A neural network assisted power amplifier automated optimization design method

Abstract

The present application belongs to the cross technical field of artificial intelligence, automatic design and power amplifier, and specifically provides a neural network assisted power amplifier automatic optimization design method to solve the problems of high calculation cost, many design iteration times, great dependence on engineer experience and performance limitation of traditional optimization method under high-dimensional and multi-objective constraints existing in the prior art power amplifier optimization design method. The present application constructs an agent prediction model based on a DNN network, substitutes a Gaussian process with an agent prediction neural network to model a target function, and combines incremental learning and efficient candidate generation to realize fast convergence and low-cost calculation of a complex, high-dimensional and multi-objective power amplifier optimization problem. Compared with Bayesian optimization, the present application has faster convergence speed, lower calculation cost and more flexible exploration ability in the iteration process, and shows better adaptability and robustness in complex high-dimensional problems and multi-objective constraint optimization.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary technical field of artificial intelligence, automated design, and power amplifiers, and specifically provides a neural network-assisted automated optimization design method for power amplifiers. Background Technology

[0002] Power amplifiers (PAs) play a crucial role in wireless communication systems, affecting not only signal coverage and transmission distance but also directly impacting system energy efficiency, linearity, and signal quality. In modern wireless communication, such as cellular networks, satellite communications, millimeter-wave radar, and 5G / 6G systems, high-efficiency, wideband, and linear power amplifiers are core components ensuring high data rates and stable communication performance.

[0003] However, designing a power amplifier is a complex process. First, it requires selecting appropriate transistor technology with sufficient margin in output power and efficiency, and determining the gate lengths and exponents of the driver and final stages. Second, due to the inherent nonlinearity of power amplifiers, engineers must carefully select input and output matching networks for different stages at different power levels by performing load-pull / source-pull simulations. For the selected transistors, full electromagnetic (EM) simulations of the input / inter- / output matching networks are required to characterize the power amplifier's performance, such as operating bandwidth, output power, efficiency, gain, and linearity. Finally, updating design parameters and weighing various power amplifier (PA) requirements typically requires multiple iterations. This process is extremely time-consuming and highly dependent on the designer's experience.

[0004] In recent years, many optimization methods have been proposed for the parameter design of power amplifiers (PAs), such as the simulated annealing particle swarm optimization (SA-PSO) method. However, SA-PSO is a heuristic model-free method, and its computational cost increases rapidly in complex cases, especially when considering electromagnetic simulations. To address the problem of long computation time, patents CN209085657U and CN114297925B proposed the Bayesian optimization (BO) framework, which was later extended to the multi-objective scenario of broadband high-efficiency power amplifiers. The BO framework uses a Gaussian process (GP) model to predict the optimal new solution in each iteration, thereby reducing the number of necessary electromagnetic simulations. However, due to the lack of search operators and the limitations of kernel functions, this method can usually only handle low-dimensional parameters, and the initialization of parameters directly affects the overall optimization result. Summary of the Invention

[0005] The purpose of this invention is to provide a neural network-assisted automated optimization design method for power amplifiers, which solves the problems of high computational cost, numerous design iterations, heavy reliance on engineer experience, and limited performance of traditional optimization methods under high-dimensional and multi-objective constraints in existing power amplifier optimization design methods.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A neural network-assisted automated optimization design method for power amplifiers, characterized by comprising the following steps:

[0008] Step 1: Obtain the circuit topology of the power amplifier to be optimized, and set the optimization parameters and expected performance indicators;

[0009] Step 2: Construct a neural network-based agent prediction model, taking the optimization parameters as input and outputting a performance index prediction matrix;

[0010] Step 3: Sample the optimized parameters within the preset value range using Latin hypercube sampling to generate Q sets of simulation samples. Pre-train the surrogate prediction model based on the initially sampled Q sets of simulation samples, and combine all simulation samples to form a historical dataset.

[0011] Step 4: With the desired performance index as the target, iteratively optimize the pre-trained agent prediction model to obtain the optimized parameters.

[0012] Furthermore, in step 1, the optimization parameters include: the length of the microstrip transmission line and the width of the capacitor.

[0013] Furthermore, in step 1, the desired performance indicators include: S11, S21, output power and power-added efficiency.

[0014] Furthermore, in step 2, the surrogate prediction model uses a DNN network, consisting of an input layer, three hidden layers, and an output layer; the number of neurons in the input layer is M, where M is the dimension of the optimization parameters; the number of neurons in the output layer is N×F, where N is the dimension of the performance index and F is the number of operating frequencies; the hidden layers use ReLU as the activation function, and the output layer uses Sigmoid as the activation function.

[0015] Furthermore, the DNN network takes the optimization parameters as input, performs maximum and minimum value normalization on each optimization parameter, and forms an input vector according to the index of the normalization result. After passing through 3 hidden layers, the output layer outputs the performance index prediction matrix.

[0016] Furthermore, in step 3, during the training process, the root mean square error is used as the loss.

[0017] Further, the specific process of step 4 is as follows:

[0018] Taking the expected performance index as the objective and the root mean square error as the objective function; selecting P samples with the minimum objective function value from the historical dataset, where P<Q, performing random variation on them to generate P candidate solutions, obtaining the performance index prediction matrix of each candidate solution through the surrogate prediction model, and calculating the predicted objective function values respectively; selecting the candidate solution with the minimum predicted objective function value to perform EM simulation, and calculating to obtain the simulated objective function value. If the simulated objective function value is less than the preset threshold, outputting the candidate solution as the optimization result; otherwise, putting the candidate solution into the historical dataset as a new sample, performing incremental training on the model through the new sample and then carrying out the next iteration until the expected performance index is satisfied.

[0019] Based on the above technical solutions, the beneficial effects of the present invention lie in that:

[0020] The present invention provides a neural network-aided automatic optimization design method for power amplifiers, which constructs a surrogate prediction model based on a DNN network to predict the performance indexes of the power amplifier; the surrogate prediction neural network is used to replace the Gaussian process for objective function modeling, and combined with incremental learning and efficient candidate generation, fast convergence and low-cost calculation for complex, high-dimensional, multi-objective power amplifier optimization problems are realized; compared with Bayesian optimization, the present invention has faster convergence speed, lower calculation cost, and more flexible exploration capability in the iteration process, and shows better adaptability and robustness in complex high-dimensional problems and multi-objective constrained optimization. Brief Description of the Drawings

[0021] Figure 1 is a schematic flow diagram of the neural network-aided automatic optimization design method for power amplifiers according to the present invention.

[0022] Figure 2 is a layout diagram of the power amplifier (PA) in the embodiment of the present invention.

[0023] Figure 3 is a diagram of S11 and S21 simulation results of the power amplifier (PA) in the embodiment of the present invention.

[0024] Figure 4 is a diagram of output power simulation results of the power amplifier (PA) in the embodiment of the present invention.

[0025] Figure 5 is a diagram of power added efficiency simulation results of the power amplifier (PA) in the embodiment of the present invention.

[0026] Figure 6 is a comparison diagram of convergence speed between the embodiment of the present invention and the comparative example. Detailed Description of the Embodiments

[0027] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0028] This embodiment provides a neural network-assisted automated optimization design method for power amplifiers, the process of which is as follows: Figure 1 As shown, the main steps include: initial sampling and pre-training of the surrogate model, and optimization of neural network auxiliary parameters. The specific steps are as follows:

[0029] Step 1: Obtain the circuit topology of the power amplifier to be optimized and set the optimization parameters;

[0030] Step 2: Construct a neural network-based agent prediction model, using a DNN network consisting of an input layer, three hidden layers, and an output layer;

[0031] like Figure 2 As shown, in this embodiment, the power amplifier includes 11 optimization parameters and sets 4 performance indicators: S11, S21, output power, and power-added efficiency. Therefore, there are 11 neurons in the input layer, 64 neurons in each of the three hidden layers, and 36 neurons in the output layer (4×9=36). Each prediction indicator corresponds to 9 frequency points. The hidden layer uses ReLU as the activation function, and the output layer uses Sigmoid as the activation function.

[0032] The DNN network takes the optimization parameters as input, performs maximum and minimum value normalization on each optimization parameter, and forms an input vector according to the index of the normalization result. After passing through 3 hidden layers, the output layer outputs the performance index prediction matrix.

[0033] The optimization parameters include: the length of the microstrip transmission line and the width of the capacitor;

[0034] Step 3: Use a Python script to call the internal API of ADS to realize fully automatic modeling, simulation and post-processing of the power amplifier to be optimized; generate 10 sets of simulation samples within a preset value range through Latin hypercube sampling; pre-train the surrogate prediction model based on the 10 sets of initial simulation samples, and combine all simulation samples to form a historical dataset; during the pre-training process, the root mean square error is used as the loss.

[0035] Step 4: Iteratively optimize the pre-trained agent prediction model; the specific process is as follows:

[0036] With the desired performance index as the objective and the root mean square error (RMSE) as the objective function, the eight samples with the smallest objective function values ​​are selected from the historical dataset and randomly mutated to generate eight candidate solutions. The performance index prediction matrix of each candidate solution is obtained through a surrogate prediction model, and the predicted objective function value is calculated for each solution. The candidate solution with the smallest predicted objective function value is selected for EM simulation, and the simulated objective function value is calculated. If the simulated objective function value is less than a preset threshold, the candidate solution is output as the optimization result; otherwise, the candidate solution is added to the historical dataset as a new sample, and the model is incrementally trained using the new sample for the next iteration until the desired performance index is met.

[0037] The beneficial effects of the present invention will be explained in detail below with reference to simulation tests.

[0038] In this embodiment, the layout structure of the power amplifier (PA) is as follows: Figure 2 As shown, the input matching circuit includes microstrip line L1, capacitor C1, microstrip line L2, capacitor C2, microstrip line L4, and microstrip line L5; the output matching circuit includes microstrip line L6, microstrip line L7, capacitor C3, and microstrip line L8. All microstrip lines are equivalent to inductors, thus forming the input and output matching circuits. It should be noted that in the optimization design of the power amplifier, for microstrip lines, the corresponding optimization parameter in the layout is the microstrip line length, which represents the inductance value of the equivalent inductance; for capacitors, the corresponding optimization parameter in the layout is the capacitor width, which represents the capacitance value.

[0039] In this embodiment, the power amplifier (PA) is based on a user-defined operating frequency band of 28 GHz to 30 GHz, and is required to have an output power ≥28 dBm, a gain ≥10 dB, and a power-added efficiency (PAE) ≥30%. Based on the above-mentioned automated optimization design method, the optimization results are shown in Table 1.

[0040] Table 1. Optimization results of the power amplifier (unit: μm)

[0041]

[0042] Under neural network-assisted multi-objective iterative optimization, the power amplifier automatically optimized by this invention has been verified by simulation, and the results are as follows: Figures 3-5 As shown in the figure, within the frequency range of 27.5~30.5 GHz, the input matching (S11) of the PA is less than -10 dB, and the bandwidth reaches 3 GHz; within this bandwidth, the gain is greater than 10 dB; when the input power is greater than 20 dBm, the output power is ≥28 dBm, and the power-added efficiency (PAE) is ≥30%, meeting all the set performance indicators and fully verifying the effectiveness and superiority of the optimization design method in the automated design of RF power amplifiers.

[0043] Furthermore, using traditional heuristic algorithms (genetic algorithms) and Bayesian optimization algorithms as comparative examples, the iterative optimization process of this invention is compared with that of the comparative examples, and the results are as follows: Figure 6 As shown, the present invention exhibits superior performance in terms of convergence speed.

[0044] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.

Claims

1. A neural network-assisted automated optimization design method for power amplifiers, characterized in that, comprising the following steps: Step 1: Acquire the circuit topology of the power amplifier to be optimized, and set optimization parameters and expected performance metrics; Step 2: Construct a neural network-based surrogate prediction model, which takes the optimization parameters as input and outputs a performance metric prediction matrix; Step 3: Sample the optimization parameters within a preset value range through Latin hypercube sampling to generate Q sets of simulation samples, pre-train the surrogate prediction model based on the initially sampled Q sets of simulation samples, and all simulation samples collectively form a historical dataset; Step 4: Taking the expected performance metrics as the goal, perform iterative optimization based on the pre-trained surrogate prediction model to obtain optimized results of the optimization parameters.

2. The neural network-assisted automated optimization design method for power amplifiers according to claim 1, characterized in that, In Step 1, the optimization parameters include: the length of the microstrip transmission line and the width of the capacitor.

3. The neural network-assisted automated optimization design method for power amplifiers according to claim 1, characterized in that, In Step 1, the expected performance metrics include: S11, S21, output power and power added efficiency.

4. The neural network-assisted automated optimization design method for power amplifiers according to claim 1, characterized in that, In Step 2, the surrogate prediction model adopts a DNN network, which consists of an input layer, 3 hidden layers and an output layer; the number of neurons in the input layer is M, where M is the dimension of the optimization parameters; the number of neurons in the output layer is N×F, where N is the dimension of the performance metrics and F is the number of operating frequency points; the hidden layers use ReLU as the activation function, and the output layer uses Sigmoid as the activation function.

5. The neural network-assisted automated optimization design method for power amplifiers according to claim 4, characterized in that, The DNN network takes the optimization parameters as input, performs maximum and minimum value normalization on each optimization parameter separately, forms an input vector from the normalized results according to indexes, and outputs the performance metric prediction matrix via the output layer after passing through the 3 hidden layers.

6. The neural network-assisted automated optimization design method for power amplifiers according to claim 1, characterized in that, In Step 3, during the training process, root mean square error is used as the loss.

7. The neural network-assisted automated optimization design method for power amplifiers according to claim 1, characterized in that, The specific process of Step 4 is as follows: Taking the expected performance metrics as the goal and the root mean square error as the objective function; selecting P samples with the smallest objective function values from the historical dataset, where P<Q, performing random variation on them to generate P candidate solutions, obtaining the performance metric prediction matrix of each candidate solution through the surrogate prediction model, and calculating the predicted objective function values respectively; Select the candidate solution with the smallest predicted objective function value to perform EM simulation, calculate to obtain a simulated objective function value; if the simulated objective function value is less than a preset threshold, output the candidate solution as the optimization result; otherwise, add the candidate solution to the historical dataset as a new sample, perform incremental training on the model with the new sample, and then carry out the next iteration until the expected performance metrics are satisfied.

Images