Bulk acoustic wave filter automatic design method and device based on machine learning

Abstract

The invention relates to an automatic design method and device for a bulk acoustic wave filter based on machine learning, and the method achieves the design of the bulk acoustic wave filter through the combination of a plurality of resonators, and comprises the following steps: S1, constructing a machine learning model, training the machine learning model through the bulk acoustic wave filter S parameter data set and the bulk acoustic wave filter zero point data set; s2, inputting an S parameter of the bulk acoustic wave filter meeting the design performance requirement to the machine learning model to obtain a zero point of the bulk acoustic wave filter meeting the design performance requirement; s3, calculating a resonance point and static capacitance of each resonator, wherein the resonance point comprises a series resonance frequency and a parallel resonance frequency; and S4, reversely deducing structural parameters of the bulk acoustic wave filter according with design performance requirements through the Mason model of the bulk acoustic wave filter. The method has the beneficial effects that the resonator parameters are directly obtained through the bulk acoustic wave filter performance, the time cost is low, and the BAW filter characteristics can be accurately described.

Description

[Technical field]

[0001] The present invention relates to the field of electronic technology, and in particular to a method and device for automatic design of a bulk acoustic wave filter based on machine learning. [Background technology]

[0002] With the development of communication technology, miniaturized filters with excellent performance are widely favored by communication equipment manufacturers. A very important type of filters is the BAW (Bulk Acoustic Wave) filter based on the microelectromechanical system (MEMS) process. It is a small and medium power product and has broad application prospects in the future fifth-generation (5G) mobile communication equipment, making the design method of BAW filter a research hotspot.

[0003] The basic principle of BAW filter is to use the piezoelectric effect of piezoelectric materials to convert electrical signals into acoustic signals and propagate them in solid media. The working principle is based on the propagation and resonance of bulk acoustic waves. It uses the piezoelectric effect of bulk acoustic wave resonators (such as FBAR or SMR) to convert RF electrical energy into mechanical energy, and realizes the filtering function through the propagation of acoustic waves in the device cavity. The acoustic wave signal is affected by the characteristics of the medium during the propagation process, thereby realizing the filtering function. At the output end, the acoustic wave signal is converted into an electrical signal again to complete the filtering process.

[0004] The simulation of a filter system composed of multiple resonators is the first problem that needs to be solved when designing a filter. In the Chinese invention patent "BAW filter design method, device and equipment based on BVD model" (Announcement No.: CN 107562990 B), it is proposed how to perform joint simulation of multiple resonators.

[0005] However, the invention method still has the following three problems. First, it is necessary to rely on the experience of the designer to provide the initial parameters of the resonator that are close to meeting the design goals. Second, the method requires multiple iterations of optimization algorithms. The optimization algorithm greatly increases the time cost. Third, the BVD (Butterworth-Van Dyke) model used in this method cannot accurately describe the characteristics of the BAW filter, especially in terms of loss.

[0006] Mason model: Mason derived the relationship between mechanical quantities and electrical quantities through the piezoelectric equation, wave equation, and mechanical boundary conditions, and based on this, drew an electromechanical equivalent circuit (Mason equivalent circuit) in the form of a six-port network.

[0007] As a two-port component, the S parameter of the filter is a key indicator to describe its performance in microwave transmission and RF measurement. The S parameter describes the transmission and reflection effect of the RF microwave network on the signal from the perspective of signal size or signal energy. For a two-port network (such as a filter), the S parameter matrix consists of four S parameters (S11, S12, S21 and S22), which define the relationship between the two ports in the RF system. The S11 parameter, also known as the input reflection coefficient or input return loss, describes the degree to which the signal is reflected back at the input port of the filter. The S21 parameter, which stands for the forward transmission coefficient, is what we usually call gain. It describes the ability of the signal to be transmitted from the input port of the filter to the output port.

[0008] The Chebyshev function is usually represented as Ψ(x), which is a function closely related to the prime number function π(x). The important property of this function is that it establishes a simple connection with the logarithmic function, thus laying the foundation for proving the prime number theorem and studying the distribution of prime numbers.

[0009] The present invention aims to solve the technical problems that the prior art needs to rely on experience to give the initial parameters of the resonator, has high iteration time and cost, and cannot accurately describe the characteristics of the BAW filter, and makes technical improvements to the automated design method and device of the BAW filter. [Summary of the invention]

[0010] The object of the present invention is to provide an automated design method for BAW filters, which can directly obtain resonator parameters from BAW filter performance, has low time cost, and can accurately describe BAW filter characteristics.

[0011] To achieve the above object, the technical solution adopted by the present invention is a method for automatically designing a bulk acoustic wave filter based on machine learning, which uses a plurality of resonators to combine and simulate the bulk acoustic wave filter to realize the bulk acoustic wave filter design, comprising the following steps:

[0012] S1. Construct a machine learning model, wherein the machine learning model is used to obtain a bulk acoustic wave filter system function of an input bulk acoustic wave filter S parameter and an output bulk acoustic wave filter zero point, and train the machine learning model through a bulk acoustic wave filter S parameter data set and a bulk acoustic wave filter zero point data set;

[0013] S2. Inputting the S parameters of the bulk acoustic wave filter that meet the design performance requirements into the machine learning model to obtain the system function of the bulk acoustic wave filter that meets the design performance requirements and the zero point of the bulk acoustic wave filter that meets the design performance requirements;

[0014] S3, comprehensively meeting the design performance requirements of the BAW filter zero point, obtaining a plurality of resonator combinations of the BAW filter system function meeting the design performance requirements, and calculating the resonance point and static capacitance of each resonator, wherein the resonance point includes a series resonance frequency and a parallel resonance frequency;

[0015] S4. Based on the resonance point and static capacitance of each resonator in the combination of the multiple resonators, the bulk acoustic wave filter Mason model is used to reversely deduce the bulk acoustic wave filter structural parameters that meet the design performance requirements, that is, the thickness and area of ​​each layer of each resonator.

[0016] Preferably, the machine learning model is a machine learning model of a back-propagation neural network architecture.

[0017] Preferably, step S1 specifically includes the following sub-steps:

[0018] S11. Generate a random array Ω between (1,5) and (-5,-1) n , as a BAW filter zero array at a normalized frequency Ω, the random array Ω n The size is determined by the number of resonators N;

[0019] S12, the BAW filter zero array Ω n Substitute the Chebyshev function to calculate the S parameters of the BAW filter, sample the BAW filter S21 parameter sequence and return loss RL sequence, and the BAW filter zero array Ω n Compose a machine learning data set and output it to the machine learning model for machine learning.

[0020] Preferably, the Chebyshev function is as follows:

[0021]

[0022] Preferably, step S2: inputting a bulk acoustic wave filter S21 parameter sequence and a return loss RL sequence that meet the design performance requirements into the machine learning model.

[0023] Preferably, step S3 specifically includes the following steps:

[0024] S31, obtain the ABCD matrix of each resonator in the combination of multiple resonators, the ABCD matrix of the series resonator ABCD Matrix of Parallel Resonators

[0025]

[0026] S32, multiplying the ABCD matrix of each resonator of the plurality of resonators to obtain the ABCD matrix of the bulk acoustic wave filter, thereby obtaining the S21 parameters of the bulk acoustic wave filter that meets the design performance requirements

[0027]

[0028] S33, the pole p of the bulk acoustic wave filter that meets the design performance requirements is obtained from the S21 parameter j,target , let BW be the filter bandwidth, p j The cost function of the synthesis algorithm is constructed for the BAW filter poles

[0029]

[0030] S34, combining the resonance point and static capacitance of each resonator of the plurality of resonators as data for the input synthesis algorithm, and using the pattern search algorithm to obtain the resonant series resonance frequency f when the cost function is minimized s , parallel resonant frequency f p And static capacitance C0.

[0031] Preferably, the bulk acoustic wave filter is a FBAR filter with a bandwidth of 2.4 GHz-2.5 GHz.

[0032] Preferably, step S4: in order to ensure process feasibility, the thickness of the bottom electrode of the resonator and the thickness and area of ​​the piezoelectric layer are set to constant values, and only the thickness and area of ​​the upper electrode are modified; the resonance point and static capacitance of each resonator in a plurality of resonator combinations are known, and the thickness and area of ​​the upper electrode of each resonator in the plurality of resonator combinations are solved by the Mason model.

[0033] Another object of the present invention is to provide an automated BAW filter design device that can directly obtain resonator parameters from BAW filter performance, has low time cost, and can accurately describe BAW filter characteristics.

[0034] To achieve the above-mentioned yet another objective, the technical solution adopted by the present invention is a bulk acoustic wave filter automatic design device based on machine learning, which is used to execute the above-mentioned bulk acoustic wave filter automatic design method based on machine learning.

[0035] The present invention provides an automatic design method and device for a bulk acoustic wave filter based on machine learning, and has the following beneficial effects: 1. The parameters of the resonator are directly obtained from the performance of the bulk acoustic wave filter, avoiding the need to rely on the experience of designers for design; 2. Avoiding the use of time-consuming optimization algorithms, reducing time costs; 3. Compared with the bulk acoustic wave filter constructed by the traditional gradient algorithm, the bulk acoustic wave filter constructed by the present invention has better out-of-band suppression, better return loss and equivalent passband performance; 4. The resonator structure is obtained using the Mason model, and the Q value of the resonator is taken into consideration during filter simulation, thereby improving the accuracy of the simulation.

Brief Description of the Drawings

[0036] Figure 1 This is a diagram of the Mason model of a bulk acoustic wave filter.

[0037] Figure 2 It is a flow chart of the automated design method of bulk acoustic wave filters based on machine learning.

[0038] Figure 3 It is a diagram of the machine learning model architecture.

[0039] Figure 4 The figure is a schematic diagram comparing the S11 parameters of a bulk acoustic wave filter designed by an automated bulk acoustic wave filter design method based on machine learning and a bulk acoustic wave filter designed by gradient optimization.

[0040] Figure 5 The figure is a schematic diagram comparing the S21 parameters of a bulk acoustic wave filter designed by an automated bulk acoustic wave filter design method based on machine learning and a bulk acoustic wave filter designed by gradient optimization. [Specific implementation method]

[0041] The features and exemplary embodiments of various aspects of the present invention will be described in detail below. In the detailed description below, many specific details are proposed to provide a comprehensive understanding of the present invention. But it is obvious to those skilled in the art that the present invention can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present invention by illustrating examples of the present invention. The present invention is by no means limited to any specific configuration and algorithm proposed below, but covers any modification, replacement and improvement of elements, parts and algorithms without departing from the concept of the present invention. In the accompanying drawings and the following description, known structures and technologies are not shown to avoid unnecessary ambiguity to the present invention.

[0042] Example 1

[0043] This embodiment implements an automated design method for bulk acoustic wave filters based on machine learning.

[0044] The technical problems to be solved by this embodiment are:

[0045] 1. The resonator parameters are directly obtained from the filter performance, avoiding the need to rely on the designer's experience for design.

[0046] 2. Avoid using time-consuming optimization algorithms and reduce time costs.

[0047] 3. Compared with the filter constructed by the traditional gradient algorithm, the filter constructed in this embodiment has better out-of-band suppression, better return loss and comparable passband performance.

[0048] 4. The Mason model is used to obtain the resonator structure, and the Q value of the resonator is taken into consideration during the filter simulation, which improves the accuracy of the simulation.

[0049] Figure 1 is the Mason model diagram of the bulk acoustic wave filter. Figure 1 As shown, the BAW filter is composed of a certain number of resonators. To simulate the resonator as a component, an equivalent circuit model is required. The Mason model is an equivalent circuit model. By inputting the structural parameters of the resonator (the thickness of each layer, the area of ​​the resonator), the impedance characteristics of the resonator can be obtained. The method of this embodiment is based on machine learning to obtain the impedance characteristics (resonant frequency and static capacitance) of each resonator, and then the BAW filter structure is inferred through the Mason model of the BAW filter.

[0050] The filter S parameters and filter zero points have a complex mathematical relationship (system function), which can be fitted through machine learning. The filter S parameters can be calculated by the Chebyshev function, and the zero point is the input variable of this function (Chebyshev function), so predicting the zero point by machine learning is equivalent to predicting the system function of the filter.

[0051] This embodiment selects BP (Back Propagation) neural network as the machine learning model architecture; by learning the S parameters and zero point distribution of the known corresponding relationship in the data set, changing the node weights and biases, so that the loss function Loss reaches a value below the required value, which means that the model has the ability to predict the zero point of the filter.

[0052] Figure 2 This is a flow chart of the automated design method of bulk acoustic wave filters based on machine learning. Figure 2 As shown, the method of this embodiment includes the following steps:

[0053] 1. Generate a random array Ω between (1,5) and (-5,-1) n , the array size is determined by the number of resonators N. This number (random array Ω n ) as the zero point of the filter at the normalized frequency Ω.

[0054] 2. Set the zero point array Ω n Substitute the Chebyshev function to calculate the S parameters of the filter. The Chebyshev function is as follows:

[0055]

[0056] 3. In order to reduce the size of the data set and prevent overfitting of machine learning, the S parameters are sampled. Take the S parameter within the normalized frequency step with a step size of 0.5.21 Maximum value of the parameter.

[0057] 4. Figure 3 This is a diagram of the machine learning model architecture. Figure 3 As shown, the sampling value in step 3 represents the out-of-band suppression and the S in the passband 11 The maximum value represents the return loss, and the corresponding zero array Ω n The data sets constitute the machine learning data sets and are output to the machine learning model for machine learning.

[0058] 5. After completing machine learning, write S according to the design goals 21 Parameters, get the predicted zero-point array.

[0059] 6. The ABCD matrix of each resonator can be obtained through the results of step 5. The ABCD matrix of the series resonator and the parallel resonator is as follows:

[0060]

[0061] 7. Multiply the ABCD matrix of each resonator to obtain the ABCD matrix of the filter, thereby obtaining the S parameters of the target filter:

[0062]

[0063] 8. Obtain the pole p of the target filter from the S parameter j,target . Let BW be the filter bandwidth, p j For the poles of the filter, the cost function of the synthesis algorithm is constructed as follows:

[0064]

[0065] 9. Use the resonant frequency and static capacitance of each resonator as the data for the input synthesis algorithm. Use the pattern search algorithm to obtain the resonant frequency f of the resonator with the minimum cost function. s 、f p And static capacitance C0.

[0066] 10. Use the performance information of the resonator to calculate the structural parameters of the BAW filter (the thickness of each layer and the area of ​​the resonator) through the Mason model.

[0067] Example 2

[0068] This embodiment implements an automated design method for bulk acoustic wave filters based on machine learning.

[0069] This example will use a FBAR filter with a bandwidth of 2.4 GHz-2.5 GHz to specifically illustrate the specific implementation process of the method of Example 1.

[0070] 1. Build a data set according to steps 1-4 of the solution of Example 1 to complete the training of the machine learning model.

[0071] 2. As in Table 1, fill in the S21 sequence and return loss RL that meet the filter performance requirements.

[0072] Table 1 Input data table of a bulk acoustic wave filter automated design method based on machine learning.

[0073]

[0074] 3. Machine learning will give the zero position of the filter, thus obtaining the mathematical function of the filter.

[0075] 4. The synthesis will obtain a resonator combination that conforms to the mathematical expression of the filter, that is, the resonance point of each resonator and its static capacitance.

[0076] 5. In order to ensure process feasibility, the thickness of the bottom electrode and the piezoelectric layer are set to fixed values, and only the thickness of the top electrode is modified.

[0077] 6. Given the resonance point and static capacitance of the resonator, the thickness of the upper electrode of the resonator can be solved using the Mason model.

[0078] 7. Figure 4 The figure is a schematic diagram comparing the S11 parameters of a bulk acoustic wave filter designed by an automated bulk acoustic wave filter design method based on machine learning and a bulk acoustic wave filter designed by gradient optimization. Figure 5 This is a schematic diagram comparing the S21 parameters of a BAW filter designed by an automated BAW filter design method based on machine learning and a BAW filter designed by gradient optimization. Figure 4 , Figure 5 As shown, the results obtained by the algorithm are verified with the results obtained by the traditional gradient optimization algorithm; the filter constructed by the algorithm of this embodiment has better out-of-band suppression, better return loss and comparable passband performance.

[0079] Those skilled in the art will appreciate that all or part of the steps to implement the above embodiments may be accomplished by hardware, or may be accomplished by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, wherein the storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0080] The above is only a preferred embodiment of the present invention. It should be pointed out that ordinary technicians in this technical field can make several improvements and supplements without departing from the principle of the present invention. These improvements and supplements should also be regarded as the scope of protection of the present invention.

Claims

1. A method for automated design of bulk acoustic wave filters based on machine learning, which uses a plurality of resonators to combine and simulate bulk acoustic wave filters to achieve bulk acoustic wave filter design, characterized in that: The following steps are involved: S1. Construct a machine learning model, wherein the machine learning model is used to obtain a bulk acoustic wave filter system function of an input bulk acoustic wave filter S parameter and an output bulk acoustic wave filter zero point, and train the machine learning model through a bulk acoustic wave filter S parameter data set and a bulk acoustic wave filter zero point data set; S2. Inputting the S parameters of the bulk acoustic wave filter that meet the design performance requirements into the machine learning model to obtain the system function of the bulk acoustic wave filter that meets the design performance requirements and the zero point of the bulk acoustic wave filter that meets the design performance requirements; S3, comprehensively meeting the design performance requirements of the BAW filter zero point, obtaining a plurality of resonator combinations of the BAW filter system function meeting the design performance requirements, and calculating the resonance point and static capacitance of each resonator, wherein the resonance point includes a series resonance frequency and a parallel resonance frequency; S4. Based on the resonance point and static capacitance of each resonator in the combination of the multiple resonators, the bulk acoustic wave filter Mason model is used to reversely deduce the bulk acoustic wave filter structural parameters that meet the design performance requirements, that is, the thickness and area of ​​each layer of each resonator.

2. The method for automated design of bulk acoustic wave filters based on machine learning according to claim 1, characterized in that: The machine learning model is a machine learning model of a back-propagation neural network architecture.

3. The method for automatic design of bulk acoustic wave filters based on machine learning according to claim 2, characterized in that: Step S1 specifically includes the following sub-steps: S11. Generate a random array Ω between (1,5) and (-5,-1) n , as a BAW filter zero array at a normalized frequency Ω, the random array Ω n The size is determined by the number of resonators N; S12, the BAW filter zero array Ω n Substitute the Chebyshev function to calculate the S parameters of the BAW filter, sample the BAW filter S21 parameter sequence and return loss RL sequence, and the BAW filter zero array Ω n Compose a machine learning data set and output it to the machine learning model for machine learning.

4. The method for automatic design of bulk acoustic wave filters based on machine learning according to claim 3, characterized in that: The Chebyshev function is as follows:

5. The method for automatic design of bulk acoustic wave filters based on machine learning according to claim 4, characterized in that: Step S2: Input the bulk acoustic wave filter S21 parameter sequence and the return loss RL sequence that meet the design performance requirements to the machine learning model.

6. The method for automatic design of bulk acoustic wave filters based on machine learning according to claim 5, characterized in that: Step S3 specifically includes the following steps: S31, obtain the ABCD matrix of each resonator in the combination of multiple resonators, the ABCD matrix of the series resonator ABCD Matrix of Parallel Resonators S32, multiplying the ABCD matrix of each resonator of the plurality of resonators to obtain the ABCD matrix of the bulk acoustic wave filter, thereby obtaining the S21 parameters of the bulk acoustic wave filter that meets the design performance requirements S33, the pole p of the bulk acoustic wave filter that meets the design performance requirements is obtained from the S21 parameter j,target , let BW be the filter bandwidth, p j The cost function of the synthesis algorithm is constructed for the BAW filter poles S34, combining the resonance point and static capacitance of each resonator of the plurality of resonators as data for the input synthesis algorithm, and using the pattern search algorithm to obtain the resonant series resonance frequency f when the cost function is minimized s , parallel resonant frequency f p And static capacitance C0.

7. The method for automated design of bulk acoustic wave filters based on machine learning according to claim 6, characterized in that: The BAW filter is a FBAR filter with a bandwidth of 2.4 GHz-2.5 GHz.

8. The method for automatic design of bulk acoustic wave filters based on machine learning according to claim 7, characterized in that: Step S4: In order to ensure process feasibility, the thickness of the bottom electrode of the resonator and the thickness and area of ​​the piezoelectric layer are set to constant values, and only the thickness and area of ​​the upper electrode are modified; the resonance point and static capacitance of each resonator in multiple resonator combinations are known, and the thickness and area of ​​the upper electrode of each resonator in the multiple resonator combinations are solved through the Mason model.

9. An automatic design device for bulk acoustic wave filters based on machine learning, characterized in that: Used to execute the automated design method of bulk acoustic wave filters based on machine learning as described in any one of claims 1 to 8.

Images