CN117272778A - Design methods and devices for microwave passive components - Google Patents

Abstract

The invention discloses a design method and device for microwave passive devices. The method includes: determining the design target and design parameter range; performing sampling to generate an initial design parameter sample set _XI and an extended sample set _XU ; performing full-wave analysis on _XI Through simulation calculation, the scattering parameters are obtained and the electromagnetic characteristic data set Y _I is extracted from it, and the initial design data set D _I is obtained; D _I is divided into a training set and a test set, and the forward multi-layer perceptron neural network and Gaussian process are trained and combined Calculate the _test error; predict the design performance _of D _I and D _E train the reversible neural network; generate design candidate values according to the design goals, conduct testing and verification to determine the final design parameters, and realize intelligent design.

Description

Design method and device for microwave passive components

Technical Field

The invention relates to a design method and device for a microwave passive device, in particular to a design method and device for a microwave passive device based on semi-supervised-reversible neural network inverse modeling, belonging to the technical field of microwave passive device design.

Background Art

With the development of microwave and radio frequency circuits, the performance requirements for microwave passive devices are gradually increasing. In order to meet the high performance requirements of microwave and radio frequency circuits, some novel microwave device structures emerge in an endless stream, making the design problems of microwave passive devices increasingly complex. The performance of microwave passive devices is usually related to their physical and dimensional parameters. The design method of traditional microwave devices relies on parameter scanning, and it is necessary to call simulation software multiple times to perform precise numerical calculations on microwave devices with different physical and dimensional parameters in order to optimize the most suitable design parameters. This greatly increases the design cycle of microwave passive devices and increases the design cost. Not only that, the design of microwave devices requires designers to have a deep understanding of electromagnetic fields and microwave technology, and the design threshold is high. Therefore, the intelligent and automated design of microwave devices has become very necessary today.

The inverse modeling design method has become an important development direction in the field of microwave passive device design. Inverse modeling is relative to forward modeling. The input of forward modeling is the design parameters, and the output is the design index. Inverse modeling refers to treating the design goal as the input of the model and the design parameters as the output of the model, and obtaining the design parameters directly from the design goal through the model. Therefore, the inverse modeling design method is the most convenient design method that does not rely on professional knowledge. At present, the inverse modeling method mainly relies on the multi-layer perceptron neural network model. Related research has shown that the multi-layer perceptron neural network inverse modeling method has certain application limitations.

The first limitation is that the use of multi-layer perceptron neural networks is not suitable for inverse modeling with one-to-many problems. The one-to-many problem in the inverse design of microwave passive devices refers to the fact that the same design goal may correspond to multiple sets of design parameters. The principle of the multi-layer perceptron neural network is to minimize the square mean error between the output predicted by the model and the true value of the data set. Therefore, when there is a one-to-many problem in the data set, using a neural network directly for inverse modeling will cause a huge error. The second limitation is that compared with forward modeling, the use of neural networks for inverse modeling requires a large number of data sets, and the acquisition of data sets needs to rely on full-wave simulation software. This process is very time-consuming, especially when the microwave device structure is complex and there are many design parameters. The data set required for neural network inverse modeling is larger, and the time consumption problem is more prominent.

Summary of the invention

In order to solve the above problems, the present invention proposes a design method and device for microwave passive components based on semi-supervised-reversible neural network inverse modeling, which can utilize reversible neural networks to improve the accuracy of inverse modeling, and proposes a semi-supervised learning method to improve the efficiency of generating data sets, so that designers do not rely on empirical knowledge and can directly obtain design parameters from design goals, thereby realizing intelligent design of microwave passive components.

The technical solution adopted by the present invention to solve the technical problem is:

In a first aspect, an embodiment of the present invention provides a method for designing a microwave passive device, comprising the following steps:

Determining a design goal of a microwave passive device to be designed, wherein the microwave passive device to be designed is a microwave passive device whose performance is to be improved;

Determining an allowable range of a design parameter of the microwave passive component to be designed according to a design goal of the microwave passive component to be designed, wherein the design parameter is a size parameter of the microwave passive component;

The Latin hypercube sampling method is used to sample the design parameters of the microwave passive device to generate an initial design parameter sample set _XI and an extended design parameter sample set _XU respectively;

Perform full-wave simulation calculation on the initial design parameter sample set _XI to obtain the scattering parameters of the initial design parameter samples, and extract the electromagnetic characteristic data set _YI from the scattering parameters as the design performance to obtain the initial design data set D _I ={ _XI , _YI };

The k-fold cross-validation method is used to divide the initial design data set D _I into a training set and a test set, and the forward multilayer perceptron neural network F _ANN and Gaussian process F _GP are trained, and the test errors of the multilayer perceptron neural network F _ANN and Gaussian process F _GP are calculated respectively;

If the test error does not meet the accuracy requirement, Latin hypercube sampling is used to continue adding samples to the initial design parameter sample set _XI , otherwise proceed to the next step;

The design performance of the extended design parameter sample set X _U is predicted by using multi-layer perceptron neural network F _ANN and Gaussian process F _GP respectively.

According to the modeling accuracy of the multi-layer perceptron neural network and the Gaussian process as weights, the weighted average of their predicted values is selected as the extended design performance; the extended design performance and the corresponding extended design parameters are used as the extended design data set _DE ;

The reversible neural network is trained using the initial design dataset D _I and the extended design dataset D _E ;

A reversible neural network is used to generate design candidate values according to the design objectives, and a multilayer perceptron neural network and Gaussian process are used for test verification. The design candidate value with the smallest test error is selected as the final design parameter.

As a possible implementation of this embodiment, the design target is the electromagnetic characteristics of the microwave passive device, including but not limited to the return loss at the working frequency band and the working frequency point. The working frequency band includes the upper cutoff frequency and the lower cutoff frequency corresponding to the working bandwidth of the microwave passive device.

As a possible implementation of this embodiment, the initial design parameter sample set _XI is:

_Xi = { _x1 , _x2 , ..., _xm }

Among them, m is the number of initial design parameter samples, x ₁ ,x ₂ ,…,x _m are initial design parameter samples;

The extended design parameter sample set X _U is:

X _U ＝{x _m+1 ,x _m+2 ,…,x _m+n }

Wherein, n is the number of extended design parameter samples, x _m+1 ,x _m+2 ,…,x _m+n are extended design parameter samples;

The electromagnetic property data set Y _I is:

Y _I ={y ₁ ,y ₂ ,…,y _m }

Among them, y ₁ ,y ₂ ,…,y _m are electromagnetic properties.

As a possible implementation of this embodiment, the calculation formula of the design performance is:

Among them, _xi is the initial design parameter sample or the extended design parameter sample, is the predicted value of the multilayer perceptron neural network, is the predicted value of the Gaussian process.

As a possible implementation of this embodiment, the calculation formula of the design performance is:

Among them, _yi is the design performance predicted by multilayer perceptron neural network and Gaussian process, _eANN is the test error of multilayer perceptron neural network, and _eGP is the test error of Gaussian process.

As a possible implementation of this embodiment, the reversible neural network is composed of reversible blocks:

hi ⁺¹ = ^gi ( ^hi )

Where, ^hi represents the input of the forward process of the i-th reversible block, and hi ⁺¹ represents the output of the forward process g ⁱ (·) of the i-th reversible block;

The specific calculation process of the forward process g ⁱ (·) is to first divide ^hi into [ ^hi _1:d , ^hi _d+1:D ] according to the dimension, and then calculate it using the following formula:

s(·) and t(·) are composed of arbitrary neural networks and are both mappings from d dimensions to D-d dimensions; ⊙ is the element-wise product symbol;

The reverse process of the reversible block is:

The variable conversion theorem defines the change in probability distribution from hi ⁺¹ to ^hi :

in, is the Jacobian of g ⁱ (h ⁱ ):

Where _Id is the d×d identity matrix, The diagonal elements are The diagonal matrix of .

As a possible implementation of this embodiment, the training of the reversible neural network is to use a back propagation algorithm to minimize the loss function of the reversible neural network to train the reversible neural network parameters θ _INN :

Among them, _XT is the training set design parameter, _Xpred is the design parameter obtained by the inverse process of the reversible neural network, b is the training set size, is the predicted value obtained by the reversible neural network forward process for the i-th design parameter in the training set, _yi is the corresponding true value; Z is the prior distribution, _Zpred is the predicted value of the reversible neural network forward process, and MMD is the maximum mean error;

Assuming data distribution P = {p ₁ ,p ₂ ,…,p _b } and Q = {q ₁ ,q ₂ ,…,q _b }, the calculation formula of MMD is:

Among them, k(·) is the kernel function;

The inverse multi-quadratic radial basis function is used as the kernel function:

Where ε is a hyperparameter, r ₁ = p _i , r ₂ = p _j , or r ₁ = q _i , r ₂ = q _j ;

When the loss function value converges, θ _INN is used as the parameter of the reversible neural network to obtain the F _INN model.

As a possible implementation of this embodiment, the method of using a reversible neural network to generate design candidate values according to a design goal includes:

Let h ¹ = x and h ^N = [y, z], where z satisfies the standard Gaussian distribution;

The reversible neural network transforms the distribution of z into a conditional probability distribution p(x|y), so that the reversible neural network can obtain multiple sets of design candidate values x by probability sampling under a given design target y;

Based on the Gaussian distribution, sample g z and the target y as the input of the reversible neural network and substitute them into the following formula:

x＝F _INN (y,z)

G design candidate values are obtained.

As a possible implementation of this embodiment, the process of testing and verifying using a multi-layer perceptron neural network and a Gaussian process is as follows:

The electromagnetic performance of the generated g design candidate values is calculated through the trained multi-layer perceptron neural network F _ANN and Gaussian process F _GP , and the error between the electromagnetic performance and the target electromagnetic performance is calculated:

in, and The output of the multilayer perceptron neural network and the output of the Gaussian process corresponding to the j-th design candidate value.

In a second aspect, an embodiment of the present invention provides a design device for a microwave passive device, comprising:

A design target determination module, used to determine a design target of a microwave passive device to be designed, wherein the microwave passive device to be designed is a microwave passive device whose performance is to be improved;

A design parameter range determination module, used to determine the allowable range of the design parameters of the microwave passive device to be designed according to the design target of the microwave passive device to be designed, wherein the design parameters are the size parameters of the microwave passive device;

A design parameter sample set generation module, used for sampling the design parameters of the microwave passive device to be designed by using the Latin hypercube sampling method, and generating an initial design parameter sample set _XI and an extended design parameter sample set _XU respectively;

A full-wave simulation calculation module is used to perform full-wave simulation calculation on the initial design parameter sample set _XI to obtain the scattering parameters of the initial design parameter samples, and extract the electromagnetic characteristic data set _YI from the scattering parameters as the design performance to obtain the initial design data set D _I ={ _XI , _YI };

A test error calculation module is used to divide the initial design data set _DI into a training set and a test set by using a k-fold cross-validation method, train a forward multi-layer perceptron neural network F _ANN and a Gaussian process F _GP , and calculate the test errors of the multi-layer perceptron neural network F _ANN and the Gaussian process F _GP respectively;

The test error judgment module is used to judge whether the test error meets the accuracy requirement. If it does not meet the accuracy requirement, Latin hypercube sampling is used to continue adding samples to the initial design parameter sample set _XI , otherwise it proceeds to the next step;

A design performance prediction module is used to predict the design performance of the extended design parameter sample set _XU by using a multi-layer perceptron neural network _FAN and a Gaussian process _FGP ;

The extended design data set determination module is used to select the weighted average of the predicted values as the extended design performance according to the modeling accuracy of the multi-layer perceptron neural network and the Gaussian process as weights; and the extended design performance and the corresponding extended design parameters are used as the extended design data set _DE ;

A reversible neural network training module, used for training a reversible neural network using an initial design data set D _I and an extended design data set D _E ;

The final design parameter determination module is used to use a reversible neural network to generate design candidate values according to the design objectives, and use a multi-layer perceptron neural network and a Gaussian process for test verification, and select the design candidate value with the smallest test error as the final design parameter.

The technical solution of the embodiment of the present invention may have the following beneficial effects:

A method for designing a microwave passive device according to the technical solution of an embodiment of the present invention,

The present invention utilizes a reversible neural network to improve the accuracy of inverse modeling, and proposes a semi-supervised learning method to improve the efficiency of generating a data set, so that the designer does not rely on empirical knowledge and can directly obtain design parameters from the design target, thereby realizing the intelligent design of microwave passive components; the present invention utilizes a semi-supervised reversible neural network to design microwave passive components, and the reversible neural network can directly generate multiple groups of design candidate values under given design performance conditions through a probability sampling method, thereby avoiding the experience participation of engineers and greatly improving the design efficiency; the present invention also proposes a semi-supervised learning method to improve the efficiency of generating a data set for reversible neural network training, and secondly proposes a new verification method to replace the use of full-wave simulation to verify the design parameters, thereby greatly improving the verification efficiency.

The application scenario of the present invention is the parameter design of microwave passive devices with fixed structures. The design parameters of microwave passive devices are determined according to different design performances. The efficiency of data set acquisition is improved through semi-supervised learning, and a reversible neural network is trained to learn the relationship between the design objectives and design parameters of microwave passive devices. The reversible neural network generates design candidate values according to the design objectives, and the design candidate values can obtain accurate design parameters after passing through a verification module.

The present invention introduces a reversible neural network into the inverse modeling of microwave passive devices, solves the one-to-many problem existing in the inverse modeling, and thus improves the accuracy of the inverse modeling. The present invention transforms the design problem of microwave passive devices into an inverse modeling problem, that is, the design target is regarded as the input of the model, the design parameters are regarded as the output of the model, and the design parameters are directly obtained from the design target through the model. The present invention adopts a reversible neural network, learns the distribution of design parameters under given design performance of microwave passive devices, generates design parameters through a probability sampling method, thereby being able to provide multiple groups of parameters, and can intelligently solve the one-to-many problem in the inverse modeling.

The present invention introduces a semi-supervised learning method to improve the efficiency of generating a data set. The present invention uses a forward model to provide a data set for inverse modeling, specifically using a small sample data set obtained from full-wave simulation software to train a multi-layer perceptron neural network and a Gaussian process as a forward model, and then uses the forward model to predict more design performance without design performance data, that is, predict the design performance of design parameters, thereby providing a larger data set for inverse modeling. The present invention obtains a data set through a semi-supervised learning method, which significantly reduces the time, thereby improving the efficiency of generating a data set.

The present invention uses the previously trained multi-layer perceptron neural network and Gaussian process as forward models, and sends the parameters generated by the reversible neural network to the two forward models for cross-validation, thereby improving the verification efficiency and accuracy.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for designing a microwave passive component according to an exemplary embodiment;

FIG2 is a schematic diagram of a reversible neural network according to an exemplary embodiment;

Fig. 3 is a schematic diagram of a mutually coupled affine layer of a basic unit of a reversible neural network according to an exemplary embodiment;

FIG4 is a flow chart showing a design of a microwave passive device using the apparatus of the present invention according to an exemplary embodiment;

FIG5 is a schematic structural diagram of a cavity filter according to an exemplary embodiment;

Fig. 6 is a schematic diagram showing a comparison between a design performance corresponding to a design parameter obtained by using the method of the present invention and a target design performance according to an exemplary embodiment.

DETAILED DESCRIPTION

The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

In order to clearly illustrate the technical features of the present solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The disclosure below provides many different embodiments or examples for realizing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below. In addition, the present invention may repeat reference numbers and/or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and/or settings discussed. It should be noted that the components illustrated in the accompanying drawings are not necessarily drawn to scale. The present invention omits the description of known components and processing techniques and processes to avoid unnecessary limitations on the present invention.

As shown in FIG1 , a design method for a microwave passive device provided by an embodiment of the present invention includes the following steps:

Determining a design goal of a microwave passive device to be designed, wherein the microwave passive device to be designed is a microwave passive device whose performance is to be improved;

Determining an allowable range of a design parameter of the microwave passive component to be designed according to a design goal of the microwave passive component to be designed, wherein the design parameter is a size parameter of the microwave passive component;

The Latin hypercube sampling method is used to sample the design parameters of the microwave passive device to generate an initial design parameter sample set _XI and an extended design parameter sample set _XU respectively;

Perform full-wave simulation calculation on the initial design parameter sample set _XI to obtain the scattering parameters of the initial design parameter samples, and extract the electromagnetic characteristic data set _YI from the scattering parameters as the design performance to obtain the initial design data set D _I ={ _XI , _YI };

The k-fold cross-validation method is used to divide the initial design data set D _I into a training set and a test set, and the forward multilayer perceptron neural network F _ANN and Gaussian process F _GP are trained, and the test errors of the multilayer perceptron neural network F _ANN and Gaussian process F _GP are calculated respectively;

If the test error does not meet the accuracy requirement, Latin hypercube sampling is used to continue adding samples to the initial design parameter sample set _XI , otherwise proceed to the next step;

The design performance of the extended design parameter sample set X _U is predicted by using multi-layer perceptron neural network F _ANN and Gaussian process F _GP respectively.

According to the modeling accuracy of the multi-layer perceptron neural network and the Gaussian process as weights, the weighted average of their predicted values is selected as the extended design performance; the extended design performance and the corresponding extended design parameters are used as the extended design data set _DE ;

The reversible neural network is trained using the initial design dataset D _I and the extended design dataset D _E ;

A reversible neural network is used to generate design candidate values according to the design objectives, and a multilayer perceptron neural network and Gaussian process are used for test verification. The design candidate value with the smallest test error is selected as the final design parameter.

The present invention adopts the Latin hypercube sampling method to effectively reflect the information of the sample space, obtains the design parameters of the initial data set and the extended data set through Latin hypercube sampling, and paves the way for subsequent modeling; obtains the scattering parameters of m design parameter samples through full-wave simulation, and extracts the electromagnetic characteristics Y _I ={y ₁ ,y ₂ ,…,y _m } from the scattering parameters, and uses them as the design performance to obtain the initial design performance data set D _I ={X _I ,Y _I }, so that the initial data set can be obtained, which provides a data set for forward modeling; the test errors of the multilayer perceptron neural network and Gaussian process are calculated in order to train the multilayer perceptron neural network and Gaussian process as forward models, which provide a basis for expanding the data set and verifying the design parameters; through the judgment of the test error, it is ensured that the multilayer perceptron neural network and Gaussian process with the required accuracy can be trained, so as to ensure the correctness of the obtained design parameters; the design performance of n design parameter samples are predicted by using the multilayer perceptron neural network and Gaussian process respectively, in order to obtain the design performance of the microwave passive device to be designed; the weighted average of the predicted values is selected as the extended design performance, which is convenient for obtaining the final extended data set and ensuring the accuracy of the data set; the reversible neural network is trained by training the reversible neural network to obtain multiple groups of design candidate values, and the design candidate value with the smallest error is selected as the final design parameter.

As a possible implementation of this embodiment, the design target is the electromagnetic characteristics of the microwave passive device, including but not limited to the return loss at the working frequency band and the working frequency point. The working frequency band includes the upper cutoff frequency and the lower cutoff frequency corresponding to the working bandwidth of the microwave passive device.

As a possible implementation of this embodiment, the initial design parameter sample set _XI is:

_Xi = { _x1 , _x2 , ..., _xm }

Among them, m is the number of initial design parameter samples, x ₁ ,x ₂ ,…,x _m are initial design parameter samples;

The extended design parameter sample set X _U is:

X _U ＝{x _m+1 ,x _m+2 ,…,x _m+n }

Wherein, n is the number of extended design parameter samples, x _m+1 ,x _m+2 ,…,x _m+n are extended design parameter samples;

The electromagnetic property data set Y _I is:

Y _I ={y ₁ ,y ₂ ,…,y _m }

Among them, y ₁ ,y ₂ ,…,y _m are electromagnetic properties.

As a possible implementation of this embodiment, the calculation formula of the design performance is:

Among them, _xi is the initial design parameter sample or the extended design parameter sample, is the predicted value of the multilayer perceptron neural network, is the predicted value of the Gaussian process.

As a possible implementation of this embodiment, the calculation formula of the design performance is:

Among them, _yi is the design performance predicted by multilayer perceptron neural network and Gaussian process, _eANN is the test error of multilayer perceptron neural network, and _eGP is the test error of Gaussian process.

As a possible implementation of this embodiment, the reversible neural network is composed of reversible blocks:

hi ⁺¹ = ^gi ( ^hi )

Where, ^hi represents the input of the forward process of the i-th reversible block, and hi ⁺¹ represents the output of the forward process g ⁱ (·) of the i-th reversible block;

The specific calculation process of the forward process g ⁱ (·) is to first divide ^hi into [ ^hi _1:d , ^hi _d+1:D ] according to the dimension, and then calculate it using the following formula:

s(·) and t(·) are composed of arbitrary neural networks and are both mappings from d dimensions to D-d dimensions; ⊙ is the element-wise product symbol;

The reverse process of the reversible block is:

The variable conversion theorem defines the change in probability distribution from hi ⁺¹ to ^hi :

in, is the Jacobian of g ⁱ (h ⁱ ):

Where _Id is the d×d identity matrix, The diagonal elements are The diagonal matrix of .

As a possible implementation of this embodiment, the training of the reversible neural network is to use a back propagation algorithm to minimize the loss function of the reversible neural network to train the reversible neural network parameters θ _INN :

Among them, _XT is the training set design parameter, _Xpred is the design parameter obtained by the inverse process of the reversible neural network, b is the training set size, is the predicted value obtained by the reversible neural network forward process for the i-th design parameter in the training set, _yi is the corresponding true value; Z is the prior distribution, _Zpred is the predicted value of the reversible neural network forward process, and MMD is the maximum mean error;

Assuming data distribution P = {p ₁ ,p ₂ ,…,p _b } and Q = {q ₁ ,q ₂ ,…,q _b }, the calculation formula of MMD is:

Among them, k(·) is the kernel function;

The inverse multi-quadratic radial basis function is used as the kernel function:

Where ε is a hyperparameter, r ₁ = p _i , r ₂ = p _j , or r ₁ = q _i , r ₂ = q _j ;

When the loss function value converges, θ _INN is used as the parameter of the reversible neural network to obtain the F _INN model.

As a possible implementation of this embodiment, the method of using a reversible neural network to generate design candidate values according to a design goal includes:

Let h ¹ = x and h ^N = [y, z], where z satisfies the standard Gaussian distribution;

The reversible neural network transforms the distribution of z into a conditional probability distribution p(x|y), so that the reversible neural network can obtain multiple sets of design candidate values x by probability sampling under a given design target y.

Based on the Gaussian distribution, sample g z and the target y as the input of the reversible neural network and substitute them into the following formula:

x＝F _INN (y,z)

G design candidate values are obtained.

As a possible implementation of this embodiment, the process of testing and verifying using a multi-layer perceptron neural network and a Gaussian process is as follows:

The electromagnetic performance of the generated g design candidate values is calculated through the trained multi-layer perceptron neural network F _ANN and Gaussian process F _GP , and the error between the electromagnetic performance and the target electromagnetic performance is calculated:

in, and The output of the multilayer perceptron neural network and the output of the Gaussian process corresponding to the j-th design candidate value.

The application scenario of the present invention is the parameter design of microwave passive devices with fixed structures. The method described in the present invention can be used to determine the design parameters of microwave passive devices according to different design performances. The design performance described in the present invention refers to the electromagnetic characteristics of microwave passive devices, including but not limited to electromagnetic performances such as return loss at the operating frequency band and operating frequency point; the design parameters refer to the size parameters of microwave passive devices. The method described in the present invention improves the efficiency of data set acquisition through semi-supervised learning, and trains a reversible neural network to learn the relationship between the design objectives and design parameters of microwave passive devices. The reversible neural network can generate design candidate values based on the design objectives, and the design candidate values can obtain accurate design parameters after passing through the verification module.

As shown in FIG2 , a microwave passive device design device provided by an embodiment of the present invention includes:

A design target determination module, used to determine a design target of a microwave passive device to be designed, wherein the microwave passive device to be designed is a microwave passive device whose performance is to be improved;

A design parameter range determination module, used to determine the allowable range of the design parameters of the microwave passive device to be designed according to the design target of the microwave passive device to be designed, wherein the design parameters are the size parameters of the microwave passive device;

A design parameter sample set generation module, used for sampling the design parameters of the microwave passive device to be designed by using the Latin hypercube sampling method, and generating an initial design parameter sample set _XI and an extended design parameter sample set _XU respectively;

A full-wave simulation calculation module is used to perform full-wave simulation calculation on the initial design parameter sample set _XI to obtain the scattering parameters of the initial design parameter samples, and extract the electromagnetic characteristic data set _YI from the scattering parameters as the design performance to obtain the initial design data set D _I ={ _XI , _YI };

A test error calculation module is used to divide the initial design data set _DI into a training set and a test set by using a k-fold cross-validation method, train a forward multi-layer perceptron neural network F _ANN and a Gaussian process F _GP , and calculate the test errors of the multi-layer perceptron neural network F _ANN and the Gaussian process F _GP respectively;

The test error judgment module is used to judge whether the test error meets the accuracy requirement. If it does not meet the accuracy requirement, Latin hypercube sampling is used to continue adding samples to the initial design parameter sample set _XI , otherwise it proceeds to the next step;

A design performance prediction module is used to predict the design performance of the extended design parameter sample set _XU by using a multi-layer perceptron neural network _FAN and a Gaussian process _FGP ;

The extended design data set determination module is used to select the weighted average of the predicted values as the extended design performance according to the modeling accuracy of the multi-layer perceptron neural network and the Gaussian process as weights; and the extended design performance and the corresponding extended design parameters are used as the extended design data set _DE ;

A reversible neural network training module, used for training a reversible neural network using an initial design data set D _I and an extended design data set D _E ;

The final design parameter determination module is used to use a reversible neural network to generate design candidate values according to the design objectives, and use a multi-layer perceptron neural network and a Gaussian process for test verification, and select the design candidate value with the smallest test error as the final design parameter.

As shown in Figures 3 and 4, the reversible neural network of the present invention is composed of reversible blocks, which are composed of mutually coupled affine layers. The calculation of the reversible block includes a forward process and a reverse process. Let the D-dimensional vector ^hi represent the input of the forward process of the i-th reversible block, and hi ⁺¹ represent the output of the forward process g ⁱ (·) of the i-th reversible block, then:

hi ⁺¹ = ^gi ( ^hi )

The specific calculation process of the forward process g ⁱ (·) is to first divide ^hi into [ ^hi _1:d , ^hi _d+1:D ] according to the dimension, and then calculate it using the following formula:

s(·) and t(·) are composed of arbitrary neural networks and are both mappings from d dimensions to D-d dimensions; ⊙ is the element-wise product symbol;

The reverse process of the reversible block is:

The variable conversion theorem defines the change in probability distribution from hi ⁺¹ to ^hi :

in is the Jacobian of g ⁱ (h ⁱ ). The Jacobian is:

where I _d is the d×d identity matrix, The diagonal elements are It can be found that the Jacobian determinant does not need to calculate the derivatives of s(·) and t(·), thus simplifying the calculation of the reversible neural network.

If h ¹ = x and h ^N = [y, z], where z satisfies the standard Gaussian distribution, then the reversible neural network can transform the distribution of z into a conditional probability distribution p(x|y), thereby enabling the reversible neural network to obtain multiple sets of design candidate values by probability sampling under given design objectives.

After determining the design target of the microwave passive device to be designed and the allowable range of the design parameters; as shown in FIG5 , the design process of the microwave passive device using the apparatus of the present invention is as follows.

Step 1: Sample the design parameters of the microwave passive device to be designed through Latin hypercube, and generate m design parameter samples _XI = { _x1 , _x2 , ..., _xm } and n design parameter samples _XU = { _xm+1 , _xm+2 , ..., xm _+n }. The Latin hypercube sampling method can effectively reflect the information of the sample space. The design parameters of the initial data set and the extended data set are obtained through Latin hypercube sampling, which lays the foundation for subsequent modeling.

Step 2: Obtain the scattering parameters of m design parameter samples through full-wave simulation software, and extract the electromagnetic characteristics Y _I = {y ₁ , y ₂ , …, y _m } from the scattering parameters, and use them as the design performance. Obtain the initial design performance data set D _I = {X _I , Y _I }. This step can obtain the initial data set, which provides a data set for forward modeling in step 3.

Step 3: Based on _DI , the k-fold cross-validation method is used to divide _DI into training set and test set to train the forward multi-layer perceptron neural network F _ANN and Gaussian process F _GP , and calculate the test error of the multi-layer perceptron neural network and Gaussian process. The purpose of this step is to train the multi-layer perceptron neural network and Gaussian process as forward models, which provides a basis for expanding the data set in step 5 and verifying the design parameters in step 9.

Step 4: If the test error does not meet the accuracy range, continue to add samples to _XI using Latin hypercube sampling; if the accuracy requirement is met, proceed to step 5. The purpose of this step is to ensure that a multilayer perceptron neural network and Gaussian process that meet the accuracy requirements can be trained, thereby ensuring that the results of steps 6 and 9 are correct.

Step 5: Use the multi-layer perceptron neural network and Gaussian process to predict the design performance of n design parameter samples respectively. The design performance is obtained using the following formula:

The purpose of this step is to obtain the design performance of the design parameters that did not have design performance in step 1.

Step 6: Based on the modeling accuracy of the multilayer perceptron neural network and the Gaussian process as weights, the weighted average of their predicted values is selected as the sample design performance. The obtained design performance y and its corresponding x are used as the extended data set _DE .

According to the test set errors e _ANN and e _GP corresponding to the multilayer perceptron neural network F _ANN and Gaussian process F _GP obtained in step 3, the sample design performance is calculated as:

The purpose of this step is to obtain the final extended dataset and ensure the accuracy of the dataset.

Step 7: Train a reversible neural network based on D _I and D _E.

Let ^hi represent the input of the forward process of the i-th reversible block, and hi ⁺¹ represent the output of the forward process g ⁱ (·) of the i-th reversible block, then:

hi ⁺¹ = ^gi ( ^hi )

The variable conversion theorem defines the change in probability distribution from hi ⁺¹ to ^hi :

in is the Jacobian determinant of g ⁱ (h ⁱ ). If h ¹ = x and h ^N = [y, z], and z satisfies the standard Gaussian distribution, then the reversible neural network is equivalent to converting the distribution of z into a conditional probability distribution p(x|y), so that the reversible neural network can obtain multiple sets of design parameters (x) by probability sampling under a given electromagnetic response (y).

The method for training the reversible neural network parameter θ _INN is to use the back propagation algorithm to minimize the loss function of the reversible neural network, that is,

Where _XT is the training set design parameter, _Xpred is the design parameter obtained by the inverse process of the reversible neural network, b is the training set size, is the predicted value obtained by the reversible neural network forward process for the i-th design parameter of the training set, and _yi is the corresponding true value. Z is the prior distribution, _Zpred is the predicted value of the reversible neural network forward process, and MMD is the maximum mean error. Assuming that the data distribution P＝{ _p1 , _p2 ,…, _pb } and Q＝{ _q1 , _q2 ,…, _qb }, the calculation formula of MMD is:

k(·) is a kernel function, which is selected as an inverse multi-quadratic radial basis function in the present invention. Its expression is:

Where ε is a hyperparameter, which can be obtained by grid search. By using the Adam algorithm to optimize the loss function of the reversible neural network, the reversible neural network parameter θ _INN can be trained. When the loss function value converges, θ _INN can be used as the parameter of the reversible neural network to obtain the F _INN model.

The purpose of this step is to train a reversible neural network for use in step 8 to generate design candidate values.

Step 8: Use a reversible neural network to generate g design candidate values according to the design goal.

First, g z are sampled based on Gaussian distribution, and the sampled z and target y are used as inputs of the reversible neural network. According to the formula x=F _INN (y, z), g design candidate values are output.

The purpose of this step is to obtain multiple sets of design candidate values for screening in step 9.

Step 9: Verify using multi-layer perceptron neural network and Gaussian process:

The electromagnetic performance of the g design candidate values generated in step 8 is calculated by the multilayer perceptron neural network F _ANN and Gaussian process F _GP trained in step 3, and the error between the electromagnetic performance and the target electromagnetic performance is calculated, that is:

in and The output of the multilayer perceptron neural network and the output of the Gaussian process corresponding to the j-th design candidate value.

The design candidate with the smallest error is selected as the final design parameter.

The above steps 1 to 7 can train a reversible neural network that can directly obtain design candidate values according to design performance under a given microwave passive device structure, and a multilayer perceptron neural network and Gaussian process that can be used to verify the design candidate values. Steps 8-9 are specific methods for how to use reversible neural networks and multilayer perceptron neural networks and Gaussian processes to obtain design parameters. The application scenario of the present invention is the parameter design of microwave passive devices with fixed structures. The design parameters of microwave passive devices are determined according to different design performances. The efficiency of data set acquisition is improved through semi-supervised learning, and a reversible neural network is trained to learn the relationship between the design objectives and design parameters of microwave passive devices; the reversible neural network generates design candidate values according to the design objectives, and the design candidate values can obtain accurate design parameters after passing through the verification module.

As shown in FIG6 , a cavity filter is designed using the method shown in the present invention. In FIG6 , L ₁ , L ₂ are the lengths of the resonant cavity of the cavity filter, W ₁ , W ₂ are the spacings between the resonant cavities, a and b are the length and width of the waveguide, which need to be fixed. In this example, a WR-75 waveguide is used, a is 19.05 mm, b is 9.525 mm; t is the width of the resonant cavity, and the cavity filter includes four design parameters [L ₁ , L ₂ , W ₁ , W ₂ , t]; the design performance index is a bandwidth of 11.85-12.15 GHz, and the maximum value of the return loss S ₁₁ within the bandwidth is -20 dB. The design steps of the cavity filter are as follows:

Step a1: Use Latin hypercube sampling to divide x = [L ₁ , L ₂ , W ₁ , W ₂ , t] into 10 layers, generate 160 sets of initial design parameters _XI and 600 sets of extended design parameters _XU , where the range of _L1 is 12-18 mm, the range of _L2 is 12-18 mm, the range of _W1 is 6-12 mm, the range of _W2 is 6-12 mm, and the range of t is 1-3 mm.

Step a2: Obtain the scattering parameters corresponding to the design parameters _XI through full-wave simulation software and extract the upper cutoff frequency, lower cutoff frequency and the maximum value of the return loss ( _S11 ) in the working frequency band corresponding to the working bandwidth as the design index y = [ _fL , _fH , _S11max ], obtain the design performance _YI of the data set, and then obtain the initial data set _DI = { _XI , _YI }.

Step a3: Based on D _I , the multilayer perceptron neural network and Gaussian process model are trained using the k-fold cross-validation method, and the test errors of the multilayer perceptron neural network and Gaussian process are calculated.

Step a4: If the test error does not meet the accuracy range, continue to add samples to _XI using Latin hypercube sampling; if the accuracy meets the requirements, go to step a5. In this specific embodiment, the initial data set can already train a multilayer perceptron neural network and Gaussian process that meet the accuracy, so there is no need to add additional samples.

Step a5: Use the trained multi-layer perceptron neural network and Gaussian process model to predict the upper cutoff frequency, lower cutoff frequency and maximum value of return loss (S ₁₁ ) corresponding to the working bandwidth of the extended design parameter X _U .

Step a6: Based on the modeling accuracy of the multilayer perceptron neural network and the Gaussian process as weights, the weighted average of their predicted values is selected as the sample design performance. The obtained design performance y and its corresponding x are used as the extended data set _DE .

Step a7: Train a reversible neural network based on D _I and D _E.

Step a8: Based on the given target design performance y = [11.85, 12.15, -20], randomly sample 100 groups of z from the Gaussian distribution, use y and z as the input of the reversible neural network, and obtain 100 groups of design candidate values. The design performance and target design performance corresponding to the design parameters obtained by the method proposed in the present invention are shown in Figure 7, in which the dotted line is the design target and the solid line is the design performance given by the proposed method.

Step a9: Using the multilayer perceptron neural network and Gaussian process verification, the design candidate value x = [13.6024, 14.9591, 9.4705, 6.3201, 1.5187] (mm) with the smallest error is selected as the final design parameter.

The comparison of the accuracy of inverse modeling using a multilayer perceptron neural network and the accuracy of the proposed method is shown in Table 1. Table 1:

The specific comparison method is to calculate the mean square error between the design performance of the design parameters obtained by each method and the target design performance. It is worth noting that the mean square error calculation method of the method shown in the present invention is to calculate the mean square error between the design performance corresponding to the design candidate value and the target performance, rather than the mean square error between the design performance corresponding to the design parameters and the target performance obtained after the final verification.

The present invention introduces a reversible neural network into the inverse modeling of microwave passive devices, solves the one-to-many problem existing in the inverse modeling, and thus improves the accuracy of the inverse modeling. The present invention converts the design problem of microwave passive devices into an inverse modeling problem, that is, the design target is regarded as the input of the model, the design parameter is regarded as the output of the model, and the design parameter is directly obtained from the design target through the model. There is a one-to-many problem in inverse modeling, that is, the same design performance corresponds to multiple sets of design parameters, and the traditional modeling method will encounter obstacles. If the multi-layer perceptron neural network in the patent with the publication number of CN109284541A and the patent name of "A neural network multi-physics modeling method for microwave passive devices" is used for inverse modeling, because its modeling process is to minimize the error between the neural network prediction value and the true value, when there is a one-to-many problem, the output value of the multi-layer perceptron neural network will be the average value of multiple sets of design parameters corresponding to the design performance, thereby causing a large error. The present invention adopts a reversible neural network, by learning the distribution of design parameters under the given design performance of microwave passive devices, generating design parameters by the method of probability sampling, so that multiple sets of parameters can be provided, and the one-to-many problem in inverse modeling can be intelligently solved, thereby improving the accuracy of inverse modeling.

The present invention introduces a semi-supervised learning method to improve the efficiency of generating a data set. The traditional method of obtaining a data set is to use a method of randomly generating design parameters, and bring all the design parameters into the full-wave simulation software for calculation to obtain the corresponding electromagnetic performance, that is, the design performance value. Since the full-wave simulation time is long, it is relatively slow to obtain the design performance corresponding to the design parameters. The present invention uses a forward model to provide a data set for inverse modeling, specifically, a multi-layer perceptron neural network and a Gaussian process are trained using a small sample data set obtained from the full-wave simulation software as a forward model, and then the forward model is used to predict more design performance without design performance data, that is, to predict the design performance of the design parameters, thereby providing a larger data set for inverse modeling. The present invention obtains a data set through a semi-supervised learning method, which significantly reduces the time, thereby improving the efficiency of generating a data set.

The present invention proposes a method for verifying design parameters. Traditional verification requires that the design parameters be sent to full-wave simulation software for verification, which is relatively time-consuming. The present invention uses a previously trained multi-layer perceptron neural network and a Gaussian process as a forward model, and sends the parameters generated by a reversible neural network to two forward models for cross-validation, thereby improving verification efficiency and accuracy.

The present invention trains a reversible neural network that can directly obtain a design candidate value according to the design performance under a given microwave passive device structure, and a multi-layer perceptron neural network and a Gaussian process that can be used to verify the design candidate value.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims (10)

1. A method for designing a microwave passive device, characterized in that it comprises the following steps: Determining a design goal of a microwave passive device to be designed, wherein the microwave passive device to be designed is a microwave passive device whose performance is to be improved; Determining an allowable range of a design parameter of the microwave passive component to be designed according to a design goal of the microwave passive component to be designed, wherein the design parameter is a size parameter of the microwave passive component; The Latin hypercube sampling method is used to sample the design parameters of the microwave passive device to generate an initial design parameter sample set _XI and an extended design parameter sample set _XU respectively; Perform full-wave simulation calculation on the initial design parameter sample set _XI to obtain the scattering parameters of the initial design parameter samples, and extract the electromagnetic characteristic data set _YI from the scattering parameters as the design performance to obtain the initial design data set D _I ={ _XI , _YI }; The k-fold cross-validation method is used to divide the initial design data set D _I into a training set and a test set, and the forward multilayer perceptron neural network F _ANN and Gaussian process F _GP are trained, and the test errors of the multilayer perceptron neural network F _ANN and Gaussian process F _GP are calculated respectively; If the test error does not meet the accuracy requirement, Latin hypercube sampling is used to continue adding samples to the initial design parameter sample set _XI , otherwise proceed to the next step; The design performance of the extended design parameter sample set X _U is predicted by using multi-layer perceptron neural network F _ANN and Gaussian process F _GP respectively. According to the modeling accuracy of the multi-layer perceptron neural network and the Gaussian process as weights, the weighted average of their predicted values is selected as the extended design performance; the extended design performance and the corresponding extended design parameters are used as the extended design data set _DE ; The reversible neural network is trained using the initial design dataset D _I and the extended design dataset D _E ; A reversible neural network is used to generate design candidate values according to the design objectives, and a multilayer perceptron neural network and Gaussian process are used for test verification. The design candidate value with the smallest test error is selected as the final design parameter. 2. The design method of a microwave passive device according to claim 1 is characterized in that the design target is the electromagnetic characteristics of the microwave passive device, including but not limited to the return loss at the operating frequency band and the operating frequency point. 3. The method for designing a microwave passive device according to claim 1, wherein the initial design parameter sample set _XI is: _Xi = { _x1 , _x2 , ..., _xm } Among them, m is the number of initial design parameter samples, x ₁ ,x ₂ ,…,x _m are initial design parameter samples; The extended design parameter sample set X _U is: X _U ＝{x _m+1 ,x _m+2 ,…,x _m+n } Wherein, n is the number of extended design parameter samples, x _m+1 ,x _m+2 ,…,x _m+n are extended design parameter samples; The electromagnetic property data set Y _I is: Y _I ={y ₁ ,y ₂ ,…,y _m } Among them, y ₁ ,y ₂ ,…,y _m are electromagnetic properties. 4. The method for designing a microwave passive device according to claim 1, wherein the calculation formula for the design performance is: Among them, _xi is the initial design parameter sample or the extended design parameter sample, is the predicted value of the multilayer perceptron neural network, is the predicted value of the Gaussian process. 5. The method for designing a microwave passive device according to claim 4, wherein the calculation formula for the design performance is: Among them, _yi is the design performance predicted by multilayer perceptron neural network and Gaussian process, _eANN is the test error of multilayer perceptron neural network, and _eGP is the test error of Gaussian process. 6. The method for designing a microwave passive device according to any one of claims 1 to 5, characterized in that the reversible neural network is composed of reversible blocks: hi ⁺¹ = ^gi ( ^hi ) Where, ^hi represents the input of the forward process of the i-th reversible block, and hi ⁺¹ represents the output of the forward process g ⁱ (·) of the i-th reversible block; The specific calculation process of the forward process g ⁱ (·) is to first divide ^hi into [ ^hi _1:d , ^hi _d+1:D ] according to the dimension, and then calculate it using the following formula: s(·) and t(·) are composed of arbitrary neural networks and are both mappings from d dimensions to D-d dimensions; ⊙ is the element-wise product symbol; The reverse process of the reversible block is: The variable conversion theorem defines the change in probability distribution from hi ⁺¹ to ^hi : in, is the Jacobian of g ⁱ (h ⁱ ): Where _Id is the d×d identity matrix, The diagonal elements are The diagonal matrix of . 7. The method for designing a microwave passive device according to claim 6, wherein the training of the reversible neural network is to use a back propagation algorithm to minimize the loss function of the reversible neural network to train the reversible neural network parameter θ _INN : Among them, _XT is the training set design parameter, _Xpred is the design parameter obtained by the inverse process of the reversible neural network, b is the training set size, is the predicted value obtained by the reversible neural network forward process for the i-th design parameter in the training set, _yi is the corresponding true value; Z is the prior distribution, _Zpred is the predicted value of the reversible neural network forward process, and MMD is the maximum mean error; Assuming data distribution P = {p ₁ ,p ₂ ,…,p _b } and Q = {q ₁ ,q ₂ ,…,q _b }, the calculation formula of MMD is: Among them, k(·) is the kernel function; The inverse multi-quadratic radial basis function is used as the kernel function: Where ε is a hyperparameter, r ₁ = p _i , r ₂ = p _j , or r ₁ = q _i , r ₂ = q _j ; When the loss function value converges, θ _INN is used as the parameter of the reversible neural network to obtain the F _INN model. 8. The method for designing a microwave passive device according to claim 7, wherein the step of generating design candidate values according to a design goal using a reversible neural network comprises: Let h ¹ = x and h ^N = [y, z], where z satisfies the standard Gaussian distribution; The reversible neural network transforms the distribution of z into a conditional probability distribution p(x|y), so that the reversible neural network can obtain multiple sets of design candidate values x by probability sampling under a given design target y; Based on the Gaussian distribution, sample g z and the target y as the input of the reversible neural network and substitute them into the following formula: x＝F _INN (y,z) G design candidate values are obtained. 9. The method for designing a microwave passive device according to claim 8, characterized in that the process of testing and verifying by using a multi-layer perceptron neural network and a Gaussian process is: The electromagnetic performance of the generated g design candidate values is calculated through the trained multi-layer perceptron neural network F _ANN and Gaussian process F _GP , and the error between the electromagnetic performance and the target electromagnetic performance is calculated: in, and The output of the multilayer perceptron neural network and the output of the Gaussian process corresponding to the j-th design candidate value. 10. A design device for a microwave passive device, comprising: A design target determination module, used to determine a design target of a microwave passive device to be designed, wherein the microwave passive device to be designed is a microwave passive device whose performance is to be improved; A design parameter range determination module, used to determine the allowable range of the design parameters of the microwave passive device to be designed according to the design target of the microwave passive device to be designed, wherein the design parameters are the size parameters of the microwave passive device; A design parameter sample set generation module, used for sampling the design parameters of the microwave passive device to be designed by using the Latin hypercube sampling method, and generating an initial design parameter sample set _XI and an extended design parameter sample set _XU respectively; A full-wave simulation calculation module is used to perform full-wave simulation calculation on the initial design parameter sample set _XI to obtain the scattering parameters of the initial design parameter samples, and extract the electromagnetic characteristic data set _YI from the scattering parameters as the design performance to obtain the initial design data set D _I ={ _XI , _YI }; A test error calculation module is used to divide the initial design data set _DI into a training set and a test set by using a k-fold cross-validation method, train a forward multi-layer perceptron neural network F _ANN and a Gaussian process F _GP , and calculate the test errors of the multi-layer perceptron neural network F _ANN and the Gaussian process F _GP respectively; The test error judgment module is used to judge whether the test error meets the accuracy requirement. If it does not meet the accuracy requirement, Latin hypercube sampling is used to continue adding samples to the initial design parameter sample set _XI , otherwise it proceeds to the next step; A design performance prediction module is used to predict the design performance of the extended design parameter sample set _XU by using a multi-layer perceptron neural network _FAN and a Gaussian process _FGP ; The extended design data set determination module is used to select the weighted average of the predicted values as the extended design performance according to the modeling accuracy of the multi-layer perceptron neural network and the Gaussian process as weights; and the extended design performance and the corresponding extended design parameters are used as the extended design data set _DE ; A reversible neural network training module, used for training a reversible neural network using an initial design data set D _I and an extended design data set D _E ; The final design parameter determination module is used to use a reversible neural network to generate design candidate values according to the design objectives, and use a multi-layer perceptron neural network and a Gaussian process for test verification, and select the design candidate value with the smallest test error as the final design parameter.