Surface acoustic wave filter end-to-end design method and system based on deep learning

Abstract

The invention discloses a surface acoustic wave filter end-to-end design method and system based on deep learning. The method comprises the following steps: constructing a high-fidelity data set of a surface acoustic wave resonator; a neural operator model is designed and trained to serve as a forward proxy model, and rapid and high-precision prediction from structure parameters to admittance response is achieved; integrating the neural operator and an ABCD matrix cascade physical model, constructing a differentiable calculation graph, and directly optimizing resonator structure parameters through a gradient descent algorithm to match target filter performance indexes; the design method is realized based on the neural operator and the differentiable physical model, the trained neural operator model packages the physical characteristics of the resonator, and the neural operator model can be used as a general design asset and is directly reused for the design of filters with different frequency bands and indexes; the defects that a traditional method and an existing AI method have one question and one proposal, and knowledge cannot be accumulated and inherited are overcome.

Description

Technical Field

[0001] This invention relates to the field of radio frequency filter design and artificial intelligence, specifically to an end-to-end design method and system for surface acoustic wave filters based on deep learning. Background Technology

[0002] Currently, the mainstream design methods for surface acoustic wave (SAW) filters heavily rely on physical simulation techniques, such as the finite element method (FEM) (Koigerov, AS "Surface Acoustic Wave Devices on Frequency Harmonics. Features of Calculating SAW Parameters by the Finite Element Method." Optics and Spectroscopy 132.1 (2024): 54-63.) and the hierarchical cascading finite element method (Koskela J, Plessky V, Willemsen B, et al. Hierarchical cascading algorithm for 2-DFEM simulation of finite SAW devices [J]. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 2018, 65(10): 1933-1942.). Although these methods have high simulation accuracy, their single calculation can take several hours, making it difficult to support the global optimization and reverse design processes that require massive iterations, resulting in a long overall design cycle and low efficiency.

[0003] To improve design efficiency, artificial intelligence (AI) technology has been gradually introduced. However, existing methods are mostly applied in a fragmented manner, failing to form complete and reusable end-to-end solutions. For example, some studies use machine learning to help extract coupled-mode model parameters or predict temperature drift characteristics, which improves the accuracy of local modeling, but has limited effect on accelerating the overall design process. Other studies have attempted to use AI models as direct replacements for simulators for end-to-end mapping, but such methods often skip key intermediate physical variables such as admittance, resulting in predictions that lack physical rationality and generalization ability.

[0004] In reverse engineering, existing deep learning-based intelligent design methods have shown some potential, but significant limitations remain. For example, some solutions using convolutional neural networks or cascaded neural network architectures, while achieving the mapping from design parameters to performance metrics, suffer from the following drawbacks: First, the trained models are mostly designed for specific tasks, lacking the ability to reuse knowledge across frequency bands and metrics, leading to a waste of resources through "one training session per problem"; second, their optimization process still relies on gradient-free optimization methods such as genetic algorithms, resulting in slow convergence and a tendency to get trapped in local optima.

[0005] In summary, existing technologies have not yet provided a smart design system for SAW filters that combines high precision, high efficiency, end-to-end automation, and knowledge reusability. To address this, this invention proposes a hybrid design framework based on neural operators and differentiable physical modeling. By constructing a fully differentiable computational graph of the "structure-admittance-system" response, it achieves end-to-end gradient optimization from performance indicators to component parameters while maintaining physical consistency, and endows the model with cross-task reusability, thereby systematically solving the aforementioned technical deficiencies. Summary of the Invention

[0006] The purpose of this invention is to provide an end-to-end design method and system for surface acoustic wave (SAW) filters based on deep learning. By integrating neural operators and differentiable physical modeling techniques, a complete differentiable computational chain from structural parameters to system performance is constructed, enabling rapid, accurate, and automated design of SAW filters. This method aims to overcome the limitations of traditional design methods, significantly improve design efficiency, reduce reliance on expert experience, and simultaneously ensure the physical rationality and superior performance of the design results.

[0007] The core of this invention lies in constructing a complete technical framework of "data-driven - surrogate modeling - physical constraints - gradient optimization". First, a training dataset is built through high-fidelity electromagnetic simulation. Then, a neural operator model is constructed using a deep residual network to learn the electromagnetic response characteristics of the surface acoustic wave resonator, establishing a high-precision forward surrogate model. Based on this, a differentiable filter model is constructed by combining a filter cascade physical model. Finally, automatic optimization of structural parameters is achieved through multi-objective gradient optimization.

[0008] The technical solution of this invention breaks through several bottlenecks of traditional design methods: first, it achieves millisecond-level performance prediction by replacing time-consuming electromagnetic simulation with neural operators; second, it realizes end-to-end gradient propagation from performance indicators to structural parameters through differentiable modeling; and third, it ensures the rationality and feasibility of the design results through the design of loss functions with physical constraints.

[0009] The present invention is achieved by at least one of the following technical solutions.

[0010] A deep learning-based end-to-end design method for surface acoustic wave (SAW) filters includes the following steps: The trained neural operator model is used as a forward surrogate model to obtain the nonlinear mapping relationship from structural parameters to admittance response; A differentiable filter model is established based on a neural operator model, and the resonator structural parameters are optimized through multi-objective gradient optimization to match the target filter performance index.

[0011] Furthermore, the neural operator model includes an input projection layer, multiple residual connection blocks, and an output layer, with each residual block containing a linear transformation and a GELU activation function.

[0012] Furthermore, the loss function of the neural operator model adopts a weighted mean square error loss function to optimize the prediction accuracy of the resonant point region and the anti-resonant point region.

[0013] Furthermore, the training data for the neural operator model is sampled using the Latin hypercube sampling method within the design range of the structural parameters of the surface acoustic wave resonator.

[0014] Furthermore, the establishment of the differentiable filter model includes: Based on the admittance results predicted by neural operators, the admittance parameters in complex form are reconstructed; Based on the filter circuit topology, the admittance parameters of the series and parallel resonators are converted into ABCD matrix form; The ABCD parameters of the overall filter are obtained by matrix cascading calculation; Calculate the return loss of the filter based on the overall ABCD parameters and characteristic impedance. S 11 and insertion loss S 21 .

[0015] Furthermore, the multi-objective gradient optimization includes multi-objective optimization of passband characteristics, stopband characteristics, and resonant frequency constraints. The corresponding multi-objective loss function is set as a weighted sum of passband constraint loss, stopband constraint loss, and resonant frequency constraint loss.

[0016] Furthermore, the passband constraint loss is:

[0017] Here, passband refers to the passband. They represent the passband area respectively. S 21 and S The loss function values ​​in part 11. S 21 refers to insertion loss, which represents the transmission efficiency from one port to another. S 11 refers to return loss, which represents the energy of a signal reflected back from one port to the same port. Indicates the number of frequency points within the passband; and They represent the passband area respectively. S 21 needs to be greater than the threshold and S 11 needs to be less than a threshold; and These represent the frequency points respectively. Predicted S 21 and S 11 parameters.

[0018] Furthermore, the stopband constraint loss is:

[0019] Where stopband represents the stopband. Indicates within the stopband S The loss function values ​​for part 21; N sb represents the number of frequency points within the stopband; T sb indicates within the stopband S 21 needs to be less than the threshold; Indicates at frequency point Predicted S 21 parameters; Furthermore, the resonant point constraint loss is calculated by estimating the resonant frequency of each resonator using the soft argmax technique, ensuring that the estimated resonant frequency falls within the set minimum and maximum limiting frequency range. The resonant point constraint loss is the sum of the squares of the deviations between the resonant frequency and the limiting frequency.

[0020] The system for implementing the deep learning-based end-to-end design method for surface acoustic wave filters includes: Resonator model construction for cascading surface acoustic wave filters in multilayer materials; The data construction module is used to build a high-fidelity dataset of surface acoustic wave resonators through electromagnetic simulation; The neural operator training module is used to design and train neural operator models to learn the mapping from structural parameters to admittance response; The differentiable model construction module is used to integrate trained neural operators with ABCD matrix cascaded physical models to construct a differentiable computation graph. The optimization module is used to optimize the resonator structure parameters using a multi-objective gradient descent algorithm to match the target filter performance metrics.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. It reduces the traditional design cycle of several weeks to several hours, with single performance predictions taking only milliseconds, representing a speedup of several orders of magnitude compared to traditional electromagnetic simulation. This overcomes the drawbacks of traditional finite element simulation methods, such as excessively high computational costs and long design cycles. 2. By embedding a differentiable physical model, we ensure that the design results strictly conform to the laws of electromagnetism, avoiding non-physical interpretations that may arise from purely data-driven methods.

[0022] 3. Gradient-based optimization methods can effectively explore a broad design space, avoid getting trapped in local optima, achieve global automatic optimization, and significantly reduce reliance on expert experience. This invention enables end-to-end reverse design from system metrics to component parameters. The design process is only limited by the iteration speed of gradient optimization, requiring no manual intervention. 4. The trained neural operator model encapsulates the physical characteristics of the surface acoustic wave resonator and can be used as a reusable design asset to support filter design for different frequency bands, performance requirements, and circuit structures. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of a 2.5D model of the surface acoustic wave resonator used in the embodiments of the present invention.

[0024] Figure 2 This is a schematic diagram of the neural operator model used in the embodiments of the present invention.

[0025] Figure 3 This is a schematic diagram of the end-to-end intelligent design process of the surface acoustic wave filter in an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram illustrating the end-to-end intelligent design optimization example of the surface acoustic wave filter in this embodiment of the invention. Detailed Implementation

[0027] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] A deep learning-based end-to-end design method for surface acoustic wave (SAW) filters includes the following steps: (1) Construct a dataset of admittance response of surface acoustic wave resonators based on heterojunctions: A quasi-3D surface acoustic wave resonator model is established through electromagnetic simulation. Compared with periodic performance, the mapping between structure and admittance curve is obtained more accurately. Script code is written to control the finite element simulation software to obtain the mapping data from structure to admittance curve in batches. The design space is defined and efficient sampling is performed. The design parameters of the surface acoustic wave resonator include key geometric parameters such as the width of the interdigital transducer fingers, the finger spacing, and the aperture length. This embodiment uses the Latin hypercube sampling method to perform efficient sample sampling within the design range of structural parameters. This method can fully explore the design space with a small number of samples, ensuring the representativeness and uniformity of the sample points. In specific implementation, the physical value range of each structural parameter is first determined. These ranges are set based on manufacturing capabilities and physical feasibility. Then, Latin hypercube sampling is used to generate hundreds of different combinations of structural parameters, each combination representing a specific resonator design scheme.

[0029] A high-precision dataset is constructed using the structural parameter combinations obtained from each sample. Accurate electromagnetic simulation calculations are performed using multiphysics simulation software. Simulation settings include: defining appropriate material parameters and boundary conditions, setting the frequency scan range, configuring mesh generation accuracy and solver parameters, and performing harmonic response analysis to obtain admittance characteristics. Batch code is written via scripts to automate a large number of simulation tasks, constructing a high-density dataset containing data from 300 different structures, with 1301 frequency points for each structure. The dataset includes admittance amplitude and phase information for different combinations of structural parameters in a given frequency band. The formula for calculating the admittance amplitude is:

[0030] In the formula, A dB This represents the admittance magnitude in dB. and These represent conductivity and susceptance, respectively.

[0031] The formula for calculating the phase is:

[0032] In the formula, Indicates phase.

[0033] Data preprocessing and feature engineering were performed on the original simulation data. Specifically, the frequency and structural parameters were normalized to the [0,1] interval using a min-max method to eliminate the influence of dimensions and accelerate model convergence. The calculation formula is as follows:

[0034] In the formula, Represents the first in the original data i One sample point, Indicates the first element in the normalized data. i One sample point, and These represent the maximum and minimum values ​​in the original dataset, respectively.

[0035] To address the resonance spikes present in the admittance data, robust normalization based on the median and interquartile range is performed on the admittance data, calculated as follows:

[0036] In the formula, A norm and A dB These are the admittance data before and after normalization, respectively. median ( A dB )yes A dB the median of IQR ( A dB )yes A dB The interquartile range. This normalization method is insensitive to outliers and is more suitable for admittance response curve data.

[0037] To convert the phase into sine and cosine components and avoid phase entanglement, the calculation formula is as follows:

[0038] In the formula, S norm and C norm These are the data after the phase has undergone sine and cosine transformations and then standardized. Indicates the original phase. and These represent the mean values ​​of the phase after sine and cosine transformations, respectively. and These represent the standard deviations of the phase after sine and cosine transformations, respectively.

[0039] (2) Training the neural operator surrogate model: Construct and train a stable residual multilayer perceptron network as a neural operator to learn the complex nonlinear mapping relationship from structural parameters to admittance response; A neural operator model for predicting admittance results is constructed using a residual multilayer perceptron network. Its core architecture includes an input projection layer, multiple residual layers, and an output layer. The input projection layer increases the input dimension from 4 to 512. This is followed by six identical residual layers, each containing a linear transformation and a GELU activation function. The residual connection mechanism effectively mitigates the vanishing gradient problem in deep networks, ensuring training stability. Finally, the output layer maps the 512-dimensional hidden features to a 3-dimensional output space, corresponding to the admittance magnitude, phase sine, and phase cosine, respectively.

[0040] The formula for inputting the projection layer is:

[0041] in, , representing the input of the projection layer, consists of normalized structural parameters (such as interdigitated electrode width, spacing, and aperture width) and normalized frequency; , The weight matrix and the bias vector are respectively; It is the 512-dimensional feature vector after projection. It represents the set of real numbers, and the upper right corner indicates the dimension of the real number matrix or vector.

[0042] The formula for calculating the residual block is:

[0043] Among them, W l , b l and h l The first l The weights, bias parameters, and output data of each hidden layer.

[0044] The formula for calculating the output layer is:

[0045] Among them, W out , b out and h L These are the weight matrix, bias vector, and th layer of the output layer, respectively. L The output data of each hidden layer. y represents the output vector, which contains three physical quantities: admittance magnitude, phase sine, and phase cosine.

[0046] The loss function and training strategy are designed, employing a weighted mean square error loss function. The focus is on optimizing the prediction accuracy in the resonance point and anti-resonance point regions, with a significantly higher weight given to the resonance peak region than to the flat region. The specific calculation formula is as follows:

[0047]

[0048] in, This represents the value of the loss function; N Indicates the number of frequency points; A pred,n , S pred,n and C pred,n These are the admittance magnitude, phase sine, and cosine predicted by the neural operator. A ture,n , S ture,n andC ture,n These are the actual admittance magnitude, phase sine, and cosine, respectively. This represents the standard deviation of the admittance dataset; w n This means that if the admittance magnitude at a given frequency point is greater than twice the standard deviation of the admittance dataset, then the admittance weight at that point is set to 50.0. As an example, the Adam optimizer is used during training, with an initial learning rate of 5e-4, decaying by a factor of 0.5 every 2000 epochs. An early stopping strategy is used to prevent overfitting; training is terminated early when the validation set loss no longer decreases for several consecutive epochs.

[0049] (3) Establish a differentiable filter model: Based on the trained neural operator model, construct a series-parallel resonator cascade network to form a differentiable mathematical relationship between structural parameters and filter performance indicators; The complex form of the admittance, including the admittance magnitude, is reconstructed based on the admittance results predicted by neural operators. The calculation formula is:

[0050] The formulas for calculating the real and imaginary parts are:

[0051] The original phase is recovered using the arctangent function, calculated as follows:

[0052] The surface acoustic wave filter is modeled using a cascaded ABCD matrix, where the ABCD matrix calculation expressions for the series resonator and the parallel resonator are as follows:

[0053] in The admittance is a complex number containing the real part. G and the virtual part B ; series and shunt represent series and parallel connections, respectively.

[0054] The overall ABCD parameters of the filter are calculated using matrix multiplication, as shown in the following formula:

[0055] in M This represents the number of resonators in the filter circuit.

[0056] After calculating the overall ABCD matrix of the cascaded filter, use this matrix to calculate the S-parameters of the filter. The formulas for calculating the S-parameters are as follows:

[0057] Characteristic impedance Z 0 = 50Ω S 11 and S 21 The parameters represent return loss and insertion loss, respectively.

[0058] (4) Optimize structural parameters: Define a multi-objective loss function that includes passband characteristics, stopband characteristics and resonant point constraints. The resonator structural parameters are automatically optimized by gradient analysis algorithm, and the optimal filter design that meets the target performance index is output.

[0059] This embodiment transforms the surface acoustic wave filter optimization design problem into a multi-objective gradient optimization problem:

[0060] Where p is the structural parameter vector, and the optimization objective is to minimize the multi-objective loss function. The total loss function takes the structural parameter vector p as input and is expressed as a weighted sum of the passband constraint loss, stopband constraint loss, and resonant frequency constraint loss. Its specific expression is as follows: Passband constraint: Ensures that within the passband... S 21 is greater than the set threshold and S 11 is less than the set threshold; Passband constraint loss ensures transmission performance within the passband, and its calculation expression is:

[0061] Here, passband refers to the passband. They represent the passband area respectively. S 21 and S The loss function values ​​for part 11; Indicates the number of frequency points within the passband; and They represent the passband area respectively. S 21 needs to be greater than the threshold and S 11 needs to be less than a threshold; and These represent the frequency points respectively. Predicted S 21 and S 11 parameters; the max(.) function is implemented using the ReLU function.

[0062] Stopband constraint: Ensures that within the stopband... S 21 is less than the set threshold; Stopband constraint loss ensures stopband suppression performance; the calculation expression is as follows:

[0063] Wherein, stopband represents the passband. Indicates within the stopband S The loss function values ​​for part 21; N sb represents the number of frequency points within the stopband; T sb indicates within the stopband S 21 needs to be less than the threshold; Indicates at frequency point Predicted S 21 parameters; Resonant frequency constraint loss: To ensure that the resonant frequencies of each resonator fall within a specified range, the resonant frequency constraint loss is calculated using the soft argmax technique to achieve a differentiable resonant frequency estimate. The calculation expression is as follows:

[0064] in, It is an adjustable coefficient used to control the sharpness of the weight distribution. , These represent the frequency points respectively. k and frequency point j The predicted admittance value, f res This represents the resonant frequency estimated using the soft argmax technique.

[0065] The expression for calculating the resonant frequency constraint loss is as follows:

[0066] in, These are the estimated resonant frequency and the set minimum and maximum resonant frequency limits, respectively. N r This represents the number of resonators. V This represents the frequency out-of-range penalty for a single resonator; This represents the total resonant frequency constraint loss value, which is the average of the squares of the normalized penalty amounts for all resonators; Indicates the first in the circuit r Out-of-bounds penalty for each resonator; Indicates the first r The design bandwidth of each resonator is used to normalize the out-of-bounds penalty, so that the dimensions of the loss function are consistent. N r This indicates the total number of resonators in the filter circuit.

[0067] The expression for calculating the total optimization loss is:

[0068] in, These are the adjustable weights for their respective parts. and These represent the loss constraints for the passband and stopband, respectively.

[0069] Gradient optimization is performed based on the constructed multi-objective loss function. The gradient of the loss function with respect to the structural parameters is calculated using the automatic differentiation technique in the PyTorch AI framework. The calculation expression is:

[0070] In the formula, Indicates the admittance magnitude. This represents the S-parameters of the matrix-computed filter. The gradient is automatically calculated through backpropagation of the neural operator and the physical model. Performance metrics are monitored in real time during optimization, and optimization terminates when all constraints are met or the maximum number of iterations is reached.

[0071] As a specific embodiment, this embodiment provides an end-to-end design method for surface acoustic wave (SAW) filters based on deep learning. By integrating neural operators and differentiable physical modeling techniques, a complete differentiable computational chain from structural parameters to system performance is constructed, enabling fast, accurate, and automated design of SAW filters. The specific steps are as follows: First, establish a simulation system in the multiphysics simulation software, such as... Figure 1 The 2.5D model of the multilayer surface acoustic wave resonator shown is composed of aluminum electrodes, lithium tantalate, silicon dioxide, polycrystalline silicon, and silicon material from top to bottom. It includes key components such as interdigital transducers (IDTs), reflective gratings, and piezoelectric substrates. The core design parameters of the resonator include the number of interdigital electrode pairs, the interdigital electrode width, and the aperture width. Based on CMOS process capabilities, the values ​​for these parameters are set to range from 80 to 125 pairs, 295 to 440 nm, and 50 to 200 μm, respectively.

[0072] The Latin hypercube sampling (LHS) method was used to generate 300 combinations of structural parameters within the design space. The LHS method ensures that all value ranges of each design variable are uniformly covered, offering better space filling compared to random sampling. In practice, the value range of each design parameter variable was divided into 300 equal intervals. A value was randomly selected from each interval, and the values ​​were randomly arranged to ensure independence. Finally, 300 different structural parameter samples were formed. Compared to full factorial design, which requires tens of thousands of samples to fully explore the 3D space, this method significantly improves efficiency.

[0073] The acoustic-electric coupling simulation was performed using finite element simulation software, with the following specific settings: [Settings to be established] Figure 1The 2.5D model, from top to bottom, consists of electrodes, piezoelectric material, SiO2, polycrystalline silicon, and silicon. In the piezoelectric effect module, the elastic matrix, piezoelectric matrix, and dielectric matrix of the 42°Y-XLiTaO3 piezoelectric substrate (the crystal tangentially rotated 42 degrees around the X-axis, with the surface acoustic wave propagation direction along the X-axis on the crystal surface) are defined. The electrostatic module is used to simulate the electric field distribution between the electrodes. The frequency scan range is set to 2000~3300MHz, and a set of admittance curves for the resonators are obtained and exported by solving the frequency domain. Due to the large number of 300 training sets, corresponding script files are written to set the structural parameters that need to be modified, as well as the solution module, enabling automated batch simulation and data collection.

[0074] The collected data underwent inverse preprocessing and data feature engineering. The frequency and structural parameters were normalized to the [0,1] interval using min-max normalization to eliminate dimensional differences and promote neural operator convergence. For the resonant peaks in the admittance curve, a robust normalization based on statistics was adopted. This normalization is insensitive to outliers and preserves the relative intensity of the resonant characteristics. The phase radians were converted into sine and cosine components and then standardized to effectively avoid phase entanglement.

[0075] Training samples are constructed using 3D normalized structural parameters plus 1D normalized frequency as input to the neural operator, and the normalized admittance amplitude and sine and cosine phase components at each frequency point as output. The samples are randomly divided into training and test sets in an 8:2 ratio. The training set is used to enable the model to learn the relationship between the implicit structural parameters and admittance data in the data, while the test set is used to verify the model's learning and evaluate its performance.

[0076] Constructing neural operator model architectures, such as Figure 2 As shown, the neural operator model consists of a multilayer perceptron with residual connections, including one input projection layer, three residual connection blocks, and one output layer. The input projection layer boosts the dimension of the input structural parameters (P1~P6) to 512 dimensions, and the output layer restores the data from 512 dimensions to 3 dimensions for output. Each residual block contains a 512-dimensional hidden layer. Residual connections are achieved through element-wise addition to avoid gradient vanishing, ensure the stability of the training process, and improve the model's learning ability. The GELU (Gaussian difference linear unit) activation function is used between each layer of the model.

[0077] During training, a weighted MSE loss function is used, assigning dynamic weights to the admittance prediction error. The weight is set to 50.0 when the admittance magnitude is greater than twice the standard deviation of the dataset, and 1.0 otherwise. This design focuses on optimizing the prediction accuracy in the resonance region. The Adam optimizer is used, with an initial learning rate of 5 × 10⁻⁶. -4 The weight decays to 1×10 -4The batch size is 1024, and the training epochs are 10,000. A step decay strategy is used for learning rate scheduling, halving the learning rate every 2000 epochs. An early stopping strategy is also implemented, terminating training when the validation set loss no longer decreases after 500 consecutive epochs. Model building and training are both based on the PyTorch AI framework.

[0078] Figure 3 The flowchart illustrates the optimization process, establishing a differentiable filter model based on a trained neural operator model. First, the complex form of the admittance parameters is reconstructed, converting the dB values ​​to linear amplitudes. The phase radians are calculated from the sin / cos components, and then the real and imaginary parts of the admittance are calculated. Next, a surface acoustic wave (SAW) filter is constructed, determining the filter circuit structure. Series and parallel resonators are represented using ABCD matrices. The overall ABCD parameters are calculated through matrix multiplication, and based on these parameters, the S-parameters of the two-port network, including the reflection coefficient, are calculated. S 11 and transmission coefficient S 21 The characteristic impedance is set to 50Ω.

[0079] Finally, multi-objective gradient optimization is performed. The filter design problem is transformed into a constrained optimization problem, defining a multi-objective loss function that includes passband constraints, stopband constraints, and resonant point constraints. The passband constraints are for the 2600-2700MHz frequency range, requiring... S 21 Greater than -2dB S 11 Less than -10dB. Stopband constraints apply to the frequency range below 2580MHz and above 2720MHz, requiring... S 21 Less than -25dB. The resonant point constraint employs soft argmax technology to achieve differentiable resonant frequency estimation. A resonant frequency range of 2650MHz-2750MHz is set for the series resonator, and 2550MHz-2650MHz for the parallel resonator. The total optimization loss function is a weighted sum of the losses from each constraint. The weighting coefficients are carefully adjusted to ensure a balance among various performance indicators. After optimization, the following can be obtained: Figure 4 The structural parameters and admittance response curves of each resonator are shown in (a) and (b), and the performance evaluation curves of S11 and S21 are obtained.

[0080] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A deep learning-based end-to-end design method for surface acoustic wave (SAW) filters, characterized in that, Includes the following steps: The trained neural operator model is used as a forward surrogate model to obtain the nonlinear mapping relationship from structural parameters to admittance response; A differentiable filter model is established based on a neural operator model, and the resonator structural parameters are optimized through multi-objective gradient optimization to match the target filter performance index.

2. The end-to-end design method for surface acoustic wave filters based on deep learning according to claim 1, characterized in that, The neural operator model consists of an input projection layer, multiple residual connection blocks, and an output layer. Each residual block contains a linear transformation and a GELU activation function.

3. The end-to-end design method for surface acoustic wave filters based on deep learning according to claim 1, characterized in that, The loss function of the neural operator model adopts the weighted mean square error loss function to optimize the prediction accuracy of the resonant point region and the anti-resonant point region.

4. The end-to-end design method for surface acoustic wave filters based on deep learning according to claim 1, characterized in that, The training data for the neural operator model were sampled using the Latin hypercube sampling method within the design range of the structural parameters of the surface acoustic wave resonator.

5. The end-to-end design method for surface acoustic wave filters based on deep learning according to claim 1, characterized in that, The establishment of a differentiable filter model includes: Based on the admittance results predicted by neural operators, the admittance parameters in complex form are reconstructed; Based on the filter circuit topology, the admittance parameters of the series and parallel resonators are converted into ABCD matrix form; The ABCD parameters of the overall filter are obtained by matrix cascading calculation; Calculate the return loss of the filter based on the overall ABCD parameters and characteristic impedance. S 11 and insertion loss S 21 .

6. The end-to-end design method for surface acoustic wave filters based on deep learning according to claim 1, characterized in that, Multi-objective gradient optimization includes multi-objective optimization of passband characteristics, stopband characteristics, and resonant frequency constraints. The corresponding multi-objective loss function is set as a weighted sum of passband constraint loss, stopband constraint loss, and resonant frequency constraint loss.

7. The end-to-end design method for surface acoustic wave filters based on deep learning according to claim 6, characterized in that, The passband constraint loss is: Here, passband refers to the passband. They represent the passband and the area within the passband, respectively. S 21 and S The loss function values ​​in part 11. S 21 refers to insertion loss, which represents the transmission efficiency from one port to another. S 11 refers to return loss, which represents the energy of a signal reflected back from one port to the same port. Indicates the number of frequency points within the passband; and They represent the passband and the area within the passband, respectively. S 21 needs to be greater than the threshold and S 11 needs to be less than the threshold; and These represent the frequency points respectively. Predicted S 21 and S 11 parameters.

8. The end-to-end design method for surface acoustic wave filters based on deep learning according to claim 6, characterized in that, The stopband constraint loss is: Where stopband represents the stopband. Indicates the stopband S The loss function values ​​for part 21; N sb represents the number of frequency points within the stopband; T sb indicates within the stopband S 21 needs to be less than the threshold; Indicates at frequency point Predicted S 21 parameters.

9. The end-to-end design method for surface acoustic wave filters based on deep learning according to claim 6, characterized in that, The resonant point constraint loss is calculated by estimating the resonant frequency of each resonator using the soft argmax technique, ensuring that the estimated resonant frequency falls within the set minimum and maximum limit frequency range. The resonant point constraint loss is the sum of the squares of the deviations between the resonant frequency and the limit frequency.

10. A system for implementing the deep learning-based end-to-end design method for surface acoustic wave filters as described in claim 1, characterized in that, include: Resonator model construction for cascading surface acoustic wave filters in multilayer materials; The data construction module is used to build a high-fidelity dataset of surface acoustic wave resonators through electromagnetic simulation; The neural operator training module is used to design and train neural operator models to learn the mapping from structural parameters to admittance response; The differentiable model construction module is used to integrate trained neural operators with ABCD matrix cascaded physical models to construct a differentiable computation graph. The optimization module is used to optimize the resonator structure parameters using a multi-objective gradient descent algorithm to match the target filter performance metrics.

Images