Resonant filter circuit design method, system, apparatus, and storage medium

Abstract

The application relates to the technical field of electronic circuits, and specifically discloses a resonant filter circuit design method, system and device and a storage medium, wherein a to-be-optimized circuit is modeled as a weighted graph based on a graph representation method weighted by electromagnetic coupling effects, modeling of a complex electromagnetic environment in the circuit is realized, that is, a node represents a resonant unit, an edge represents electromagnetic coupling effects between resonant units, and an edge weight represents differentiated electromagnetic coupling strength; the above graph structure data is input into a fusion graph representation model (EMGNN) of a graph isomorphism network cascade graph attention network to generate a transfer function, a cross-electromagnetic topology graph migration learning mechanism driven by a model and data is introduced to perform migration and electromagnetic parameter reverse optimization on the fusion graph representation model (EMGNN), and deep representation across electromagnetic topologies, electromagnetic parameter iterative optimization and lightweight design across topological structures are realized, so that the method can be applied to isomorphic component circuit design of various different electromagnetic coupling types and different device types.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electronic circuit, and particularly relates to a resonant filter circuit design method, system, device and storage medium. BACKGROUND

[0002] The essence of resonant filter circuit, such as resonant filter circuit forward design, is a performance index driven electromagnetic parameter and layout optimization problem, which is mostly an NP-complete electromagnetic inverse problem. The traditional method iteratively optimizes by a large number of simulations and parameter adjustments of the initial template and relies on manual work, which has the problems of long design cycle and high computational complexity.

[0003] Graph representation learning is an effective method to realize intelligent forward design of circuit. However, there is currently no effective graph modeling research on the complex electromagnetic environment of coupled circuits, resulting in low performance prediction accuracy of the circuit; at the same time, the existing research is mostly limited to internal deep representation and learning of the database composed of the same type of circuit topology structure, and relies on a large amount of circuit simulation data, which is difficult to realize cross-topology migration, fast automation and lightweight computation of intelligent design of resonant filter circuit. SUMMARY

[0004] The present application provides a resonant filter circuit design method, system, device and storage medium, which solves the technical problem of how to realize cross-topology migration, fast automation and lightweight computation of intelligent design of resonant filter circuit.

[0005] To solve the above technical problems, the present application provides a resonant filter circuit design method, which comprises the following steps:

[0006] S1, input the source data composed of a plurality of resonant filter circuits into a graph representation module weighted by electromagnetic coupling effect;

[0007] S2, the graph representation module represents the resonant unit as a node, represents the electromagnetic coupling effect between the resonant units as an edge, and represents the electromagnetic coupling strength as an edge weight, and models the source data as a weighted graph input into a fusion graph representation model introducing graph isomorphism network cascade graph attention network;

[0008] S3, the fusion graph representation model generates a corresponding transfer function prediction value based on the input weighted graph;

[0009] S4, update the parameters of the fusion graph representation model based on the transfer function prediction value and the actual value corresponding to the source data, to obtain a pre-trained fusion graph representation model;

[0010] S5, perform transfer learning on the pre-trained fusion graph representation model driven by small data of the target electromagnetic specification of the circuit to be optimized, to obtain a transferred fusion graph representation model;

[0011] S6. For the circuit to be optimized, M resonant filter circuits are randomly generated, and after graph modeling, they are input into the fused graph representation model after migration. The target transfer function is used as the approximation target. The electromagnetic parameters of each resonant filter circuit are automatically updated in the fused graph representation model after migration through the gradient optimization algorithm. Finally, the resonant filter circuit with the smallest loss value after the electromagnetic parameter update is selected as the best resonant filter circuit.

[0012] Furthermore, each of the resonant filter circuits consists of N microstrip open-loop square resonators, i.e., resonant units, with each resonant unit serving as a node;

[0013] The weighted graph is characterized by the electromagnetic topology of the resonant filter circuit, the node features of each node, the edge features between node features, and the weights corresponding to the edge features; the electromagnetic topology of the resonant filter circuit, the node features of each node, the edge features between node features, and the weights corresponding to the edge features are determined according to the physical parameters of the resonant filter circuit.

[0014] Furthermore, by obtaining the first part of the physical parameters of each node as point parameters, and then by mapping the point parameters to a high-dimensional space for encoding and learning, the node features of each node are obtained.

[0015] The second part of the physical parameters between nodes is obtained as edge parameters, and the edge features corresponding to the edge parameters are calculated.

[0016] The electromagnetic topology is represented by an N×N electromagnetic topology matrix, where the elements of the electromagnetic topology matrix in the i-th row and j-th column of the k-th layer of the multilayer perceptron are... This represents the electromagnetic coupling effect between node i and node j in the k-th layer. The values ​​i,j = 1, 2, 3…N are determined based on the second part of the physical parameters of each node, the node characteristics, and the edge characteristics.

[0017] The weights corresponding to the edge features are determined based on the node features and the edge features.

[0018] Furthermore, the first part of the physical parameters includes the input / output ports, side length, opening orientation, and opening angle of the resonant unit; the second part of the physical parameters includes the difference between the horizontal and vertical coordinates between the center positions of the resonant units, the gap between the resonant units, and the vertical displacement between the resonant units.

[0019] The electromagnetic topology matrix in the i-th row and j-th column of the k-th layer of the multilayer perceptron. Equal to the normalized adjacency matrix of the resonant filter circuit The element a in the i-th row and j-th column ij Multiply by the edge encoding between node i and node j in the k-th layer

[0020] a ij is defined as: when the distance Δi,j between the different nodes i and j is less than the farthest distance L between the nodes i and j that produces a strong enough electromagnetic coupling effect, a ij is equal to 1, otherwise, a ij is equal to relu(·) represents a relu activation function; and when i = j, a ij = 1.

[0021] is equal to respectively represent the node features of the nodes i and j at the k-th layer, represents the edge feature between the nodes i and j at the k-th layer, η k () represents the k-th layer of a multi-layer perception with a LeakyReLU activation function;

[0022] the edge feature corresponds to the weight is equal to represents the importance of the node j to the node i, and respectively represent the fusion features of the node features and the corresponding edge features of the nodes i and j, softmax j () represents a softmax function;

[0023] is equal to δ represents a learnable weight vector, W represents a learnable weight matrix, (·) T represents a matrix transposition operation, LeakyReLU() represents a LeakyReLU activation function, and || is a vector concatenation operation.

[0024] Further, the fusion graph representation model comprises a cascaded 1-layer graph isomorphic network, K-1-layer graph attention network, graph pooling network and fully connected network; the graph isomorphic network is based on the graph input of the input containing the initial node feature h (0) and the adjacency matrix A to perform neighborhood aggregation and update to obtain the first layer of updated node features h (1) input to the first layer of the graph attention network; the K-1-layer graph attention network is based on h (1) its own features and the weighted electromagnetic coupling effect of the neighborhood to perform message passing and node update, and the K-1-layer graph attention network outputs the updated node features h (K) ; the graph pooling network is used to pool the output node features of the K-1-layer graph attention network and input node features perform splicing to obtain a graph embedding; the fully connected network predicts an output corresponding transfer function according to the graph embedding;

[0025] In the pre-training of the fusion graph representation model, the iteratively trained are the learnable parameters W and δ.

[0026] Further, in the step S5, when designing across topologies, the optimal migration is to fine-tune all layers of the pre-trained fusion graph representation model; in the two different cell number tasks designed across cell numbers, the optimal settings are to fine-tune all layers of the pre-trained fusion graph representation model, freeze the first layer of the pre-trained fusion graph representation model and fine-tune the remaining layers, respectively.

[0027] Further, in the step S6, the target transfer function is defined as and a linear combination of λ2·|w * | 2 y i represents the transfer function of the i-th resonant filter circuit predicted by the migrated fusion graph representation model, w BP represents a set of passband frequencies, w i represents the passband frequency of the i-th resonant filter circuit, w * represents the transition bandwidth, w * calculated from the difference between the passband and stopband widths, λ1 and λ2 are learnable weight parameters.

[0028] The present application also provides a resonant filter circuit design system, which is characterized by comprising an input module, a graph representation module, a fusion graph representation module, a migration module and an optimization module, which are respectively used to perform steps S1, S2, S3-S4, S5 and S6 in the above resonant filter circuit design method.

[0029] The present application also provides a resonant filter circuit design device, which is characterized by being provided with an intelligent agent, which has an input unit, a graph representation unit, a fusion graph representation unit, a migration unit and an optimization unit, which are respectively used to perform steps S1, S2, S3-S4, S5 and S6 in the above resonant filter circuit design method.

[0030] The present application also provides a storage medium, which is characterized by storing a computer program, which is executed to implement steps S1 to S6 in the above resonant filter circuit design method.

[0031] The resonant filter circuit design method, system and device and storage medium provided by the present application are based on an electromagnetic coupling effect based on graph attention (EMCoE-GAT) to model the circuit to be optimized as a weighted graph, realize modeling of a complex electromagnetic environment inside the circuit, that is, a node represents a resonant unit, an edge represents an electromagnetic coupling effect between resonant units, and an edge weight represents differentiated electromagnetic coupling strength; input the above graph structure data to a fusion graph representation model (EMGNN) of a graph isomorphism network cascade graph attention network to generate a transfer function, and introduce a model and data double-driven cross electromagnetic topology graph migration learning mechanism to migrate and optimize the electromagnetic parameters of the fusion graph representation model (EMGNN), realize deep representation across electromagnetic topologies, iterative optimization of electromagnetic parameters, and lightweight design across topologies, and can be applied to isomorphic component circuit design of various different electromagnetic coupling types and different device types. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 is a principle architecture diagram of the resonant filter circuit design method, system, device and storage medium provided by the embodiment of the present application;

[0033] Figure 2 is an example graph of an electromagnetic coupling effect based on graph attention provided by the embodiment of the present application;

[0034] Figure 3 is a graph representation principle example graph of a resonant unit provided by the embodiment of the present application;

[0035] Figure 4 is a graph representation principle diagram example graph of an electromagnetic coupling effect provided by the embodiment of the present application;

[0036] Figure 5 is an architecture diagram of a fusion graph representation model (EMGNN) provided by the embodiment of the present application;

[0037] Figure 6 is a CircuitGTL cross-topology visualization prediction result example graph provided by the embodiment of the present application;

[0038] Figure 7 is a transfer function graph corresponding to a filter designed by the embodiment of the present application CircuitGTL, CircuitGNN and artificial design, wherein the green area represents a target passband, and the design task is a bandpass resonant filter with a passband of [220GHz, 250GHz] composed of four resonant units;

[0039] Figure 8is an example diagram of CircuitGTL cross-cell quantity visualization prediction result provided by the embodiment of the present application.

[0040] Figure 9 is a transfer function diagram corresponding to CircuitGTL, CircuitGNN and artificial design result provided by the embodiment of the present application, wherein the yellow area represents the target passband, and the design task is a bandpass resonant filter with a passband of [220GHz, 250GHz] composed of 5 resonant cells. DETAILED DESCRIPTION

[0041] The embodiments of the present application will be described in detail below with reference to the accompanying drawings. The embodiments are given only for illustrative purposes and cannot be understood as limiting the present application. The accompanying drawings are only for reference and illustration and do not constitute a limitation on the scope of patent protection of the present application, because many changes can be made to the present application without departing from the spirit and scope thereof.

[0042] The resonant filter circuit design method provided by the embodiment of the present application is referred to as CircuitGTL method / algorithm, and the working principle framework is as shown in Figure 1 , mainly including the following steps:

[0043] S1, input the source data (training data) into the graph representation module (EMCoE-GAT) using electromagnetic coupling effect weighting, that is,① in Figure 1

[0044] S2, the graph representation module (EMCoE-GAT) models the source data as a weighted graph, in which the nodes represent resonant cells, the edges represent the electromagnetic coupling effect between the resonant cells, and the edge weights represent the electromagnetic coupling strength, and inputs the weighted graph into the fusion graph representation model (EMGNN) of the Graph Isomorphic Network, GIN, cascaded Graph Attention Network, GAT), that is,② in Figure 1

[0045] S3, the fusion graph representation model (EMGNN) generates a corresponding transfer function prediction value based on the input weighted graph, that is,③ in Figure 1

[0046] S4, fit the prediction value with the actual value, if the fitting is successful, the training is ended, and the pre-trained fusion graph representation model (EMGNN) is obtained, if the fitting is not successful, update the parameters of the fusion graph representation model (EMGNN) based on the prediction value and the actual value (i.e. back propagation), and obtain the pre-trained fusion graph representation model (EMGNN), that is,④ in Figure 1

[0047] S5, performing transfer learning on the pre-trained fusion graph representation model (EMGNN) driven by small data of target electromagnetic specifications of the circuit to be optimized, to obtain a migrated fusion graph representation model (EMGNN), i.e. Figure 1 ⑤ in the method;

[0048] S6, randomly generating M initial resonant filter circuits for the circuit to be optimized, performing graph modeling and inputting the migrated fusion graph representation model (EMGNN), taking the target transfer function as the approximation target, and automatically updating the electromagnetic parameters of each resonant filter circuit in the EMGNN through a gradient optimization algorithm, and finally taking the resonant filter circuit with the smallest loss value after updating the electromagnetic parameters as the best resonant filter circuit, i.e. Figure 1 ⑥ in the method.

[0049] Firstly, the example proposes an electromagnetic coupling effect weighted graph representation method (EMCoE-GAT), i.e. the nodes represent resonant units, the edges represent the electromagnetic coupling effect between the resonant units, and the edge weights represent the electromagnetic coupling strength; secondly, a fusion graph representation model (EMGNN) of a graph isomorphic network (GIN) cascaded graph attention network is introduced to realize cross-domain deep representation of source domain and target domain electromagnetic graph data; finally, an optimization objective of joint regularization of passband and transition band transfer function is proposed, and iterative optimization of electromagnetic parameters of the target domain resonant filter is realized based on the shared parameters of the EMGNN model. The CircuitGTL deeply integrates the graph representation capability of the graph neural network and the lightweight capability of the graph transfer learning, and realizes the intelligent design of the cross electromagnetic topology transfer learning circuit.

[0050] An example of the electromagnetic coupling effect weighted graph representation (EMCoE-GAT) is shown in Figure 2 The weighted graph is represented by the electromagnetic topology of the resonant filter circuit, the node features of each node, the edge features between the node features, and the weights corresponding to the edge features. The electromagnetic topology of the resonant filter circuit, the node features of each node, the edge features between the node features, and the weights corresponding to the edge features are determined according to the physical parameters of the resonant filter circuit.

[0051] The example takes the Massachusetts Institute of Technology (MIT) open-source resonant filter circuit data MIT-CircuitGNN as the test object, and each resonant filter circuit is composed of several microstrip open-loop square resonators (resonant units), and the physical parameters are as shown in Figure 2and Table 1, the resonant filter includes 3 resonant units from left to right, the center coordinates of which are (x1, y1), (x2, y2), (x3, y3) respectively, wherein the first resonant unit is provided with an input port (Input), the third resonant unit is provided with an output port (Output), the side length of the three resonant units is l, the gap between the resonant units is S, the vertical displacement between the resonant units is D, the opening angle of the resonant unit is θ, and the opening direction of the resonant unit (left or down or right or up). Figure 2 The corresponding resonant filter is shown in the graph of Figure 2 (b).

[0052] Table 1 Point\Edge Features of Resonant Filter Graph Modeling

[0053]

[0054] The EMCoE-GAT includes a node-level modeling network and an edge-level modeling network, and models the resonant filter as graph data, that is, the nodes represent the resonant units, the edges represent the electromagnetic coupling effect between the resonant units, and the edge weight represents the electromagnetic coupling strength; the point features (node features) and the edge features fully model the size information and the structural characteristics of the resonant filter.

[0055] By obtaining the first part of the physical parameters of each node as the point parameters, and then mapping the point parameters to a high-dimensional space for encoding and learning, the node features of each node are obtained;

[0056] By obtaining the second part of the physical parameters between the nodes as the edge parameters, and calculating the edge features corresponding to the edge parameters;

[0057] The electromagnetic topology is represented by an N×N electromagnetic topology matrix, and the element of the electromagnetic topology matrix in the i-th row and the j-th column of the k-th layer of the multi-layer perceptron represents the electromagnetic coupling effect between the k-th layer node i and the node j, determined according to the second part of the physical parameters of each node, the node features and the edge features, i, j = 1, 2, 3…N;

[0058] The weight corresponding to the edge features is determined according to the node features and the edge features.

[0059] As shown in Table 1, the point parameters of the example for the resonant filter include input / output port I / O = [1, 0] or [0, 1] or [0, 0] ([1, 0] represents an input port, [0, 1] represents an output port, and [0, 0] represents neither an input port nor an output port), the side length l of the resonant unit, the opening direction dir = Binary[a, b, c, d] (where [a, b, c, d] are binary elements representing the left, bottom, right, and top four directions, respectively, such as the opening direction left, a is 1, and b, c, and d are all 0), and the resonant unit opening angle The physical information of a single resonant unit is modeled, and then the node features are mapped to a high-dimensional space by a multilayer perceptron (MLP) with LeakyReLU activation to realize encoding and learning, as shown in the node feature generation process of the resonant unit Figure 3 , where the node features after encoding and learning are represented by h i . So far, the node features have been mapped from the initial dimension dim = 8 ([I / O, l, θ, dir]) to the dimension dim = 800.

[0060] As shown in Table 1, the edge parameters of the example for the resonant filter include the gap S between resonant units and the vertical displacement D between resonant units.

[0061] The electromagnetic coupling effect comprehensively reflects the influencing factors of the electromagnetic coupling strength, such as distance, coupling mode, orientation, and angle of the unit opening. EMCoE-GAT models the influence degree of each factor to represent differentiated electromagnetic coupling strength and takes it as the edge weight value of electromagnetic signal aggregation, which realizes the graph representation of the complex electromagnetic environment inside the resonant filter circuit and improves the graph representation ability of the traditional sum neighborhood aggregation. As shown in the graph representation principle diagram of the electromagnetic coupling effect Figure 4 .

[0062] The influence of the distance of the electromagnetic coupling effect adopts the modeling method of the normalized adjacency matrix. Assuming that there are N nodes (N resonant units), let a ij be the normalized adjacency matrix , the element corresponding to node i and node j in the normalized adjacency matrix a ij , can be expressed as:

[0063]

[0064] wherein, represents the distance between node i and node j (Δx and Δy represent the difference between the horizontal and vertical coordinates of node i and node j, respectively), L represents the farthest distance between node i and node j that produces a strong enough electromagnetic coupling effect, and relu(·) represents the relu activation function. a ijThe electromagnetic coupling strength can be fully characterized with respect to distance.

[0065] The influence of the electromagnetic coupling effect coupling mode, opening orientation and angle is modeled by an attention mechanism-based modeling method that sets represents the electromagnetic topology matrix of the kth layer, where represents the electromagnetic coupling effect between nodes i and j, and C can be represented as:

[0066]

[0067] where η k represents the kth layer of the MLP with a LeakyReLU activation function (LeakyReLU()), and equation (4) concatenates the edge features with the node features of nodes i and j at the kth layer , and maps the features to a high-dimensional space through η k to obtain edge encoding represents the edge features of the kth layer containing the gap S and the vertical displacement D between the two nodes i and j. Equation (3) multiplies and the corresponding adjacency matrix element value a ij to realize the construction of the electromagnetic topology matrix.

[0068] Then, the self-attention mechanism is calculated on nodes i and j to obtain the normalized attention coefficient , and is taken as the edge weight to represent the differentiation degree of the electromagnetic coupling of each neighborhood node to node i:

[0069]

[0070] where and represent the fusion features of the node features and the corresponding edge features of nodes i and j, respectively, || is the vector concatenation operation, represents the importance of node j to node i, represents a learnable weight matrix (F represents the number of node features, and F' represents the number of edge features), and N(i) represents the neighborhood of node i, represents a learnable weight vector, and (·) T represents the matrix transposition operation, and softmaxj() represents the softmax function. Thus, the joint equation (3), (7) can obtain the graph representation method of the electromagnetic coupling effect weighted by the node features represented by , the electromagnetic topology represented by the matrix C, the edge features represented by , and the edge weight represented by The initial dimension dim = 4 ( [[Delta x, Delta y], S, D] ) is mapped to a dimension dim = 800.

[0071] Figure 5 The GIN (graph isomorphism network) and K-layer GAT (graph attention network) cascaded fusion graph representation model EMGNN architecture diagram is given, Figure 5 Taking K equal to 3 as an example in the middle. The fusion graph representation model includes a cascaded 1-layer graph isomorphism network, K-1-layer graph attention network, graph pooling network and fully connected network; the graph isomorphism network is based on the input containing the initial node feature h (0) The neighborhood aggregation and update are performed with the graph input of the adjacency matrix A, and the first layer updated node feature h (1) is input to the first layer of the graph attention network; the K-1-layer graph attention network is based on h (1) The weighted electromagnetic coupling effect of its own feature and neighborhood is used for message passing and node updating, and the K-1-layer graph attention network outputs the updated node feature h (K) ; the graph pooling network is used to splice the output node feature of the K-1-layer graph attention network with the input node feature to obtain the graph embedding; the fully connected network predicts the output corresponding transfer function according to the graph embedding. The GIN layer is placed before the GAT layer, which improves the graph representation ability of the model. The electromagnetic coupling effect weighted circuit graph representation given by EMCoE-GAT is input into the EMGNN model, and the GIN and GAT cascaded network layer is used to complete the iterative training of the learnable parameters W and delta in formula (6), so as to realize the graph propagation and learning of electromagnetic signals. Finally, the transfer function of the circuit is output. Specifically, in the GAT layer, the weighted electromagnetic coupling effect based on the node itself feature and neighborhood is used for message passing and node updating, which is represented as follows:

[0072]

[0073] Wherein, Based on formula (3) and formula (7), the total electromagnetic coupling effect received by node i is aggregated. Gamma k (·) represents an update function. Taking Figure 4 as an example, based on formula (8), the total electromagnetic coupling effect received by node 2 at the kth layer can be represented as:

[0074]

[0075] The transfer function output by EMGNN reflects the electromagnetic characteristics of the resonant circuit, achieving a deep representation of electromagnetic graph data. Thus, the input to EMGNN is a graph whose features have been represented in a high-dimensional space (node ​​and edge encoding dimensions are both 800), and the output value through the fully connected network has a dimension equal to the number of discrete points in the transfer function, set to 6000 in this example.

[0076] Based on the deep representation of electromagnetic graph data, this example proposes an optimization objective of the joint regularized transfer function of the stopband and transition band to drive the iterative optimization of electromagnetic parameters, thereby achieving high-performance filtering of the target circuit. Furthermore, it introduces a graph representation network weighted by electromagnetic coupling effect with strong representational capabilities and small data of the target electromagnetic specifications as dual-driven cross-electromagnetic topology high-performance graph transfer learning, which solves the problem of simulation of expensive and scarce target circuit labeled data and realizes lightweight design of the target circuit.

[0077] Firstly, this example uses the target transfer function as the approximation objective and automatically updates the electromagnetic parameters in EMGNN using a gradient optimization algorithm. To address the issue of transition band broadening caused by considering only the passband and neglecting transition band fitting, this example proposes a joint regularization method for the stopband and transition band, improving the objective function. Can be defined as With λ2·|w * | 2 A linear combination, more specifically, in this example, is defined as:

[0078]

[0079] Among them, y i Let w represent the transfer function of the i-th resonant filter circuit predicted by the EMGNN model. BP Let w represent the set of passband frequencies, where λ1 and λ2 are learnable weight parameters that balance the transition band width and out-of-band attenuation. i w represents the passband frequency of the i-th resonant filter circuit. * This represents the transition bandwidth, which is calculated from the difference between the passband and stopband widths:

[0080] w * =|w1-w2| (12)

[0081] Wherein, the passband w1 is the frequency region where the transfer function amplitude is less than 3dB of its maximum amplitude, and the stopband w2 is defined as the frequency region where the transfer function amplitude is less than 10dB of its maximum amplitude. By iteratively optimizing the electromagnetic parameters for the target, the attenuation of the transition band was improved, resulting in better filtering performance.

[0082] On the basis of guaranteeing the high-performance filtering of the designed circuit, in order to realize the lightweight design of the cross electromagnetic topology, the example further introduces a cross electromagnetic topology graph transfer learning mechanism driven by the shared parameters of the EMGNN model and the small data of the target electromagnetic specification. Specifically, the source domain is defined as (P source domain samples), and the pth sample p∈P represents a filter data composed of X resonant units, and the target domain is (Q target domain samples), and the qth sample q∈Q represents a filter data composed of Y resonant units, wherein Q << P, and X ≠ Y. The example adopts an improved graph transfer learning algorithm to realize the transfer of the structure and feature knowledge learned by the graph representation model EMGNN in the source domain D s to the target domain D t filter design without negative transfer. The algorithm realizes the further enhancement of the electromagnetic representation ability of the EMGNN specific to the target domain filter while retaining the general electromagnetic features learned in the source domain by selectively freezing and fine-tuning the network layer weights of the pre-trained EMGNN, without relying on a large amount of target domain data to retrain the model.

[0083] Table 2 gives the best freezing-fine-tuning settings of the model. That is, the best settings of the above transfer algorithm under different transfer design tasks are obtained through sequential experiments: when designing across topology structures, the best transfer setting is to fine-tune all layers of the EMGNN, and the number of frozen layers is 0. In the two tasks of designing across the number of units, the best settings are to fine-tune all layers of the EMGNN, freeze the first layer of the EMGNN (i.e. the GIN layer) and fine-tune the remaining layers.

[0084] Table 2 Best freezing-fine-tuning settings of the transfer algorithm under different transfer tasks

[0085]

[0086] In summary, the algorithm flow of the resonant filter circuit design method specifically includes the following steps:

[0087] 1. Construct the EMCoE-GAT network: the nodes represent resonant units, the edges represent electromagnetic coupling effects, and the edge weights represent the coupling strength.

[0088] 2. Construct the EMGNN network and pre-train the EMGNN network based on the source data, and the input is the graph input containing the initial features h (0) of the filter and the adjacency matrix A. The specific training process includes the following steps:

[0089] 1) input h (0) , A;

[0090] 2) For each of the N resonant units, the first layer, i.e., the GIN network, performs neighborhood aggregation and update, and outputs the node features h (1) ; the 2nd to Kth layers, update the adjacency matrix A based on formula (1), and obtain Then based on formulas (3)-(4), obtain the electromagnetic topology matrix C, then based on formulas (5)-(7), obtain the attention coefficient a, and finally based on C, the attention coefficient a, and formula (8), aggregate and update in the EMGNN, and output the node features h (K) ;

[0091] 3) Graph pooling and prediction of transfer function:

[0092] (the output node features of the Kth layer and the input node features are spliced), and the predicted transfer function = the output of the fully connected network;

[0093] 4) Calculate the l1 loss of the predicted value and the true value of the transfer function;

[0094] 5) Backpropagation, update the parameters of the EMGNN model;

[0095] 6) Repeat steps 1) to 5) until the l1 loss of the predicted value and the true value of the transfer function is minimized, and the training is completed;

[0096] 3, Migration of the pre-trained EMGNN model: migration of the EMGNN model based on the settings of Table 2, to obtain the fine-tuned (migrated) EMGNN model;

[0097] 4, Resonant filter lightweight design, including steps:

[0098] 1) Input the target transfer function and other electromagnetic specifications;

[0099] 2) Randomly generate M resonant filter circuits (templates) and perform graph modeling;

[0100] 3) Load the fine-tuned EMGNN model;

[0101] 4) Input the initial features h (0) and the adjacency matrix A of the filter;

[0102] 5) For each randomly generated resonant filter circuit, use the fine-tuned EMGNN model to generate a predicted transfer function, and then calculate and minimize the objective function based on formula (11), and then based on local reparameterization, optimize the input electromagnetic parameters in reverse until is minimized;

[0103] 6) Select the resonant filter circuit corresponding to the smallest loss value, that is, the optimal design result;

[0104] 7) Output the optimized electromagnetic parameters and layout.

[0105] Based on the above resonant filter circuit design method, the embodiment further provides a resonant filter circuit design system, which comprises an input module, a graph representation module, a fusion graph representation module, a migration module and an optimization module, which are respectively used to execute steps S1, S2, S3-S4, S5, S6 in the above resonant filter circuit design method.

[0106] Based on the above resonant filter circuit design method, the embodiment further provides a resonant filter circuit design device, which is provided with an intelligent agent, and the intelligent agent has an input unit, a graph representation unit, a fusion graph representation unit, a migration unit and an optimization unit, which are respectively used to execute steps S1, S2, S3-S4, S5, S6 in the above resonant filter circuit design method.

[0107] Based on the above resonant filter circuit design method, the embodiment further provides a storage medium, which stores a computer program, and the computer program is executed to realize steps S1 to S6 in the above resonant filter circuit design method.

[0108] In summary, the resonant filter circuit design method, system, device and storage medium provided by the embodiment of the present application are based on the graph representation method (EMCoE-GAT) weighted by electromagnetic coupling effect, which models the circuit to be optimized as a weighted graph, realizes the modeling of the complex electromagnetic environment inside the circuit, that is, the node represents the resonant unit, the edge represents the electromagnetic coupling effect between the resonant units, and the edge weight represents the differentiated electromagnetic coupling strength; the above graph structure data is input into the fusion graph representation model (EMGNN) of the graph isomorphism network cascade graph attention network to generate a transfer function, and a model and data double-driven cross-electromagnetic topology graph migration learning mechanism is introduced to migrate and optimize the electromagnetic parameters of the fusion graph representation model (EMGNN), realizing deep representation across electromagnetic topologies, iterative optimization of electromagnetic parameters, and lightweight design across topological structures, which can be applied to the design of isomorphic component circuits of various electromagnetic coupling types and different device types.

[0109] The following experiments are carried out for verification.

[0110] In this example, CircuitGNN proposed by MIT is taken as a baseline model to verify the superiority of the CircuitGTL algorithm; the data set adopts the MIT open-source resonant filter circuit data MIT-CircuitGNN, which contains about 400,000 circuit data of different topologies composed of 3-6 square resonant units, and the feature description is shown in Table 3; when simulating and verifying, the sampling data is divided into training set, verification set and test set according to the ratio of 8:1:1.

[0111] Table 3. Description of Dataset Features

[0112]

[0113] Tables 4 and 5 present the parameter settings for the CircuitGTL algorithm and the performance evaluation metrics for the CircuitGTL algorithm and the resulting filter, respectively. It is worth noting that this example uses the time saved in model convergence to reflect the improvement in computational efficiency; simultaneously, considering the time overhead of establishing source data and migrating data, and due to the difficulty in standardizing quantification, this example uses the reduction in the amount of circuit data required by the model to reflect the time and resource savings in the data establishment phase.

[0114] Table 4 CircuitGTL Algorithm Parameter Settings

[0115]

[0116] Table 5 Evaluation Indicators

[0117]

[0118] Table 6 compares the prediction accuracy and execution time of CircuitGTL and CircuitGNN on the source domain dataset. As shown in Table 6, compared to CircuitGNN, CircuitGTL improves prediction accuracy by 19.4% and reduces convergence time by 36%. These results indicate that CircuitGTL can learn the electromagnetic properties of existing filter data with higher accuracy and efficiency, providing support for subsequent high-precision transfer learning across topologies and unit numbers.

[0119] Table 6 Comparison of prediction performance of CircuitGTL and CircuitGNN on the source domain dataset

[0120]

[0121] The following section completes the performance verification of CircuitGTL's cross-topology transfer learning with the same number of resonant units, as shown in Table 2, Task 1. Table 7 presents the performance prediction results of CircuitGTL and CircuitGNN on the target domain test set. As shown in Table 7, CircuitGTL achieves a lower error than CircuitGNN with 50,000 target domain data points when trained with only 10,000 data points, while saving 31.8% of the time to convergence. The above results indicate that: (1) The electromagnetic coupling effect-weighted graph representation network introduced by CircuitGTL improves the accuracy and computational efficiency of cross-topology performance prediction with the same number of units. (2) The introduced cross-electromagnetic topology graph transfer learning mechanism reduces the target domain data requirement by 80%.

[0122] Table 7 Performance comparison of CircuitGTL and CircuitGNN across topology migration

[0123]

[0124] Figure 6 The transfer function visualization prediction results of CircuitGTL across topology structure are given under the same number of resonant units. It can be seen that: Figure 6 CircuitGTL accurately predicts the transfer function across the topology structure, and the results show the powerful electromagnetic structure graph representation ability of the model.

[0125] Figure 7 The transfer functions of the filters designed by CircuitGTL, CircuitGNN and artificial design (corresponding to (a) (b) (c) respectively) are given. The design task is a bandpass resonant filter with a passband of [220GHz, 250GHz] composed of 4 resonant units. It can be seen that: Figure 7 The passbands of the three design methods are accurately concentrated in the target passband.

[0126] Table 8 gives the average filter performance of CircuitGTL, CircuitGNN and artificial design under 10 different target passband design tasks. From Table 8, it can be seen that: (1) the filters designed by CircuitGTL across topology have the best performance in the three indicators. (2) The design process of CircuitGTL is as fast as CircuitGNN, which can be completed in less than 1 minute on a single GPU (NVIDIA GeForce RTX4090), while artificial design takes several days. The above results show that the filters designed by CircuitGTL have the best filter performance and design efficiency, and the required data amount is only 20% of CircuitGNN, realizing the iterative optimization of electromagnetic parameters and the lightweight of cross-topology design.

[0127] Table 8 Comparison of average performance of CircuitGTL, CircuitGNN and artificial design

[0128]

[0129] The performance verification of CircuitGTL migration across different resonator unit numbers is completed below, i.e., Table 2 Task 2. Table 9 and Table 10 give the performance prediction performance of CircuitGTL, CircuitGNN in different target domain test sets. From Table 9, Table 10, it can be seen that: (1) Migration to a filter with 5 resonant units: CircuitGTL only needs 10,000 data to achieve a 5.4% performance prediction accuracy improvement than CircuitGNN trained on 50,000 target domain data, while the time to run to convergence is saved by 90.9%. (2) Migration to a filter with 6 resonant units: CircuitGTL only needs 10,000 data to achieve a 9.2% performance prediction accuracy improvement than CircuitGNN trained on 50,000 target domain data, while the time to run to convergence is saved by 83.9%. The above results further show that: when facing migration of more complex topology (i.e., different number of resonant units), (1) CircuitGTL introduces an electromagnetic coupling effect weighted graph representation network to improve the accuracy and computational efficiency of performance prediction. (2) The introduced cross-electromagnetic topology graph migration learning mechanism reduces the target domain data requirement by 80%.

[0130] Table 9 Performance comparison of CircuitGTL and CircuitGNN when migrating to 5 resonant units

[0131]

[0132] Table 10 Performance comparison of CircuitGTL and CircuitGNN when migrating to 6 resonant units

[0133]

[0134] Figure 8 The transfer function visualization prediction results of CircuitGTL across unit numbers are given. It can be seen from Figure 8 that CircuitGTL accurately predicts the transfer function, although there are some minor errors in some frequency ranges far from the passband, but the overall results show the strong electromagnetic structure graph representation ability of the model.

[0135] Figure 9 The transfer functions of CircuitGTL, CircuitGNN and artificially designed filters (corresponding to (a) (b) (c) respectively) are given. Among them, the design task is a bandpass resonant filter with a passband of [220GHz, 250GHz] composed of 5 resonant units. It can be seen from Figure 9 that the passbands of the three design methods are accurately concentrated in the target passband.

[0136] Table 11 shows the average filter performance of CircuitGTL, CircuitGNN, and human design (corresponding to (a) (b) (c) among them respectively) in 10 different target passband design tasks. From Table 11, it can be seen that: (1) the filter designed by CircuitGTL performs best in passband coincidence degree and out-of-band suppression ratio, and the insertion loss is slightly lower than CircuitGNN. (2) The CircuitGTL design process is as fast as CircuitGNN, which can be completed in less than 1 minute on a single GPU (NVIDIA GeForce RTX 4090), while an engineer needs to spend several days. The above results show that in the face of more complex topological structure (i.e. different number of resonant units), CircuitGTL only needs 20% of the data amount of CircuitGNN, and realizes the iterative optimization of electromagnetic parameters and lightweight design across the number of units.

[0137] Table 11 Comparison of average performance of CircuitGTL, CircuitGNN and human design

[0138]

[0139] From the above experiments, it can be seen that the numerical results on the MIT-CircuitGNN experimental data set show that: compared with CircuitGNN, CircuitGTL realizes a 9.2% improvement in cross-topology performance prediction accuracy with only 20% of the data amount, and saves 90.9% of the model convergence time; compared with human design, CircuitGTL realizes better filter effect and 4 orders of magnitude design acceleration. The above research can provide a reference for lightweight high-precision electromagnetic structure intelligent forward design.

[0140] In summary, the embodiment proposes a resonant filter circuit design method, system, device and storage medium across electromagnetic topological transfer learning, i.e. CircuitGTL, which realizes graph modeling of complex electromagnetic environment inside the circuit, improves performance prediction accuracy, and realizes intelligent design of the circuit characterized by cross-topology transfer, fast automation and lightweight calculation. Taking the resonant filter as an example, the effectiveness of CircuitGTL is verified. It is worth noting that CircuitGTL is universal because EMCoE-GAT has better electromagnetic coupling expression ability and scalability than traditional electromagnetic formula analysis method, can fully represent the complex electromagnetic environment between devices, and can fully learn the electromagnetic propagation law between multiple types of devices through end-to-end training of EMGNN model; in addition, the graph transfer learning mechanism driven by model data is also independent of specific circuit device types. This means that CircuitGTL can be applied to the design of isomorphic component circuits of different electromagnetic coupling types and different device types.

[0141] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations, simplifications, etc. made without departing from the spirit and principles of the present application should be equivalent replacement manners and should be included in the protection scope of the present application.

Claims

1. A resonant filter circuit design method, characterized in that, Including the following steps: S1. Input the source data composed of multiple resonant filter circuits into the graph representation module weighted by electromagnetic coupling effect; each of the resonant filter circuits is composed of N microstrip open-loop square resonators, i.e., resonant units, and each resonant unit is a node. S2. The graph representation module uses nodes to represent resonant units, edges to represent the electromagnetic coupling effect between resonant units, and edge weights to represent the electromagnetic coupling strength. It models the source data as a weighted graph and inputs it into the fusion graph representation model that introduces a graph isomorphic network cascaded graph attention network. The weighted graph is characterized by the electromagnetic topology of the resonant filter circuit, the node features of each node, the edge features between the node features, and the weights corresponding to the edge features. The node features of each node are obtained by acquiring the first part of the physical parameters of each node as point parameters and then encoding and learning the point parameters by mapping them to a high-dimensional space. The second part of the physical parameters between nodes is obtained as edge parameters, and the edge features corresponding to the edge parameters are calculated; the electromagnetic topology is represented by an N×N electromagnetic topology matrix, and the elements of the electromagnetic topology matrix in the i-th row and j-th column of the k-th layer of the multilayer perceptron are... This represents the electromagnetic coupling effect between node i and node j in the k-th layer. The values ​​of i,j = 1, 2, 3…N are determined based on the second part of the physical parameters of each node, the node features, and the edge features; the weights corresponding to the edge features are determined based on the node features and the edge features. The first part of the physical parameters includes the input / output ports, side lengths, opening orientation, and opening angle of the resonant unit; the second part of the physical parameters includes the difference between the horizontal and vertical coordinates of the center positions of the resonant units, the gap between the resonant units, and the vertical displacement between the resonant units; the electromagnetic topology matrix is ​​the element in the i-th row and j-th column of the k-th layer of the multilayer perceptron. Equal to the normalized adjacency matrix of the resonant filter circuit The element in the i-th row and j-th column Multiply by the edge encoding between node i and node j in the k-th layer ; Defined as: the distance between dissimilar nodes i and j When the distance L is less than the farthest distance L that produces a sufficiently strong electromagnetic coupling effect between node i and node j Equals 1, and vice versa. equal relu(·) represents the ReLU activation function; And when i=j, =1; equal , , Let i and j represent the node features at layer k, respectively. This represents the edge characteristics between node i and node j in the k-th layer. Represents the k-th layer of a multilayer perceptron with the LeakyReLU activation function; edge features The corresponding weight equal , This indicates the importance of node j to node i. and Let represent the fused features of node features and corresponding edge features of nodes i and j, respectively. This represents the softmax function; equal , This represents a learnable weight vector. This represents the learnable weight matrix. This represents the matrix transpose operation. This represents the LeakyReLU activation function. It is a vector concatenation operation; S3. The fusion graph representation model generates the corresponding transfer function prediction value based on the input weighted graph; S4. Update the parameters of the fusion graph representation model based on the predicted value of the transfer function and the actual value corresponding to the source data to obtain a pre-trained fusion graph representation model; S5. Using small data of the target electromagnetic specifications of the circuit to be optimized as the driving force, transfer learning is performed on the pre-trained fusion graph representation model to obtain the transferred fusion graph representation model. S6. For the circuit to be optimized, generate randomly. A resonant filter circuit is used, and after graph modeling, it is input into the transferred fused graph representation model. The target transfer function is used as the approximation target. The electromagnetic parameters of each resonant filter circuit are automatically updated in the transferred fused graph representation model through gradient optimization algorithm. Finally, the resonant filter circuit with the minimum loss value after electromagnetic parameter update is selected as the optimal resonant filter circuit.

2. The resonant filter circuit design method according to claim 1, characterized in that, The fusion graph representation model includes a cascaded 1-layer graph isomorphic network, a K-1 layer graph attention network, a graph pooling network, and a fully connected network; the graph isomorphic network is based on the input features including initial nodes. Adjacency Matrix The graph input is used for neighborhood aggregation and updating to obtain the first layer of updated node features. Input is fed into the first layer of the graph attention network; the K-1 layer graph attention network is based on The weighted electromagnetic coupling effect between the node's own features and its neighborhood is used for message passing and node updates. The K-1 layer graph attention network outputs the updated node features. The graph pooling network is used to combine the output node features of the (K-1)th layer graph attention network. Features of input nodes The graph is spliced ​​together to obtain a graph embedding; the fully connected network predicts and outputs the corresponding transfer function based on the graph embedding. In the pre-training of the fusion graph representation model, the learnable parameters are trained iteratively. .

3. The resonant filter circuit design method according to claim 1 or 2, characterized in that, In step S5, when designing across topologies, the optimal transfer setting is to fine-tune all layers of the pre-trained fusion graph representation model; when designing across two tasks with different numbers of units, the optimal settings are respectively to fine-tune all layers of the pre-trained fusion graph representation model and freeze the first layer of the pre-trained fusion graph representation model and fine-tune the remaining layers.

4. The resonant filter circuit design method according to claim 3, characterized in that, In step S6, the target transfer function is defined as , and linear combination, The fused graph after migration represents the transfer function of the i-th resonant filter circuit predicted by the model. The set representing passband frequencies. Let represent the passband frequency of the i-th resonant filter circuit. Indicates transition bandwidth. It is calculated from the difference between the passband and stopband widths. These are learnable weight parameters.

5. A resonant filter circuit design system, characterized in that, It includes an input module, a graph representation module, a fusion graph representation module, a migration module, and an optimization module, which are respectively used to execute steps S1, S2, S3-S4, S5, and S6 in the resonant filter circuit design method according to any one of claims 1 to 4.

6. A resonant filter circuit design device, characterized in that, An intelligent agent is provided, which has an input unit, a graph representation unit, a fusion graph representation unit, a migration unit, and an optimization unit, respectively used to execute steps S1, S2, S3-S4, S5, and S6 in the resonant filter circuit design method according to any one of claims 1 to 4.

7. A storage medium, characterized in that, It stores a computer program that is executed to implement steps S1 to S6 in the resonant filter circuit design method according to any one of claims 1 to 4.

Images