A radio frequency circuit optimization design method based on high-performance computing

Abstract

A radio frequency circuit optimization design method based on high-performance computing, comprising the following steps: sub-modularizing the radio frequency circuit; constructing a circuit sub-module neural network; using GPU for parallel computing based on the circuit sub-module neural network; cascading the scattering parameter results of each level of sub-module circuit to obtain the scattering parameter results of the overall circuit; searching for an optimization scheme that meets the performance index condition from the scattering parameter results of the overall circuit; finding the Pareto optimal solution for the pros and cons of the search results, and screening out the circuit that meets the actual engineering index. Compared with the traditional artificial neural network used for modeling the relationship between the design size and electrical behavior of the full-stage radio frequency circuit, the method is more flexible, can improve the reusability and accuracy of the artificial neural network model, and can accelerate the calculation by using the high parallel computing characteristics of the neural network model.

Description

Technical Field

[0001] This invention is applied to the field of radio frequency circuit simulation, and in particular, it is a radio frequency circuit optimization design method based on high-performance computing. Background Technology

[0002] Radio frequency integrated circuits (RFICs) refer to circuit designs that operate in the 30kHz to 300GHz frequency band. Common circuit types include low-noise amplifiers, power amplifiers, oscillators, mixers, and filters. RFICs are widely used in various fields, including but not limited to radio broadcasting, satellite navigation, wireless communication, radar remote sensing, and human body imaging.

[0003] With the rapid development of smartphones, 5G, and IoT technologies in recent years, analog integrated circuits, especially radio frequency (RF) integrated circuits, have attracted increasing attention. However, compared to the rapid development of digital integrated circuit technology, the technological progress of analog RF integrated circuits has been relatively slow, and their design difficulty is also higher. This is mainly due to the presence of numerous parasitic effects and crosstalk in high-frequency circuits. Practical design relies heavily on the experience of RF engineers. Furthermore, with the development of communication technology, the operating frequencies of various communication devices are constantly increasing, further exacerbating the challenges and difficulties in RF circuit design. One reason for this is the lack of more mature electronic design automation (EDA) tools to support the design process in RF circuit design compared to digital circuit design.

[0004] The rapid development of artificial intelligence (AI) technology in recent years has given industry professionals great expectations and confidence in the prospects of AI technology empowering EDA tools. This has brought new possibilities and broader design space to analog RF circuit design, and provided EDA technicians with a new perspective on how intelligent EDA tools can assist circuit engineers in designing and improving work efficiency.

[0005] Currently, a relatively mature solution to circuit optimization problems is based on several natural heuristic algorithms (evolutionary algorithms, particle swarm optimization, and reinforcement learning methods, etc.) to explore the circuit design space and find the optimal solution for a given circuit problem. However, the vast circuit design space requires a large number of simulation iterations for exploration, which consumes a significant amount of time. Although the large amount of data generated during simulation iterations is used to train artificial neural network (ANN) models, hoping that the neural network models can replace simulators to save time, the neural network models obtained in this way can only be applied to a single fixed circuit, and their models lack reusability and flexibility; at the same time, the large design dimensionality of the circuit also easily leads to the problem of high training difficulty for neural network models. Summary of the Invention

[0006] To address the problems of high time cost, low reusability and flexibility of neural network models, and high training difficulty in RF circuit optimization, this invention provides a method for RF circuit design based on high-performance computing.

[0007] The present invention adopts the following technical solution:

[0008] A method for optimizing radio frequency circuit design based on high-performance computing includes the following steps:

[0009] S1. The radio frequency circuit is divided according to its function to obtain several circuit sub-modules, which include: amplifier circuit module, matching circuit module, and filter circuit module.

[0010] S2, Constructing a neural network for circuit sub-modules

[0011] Each circuit sub-module is sampled in the design space, and the circuit is simulated for possible design parameters in the design space to obtain a dataset of circuit simulation results covering the design space.

[0012] The dataset includes an input portion and an output portion:

[0013] The input section includes: the size design parameters of all passive and active components in the circuit submodule and the simulation frequency points. The passive components include: resistors, inductors, capacitors, transmission lines and T-junctions. The active components include transistors. The size design parameters include the length and width of the components, the number of turns and inner diameter of the inductor coil, and the simulation frequency points refer to the frequency points sampled according to the set simulation frequency step size on the simulation frequency band.

[0014] The output section reflects the performance metrics of its circuit, including: scattering parameters,

[0015] A neural network model is established for each circuit submodule; the input of the neural network model is the device parameters and design dimensions of the circuit submodule; the output of the neural network model is the scattering parameters as the circuit simulation result; the basic unit of the neural network model is a perceptron.

[0016] The dataset is divided into a training set and a validation set. The neural network model is trained using the training set.

[0017] In the trained neural network, the error between the predicted and actual values ​​of the samples is determined using a validation set. If the error meets the preset requirements, the neural network training is completed; otherwise, training continues. After training and validation are completed, the circuit submodule neural network is obtained.

[0018] S3. Based on a neural network of circuit submodules, parallel computing is performed using a GPU:

[0019] S3.1. Sample the possible design dimension parameters of each circuit submodule within the specified design space to form an input parameter matrix. :

[0020]

[0021] in Representing the The first sampling of design dimensional parameters One design parameter, It represents the first The set of design parameters for the sub-circuit module is given in the above formula. Represents the common The sampling of design parameters is done in groups, with each group of design parameters being... A dimensional vector, that is, one that contains Each design dimension;

[0022] S3.2. Set the optimization frequency range and step size to form an input parameter matrix with frequency point information. :

[0023]

[0024] in To be the frequency points sampled within the specified frequency range and step size. The input parameter matrix contains frequency point information. One of the submatrices represents a set of design schemes that incorporates frequency point information as input parameters at multiple simulation frequency points. Each submatrix originates from S3.1. The row vectors in the matrix are repeatedly expanded and then include frequency information at the end;

[0025] S3.3. Input this matrix into the neural network model, and use a GPU to accelerate the neural network model to achieve high-concurrency operation, and obtain the scattering parameter matrix output:

[0026]

[0027] in Here is the scattering parameter matrix, with dimension . Among them For example, its meaning represents a total The first in the group design dimension sampling Group sampling, corresponding The sampling frequency point of the nth sampling frequency point Scattering parameters at each frequency point ;

[0028] To facilitate understanding of steps S3 and S4 of this invention, the cascade calculations involved in this invention are explained below:

[0029] For cascaded front-end networks for Port network, cascaded network for Port network, where The ports are cascaded, among which Port Network The scattering parameter matrix can be expressed as:

[0030]

[0031] in The definition of is:

[0032]

[0033] Its physical meaning is the use of incident wave voltage. excitation Port and measurement reflected wave voltage from the port The ratio of the two. This definition also requires that, except for The incident wave on all other ports except the main port is set to zero, i.e., a matched load is connected to avoid reflection. That is, when all ports are connected to the matching load, The reflection coefficient as seen from the port, and When all other ports are connected to load, from port to The port's transmission coefficient.

[0034] and Port Network The scattering parameter matrix can be expressed as:

[0035]

[0036] Port Network and Port Network Cascaded network Among them, there are a total of Each port is cascaded, network have The scattering parameter matrix of each port can be expressed as:

[0037]

[0038] in, , , ,and The definition is as follows:

[0039]

[0040]

[0041]

[0042]

[0043] in, It is an identity matrix.

[0044] when , When this happens, the cascade formula can be simplified to:

[0045]

[0046]

[0047]

[0048]

[0049] S4. Cascade the scattering parameter results of each sub-module circuit to obtain the scattering parameter results of the overall circuit:

[0050] S4.1. Obtain the scattering parameter matrices of the neural network for each level of circuit sub-modules from S3.3. The scattering parameters at each frequency point are divided into four block matrices, namely... , , , Information on the real and imaginary parts of the four scattering parameters:

[0051]

[0052] in, Representing the The block matrix of the multi-level network at the set frequency band range and step size sampling frequency points ;

[0053] S4.2. The neural networks of each circuit submodule are cascaded in the manner shown by the following four formulas:

[0054]

[0055]

[0056]

[0057]

[0058] in , representing the preceding cascaded network of Scattering parameters, , representing the preceding cascaded network of Scattering parameters, , representing the network With the network The network formed by cascading of Scattering parameters;

[0059] Cascaded computation It is a complex number. ;

[0060] S5. From the scattering parameter results of the overall circuit, search for an optimized solution that meets the performance index conditions; compare the quality of the search results to find the Pareto optimal solution, and select the circuit that meets the actual engineering index.

[0061] Preferably, the radio frequency circuit is a low-noise amplifier of a two-port network.

[0062] Preferably, in step S2, the scattering parameters are output in vector form, including: input reflection coefficient. Output reflection coefficient Transmission coefficient It consists of the real and imaginary parts of four scattering parameters, and has a total of 8 dimensions, which is... real part , imaginary part , real part , imaginary part , real part , imaginary part , real part , imaginary part .

[0063] Preferably, the performance index conditions in step S5 include the following four requirements:

[0064] Requirement 1: In order to reduce the input return loss, the input reflection coefficient in the operating frequency band must be less than a fixed circuit design specification.

[0065] Requirement 2: The gain flatness should be less than a fixed circuit design specification. Ideally, the gain of an amplifier circuit should remain constant within the passband.

[0066] Requirement 3: The forward voltage gain must be as large as possible, exceeding a fixed circuit design specification;

[0067] Requirement 4: In order to reduce the return loss of the output, the output reflection coefficient in the operating frequency band must be less than a fixed circuit design specification.

[0068] Preferably, the four requirements are represented by the following four objective functions as constraints:

[0069]

[0070]

[0071]

[0072]

[0073] in This represents the spatial distribution of frequency points sampled within a specified frequency band according to a specified step size. Represents the frequency point The corresponding scattering parameters on, The objective function represents the frequency band selected. On The maximum value, based on requirement one, when the selected frequency band On When the maximum value also meets the design specifications, then the selected frequency band Other frequency points on the selected frequency band also meet the requirements. On The smaller the maximum value, the better the circuit performance is considered to be; the objective function should be minimized as much as possible.

[0074] It is the objective function that represents Flatness, used in the selected frequency band On The difference between the maximum and minimum values ​​is used as a measure. Based on requirement two, the lower the flatness, the smoother the gain changes in the specified frequency band, and the better the performance. This objective function should be minimized as much as possible.

[0075] The objective function represents the frequency band selected. superior The minimum value, based on requirement three, when the selected frequency band On When the minimum value also meets the design specifications, then the selected frequency band Other frequency points on the selected frequency band also meet the requirements. On The larger the minimum value, the greater the circuit gain, and the better the circuit performance is considered. This objective function should be maximized as much as possible.

[0076] The objective function represents the frequency band selected. On The maximum value, based on requirement four, is when the selected frequency band... On When the maximum value also meets the design specifications, then the selected frequency band Other frequency points on the selected frequency band also meet the requirements. On The smaller the maximum value, the smaller the output return loss, and the better the circuit performance is considered. This objective function should be minimized as much as possible.

[0077] Preferably, the constraint values ​​for the four objective functions are as follows:

[0078] , , , ;

[0079] The scattering parameters of the overall circuit obtained through cascading in step S4 are first filtered using the four constraint values ​​mentioned above, and then further optimized by searching for superior values. and Come and find the frontiers of Pareto.

[0080] The beneficial effects of this technical solution are:

[0081] Compared to traditional artificial neural networks used to model the relationship between the size and electrical behavior of full-level RF circuit designs, this method is more flexible, improves the reusability and accuracy of artificial neural network models, and leverages the high parallel computing capability of neural network models by using graphics processing units (GPUs) to accelerate computation and perform direct cascade calculations on scattering parameters. Attached Figure Description

[0082] Figure 1 This is a flowchart for intelligent optimization design of radio frequency circuits based on high-performance computing;

[0083] Figure 2 This is a schematic diagram of a low-noise amplifier circuit used in the implementation scheme;

[0084] Figure 3 It is the amplifier circuit sub-module one;

[0085] Figure 4 It is the second sub-module of the amplifier circuit;

[0086] Figure 5 It is the first sub-module of the matching circuit;

[0087] Figure 6 This is a diagram of a network cascading system;

[0088] Figure 7 This is a diagram showing the training results of the neural network model for the amplifier circuit submodule.

[0089] Figure 8 This is a diagram showing the training results of the neural network model for the amplifier circuit submodule 2;

[0090] Figure 9 This is a diagram showing the training results of the neural network model for the matching circuit submodule.

[0091] Figure 10 This is a comparison chart of the time costs of CPU and GPU search solutions;

[0092] Figure 11 It is a Pareto front plot that uses forward voltage gain and output reflection coefficient as indicators;

[0093] Figure 12 This is a comparison chart of the S11 results between the initial and optimized schemes;

[0094] Figure 13 This is a comparison chart of the S21 results between the initial and optimized schemes;

[0095] Figure 14 This is a comparison chart of the initial and optimized schemes for S22 results. Detailed Implementation

[0096] The following specific implementation examples illustrate the embodiments of the present invention. Those skilled in the art can intuitively understand other advantages of the present invention from the content described in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0097] Accuracy verification of this invention:

[0098] Reference Figure 1 This embodiment describes the specific steps for intelligent optimization design of radio frequency circuits based on high-performance computing:

[0099] S1. Sub-modularization of RF circuit functionalities: An RF circuit design may contain multiple amplifier and matching circuit modules. These different circuits are sub-modularized. In this example, the implementation object is a low-noise amplifier, such as... Figure 2 The circuit structure was analyzed and broken down into three sub-modules, the structures of which are as follows: Figure 3 , Figure 4 , Figure 5 .

[0100] S2. Establish a neural network model for the sub-module circuit; use the device parameters and design dimensions of the sub-module as the input of the neural network model, and the circuit simulation results of the scattering parameters as the output, and model the mathematical relationship between the input and output.

[0101] In step 2, the specific process of establishing a neural network model for the sub-module circuit is as follows:

[0102] S2.1. Sample each sub-module circuit in the design space, and use computer-aided design tools to perform extensive simulations of the circuit for possible design parameters, generating a dataset of circuit simulation results that fully covers the design space. This dataset takes the size design parameters of all passive and active components in the sub-circuit module, as well as the simulation frequency points, as inputs. Passive components include resistors, inductors, capacitors, transmission lines, and T-junctions, while active components include transistors. Size design parameters include the length and width of each component, and the number of turns and inner diameter of the inductor. Simulation frequency points refer to the frequency points sampled according to the set simulation frequency step size within the simulation frequency band. The circuit's performance metrics (scattering parameters) are outputs. Taking the two-port scattering parameters as an example, its output vector is derived from the input reflection coefficient. Output reflection coefficient Transmission coefficient It consists of the real and imaginary parts of four scattering parameters, and has a total of 8 dimensions, which is... (Right now (the real part) (Right now (the imaginary part) (Right now (the real part) (Right now (the imaginary part) (Right now (the real part) (Right now (the imaginary part) (Right now (the real part) (Right now (The imaginary part).

[0103] S2.2. Perform preprocessing on the training data of the dataset in S2.1 (optional, such as normalization, standardization, etc.). Taking standardization as an example:

[0104]

[0105]

[0106]

[0107] in Represents the data vector in the original dataset. Represents the number of samples in the dataset. This represents the average of all sample data vectors in the dataset. This represents the standard deviation of all sample data vectors in the dataset. This represents the new dataset after the original dataset has undergone data standardization.

[0108] S2.3. Divide the preprocessed dataset into a training set and a validation set according to a set ratio.

[0109] S2.3. Input the sample training set into the artificial neural network for training. The basic unit of the artificial neural network is the perceptron, and its mathematical model can be expressed as:

[0110]

[0111] in The input vector values ​​represent the training set of samples. Represents the activation function. Represents the weight matrix. Represents the bias matrix. This represents the output of the basic unit of the sensor.

[0112] The training method is as follows:

[0113]

[0114]

[0115] Here Represents the output of the neural network. The label vectors represent the actual sample training set. For loss function, This represents the set learning rate. It represents the set of all weights and biases in a neural network.

[0116] S2.4. Input the sample validation set into the trained neural network and determine whether the error between the sample predicted value and the true value meets the preset requirements. If yes, complete the neural network training; otherwise, return to S2.3 and continue training. The training results of the three sub-modules are as follows: Figure 6 , Figure 7 , Figure 8 .

[0117] S3 and GPUs accelerate highly parallel computing for neural network models;

[0118] In step 3, the specific operations for accelerating the highly parallel computing process of neural network models using GPUs are as follows:

[0119] S3.1. Sample the possible design dimension parameters of each sub-module circuit within the specified design space to form an input parameter matrix:

[0120]

[0121] in Representing the The first sampling of design dimensional parameters One design parameter, It represents the first The set of design parameters for the sub-circuit module is given in the above formula. Represents the common The sampling of design parameters is done in groups, with each group of design parameters being... A dimensional vector, that is, one that contains Each design dimension.

[0122] S3.2. Define the desired optimization frequency range and step size to form an input parameter matrix with frequency point information:

[0123]

[0124] in To be the frequency points sampled within the specified frequency range and step size. This is an input parameter matrix containing frequency point information.

[0125] S3.3. Input this matrix into the neural network model and use GPU acceleration to achieve high-concurrency operation. Taking a two-port network as an example, the scattering parameter matrix is ​​output:

[0126]

[0127] in The output scattering parameter matrix, since it is a two-port network, has the following dimensions. Among them For example, its meaning represents a total The first in the group design dimension sampling Group sampling, corresponding The sampling frequency point of the nth sampling frequency point Scattering parameters at each frequency point .

[0128] S4. Perform cascade calculations on the circuit models of each sub-module. The cascade diagram is shown below. Figure 9 As shown, the scattering parameter results of each sub-module circuit are cascaded and calculated to obtain the scattering parameter results of the entire circuit.

[0129] In step 4, the specific operations for the cascading calculation of the circuit models of each sub-module are as follows:

[0130] S4.1. The scattering parameter matrix of the neural network model of each sub-circuit module obtained in S3.3. The scattering parameters at each frequency point are divided into four block matrices. For a two-port network, these four block matrices are... , , , Information on the real and imaginary parts of the four scattering parameters:

[0131]

[0132] Representing the The multi-level network samples the frequency points according to the given desired frequency band range and step size. Block matrix.

[0133] S4.2. When two-port networks are cascaded together, the cascading formula is:

[0134]

[0135]

[0136]

[0137] in , representing the preceding cascaded network of Scattering parameters, , representing the preceding cascaded network of Scattering parameters, , representing the network With the network The network formed by cascading of Scattering parameters.

[0138] S4.3. Cascade Calculation It is a complex number. The rules for complex field calculations should be used:

[0139]

[0140]

[0141]

[0142] S5. Search for optimization schemes that meet the performance index conditions; analyze the final scattering parameter results, compare the search results to find the Pareto optimal solution, and select circuits that meet the actual engineering indicators.

[0143] In step 5, the specific steps for searching for optimization solutions that meet the performance metric conditions are as follows:

[0144] This example is a low-noise amplifier circuit, which has four requirements for the scattering parameters in general:

[0145] Requirement 1: In order to reduce the input return loss, the input reflection coefficient in the operating frequency band must be less than a fixed circuit design specification.

[0146] Requirement 2: The gain flatness should be less than a fixed circuit design specification. Ideally, the gain of an amplifier circuit should remain constant within the passband.

[0147] Requirement 3: The forward voltage gain must be as large as possible, exceeding a fixed circuit design specification;

[0148] Requirement 4: To reduce the return loss of the output, the output reflection coefficient in the operating frequency band must be less than a fixed circuit design specification. We use four objective functions as constraints to represent these four requirements.

[0149] S5.1. For this example, four objective functions are defined:

[0150]

[0151]

[0152]

[0153]

[0154] in This represents the spatial distribution of frequency points sampled within a specified frequency band according to a specified step size. Represents the frequency point The corresponding scattering parameters on, The objective function represents the frequency band selected. On The maximum value, based on requirement one, when the selected frequency band On When the maximum value also meets the design specifications, then the selected frequency band Other frequency points on the selected frequency band also meet the requirements. On The smaller the maximum value, the better the circuit performance is considered, so the objective function should be minimized as much as possible; It is the objective function that represents Flatness, used in the selected frequency band On The difference between the maximum and minimum values ​​is used as a measure. Based on requirement two, the lower the flatness, the smoother the gain changes in the specified frequency band, and the better the performance. Therefore, this objective function should be minimized as much as possible. The objective function represents the frequency band selected. superior The minimum value, based on requirement three, when the selected frequency band On When the minimum value also meets the design specifications, then the selected frequency band Other frequency points on the selected frequency band also meet the requirements. On The larger the minimum value, the greater the circuit gain, and the better the circuit performance is considered. Therefore, the objective function should be maximized as much as possible. The objective function represents the frequency band selected. On Maximum value, based on requirement four, when the selected frequency band On When the maximum value also meets the design specifications, then the selected frequency band Other frequency points on the selected frequency band also meet the requirements. On The smaller the maximum value of the objective function, the smaller the output return loss, and the better the circuit performance is considered. Therefore, the objective function should be minimized as much as possible.

[0155] S5.2. Multi-objective optimization problems involve trade-offs between multiple performance indicators. First, constraints are set according to the actual circuit design requirements. , , , .in and Having met the circuit design specifications, the scattering parameters of the cascaded circuit in S4 were first screened against four constraints, and then further optimized by searching for superior parameters. and Come and find the frontiers of Pareto.

[0156] Performance verification of this invention:

[0157] As described above, the method mentioned earlier leverages the high concurrency capabilities of neural network models and the advantages of GPU multithreading and parallel computing to accelerate neural network model computation and cascaded calculation of scattering parameters. Furthermore, by dividing the circuit into multiple sub-modules, the input dimension of the neural network model is reduced, the model training difficulty is decreased, and the model accuracy is improved. The training results are as follows: Figure 7 , Figure 8 , Figure 9 As shown. Furthermore, the cascading computation of submodules with any number of ports improves the reusability of the submodule model and enhances the flexibility of the method. Therefore, to verify the performance of the proposed method, based on... Figure 2 The low-noise amplifier circuit was optimized and verified, and the following conclusions were drawn:

[0158] First, comparing the time cost of calculating scattering parameters on a central processing unit (CPU) versus on a GPU, the results are as follows: Figure 10 As shown in the figure, it can be seen that as the data scale increases, the time cost on the CPU also increases linearly. It can be seen that the increase in time cost of the high-concurrency computing method of the GPU is very limited with the increase in data scale. Its time cost is mainly sacrificed in the search scheme. The time cost of the CPU can be 5 to 8 times that of the GPU.

[0159] Secondly, in the optimization of this circuit, after searching for optimization solutions for 1138 seconds using this method, 1129 optimization solutions that meet the constraints were found. , , Further efforts are needed to find better... and Searching for the Pareto frontier, such as Figure 11 .

[0160] Finally, an optimized scheme was selected from the Pareto frontier, and the simulation results of the original scheme were compared with those of the optimized scheme. Simultaneously, the simulation results of the optimized scheme were compared with the prediction results of the neural network model. Figure 12 , Figure 13 and Figure 14 From the image, we can see that... , The performance indicators have all been significantly improved. and The maximum values ​​of all were reduced), while for Performance metrics, while meeting hard constraints In this case, it sacrificed a certain amount Performance was improved, but flatness was reduced. At the same time, it can be seen that the neural network model's prediction results can predict the trend and magnitude of scattering parameter changes to a certain extent.

[0161] This invention applies to the field of radio frequency (RF) circuit optimization design and specifically describes its design and implementation process. This invention can be well applied to RF circuit optimization design. The method for intelligent optimization design of RF circuits based on high-performance computing proposed in this invention has significant advantages: it decomposes the circuit into multiple sub-circuit modules, reducing the difficulty of model training, improving model accuracy, increasing model reusability, and enhancing algorithm flexibility; compared to CPU processing and calculating scattering parameters, the high-concurrency operation of the GPU greatly improves the efficiency of searching for optimization solutions and reduces time costs; simultaneously, the final optimization solution for the original solution is also more obvious, showing significant performance improvement. Therefore, the method of this invention can be well applied to the field of RF circuit design and plays a promoting role in the development of its optimization design.

[0162] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention. The scope of the present invention is determined by the scope of the appended claims.

Claims

1. A radio frequency circuit optimization design method based on high-performance computing, characterized in that, Includes the following steps: S1. The radio frequency circuit is divided according to its function to obtain several circuit sub-modules, which include: amplifier circuit module, matching circuit module, and filter circuit module. S2, Constructing a neural network for circuit sub-modules Each circuit sub-module is sampled in the design space, and the circuit is simulated for possible design parameters in the design space to obtain a dataset of circuit simulation results covering the design space. The dataset includes an input portion and an output portion: The input section includes: the size design parameters of all passive and active components in the circuit submodule and the simulation frequency points. The passive components include: resistors, inductors, capacitors, transmission lines and T-junctions. The active components include transistors. The size design parameters include the length and width of the components, the number of turns and inner diameter of the inductor coil, and the simulation frequency points refer to the frequency points sampled according to the set simulation frequency step size on the simulation frequency band. The output section reflects the performance metrics of its circuit, including: scattering parameters, A neural network model is established for each circuit submodule; the input of the neural network model is the device parameters and design dimensions of the circuit submodule; the output of the neural network model is the scattering parameters as the circuit simulation result; the basic unit of the neural network model is a perceptron. The dataset is divided into a training set and a validation set. The neural network model is trained using the training set. In the trained neural network, the error between the predicted and actual values ​​of the samples is determined using a validation set. If the error meets the preset requirements, the neural network training is completed; otherwise, training continues. After training and validation are completed, the circuit submodule neural network is obtained. S3. Based on a neural network of circuit submodules, parallel computing is performed using a GPU: S3.

1. Sample the possible design dimension parameters of each circuit submodule within the specified design space to form an input parameter matrix. : ； in Representing the The first sampling of design dimensional parameters One design parameter, It represents the first The set of design parameters for the sub-circuit module is given in the above formula. Represents the common The sampling of design parameters is done in groups, with each group of design parameters being... A dimensional vector, that is, one that contains Each design dimension; S3.

2. Set the optimization frequency range and step size to form an input parameter matrix with frequency point information. : ； in To be the frequency points sampled within the specified frequency range and step size. The input parameter matrix contains frequency point information. One of the submatrices represents a set of design schemes that incorporates frequency point information as input parameters at multiple simulation frequency points. Each submatrix originates from S3.

1. The row vectors in the matrix are repeatedly expanded and then include frequency information at the end. S3.

3. Input this matrix into the neural network model, and use a GPU to accelerate the neural network model to achieve high-concurrency operation, and obtain the scattering parameter matrix output: ； in Here is the scattering parameter matrix, with dimension . ,in Its meaning represents the common The first in the group design dimension sampling Group sampling, corresponding The sampling frequency point of the nth sampling frequency point Scattering parameters at each frequency point ; S4. Cascade the scattering parameter results of each sub-module circuit to obtain the scattering parameter results of the overall circuit: S4.

1. Obtain the scattering parameter matrices of the neural network for each level of circuit sub-modules from S3.

3. The scattering parameters at each frequency point are divided into four block matrices, namely... , , , Information on the real and imaginary parts of the four scattering parameters: ； in, Representing the The block matrix of the multi-level network at the set frequency band range and step size sampling frequency points ; S4.

2. The neural networks of each circuit submodule are cascaded in the manner shown by the following four formulas: ； ； ； ； in , representing the preceding cascaded network of Scattering parameters, , representing the preceding cascaded network of Scattering parameters, , representing the network With the network The network formed by cascading of Scattering parameters; Cascaded computation It is a complex number. ; S5. From the scattering parameter results of the overall circuit, search for an optimized solution that meets the performance index conditions; compare the quality of the search results to find the Pareto optimal solution, and select the circuit that meets the actual engineering index.

2. The radio frequency circuit optimization design method based on high-performance computing as described in claim 1, characterized in that, The radio frequency circuit is a low-noise amplifier for a two-port network.

3. The radio frequency circuit optimization design method based on high-performance computing as described in claim 2, characterized in that, In step S2, the scattering parameters are output in vector form, including: input reflection coefficient. Output reflection coefficient Transmission coefficient It consists of the real and imaginary parts of four scattering parameters, and has a total of 8 dimensions, which is... real part , imaginary part , real part , imaginary part , real part , imaginary part , real part , imaginary part .

4. The radio frequency circuit optimization design method based on high-performance computing as described in claim 3, characterized in that, The performance index conditions in step S5 include the following four requirements: Requirement 1: In order to reduce the input return loss, the input reflection coefficient in the operating frequency band must be less than a fixed circuit design specification. Requirement 2: The gain flatness should be less than a fixed circuit design specification. Ideally, the gain of an amplifier circuit should remain constant within the passband. Requirement 3: The forward voltage gain must be as large as possible, exceeding a fixed circuit design specification; Requirement 4: In order to reduce the return loss of the output, the output reflection coefficient in the operating frequency band must be less than a fixed circuit design specification.

5. The radio frequency circuit optimization design method based on high-performance computing as described in claim 4, characterized in that, The four requirements are represented by the following four objective functions as constraints: ； ； ； ； in This represents the spatial distribution of frequency points sampled within a specified frequency band according to a specified step size. Represents the frequency point The corresponding scattering parameters on, The objective function represents the frequency band selected. On The maximum value, based on requirement one, when the selected frequency band On When the maximum value also meets the design specifications, then the selected frequency band Other frequency points on the selected frequency band also meet the requirements. On The smaller the maximum value, the better the circuit performance is considered to be; the objective function should be minimized as much as possible. It is the objective function that represents Flatness, used in the selected frequency band On The difference between the maximum and minimum values ​​is used as a measure. Based on requirement two, the lower the flatness, the smoother the gain changes in the specified frequency band, and the better the performance. This objective function should be minimized as much as possible. The objective function represents the frequency band selected. superior The minimum value, based on requirement three, when the selected frequency band On When the minimum value also meets the design specifications, then the selected frequency band Other frequency points on the selected frequency band also meet the requirements. On The larger the minimum value, the greater the circuit gain, and the better the circuit performance is considered. This objective function should be maximized as much as possible. The objective function represents the frequency band selected. On The maximum value, based on requirement four, is when the selected frequency band... On When the maximum value also meets the design specifications, then the selected frequency band Other frequency points on the selected frequency band also meet the requirements. On The smaller the maximum value, the smaller the output return loss, and the better the circuit performance is considered. This objective function should be minimized as much as possible.

6. The radio frequency circuit optimization design method based on high-performance computing as described in claim 5, characterized in that, The constraint values ​​for the four objective functions are as follows: ， ， ， ； The scattering parameters of the overall circuit obtained through cascading in step S4 are first filtered using the four constraint values ​​mentioned above, and then further optimized by searching for superior values. and Come and find the frontiers of Pareto.

Images