A power amplifier multi-objective optimization method and system based on a hybrid graph neural network and a trust region Bayesian optimization algorithm, a computer device, and a storage medium

Abstract

The application relates to a power amplifier multi-target optimization method based on a hybrid graph neural network and a trust region Bayesian optimization algorithm, relates to the technical field of power amplifier circuit optimization, and comprises the following steps: determining design variables in a power amplifier matching network, taking a parameter combination of the design variables as a sampling point, and taking a parameter range of the design variables as an exploration region; obtaining a plurality of sampling points in the exploration region, inputting the sampling points into a pre-trained surrogate model for prediction to obtain performance index prediction values, and constructing a comprehensive fitness function; updating a trust region based on the comprehensive fitness function, finding new sampling points in the updated trust region and performing iterative processing, and when a preset convergence condition is met, outputting a parameter combination of the design variables corresponding to a maximum comprehensive fitness score; and the parameter optimization process in traditional power amplifier design, which needs to be manually optimized for hours, is shortened to within minutes, so that the design optimization efficiency of the power amplifier is greatly improved.

Description

Technical Field

[0001] This invention relates to the field of power amplifier circuit optimization technology, specifically to a multi-objective optimization method, system, computer device, and storage medium for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithms. Background Technology

[0002] High-efficiency power amplifiers are a core component of modern radio frequency front-end systems. The energy efficiency of a high-efficiency power amplifier directly determines the coverage, endurance, detection range of a wireless communication system, and the thermal management level of a radar system. In practical engineering design, power-added efficiency (PAE) and output power (…) are crucial parameters. Conversion gain and conversion gain are the core metrics for evaluating high-efficiency power amplifiers. However, there is a clear coupling and constraint relationship between these three metrics, making it difficult to achieve optimal performance simultaneously. In addition, high-efficiency power amplifiers exhibit significant nonlinear characteristics under large-signal excitation conditions, and the design space of key modules in high-efficiency power amplifiers is usually high-dimensional, which makes the design of high-efficiency power amplifiers very challenging.

[0003] Currently, the circuit design of high-efficiency power amplifiers in existing technologies heavily relies on engineers' experience and manual tuning. This design method requires repeated electromagnetic simulations and parameter adjustments, resulting in a long design cycle. Furthermore, manual tuning struggles to obtain globally optimal parameter combinations and is prone to getting trapped in local optima. In recent years, with the development of deep learning, some deep learning-based prediction models have emerged. However, existing prediction models often treat the circuit as a black box, directly inputting the parameter combinations of design variables into a multilayer perceptron for performance prediction. This ignores the circuit's topology, losing information about active devices, transmission lines, and the connections and structural dependencies between branches, leading to low prediction accuracy and poor generalization ability. Summary of the Invention

[0004] To address the shortcomings of existing predictive models, such as low prediction accuracy and poor generalization ability, this invention provides a multi-objective optimization method, system, computer device, and storage medium for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithms.

[0005] To achieve the above objectives, the specific solution adopted by this invention is as follows: a multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm, comprising the following steps: The design variables in the power amplifier matching network are determined, and the parameter combinations of the design variables are used as sampling points for the trust domain Bayesian optimization algorithm, while the parameter range of the design variables is used as the exploration region of the trust domain Bayesian optimization algorithm. Several sampling points are obtained within the exploration area. These sampling points are then input into a pre-trained surrogate model to predict performance metrics. Based on these performance metric predictions, a comprehensive fitness function is constructed to evaluate the overall fitness score of the sampling points. ; The trust domain is updated based on the sampling points corresponding to the maximum comprehensive fitness score. New sampling points are found in the updated trust domain and iterative processing is performed. When the preset convergence condition is met, the parameter combination of the design variables corresponding to the maximum comprehensive fitness score is output. The surrogate model is constructed based on a hybrid graph neural network. The surrogate model includes a backbone network and an output layer. The backbone network includes two parallel topology graph branches and a global feature branch. The topology graph branches are used to extract the topological structure features of the power amplifier matching network and output graph embedding features. The global feature branch is used to extract the nonlinear coupling relationship characteristics between design variables and output global embedding features. The output layer is used to concatenate and fuse the graph embedding features and the global embedding features to obtain the predicted value of the performance index.

[0006] As an optimization scheme for the aforementioned multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm, the method for extracting the topological structure features of the power amplifier matching network and outputting graph embedding features includes: The matching network of the power amplifier is converted into an undirected graph containing several nodes. The undirected graph is used as input data to the topology graph branch. The topology graph branch includes a first graph feature extraction module, a second graph feature extraction module, and a global average pooling layer arranged in sequence. Both the first graph feature extraction module and the second graph feature extraction module include a GAT convolutional layer, a batch normalization layer, and an ELU activation function arranged in sequence. The first image feature extraction module is used to process the input data to generate intermediate image features; The second image feature extraction module is used to further process the intermediate image features to generate transition image features; Global average pooling layers are used to process the transition graph features to generate graph embedding features.

[0007] As another optimization scheme for the aforementioned multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm, the method for extracting the nonlinear coupling relationship characteristics between design variables and outputting global embedded features includes: The data from the sampling points are expanded into high-dimensional features through second-order polynomial operations. Then, principal component analysis is used to reduce the dimensionality of the high-dimensional features to obtain the core features. The core features are then input into the global feature branch. The global feature branch includes a first relation feature extraction module and a second relation feature extraction module set in sequence. Both the first relation feature extraction module and the second relation feature extraction module include a fully connected layer, a batch normalization layer, a SiLU activation function, and a Dropout layer set in sequence. The core features are processed using the first relation feature extraction module to generate intermediate relation features; The second relation feature extraction module is used to process the intermediate relation features to generate global embedded features.

[0008] As another optimization scheme for the power amplifier multi-objective optimization method based on hybrid graph neural network and trust domain Bayesian optimization algorithm mentioned above, a Monte Carlo random deactivation mechanism is introduced into the output layer of the surrogate model, and the mean and variance of the performance index prediction values ​​are obtained through multiple random forward propagations.

[0009] As another optimization scheme for the aforementioned multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm, the method for finding new sampling points within the updated trust domain includes: based on the acquisition function. To determine new sampling points, the acquisition function for: , , in, Sampling points Expected improvement value This represents the overall fitness score corresponding to the globally optimal sampling point obtained in the current iteration. To obtain sampling points using a Gaussian process Predict the mean of performance metrics. To obtain sampling points using a Gaussian process The standard deviation of the predicted performance index For the standardized distance to be improved, The cumulative distribution function of the standard normal distribution. It is the probability density function of the standard normal distribution.

[0010] As another optimization scheme for the aforementioned multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm: the comprehensive fitness function for: , , in, Sampling points The overall fitness score, Sampling points The weighted performance score, Sampling points In the Normalized predicted values ​​for each performance metric These are the weighting coefficients. For the surrogate model at the sampling point The uncertainty of the forecast, This represents the uncertainty penalty coefficient.

[0011] As another optimization scheme for the power amplifier multi-objective optimization method based on hybrid graph neural networks and trust domain Bayesian optimization algorithm mentioned above, the method for training the surrogate model includes the following steps: Within the search area, construct a set of parameter combinations for the design variables using a preset step size; The sample set is obtained by removing parameter combinations that cause abnormal physical responses due to electromagnetic simulation non-convergence or severe impedance mismatch. The entire sample set is divided into training set, validation set and test set according to the preset division ratio; Initialize the proxy model, train the proxy model using the training set, divide the training process of the proxy model into two training stages, and set the corresponding training parameters for each training stage. The trained agent model is evaluated using a validation set and a test set. Training ends when the evaluation is passed.

[0012] A multi-objective optimization system for a power amplifier based on a hybrid graph neural network and a trust domain Bayesian optimization algorithm includes: The data processing module is used to iteratively optimize the parameter combination of the design variables using a pre-trained surrogate model and a trust domain Bayesian optimization algorithm to obtain the parameter combination of the design variables corresponding to the maximum comprehensive fitness score. The data visualization module is used to visually display the parameter combinations of design variables corresponding to the maximum comprehensive fitness score.

[0013] Computer equipment, including: Memory, used to store computer programs; The processor is used to read and execute computer programs to implement the aforementioned multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithms.

[0014] A storage medium for storing a computer program that, when executed, implements the aforementioned multi-objective optimization method for power amplifiers based on a hybrid graph neural network and a trust domain Bayesian optimization algorithm.

[0015] Compared with existing technologies, the present invention has the following advantages: The surrogate model of the present invention is constructed based on a hybrid graph neural network. The surrogate model includes a backbone network and an output layer. The backbone network includes two parallel topology graph branches and a global feature branch. The topology graph branch is used to extract the topological structure features of the power amplifier matching network and output graph embedding features. The global feature branch is used to extract the nonlinear coupling relationship characteristics between design variables and output global embedding features. The output layer is used to concatenate and fuse the graph embedding features and the global embedding features to obtain the predicted value of the performance index. This surrogate model can still achieve high-precision prediction under the condition of a small number of data samples. At the same time, a trust domain Bayesian optimization algorithm is adopted, combined with a comprehensive fitness function. By dynamically adjusting the trust domain, efficient and accurate parameter search is achieved in a high-dimensional design space, effectively avoiding local optima. This invention shortens the parameter optimization process, which typically requires hours of manual tuning in traditional power amplifier design, to within minutes, significantly improving the design optimization efficiency of power amplifiers. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multi-objective optimization method for power amplifiers according to the present invention; Figure 2 This is an overall framework diagram of the multi-objective optimization method for power amplifiers of the present invention; Figure 3 This is a schematic diagram of the proxy model of the present invention; Figure 4 This is a flowchart of the trust domain Bayesian optimization algorithm of this invention; Figure 5 This is a bar chart comparing the coefficients of determination of different models on the predictive performance of PAE; Figure 6 This is a bar chart comparing the root mean square error of different models in predicting PAE. Figure 7 Different models A bar chart comparing the coefficients of determination for predictive performance; Figure 8 Different models A bar chart comparing the root mean square error of prediction performance; Figure 9 This is a bar chart comparing the coefficients of determination of different models on the performance of Gain prediction; Figure 10 This is a bar chart comparing the root mean square error of different models in predicting Gain. Figure 11 These are the optimization results of different optimization methods on performance indicators; Figure 12 These are fitness convergence curves based on normalized comprehensive scores for different optimization methods; Figure 13 This is a bar chart comparing the performance indicators of the present invention before and after optimization; Figure 14 This is a graph showing the simulated performance results of the performance indicators of this invention. Figure 15 This is the circuit diagram of the power amplifier to be optimized in this invention; Figure 16 This is the circuit diagram of the optimized power amplifier of this invention. Detailed Implementation

[0017] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. Parts not described or disclosed in detail in the following embodiments of the present invention should be understood as prior art known or should be known by those skilled in the art.

[0018] like Figures 1 to 4 As shown, the present invention provides a multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm, including S1 to S3.

[0019] S1. Determine the design variables in the power amplifier matching network, and use the parameter combinations of the design variables as sampling points for the Trust-Region Bayesian Optimization (TuRBO) algorithm, and use the parameter range of the design variables as the exploration region of the Trust-Region Bayesian Optimization algorithm. In one embodiment of the present invention, the physical dimensions of several microstrip lines in the power amplifier matching network are selected as design variables.

[0020] S2. Obtain several sampling points within the exploration area, input the sampling points into a pre-trained surrogate model (RobustGNN) for prediction to obtain performance index prediction values, and construct a comprehensive fitness function based on the performance index prediction values ​​to evaluate the comprehensive fitness score of the sampling points. The method for obtaining several sampling points within the exploration area is as follows: Several sampling points are generated randomly within the exploration area. Performance indicators include power-added efficiency (PAE), output power (…). The surrogate model is constructed based on a hybrid graph neural network. The surrogate model includes a backbone network and an output layer. The backbone network consists of two parallel branches: a topology graph branch and a global feature branch. The topology graph branch extracts the topological features of the power amplifier matching network and outputs graph embedding features. The global feature branch extracts the nonlinear coupling characteristics between design variables and outputs global embedding features. The output layer concatenates and fuses the graph embedding features and the global embedding features to obtain the predicted performance index. Methods for inputting sampling points into the pre-trained surrogate model to obtain the predicted performance index include S21 to S23.

[0021] S21. The topology graph branch is used to extract the topological features of the power amplifier matching network and output graph embedding features. The topology graph branch includes a first graph feature extraction module, a second graph feature extraction module, and a global average pooling layer, arranged sequentially. Both the first and second graph feature extraction modules include a GAT convolutional layer, a batch normalization layer, and an ELU activation function, arranged sequentially. The matching network of the power amplifier is converted into an undirected graph containing several nodes. The undirected graph is used as input data into the topology graph branch. The first graph feature extraction module processes the input data to generate intermediate graph features, the second graph feature extraction module processes the intermediate graph features to generate transition graph features, and the global average pooling layer processes the transition graph features to generate graph embedding features.

[0022] More specifically, the undirected graph is used to characterize the topological dependencies between microstrip lines. To enhance the surrogate model's ability to perceive signal transmission paths and structural order, positional encoding information is introduced into the initial features of each node, enabling the model to better identify the relative positions of microstrip lines in the matching network. The hidden layer dimension of the GAT convolutional layer is set to 64, and four independent attention heads are used. This setting satisfies the need for feature subspace partitioning in multi-head attention mechanisms while avoiding the introduction of too many redundant parameters in an undirected graph with only a few nodes, thus reducing the risk of overfitting. The GAT convolutional layer adaptively learns the influence of each parameter on the target performance by assigning different attention weights to different neighboring nodes. The node feature update rule in the GAT convolutional layer is as follows: ; in, for Layer nodes Feature representation, For the first Neighbor nodes in the layer Input features, For the first The learnable weight matrix of the layer, The attention coefficient is used to measure the attention of neighboring nodes. For nodes The importance of For nodes The neighborhood set, It is a non-linear activation function.

[0023] S22. The global feature branch is used to extract the nonlinear coupling relationship characteristics between design variables and output global embedded features. The global feature branch includes a first relationship feature extraction module and a second relationship feature extraction module set sequentially. Both the first and second relationship feature extraction modules include a fully connected layer, a batch normalization layer, a SiLU activation function, and a Dropout layer set sequentially. The data of the sampling points are expanded into high-dimensional features through second-order polynomial operations. Then, principal component analysis is used to reduce the dimensionality of the high-dimensional features to obtain core features. The core features are input into the global feature branch. The first relationship feature extraction module processes the core features to generate intermediate relationship features, and the second relationship feature extraction module processes the intermediate relationship features to generate global embedded features.

[0024] More specifically, the design variables of several microstrip lines in the matching network are expanded to 27 dimensions through second-order polynomial operations to enhance the surrogate model's ability to represent nonlinear relationships. Considering that polynomial expansion introduces strong feature correlation and redundant information, principal component analysis is used to reduce the dimensionality of the high-dimensional features to obtain core features while retaining 99% of the variance information. This process can alleviate multicollinearity problems while preserving as much original information as possible, significantly improving generalization ability under limited sample conditions. The core features are input into the global feature branch. The fully connected layer in the first relationship feature extraction module contains 512 neurons, and the fully connected layer in the second relationship feature extraction module contains 256 neurons. The core features are then nonlinearly mapped through these two fully connected layers. Each fully connected layer is followed by a batch normalization layer, a SiLU activation function, and a Dropout layer to further enhance the nonlinear modeling ability of the surrogate model and, to some extent, suppress overfitting.

[0025] S23. The output layer concatenates and fuses the graph embedding features with the global embedding features to obtain the predicted performance index (PAE). The output layer includes a fusion layer with 128 neurons. The graph embedding features and global embedding features are concatenated and input into the fusion layer for fusion. Based on the fused features, the performance index PAE is output. Predicted performance metrics for Gain.

[0026] S3. Update the trust domain based on the sampling point corresponding to the maximum comprehensive fitness score, find new sampling points in the updated trust domain and perform iterative processing. When the preset convergence condition is met, output the parameter combination of the design variables corresponding to the maximum comprehensive fitness score.

[0027] More specifically, methods for finding new sampling points within the updated trust domain include: based on the sampling function. To determine new sampling points, the acquisition function for: ; ; in, Sampling points The expected improvement value is used to measure the expected magnitude of the performance improvement brought about by this sampling point. This represents the overall fitness score corresponding to the globally optimal sampling point obtained in the current iteration. To obtain sampling points using a Gaussian process Predict the mean of performance metrics. To obtain sampling points using a Gaussian process The standard deviation of the predicted performance index For the standardized distance to be improved, The cumulative distribution function of the standard normal distribution. It is the probability density function of the standard normal distribution.

[0028] In constructing a comprehensive fitness function to evaluate the comprehensive fitness score of sampling points. An uncertainty penalty term is introduced, therefore a Monte Carlo random deactivation mechanism is introduced in the output layer of step S23. The mean and variance of the predicted performance index values ​​are obtained through multiple random forward propagations, and the fitness function is then used to calculate the result. for: ; ; in, Sampling points The overall fitness score, Sampling points The weighted performance score, Sampling points In the Normalized predicted values ​​for each performance metric These are the weighting coefficients. For the surrogate model at the sampling point The uncertainty of the forecast, This is the uncertainty penalty coefficient, used to strike a balance between optimization performance and search stability.

[0029] Considering that the response relationship of PAE is influenced by multiple factors such as output power, input power, and DC power consumption, and its response is complex and difficult to optimize, the weighting coefficient of PAE is appropriately increased while ensuring that multiple objectives can be comprehensively balanced and the overall equilibrium is maintained. In one embodiment of the present invention, the weighting coefficient of PAE is... The value is 0.4. Weighting coefficients The value is 0.3, which is the weighting coefficient of Gain. The value is 0.3. Uncertainty penalty coefficient. The value of needs to be reasonably balanced. If the value is too small, the uncertainty penalty effect will be insufficient; if the value is too large, it will limit the effective exploration of potential optimal regions. Based on the above trade-offs, in one embodiment of the present invention, the uncertainty penalty system... The value is 0.5.

[0030] When a preset convergence condition is met, the optimal parameter combination is output, which corresponds to the design variable with the highest overall fitness score obtained during the entire optimization process. In one embodiment of the present invention, the preset convergence condition is reaching a preset maximum number of iterations.

[0031] The method for training the agent model includes the following steps.

[0032] Within the search area, construct a set of parameter combinations for the design variables using a preset step size. For example... Figure 15 As shown, in one embodiment of the present invention, the selected power amplifier circuit belongs to Class F, and the core active device is a Cree-CGH40010F transistor. Parameter combinations of microstrip lines TL3, TL4, TL5, TL6, TL7, and TL8 in the matching network are selected for training the surrogate model. Keeping the width of the microstrip lines constant, the lengths of the selected six microstrip lines are scanned using the multi-dimensional parameter scanning module built into the ADS platform. To cover a larger exploration area as much as possible and reduce the risk of getting trapped in local optima, the parameters of each microstrip line are set with scanning ranges and step sizes. Specifically, the scanning ranges for TL3, TL4, and TL8 are 3–15 mm with a step size of 3; the scanning range for TL5 is 2–22 mm with a step size of 5; the scanning range for TL6 is 1–13 mm with a step size of 3; and the scanning range for TL7 is 12–28 mm with a step size of 4. All simulations are performed with the transistors under Class AB bias conditions and a center frequency of 2.4 GHz, ultimately resulting in a parameter combination set containing 15,625 parameter combinations.

[0033] The sample set is obtained by removing parameter combinations that cause abnormal physical response due to non-convergence of electromagnetic simulation or severe impedance mismatch. The sample set contains 15434 parameter combinations.

[0034] The entire sample set is divided into a training set, a validation set, and a test set according to a preset ratio of 7:1.5:1.5.

[0035] Initialize the proxy model and train it using the training set. The training process is divided into two phases, with corresponding training parameters set for each phase. These two phases are called the first training phase and the second training phase.

[0036] The first training phase involves training the surrogate model on the full training set using the standard mean squared error loss function. The training parameters for this phase are: a learning rate of 0.001 and 150 training rounds.

[0037] In the second training phase, a weighted mean squared error loss function is introduced: ; in, This represents the total number of samples in a single training batch. For the performance metrics dimensions of network output, For the first The sample at the th Electromagnetic simulation values ​​for each performance index For the first The sample at the th Predicted values ​​of performance indicators for each performance metric. It is an adaptive weight mask.

[0038] The training parameters for the second training phase are: a learning rate of 0.00005 and 50 training rounds.

[0039] The trained agent model is evaluated using a validation set and a test set. Training ends when the evaluation is passed.

[0040] This invention employs root mean square error (RMSE) and coefficient of determination (...). The surrogate model's predictive performance is evaluated using RMSE. RMSE represents the square root of the average of the squared differences between the predicted and actual values. RMSE reflects the dispersion of the surrogate model's prediction error; the smaller the RMSE, the closer the surrogate model's prediction is to the true value. This indicates the surrogate model's ability to explain the variance of the data. Used to reflect how well the predicted value fits the actual value change. The closer the value is to 1, the better the surrogate model fits.

[0041] RMSE and The formula is as follows: ; ; in, The number of samples in the test set. These are the actual values ​​of the performance metrics obtained from ADS simulation. The surrogate model is used to predict the performance metrics. This represents the average of the actual values ​​of the performance indicators.

[0042] The applicant conducted multiple independent experiments and calculated the average value of the experimental results, as shown in Table 1.

[0043] Table 1. Fitting results for each performance index The results show that the proxy model of the present invention performs well in PAE, High fitting accuracy was achieved on both the training and testing sets. Furthermore, the results on the training and testing sets were very close, indicating that the surrogate model has good generalization ability and does not exhibit overfitting.

[0044] To further verify the performance of the proxy model, the following tests will be conducted on PAE, The three performance metrics, Gain, and Gain, respectively illustrate the comparison of the performance of different network structures on the training and test sets under two conditions: 100% full sample data and 3% small sample data.

[0045] like Figures 5 to 10 As shown, under the condition of 100% full sample data, the pure MLP model performs well on PAE, Gain achieved high predictions. Compared to a lower RMSE, the surrogate model of this invention further improves prediction performance and demonstrates better fit across the three performance metrics. Furthermore, the performance difference between the training and test sets of the surrogate model is small, indicating good robustness under 100% full sample data conditions. In contrast, the pure GNN model performs relatively weakly, suggesting that relying solely on graph topological features is insufficient to fully describe the complex relationship between design variables and performance metrics; global feature branches are necessary in the surrogate model.

[0046] With only 3% of the sample data, the pure MLP model is effective for PAE. While Gain's predictions still achieve high accuracy on the training set, their performance drops significantly on the test set. The surrogate model of this invention maintains relatively stable prediction capabilities even with only 3% of the samples, and performs well in PAE and other tests. More reliable test performance was achieved on both the Gain and Gain tasks.

[0047] To verify the superiority of the trust domain Bayesian optimization algorithm, a comparative experiment was conducted with several other optimization algorithms, including the trust domain Bayesian optimization algorithm with uncertainty penalty term removal, Latin hypercube sampling (LHS), differential evolution (DE), and the ADS built-in optimizer. The comparison results are as follows: Figure 11 As shown, LHS struggles to effectively approximate the high-performance region under a limited sampling budget; DE, while finding some relatively good solutions, suffers from large fluctuations and significant oscillations during convergence; ADS's built-in optimizer is prone to getting trapped in local optima, exhibiting premature convergence; and while the trust-domain Bayesian optimization algorithm, which removes uncertainty penalties, can enter the higher-performance region, its solution set is scattered, with poor clustering near the optimal solution. In contrast, the full trust-domain Bayesian optimization algorithm employed in this invention achieves a better balance between global exploration and local exploitation, more stably approximating the high-performance solution region and obtaining superior overall performance results.

[0048] To further evaluate the dynamic search efficiency of different algorithms under the same evaluation budget, such as Figure 12 As shown, Figure 12 Fitness convergence curves based on the normalized comprehensive score for different optimization methods are presented. This comprehensive score integrates PAE, The three performance metrics—Gain, Convergence, and Gain—were used as benchmarks, with the built-in optimizer in ADS software as the baseline. Results show that the Trust Domain Bayesian optimization algorithm employed in this invention exhibits faster performance improvement in the early stages of optimization, gradually stabilizing after approximately 30 iterations. In contrast, while the Trust Domain Bayesian optimization algorithm with the uncertainty penalty term removed still outperforms the DE and LHS algorithms, its convergence speed and final performance both decrease. Experiments demonstrate that the introduction of the uncertainty penalty term not only improves the final optimization result but also effectively enhances the stability of the search process.

[0049] In summary, compared with traditional optimization algorithms that treat the matching network circuit as a pure black box model lacking physical topology, the method proposed in this invention incorporates circuit topology features, exhibiting stronger targeting and physical rationality in multi-objective optimization in complex parameter spaces. Meanwhile, the uncertainty penalty term plays a crucial role in the trust domain Bayesian optimization algorithm, both suppressing the optimizer's overexploration in low-confidence regions of the surrogate model and guiding the search to focus on more reliable candidate regions, effectively improving search stability, convergence efficiency, and final optimization performance.

[0050] To verify the feasibility of the optimization results in actual circuits, the optimal parameter combination obtained by the method of this invention was imported into ADS for simulation. The parameter comparison before and after optimization is shown in Table 2.

[0051] Table 2 Comparison results before and after optimization Performance metrics before and after optimization, such as Figure 13 As shown, the optimized PAE, The three performance indicators, Gain, and Gain, were improved by 1.5%, 3.4%, and 2.6% respectively, indicating that the optimization method of this invention can achieve a relatively effective comprehensive improvement among multiple performance indicators.

[0052] like Figure 14 As shown, Figure 14 This is a simulation performance result curve of the performance indicators of this invention, showing the Gain and PAE as a function of the invention. The PAE curve shows a smooth upward trend with increasing output power, reaching a peak at a specific operating point. The Gain curve also exhibits a variation pattern consistent with the device's operating characteristics, indicating that the optimized parameter combination not only performs excellently under the surrogate model evaluation but also maintains a reasonable physical response in ADS simulation. The final optimized parameter combination in this embodiment is as follows: Figure 16 As shown.

[0053] This invention leverages the computational power of a proxy model to shorten the parameter optimization process, which typically requires hours of manual tuning in traditional power amplifier design, to within minutes, significantly improving the design optimization efficiency of power amplifiers.

[0054] This invention provides a multi-objective optimization system for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm, including a data processing module and a data display module.

[0055] The data processing module is used to iteratively optimize the parameter combination of the design variables using a pre-trained surrogate model and a trust domain Bayesian optimization algorithm to obtain the parameter combination of the design variables corresponding to the maximum comprehensive fitness score. The data visualization module is used to visually represent the parameter combinations of design variables corresponding to the maximum overall fitness score. The present invention also provides a computer device, including a memory and a processor.

[0056] Memory is used to store computer programs.

[0057] A processor is used to read and execute the computer program to implement the above-described multi-objective optimization method for power amplifiers based on a hybrid graph neural network and a trust domain Bayesian optimization algorithm.

[0058] Finally, the present invention provides a storage medium for storing a computer program that, when executed, implements the aforementioned multi-objective optimization method for a power amplifier based on a hybrid graph neural network and a trust domain Bayesian optimization algorithm.

[0059] Memory, as a carrier of resources, can be read-only memory, random access memory, disk, or optical disk, etc. The resources stored on it can include operating systems, computer programs, etc., and the storage method can be temporary or permanent storage. The processor can be a general-purpose processor product based on architectures such as x86, IA64, RISC, MIPS, and ARM.

[0060] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm, characterized in that: Includes the following steps: The design variables in the power amplifier matching network are determined, and the parameter combinations of the design variables are used as sampling points for the trust domain Bayesian optimization algorithm, while the parameter range of the design variables is used as the exploration region of the trust domain Bayesian optimization algorithm. Several sampling points are obtained within the exploration area. These sampling points are then input into a pre-trained surrogate model to predict performance metrics. Based on these performance metric predictions, a comprehensive fitness function is constructed to evaluate the overall fitness score of the sampling points. ; The trust domain is updated based on the sampling points corresponding to the maximum comprehensive fitness score. New sampling points are found in the updated trust domain and iterative processing is performed. When the preset convergence condition is met, the parameter combination of the design variables corresponding to the maximum comprehensive fitness score is output. The surrogate model is constructed based on a hybrid graph neural network. The surrogate model includes a backbone network and an output layer. The backbone network includes two parallel topology graph branches and a global feature branch. The topology graph branches are used to extract the topological structure features of the power amplifier matching network and output graph embedding features. The global feature branch is used to extract the nonlinear coupling relationship characteristics between design variables and output global embedding features. The output layer is used to concatenate and fuse the graph embedding features and the global embedding features to obtain the predicted value of the performance index.

2. The multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm according to claim 1, characterized in that: The method for extracting the topological features of the power amplifier matching network and outputting graph embedding features using the topological graph branch includes: The matching network of the power amplifier is converted into an undirected graph containing several nodes. The undirected graph is used as input data to the topology graph branch. The topology graph branch includes a first graph feature extraction module, a second graph feature extraction module, and a global average pooling layer arranged in sequence. Both the first graph feature extraction module and the second graph feature extraction module include a GAT convolutional layer, a batch normalization layer, and an ELU activation function arranged in sequence. The first image feature extraction module is used to process the input data to generate intermediate image features; The second image feature extraction module is used to further process the intermediate image features to generate transition image features; Global average pooling layers are used to process the transition graph features to generate graph embedding features.

3. The multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm according to claim 1, characterized in that: The method for extracting the nonlinear coupling relationship characteristics between design variables and outputting global embedded features by the global feature branch includes: The data from the sampling points are expanded into high-dimensional features through second-order polynomial operations. Then, principal component analysis is used to reduce the dimensionality of the high-dimensional features to obtain the core features. The core features are then input into the global feature branch. The global feature branch includes a first relation feature extraction module and a second relation feature extraction module set in sequence. Both the first relation feature extraction module and the second relation feature extraction module include a fully connected layer, a batch normalization layer, a SiLU activation function, and a Dropout layer set in sequence. The core features are processed using the first relation feature extraction module to generate intermediate relation features; The second relation feature extraction module is used to process the intermediate relation features to generate global embedded features.

4. The multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm according to claim 1, characterized in that: A Monte Carlo random deactivation mechanism is introduced into the output layer of the surrogate model, and the mean and variance of the predicted performance index values ​​are obtained through multiple random forward propagations.

5. The multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm according to claim 4, characterized in that: The method for finding new sampling points within the updated trust domain includes: based on a sampling function. To determine new sampling points, the acquisition function for: ， ， in, Sampling points Expected improvement value This represents the overall fitness score corresponding to the globally optimal sampling point obtained in the current iteration. To obtain sampling points using a Gaussian process Predict the mean of performance metrics. To obtain sampling points using a Gaussian process The standard deviation of the predicted performance index For the standardized distance to be improved, The cumulative distribution function of the standard normal distribution. It is the probability density function of the standard normal distribution.

6. The multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm according to claim 1, characterized in that: The comprehensive fitness function for: ， ， in, Sampling points The overall fitness score, Sampling points The weighted performance score, Sampling points In the Normalized predicted values ​​for each performance metric These are the weighting coefficients. For the surrogate model at the sampling point The uncertainty of the forecast, This represents the uncertainty penalty coefficient.

7. The multi-objective optimization method for power amplifiers based on hybrid graph neural networks and trust domain Bayesian optimization algorithm according to claim 1, characterized in that: The method for training a proxy model includes the following steps: Within the search area, construct a set of parameter combinations for the design variables using a preset step size; The sample set is obtained by removing parameter combinations that cause abnormal physical responses due to electromagnetic simulation non-convergence or severe impedance mismatch. The entire sample set is divided into training set, validation set and test set according to the preset division ratio; Initialize the proxy model, train the proxy model using the training set, divide the training process of the proxy model into two training stages, and set the corresponding training parameters for each training stage. The trained agent model is evaluated using a validation set and a test set. Training ends when the evaluation is passed.

8. A multi-objective optimization system for a power amplifier based on a hybrid graph neural network and a trust domain Bayesian optimization algorithm, characterized in that: include: The data processing module is used to iteratively optimize the parameter combination of the design variables using a pre-trained surrogate model and a trust domain Bayesian optimization algorithm to obtain the parameter combination of the design variables corresponding to the maximum comprehensive fitness score. The data visualization module is used to visually display the parameter combinations of design variables corresponding to the maximum comprehensive fitness score.

9. A computer device, characterized in that: include: Memory, used to store computer programs; A processor for reading and executing computer programs to implement a multi-objective optimization method for power amplifiers based on a hybrid graph neural network and a trust domain Bayesian optimization algorithm as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: Used to store a computer program, which, when executed, implements the multi-objective optimization method for a power amplifier based on a hybrid graph neural network and a trust domain Bayesian optimization algorithm as described in any one of claims 1 to 7.

Images