frequency response sub-neural network and the pre-trained passive device frequency response reconstruction neural network include: an input layer, a hidden layer, and an output layer connected in sequence. The hidden layer includes a shared feature layer, which is a fully connected network.

3. The passive device synthesis method based on frequency response performance segmentation and reconstruction according to claim 2, characterized in that, The component parameters are input into a pre-trained passive device frequency response sub-neural network to obtain the predicted S-parameters of each passive device topology, specifically including: The component parameters are transmitted to the shared feature layer through the input layer to extract the topological common features of the component parameters; The common topological features are linearly transformed through the output layer to obtain the predicted S-parameters corresponding to different passive device topologies.

4. The passive device synthesis method based on frequency response performance segmentation and reconstruction according to claim 2, characterized in that, The predicted S-parameters of each passive device topology are concatenated and input into a pre-trained passive device frequency response reconstruction neural network to obtain the predicted S-parameters of the cascaded structure, specifically including: The predicted S-parameters corresponding to the topology of each passive device are concatenated to obtain a cascaded feature vector. The cascaded feature vector is transmitted to the shared feature layer through the input layer to extract the cascaded coupling features of the cascaded feature vector; The cascaded coupling features are linearly transformed through the output layer to obtain the predicted S-parameters of the cascaded structure.

5. The passive device synthesis method based on frequency response performance segmentation and reconstruction according to claim 1, characterized in that, Global optimization is performed on the predicted S-parameters of the cascaded structure to obtain the optimal combination of component parameters corresponding to the optimal predicted S-parameters, specifically including: An objective function is constructed with the goal of minimizing the error between the predicted S-parameters and the target S-parameters of the cascaded structure. The objective function is the amplitude error between the predicted S-parameters and the preset fixed threshold in the passband and stopband, respectively, and the amplitude error between the predicted S-parameters and the target S-parameters in the transition band, which is constructed by weighted sum of squares. The upper and lower limits of preset component parameters are used as constraints. Based on the particle swarm optimization algorithm, the objective function is minimized and iteratively optimized to obtain the optimal combination of component parameters corresponding to the optimal predicted S-parameters.

6. The passive device synthesis method based on frequency response performance segmentation and reconstruction according to claim 1, characterized in that, The training process of the pre-trained passive device frequency response sub-neural network includes the following steps: Sampling is performed within the component parameter range of each preset passive device topology to generate multiple sets of component parameter combinations. The true S-parameters of each set of component parameter combinations in the target frequency band are obtained through electromagnetic simulation. A training dataset is constructed by mapping the component parameter combinations and the true S-parameters. The true S-parameters include the amplitude and phase information of S11, S21 and S22 parameters. During training, the combination of component parameters is used as the input of the passive device frequency response sub-neural network, the corresponding real S-parameters are used as the training target, and the predicted S-parameters of the passive device topology are used as the output. The weights of the passive device frequency response sub-neural network are optimized by minimizing the corresponding complex loss function, and the training is completed until convergence to obtain the pre-trained passive device frequency response sub-neural network. The complex loss function of the passive device frequency response sub-neural network consists of the amplitude error and phase error between the actual S-parameters and the predicted S-parameters of the passive device topology.

7. The passive device synthesis method based on frequency response performance segmentation and reconstruction according to claim 6, characterized in that, The training process of the pre-trained passive device frequency response reconstruction neural network includes the following steps: The total S-parameters of the cascaded structure are obtained by combining the parameters of each group of components in the target frequency band and using electromagnetic simulation. During training, the actual S-parameters are used as the input to the passive device frequency response reconstruction neural network, the total S-parameters of the corresponding cascaded structure are used as the training target, and the predicted S-parameters of the cascaded structure are used as the output. The weights of the passive device frequency response reconstruction neural network are optimized by minimizing the corresponding complex loss function, and the training is completed until convergence to obtain the pre-trained passive device frequency response reconstruction neural network. The complex loss function of the passive device frequency response reconstruction neural network consists of the amplitude error and phase error between the total S-parameters of the cascaded structure and the predicted S-parameters of the cascaded structure.

8. A passive device synthesis system based on frequency response performance segmentation and reconstruction, characterized in that, include: The component parameter acquisition module is used to acquire component parameters for electromagnetic simulation of various passive component topologies; The segmentation and reconstruction module is used to input the component parameters into a pre-trained passive device frequency response sub-neural network to obtain the predicted S-parameters of each passive device topology, and to concatenate the predicted S-parameters of each passive device topology and input them into the pre-trained passive device frequency response reconstruction neural network to obtain the predicted S-parameters of the cascaded structure. The global optimization module is used to perform global optimization on the predicted S-parameters of the cascaded structure to obtain the optimal combination of component parameters corresponding to the optimal predicted S-parameters.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the passive device synthesis method based on frequency response performance segmentation and reconstruction as described in any one of claims 1 to 7.

10. A computer program product, the computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the passive device synthesis method based on frequency response performance segmentation and reconstruction as described in any one of claims 1 to 7.

Images