Neural network-based TGV model S parameter rapid prediction method and system

Abstract

The invention provides a neural network-based TGV model S parameter rapid prediction method and system. The method comprises the steps of setting TGV structure parameters of multiple groups of combination schemes, obtaining corresponding S parameter data through an electromagnetic simulation tool, constructing a simulation data set, training a neural network model after preprocessing and feature extraction, and inputting the TGV structure parameters into the trained model to quickly obtain corresponding S parameters. And meanwhile, through a neural network reverse optimization module, in combination with a genetic algorithm and a fitness function, an optimal TGV structure parameter is reversely deduced according to a target performance demand. The system comprises a simulation data set module, a neural network model module and a reverse optimization module. According to the method and the system, full-wave simulation does not need to be repeatedly carried out, S parameter prediction can be completed only through millisecond-level calculation time, the design time is greatly saved, the TGV structure design and optimization efficiency is improved, multiple performance index requirements can be accurately matched in a large-scale parameter space, and efficient adaptation of design and machining is achieved.

Description

Technical Field

[0001] This invention mainly relates to the fields of electromagnetic simulation and high-speed interconnect design. Specifically, this invention relates to a method and system for fast prediction of S-parameters of a TGV model based on a neural network. Background Technology

[0002] Through-Glass Vias (TGVs), as vertical interconnect vias in glass substrates, are key structures in high-density packaging and RF systems. Existing methods for analyzing the electromagnetic characteristics of TGVs mainly fall into three categories: The first is the full-wave simulation method, represented by HFSS, which solves the electromagnetic field numerically through adaptive mesh generation, accurately calculating S-parameters under different parameter combinations. However, this method suffers from significant efficiency bottlenecks; a single simulation takes several minutes, and in multi-parameter optimization scenarios, repeated modeling iterations are required, leading to a substantial extension of the design cycle and making it difficult to meet the rapid development needs of high-density packaging. The second category is the equivalent circuit model method, which extracts the lumped parameters of the TGV to construct a π-type circuit and then uses tools such as HSPICE for simulation. While this method shortens the computation time, the modeling process depends on the accuracy of parameter extraction, which cannot match the accuracy requirements of high-frequency scenarios. The third category is analytical numerical methods, which directly calculate transmission characteristics by deriving mathematical models. However, the application scenarios of this type of method are limited, and it can only cover simple hole structures. It lacks adaptability to actual structural features such as conical holes and sidewall roughness caused by processes such as laser ablation. In the millimeter wave band, the skin effect is aggravated, and the deviation from the measured results is further expanded.

[0003] Existing neural network-based TGV model S-parameter prediction and auxiliary optimization schemes struggle to balance the dual technical requirements of accuracy and speed. Furthermore, a method for deriving optimal TGV structural parameters from performance requirements is lacking. These technical bottlenecks severely restrict the large-scale application of TGV technology. Therefore, a method and system for rapidly predicting TGV model S-parameters based on neural networks and deriving optimal TGV structural parameters from performance requirements is urgently needed. Summary of the Invention

[0004] Based on existing technologies, the objective of this invention is to propose a method and system for rapid prediction of S-parameters of TGV models based on neural networks. This method can quickly obtain the S-parameters of the TGV structure without the need for full-wave simulation, by pre-establishing a TGV simulation dataset and using neural networks for learning and modeling. This saves design time and accelerates the determination of the optimal transmission line scheme.

[0005] According to the present invention, the above-mentioned problem is solved by a method and system for fast prediction of S-parameters of a TGV model based on a neural network.

[0006] In a first aspect, this invention proposes a fast prediction method for S-parameters of a TGV model based on a neural network, comprising the following steps: Step S100: Provide a parameterized TGV model and set the TGV structural parameters for multiple combination schemes; Step S200: For the TGV structural parameter combination scheme in step S100, obtain the corresponding S-parameter data through electromagnetic simulation tools; Step S400: Provide a simulation dataset, wherein each sample in the simulation dataset includes a set of TGV structural parameters and corresponding S-parameter data; Step S500: Provide a neural network model and train the neural network model using the simulation dataset; and Step S600: Input the TGV structure parameters into the neural network model trained in step S400 to obtain the corresponding S parameters.

[0007] Furthermore, the TGV structural parameters include: aperture, aperture spacing, dielectric thickness, dielectric constant, and / or filling material.

[0008] Furthermore, the setting range of the TGV structural parameters includes: The pore size range was set to 5 μm to 100 μm; The aperture spacing range was set to 55μm to 200μm; The medium thickness range is set to 50 μm to 500 μm; The dielectric constant is set to a range of 2 to 8; and The types of filler materials include copper, tungsten, and / or carbon nanotubes.

[0009] Further, in step S100, the design method for the TGV structural parameter combination scheme includes: Step S101: The structural parameters of the TGV are sampled at equal intervals within the set range to obtain the first sampling point; Step S102: Add random perturbation to a portion of the first sampling points to obtain second sampling points; Step S103: Cross-combine the second sampling points to form an initial TGV parameter combination; Step S104: Based on actual process constraints, the initial TGV parameter combination is screened to obtain the TGV parameter combination used to construct the simulation dataset.

[0010] Furthermore, the method also includes: Step S300: Preprocess and / or extract features from the S-parameter data in step S200.

[0011] Furthermore, the neural network model is a multilayer perceptron (MLP) neural network model, comprising: The input layer has n neurons, where n is the number of types of TGV structural parameters; Hidden layers, including first to third hidden layers, with ReLU activation function; and The output layer is configured to output the S-parameters at the target frequency or the S-parameter characteristics across the entire frequency band.

[0012] Furthermore, the method also includes: Step S701: Input the target performance requirement index into the neural network reverse optimization module; Step S702: Initialize a set of TGV structural parameters; Step S703: Input the TGV structural parameters into the neural network model to obtain the predicted S parameters; Step S704: Provide a fitness function and calculate the fitness function based on the deviation from the target performance index requirement; and Step S705: Use a genetic algorithm for iterative optimization, and output new TGV structural parameters through crossover and / or mutation to gradually approach the target. Step S706: After several iterations, output the optimal TGV structure parameters.

[0013] Further, in step S701, the target performance requirement indicators include: Insertion loss, return loss, bandwidth and / or impedance matching.

[0014] Further, in step S704, the fitness function is: in: , The predicted performance parameter values ​​output by the neural network model; , The target performance index value; , For predicted and target values ​​of bandwidth and other performance indicators; Wi(i=1,2,3,...) represents the weight coefficient of the corresponding indicator.

[0015] A second aspect of the present invention also proposes a fast prediction system for S-parameters of a TGV model based on a neural network, characterized in that it comprises: The simulation dataset module is configured to store TGV structural parameters of multiple combination schemes, as well as S-parameter data obtained by inputting the TGV structural parameters into an electromagnetic simulation tool, providing a dataset for neural network model training and verification. A neural network model module is configured to include a multilayer perceptron neural network model, the neural network model comprising: The input layer has n neurons, where n is the number of types of TGV structural parameters; Hidden layers, including first to third hidden layers, with ReLU activation function; and The output layer is configured to output S-parameters at the target frequency or S-parameter characteristics across the entire frequency band; and The neural network inverse optimization module is configured to back-calculate the optimal TGV structure parameters when given input performance requirements.

[0016] The present invention proposes a fast prediction method and system for S-parameters of a TGV model based on neural networks, which has at least the following advantages: First, the method employs a stratified sampling approach with added random perturbations to generate TGV parameter combination samples. These TGV parameter combination samples cover the entire TGV parameter space, avoiding learning bias in the neural network model. They are uniformly distributed, conforming to manufacturability requirements, and the addition of random perturbations significantly improves generalization ability, ensuring sufficient and reliable data for training the neural network model. A TGV electromagnetic simulation dataset is constructed, and the neural network model is used for learning and modeling. In practical applications, simply inputting the TGV structural parameters (aperture, spacing, material, etc.) into the trained neural network model directly yields the corresponding S-parameter results. Compared to traditional HFSS simulations, this process requires only milliseconds of computation time, eliminating the need for hours of electromagnetic solution work, simplifying complex geometric modeling and parameter setting processes, and greatly improving computational efficiency.

[0017] Secondly, the method can quickly obtain S-parameter prediction results in a large-scale parameter space, which improves design and optimization efficiency, thereby saving design time and accelerating the determination of the optimal transmission line scheme.

[0018] Furthermore, the method and system described above can reverse-engineer the combination of TGV structural parameters that meet the performance requirements from the target performance indicators. This combination of parameters can be directly used in subsequent simulation verification or process design stages, realizing automatic reverse-engineering optimization from performance requirements to structural parameters. Attached Figure Description

[0019] To further illustrate the advantages and other features of the various embodiments of the present invention, a more specific description of the embodiments of the present invention will be presented with reference to the accompanying drawings. It is understood that these drawings depict only typical embodiments of the invention and are therefore not intended to limit its scope. In the drawings, identical or corresponding parts will be indicated by the same or similar reference numerals for clarity.

[0020] Figure 1 A flowchart illustrating a fast prediction method for S-parameters of a TGV model based on a neural network is shown.

[0021] Figure 2 A schematic diagram illustrating the specific functions of a fast prediction method for S-parameters of a TGV model based on a neural network is shown. Detailed Implementation

[0022] It should be noted that the components in the various figures may be shown exaggeratedly for illustrative purposes and are not necessarily to scale. In each figure, the same reference numerals are used for components that are identical or have the same function.

[0023] In this invention, the various embodiments are merely intended to illustrate the solutions of the invention and should not be construed as limiting.

[0024] In this invention, unless otherwise specified, the quantifiers “a” and “one” do not exclude scenarios involving multiple elements.

[0025] It should also be noted that, for clarity and simplicity, only a portion of the components may be shown in the embodiments of the present invention. However, those skilled in the art will understand that, under the teachings of the present invention, the required components can be added according to specific needs. Furthermore, unless otherwise stated, features in different embodiments of the present invention can be combined with each other. For example, a feature in the second embodiment can replace a corresponding or functionally identical or similar feature in the first embodiment, and the resulting embodiment will also fall within the scope of disclosure or description of this application.

[0026] It should also be noted that within the scope of this invention, the terms "same", "equal", and "equal to" do not mean that the two values ​​are absolutely equal, but allow for a certain reasonable error. In other words, the terms also cover "substantially the same", "substantially equal", and "substantially equal to".

[0027] Furthermore, the numbering of the method steps in this invention does not limit the execution order of the method steps. Unless otherwise specified, the method steps may be executed in different orders.

[0028] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0029] This invention proposes a fast prediction method for S-parameters of a TGV model based on neural networks, comprising the following steps: Step S100: Establish a parameterized TGV model and set the TGV structural parameters for multiple combination schemes; Step S200: Obtain the corresponding S-parameter data for the TGV structural parameters in step S100 using an electromagnetic simulation tool; Step S300: Preprocess and / or extract features from the S-parameter data in step S200; Step S400: Construct a simulation dataset, wherein each sample in the simulation dataset includes a set of TGV structural parameters and corresponding S-parameter data; Step S500: Construct a neural network model and train the neural network model using the simulation dataset; and Step S600: Input the TGV structure parameters into the neural network model trained in step S400 to obtain the corresponding S parameters.

[0030] Figure 1 A flowchart illustrating a fast prediction method for S-parameters of a TGV model based on a neural network is shown. Figure 1 As shown, in one embodiment of the present invention, Ansys HFSS (hereinafter referred to as HFSS) is used as the electromagnetic simulation tool. Multiple combinations of TGV structural parameters are set, and HFSS is used to perform full-wave simulation on the multiple combinations of TGV structural parameters.

[0031] In one embodiment of the present invention, the TGV structural parameters include: Aperture refers to the inner diameter of the through-hole that vertically penetrates the glass substrate in a TGV (Transient Voltage Regulator), i.e., the maximum distance between the inner walls of the through-hole. It is one of the most fundamental geometric parameters of a TGV. The aperture directly determines the cross-sectional area of ​​the filler metal. A larger aperture allows for more metal filler, resulting in lower DC resistance of the TGV and reduced ohmic losses during signal transmission. Simultaneously, the aperture affects the parasitic capacitance of the TGV (as the aperture increases, the coupling area between the aperture wall and the surrounding structure increases, potentially increasing parasitic capacitance), thus impacting the delay and bandwidth characteristics of high-frequency signals. The adjustable range of the aperture is set from 5μm to 100μm.

[0032] The via pitch, referring to the distance between the centers of two adjacent TGV vias, is a key parameter for measuring TGV packaging density and crosstalk control. A smaller via pitch results in stronger electromagnetic field coupling between adjacent TGVs, leading to a decrease in isolation in the S-parameters due to crosstalk signals (such as interference noise), thus affecting signal integrity. Conversely, an excessively large pitch occupies more glass substrate area, contradicting the purpose of high-density packaging. The adjustable range for the via pitch is set from 55μm to 200μm.

[0033] Dielectric thickness, referring to the overall thickness of the glass substrate containing the TGV (i.e., the vertical length of the TGV via), is a key parameter determining the TGV transmission path length. Dielectric thickness is equivalent to the signal transmission path length of the TGV; a larger dielectric thickness results in a longer signal transmission distance within the TGV, but also a more pronounced skin effect and dielectric loss (dielectric loss of the glass material) at high frequencies. Simultaneously, dielectric thickness affects the characteristic impedance of the TGV. If the thickness does not match the designed impedance, it will lead to increased signal reflection, affecting transmission integrity. The adjustable range for dielectric thickness is set from 50μm to 500μm.

[0034] The dielectric constant, an inherent electrical property of glass dielectric materials, describes the material's ability to store charge in an electric field and is a core material parameter determining the electromagnetic characteristics of a TGV (Transient Voltage Regulator). The dielectric constant directly determines the parasitic capacitance of the TGV. A smaller dielectric constant results in smaller parasitic capacitance, significantly reducing signal delay (RC delay) at high frequencies and decreasing dielectric loss, thus optimizing the bandwidth characteristics of the S-parameters. Conversely, an excessively large dielectric constant leads to a surge in parasitic capacitance, limiting the signal transmission rate. The adjustable range of the dielectric constant is set to 2-8.

[0035] Filler metal is the core functional material for achieving "vertical conductivity" in TGVs. Since glass is an insulator, conductive material must be filled within the vias to allow signals / currents to vertically penetrate the glass substrate, connecting the upper and lower circuit layers. The differences in conductivity, high-temperature resistance, and mechanical strength of different metals directly determine the transmission loss, lifespan, and application scenarios of TGVs. Filler metals can be selected from conductive materials such as copper, tungsten, and carbon nanotubes.

[0036] In one embodiment of the present invention, within the adjustable range of the aforementioned parameters, multiple TGV parameter combination schemes are designed and full-wave simulation is performed using HFSS to extract their corresponding S-parameters. A TGV electromagnetic simulation dataset is constructed, wherein each sample in the dataset contains the input TGV structural parameters and the corresponding S-parameter curves. In one embodiment of the present invention, the design method of the TGV structural parameter combination scheme includes: In one embodiment of the present invention, to cover the electromagnetic responses of different TGV structures across the entire parameter space, a layered sampling method incorporating random perturbations is used to generate TGV structure parameter combination samples. Specifically, this includes the following steps: Step S101: The TGV structural parameters are sampled uniformly at equal intervals within the set range to obtain a first sampling point. Specifically, in one embodiment of the present invention, taking the aperture d as an example, uniform sampling can be performed within the range, and the first sampling point can be: d=10μm, 20μm, 30μm, 40μm, 50μm, 60μm… In one embodiment of the present invention, the first sampling point of the spacing p can be: p=60μm, 80μm, 100μm, 120μm, 150μm… In one embodiment of the present invention, the first sampling point of thickness h can be: h=100μm, 150μm, 200μm, 300μm, 400μm… By using the equally spaced sampling in the above embodiments, the first sampling point of each TGV structural parameter is obtained as the basic sampling point.

[0037] Step S102: Add random perturbations to a portion of the first sampling points to obtain second sampling points. Specifically, in one embodiment of the present invention, based on the uniform sampling in step S101, small perturbations are added to some parameters to simulate actual processing errors and / or process fluctuations. For example: d=30μm±3μm p=100μm±5μm h = 200μm ± 10μm Step S103: Cross-combine the second sampling points and the type of filling material to form an initial TGV parameter combination. After sampling the second sampling points of each parameter, add the material parameters, and cross-combine the sampling points and material parameters of each parameter to obtain the initial TGV structural parameter combination.

[0038] Step S104: Based on actual process constraints, the initial TGV parameter combinations are filtered to obtain TGV parameter combinations used to construct the simulation dataset. Specifically, in one embodiment of the present invention, the actual process constraints include: There is a limiting relationship between aperture and thickness; The spacing generally needs to be at least 40 μm larger than the aperture; and The dielectric constant must be set within the range that the type of filler material can satisfy.

[0039] The selection process ensures that all generated structural parameter combinations are manufacturable and effective structures. After the above sampling, combination, and selection, effective TGV structural parameter combinations are obtained and used to construct the training dataset.

[0040] In one embodiment of the present invention, in step S300, in order to ensure and / or improve the training effect of the neural network model, data preprocessing and feature extraction operations can be performed on the TGV simulation dataset.

[0041] In one embodiment of the present invention, the collected S-parameter data is preprocessed as follows: Step S301: Frequency Range Unification and Data Alignment. To address the potential inconsistency in frequency scanning steps across different simulation models, the frequency range of the S-parameter curves for all samples is first unified. For example, in one embodiment of this invention, the frequency range is unified to 0.1 GHz to 40 GHz, and sampling is performed at fixed steps. For samples with inconsistent frequency points, their S-parameters are mapped to a unified frequency grid using interpolation, ensuring that all samples have a consistent dimension at each frequency point, thus maintaining a consistent input format for subsequent neural network models.

[0042] Step S302: Separation and normalization of amplitude and phase. Since the S-parameters are in complex form, in one embodiment of the present invention, they are decomposed into amplitude and phase. The amplitude can be represented by modulus or dB, and the phase is extracted from the complex angle and the phase transitions are expanded to make them present a continuous and monotonic variation trend across the entire frequency band, avoiding interference to training due to abrupt phase changes.

[0043] Step S303, outlier screening and data smoothing: In some simulation results, local outliers may occur due to factors such as mesh generation and solution settings. To ensure the quality of training data, the amplitude and phase curves will be checked. If obvious peaks or abrupt changes occur at certain frequency points, smoothing or removal of outlier data points will be performed to ensure the continuity and trainability of the input data curves.

[0044] Step S304: Normalization of structural parameters. TGV structural parameters (such as aperture, aperture spacing, dielectric thickness, dielectric constant, etc.) typically have different physical dimensions and orders of magnitude. To eliminate the influence of dimensions, in one embodiment of the present invention, all input parameters are normalized to ensure they are distributed within the same numerical range (e.g., [0,1]). Normalization reduces gradient differences during training and improves model convergence speed.

[0045] Step S305, vectorization of S-parameter features: After unifying the frequency points, separating amplitude and phase, and completing preprocessing, the entire S-parameter curve is converted into a fixed-dimensional feature vector. For example, the amplitude and phase of each frequency point are concatenated sequentially to form a complete one-dimensional vector. If multiple parameters such as S11 and S21 are included, they are processed separately and then concatenated as a whole to ensure that each sample has a feature representation of the same dimension.

[0046] Step S306: In one embodiment of the present invention, optionally, if the dataset is large (e.g., a large number of samples, dense sampling frequency leading to high feature dimensionality), then principal component analysis (PCA) is used to perform dimensionality reduction on the extracted features, while retaining the main information (core features) of the data, and removing redundant or highly correlated features.

[0047] In one embodiment of the present invention, a multilayer perceptron (MLP) neural network model is constructed, specifically, its network structure includes: The input layer is configured to receive n TGV geometry and material parameters.

[0048] The hidden layers include first to third hidden layers, each of which includes: The linear layer performs a linear transformation on the output features of the previous layer. By multiplying the weight parameters with the matrix output of the previous layer and then adding the bias parameters, the features of the previous layer are mapped to a new feature space. This provides a basic linear feature representation for deeper feature extraction in subsequent hidden layers, ensuring that feature information can be effectively transmitted and initially transformed in the network.

[0049] Batch normalization layers normalize the batch data output from linear layers. The function of batch normalization layers is to maintain the data distribution within a stable range, reduce model training fluctuations caused by differences in data distribution between layers, reduce the model's dependence on changes in the data distribution of the previous layer, thereby accelerating the convergence speed of the neural network model and avoiding gradient vanishing or exploding problems during training due to data distribution shifts, ensuring the stability and efficiency of model training.

[0050] The Dropout layer, through a pre-defined random discarding mechanism, randomly discards a portion of the neurons output by the batch normalization layer. The function of the Dropout layer is to reduce excessive dependence between neurons, prevent the model from overfitting the training data during training, and simulate different network structures by randomly discarding some neuron outputs. This enhances the model's generalization ability and robustness to TGV parameter samples, ensuring that the model maintains stable prediction accuracy in real-world prediction scenarios.

[0051] The activation layer employs the ReLU activation function to perform a nonlinear transformation on the output of the Dropout layer. The function of the activation layer is to introduce nonlinear mapping capabilities into the neural network, overcoming the limitations of linear transformations. This enables the neural network model to capture the complex nonlinear relationship between TGV structural parameters and S-parameters, improving the model's feature extraction depth and fitting accuracy for TGV electromagnetic properties. This ensures the model can accurately learn and characterize the comprehensive influence of the TGV structure on the electromagnetic response, meeting the accuracy requirements for rapid S-parameter prediction.

[0052] In one embodiment of the present invention, the neural network model further includes: The output layer is configured to output S11 and S21 (or full-band S-parameter characteristics) at the target frequency.

[0053] In one embodiment of the present invention, the connection method between layers is as follows: Each node in the input layer is connected to all nodes in the first hidden layer. Each connection corresponds to a weight parameter. The input vector is first multiplied by the weight matrix, then the bias vector is added, and finally the ReLU activation function is used to obtain the output vector of the first hidden layer. .

[0054] The output value of each node in the first hidden layer is used as the input to the linear layer of the second hidden layer, and intermediate calculation results are obtained through a fully connected operation. The results are first processed through a batch normalization layer to stabilize their distribution; then they enter a dropout layer to randomly discard some neuron outputs, enhancing the model's generalization ability. Finally, the ReLU activation function is used to obtain the output vector of the second hidden layer. .

[0055] The second and third hidden layers are connected using the same method as the connection between the first and second hidden layers. The output vector of the second hidden layer... Obtained through linear transformation The distribution is then adjusted by a batch normalization layer, and some node outputs are randomly discarded using Dropout. Finally, the output vector of the third hidden layer is obtained by applying the ReLU activation function. .

[0056] All nodes in the third hidden layer are fully connected to all nodes in the output layer. Each connection corresponds to a weight parameter. The output layer maps the output vector of the third hidden layer to the output dimension to obtain the predicted output. The output layer no longer sets an activation function, so that the prediction results remain in continuous value form, to adapt to regression tasks with S-parameter amplitude or phase.

[0057] In one embodiment of the present invention, the output of each layer in the neural network model is used as the input of the next layer. The output of each linear layer is first passed through a batch normalization layer to stabilize the distribution. The batch normalized result is then passed through a Dropout layer to enhance robustness. Finally, nonlinearity is introduced through the ReLU activation function, enabling the model to fit complex S-parameter change trends.

[0058] Although the neural network model described in one embodiment of the present invention is a multilayer perceptron (MLP) neural network model, it should be understood that it is presented merely as an example and not as a limitation. It will be apparent to those skilled in the art that other types of neural network models, such as convolutional neural network (CNN) models and / or recurrent neural network (RNN) models, can be constructed to achieve the same technical effects as the multilayer perceptron neural network model in the embodiments of the present invention, and such models also fall within the scope of protection of the present invention.

[0059] In one embodiment of the present invention, the training of the neural network model employs the Adam (Adaptive Moment Estimation) adaptive optimization algorithm, and its core parameters are set as follows: Learning rate: Fixed at 0.001. A learning rate of 0.001 can avoid the loss function oscillation caused by an excessively large learning rate (failing to converge to the optimal solution) or the training cycle being too long due to an excessively small learning rate (failing to converge sufficiently after multiple iterations). Momentum parameters: β1=0.9 (first-order momentum decay coefficient) and β2=0.999 (second-order momentum decay coefficient), which are the standard configuration of the Adam algorithm. The learning rate of each parameter can be dynamically adjusted by accumulating historical gradient information.

[0060] The specific training process includes the following steps: Parameter initialization: Before training begins, the weight parameters and bias parameters of each layer in the neural network model are initialized. The weights are initialized using the Xavier method to ensure a reasonable parameter distribution and avoid gradient vanishing or gradient exploding during training. The biases are initialized to zero.

[0061] Forward propagation computation takes the preprocessed TGV structure parameters as input and passes them sequentially to each hidden layer. After linear transformation and ReLU activation function calculation, feature information is output layer by layer, and finally the neural network model's predicted values ​​of S parameters are obtained.

[0062] The loss function is calculated by comparing the predicted values ​​with the actual S-parameters obtained from the simulation, and using the mean squared error function to calculate the error, which is used to measure the model's prediction accuracy. Its expression is: in: N is the number of samples; These are the actual S-parameter values ​​obtained from the simulation; These are the predicted S-parameter values ​​output by the neural network model.

[0063] Backpropagation and parameter update: The gradient of each layer parameter is calculated by backpropagation based on the loss function, and the weights and biases are updated using the Adam optimization algorithm to gradually reduce the loss function value, thereby improving the model prediction accuracy.

[0064] The training and convergence judgment process involves iterating through batches of data. The training process terminates when the preset number of training rounds is reached or the loss function drops to a set threshold, generating the final trained neural network model.

[0065] Model solidification and validation: After training, the neural network model is tested on the validation dataset to evaluate its prediction error. If the accuracy requirements are met, the current model parameters are saved as the final model for subsequent TGV structure S-parameter prediction. The training dataset is divided into training and validation sets in an 8:2 ratio, and the number of iterations is set to 1000.

[0066] In one embodiment of the present invention, with a sample size of 200, the prediction error of the neural network model can be controlled within 5%.

[0067] In one embodiment of the present invention, the structural parameters of the TGV are input, and the corresponding S-parameter results can be directly obtained through a trained neural network model. Compared with traditional HFSS simulation, this process only requires millisecond-level computation time, eliminating the need for hours of electromagnetic solution.

[0068] Furthermore, in one embodiment of the present invention, the method further includes the following steps: Step S701: Input the target performance requirement index into the neural network reverse optimization module; Step S702: Initialize a set of TGV structural parameters; Step S703: Input the TGV structural parameters into the neural network model to obtain the predicted S parameters; Step S704: Calculate the fitness function based on the deviation from the target performance index requirements; and Step S705: Use a genetic algorithm for iterative optimization, and output new TGV structural parameters through crossover and / or mutation to gradually approach the target. Step S706: After several iterations, output the optimal TGV structure parameters.

[0069] In one specific embodiment of the present invention, a reverse optimization design method for TGV structural parameters based on performance requirements derived from a neural network model is further implemented. In this embodiment, when the designer provides a target performance index, a genetic algorithm is used to search and optimize the TGV structural parameters to obtain the optimal combination of structural parameters that satisfies the target performance index.

[0070] In one specific embodiment of the present invention, when multiple performance indicators are input, a mechanism of indicator priority and weighting coefficients is introduced to quantify the importance of different performance indicators. The performance indicators include insertion loss S21, return loss S11, bandwidth, impedance matching, etc.

[0071] The weighting coefficients of each performance index are denoted as w1, w2, ..., wn, respectively, and satisfy: w1 + w2 + ... + wn = 1 The weight values ​​are set by the designer according to the actual application requirements. For example, in high-speed signal transmission scenarios, the weight of insertion loss S21 can be increased to meet the requirements of low signal attenuation and long-distance / high-speed transmission in this scenario. In scenarios with high impedance matching requirements, the weight of return loss S11 can be increased to meet the requirement of minimizing signal reflection in this scenario and ensure that the impedance matching degree meets the design standards.

[0072] In one specific embodiment of the present invention, a fitness function is constructed based on the deviation between the input target performance index and the predicted output of the neural network model, which is used to evaluate the quality of the currently predicted TGV structural parameters. The fitness function F can be expressed as: in: , The values ​​of return loss S11 and insertion loss S21 are the performance parameters predicted by the neural network. , To design target value; , For predicted and target values ​​of bandwidth and other performance indicators; Wi(i=1,2,3,...) represents the weight coefficient of the corresponding indicator.

[0073] The smaller the fitness function value, the higher the fitness, and the closer the current structural parameters are to the target performance requirements.

[0074] In a specific embodiment of the present invention, the reverse optimization process employs a genetic algorithm as the solution strategy, and its steps include: Initialize the population by randomly generating several sets of TGV structural parameters within a set parameter space as initial population individuals. Each individual includes structural parameters such as aperture, aperture spacing, dielectric thickness, dielectric constant, and material parameters.

[0075] Performance prediction involves inputting the structural parameters of each individual into a trained neural network model, which then outputs the corresponding S-parameter prediction results.

[0076] Fitness calculation: According to the fitness function in the above embodiments of the present invention, the fitness value of each individual is calculated to measure the degree of deviation between its performance and the target performance.

[0077] The selection process, based on fitness values, employs a roulette wheel or tournament selection mechanism, prioritizing individuals with high fitness for the next generation.

[0078] Crossover operation involves randomly selecting two individuals from the chosen individuals and exchanging some of their structural parameter genes to generate new offspring individuals. For example, in a specific embodiment of the present invention, aperture and spacing parameters are cross-combined.

[0079] Mutation operations involve making small perturbations to the parameters of some individuals with a preset probability to increase population diversity and avoid getting trapped in local optima. For example, in a specific embodiment of the present invention, the aperture parameter is randomly varied by ±1% to ±5%.

[0080] Iterative updates, repeatedly performing the performance prediction to mutation operation steps, until at least one of the following termination conditions is met: Reaching the maximum number of iterations; and / or The fitness function value is lower than a set threshold; and / or The fitness variation over multiple generations is less than the set precision.

[0081] The present invention proposes a fast prediction method for S-parameters of TGV models based on neural networks, which is applicable to fields such as high-speed interconnects, RF front-end modules, and 3D IC packaging, and is especially suitable for large-scale optimization design of TGV in glass substrates.

[0082] For example, in one application embodiment of the present invention, in a 5G base station chip, the TGV needs to meet the requirements of low loss and wide bandwidth. The present invention proposes a fast prediction method for S-parameters of the TGV model based on a neural network, which can quickly predict the S-parameters under different designs. Through back-calculation optimization, a TGV design that meets system specifications can be obtained, providing direct guidance for device fabrication.

[0083] Although various embodiments of the invention have been described above, it should be understood that they are presented by way of example only and not as limitations. It will be apparent to those skilled in the art that various combinations, modifications, and alterations can be made without departing from the spirit and scope of the invention. Therefore, the breadth and scope of the invention disclosed herein should not be limited by the exemplary embodiments disclosed above, but should be defined solely by the appended claims and their equivalents.

Claims

1. A fast prediction method for S-parameters of a TGV model based on a neural network, characterized in that, Includes the following steps: Step S100: Provide a parameterized TGV model and set the TGV structural parameters for multiple combination schemes; Step S200: For the TGV structural parameter combination scheme in step S100, obtain the corresponding S-parameter data through electromagnetic simulation tools; Step S400: Provide a simulation dataset, wherein each sample in the simulation dataset includes a set of TGV structural parameters and corresponding S-parameter data; Step S500: Provide a neural network model and train the neural network model using the simulation dataset; and Step S600: Input the TGV structure parameters into the neural network model trained in step S400 to obtain the corresponding S parameters.

2. The method according to claim 1, characterized in that, The structural parameters of the TGV include: aperture, aperture spacing, dielectric thickness, dielectric constant and / or filling material.

3. The method according to claim 2, characterized in that, The setting range of the TGV structural parameters includes: The pore size range was set to 5 μm to 100 μm; The aperture spacing range was set to 55μm to 200μm; The medium thickness range is set to 50 μm to 500 μm; The dielectric constant is set to a range of 2 to 8; and The types of filler materials include copper, tungsten, and / or carbon nanotubes.

4. The method according to claim 1, characterized in that, In step S100, the design method for the TGV structural parameter combination scheme includes: Step S101: The structural parameters of the TGV are sampled at equal intervals within the set range to obtain the first sampling point; Step S102: Add random perturbation to a portion of the first sampling points to obtain second sampling points; Step S103: Cross-combine the second sampling points to form an initial TGV parameter combination; Step S104: Based on actual process constraints, the initial TGV parameter combination is screened to obtain the TGV parameter combination used to construct the simulation dataset.

5. The method according to claim 1, characterized in that, The method further includes: Step S300: Preprocess and / or extract features from the S-parameter data in step S200.

6. The method according to claim 1, characterized in that, The neural network model is a multilayer perceptron (MLP) neural network model, including: The input layer has n neurons, where n is the number of types of TGV structural parameters; Hidden layers, including first to third hidden layers, with ReLU activation function; and The output layer is configured to output the S-parameters at the target frequency or the S-parameter characteristics across the entire frequency band.

7. The method according to claim 1, characterized in that, The method further includes: Step S701: Input the target performance requirement index into the neural network reverse optimization module; Step S702: Initialize a set of TGV structural parameters; Step S703: Input the TGV structural parameters into the neural network model to obtain the predicted S parameters; Step S704: Provide a fitness function and calculate the fitness function based on the deviation from the target performance index requirement; and Step S705: Use a genetic algorithm for iterative optimization, and output new TGV structural parameters through crossover and / or mutation to gradually approach the target. Step S706: After several iterations, output the optimal TGV structure parameters.

8. The method according to claim 7, characterized in that, In step S701, the target performance requirement indicators include: Insertion loss, return loss, bandwidth and / or impedance matching.

9. The method according to claim 7, characterized in that, In step S704, the fitness function is: in: , The predicted performance parameter values ​​output by the neural network model; , The target performance index value; , For predicted and target values ​​of bandwidth and other performance indicators; Wi(i=1,2,3,...) represents the weight coefficient of the corresponding indicator.

10. A fast prediction system for S-parameters of a TGV model based on a neural network, characterized in that, include: The simulation dataset module is configured to store TGV structural parameters of multiple combination schemes, as well as S-parameter data obtained by inputting the TGV structural parameters into an electromagnetic simulation tool, providing a dataset for neural network model training and verification. A neural network model module is configured to include a multilayer perceptron neural network model, the neural network model comprising: The input layer has n neurons, where n is the number of types of TGV structural parameters; Hidden layers, including first to third hidden layers, with ReLU activation function; and The output layer is configured to output S-parameters at the target frequency or S-parameter characteristics across the entire frequency band; and The neural network inverse optimization module is configured to back-calculate the optimal TGV structure parameters when given input performance requirements.

Images