CN-121980899-A - Surface acoustic wave filter end-to-end design method and system based on deep learning

CN121980899ACN 121980899 ACN121980899 ACN 121980899ACN-121980899-A

Abstract

The invention discloses an end-to-end design method and system of a surface acoustic wave filter based on deep learning, comprising the following steps of constructing a high-fidelity data set of a surface acoustic wave resonator, designing and training a nerve operator model as a forward proxy model to realize rapid and high-precision prediction from structural parameters to admittance response, integrating the nerve operator and an ABCD matrix cascade physical model to construct a differentiable calculation map, directly optimizing the structural parameters of the resonator by a gradient descent algorithm to match with the performance index of a target filter, realizing the design method based on the nerve operator and the differentiable physical model, and directly multiplexing the physical characteristics of the resonator packaged by the trained nerve operator model as a universal design asset to filter designs of different frequency bands and indexes, thereby overcoming the defect that the conventional method and the existing AI method cannot accumulate and inherit the first question and knowledge.

Inventors

ZHOU CHANGJIAN
WANG SHITING

Assignees

华南理工大学

Dates

Publication Date: 20260505
Application Date: 20251201

Claims (10)

1. The end-to-end design method of the surface acoustic wave filter based on deep learning is characterized by comprising the following steps of: Using the trained nerve operator model as a forward proxy model to acquire a nonlinear mapping relation from the structural parameters to admittance response; And establishing a differentiable filter model based on the nerve operator model, and optimizing resonator structure parameters through multi-objective gradients to match objective filter performance indexes.
2. The end-to-end design method of the deep learning-based surface acoustic wave filter of claim 1, wherein the neural operator model comprises an input projection layer, a plurality of residual blocks, and an output layer, each residual block comprising a linear transformation and GELU activation functions.
3. The end-to-end design method of the deep learning-based surface acoustic wave filter according to claim 1, wherein the loss function of the neural operator model adopts a weighted mean square error loss function for optimizing prediction accuracy of the resonance point region and the antiresonance point region.
4. The end-to-end design method of the surface acoustic wave filter based on deep learning of claim 1, wherein training data of the neural operator model is sampled within a structural parameter design range of the surface acoustic wave resonator by using latin hypercube sampling.
5. The end-to-end design method of a deep learning-based surface acoustic wave filter according to claim 1, wherein the establishing of the differentiable filter model comprises: Reconstructing admittance parameters in a complex form based on admittance results predicted by the nerve operators; According to the topology of the filter circuit, the admittance parameters of the series resonators and the parallel resonators are converted into an ABCD matrix form; obtaining ABCD parameters of the integral filter through matrix cascading calculation; the return loss S 11 and insertion loss S 21 of the filter are calculated based on the overall ABCD parameters and the characteristic impedance.
6. The deep learning-based surface acoustic wave filter end-to-end design method of claim 1, wherein the multi-objective gradient optimization comprises multi-objective optimization of passband characteristics, stopband characteristics, and resonance point constraints, and the corresponding multi-objective loss function is set as a weighted sum of passband constraint loss, stopband constraint loss, and resonance frequency constraint loss.
7. The end-to-end design method of the deep learning-based surface acoustic wave filter of claim 6, wherein the passband constraint loss is: wherein passband denotes the pass band, Loss function values of the S21 and S11 portions within the passband are respectively represented, S21 denotes insertion loss, denotes transmission efficiency from one port to the other port, and S11 denotes return loss, denotes energy of a signal reflected from one port back to the port; Representing the number of frequency points within the passband; And (3) with Respectively representing a threshold value which is required to be larger than S21 and a threshold value which is required to be smaller than S11 in the passband; And (3) with Respectively expressed at frequency points Predicted S21 and S11 parameters.
8. The end-to-end design method of the deep learning-based surface acoustic wave filter of claim 6, wherein the stopband constraint loss is: Wherein stopband denotes the stop band, Representing the loss function value of the S21 part in the stop band, wherein Nsb represents the number of frequency points in the stop band, and Tsb represents the threshold value which S21 in the stop band needs to be smaller than; Expressed at a frequency point Predicted S21 parameters.
9. The deep learning-based surface acoustic wave filter end-to-end design method of claim 6, wherein the resonance point constraint loss estimates the resonance frequency of each resonator by a soft argmax technique and ensures that the estimated resonance frequency falls within a set minimum and maximum limiting frequency range, the resonance point constraint loss being the sum of squares of deviations of the resonance frequency from the limiting frequency.
10. A system for implementing the deep learning-based surface acoustic wave filter end-to-end design method of claim 1, comprising: constructing a resonator model, wherein the resonator model is used for cascading acoustic surface wave filters made of multiple layers of materials; the data construction module is used for constructing a high-fidelity data set of the surface acoustic wave resonator through electromagnetic simulation; the nerve operator training module is used for designing and training a nerve operator model and learning the mapping from the structural parameters to admittance responses; The differentiable model construction module is used for integrating the trained nerve operator and the ABCD matrix cascade physical model to construct a differentiable calculation map; And the optimizing module is used for optimizing the resonator structure parameters through a multi-target gradient descent algorithm so as to match the target filter performance index.

Description

Surface acoustic wave filter end-to-end design method and system based on deep learning Technical Field The invention relates to the technical field of radio frequency filter design and artificial intelligence intersection, in particular to a surface acoustic wave filter end-to-end design method and system based on deep learning. Background Currently, the mainstream design method of the surface acoustic wave filter relies heavily on physical simulation technology, and although the finite element method （FEM）（Koigerov,A.S."SurfaceAcousticWaveDevicesonFrequencyHarmonics.FeaturesofCalculatingSAWParametersbytheFiniteElementMethod."OpticsandSpectroscopy132.1(2024):54-63.） and the layered cascading finite element method （KoskelaJ,PlesskyV,WillemsenB,etal.Hierarchicalcascadingalgorithmfor2-DFEMsimulationoffiniteSAWdevices[J].IEEETransactionsonUltrasonics,Ferroelectrics,andFrequencyControl,2018,65（10）:1933-1942.）. have higher simulation precision, single calculation of the method can take up to several hours, and it is difficult to support global optimization and reverse design flow requiring massive iteration, so that the overall design period is long and the efficiency is low. To improve design efficiency, artificial intelligence techniques are gradually introduced. However, existing approaches are often "fragmented" applications, and do not form a complete, reusable end-to-end solution. For example, some studies use machine learning to assist in extracting the coupled mold model parameters or predicting the temperature drift characteristics, while improving the local modeling accuracy, the acceleration effect on the overall design flow is limited. There have been other studies attempting to end-to-end map AI models as a direct substitute for simulators, but such methods tend to skip key intermediate physical variables such as admittance, resulting in a lack of physical rationality and generalization ability of the predicted results. In the aspect of reverse design, the existing intelligent design method based on deep learning has shown a certain potential, but has obvious limitations. For example, some schemes adopting convolutional neural networks or series neural network architecture can realize mapping from design parameters to performance indexes, but have the defects that firstly, a trained model aims at a specific design task, and lacks of knowledge multiplexing capability of cross frequency bands and cross indexes, so that 'one-problem training' resource is wasted, and secondly, the optimization process still depends on a gradient-free optimization method such as a genetic algorithm, and the like, so that the problem of slow convergence speed and easiness in sinking into local optimum exists. As described above, the prior art has not yet provided an intelligent SAW filter design system with high accuracy, high efficiency, end-to-end automation and knowledge reusability. Therefore, the invention provides a hybrid design framework based on nerve operators and differentiable physical modeling, and the technical defects are systematically solved by constructing a full-link differentiable computation graph of 'structure-admittance-system' response, realizing end-to-end gradient optimization from performance indexes to component parameters while maintaining physical consistency, and endowing the model with the capacity of cross-task multiplexing. Disclosure of Invention The invention aims to provide an end-to-end design method and system of a surface acoustic wave filter based on deep learning, which are used for constructing a complete differentiable calculation link from structural parameters to system performance by fusing a nerve operator and differentiable physical modeling technology, so as to realize rapid, accurate and automatic design of the surface acoustic wave filter. The method aims to overcome the limitation of the traditional design method, greatly improve the design efficiency, reduce the dependence on expert experience, and simultaneously ensure the physical rationality and the performance superiority of the design result. The core of the invention is to construct a complete technical framework of data driving-agent modeling-physical constraint-gradient optimization. Firstly, constructing a training data set through high-fidelity electromagnetic simulation, then constructing a neural operator model by utilizing a depth residual error network to learn electromagnetic response characteristics of the surface acoustic wave resonator, and establishing a high-precision forward proxy model. On the basis, a differential filter model is constructed by combining a filter cascade physical model, and finally, the automatic optimization of the structural parameters is realized through multi-objective gradient optimization. The technical scheme breaks through a plurality of bottlenecks of the traditional design method, namely firstly, the time-consuming electromagnetic simulation is replaced by a nerve opera