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

Abstract

The invention relates to the technical field of electronic circuits, and particularly discloses a resonant filter circuit design method, system and device and a storage medium, a to-be-optimized circuit is modeled into a weighted graph on the basis of an electromagnetic coupling effect weighted graph representation method, modeling of a complex electromagnetic environment in the circuit is achieved, and the reliability of the circuit is improved. The nodes represent the resonance units, the edges represent the electromagnetic coupling effect between the resonance units, and the edge weights represent the differentiated electromagnetic coupling strength; inputting the graph structure data into a fusion graph representation model (EMGNN) of a graph isomorphic network and a cascade graph attention network to generate a transfer function, and introducing a model and data dual-driven cross-electromagnetic topological graph transfer learning mechanism to carry out transfer and electromagnetic parameter reverse optimization on the fusion graph representation model (EMGNN); the cross-electromagnetic topology depth representation, electromagnetic parameter iterative optimization and cross-topological structure lightweight design are realized, and 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 invention relates to the technical field of electronic circuits, and in particular to a resonant filter circuit design method, system, device and storage medium. Background Art

[0002] The forward design of resonant filter circuits is essentially an electromagnetic parameter and layout optimization problem driven by performance indicators, which is mostly an NP-complete electromagnetic inverse problem. Traditional methods use a large number of simulations and parameter adjustments of the initial template to iterate and optimize, and rely on manual work, which has the problems of long design cycle and high computational complexity.

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

[0004] The present invention provides a resonant filter circuit design method, system, device and storage medium, and solves the technical problem of how to achieve cross-topology migration, rapid automation, and lightweight computational intelligent design of a resonant filter circuit.

[0005] In order to solve the above technical problems, the present invention provides a resonant filter circuit design method, comprising the steps of:

[0006] S1, inputting source data composed of multiple resonant filter circuits into a graph representation module using electromagnetic coupling effect weighting;

[0007] S2, the graph representation module uses nodes to represent resonance units, edges to represent electromagnetic coupling effects between resonance units, and edge weights to represent electromagnetic coupling strength, and models the source data as a weighted graph and inputs it into a fusion graph representation model that introduces a 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, updating 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, using the small data of the target electromagnetic specifications of the circuit to be optimized as a driver to perform transfer learning on the pre-trained fusion graph representation model 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 migrated fusion 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 migrated fusion graph representation model through the gradient optimization algorithm. Finally, the resonant filter circuit with the smallest loss value after the electromagnetic parameters are updated is taken as the optimal resonant filter circuit.

[0012] Furthermore, each of the resonant filter circuits is composed of N microstrip open-loop square resonators, i.e., resonant units, and each resonant unit serves as a node;

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

[0014] Furthermore, the node features of each node are obtained 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;

[0015] By obtaining the second part of the physical parameters between the nodes as edge parameters, and calculating the edge features corresponding to the edge parameters;

[0016] 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 j-th column of the k-th layer of the multilayer perceptron is represents the electromagnetic coupling effect between nodes i and j in the kth layer, Determine according to the second part of the physical parameters of each node, the node characteristics, and the edge characteristics, i, j = 1, 2, 3 ... N;

[0017] The weight corresponding to the edge feature is determined according to the node feature and the edge feature.

[0018] Further, the first part of the physical parameters includes the input / output port, side length, opening direction and opening angle of the resonance unit; the second part of the physical parameters includes the difference between the horizontal and vertical coordinates of the center positions of the resonance units, the gap between the resonance units and the vertical displacement between the resonance units;

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

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

[0021] equal Respectively represent the node features of node i and node j in the kth layer, represents the edge feature between node i and node j in the kth layer, η k () represents the kth layer of the multilayer perceptron with LeakyReLU activation function;

[0022] Edge Features The corresponding weight equal Indicates the importance of node j to node i, and Respectively represent the fusion features of node features of nodes i and j and the corresponding edge features, softmax j () represents the softmax function;

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

[0024] Furthermore, the fusion graph representation model includes a cascaded 1-layer graph isomorphism network, a K-1-layer graph attention network, a graph pooling network and a fully connected network; the graph isomorphism network is based on the input containing the initial node feature h (0) Perform neighborhood aggregation and update with the graph input of the adjacency matrix A to obtain the node features h after the first layer update (1) Input to the first layer of the graph attention network; K-1 layer graph attention network is based on h (1) The weighted electromagnetic coupling effect of its own features and the neighborhood is used for message transmission and node update. The K-1th layer graph attention network outputs the updated node feature h (K) ; The graph pooling network is used to pool the output node features of the K-1th layer graph attention network With input node features splicing to obtain a graph embedding; the fully connected network predicts the transfer function corresponding to the output according to the graph embedding;

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

[0026] Furthermore, in step S5, when designing across topology structures, the optimal migration setting is to fine-tune all layers of the pre-trained fusion graph representation model; in tasks with two different numbers of units designed across the number of units, 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.

[0027] Further, in step S6, the target transfer function is defined as With λ 2 ·|w * | 2 The linear combination of i The fusion graph after migration represents the transfer function of the i-th resonant filter circuit predicted by the model, w BP represents the 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, λ 1 and λ 2 is a learnable weight parameter.

[0028] The present invention also provides a resonant filter circuit design system, which mainly 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 above-mentioned resonant filter circuit design method.

[0029] The present invention also provides a resonant filter circuit design device, the key of which is that an intelligent agent is provided, 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, and S6 in the above-mentioned resonant filter circuit design method.

[0030] The present invention also provides a storage medium, the key of which is that it stores a computer program, and the computer program is executed to implement steps S1 to S6 in the above-mentioned resonant filter circuit design method.

[0031] The resonant filter circuit design method, system, device and storage medium provided by the present invention model the circuit to be optimized as a weighted graph based on the weighted graph representation method of electromagnetic coupling effect (Electromagnetic Coupling Effect based on Graph Attention, EMCoE-GAT), thereby realizing the modeling of the complex electromagnetic environment inside the circuit, that is, the nodes represent the resonant units, the edges represent the electromagnetic coupling effects between the resonant units, and the edge weights represent the differentiated electromagnetic coupling strengths; the above-mentioned graph structure data is input into the fusion graph representation model (EMGNN) of the graph isomorphic network cascade graph attention network to generate a transfer function, and a cross-electromagnetic topology graph transfer learning mechanism driven by both the model and the data is introduced to migrate the fusion graph representation model (EMGNN) and reversely optimize the electromagnetic parameters, thereby realizing deep characterization across electromagnetic topologies, iterative optimization of electromagnetic parameters, and lightweight design across topology structures, and can be applied to the design of isomorphic component circuits of various different electromagnetic coupling types and different device types. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0033] Figure 2 is an example diagram of a weighted graph representation of electromagnetic coupling effect provided by an embodiment of the present invention;

[0034] Figure 3 is a diagram showing an example of a principle of a resonance unit provided by an embodiment of the present invention;

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

[0036] Figure 5 It is an architecture diagram of the fusion graph representation model (EMGNN) provided in an embodiment of the present invention;

[0037] Figure 6 This is an example diagram of the CircuitGTL cross-topology visualization prediction result provided by an embodiment of the present invention;

[0038] Figure 7 It is a transfer function diagram corresponding to CircuitGTL, CircuitGNN and the manually designed filter provided by the embodiment of the present invention, wherein the green area represents the target passband, and the design task is a bandpass resonant filter composed of 4 resonant units with a passband of [220 GHz, 250 GHz];

[0039] Figure 8This is an example diagram of a CircuitGTL cross-unit quantity visualization prediction result provided by an embodiment of the present invention;

[0040] Fig. 9 It is a transfer function diagram corresponding to the CircuitGTL, CircuitGNN and the manual design results provided by the embodiment of the present invention, wherein the yellow area represents the target passband, and the design task is a bandpass resonant filter with a passband of [220 GHz, 250 GHz] composed of 5 resonant units. DETAILED DESCRIPTION

[0041] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0042] The resonant filter circuit design method provided in the embodiment of the present invention is referred to as the CircuitGTL method / algorithm in this example. The working principle framework is as follows: Figure 1 As shown, the main steps include:

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

[0044] S2, the graph representation module (EMCoE-GAT) 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. The source data is modeled as a weighted graph and input into the fusion graph representation model (EMGNN) that introduces the Graph Isomorphic Network (GIN) cascaded graph attention network (GAT), that is, Figure 1 ② in the;

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

[0046] S4, fit the predicted value with the actual value. If they can fit, the training ends and the pre-trained fusion graph representation model (EMGNN) is obtained. If they cannot fit, the parameters of the fusion graph representation model (EMGNN) are updated based on the predicted value and the actual value (i.e., back propagation) to obtain the pre-trained fusion graph representation model (EMGNN), i.e. Figure 1 ④ in the;

[0047] S5. Driven by the small data of the target electromagnetic specifications of the circuit to be optimized, the pre-trained fusion graph representation model (EMGNN) is transferred to obtain the transferred fusion graph representation model (EMGNN), that is, Figure 1 ⑤ in the;

[0048] S6. For the circuit to be optimized, M initial resonant filter circuits are randomly generated, and after graph modeling, they are input into the migrated fusion graph representation model (EMGNN). The target transfer function is used as the approximation target, and the electromagnetic parameters of each resonant filter circuit are automatically updated in the EMGNN through the gradient optimization algorithm. Finally, the resonant filter circuit with the smallest loss value after the electromagnetic parameters are updated is taken as the optimal resonant filter circuit, that is, Figure 1 ⑥ in the.

[0049] First, this example proposes a weighted graph representation method for electromagnetic coupling effects (EMCoE-GAT), that is, nodes represent resonant units, edges represent electromagnetic coupling effects between resonant units, and edge weights represent electromagnetic coupling strength; then, a fusion graph representation model (EMGNN) of a cascaded graph attention network (Graph Isomorphic Network, GIN) is introduced to achieve cross-domain deep representation of electromagnetic graph data in the source and target domains; finally, an optimization goal of the joint regularized transfer function of the stopband and transition band is proposed, and the iterative optimization of the electromagnetic parameters of the target domain resonant filter is achieved based on the shared parameters of the EMGNN model. CircuitGTL deeply integrates the graph representation capability of graph neural networks and the lightweight capability of graph transfer learning, and realizes the intelligent design of circuits across electromagnetic topology transfer learning.

[0050] An example of a weighted graphical representation of electromagnetic coupling effects (EMCoE-GAT) is shown below: Figure 2 As shown. The weighted graph is characterized by the electromagnetic topology of the resonant filter circuit, the node characteristics of each node, the edge characteristics between the node characteristics, and the weights corresponding to the edge characteristics. The electromagnetic topology of the resonant filter circuit, the node characteristics of each node, the edge characteristics between the node characteristics, and the weights corresponding to the edge characteristics are determined according to the physical parameters of the resonant filter circuit.

[0051] In this example, the resonant filter circuit data MIT-CircuitGNN, which is open source from the Massachusetts Institute of Technology (MIT), is used as the test object. Each resonant filter circuit consists of several microstrip open-loop square resonators (resonant units), and its physical parameters are as follows: Figure 2 As shown in (a) and Table 1, the resonant filter includes three resonant units from left to right, and their center coordinates are (x 1 ,y 1 )、(x 2 ,y2 )、(x 3 ,y 3 ), where the first resonant unit has an input port (Input), the third resonant unit has an output port (Output), the side lengths of the three resonant units are all l, the gaps between the resonant units are S, the vertical displacements between the resonant units are D, the opening angles of the resonant units are θ, and the opening directions of the resonant units are (left or bottom or right or top). Figure 2 The corresponding resonant filter diagram of (a) is shown as Figure 2 as shown in (b).

[0052] Table 1 Vertex and edge features of resonant filter graph modeling

[0053]

[0054] EMCoE-GAT includes node-level modeling networks and edge-level modeling networks, and models the resonant filter as graph data, 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 electromagnetic coupling strength; the point features (node ​​features) and edge features fully model the size information and structural characteristics of the resonant filter.

[0055] By obtaining the first part of the physical parameters of each node as point parameters, and then mapping the point parameters to 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 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, where the element in the i-th row and j-th column of the k-th layer of the multilayer perceptron is represents the electromagnetic coupling effect between nodes i and j in the kth layer, Determine according to the second part of physical parameters, node characteristics, and edge characteristics of each node, i, j = 1, 2, 3 ... N;

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

[0059] As shown in Table 1, this example is for the resonant filter, and its point parameters include input\output port I / O = [1,0] or [0,1] or [0,0] ([1,0] means with input port, [0,1] means with output port, [0,0] means neither with input port nor 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 four directions of left, bottom, right, and top, respectively. If the opening faces left, a is 1, and b, c, and d are all 0) and the opening angle of the resonant unit. The physical information of a single resonant unit is modeled, and then the node features are mapped to a high-dimensional space through a multilayer perceptron (MLP) with LeakyReLU activation to achieve encoding and learning, such as Figure 3 The node feature generation process of the resonance unit is shown in the figure, where the encoded and learned node features are represented by h i At this point, the node feature is mapped from the initial dimension dim=8([I / O,l,θ,dir]) to the dimension dim=800.

[0060] As shown in Table 1, this example is for a resonant filter, and its side parameters include a gap S between resonant units and a vertical displacement D between resonant units.

[0061] The electromagnetic coupling effect comprehensively reflects the factors affecting the electromagnetic coupling strength, such as distance, coupling mode, orientation and angle of the unit opening. EMCoE-GAT characterizes the differentiated electromagnetic coupling strength by modeling the influence of each factor separately, and uses it as the edge weight of electromagnetic signal aggregation, which can realize the graph representation of the complex electromagnetic environment inside the resonant filter circuit, and improve the graph representation ability of the traditional sum neighborhood aggregation. Figure 4 A diagram of the electromagnetic coupling effect is shown in the schematic diagram.

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

[0063]

[0064] in, 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 sufficiently strong electromagnetic coupling effect, and relu(·) represents the relu activation function. ijThe characteristics of electromagnetic coupling strength attenuation with distance can be fully characterized.

[0065] The electromagnetic coupling effect coupling mode, opening direction and angle influence are modeled based on the attention mechanism to characterize the influence degree. represents the electromagnetic topology matrix of the kth layer, where represents the electromagnetic coupling effect between nodes i and j, and C can be expressed as:

[0066]

[0067] Among them, η k () represents the kth layer of the MLP with the LeakyReLU activation function (LeakyReLU()), and Equation (4) combines the edge features with the node features of nodes i, j at the kth layer Perform splicing operation and pass η k () Map features to high-dimensional space to obtain edge encoding represents the edge feature of the kth layer including the gap S and vertical displacement D between two nodes i and j. and the corresponding adjacency matrix element value a ij Multiply to realize the construction of electromagnetic topology matrix.

[0068] Then, the self-attention mechanism is executed on nodes i and j to calculate the standardized attention coefficient and will As the edge weight, it represents the degree of electromagnetic coupling differentiation of each neighboring node to node i:

[0069]

[0070] in, and They represent the fusion features of the node features of nodes i and j and the corresponding edge features, respectively. || is a vector concatenation operation. Indicates the importance of node j to node i, represents the learnable weight matrix (F represents the number of node features, F' represents the number of edge features), N(i) represents the neighborhood of node i, represents the learnable weight vector, (·) T represents the matrix transpose operation, and softmaxj() represents the softmax function. So far, the combined equations (3) and (7) can be obtained as The node characteristics represented by the matrix C represent the electromagnetic topology. The edge features represented by The weighted graph representation method of electromagnetic coupling effect composed of edge weights is represented. The initial dimension dim=4([[Δx,Δy],S,D]) is mapped to dimension dim=800.

[0071] Figure 5 The EMGNN architecture diagram of the fusion graph representation model of GIN (graph isomorphism network) and K-layer GAT (graph attention network) cascade is given. Figure 5 In the example, K is equal to 3. The fusion graph representation model includes a cascaded 1-layer graph isomorphism network, a K-1-layer graph attention network, a graph pooling network, and a fully connected network; the graph isomorphism network is based on the input containing the initial node feature h (0) Perform neighborhood aggregation and update with the graph input of the adjacency matrix A to obtain the node features h after the first layer update (1) Input to the first layer of the graph attention network; K-1 layer graph attention network is based on h (1) The weighted electromagnetic coupling effect of its own features and the neighborhood is used for message transmission and node update. The K-1th layer graph attention network outputs the updated node feature h (K) ; The graph pooling network is used to pool the output node features of the K-1th layer graph attention network With input node features The graph embedding is obtained by splicing; the fully connected network predicts the transfer function corresponding to the output based on the graph embedding. Placing the GIN layer before the GAT layer improves the graph representation ability of the model. The weighted circuit diagram representation of the electromagnetic coupling effect given by EMCoE-GAT is input into the EMGNN model. After the network layer of the cascade of GIN and GAT, the iterative training of the learnable parameters W and δ in formula (6) is completed to realize the propagation and learning of electromagnetic signals on the graph, and finally output the transfer function of the circuit. Specifically, in the GAT layer, the weighted electromagnetic coupling effect based on the node's own characteristics and the neighborhood is used for message transmission and node update, which is expressed as follows:

[0072]

[0073] in, Based on equations (3) and (7), the total electromagnetic coupling effect received by node i is aggregated. k (·) represents the update function. Figure 4 For example, at the kth layer and based on equation (8), the total electromagnetic coupling effect received by node 2 can be expressed as:

[0074]

[0075] The transfer function output by EMGNN reflects the electromagnetic characteristics of the resonant circuit and realizes the deep characterization of electromagnetic graph data. So far, the input of EMGNN is a graph represented by features in high-dimensional space (node ​​and edge encoding dimensions are both 800), and the value dimension output by the fully connected network is the number of discrete points of the transfer function, which is set to 6000 in this example.

[0076] Based on the deep characterization of electromagnetic graph data, this example proposes an optimization target for the joint regularized transfer function of the stopband and transition band to drive the iterative optimization of electromagnetic parameters and achieve high-performance filtering of the target circuit. It further introduces a graph representation network weighted by electromagnetic coupling effects with powerful characterization capabilities and small data of target electromagnetic specifications as dual-driven cross-electromagnetic topology high-performance graph transfer learning, which solves the problem of expensive and scarce target circuit labeled data simulation and achieves lightweight design of the target circuit.

[0077] First, this example takes the target transfer function as the approximation target and automatically updates the electromagnetic parameters in EMGNN through the gradient optimization algorithm. In order to solve the problem that only the passband is considered without considering the transition band fitting, which leads to the widening of the transition band, this example proposes a joint regularization method of the stopband and the transition band, and the improved objective function Can be defined as With λ 2 ·|w * | 2 The linear combination of , more specifically, in this case is defined as:

[0078]

[0079] Among them, y i represents the transfer function of the i-th resonant filter circuit predicted by the EMGNN model, w BP represents the set of passband frequencies, λ 1 and λ 2 is a learnable weight parameter that weighs the transition band width and out-of-band attenuation bonus and penalty, w i represents the passband frequency of the i-th resonant filter circuit, w * represents the transition bandwidth, which is calculated from the difference between the passband and stopband widths:

[0080] w * =|w 1 -w 2 | (12)

[0081] Among them, the passband w 1 The stopband is the frequency region where the transfer function amplitude is 3dB less than its maximum amplitude. 2 Defined as the frequency region where the transfer function amplitude is 10dB less than its maximum amplitude. The iterative optimization of the target traction electromagnetic parameters improves the attenuation of the transition zone and achieves a better filtering effect.

[0082] On the basis of ensuring high-performance filtering of the designed circuit, in order to achieve lightweight design across electromagnetic topologies, this example further introduces a cross-electromagnetic topology graph transfer learning mechanism driven by the shared parameters of the EMGNN model and small data of the target electromagnetic specifications. Specifically, the source domain is defined as (P source domain samples), where the pth sample p∈P represents a filter data consisting of X resonant units, and the target domain is (Q target domain samples), where the qth sample q∈Q represents a filter data consisting of Y resonant units, where Q<<P and X≠Y. This example uses an improved graph transfer learning algorithm to achieve the transfer of the graph representation model EMGNN in the source domain D while avoiding negative transfer. s The structure and feature knowledge learned in the t The algorithm selectively freezes and fine-tunes the weights of each network layer of the pre-trained EMGNN, thereby further enhancing the electromagnetic representation capability of the EMGNN filter specific to the target domain while retaining the general electromagnetic features learned in the source domain, without relying on a large amount of target domain data to retrain the model.

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

[0084] Table 2 Optimal freeze-fine-tune settings for migration algorithms under different migration tasks

[0085]

[0086] Based on the above content, the algorithm flow of the resonant filter circuit design method specifically includes the following steps:

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

[0088] 2. Construct the EMGNN network and pre-train it based on the source data. The input contains the initial feature h of the filter. (0) With the graph input of 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 updates, and outputs the node features h of the first layer (1) ; From the 2nd to the Kth layer, the adjacency matrix A is updated based on formula (1) to obtain Then, based on formulas (3)-(4), the electromagnetic topology matrix C is obtained, and then based on formulas (5)-(7), the attention coefficient α is obtained. Finally, based on C, the attention coefficient α and formula (8), it is aggregated and updated in EMGNN to output the node feature h of the Kth layer. (K) ;

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

[0092] (The output node features of the Kth layer are concatenated with the input node features), prediction transfer function = the output of the fully connected network;

[0093] 4) Calculate the l between the predicted value and the true value of the transfer function 1 loss;

[0094] 5) Back propagation, updating the parameters of the EMGNN model;

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

[0096] 3. Migrate the pre-trained EMGNN model: Migrate the EMGNN model based on the settings in Table 2 to obtain the fine-tuned (migrated) EMGNN model;

[0097] 4. Lightweight design of resonant filter, including the following steps:

[0098] 1) Input electromagnetic specifications such as target transfer function;

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

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

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

[0102] 5) For each randomly generated resonant filter circuit, the fine-tuned EMGNN model is used to generate the predicted transfer function, and then the objective function is calculated and minimized based on formula (11): Then, based on the local reparameterization, the input electromagnetic parameters are optimized inversely until minimize;

[0103] 6) Select the resonant filter circuit corresponding to the minimum loss value, which is the best design result;

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

[0105] Based on the above-mentioned resonant filter circuit design method, this embodiment also provides a resonant filter circuit design system, which 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 above-mentioned resonant filter circuit design method.

[0106] Based on the above-mentioned resonant filter circuit design method, this embodiment also provides a resonant filter circuit design device, which is 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 execute steps S1, S2, S3-S4, S5, and S6 in the above-mentioned resonant filter circuit design method.

[0107] Based on the above-mentioned method for designing a resonant filter circuit, this embodiment further provides a storage medium storing a computer program, which is executed to implement steps S1 to S6 in the above-mentioned method for designing a resonant filter circuit.

[0108] In summary, the resonant filter circuit design method, system, device and storage medium provided by the embodiments of the present invention model the circuit to be optimized as a weighted graph based on the weighted graph representation method of electromagnetic coupling effect (EMCoE-GAT), thereby realizing the modeling of the complex electromagnetic environment inside the circuit, that is, the nodes represent the resonant units, the edges represent the electromagnetic coupling effects between the resonant units, and the edge weights represent the differentiated electromagnetic coupling strengths; the above-mentioned graph structure data is input into the fusion graph representation model (EMGNN) of the graph isomorphic network cascade graph attention network to generate a transfer function, and a cross-electromagnetic topology graph transfer learning mechanism driven by both the model and data is introduced to migrate the fusion graph representation model (EMGNN) and reversely optimize the electromagnetic parameters, thereby realizing deep characterization across electromagnetic topologies, iterative optimization of electromagnetic parameters, and lightweight design across topology structures, which can be applied to the design of isomorphic component circuits of various different electromagnetic coupling types and different device types.

[0109] The following is an experimental verification.

[0110] In this example, the CircuitGNN proposed by MIT is used as the baseline model to verify the superiority of the CircuitGTL algorithm. The data set uses MIT's open source resonant filter circuit data MIT-CircuitGNN, which contains about 400,000 circuit data of different topological structures composed of 3-6 square resonant units. The feature description is shown in Table 3. During simulation verification, the sampled data is divided into training set, verification set and test set in a ratio of 8:1:1.

[0111] Table 3. Dataset feature description

[0112]

[0113] Tables 4 and 5 give the parameter settings of the CircuitGTL algorithm, the performance evaluation indicators of the CircuitGTL algorithm and the obtained filter. It is worth noting that this example uses the time saved from running the model to convergence to reflect the degree of improvement in computational efficiency; at the same time, considering the time overhead of establishing source data and migrating data, since it is difficult to quantify the standard, this example uses the reduction in the amount of circuit data required by the model to reflect the degree of time and resource saving in the data establishment stage.

[0114] Table 4 CircuitGTL algorithm parameter settings

[0115]

[0116] Table 5 Evaluation indicators

[0117]

[0118] Table 6 compares the prediction accuracy and time consumption of CircuitGTL and CircuitGNN in the source domain data set. It can be seen from Table 6 that compared with CircuitGNN, CircuitGTL's prediction accuracy is improved by 19.4% and the time to convergence is reduced by 36%. The above results show that CircuitGTL can learn the electromagnetic characteristics of existing filter data with higher accuracy and efficiency, providing support for subsequent high-precision cross-topology structure and unit number transfer learning.

[0119] Table 6 Comparison of prediction performance between CircuitGTL and CircuitGNN on source domain datasets

[0120]

[0121] Next, we will complete the performance verification of CircuitGTL's cross-topology migration when the number of resonant units is the same, that is, Task 1 in Table 2. Table 7 shows the performance prediction of CircuitGTL and CircuitGNN on the target domain test set. As can be seen from Table 7: CircuitGTL only needs 10,000 data to achieve a lower error than CircuitGNN under 50,000 target domain data training, and the time to convergence is saved by 31.8%. The above results show that: (1) The electromagnetic coupling effect weighted graph representation network introduced by CircuitGTL improves the accuracy and computational efficiency of cross-topology performance prediction under 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 between CircuitGTL and CircuitGNN during cross-topology migration

[0123]

[0124] Figure 6 The transfer function visualization prediction results of CircuitGTL across topologies with the same number of resonant units are given. Figure 6 It can be seen that CircuitGTL accurately predicts the transfer function across topologies, and the results show the model's powerful ability to represent electromagnetic structure diagrams.

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

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

[0127] Table 8 Average performance comparison of CircuitGTL, CircuitGNN and manual design

[0128]

[0129] Next, we will complete the performance verification of CircuitGTL migration across different numbers of resonant units, i.e., Task 2 in Table 2. Tables 9 and 10 show the performance prediction performance of CircuitGTL and CircuitGNN in different target domain test sets. It can be seen from Tables 9 and 10 that: (1) Migrating to a filter with 5 resonant units: CircuitGTL only needs 10,000 data points to achieve a 5.4% improvement in performance prediction accuracy over CircuitGNN trained with 50,000 target domain data, and saves 90.9% of the time required to run to convergence. (2) Migrating to a filter with 6 resonant units: CircuitGTL only needs 10,000 data points to achieve a 9.2% improvement in performance prediction accuracy over CircuitGNN trained with 50,000 target domain data, and saves 83.9% of the time required to run to convergence. The above results further show that: when faced with the migration of more complex topological structures (i.e., different numbers of resonant units), (1) the electromagnetic coupling effect weighted graph representation network introduced by CircuitGTL improves the accuracy and computational efficiency of performance prediction. (2) The introduced cross-electromagnetic topological graph transfer learning mechanism reduces the target domain data requirement by 80%.

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

[0131]

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

[0133]

[0134] Figure 8 The transfer function visualization prediction results of CircuitGTL across the number of units are given. Figure 8 It can be seen that CircuitGTL accurately predicts the transfer function. Although there are some minor errors in certain frequency ranges far away from the passband, the overall results show the model's powerful ability to represent electromagnetic structure diagrams.

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

[0136] Table 11 shows the average filter performance of CircuitGTL, CircuitGNN, and manual design (corresponding to (a), (b), and (c) respectively) under 10 different target passband design tasks. It can be seen from Table 11 that: (1) The filter designed by CircuitGTL performs best in passband overlap and out-of-band suppression ratio, and the insertion loss is slightly lower than that of CircuitGNN. (2) The design process of CircuitGTL is as fast as CircuitGNN, and can be completed in less than 1 minute on a single GPU (NVIDIA GeForce RTX 4090), while engineers need several days. The above results show that: in the face of more complex topological structures (i.e., different numbers of resonant units), CircuitGTL only needs 20% of the data volume of CircuitGNN, achieving iterative optimization of electromagnetic parameters and lightweight design across the number of units.

[0137] Table 11 Average performance comparison of CircuitGTL, CircuitGNN and manual design

[0138]

[0139] From the above experiments, we can see that the numerical results on the MIT-CircuitGNN experimental data set show that: compared with the CircuitGNN algorithm, CircuitGTL achieves a 9.2% improvement in cross-topology performance prediction accuracy with 20% of the data volume, and saves 90.9% of the model convergence time; compared with manual design, CircuitGTL achieves better filtering effects and 4 orders of magnitude design acceleration. The above research can provide a reference for the intelligent forward design of lightweight and high-precision electromagnetic structures.

[0140] In summary, this embodiment proposes a resonant filter circuit design method, system, device and storage medium for cross-electromagnetic topology transfer learning, namely CircuitGTL, which realizes the graph modeling of the complex electromagnetic environment inside the circuit and improves the performance prediction accuracy; it also realizes the intelligent design of circuits characterized by cross-topology migration, rapid automation, and lightweight calculation. The effectiveness of CircuitGTL was verified using the resonant filter as a case study; it is worth noting that CircuitGTL is universal, because EMCoE-GAT has better electromagnetic coupling expression ability and scalability than the traditional electromagnetic formula analysis method, can fully characterize the complex electromagnetic environment between devices, and can fully learn the electromagnetic propagation laws between multiple types of devices through end-to-end training of the EMGNN model; in addition, the graph transfer learning mechanism driven by both model data and circuit data does not depend on specific circuit device types. This means that CircuitGTL can be applied to the design of isomorphic component circuits with different electromagnetic coupling types and different device types.

[0141] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A resonant filter circuit design method, characterized in that: Includes steps: S1, inputting source data composed of multiple resonant filter circuits into a graph representation module using electromagnetic coupling effect weighting; S2, the graph representation module uses nodes to represent resonance units, edges to represent electromagnetic coupling effects between resonance units, and edge weights to represent electromagnetic coupling strength, and models the source data as a weighted graph and inputs it into a fusion graph representation model that introduces a graph isomorphism network cascade graph attention network; S3, the fusion graph representation model generates a corresponding transfer function prediction value based on the input weighted graph; S4, updating 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; S5, using the small data of the target electromagnetic specifications of the circuit to be optimized as a driver to perform transfer learning on the pre-trained fusion graph representation model to obtain a transferred fusion graph representation model; S6. For the circuit to be optimized, M resonant filter circuits are randomly generated, and after graph modeling, they are input into the migrated fusion 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 migrated fusion graph representation model through the gradient optimization algorithm. Finally, the resonant filter circuit with the smallest loss value after the electromagnetic parameters are updated is taken as the optimal resonant filter circuit.

2. The resonant filter circuit design method according to claim 1, characterized in that: Each of the resonant filter circuits is composed of N microstrip open-loop square resonators, i.e., resonant units, and each resonant unit serves as a node; The weighted graph is characterized by the electromagnetic topology of the resonant filter circuit, the node characteristics of each node, the edge characteristics between the node characteristics, and the weights corresponding to the edge characteristics; the electromagnetic topology of the resonant filter circuit, the node characteristics of each node, the edge characteristics between the node characteristics, and the weights corresponding to the edge characteristics are determined according to the physical parameters of the resonant filter circuit.

3. The method for designing a resonant filter circuit according to claim 2, wherein: 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; By obtaining the second part of the physical parameters between the nodes as edge parameters, and calculating the edge features corresponding to the edge parameters; 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 j-th column of the k-th layer of the multilayer perceptron is represents the electromagnetic coupling effect between nodes i and j in the kth layer, Determine according to the second part of the physical parameters of each node, the node characteristics, and the edge characteristics, i, j = 1, 2, 3 ... N; The weight corresponding to the edge feature is determined according to the node feature and the edge feature.

4. The method for designing a resonant filter circuit according to claim 3, wherein: The first part of the physical parameters includes the input / output port, side length, opening direction and opening angle of the resonance unit; the second part of the physical parameters includes the difference between the horizontal and vertical coordinates between the center positions of the resonance units, the gap between the resonance units and the vertical displacement between the resonance units; The element of the electromagnetic topology matrix in the i-th row and j-th column of the k-th layer of the multilayer perceptron is Equal to the normalized adjacency matrix of the resonant filter circuit The element a in the i-th row and j-th column of ij Multiply by the edge code between nodes i and j at the kth layer a ij It is defined as: when the distance Δl between different nodes i and j is less than the farthest distance L between nodes i and j that produces a sufficiently strong electromagnetic coupling effect, a ij is equal to 1, otherwise, a ij equal relu(·) represents the relu activation function; And when i=j, a ij =1; equal Respectively represent the node features of node i and node j in the kth layer, represents the edge feature between node i and node j in the kth layer, η k () represents the kth layer of the multilayer perceptron with LeakyReLU activation function; Edge Features The corresponding weight equal Indicates the importance of node j to node i, and Respectively represent the fusion features of node features of nodes i and j and the corresponding edge features, softmax j () represents the softmax function; equal δ represents the learnable weight vector, W represents the learnable weight matrix, (·) T represents the matrix transpose operation, LeakyReLU() represents the LeakyReLU activation function, and || is the vector concatenation operation.

5. The method for designing a resonant filter circuit according to claim 4, characterized in that: The fusion graph representation model includes a cascaded 1-layer graph isomorphism network, a K-1-layer graph attention network, a graph pooling network and a fully connected network; the graph isomorphism network is based on the input containing the initial node feature h (0) Perform neighborhood aggregation and update with the graph input of the adjacency matrix A to obtain the node features h after the first layer update (1) Input to the first layer of the graph attention network; K-1 layer graph attention network is based on h (1) The weighted electromagnetic coupling effect of its own features and the neighborhood is used for message transmission and node update. The K-1th layer graph attention network outputs the updated node feature h (K) ; The graph pooling network is used to pool the output node features of the K-1th layer graph attention network With input node features splicing to obtain a graph embedding; the fully connected network predicts the transfer function corresponding to the output according to the graph embedding; In the pre-training of the fusion graph representation model, the learnable parameters W and δ are iteratively trained.

6. The method for designing a resonant filter circuit according to any one of claims 1 to 5, characterized in that: In step S5, when designing across topology structures, the optimal migration setting is to fine-tune all layers of the pre-trained fusion graph representation model; in tasks with two different numbers of units designed across the number of units, 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.

7. The method for designing a resonant filter circuit according to claim 6, characterized in that: In step S6, the target transfer function is defined as and λ2·|w * | 2 The linear combination of i The fusion graph after migration represents the transfer function of the i-th resonant filter circuit predicted by the model, w BP represents the set of passband frequencies, w i represents the passband frequency of the i-th resonant filter circuit, w * represents the transition bandwidth, w * It is calculated from the difference between the passband and stopband widths. λ1 and λ2 are learnable weight parameters.

8. 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 5.

9. 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, which are respectively used to execute steps S1, S2, S3-S4, S5, and S6 in the resonant filter circuit design method described in any one of claims 1 to 5.

10. A storage medium, characterized in that A computer program is stored therein, and the computer program is executed to implement steps S1 to S6 in the method for designing a resonance filter circuit according to any one of claims 1 to 5.

Images