Radio frequency device modeling method and device based on electromagnetic coupling deep learning model

Abstract

This invention discloses a method and apparatus for modeling radio frequency (RF) devices based on an electromagnetic coupling deep learning model. Multiple RF devices are placed at multiple different relative positions, and the relative positions of their center points and frequencies are recorded for parameter scanning. A batch of Y parameters is generated using a full-wave electromagnetic simulator. em Y em When placed in a circuit simulator to have the same physical meaning as full-wave electromagnetic simulation, the simulation yields Y. Spice Y em With Y Spice The difference is calculated, generating a difference value ΔY. This difference relationship is defined as a coupling effect, and an electromagnetic coupling deep learning model is constructed. The electromagnetic coupling deep learning model is cascaded with a circuit simulator to obtain the overall simulation result. This invention utilizes a circuit simulator to introduce an electromagnetic coupling deep learning model, combining the coupling effect with the originally accurate and fast single RF device AI model, resulting in results that more closely approximate full-wave electromagnetic simulation results.

Description

Technical Field

[0001] This invention belongs to the field of microwave radio frequency engineering technology, and relates to a method and apparatus for modeling radio frequency devices based on an electromagnetic coupling deep learning model. Background Technology

[0002] Modeling radio frequency (RF) devices is an important and common technique. The finite element method (FEM) for RF devices transforms Maxwell's equations into differential equations. FEM solvers can handle structures of arbitrary shapes, such as bond wires, tapered vias, or finite-size substrates. While EM simulation is very useful for RF device analysis, it suffers from a significant drawback: it is extremely time-consuming. For example, the applicant's earlier Chinese patent CN108959728 A proposed a deep learning-based RF circuit modeling and optimization technique, which can significantly reduce the time consumed by EM simulation. However, it focuses on the accurate generation and optimization of scattering parameters for a single RF device. This RF device modeling only considers the case of a single RF device, neglecting the coupling effect when multiple RF device models are cascaded in the circuit design, leading to inaccurate overall design simulation results.

[0003] Therefore, this invention proposes a method and apparatus for modeling radio frequency devices based on a deep learning model of electromagnetic coupling. This method parameterizes the coupling effect between two radio frequency devices to generate a coupling model, which is then used in layout simulation, effectively solving the aforementioned problems. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a method and apparatus for modeling radio frequency devices based on an electromagnetic coupling deep learning model, which effectively improves the errors caused by the lack of consideration of coupling effects after deep learning model cascading.

[0005] The technical solution adopted in this invention is as follows:

[0006] Secondly, the present invention provides a method for modeling radio frequency devices based on an electromagnetic coupling deep learning model, the method comprising the following steps:

[0007] Step (1) Obtain data

[0008] Multiple radio frequency (RF) devices are placed in multiple different relative positions. Then, the relative positions Δx of the center points of two RF devices in the layout design are obtained, as well as the multiple signal frequencies f added to the aforementioned RF devices. req Then, the relative position Δx and signal frequency f req The parameters are combined in pairs, and then a parameter scan batch simulation is performed in a full-wave electromagnetic simulator to obtain the corresponding batch Y parameters (admittance parameters), denoted as Y. em ;

[0009] Y em Substituting the equivalent circuit into the circuit simulator, so that it has the same physical meaning as the layout design, the final simulation yields Y. spice ; Calculate T em With Y spice The difference ΔY is defined as the coupling effect between the two radio frequency devices;

[0010] The coupling effect ΔY is used as the corresponding relative position Δx and frequency information f. req Tags;

[0011] Step (2): Preprocess the data and construct the dataset;

[0012] 2.1 For relative position Δx and frequency information f req The deviation standardization method is used to normalize the data and map it to the interval [-1,1].

[0013] The deviation standardization method is given by formula (1):

[0014]

[0015] Where x min Let x be the minimum value of x. max The maximum value of x is given by x = relative position Δx or frequency information f. req ;

[0016] 2.2 The coupling effect ΔY has a large range of variation in amplitude, phase, or real and imaginary parts, and is prone to individual prominent values; therefore, the Z-Scores method is used to normalize the coupling effect ΔY.

[0017] The Z-Scores method converts scores to standard scores in the following way:

[0018]

[0019] Where μ is the mean of ΔY and σ is the standard deviation of ΔY;

[0020] Step (3): Construct an electromagnetic coupling deep learning model and train, validate and test it using a dataset;

[0021] The electromagnetic coupling deep learning model connects two or more hidden layers between the input and output layers of a multi-layer neural network.

[0022] Each hidden layer contains 100 neurons. A ReLU activation function is added after each hidden layer to introduce non-linearity into the neural network. A dropout layer is added after the ReLU activation function. The entire network uses a fully connected architecture. The output layer has 8 neurons, corresponding to the real part of ΔY11, the imaginary part of ΔY11, the real part of ΔY21, the imaginary part of ΔY21, the real part of ΔY12, the imaginary part of ΔY12, the real part of ΔY22, and the imaginary part of ΔY22.

[0023] The relative position Δx and frequency information f after preprocessing in step (2) above req As the input to the electromagnetic coupling deep learning model, the coupling effect ΔY corresponding to the above radio frequency devices is used as the output of the electromagnetic coupling deep learning model.

[0024] The loss function of the electromagnetic coupling deep learning model calls an optimization algorithm to change the input weight parameter Wi and bias parameter b through backpropagation, thereby reducing the loss value until the termination condition is met.

[0025] The loss function described uses the mean squared error (MSE) function:

[0026]

[0027] Where i represents the i-th sample, y i This represents the true value of the coupling effect ΔY, y' i The model output represents the predicted coupling effect ΔY, where n represents the total number of samples.

[0028] Step (4): The trained, verified and tested electromagnetic coupling deep learning model is placed into the circuit simulator as a circuit component to obtain the circuit simulation results.

[0029] In a second aspect, the present invention provides a radio frequency device modeling apparatus for implementing the method, comprising:

[0030] The trained, validated, and tested electromagnetic coupling deep learning model is responsible for determining the relative positions Δx of the center points of two RF devices in the layout design, as well as the multiple signal frequencies f added to the aforementioned RF devices. req The coupling effect ΔY between the two radio frequency devices was predicted.

[0031] The circuit simulator is responsible for simulating the equivalent circuit based on the layout design requirements and the coupling effect ΔY output by the electromagnetic coupling deep learning model.

[0032] Thirdly, the present invention provides an electronic device including a processor and a memory, the memory storing machine-executable instructions executable by the processor, the processor executing the machine-executable instructions to implement the method.

[0033] Fourthly, the present invention provides a machine-readable storage medium, characterized in that the machine-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the method.

[0034] The beneficial effects of this invention are:

[0035] This invention proposes a novel calculation method for the coupling effect of radio frequency (RF) devices, filling the technical gap in RF device modeling that did not consider coupling issues, and significantly improving the accuracy of RF device modeling and simulation results. This invention utilizes a circuit simulator to introduce a deep learning model of electromagnetic coupling, combining the previously accurate and fast single RF device AI model with coupling effects, resulting in an overall deep learning model design that more closely approximates the full-wave electromagnetic simulation results. Attached Figure Description

[0036] Figure 1 This is a flowchart of the method of the present invention;

[0037] Figure 2 This is a diagram of a spiral inductor structure.

[0038] Figure 3 It is a deep multilayer neural network;

[0039] Figure 4 (a) is the circuit-level spiral inductor AI model structure without coupling effect, and (b) is the overall circuit-level model structure with coupling effect.

[0040] Figure 5 This is a schematic diagram of S-parameter synthesis;

[0041] Figure 6 A comparison chart of the dB values ​​of the S-parameter results of the coupled model and the EM simulation results; Detailed Implementation

[0042] The following detailed description of this technology, in conjunction with the accompanying drawings and implementation examples, will further illustrate the technology.

[0043] like Figure 1 The diagram shown is a flowchart of the radio frequency device modeling method based on an electromagnetic coupling deep learning model according to the present invention. An example of two spiral inductors is provided. The structural diagrams of the two spiral inductors are shown below. Figure 2 As shown.

[0044] Training was performed using Google's deep learning framework TensorFlow. In this case study, a coupled model was constructed using two microstrip rectangular spiral inductors. The two rectangular spiral inductors were arranged vertically on two different metal layers. Multiple relative positions Δx of the two rectangular spiral inductors and multiple signal frequencies f added to the two rectangular spiral inductors were obtained.req The parameters are then combined in pairs and batch simulated using a full-wave electromagnetic simulator to obtain the relative position Δx and frequency f. req The corresponding Y parameter is denoted as Y. em ;

[0045] The Y parameter result after parameter scanning em Substitute the equivalent circuit into a circuit simulator (such as FDSpice and EMOptimizer) so that it has the same physical meaning as the layout design (e.g., Figure 4 (a) shows that the final simulation yields the Y parameter Y. spice ;

[0046] Calculate Y em With Y spice The difference ΔY is defined as the coupling effect between the two radio frequency devices.

[0047] The relative position Δx and frequency information are used as the input value x of the neural network, and the corresponding coupling effect ΔY is used as the label value corresponding to the output of the neural network.

[0048] For position Δx and frequency information f req The coupling effect ΔY continues with data preprocessing, specifically normalizing the input term x using deviation standardization, with the formula:

[0049]

[0050] Where x min Let x be the minimum value of x. max To find the maximum value of x, the input vector x is limited to the interval [-1, 1].

[0051] The coupling effect ΔY is normalized using the Z-Scores method. For the Z-Scores algorithm, if the population mean and population standard deviation are known, the original score ΔY can be converted into a standard score in the following way:

[0052]

[0053] Where μ is the mean of ΔY and σ is the standard deviation of ΔY;

[0054] The purpose is to prevent the learning rate from being constantly adjusted during subsequent optimization along different gradient directions, thus avoiding an inefficient "zigzag" descent process for finding the minimum loss function. This normalization operation also ensures that the Adam optimization algorithm used later can achieve its best results.

[0055] After data preprocessing, the raw data needs to be split. Here, we can directly call the `tarin_test_split` function of the `model_selection` module in the sklearn library to split the true value y, which is composed of the difference between the input terms x and Y parameters ΔY, into 70% of the training dataset x_train and y_train, and 30% of the test dataset x_test and y_test.

[0056] After the data is segmented, a deep multi-layer neural network is constructed as an electromagnetic coupling deep learning model; the multi-layer neural network is represented by the following formula:

[0057] y=σ(W n …σ(W 2 σ(W 1 x+b 1 )+b 2 )...+b n (3)

[0058] Where W is the weight parameter, b is the bias parameter, and σ() is the activation function. After multiple activation functions, the neural network has more nonlinear features. The input term x of the RF device to be trained is set as the input layer of the neural network, and its matrix is ​​determined to be n rows and 3 columns, where n is the number of samples fed into the neural network each time, which is determined according to the batch_size of the samples fed into the neural network (batch_size represents the number of training samples in each batch), and 3 represents the parameters l1, l2, freq of the spiral inductor, that is, there are 3 neurons in the input layer. The scattering S parameter value is set as the output layer in the neural network, and its matrix is ​​determined to be 100 rows and 8 columns (100 represents the number of neurons connected to the previous hidden layer, and 8 represents the real part of ΔY11, the imaginary part of ΔY11, the real part of ΔY21, the imaginary part of ΔY21, the real part of ΔY12, the imaginary part of ΔY12, the real part of ΔY22, and the imaginary part of ΔY22), that is, the number of neurons in the output layer is 8. Next, hidden layers are connected between the input and output layers. Here, we take two hidden layers, each containing 100 neurons. After each neuron, a ReLU activation function is added to introduce a non-linear element into the neural network. The formula for the ReLU activation function is:

[0059] f(x)=max(0,x) (4) This suppresses all inputs less than 0 to the activation function, resulting in an output value of 0. This allows the neural network to fit nonlinear models. It can be seen that the derivative of the ReLU activation function is very simple, with a derivative of 1 for the part greater than 0. This activation function has a fast fitting speed and does not exhibit gradient vanishing during backpropagation. To prevent overfitting and enhance the model's generalization ability, a dropout layer is added after each hidden layer's activation function. The dropout operation randomly discards some neurons from a hidden layer during neural network training, with a dropout ratio of 50%. The entire network uses a fully connected neural network architecture, such as... Figure 3 As shown.

[0060] After building the neural network, define the loss function, using the mean squared error (MSE) function:

[0061]

[0062] Where i represents the i-th sample, y i x represents the input term for training the model. i The corresponding true value, y' i Indicates input item x i The predicted value is obtained through neural network calculation, where n represents the number of samples input into the neural network.

[0063] After defining the mean squared error loss function, call the Adam adaptive optimization algorithm from the Tensorflow library: tf.train.AdamOptimizer(). The formula for the Adam algorithm is:

[0064] Estimate the gradient mean:

[0065] m t =β1m t-1 +(1-β1)g t (6)

[0066] Estimate the squared mean of the gradient:

[0067]

[0068] Consider correcting the gradient mean by the number of iterations:

[0069]

[0070] Consider correcting the mean squared gradient based on the number of iterations:

[0071]

[0072] Update weights and biases:

[0073]

[0074] Where g t Let m be the gradient of the loss function in the t-th iteration, β1 and β2 be the delay factors, and m be the gradient of the loss function in the t-th iteration. t ,v t These are the first-order moment estimates and second-order moment estimates of the gradient, respectively. It is for m t and v t The correction can be applied to approximate an unbiased estimate of the expectation. θ t+1 It is θ t The updated parameters, η and ε, are two constants. During the descent optimization of the loss function, the Adam optimization algorithm uses the number of iterations and a delay factor to correct the mean and mean squared gradients, accelerating the learning speed and efficiency. It can also adjust the learning rate itself. Here, the learning rate base α = 0.001, delay factors β1 = 0.9, β2 = 0.999, and ε = 10 are set. -8 Initially, the iteration count t = 0, and the bias and weight matrix θ0 is a random matrix with a mean of 0 and a very small variance, m0 = 0, v0 = 0. The Adam algorithm makes it less likely that the loss function will get trapped in local minima or saddle points during optimization, causing training to stop.

[0075] Once the optimization algorithm is determined, a session is started in the Tensorflow deep learning framework. The `tf.global_variables_initializer()` function is used to initialize all parameters in the session. The training data `x_train` and `y_train` are passed to the deep neural network model using the `feed_dic` assignment operation in `Session.run()`. Mini-batches learning is used with a batch size of 100, iterating through 100 random samples from all samples each time. After one iteration, the backpropagation algorithm integrated into the AdamOptimizer() optimization algorithm automatically updates the weight parameters W and the bias term b. Thus, after training all samples (one epoch), the W parameters and bias term b have been updated multiple times. During training, epochs are continuously iterated until the error requirement f(θ) < 1 × 10⁻⁶ is met. -2 .

[0076] When the loss value f(θ) of the trained model is less than 1×10 -2If the test set data x_test and y_test are fed into the neural network, and the loss value calculated after feeding the test set data into the neural network is less than the termination condition 1×10, then... -4 Training ends when the time is right, and the change in the loss value is as follows: Figure 4 As shown, the initial loss value fluctuates significantly because each gradient descent iteration is performed on the min-batch dataset rather than the entire dataset. This data variation can cause this effect. However, as the number of training iterations increases, the fluctuation in the loss value decreases, and the loss value converges and continuously approaches 1×10⁻⁶. -4 After training is complete, the `tf.train.Saver()` function is called to save the session containing the deep learning network as a `.ckpt` file.

[0077] Once the electromagnetic coupling deep learning model is obtained, its various ports are compared with the AI ​​model of the original circuit simulator (e.g., Figure 4 (a) The ports are connected sequentially to obtain the overall coupled model (e.g. Figure 4 (b) is shown), and the overall Y parameters are calculated using the Y parameter synthesis method.

[0078] The Y-parameter synthesis process is as follows: First, the S-parameters are converted into Y-parameters. Then, based on Kirchhoff's laws, all Y-parameter matrices are constructed into a large matrix for solving. This large matrix contains information such as excitation, external ports, and interconnections. Finally, the current and voltage of the external ports are obtained, and then converted back to Y-parameters. Finally, the S-parameters of the entire system are obtained through the conversion between S- and Y-parameters. Figure 5 As shown, the equivalent diagram of the multi-port interconnect network circuit simulated by the circuit simulator is a large interconnect network composed of M small sub-networks. Each small sub-network is interconnected with other sub-networks, and each sub-network can be characterized by S-parameters.

[0079] For example Figure 5 For example, suppose the m-th subnetwork has N m Given a network of ports, all ports of all subnets are arranged in a specific order, and a global port number is defined for each port. Then, the global port number k of the p-th port of the m-th subnet is:

[0080] k = t(m) + p (11)

[0081]

[0082] The total number of ports K is then shown in the formula:

[0083]

[0084] Assuming the entire large interconnected network consists of H networks, circuit theory dictates that on each network, all different network ports connected to the same node have the same voltage, and the sum of the currents flowing to that node must be zero. The interconnection between multiple sub-networks typically results in a total number of equations greater than twice H; let's denote the total number of equations as L. Let the current and voltage at the k-th global port be expressed as i... k and v k Therefore, all port current matrices I and all port voltage matrices V can be represented as follows:

[0085]

[0086] Then, the following system equations can be established using I and V:

[0087]

[0088] In the K×K matrix, A1 and A2 represent matrices constructed from all Y parameters of the sub-networks, while the L×K matrices B1 and B2 are constructed from the interconnection relationships between the sub-networks. The unconnected system matrix A and interconnection matrix B are defined as follows:

[0089] A = [A1 A2] (16)

[0090] B = [B1 B2] (17)

[0091] Then, the unconnected system matrix A can be obtained from the Y-matrix of all matrices, and the interconnection matrix B can be obtained from the subnetworks based on their interconnection relationships. It is particularly noteworthy that here, the unconnected system matrix A depends only on the Y-parameter matrices of the subnetworks, while the interconnection matrix B depends only on the interconnection relationships between multiple subnetworks. Matrix A and B are independent of the port excitation signals. In this sense, Equation (14) is a very general description of the system (overall) network; it is a function of the (overall) network topology. Mathematically, Equation (14) is indeterminate because the equation has fewer variables (i.e., the number of networks is usually less than the number of ports), meaning its solution is not unique. Physically, this corresponds to a system with many possible excitations. However, for each set of specified excitations, Equation (14) can give a unique solution. If certain port voltages are specified, the corresponding voltage values ​​will be replaced by other rows (equations) and moved to the right side of Equation (14), and the corresponding variables will be eliminated from Equation (14). After this series of operations, the constructed matrix equations can be solved to obtain a unique solution, that is, the final result of the Y-parameters can be calculated from Equation (14).

[0092] In a special case where all subnets are unconnected, the interconnection matrix B becomes an empty matrix, specifying all port voltages (i.e., energized or grounded). The equations in formula (14) and the total number of unknowns (i.e., port currents) become K. Then, the unconnected system matrix A subnets are calculated based on the Y parameters of A. For the m-th subnet, the equations for port current and voltage are as follows:

[0093] I m =Y m V m (18)

[0094] in:

[0095]

[0096] Y m Let be the m-th subnetwork of the Y matrix, where t(m) is defined by the formula above. The following formula can be written as:

[0097]

[0098] By combining all the subnetworks, the following matrix can be obtained:

[0099]

[0100] In other words, A1 is a K×K identity matrix, A2 is a K×K matrix, and its diagonal blocks are the Y-matrices of the subnetwork.

[0101] Based on matrix B obtained from the internal interconnection of subnetworks, all ports on the same network should have the same voltage, and the sum of the currents of all ports on the same network must be zero. Assume the total number of ports on the nth network is P. n The global port number of the nth network can be represented as s(n,1),…,s(n,P) n Then there is:

[0102]

[0103] v s(n,j) -v s(n,1) =0(j=2,…,P) n ) (twenty four)

[0104] Equation (23) is derived from Kirchhoff's law that the sum of all currents to the same node equals zero. Equation (24) shows that without any excitation at the ports, all port voltages of each subnetwork are equal. For cases with excitation, additional equations will be given in the derivation below. Based on the above two equations: Having P n The nth network with _ ports can obtain a total of P_. nThe equations, that is, the independent equations from any network, are equal to the number of ports on that network. If two networks are connected to the same port, then these interconnected networks must be unified into one. Therefore, the total number of ports on all networks is equal to the total number of equations on all networks. Any interconnection (i.e., network) must connect at least two ports, that is, P n The value must be no less than 2. The equation for the p-th equation from the n-th network is given by the following formula:

[0105]

[0106] The total number of equations across all networks (i.e., the total number of ports across all networks) can be obtained as follows:

[0107]

[0108] Then, the interconnection matrix B1 can be obtained from the current formula (23). Specifically, except for the following non-zero cases, all elements in matrix B1 are zero:

[0109] B1(q(n)+1,s(n,j))=1, j=1,…,P n ;n=1,…,H (27)

[0110] Furthermore, the interconnection matrix B2 can be obtained from formula (24). Specifically, except for the following non-zero cases, all elements in matrix B2 are zero:

[0111] B2(q(n)+j,s(n,j))=1,B2(q(n)+j,s(n,1))=-1,j=2,…,P n ;n=1,…,H (28) In summary, all non-zero terms of the interconnection matrix B can be determined by the above two formulas. The non-zero terms of the interconnection matrix B are determined by the formula B=[B1 B2].

[0112] After filling the above matrices, the filling of the excitation signal in the matrices needs to be processed. At this point, equation (3) has been established, and all submatrices A1, A2, B1, and B2 are known. Based on the above discussion, the interconnect set fully connects L ports, and all other (KL) ports are external ports, which must be excited or grounded to calculate the overall Y parameters. Finally, the port voltages on all (KL) ports are known, and they must be replaced and moved to the right side of equation (14), i.e., the total number of unknowns becomes [2(K-(KL)]=(K+L), equal to the total number of independent equations in equation (14). Figure 6 A comparison of the EM simulation results and the coupled model results shows that the overall coupled model can obtain results that are very close to those of the EM simulation.

[0113] In summary, the AI ​​models of the original circuit-level simulator can already achieve rapid prediction of S-parameters. Introducing coupled models into the overall design will cause the structure obtained by cascading multiple AI models to no longer be independent of each other, further achieving a realistic EM simulation effect on the basis of rapid prediction. The above embodiments are not intended to limit the present invention, and the present invention is not limited to the above embodiments. As long as they meet the requirements of the present invention, they are all within the protection scope of the present invention.

Claims

1. A method for modeling radio frequency devices based on an electromagnetic coupling deep learning model, characterized in that... The method includes the following steps: Step (1), Obtain data: Multiple radio frequency (RF) devices are placed in multiple different relative positions. Multiple relative positions Δx of the center points of two RF devices in the layout design are obtained, along with multiple signal frequencies f added to the aforementioned RF devices. req Then, the relative position Δx and signal frequency f req The parameters are then combined in pairs, and then a parameter scanning batch simulation is performed in a full-wave electromagnetic simulator to obtain the corresponding batch Y parameters, denoted as Y. em ; Y em Substituting the equivalent circuit into the circuit simulator, the equivalent circuit is made to have the same physical meaning as the layout design, and the Y parameter is finally obtained through simulation. spice ; Calculate Y em With Y spice The difference ΔY is defined as the coupling effect between the two radio frequency devices; The coupling effect ΔY is used as the corresponding relative position Δx and frequency information f. req Tags; Step (2): For the relative position Δx and frequency information f req The coupling effect ΔY is preprocessed to construct a dataset; Step (3): Construct an electromagnetic coupling deep learning model and train, validate, and test it using a dataset; use the preprocessed relative position Δx and frequency information f from step (2) above. req As the input to the electromagnetic coupling deep learning model, the coupling effect ΔY corresponding to the above radio frequency devices is used as the output of the electromagnetic coupling deep learning model. Step (4): The trained, verified and tested electromagnetic coupling deep learning model is placed into the circuit simulator as a circuit component to obtain the circuit simulation results.

2. The method according to claim 1, characterized in that... For relative position Δx and frequency information f req The preprocessing process involves using a deviation normalization method to process the relative position Δx and frequency information f. req Normalization is performed, mapping it to the interval [-1, 1].

3. The method according to claim 1, characterized in that... The process of preprocessing the coupling effect ΔY is as follows: the coupling effect ΔY is normalized using the Z-Scores method.

4. The method according to claim 1, characterized in that... The electromagnetic coupling deep learning model employs a multi-layer neural network.

5. The method according to claim 4, characterized in that... The electromagnetic coupling deep learning model connects two or more hidden layers between the input and output layers of a multi-layer neural network.

6. The method according to claim 5, characterized in that... In the electromagnetic coupling deep learning model, multiple neurons are set in each hidden layer, and a ReLU activation function and a dropout layer are added sequentially after each hidden layer.

7. The method according to claim 5, characterized in that... The entire network of the electromagnetically coupled deep learning model uses a fully connected architecture.

8. A radio frequency device modeling apparatus for implementing the method of any one of claims 1-7, characterized in that... include: The trained, validated, and tested electromagnetic coupling deep learning model is responsible for placing multiple RF devices in multiple different relative positions, based on the multiple relative positions Δx of the center points of two RF devices in the layout design, and the multiple signal frequencies f added to the aforementioned RF devices. req The coupling effect ΔY between the two radio frequency devices was predicted. The circuit simulator is responsible for simulating the equivalent circuit based on the layout design requirements and the coupling effect ΔY output by the electromagnetic coupling deep learning model.

9. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to implement the method of any one of claims 1-7.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores machine-executable instructions that, when invoked and executed by a processor, cause the processor to implement the method of any one of claims 1-7.

Images