Machine learning assisted microwave circuit optimization method based on high-dimensional output proxy model

Abstract

The invention discloses a machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model, and the method comprises the steps: carrying out the learning of a plurality of performance parameters in a working frequency band of a microwave circuit through the high-dimensional output Gaussian process regression (HOGP), and obtaining a proxy model; and a collection function based on Monte Carlo sampling is optimized to the maximum, and an existing optimal solution is explored. The HOGP maps high-dimensional output to a low-dimensional potential space, and the correlation is modeled through a covariance structure of potential variables. Compared with a traditional machine learning optimization method, the method has the advantages that the convergence speed is higher, the optimization result is better, and meanwhile robustness is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of microwave circuit design and relates to a machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model. Background Art

[0002] Modern wireless communication systems feature high transmission rates, high spectral efficiency, and low power consumption, placing increasingly stringent demands on the bandwidth, efficiency, and linearity of microwave circuit systems. Over the past few years, harmonic control technologies, including harmonic termination and reflection suppression, have been the primary means of achieving high-performance microwave circuits. To expand operating bandwidth, advanced broadband matching techniques, such as multimode resonance and asymmetric coupling structures, have been extensively researched. In practical microwave circuit design, the desired frequency response characteristics are typically achieved through optimization of distributed parameter networks. To balance broadband performance and size constraints, multilayer structures or composite left-handed and right-handed transmission lines are often required. This sophisticated design requires extremely high engineering expertise. While modern EDA tools such as ADS and HFSS can accurately simulate complex electromagnetic field problems, they still have significant shortcomings in the synthesis of intelligent matching networks.

[0003] In recent years, real-frequency and simplified real-frequency techniques have become popular optimization methods for designing broadband matching networks. However, their optimization process is simple and ideal microstrip lines are generated in the matching network. Due to the presence of step microstrip lines, there are certain differences between the results of the layout and the schematic, making it difficult to achieve good results at the layout level. Therefore, it is necessary to further optimize the layout. Among advanced optimization algorithms, machine learning-assisted optimization (MLAO) is widely used in microwave circuits. During the optimization process, the relationship between input parameters and response results is predicted by training a Gaussian process regression (GPR) model, replacing simulation, which significantly reduces computational costs and improves the efficiency of power amplifier optimization. Currently, the vast majority of machine learning-assisted power amplifier optimization methods use a single-output surrogate model and only use acquisition functions calculated by formulas, making it difficult to accurately explore optimization directions for complex fitness functions. Summary of the Invention

[0004] Purpose of the invention: In response to the problems existing in the prior art, the purpose of the present invention is to provide a machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model, which can ensure the effectiveness and robustness of the algorithm and ensure optimization performance and computational efficiency.

[0005] Technical Solution: To achieve the above-mentioned purpose, the present invention provides a machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model, comprising the following steps:

[0006] Initialize and set the optimization target, frequency band to be optimized, structural parameters and optimization range of the microwave circuit;

[0007] Use Latin hypercube sampling to generate samples that evenly distribute the exploration space;

[0008] The performance data of the microwave circuit samples obtained by simulation and the parameter data of the samples are combined into a training set and the fitness function is calculated; the fitness function combines all the performance data of all the frequency points to be optimized;

[0009] The training set is trained using a high-dimensional Gaussian process regression machine learning method to obtain a proxy model; the model outputs all performance data of all frequency points to be optimized;

[0010] Maximize the expected improvement based on Monte Carlo sampling using optimization algorithms, and use surrogate models to predict the response between microwave circuit structural parameters and performance;

[0011] The optimized microwave circuit structural parameters are simulated to verify the results of the algorithm exploration and obtain real performance data;

[0012] Determine whether the simulation results meet the optimization goal. If not, add the current optimized performance data to the dataset and repeat the surrogate model training to this step until the loop stops.

[0013] During specific implementation, the optimization target of the microwave circuit includes one or more performance targets within the operating frequency band.

[0014] Preferably, the fitness function fitness of the microwave circuit performance is calculated as follows:

[0015]

[0016] in f j The performance of the i-th optimization target under the frequency point, m is the number of frequency points to be optimized, s is the number of targets to be optimized, and the fitness function is suitable for maximizing the performance of each target to be optimized. The larger the fitness function, the better the performance of the target to be optimized.

[0017] As a preference, each sample in the training set X={x1,…,x n}, n is the number of parameters to be optimized, and the corresponding output data is

[0018] Preferably, Monte Carlo sampling is used to generate N samples f from the posterior distribution of the surrogate model (1) (x),…,f (N) (x), and according to the current best fitness function value f best Calculate the expected improvement:

[0019]

[0020] And use the gradient optimization algorithm to maximize EI MC (x):

[0021] x opt =argmaxEI MC (x)

[0022] f (i) (x) is the fitness function value corresponding to the i-th sample structure parameter x, x opt is the optimal structural parameter.

[0023] In some embodiments, the microwave circuit is a power amplifier, and the optimization objectives include one or more of power added efficiency, drain efficiency, and saturated output power within the operating frequency band; the parameters to be optimized include the length and width of each microstrip line in the matching network.

[0024] A computer system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the steps of the machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model are implemented.

[0025] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model.

[0026] A computer program product includes a computer program, which, when executed by a processor, implements the steps of the machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model.

[0027] Beneficial effects: Compared with the prior art, the present invention takes into account that microwave circuits, such as power amplifiers, have many performance indicators and significant changes at different frequencies. It is necessary to consider the performance of all performance indicators at all frequencies, and the target space to be optimized is further expanded. The high-dimensional Gaussian process regression adopted by the present invention can capture the correlation between outputs through the covariance structure of the latent variables. Only one proxy model needs to be trained, which reduces the computational complexity. The expected improvement estimation based on Monte Carlo sampling adopted by the present invention, facing a composite fitness function composed of multiple performance objectives, calculates the expected improvement using the sample distribution generated by the posterior distribution of the proxy model, can more accurately find the optimization iteration direction, is not easily trapped in the local optimal solution, and can accelerate the optimization convergence speed. Experiments show that the optimization method using the present invention converges faster than the traditional machine learning method and can achieve better performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1It is a flow chart of the auxiliary microwave circuit optimization method of the present invention.

[0029] Figure 2 It is a schematic diagram of the power amplifier structure used to verify the algorithm proposed in the present invention.

[0030] Figure 3 It is a schematic diagram of the power amplifier layout used to verify the algorithm proposed in the present invention.

[0031] Figure 4 It is an optimization iteration curve chart using different strategies.

[0032] Figure 5 The PAE simulation curves before and after optimization are shown in Figure 2, where (a) is the power added efficiency (PAE) curve and (b) is the saturated output power (P out )curve. DETAILED DESCRIPTION

[0033] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0034] In this embodiment, taking a power amplifier as an example, a machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model is disclosed. The main steps are as follows:

[0035] Step S101: Initialize and set the optimization target, frequency band to be optimized, structural parameters and optimization range of the power amplifier. The optimization target includes one or more of power added efficiency (PAE), drain efficiency (DE) and saturated output power (Pout) within the operating frequency band.

[0036] Step S102: Based on the parameter settings initialized in step S101, Latin hypercube sampling is used to generate samples that are evenly distributed in the exploration space.

[0037] Step S103: Using the EM simulation model to obtain the performance data of the power amplifier sample, and together with the parameter data of the sample, form a training set and calculate the fitness function; wherein the fitness function combines all the performance data of all the frequency points to be optimized. Specifically, the fitness function fitness can be calculated according to the following formula:

[0038]

[0039] in f j The performance of the i-th optimization target under the frequency point, m is the number of frequency points to be optimized, s is the number of targets to be optimized, and the fitness function is suitable for maximizing the performance of each target to be optimized. The larger the fitness function, the better the performance of the target to be optimized.

[0040] Step S104: The training level is trained using a high-dimensional Gaussian process regression machine learning method to obtain a proxy model; the proxy model outputs all performance data of all frequency points to be optimized. Specifically, each sample X in the training set is {x1,…,x n}, n is the number of parameters to be optimized, and the corresponding output data is

[0041] Step S105: maximizing the expected improvement based on Monte Carlo sampling using an optimization algorithm, and using the proxy model in step S104 to predict the response between the power amplifier structural parameters and performance.

[0042] Specifically, in this step, Monte Carlo sampling is used to generate N samples f from the posterior distribution of the proxy model. (1) (x),…,f (N) (x), and according to the current best fitness function value f best Calculate expected improvement:

[0043]

[0044] And use the gradient optimization algorithm to maximize EI MC (x):

[0045] x opt =argmaxEI MC (x)

[0046] f (i) (x) is the fitness function value corresponding to the i-th sample structure parameter x, x opt is the optimal structural parameter.

[0047] Step S106: Performing EM simulation on the optimal structural parameters of the power amplifier obtained by optimization in step S105 to verify the results of the algorithm exploration and obtain real performance data.

[0048] Step S107: Determine whether the simulation result reaches the optimization target. If not, add the currently optimized performance data to the data set and repeat steps S104 to S107 until the loop stops.

[0049] The following combination Figures 1 to 5 , a real power amplifier structure example is used to illustrate the superiority and robustness of the proposed method. Figure 2 As shown in the figure, a schematic diagram of the power amplifier structure based on the CGH40010F power tube is given. The optimized frequency band is 1.0~2.0GHz. The parameters to be optimized are shown in Table 1, including the length (L) and width (W) of the microstrip line of each stage of the matching network.

[0050] Table 1

[0051] parameter <![CDATA[L1]]> <![CDATA[L2]]> <![CDATA[L3]]> <![CDATA[L4]]> <![CDATA[L5]]> <![CDATA[L6]]> <![CDATA[L7]]> <![CDATA[L8]]> <![CDATA[L9]]> <![CDATA[L 10 ]]> <![CDATA[L 11 ]]> Size / mm 7.53 7.03 7.19 7.46 1.5 33 7.81 7.27 7.30 7.43 7.33 parameter <![CDATA[W1]]> <![CDATA[W2]]> <![CDATA[W3]]> <![CDATA[W4]]> <![CDATA[W5]]> <![CDATA[W6]]> <![CDATA[W7]]> <![CDATA[W8]]> <![CDATA[W9]]> <![CDATA[W 10 ]]> <![CDATA[W 11 ]]> Size / mm 2.00 10.17 5.65 2.50 2.00 1.00 0.84 4.32 3.93 2.73 1.00

[0052] Based on the above power amplifier structure, the machine learning-assisted power amplifier optimization method based on the high-dimensional output proxy model has the following main implementation steps:

[0053] Step S201: Constructing an initial sample set: Selecting the optimization parameters and optimization interval of the power amplifier, randomly sampling within the optimization interval, and then performing calculations using EM simulation; Figure 2 Taking the power amplifier in as an example, the optimization parameters are the 22 listed in Table 1. Latin Hypercube Sampling (LHS) is used to obtain 50 sets of input parameter combinations X, where the dimension of X is 50×22. The power amplifier model with the structural parameter X is simulated by EM, and the power added efficiency (PAE) and saturated output power (P out ), PAE and P out The dimension is m×50, where m is the number of frequency points. The power amplifier model is simulated in ADS.

[0054] Step S202: Set the maximum number of iterations to 100, and process the power amplifier performance data obtained in step S201 accordingly according to the optimization goal: the optimization goal is to maximize PAE and P in the range of 1.0-2.0 GHz. out , together with X in step S201, form a training set D = [XY], where Calculate the fitness function for each sample's performance:

[0055]

[0056]

[0057] This is used to judge the performance of the power amplifier.

[0058] Step S203: Using a machine learning algorithm to learn the training set data to obtain a cheap proxy model. Specifically, using high-order Gaussian process regression (HOGP) for training to learn the relationship between the power amplifier structure parameters and the corresponding output targets to obtain the proxy model U.

[0059] Step S204: Optimize the proxy model U in step S203 using a gradient optimization algorithm, and generate N samples f from the posterior distribution of the proxy model using Monte Carlo sampling. (1) (x),…,f (N) (x), and according to the current best fitness function value f best Calculate expected improvement:

[0060]

[0061] And use the gradient optimization algorithm to maximize EI MC (x) Get the optimal parameter combination:

[0062] x opt =argmaxEI MC (x)

[0063] Step S205: Combine the optimal structural parameters obtained in this generation into X opt Bring in EM simulation to verify the results of algorithm exploration and obtain real performance data Y opt .

[0064] Step S206: Determine whether the current simulation result meets the optimization index. If not, add the optimal solution of this generation to the training set in step S202, and repeat steps S203 to S205 until the number of iterations reaches the set maximum value of 100.

[0065] Optimization results using the proposed algorithm:

[0066] Figure 4 The comparison results of the optimization strategy based on high-dimensional Gaussian process regression and Monte Carlo sampling expectation improvement and single-output Gaussian process regression and expectation improvement strategy (GPR+EI) and probability improvement (GPR+PI) are given. It can be seen that HOGP+EI MC At the 17th iteration, the optimization strategy surpassed the performance of a single-output Gaussian process regression strategy. In subsequent iterative optimization processes, the proposed strategy, after 50 iterations, achieved superior power amplifier performance compared to the traditional strategy after 100 iterations.

[0067] Figure 5 The final optimization results of the present invention are given. (a) is the comparison before and after PAE optimization. The PAE of the power amplifier optimized by the algorithm of the present invention is 67% to 77% in the 1-2GHz range, which is better than the initial design and the power amplifier optimized by the traditional machine learning algorithm; (b) is P out Comparing before and after optimization, the saturated output power of the power amplifier optimized by the algorithm of the present invention is greater than 38.7 dBm, which is better than the 36.4 dBm of the initial design.

[0068] An embodiment of the present invention also discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the steps of the aforementioned machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model are implemented.

[0069] An embodiment of the present invention also discloses a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the aforementioned machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model.

[0070] The embodiment of the present invention also discloses a computer program product, including a computer program, which, when executed by a processor, implements the steps of the aforementioned machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model.

[0071] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model, characterized in that: The following steps are involved: Initialize and set the optimization target, frequency band to be optimized, structural parameters and optimization range of the microwave circuit; Use Latin hypercube sampling to generate samples that evenly distribute the exploration space; The performance data of the microwave circuit samples obtained by simulation and the parameter data of the samples are combined into a training set and the fitness function is calculated; the fitness function combines all the performance data of all the frequency points to be optimized; The training set is trained using a high-dimensional Gaussian process regression machine learning method to obtain a proxy model; the model outputs all performance data of all frequency points to be optimized; Maximize the expected improvement based on Monte Carlo sampling using optimization algorithms, and use surrogate models to predict the response between microwave circuit structural parameters and performance; The optimized microwave circuit structural parameters are simulated to verify the results of the algorithm exploration and obtain real performance data; Determine whether the simulation results meet the optimization goal. If not, add the current optimized performance data to the dataset and repeat the surrogate model training to this step until the loop stops.

2. The machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model according to claim 1 is characterized in that: The optimization objectives of microwave circuits include one or more performance objectives within the operating frequency band.

3. The machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model according to claim 1, characterized in that: The fitness function fitness of microwave circuit performance is calculated as follows: in f j The performance of the i-th optimization target under the frequency point, m is the number of frequency points to be optimized, s is the number of targets to be optimized, and the fitness function is suitable for maximizing the performance of each target to be optimized. The larger the fitness function, the better the performance of the target to be optimized.

4. The machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model according to claim 1, characterized in that: Each sample in the training set X={x1,…,x n }, n is the number of parameters to be optimized, and the corresponding output data is s is the number of targets to be optimized, m is the number of frequency points to be optimized, f j The performance of the i-th optimization objective under the frequency point, i∈{1,2,...,s},j∈{1,2,...,m}.

5. The machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model according to claim 1, characterized in that: Generate N samples f from the posterior distribution of the surrogate model using Monte Carlo sampling (1) (x),…,f (N) (x), and according to the current best fitness function value f best Calculate expected improvement: And use the gradient optimization algorithm to maximize EI MC (x): x opt =argmaxEI MC (x) f (i) (x) is the fitness function value corresponding to the i-th sample structure parameter x, x opt is the optimal structural parameter.

6. The machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model according to claim 1, characterized in that: The microwave circuit is a power amplifier, and the optimization objectives include one or more of power added efficiency, drain efficiency, and saturated output power within the operating frequency band.

7. The machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model according to claim 6, characterized in that: The parameters to be optimized include the length and width of the microstrip line at each stage of the matching network.

8. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is loaded into a processor, the steps of the machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model are implemented according to any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model according to any one of claims 1 to 7 are implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the machine learning-assisted microwave circuit optimization method based on a high-dimensional output proxy model according to any one of claims 1 to 7 are implemented.

Images