Radio frequency filter parameterization inverse design method and device based on loop network

Abstract

The invention discloses a radio frequency filter parameterization inverse design method and device based on a loop network, and the method comprises the steps: constructing a loop network architecture composed of a forward network and a reverse network, and collecting an initial training data set comprising radio frequency filter element parameters and corresponding design indexes; training a forward network based on the initial training data set; newly added radio frequency filter element parameters and data of corresponding design indexes are generated by using the trained forward network, and the initial training data set is expanded; training a reverse network based on the expanded initial training data set, constructing a loss function in combination with a real design index, and enabling the reverse network to have the capability of mapping from the design index to an element parameter through iterative optimization; and inputting a target design index into the trained reverse network, and outputting the radio frequency filter element parameters conforming to the target design index by the reverse network to complete the reverse design. The design period is shortened, the design freedom degree and precision are improved, and the method is suitable for various radio frequency device inverse design scenes.

Description

Technical Field

[0001] This invention belongs to the field of RF filter and deep learning fusion technology, specifically relating to a parametric inverse design method and apparatus for RF filters based on recurrent networks. Background Technology

[0002] With the rapid development of satellite communication, millimeter-wave technology, and wireless communication systems, radio frequency (RF) filters, as core components, must simultaneously meet requirements for amplitude-frequency characteristics, group delay, and other phase characteristics to avoid modulation signal distortion, EVM (Error Vector Magnitude) degradation, and ACLR (Adjacent Channel Leakage Ratio) decline. Their design accuracy and efficiency directly affect the system's signal integrity and reliability. However, current mainstream design methods have significant limitations and are difficult to adapt to the rapidly evolving needs of the industry. Traditional design methods rely heavily on human experience, requiring either repeated parameter adjustments using simulation software or iterative solutions through optimization algorithms. Both suffer from long design cycles and high computational costs. While deep learning-based forward fitting methods can replace some simulation processes, they still require iterative parameter finding through optimization algorithms and cannot directly output results, resulting in limited efficiency improvements. Deep learning-based inverse design methods attempt to directly construct a mapping between design metrics and component parameters, but are hampered by the multiple solutions to the inverse problem, requiring massive amounts of data and manual cleaning, leading to high costs, significant data waste, and poor model generalization.

[0003] In addition, the scarcity of spectrum resources has led to an increase in the target frequency band of filters and a larger component size relative to the wavelength. Coupled with the need for multiphysics simulation, this has further increased the computational complexity and extended the design verification time. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method and apparatus for parametric inverse design of radio frequency filters based on cyclic networks, so as to solve or partially solve the problems mentioned in the background art.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for parametric inverse design of radio frequency filters based on cyclic networks, comprising: Step 1: Construct a recurrent network architecture consisting of a forward network and a reverse network. The forward network is used to realize data mapping and dataset expansion, and the reverse network is used to realize the mapping from design specifications to RF filter component parameters. The training process of the reverse network depends on the prediction results output by the forward network for optimization. Step 2: Use optimization algorithms combined with forward solvers or simulation software to filter effective data and collect an initial training dataset including RF filter component parameters and corresponding design specifications. Step 3: Train the forward network based on the initial training dataset, so that the forward network has the ability to map the parameters of the radio frequency filter components to the corresponding design specifications; Step 4: Use the trained forward network to generate data on the parameters of newly added RF filter components and their corresponding design specifications to expand the initial training dataset; Step 5: Train the inverse network based on the expanded initial training dataset. During training, input the component parameters output by the inverse network into the trained forward network to obtain the predicted design metrics. Combine the actual design metrics to construct a loss function. Through iterative optimization, enable the inverse network to map design metrics to component parameters. Step 6: Input the target design index into the trained reverse network, and the reverse network outputs the RF filter element parameters that meet the target design index to complete the reverse design.

[0006] As a preferred embodiment of the parametric inverse design method for RF filters based on cyclic networks, in step 1, both the forward network and the inverse network adopt the Vision Permutator network architecture; the Vision Permutator network architecture includes an input block processing module, a rearrangement module, and a result output module; The input block processing module divides the network input into sub-blocks according to preset rules, and performs zero-padding on the sub-blocks of different types of RF filter element parameters to unify the dimensions; The rearrangement module includes multiple rearrangement sub-blocks, each of which includes a Permute-MLP layer for processing spatial information, a Channel-MLP layer for mixing channel information, and two LayerNorm layers for normalization. The output module maps the features extracted by the network to the final output result through a fully connected layer and a global average pooling layer.

[0007] As a preferred scheme for the parametric inverse design method of RF filter based on cyclic networks, in step 2, the optimization algorithm is the Cuckoo Optimization Algorithm. When screening the initial training data through the optimization algorithm, the design indicators of the RF filter are used as constraints to eliminate invalid component parameter data that do not meet the indicators.

[0008] As a preferred embodiment of the parametric inverse design method for RF filters based on cyclic networks, in steps 1 to 6, the design index is the S21 parameter of the RF filter. In step 3, the input of the forward network is the component parameters of the RF filter, and the output is the S21 parameter under the corresponding component parameters; In step 5, the input of the reverse network includes the target S21 parameter of the RF filter and a string of random numbers. The random numbers are used to distinguish different component parameter solutions corresponding to the same S21 parameter. The output of the reverse network is the component parameters of the RF filter.

[0009] As a preferred embodiment of the parametric inverse design method for RF filters based on recurrent networks, in step 5, the loss function includes a first error term and a second error term. The first error term is the error between the predicted S21 parameters output by the forward network and the actual S21 parameters, and the second error term is the error between the component parameters output by the inverse network and the actual component parameters. During training, the weights of the first error term and the second error term are dynamically adjusted so that the weight of the first error term gradually becomes higher than the weight of the second error term.

[0010] As a preferred embodiment of the parametric inverse design method for RF filters based on recurrent networks, the processing of the Permute-MLP layer in the rearrangement module satisfies the following mathematical model: Assume the output of the input block processing module is:

[0011] In the formula, For the sample size, For the number of channels, The parameter dimensions for each sample; For input Perform channel grouping, let the number of groups be . The grouped data is In the formula, , The height dimension after grouping; Grouped data By performing dimensional permutation, we obtain And then By reshaping the dimensions, we can obtain ; The height direction is encoded using a fully connected layer to obtain the height direction output:

[0012] In the formula, Here is the weight matrix of the fully connected layer in the height direction. For the height-direction fully connected layer bias term; The width direction output is obtained sequentially. Output in the direction of the channel ,in Weight matrix of the fully connected layer in the width direction With bias term , Weight matrix of fully connected layer in the corresponding channel direction With bias term ; The output of the Permute-MLP layer is Then it is processed through a fully connected layer to obtain In the formula, This is the final weight matrix of the fully connected layer. This is the bias term for the final fully connected layer.

[0013] As a preferred embodiment of the parametric inverse design method for RF filters based on cyclic networks, the rearrangement module includes a Channel-MLP layer comprising two fully connected layers and a GELU activation function. The Channel-MLP layer is used to mix channel information by performing a fully connected operation on the channels. Let the input of the Channel-MLP layer be First, perform a dimensionality transformation on the input, placing the channel dimension as the last dimension, to obtain... get Next, parallel fully connected layer processing is performed, mixing the channel information for each W-dimensional element to obtain... In the formula, This is the weight matrix of the first fully connected layer. This is the bias term for the first fully connected layer; Next, the GELU activation function is applied to... Perform a nonlinear transformation to obtain =GELU ( ); Finally, it is processed through a second fully connected layer to obtain... In the formula, This is the weight matrix for the second fully connected layer. This is the bias term for the second fully connected layer, and the second activation function. =GELU ( Next, the matrix dimensions are restored to their original form to obtain the output of the Channel-MLP layer. .

[0014] As a preferred embodiment of the parametric inverse design method for RF filters based on cyclic networks, the RF filter component parameters include inductance parameters and capacitance parameters; the inductor is a spiral inductor, including the length, width, line width, and spacing of the inductor; The capacitor is a MIM capacitor, including the length and width of the capacitor metal sheet.

[0015] As a preferred embodiment of the parametric inverse design method for RF filters based on cyclic networks, the length of the random number input to the inverse network is consistent with the length of the sub-block of the input block processing module; During the input processing of the reverse network, the target S21 parameter is first processed into a fixed-dimensional vector according to a preset sampling rule, and then concatenated with a random number vector before being input to the input block processing module. Through the randomness of the random number, the reverse network outputs multiple different combinations of component parameters for the same S21 parameter input.

[0016] In a second aspect, the present invention provides a parameterized inverse design apparatus for radio frequency filters based on cyclic networks, employing the parameterized inverse design method for radio frequency filters based on cyclic networks according to the first aspect or any possible implementation thereof, including: A recurrent network construction unit is used to construct a recurrent network architecture consisting of a forward network and a reverse network. The forward network is used to realize data mapping and dataset expansion, and the reverse network is used to realize the mapping from design indicators to RF filter component parameters. The training process of the reverse network depends on the prediction results output by the forward network for optimization. The data collection unit is used to filter effective data using optimization algorithms combined with forward solvers or simulation software, and to collect an initial training dataset including RF filter component parameters and corresponding design specifications. A forward network training unit is used to train the forward network based on the initial training dataset, so that the forward network has the ability to map the parameters of the radio frequency filter components to the corresponding design specifications; The dataset expansion unit is used to expand the initial training dataset by generating data on the parameters of newly added radio frequency filter components and their corresponding design specifications using the trained forward network. The inverse network training unit is used to train the inverse network based on the expanded initial training dataset. During training, the component parameters output by the inverse network are input into the trained forward network to obtain the predicted design index. The loss function is constructed by combining the actual design index and iteratively optimizing the inverse network to enable it to map the design index to the component parameters. The design output unit is used to input the target design index into the trained inverse network, and the inverse network outputs the RF filter element parameters that meet the target design index, thus completing the inverse design.

[0017] As a preferred embodiment of the parametric inverse design device for RF filters based on cyclic networks, the forward network and the inverse network in the cyclic network construction unit both adopt the Vision Permutator network architecture; the Vision Permutator network architecture includes an input block processing module, a rearrangement module, and a result output module; The input block processing module divides the network input into sub-blocks according to preset rules, and performs zero-padding on the sub-blocks of different types of RF filter element parameters to unify the dimensions; The rearrangement module includes multiple rearrangement sub-blocks, each of which includes a Permute-MLP layer for processing spatial information, a Channel-MLP layer for mixing channel information, and two LayerNorm layers for normalization. The output module maps the features extracted by the network to the final output result through a fully connected layer and a global average pooling layer.

[0018] As a preferred scheme for the parameterized inverse design method of RF filter based on cyclic networks, the optimization algorithm is the Cuckoo Optimization Algorithm. When screening the initial training data through the optimization algorithm, the design indicators of the RF filter are used as constraints to eliminate invalid component parameter data that do not meet the indicators.

[0019] As a preferred embodiment of the parameterized inverse design device for RF filters based on cyclic networks, the design parameters are the S21 parameters of the RF filter. The input of the forward network is the component parameters of the radio frequency filter, and the output is the S21 parameter under the corresponding component parameters; The input of the reverse network includes the target S21 parameter of the RF filter and a string of random numbers. The random numbers are used to distinguish different component parameter solutions corresponding to the same S21 parameter. The output of the reverse network is the component parameters of the RF filter.

[0020] As a preferred embodiment of the RF filter parameterization inverse design device based on a recurrent network, in the inverse network training unit, the loss function includes a first error term and a second error term. The first error term is the error between the predicted S21 parameters output by the forward network and the actual S21 parameters, and the second error term is the error between the component parameters output by the inverse network and the actual component parameters. During training, the weights of the first error term and the second error term are dynamically adjusted so that the weight of the first error term gradually becomes higher than the weight of the second error term.

[0021] As a preferred embodiment of the parametric inverse design device for RF filters based on cyclic networks, the processing of the Permute-MLP layer in the rearrangement module satisfies the following mathematical model: Assume the output of the input block processing module is:

[0022] In the formula, For the sample size, For the number of channels, The parameter dimensions for each sample; For input Perform channel grouping, let the number of groups be . The grouped data is In the formula, , The height dimension after grouping; Grouped data By performing dimensional permutation, we obtain And then By reshaping the dimensions, we can obtain ; The height direction is encoded using a fully connected layer to obtain the height direction output:

[0023] In the formula, Here is the weight matrix of the fully connected layer in the height direction. For the height-direction fully connected layer bias term; The width direction output is obtained sequentially. Output in the direction of the channel ,in Weight matrix of the fully connected layer in the width direction With bias term , Weight matrix of fully connected layer in the corresponding channel direction With bias term ; The output of the Permute-MLP layer is Then it is processed through a fully connected layer to obtain In the formula, This is the final weight matrix of the fully connected layer. This is the bias term for the final fully connected layer.

[0024] As a preferred embodiment of the parametric inverse design device for RF filters based on cyclic networks, the rearrangement module includes a Channel-MLP layer comprising two fully connected layers and a GELU activation function. The Channel-MLP layer is used to mix channel information by performing a fully connected operation on the channels. Let the input of the Channel-MLP layer be First, perform a dimensionality transformation on the input, placing the channel dimension as the last dimension, to obtain... get Next, parallel fully connected layer processing is performed, mixing the channel information for each W-dimensional element to obtain... In the formula, This is the weight matrix of the first fully connected layer. This is the bias term for the first fully connected layer; Next, the GELU activation function is applied to... Perform a nonlinear transformation to obtain =GELU ( ); Finally, it is processed through a second fully connected layer to obtain... In the formula, This is the weight matrix for the second fully connected layer. This is the bias term for the second fully connected layer, and the second activation function. =GELU ( Next, the matrix dimensions are restored to their original form to obtain the output of the Channel-MLP layer. .

[0025] As a preferred embodiment of the parametric inverse design device for RF filters based on cyclic networks, the RF filter element parameters include inductance parameters and capacitance parameters; the inductor is a spiral inductor, including the length, width, line width, and spacing of the inductor; The capacitor is a MIM capacitor, including the length and width of the capacitor metal sheet.

[0026] As a preferred embodiment of the RF filter parameterization inverse design device based on a cyclic network, the length of the random number input to the inverse network is consistent with the length of the sub-block of the input block processing module; During the input processing of the reverse network, the target S21 parameter is first processed into a fixed-dimensional vector according to a preset sampling rule, and then concatenated with a random number vector before being input to the input block processing module. Through the randomness of the random number, the reverse network outputs multiple different combinations of component parameters for the same S21 parameter input.

[0027] Thirdly, 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, when executing the program, implements the first aspect or any possible implementation thereof, a parameterized inverse design method for a radio frequency filter based on a cyclic network.

[0028] Fourthly, the present invention provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform steps in the cyclic network-based parameterized inverse design method for radio frequency filters according to the first aspect or any possible implementation thereof.

[0029] From the above description, the beneficial effects of the present invention can be summarized as follows: After training, the recurrent network constructed by this invention allows users to directly output the required RF filter component parameters within seconds by simply inputting the target design specifications. This eliminates the need for repeated manual parameter adjustments or reliance on empirical formulas for iteration, completely changing the lengthy process of traditional design methods, significantly reducing the dependence on human experience in design, and adapting to the rapidly developing needs of fields such as integrated circuits.

[0030] Unlike traditional optimization algorithms and deep learning-based forward fitting methods, this invention directly fits the mapping relationship between design indicators and component parameters through an inverse network. With the assistance of a forward network, it eliminates the need for repeated iterations to approximate the optimal solution using additional optimization algorithms. This fundamentally solves the pain point of existing solutions requiring multiple iterations and failing to obtain a direct solution, thus significantly improving design efficiency.

[0031] This invention employs cascading random number encoding at the input of the inverse network, effectively distinguishing different component parameter solutions corresponding to the same design specification and avoiding the problem of model non-convergence caused by multiple solutions. Simultaneously, the initial data collection utilizes the Cuckoo Optimization algorithm to filter valid data, combined with the forward network to rapidly expand the dataset, eliminating the need for massive amounts of raw data and manual cleaning, thus reducing data waste and lowering the time and labor costs of data collection. Through random number encoding, the inverse network can output multiple compliant component parameter combinations for the same design specification, breaking the limitation of the traditional "single solution" in inverse design methods, providing engineers with more design options and increasing design freedom. Furthermore, the recurrent network architecture is not limited to specific types of filters and can adapt to the parametric inverse design of various devices such as IPD filters and patch antennas, resulting in stronger model generalization capabilities.

[0032] During training, the loss function of the inverse network combines the error between the predicted S21 parameters and the actual S21 parameters, as well as the component parameter error, and gradually increases the weight of the S21 parameter error to ensure that the optimization is centered on the accuracy of the design specifications, avoiding the situation where the component parameters are qualified but the specifications are not. In actual verification, the error between the S21 parameters corresponding to the output parameters and the target specifications can be controlled within a low range, which can accurately meet the phase characteristics requirements such as amplitude-frequency characteristics and group delay, and effectively avoid problems such as modulation signal distortion, EVM deterioration and ACLR decrease.

[0033] The network architecture adopted in this invention has low parameter count and high computational efficiency. Furthermore, the forward network can replace traditional simulation software and forward solvers, avoiding the problem of increased simulation complexity caused by the increase in the ratio of component size to wavelength in high-frequency scenarios. At the same time, it eliminates the need for repeated simulation verification using multiple physics fields, further reducing computational costs and adapting to the design requirements of increasing the target frequency band of filters under the background of scarce spectrum resources. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in this invention or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1A schematic diagram of the parameterized inverse design method for RF filters based on cyclic networks provided in an embodiment of the present invention; Figure 2 The IPD filter structure diagram is provided as a verification example in the embodiments of the present invention; Figure 3 A diagram of a cyclic network architecture provided in an embodiment of the present invention; Figure 4 A flowchart of recurrent network training provided in an embodiment of the present invention; Figure 5 The design result diagram of the reverse model provided in the embodiment of the present invention; Figure 6 The following are different results of the reverse model under the same input index provided in the embodiments of the present invention; Figure 7 A schematic diagram of the parameterized inverse design device architecture for a RF filter based on a cyclic network provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0037] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this invention should have the ordinary meaning understood by those skilled in the art to which this invention pertains. The terms "comprising" or "including," or similar words used in the embodiments of this invention, mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects.

[0038] With the rapid development of satellite communication and millimeter-wave technology, radio frequency (RF) filters are playing an increasingly important role in the industry. As the electromagnetic environment becomes increasingly complex, the design requirements for RF filters are becoming more stringent. They not only need to meet design specifications such as amplitude-frequency characteristics, but also precisely control phase characteristics such as group delay to avoid problems such as modulation signal distortion, EVM degradation, and ACLR reduction. Similarly, as a core component of wireless communication systems, the performance of RF filters directly affects signal integrity and system reliability, playing a crucial role in the entire system.

[0039] However, with the increasing number of applications for RF filters today, significant limitations exist in their design: The first category is traditional design methods, which heavily rely on engineers' accumulated experience, resulting in lengthy design processes and high iteration costs. One type of design approach requires repeatedly adjusting the parameters of each filter component using simulation software or forward solvers, conducting multiple simulation verifications, and relying on human experience to determine the direction of parameter optimization, ultimately selecting component parameters that meet the criteria. The other type introduces optimization algorithms such as the Cuckoo Algorithm and Genetic Algorithms to automate parameter iteration by designing fitness functions, but the construction of the fitness function still relies on empirical formulas, and the iteration process requires repeated calls to simulation software or forward solvers, leading to high computational costs and making it difficult to shorten the design cycle.

[0040] The second category is the forward fitting design method based on deep learning. This type of method attempts to replace traditional simulation software or forward solvers with the nonlinear fitting capabilities of deep learning algorithms to accelerate the parametric simulation process. Specifically, it maps component parameters to design specifications by training a model, and then iteratively searches for component parameters that meet the target specifications using optimization algorithms. While this approach reduces the time cost of a single simulation to some extent, it still requires multiple iterations to approximate the optimal solution and cannot directly output parameter results that meet the design specifications, thus limiting the improvement in design efficiency.

[0041] The third category is deep learning-based inverse design methods, which aim to directly construct the mapping relationship between design metrics and component parameters. Theoretically, this can achieve a rapid response to design requirements. However, due to the inherent pathological nature and multiple solutions of the inverse problem, there are obvious bottlenecks in practical applications: On the one hand, in order to avoid the model failing to converge due to "the same input corresponding to multiple outputs", massive amounts of training data need to be collected and manually cleaned. This not only incurs extremely high time and manpower costs, but also results in a large amount of data waste (only a small amount of the uncleaned data can be used for training); on the other hand, the solution space is often artificially reduced during the manual cleaning process, resulting in poor generalization ability of the trained model, making it difficult to adapt to the design requirements of different frequency bands and different types of RF filters.

[0042] Meanwhile, with increasingly scarce spectrum resources, the target frequency bands for RF filters are constantly rising, and the component size relative to wavelength is increasing, further increasing the computational complexity of simulation software and forward solvers, leading to slower simulation speeds. Furthermore, complex application scenarios require consideration of multiphysics effects, and the introduction of multiphysics simulation further prolongs design verification time. These issues collectively result in existing design methods being unable to meet the demands of rapid industry development in terms of cycle time and cost.

[0043] In view of this, the present invention proposes a parametric inverse design method and apparatus for RF filters based on cyclic networks. While ensuring design accuracy, it can improve design freedom, simultaneously provide multiple design results, reduce the RF filter design cycle, and decrease the reliance on human experience in the design process. The following are the specific details of the embodiments of the present invention.

[0044] See Figure 1 This invention provides a parameterized inverse design method for radio frequency filters based on cyclic networks, comprising the following steps: Step 1: Construct a recurrent network architecture consisting of a forward network and a reverse network. The forward network is used to realize data mapping and dataset expansion, and the reverse network is used to realize the mapping from design specifications to RF filter component parameters. The training process of the reverse network depends on the prediction results output by the forward network for optimization. Step 2: Use optimization algorithms combined with forward solvers or simulation software to filter effective data and collect an initial training dataset including RF filter component parameters and corresponding design specifications. Step 3: Train the forward network based on the initial training dataset, so that the forward network has the ability to map the parameters of the radio frequency filter components to the corresponding design specifications; Step 4: Use the trained forward network to generate data on the parameters of newly added RF filter components and their corresponding design specifications to expand the initial training dataset; Step 5: Train the inverse network based on the expanded initial training dataset. During training, input the component parameters output by the inverse network into the trained forward network to obtain the predicted design metrics. Combine the actual design metrics to construct a loss function. Through iterative optimization, enable the inverse network to map design metrics to component parameters. Step 6: Input the target design index into the trained reverse network, and the reverse network outputs the RF filter element parameters that meet the target design index to complete the reverse design.

[0045] In this embodiment, in step 1, both the forward network and the reverse network adopt the Vision Permutator network architecture. The Vision Permutator network architecture includes an input block processing module, a rearrangement module, and a result output module. The input block processing module divides the network input into sub-blocks according to preset rules and performs zero-padding on the sub-blocks of different types of RF filter element parameters to unify the dimensions. The rearrangement module includes multiple rearrangement sub-blocks, each of which includes a Permute-MLP layer for processing spatial information, a Channel-MLP layer for mixing channel information, and two LayerNorm layers for normalization. The result output module maps the features extracted by the network to the final output result through a fully connected layer and a global average pooling layer.

[0046] The parametric design in this embodiment refers to extracting several component parameters that significantly impact performance as independent variables, and adjusting these important component parameters to ensure the designed device meets performance specifications. Specifically, the recurrent network architecture in step 1 consists of two parts: an inverse network and a forward network. The forward network has two main functions: one is to quickly generate and expand the dataset, and the other is to act as an auxiliary training module for the forward solver. The inverse network utilizes the strong nonlinear fitting capability of deep learning to map design specifications to component parameters. In each training iteration, it calls the trained forward algorithm to quickly predict the S21 of the inverse network. It is called a recurrent network because during the training process of the inverse network, the output of the inverse network is input into the forward network to obtain the predicted design specifications, and the predicted design specifications, together with the actual design specifications, optimize the model.

[0047] This embodiment focuses on the S21 parameter as the design metric, but it can also be used to achieve other design metrics. The loss function of the inverse network consists of the errors of the predicted S21 and the actual S21, as well as the errors of the component parameters and the actual component parameters. As training progresses, the weight of the S21 error will be higher than that of the component error. Furthermore, to alleviate the ambiguity of the inverse design problem, a random number encoding is concatenated after the input of the inverse network, i.e., the design metric S21, so that it can output multiple parameters for the same metric while quickly outputting the component parameters.

[0048] In this embodiment, in step 2, the optimization algorithm is the Cuckoo Optimization Algorithm. When filtering the initial training data using the optimization algorithm, the design specifications of the radio frequency filter are used as constraints to remove invalid component parameter data that do not meet the specifications.

[0049] Specifically, due to the time-consuming nature of collecting data using random parameters and the presence of a significant amount of useless data, the use of the Cuckoo Optimization Algorithm combined with a solver or simulation software can significantly reduce the time spent manually removing useless data. After data collection, the forward network is trained. Both the forward and inverse networks are constructed using the Vision Permutator network architecture. Then, using a forward model to better illustrate the importance of the invention, a fourth-order IPD filter is used as an example, such as... Figure 2As shown, the device consists of four inductors and five capacitors, capable of filtering any frequency band within the 1-6GHz range. The entire device is symmetrical around capacitor C3. To better demonstrate the model's performance, the layout parameters of the capacitors and inductors, rather than their inductance values, are used as design specifications. Each inductor has four adjustable parameters: length, width, line width, and spacing. Each capacitor has two adjustable parameters: the length and width of the metal sheet. All inductors are spiral inductors, and all capacitors are MIM capacitors. Therefore, there are a total of 14 adjustable parameters. The final model takes the design specifications (in this example, the filter's S21 parameters) as input and outputs the 14 capacitor and inductor parameters. By sampling the S21 parameters at 0.01GHz intervals, the target frequency band of 1-6GHz is divided into 501 frequency points. For better subsequent calculations, the data of the last frequency point is discarded. Therefore, the input of the inverse model is the S21 parameters of 500 frequency points, and the output is the 14 capacitor and inductor parameters. Similarly, the input to the forward model is 14 capacitor and inductor parameters, and the output is the S21 parameter.

[0050] In this embodiment, the core idea of ​​the Vision Permutator network architecture lies in its handling of spatial information in the input feature representation. Unlike traditional methods that flatten the input feature representation before encoding spatial information, the Vision Permutator network architecture encodes the feature representation using linear projection along both the height and width dimensions. This allows the model to capture long-range dependencies along one spatial direction while preserving precise location information along another. This design enables the Vision Permutator network architecture to achieve performance comparable to CNNs and ViTs while maintaining a low parameter count, without relying on traditional convolutional or attention mechanisms. This architecture provides an alternative approach in computer vision, especially in scenarios requiring efficient, low-parameter models.

[0051] like Figure 3As shown, the Vision Permutator network architecture can be divided into three parts: an input block processing module, a rearrangement module, and a result output module. The input block processing module splits the output into uniform sub-segments. For the forward network, the input consists of the parameters of 14 capacitors and inductors. This is divided into five sub-blocks based on the components. Since inductors have four parameters and capacitors have two, zeros are added to the parameters of each capacitor before further division. Therefore, for the forward network, the input dimension of the Vision Permutator network architecture is (N, C, H), where N is the number of samples, C is the number of channels, and H is the parameter of each sample. Thus, in the forward module, C=1 and H=14. There are three capacitors; after further division, two zeros are added to the end of each capacitor. The output dimension of the first sub-module of the forward network is then C=5 and N=4. Similarly, the inverse network has C=1 and N=500.

[0052] The key point to note is that in the reverse network, since the input is the S21 parameter and the inverse problem has multiple solutions, a string of random numbers needs to be concatenated with the S21 parameter to distinguish between different outputs from the same input. This is feasible because all component parameters conforming to the S21 parameter distribution are viable. The length of the random numbers is the length of one sub-module, which is then input to the input block processing module for segmentation. In the verification example of this invention, the segmentation length is 10 (which can also be replaced with other dimensions divisible by the original sequence length), resulting in an output dimension of C=101 and N=5. The segmented data is then input to a parallel fully connected network, and the output of the parallel fully connected network serves as the input to the rearrangement module.

[0053] In this embodiment, the rearrangement module is constructed from a series of rearrangement sub-blocks. Each rearrangement sub-block consists of a Channel-MLP (fully connected channel), a (PerMute-MLP) rearrangement MLP, and two Layernorms (layer normalizations). The rearrangement module is the core module of the Vision Permutator network architecture, and will be described in detail below.

[0054] First, we introduce the PerMute-MLP layer, whose main function is to process spatial information. Permute-MLP introduces a three-dimensional information encoding mechanism of height, width, and channel. To calibrate the importance of different branches, a weighted fusion method is proposed. The input information is fed into each of the three branches separately. In the channel direction, a simple fully connected layer is used to linearly project the input, while in the width and height directions, the features are reordered and integrated, thus performing independent encoding in each direction. Specifically, let the output of the input block processing module be:

[0055] In the formula, For the sample size, For the number of channels, The parameter dimensions for each sample; First, the input Perform channel grouping, let the number of groups be . The grouped data is In the formula, , The height dimension after grouping; Next, the grouped data By performing dimensional permutation, we obtain And then By reshaping the dimensions, we can obtain ; Then, the height direction is encoded through a fully connected layer to obtain the height direction output:

[0056] In the formula, Here is the weight matrix of the fully connected layer in the height direction. For the height-direction fully connected layer bias term; Similarly, the width direction output can be obtained. Output in the direction of the channel ,in Weight matrix of the fully connected layer in the width direction With bias term , Weight matrix of fully connected layer in the corresponding channel direction With bias term ; The output of the Permute-MLP layer is Then it is processed through a fully connected layer to obtain In the formula, This is the final weight matrix of the fully connected layer. This is the bias term for the final fully connected layer.

[0057] In the rearrangement module, the Channel-MLP layer includes two fully connected layers and a GELU activation function, which aims to mix channel information by fully connecting the channels. Let the input of the Channel-MLP layer be First, perform a dimensionality transformation on the input, placing the channel dimension as the last dimension, to obtain... get Next, parallel fully connected layer processing is performed, where the channel information of each W-dimensional element is mixed to obtain... In the formula, This is the weight matrix for the first fully connected layer. This is the bias term for the first fully connected layer; Next, the GELU activation function is applied to... Perform a nonlinear transformation to obtain =GELU ( ); Finally, it is processed through a second fully connected layer to obtain... In the formula, This is the weight matrix for the second fully connected layer. This is the bias term for the second fully connected layer, and the second activation function. =GELU ( Next, the matrix dimensions are restored to their original form to obtain the output of the Channel-MLP layer. .

[0058] By encoding from three different dimensions, the Vision Permutator network architecture can better represent the high-dimensional features of the input. The Channel-MLP layer, which mixes channel information, consists of two fully connected layers and the GELU activation function, aiming to achieve the interaction of channel information through full connectivity. The final module is the output module, constructed using a fully connected layer and global average pooling, mapping the feature map to the network model's output.

[0059] like Figure 4 As shown, the training process of the model first involves training a forward model using the dataset. After the forward network model is trained, new data is quickly generated to expand the dataset. Next, an inverse network model is trained. The input to the inverse network model is the S21 parameters plus a string of random numbers, and the output is the capacitance and inductance parameters required for the fourth-order IPD filter. The output of the inverse network model is then fed into the trained forward network model to obtain the predicted S21 parameters, which, along with the design metrics, are input into the loss function to iteratively optimize the inverse model. The trained inverse network model can then be used to quickly design the required IPD filter. The forward network trained on 20,000 sets of data achieves an error of 1.07% on 2,000 test sets. Combining these 20,000 sets of data with the 40,000 sets of data generated by the forward network to train the inverse model, the accuracy on 10,000 test samples is 1.26%.

[0060] See Figure 5 The paper presents six sets of S21 parameters and the accurate values ​​obtained by substituting the capacitor and inductor components output by the network into the forward solver. It can be seen that the design results provided by the model closely match the target. To further verify the superiority of this invention, Figure 6The different results given by the network under the same design index input are presented, which shows that the model can provide multiple different component parameters that meet the design index, greatly improving the degree of design freedom.

[0061] Furthermore, it should be emphasized that the method proposed in this invention can be used in the parametric modeling of various RF filters, not just the IPD filter used as an example in this invention. This invention can significantly reduce the design cycle of RF filters.

[0062] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and these multiple devices will interact with each other to complete the described RF filter parameterization inverse design method based on a cyclic network.

[0063] It should be noted that the above description describes some embodiments of the present invention. In some cases, the described actions or steps can be performed in a different order than that shown in the above embodiments and the desired result can still be achieved. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0064] See Figure 7 Based on the same inventive concept, corresponding to any of the above embodiments, this invention also provides a RF filter parameterization inverse design apparatus based on cyclic networks. The RF filter parameterization inverse design method based on cyclic networks, employing the above embodiments or any possible implementation thereof, includes: The recurrent network construction unit 100 is used to construct a recurrent network architecture consisting of a forward network and a reverse network. The forward network realizes data mapping and dataset expansion, and the reverse network realizes the mapping from design indicators to RF filter component parameters. The training process of the reverse network depends on the prediction results output by the forward network for optimization. The data collection unit 200 is used to use optimization algorithms combined with forward solvers or simulation software to filter effective data and collect an initial training dataset including RF filter component parameters and corresponding design specifications. The forward network training unit 300 is used to train the forward network based on the initial training dataset, so that the forward network has the ability to map the parameters of the 300 RF filter components to the corresponding design specifications. The dataset expansion unit 400 is used to expand the initial training dataset by generating data on the parameters of newly added radio frequency filter components and their corresponding design specifications using the trained forward network. The inverse network training unit 500 is used to train the inverse network based on the expanded initial training dataset. During training, the component parameters output by the inverse network are input into the trained forward network to obtain the predicted design index. The loss function is constructed by combining the actual design index and iteratively optimizing the inverse network to enable it to map the design index to the component parameters. The design output unit 600 is used to input the target design index into the trained inverse network, and the inverse network outputs the RF filter element parameters that meet the target design index, thereby completing the inverse design.

[0065] In this embodiment, in the recurrent network construction unit 100, both the forward network and the reverse network adopt the Vision Permutator network architecture; the Vision Permutator network architecture includes an input block processing module, a rearrangement module, and a result output module; The input block processing module divides the network input into sub-blocks according to preset rules, and performs zero-padding on the sub-blocks of different types of RF filter element parameters to unify the dimensions; The rearrangement module includes multiple rearrangement sub-blocks, each of which includes a Permute-MLP layer for processing spatial information, a Channel-MLP layer for mixing channel information, and two LayerNorm layers for normalization. The output module maps the features extracted by the network to the final output result through a fully connected layer and a global average pooling layer.

[0066] In this embodiment, the optimization algorithm is the Cuckoo Optimization Algorithm. When filtering the initial training data using the optimization algorithm, the design specifications of the radio frequency filter are used as constraints to eliminate invalid component parameter data that do not meet the specifications.

[0067] In this embodiment, the design specification is the S21 parameter of the radio frequency filter; The input of the forward network is the component parameters of the radio frequency filter, and the output is the S21 parameter under the corresponding component parameters; The input of the reverse network includes the target S21 parameter of the RF filter and a string of random numbers. The random numbers are used to distinguish different component parameter solutions corresponding to the same S21 parameter. The output of the reverse network is the component parameters of the RF filter.

[0068] In this embodiment, in the inverse network training unit 500, the loss function includes a first error term and a second error term. The first error term is the error between the predicted S21 parameters output by the forward network and the actual S21 parameters, and the second error term is the error between the component parameters output by the inverse network and the actual component parameters. During training, the weights of the first error term and the second error term are dynamically adjusted so that the weight of the first error term gradually becomes higher than the weight of the second error term.

[0069] In this embodiment, the processing of the Permute-MLP layer in the rearrangement module satisfies the following mathematical model: Assume the output of the input block processing module is:

[0070] In the formula, For the sample size, For the number of channels, The parameter dimensions for each sample; For input Perform channel grouping, let the number of groups be . The grouped data is In the formula, , The height dimension after grouping; Grouped data By performing dimensional permutation, we obtain And then By reshaping the dimensions, we can obtain ; The height direction is encoded using a fully connected layer to obtain the height direction output:

[0071] In the formula, Here is the weight matrix of the fully connected layer in the height direction. For the height-direction fully connected layer bias term; The width direction output is obtained sequentially. Output in the direction of the channel ,in Weight matrix of the fully connected layer in the width direction With bias term , Weight matrix of fully connected layer in the corresponding channel direction With bias term ; The output of the Permute-MLP layer is Then it is processed through a fully connected layer to obtain In the formula, This is the final weight matrix of the fully connected layer. This is the bias term for the final fully connected layer.

[0072] In this embodiment, the rearrangement module includes two fully connected layers and a GELU activation function. The Channel-MLP layer is used to mix channel information by performing a fully connected operation on the channels. Let the input of the Channel-MLP layer be First, perform a dimensionality transformation on the input, placing the channel dimension as the last dimension, to obtain... get Next, parallel fully connected layer processing is performed, mixing the channel information for each W-dimensional element to obtain... In the formula, This is the weight matrix of the first fully connected layer. This is the bias term for the first fully connected layer; Next, the GELU activation function is applied to... Perform a nonlinear transformation to obtain =GELU ( ); Finally, it is processed through a second fully connected layer to obtain... In the formula, This is the weight matrix for the second fully connected layer. This is the bias term for the second fully connected layer, and the second activation function. =GELU ( Next, the matrix dimensions are restored to their original form to obtain the output of the Channel-MLP layer. .

[0073] In this embodiment, the parameters of the radio frequency filter element include inductance parameters and capacitance parameters; the inductor is a spiral inductor, including the length, width, line width and spacing of the inductor; The capacitor is a MIM capacitor, including the length and width of the capacitor metal sheet.

[0074] In this embodiment, the length of the random number input to the reverse network is consistent with the length of the sub-block of the input block processing module; During the input processing of the reverse network, the target S21 parameter is first processed into a fixed-dimensional vector according to a preset sampling rule, and then concatenated with a random number vector before being input to the input block processing module. Through the randomness of the random number, the reverse network outputs multiple different combinations of component parameters for the same S21 parameter input.

[0075] The apparatus of the above embodiments is used to implement a corresponding RF filter parameterized inverse design method based on a cyclic network in any of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0076] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the parameterized inverse design method for radio frequency filters based on cyclic networks as described in any of the above embodiments.

[0077] Figure 8 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 710, a memory 720, an input / output interface 730, a communication interface 740, and a bus 750. The processor 710, memory 720, input / output interface 730, and communication interface 740 are interconnected internally via the bus 750.

[0078] The processor 710 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0079] The memory 720 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 720 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 720 and is called and executed by the processor 710.

[0080] The input / output interface 730 is used to connect input / output modules to enable information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0081] The communication interface 740 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0082] Bus 750 includes a pathway for transmitting information between various components of the device, such as processor 710, memory 720, input / output interface 730, and communication interface 740.

[0083] It should be noted that although the above-described device only shows the processor 710, memory 720, input / output interface 730, communication interface 740, and bus 750, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0084] The electronic devices described above are used to implement a corresponding RF filter parameterized inverse design method based on a cyclic network in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0085] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute a parameterized inverse design method for a radio frequency filter based on a cyclic network as described in any of the above embodiments.

[0086] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0087] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute a parameterized inverse design method for a radio frequency filter based on a cyclic network as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0088] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the embodiments of the invention as described above, which are not provided in detail for the sake of brevity.

[0089] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of the invention, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of the invention, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of the invention will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of the invention, it will be apparent to those skilled in the art that the embodiments of the invention may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0090] Although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAMDRAM) may be used with the embodiments discussed.

[0091] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of this invention should be included within the scope of protection of this invention.

Claims

1. A parameterized inverse design method for RF filters based on cyclic networks, characterized in that, include: Step 1: Construct a recurrent network architecture consisting of a forward network and a reverse network. The forward network is used to realize data mapping and dataset expansion, and the reverse network is used to realize the mapping from design specifications to RF filter component parameters. The training process of the reverse network depends on the prediction results output by the forward network for optimization. Step 2: Use optimization algorithms combined with forward solvers or simulation software to filter effective data and collect an initial training dataset including RF filter component parameters and corresponding design specifications. Step 3: Train the forward network based on the initial training dataset, so that the forward network has the ability to map the parameters of the radio frequency filter components to the corresponding design specifications; Step 4: Use the trained forward network to generate data on the parameters of newly added RF filter components and their corresponding design specifications to expand the initial training dataset; Step 5: Train the inverse network based on the expanded initial training dataset. During training, input the component parameters output by the inverse network into the trained forward network to obtain the predicted design metrics. Combine the actual design metrics to construct a loss function. Through iterative optimization, enable the inverse network to map design metrics to component parameters. Step 6: Input the target design index into the trained reverse network, and the reverse network outputs the RF filter element parameters that meet the target design index to complete the reverse design.

2. The parameterized inverse design method for RF filters based on cyclic networks according to claim 1, characterized in that, In step 1, both the forward network and the reverse network adopt the Vision Permutator network architecture; the Vision Permutator network architecture includes an input block processing module, a rearrangement module, and a result output module; The input block processing module divides the network input into sub-blocks according to preset rules, and performs zero-padding on the sub-blocks of different types of RF filter element parameters to unify the dimensions; The rearrangement module includes multiple rearrangement sub-blocks, each of which includes a Permute-MLP layer for processing spatial information, a Channel-MLP layer for mixing channel information, and two LayerNorm layers for normalization. The output module maps the features extracted by the network to the final output result through a fully connected layer and a global average pooling layer.

3. The parameterized inverse design method for RF filters based on cyclic networks according to claim 1, characterized in that, In step 2, the optimization algorithm is the Cuckoo Optimization Algorithm. When filtering the initial training data using the optimization algorithm, invalid component parameter data that does not meet the design specifications of the radio frequency filter are eliminated as constraints.

4. The parameterized inverse design method for RF filters based on cyclic networks according to claim 1, characterized in that, In steps 1 to 6, the design specifications are the S21 parameters of the radio frequency filter; In step 3, the input of the forward network is the component parameters of the RF filter, and the output is the S21 parameter under the corresponding component parameters; In step 5, the input of the reverse network includes the target S21 parameter of the RF filter and a string of random numbers. The random numbers are used to distinguish different component parameter solutions corresponding to the same S21 parameter. The output of the reverse network is the component parameters of the RF filter.

5. The parameterized inverse design method for RF filters based on cyclic networks according to claim 1, characterized in that, In step 5, the loss function includes a first error term and a second error term. The first error term is the error between the predicted S21 parameters output by the forward network and the actual S21 parameters, and the second error term is the error between the component parameters output by the reverse network and the actual component parameters. During training, the weights of the first error term and the second error term are dynamically adjusted so that the weight of the first error term gradually becomes higher than the weight of the second error term.

6. The parameterized inverse design method for RF filters based on cyclic networks according to claim 2, characterized in that, In the rearrangement module, the processing of the Permute-MLP layer satisfies the following mathematical model: Assume the output of the input block processing module is: ； In the formula, For the sample size, For the number of channels, The parameter dimensions for each sample; For input Perform channel grouping, let the number of groups be . The grouped data is In the formula, , The height dimension after grouping; Grouped data By performing dimensional permutation, we obtain And then By reshaping the dimensions, we can obtain ; The height direction is encoded using a fully connected layer to obtain the height direction output: ； In the formula, Here is the weight matrix of the fully connected layer in the height direction. For the height-direction fully connected layer bias term; The width direction output is obtained sequentially. Output in the direction of the channel ,in Weight matrix of the fully connected layer in the width direction With bias term , Weight matrix of fully connected layer in the corresponding channel direction With bias term ; The output of the Permute-MLP layer is Then it is processed through a fully connected layer to obtain In the formula, This is the final weight matrix of the fully connected layer. This is the bias term for the final fully connected layer.

7. The parameterized inverse design method for RF filters based on cyclic networks according to claim 2, characterized in that, In the rearrangement module, the Channel-MLP layer includes two fully connected layers and a GELU activation function. The Channel-MLP layer is used to mix channel information by performing a fully connected operation on the channels. Let the input of the Channel-MLP layer be First, perform a dimensionality transformation on the input, placing the channel dimension as the last dimension, to obtain... get Next, parallel fully connected layer processing is performed, mixing the channel information for each W-dimensional element to obtain... In the formula, This is the weight matrix of the first fully connected layer. This is the bias term for the first fully connected layer; Next, the GELU activation function is applied to... Perform a nonlinear transformation to obtain =GELU ( ); Finally, it is processed through a second fully connected layer to obtain... In the formula, This is the weight matrix for the second fully connected layer. For the bias term of the second fully connected layer, and the second activation function =GELU ( Then, the matrix dimensions are restored to their original form to obtain the output of the Channel-MLP layer. .

8. The parameterized inverse design method for RF filters based on cyclic networks according to claim 4, characterized in that, The parameters of the radio frequency filter components include inductance parameters and capacitance parameters; the inductor is a spiral inductor, including the length, width, line width and spacing of the inductor; The capacitor is a MIM capacitor, including the length and width of the capacitor metal sheet.

9. The parameterized inverse design method for RF filters based on cyclic networks according to claim 4, characterized in that, The length of the random number input to the reverse network is the same as the length of the sub-block in the input block processing module; During the input processing of the reverse network, the target S21 parameter is first processed into a fixed-dimensional vector according to a preset sampling rule, and then concatenated with a random number vector before being input to the input block processing module. Through the randomness of the random number, the reverse network outputs multiple different combinations of component parameters for the same S21 parameter input.

10. A parametric inverse design apparatus for RF filters based on cyclic networks, employing the parametric inverse design method for RF filters based on cyclic networks as described in any one of claims 1 to 9, characterized in that, include: A recurrent network construction unit is used to construct a recurrent network architecture consisting of a forward network and a reverse network. The forward network is used to realize data mapping and dataset expansion, and the reverse network is used to realize the mapping from design indicators to RF filter component parameters. The training process of the reverse network depends on the prediction results output by the forward network for optimization. The data collection unit is used to filter effective data using optimization algorithms combined with forward solvers or simulation software, and to collect an initial training dataset including RF filter component parameters and corresponding design specifications. A forward network training unit is used to train the forward network based on the initial training dataset, so that the forward network has the ability to map the parameters of the radio frequency filter components to the corresponding design specifications; The dataset expansion unit is used to expand the initial training dataset by generating data on the parameters of newly added radio frequency filter components and their corresponding design specifications using the trained forward network. The inverse network training unit is used to train the inverse network based on the expanded initial training dataset. During training, the component parameters output by the inverse network are input into the trained forward network to obtain the predicted design index. The loss function is constructed by combining the actual design index and iteratively optimizing the inverse network to enable it to map the design index to the component parameters. The design output unit is used to input the target design index into the trained inverse network, and the inverse network outputs the RF filter element parameters that meet the target design index, thus completing the inverse design.

Images