Wide-area intelligent design method and system for ultra-wideband low-noise amplifier

Abstract

The invention discloses a wide-area intelligent design method and system for an ultra-wideband low-noise amplifier, and belongs to the field of radio frequency integrated circuit design, and the method comprises the following steps: carrying out the pixelated decomposition of a topological structure of an LNA matching network, and obtaining a pixelated matrix; processing the pixelated matrix by using a pre-trained deep learning model to obtain a prediction S parameter corresponding to the topological structure; a target S parameter is preset, a global optimization algorithm is adopted to carry out iterative optimization on the pixelated matrix, the error of the prediction S parameter and the target S parameter between a real part and an imaginary part is minimized in each iteration process, and an optimal pixelated matrix is obtained after iteration is completed; and carrying out reverse synthesis on the optimal pixelated matrix to obtain an optimal topological structure of the LNA matching network. According to the method, the problem that the global optimal matching network structure is difficult to search through performance target constraint due to manual parameter adjustment in the existing design method can be solved.

Description

Technical Field

[0001] This invention belongs to the field of radio frequency integrated circuit design, specifically relating to a wide-area intelligent design method and system for ultra-wideband low-noise amplifiers. Background Technology

[0002] In the field of radio frequency integrated circuit design, the low-noise amplifier (LNA) is a key component, and its performance directly affects the signal quality and efficiency of the entire communication system. The design of the matching network is a core aspect of the LNA. Traditional methods generally rely on the designer's professional experience and fundamental knowledge of electromagnetic theory, manually deriving circuit formulas and combining them with electronic design automation (EDA) tools for parameter adjustment and simulation optimization. A typical design flow usually employs a fixed topology and performs iterative simulations near the center frequency to gradually approximate the expected noise figure, gain, and reflection coefficient. This process is considered the mainstream solution in the industry, aiming to achieve the basic functional requirements of the LNA and provide a foundation for subsequent circuit integration.

[0003] However, the existing problems are that the above methods exhibit significant limitations when dealing with ultra-wideband and multi-objective performance optimization applications. Because they rely on manual experience to select the topology and set initial parameters, the design process requires extensive repetitive simulation trials, resulting in low overall efficiency and a long cycle time. Simultaneously, optimization strategies based on the center frequency struggle to effectively account for the frequency variation characteristics of component parameters across a wide bandwidth, often leading to a lack of coordinated optimization between noise matching and power matching within the bandwidth, thus limiting the overall performance improvement of the LNA. Furthermore, the multi-component structure introduced to extend the operating bandwidth not only increases the insertion loss and physical area of ​​the matching network but also makes gain flatness control more complex, ultimately affecting the stability and applicability of the circuit under wideband conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a wide-area intelligent design method and system for ultra-wideband low-noise amplifiers, so as to solve the problem that existing design methods, which manually adjust parameters, are unable to find the globally optimal matching network structure through performance target constraints.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a wide-area intelligent design method for ultra-wideband low-noise amplifiers includes the following steps: The topology of the LNA matching network is decomposed into pixels to obtain a pixelated matrix; The pixelated matrix is ​​processed using a pre-trained deep learning model to obtain the predicted S-parameters corresponding to the topological structure. A target S-parameter is preset, and a global optimization algorithm is used to iteratively optimize the pixelated matrix. In each iteration, the error between the predicted S-parameter and the target S-parameter in the real and imaginary parts is minimized. After the iteration is completed, the optimal pixelated matrix is ​​obtained. By inversely synthesizing the optimal pixelation matrix, the optimal topology of the LNA matching network is obtained.

[0006] In some implementations, the pre-trained deep learning model is a convolutional neural network, which includes: an input layer, a plurality of convolutional layer modules, a plurality of fully connected layer modules, and an output layer connected in sequence. The convolutional layer module includes a convolutional layer, a normalization layer, and an activation layer connected in sequence, and the fully connected layer module includes a fully connected layer, a normalization layer, an activation layer, and a random deactivation layer connected in sequence.

[0007] In some implementations, the training process of the pre-trained deep learning model includes the following steps: The pixelated matrix is ​​filtered using an electrical connectivity judgment algorithm, and electromagnetic simulation is performed to obtain an S-parameter dataset, which includes an S11 parameter dataset, an S21 parameter dataset, an S12 parameter dataset, and an S22 parameter dataset. The convolutional neural network is trained based on the S11 parameter dataset to obtain a pre-trained model; The parameters of the convolutional layers of the pre-trained model are frozen using transfer learning, and the parameters of the fully connected layers of the pre-trained model are fine-tuned based on the S21 parameter dataset, S12 parameter dataset, and S22 parameter dataset, respectively. The pre-trained model is then trained to obtain a pre-trained deep learning model.

[0008] In some implementations, the preset target S-parameter specifically includes: For the input of the LNA matching network, the first target S-parameters corresponding to the best noise matching are preset; Alternatively, for the output of the LNA matching network, preset the second target S-parameter corresponding to the maximum power matching.

[0009] In some implementations, when the first target S-parameter corresponding to the optimal noise match is preset, the pixelation matrix is ​​iteratively optimized using a global optimization algorithm. During each iteration, the error between the predicted S-parameter and the target S-parameter in the real and imaginary parts is minimized. After the iteration is completed, the optimal pixelation matrix is ​​obtained. Specifically, this includes: Calculate the mean absolute error between the real and imaginary parts of the first target S-parameters and the predicted S-parameters, and use it as the input matching cost function; The input matching cost function is minimized, and the pixelated matrix is ​​iteratively optimized using a genetic algorithm until a preset maximum number of iterations is reached, resulting in the best noise-matching two-dimensional matrix, which is then used as the optimal pixelated matrix.

[0010] In some implementations, when the second target S-parameter corresponding to the maximum power matching is preset, the pixelation matrix is ​​iteratively optimized using a global optimization algorithm. During each iteration, the error between the real and imaginary parts of the predicted S-parameter and the target S-parameter is minimized. After the iteration is completed, the optimal pixelation matrix is ​​obtained. Specifically, this includes: The mean absolute error of the real and imaginary parts of the second target S-parameters and the predicted S-parameters is calculated and used as the output matching cost function; The output matching cost function is minimized, and the pixelation matrix is ​​iteratively optimized using a genetic algorithm until a preset maximum number of iterations is reached, resulting in a two-dimensional matrix with maximum power matching, which is taken as the optimal pixelation matrix.

[0011] Secondly, a wide-area intelligent design system for an ultra-wideband low-noise amplifier includes: The topology representation module is used to decompose the topology of the LNA matching network into a pixelated matrix. The performance prediction module is used to process the pixelated matrix using a pre-trained deep learning model to obtain the predicted S-parameters corresponding to the topology. The intelligent optimization module is used to preset the target S-parameters and use a global optimization algorithm to iteratively optimize the pixelated matrix. In each iteration, the error between the predicted S-parameters and the target S-parameters in the real and imaginary parts is minimized, and the optimal pixelated matrix is ​​obtained after the iteration is completed. The reverse synthesis module is used to reverse synthesize the optimal pixelated matrix to obtain the optimal topology of the LNA matching network.

[0012] Thirdly, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable in the processor, wherein the processor, when executing the computer program, implements the steps of the ultra-wideband low-noise amplifier wide-area intelligent design method.

[0013] Fourthly, a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the ultra-wideband low-noise amplifier wide-area intelligent design method.

[0014] Fifthly, a computer program product comprising a computer program that, when executed by a processor, implements the steps of the ultra-wideband low-noise amplifier wide-area intelligent design method.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a wide-area intelligent design method for ultra-wideband low-noise amplifiers (UNAs). First, by pixelating the LNA matching network, a search can be performed in an infinite topology space, breaking the limitations of traditional fixed topologies. Second, a pre-trained deep learning model is used to process the pixelated matrix to obtain the predicted S-parameters corresponding to the topology. This replaces the time-consuming electromagnetic simulations and manual trial-and-error in traditional design, shortening the design cycle, greatly improving design efficiency, and reducing over-reliance on the designer's personal experience. Finally, by presetting target S-parameters, a global optimization algorithm automatically iterates to find the optimal solution, ensuring that the synthesized non-classical topology can simultaneously match the performance requirements reflected by the target S-parameters across a wide frequency band.

[0016] Furthermore, the convolutional neural network includes: an input layer, several convolutional layer modules, several fully connected layer modules, and an output layer connected in sequence. Its structure is optimized for pixelated matrix input. The convolutional layer modules can extract key local electromagnetic features in the topology corresponding to the pixelated matrix step by step, thereby accurately capturing the nonlinear mapping relationship between the topological structure and complex S-parameters. At the same time, several fully connected layers are also set, where the normalization layer ensures the stability and convergence speed of the training process, and the random deactivation layer effectively prevents the model from overfitting to the training data, enhances the model's generalization prediction ability for new topological structures, and ensures reliability and robustness in practical applications.

[0017] Furthermore, by using an electrical connectivity judgment algorithm to filter the pixelated matrix, a large number of invalid electrical connections in the topology can be pre-screened, improving the efficiency and quality of dataset construction. After training the convolutional neural network with the S11 parameter dataset, other S parameters are trained through transfer learning. This fully utilizes the sensitivity of the S11 parameters to topological changes to allow the model to learn general features first. By freezing the convolutional layers and only fine-tuning the fully connected layers, the model can converge quickly with extremely low data cost and computational overhead when training other S parameters, significantly shortening the training time and reducing the validation error, thus achieving efficient and high-precision collaborative modeling of multiple S parameters.

[0018] Furthermore, for the input end of the LNA matching network, a first target S-parameter corresponding to the optimal noise matching is preset; or for the output end of the LNA matching network, a second target S-parameter corresponding to the maximum power matching is preset; this can meet the drastically different performance requirements of the input and output ends in the LNA design. By flexibly configuring the target S-parameters, the optimal matching network for the input and output ports can be synthesized in a targeted manner.

[0019] Furthermore, using the average absolute error of the real and imaginary parts of the predicted S-parameters and the target S-parameters as the cost function can more robustly measure the deviation between the prediction performance and the optimal noise matching target, and is insensitive to outliers. This guides the optimization process to more stably and directly approximate the optimal noise impedance point of the transistor. Combined with the global search capability of the genetic algorithm, it can automatically discover and evolve non-classical matching topologies that can achieve ultra-low noise figures in a wide pixelated matrix space, ultimately ensuring that the generated input matching network achieves optimal noise performance over a wide bandwidth. This is something that traditional local optimization methods based on formulas and experience cannot achieve.

[0020] Furthermore, the mean absolute error of the real and imaginary parts of the second target S-parameter and the predicted S-parameter is calculated as the output matching cost function, which can clearly guide to maximum power matching. By minimizing the error between the real and imaginary parts of the predicted S-parameter and the target S-parameter, the gain and output power capability of the LNA are directly improved. This results in an output matching network with high gain, good flatness and low return loss over a wide bandwidth. Compared with traditional methods that introduce too many components to cover the broadband, leading to increased insertion loss, the topology synthesized in this invention achieves a better balance between performance and structural simplicity. Attached Figure Description

[0021] Figure 1 A flowchart illustrating the wide-area intelligent design method for ultra-wideband low-noise amplifiers provided in this embodiment of the invention; Figure 2 This is a structural diagram of the ultra-wideband low-noise amplifier wide-area intelligent design system provided in an embodiment of the present invention; Figure 3 This is a technical roadmap for the wide-area intelligent design method of ultra-wideband low-noise amplifiers provided in the embodiments of the present invention. Figure 4 A structural diagram of a convolutional neural network provided in an embodiment of the present invention; Figure 5 A flowchart of the genetic algorithm provided in an embodiment of the present invention; Figure 6 The circuit schematic diagram of the LNA matching network provided in the embodiment of the present invention; Figure 7 The following is a simulation result diagram of the low-noise amplifier provided in the embodiment of the present invention, wherein (a) is the characteristic curve of the forward transmission gain of the LNA as a function of frequency, (b) is the characteristic curve of the noise figure of the LNA as a function of frequency, (c) is the characteristic curve of the maximum gain of the LNA as a function of frequency, and (d) is the characteristic curve of the input reflection coefficient of the LNA as a function of frequency. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. The content described herein is for explanation rather than limitation of the present invention.

[0023] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification of this invention are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, systems, products or devices.

[0024] The following explains the relevant technical terms: Convolutional Neural Network (CNN); High Frequency Structure Simulator (HFSS); Mean Absolute Error (MAE); Linear rectifier unit (ReLU); Leaky Rectified Linear Unit (LeakyReLU); Breadth-First Search (BFS); The S-parameters include: S11, S21, S12, and S22. S11 is the input reflection coefficient, S21 is the forward transmission gain, S12 is the reverse transmission coefficient, and S22 is the output reflection coefficient.

[0025] This invention aims to solve the problems of traditional low-noise amplifier (LNA) design methods, which rely entirely on the designer's experience, require manual parameter trial and error using RF simulation software, have long design cycles, low efficiency, and lack systematic optimization logic, making it difficult to find the globally optimal matching network structure under the constraints of multiple performance objectives such as "low noise, high gain, and large bandwidth".

[0026] The core concept of this invention is a reverse design paradigm that directly drives topology generation based on performance metrics. This paradigm primarily protects the following: First, targeting desired LNA performance metrics (such as optimal noise matching and maximum power matching), a convolutional neural network model is used as a performance predictor to drive a genetic algorithm to perform a global search in the pixelated matrix space, automatically generating high-performance, non-classical LNA matching network structures. Second, the convolutional neural network model for the LNA matching network is trained through transfer learning. Third, before training the convolutional neural network model, efficient training data generation for the LNA matching network is required. After generating the pixelated matrix, a screening process is first performed based on an electrical connectivity judgment algorithm, only performing electromagnetic simulation on structures with effective electrical connections between ports. Fourth, LNA matching networks with different emphases are achieved by modifying the target S-parameters and cost function.

[0027] Based on the above core concept, the implementation scheme of this invention is achieved through three organically connected stages: Phase 1: Precise Definition of Active Core Performance Boundaries of LNA. Based on the selected active devices, through stability design, bias optimization, and layout-level electromagnetic simulation, the optimal noise impedance and maximum gain impedance within the target frequency band are precisely extracted as targets.

[0028] Phase Two: After generating a large number of random topology matrices, connectivity filtering is first performed using algorithms such as breadth-first search. Electromagnetic simulation is then conducted only on topologies that constitute valid two-port networks to obtain S-parameter labels. This method significantly improves the efficiency and quality of LNA matching network dataset construction by pre-eliminating invalid structures.

[0029] Phase 3: A mapping model from pixelated topology to S-parameters is constructed using a convolutional neural network, enabling instantaneous evaluation of the electromagnetic performance of any structure within the LNA band. Optimal noise matching and maximum power matching are used as optimization objectives to drive a genetic algorithm for targeted search, ultimately synthesizing a high-performance matching network structure.

[0030] like Figure 3As shown, the core process of this embodiment is completed collaboratively through four steps: "topology representation, data augmentation, model fitting, and intelligent optimization." Using the image pixelation approach, any LNA matching network structure is decomposed into pixels, quantifying the topological features of the matching network in the form of a two-dimensional matrix. For the pixelated matching network, electrical connectivity is first determined, and then electromagnetic simulation is performed on the selected data, significantly improving the efficiency and quality of LNA matching network dataset construction from the source. The selected data is then horizontally flipped to expand the dataset. During model training, a sub-model is first trained using the S11 dataset of the LNA matching network and saved as a pre-trained model. Then, transfer learning is used to freeze the convolutional layer parameters of the pre-trained model, and the fully connected layers are fine-tuned to train other S-parameters. This method can shorten training time and reduce validation error, achieving high-precision mapping from topology to S-parameters. Taking optimal noise matching and maximum power matching of LNA as optimization objectives, the target S-parameters and cost functions of optimal noise matching and maximum power matching are different. The average absolute error of the real and imaginary parts of the S-parameters predicted by CNN and the target S-parameters is used as the cost function of the genetic algorithm. The optimal solution is found by iteratively searching the two-dimensional matrix topology of the matching network. Finally, a high-performance LNA matching network topology is automatically synthesized in reverse, with 13 frequency points in the range of 10GHz~16GHz at 0.5GHz intervals.

[0031] Based on the above technical approach, such as Figure 1 As shown, this embodiment provides a wide-area intelligent design method for ultra-wideband low-noise amplifiers, including the following steps: S1, decompose the topology of the LNA matching network into a pixelated matrix; This step employs the concept of image pixelation to decompose any LNA matching network structure into pixels, quantifying the topological features of the matching network in the form of a 26×26 two-dimensional matrix. This representation method unifies the description of arbitrary LNA matching network structures, laying the foundation for subsequent intelligent processing.

[0032] S2, The pixelated matrix is ​​processed using a pre-trained deep learning model to obtain the predicted S-parameters corresponding to the topological structure; To achieve fast and accurate prediction of the S-parameters of the LNA matching network, this invention designs and trains a convolutional neural network model, the architecture of which is as follows: Figure 4 As shown, the model takes a 26×26 two-dimensional pixelated matrix as input, which represents the topology of the LNA matching network. The first row, first column, 26th row, and 26th column are fixedly defined as the positions of the excitation ports. The network employs a deep architecture of "12 convolutional layers + 5 fully connected layers," designed and optimized to address the complex nonlinear mapping relationship between the topological features and S-parameters of the LNA matching network. Figure 4Re(S11) is the real part of the S11 parameter, and Im(S11) is the imaginary part of the S11 parameter, for a total of 29 outputs.

[0033] Table 1 shows the specific design of 12 convolutional layers, and Table 2 shows the specific design of 5 fully connected layers. Table 1 Convolutional Layers

[0034] Table 2 Fully Connected Layers

[0035] Specifically, the pre-trained deep learning model is a convolutional neural network (CNN), which includes: an input layer, several convolutional layer modules, several fully connected layer modules, and an output layer connected in sequence. Each convolutional layer module includes a convolutional layer, a normalization layer, and an activation layer connected in sequence; each fully connected layer module includes a fully connected layer, a normalization layer, an activation layer, and a random deactivation layer connected in sequence. The convolutional layer modules, through multiple convolutions, normalization, and LeakyReLU activation layers, are responsible for extracting key electromagnetic features layer by layer from the pixelated topology; the fully connected layer modules, through multi-layer perceptron and random deactivation techniques, map high-level features to the final S-parameter prediction values, effectively preventing overfitting and improving the model's generalization ability. The trained CNN model can instantly and accurately predict the performance of the matching network, replacing a large amount of time-consuming electromagnetic simulation. This allows subsequent genetic algorithms to perform efficient global optimization, greatly improving the design efficiency of the LNA.

[0036] The training process of the pre-trained deep learning model includes the following steps: First, a large LNA matching network dataset was constructed. 480,000 26×26 two-dimensional matrices were randomly generated using an algorithm. Crucially, before performing the time-consuming electromagnetic simulation using HFSS, an electrical connectivity judgment algorithm (such as breadth-first search) was used to filter the pixelated matrices, performing electromagnetic simulations only on structures with valid electrical connections between ports to obtain S-parameter labels. The S-parameter dataset included S11, S21, S12, and S22 parameter datasets. This significantly improved the efficiency and quality of LNA matching network dataset construction from the outset, concentrating valuable electromagnetic simulation resources on effective samples.

[0037] Subsequently, the convolutional neural network is trained based on the S11 parameter dataset to obtain a pre-trained model.

[0038] Next, transfer learning is used to freeze the parameters of the convolutional layers of the pre-trained model, and the parameters of the fully connected layers of the pre-trained model are fine-tuned based on the S21 parameter dataset, S12 parameter dataset, and S22 parameter dataset, respectively. The pre-trained model is then trained to obtain a pre-trained deep learning model. This method can shorten training time and reduce validation error, achieving a high-precision mapping from topology to S-parameters.

[0039] S3, preset target S parameters, and use a global optimization algorithm to iteratively optimize the pixelated matrix. In each iteration, minimize the error between the predicted S parameters and the target S parameters in the real and imaginary parts. After the iteration is completed, the optimal pixelated matrix is ​​obtained. The preset target S-parameter specifically includes: For the input of the LNA matching network, a first objective S-parameter corresponding to optimal noise matching is preset; or for the output of the LNA matching network, a second objective S-parameter corresponding to maximum power matching is preset. The objective S-parameters and cost functions for optimal noise matching and maximum power matching are different. This provides precise and differentiated design capabilities for the LNA input and output ports. By flexibly configuring the objective S-parameters in the same inverse design system, the optimal performance input and output matching networks can be synthesized separately.

[0040] The genetic algorithm process is as follows: Figure 5 As shown, with the topology synthesis of LNA with optimal noise matching and maximum power matching as the core objective, the LNA matching network design is completed according to the logic of "population initialization - iterative optimization - optimal topology output".

[0041] First, the target S-parameters of the LNA matching network are set, and then 2048 pixelated two-dimensional matrices are randomly generated, each matrix uniquely corresponding to a topology of an LNA matching network.

[0042] Predict the S-parameters of the matching network for all two-dimensional matrices in the current population using a trained convolutional neural network.

[0043] The specific optimization process is as follows: When the first target S-parameters corresponding to the preset optimal noise matching are determined, the average absolute error of the real and imaginary parts of the first target S-parameters and the predicted S-parameters is calculated as the input matching cost function. The input matching cost function is minimized, and the pixelated matrix is ​​iteratively optimized using a genetic algorithm until the preset maximum number of iterations is reached, resulting in the optimal two-dimensional matrix of the noise matching, which is then used as the optimal pixelated matrix.

[0044] The input matching cost function is:

[0045] in, For the input matching cost function value, These are the weighting coefficients. Let Γ be the predicted real part of the S11 parameter. inreal目标 The target value for the real part of the LAN input reflection coefficient. Γ represents the predicted imaginary part of the S11 parameter. inimag目标 The target value for the imaginary part of the LAN input reflection coefficient.

[0046] When the second target S-parameters corresponding to the preset maximum power matching are calculated, the average absolute error of the real and imaginary parts of the second target S-parameters and the predicted S-parameters is calculated as the output matching cost function. The output matching cost function is minimized, and the pixelation matrix is ​​iteratively optimized by a genetic algorithm until the preset maximum number of iterations is reached to obtain the two-dimensional matrix of the maximum power matching, which is the optimal pixelation matrix.

[0047] The output matching cost function is:

[0048] in, To output the matching cost function value, These are the weighting coefficients. The predicted real part of the S22 parameter is... The target value is the real part of the conjugate of the LAN output reflection coefficient. The predicted value of the imaginary part of the S22 parameter. The target value for the imaginary part of the LAN output reflection coefficient conjugate.

[0049] In each iteration of the genetic algorithm, 248 samples are randomly selected from the population. The two samples with the smallest error to the target S-parameters are selected as the "parent matrix" and "mother matrix." A new offspring matrix is ​​generated through crossover, corresponding to a new LNA matching network topology. The error between the S-parameters predicted by the convolutional neural network and the target S-parameters of the LNA matching network is calculated, and the eight matrices with the smallest error are directly inherited to the offspring. At the same time, the offspring matrix generated by crossover is added to the new population, completing the population update. After 50 iterations, the best-performing two-dimensional matrix in the population is output, and its corresponding structure is the optimal input / output matching network topology that meets the performance requirements of LNA.

[0050] This step achieves coordinated optimization of noise and gain in the LNA matching network over a wide bandwidth. Traditional designs, limited by fixed topologies such as L-type and π-type, struggle to simultaneously achieve low noise and high gain over a wide bandwidth. This invention uses a pixelated matrix to uniformly represent arbitrary structures and employs a convolutional neural network to construct a high-precision mapping between these structures and S-parameters, completely eliminating dependence on classical topologies. This system can automatically generate non-classical topologies in a broad solution space, and the resulting matching network achieves better optimal noise matching and maximum power matching over a wide bandwidth, thereby simultaneously improving the LNA's core performance in terms of low noise, high gain, and wide bandwidth.

[0051] S4. The optimal pixelated matrix is ​​inversely synthesized to obtain the optimal topology of the LNA matching network.

[0052] Finally, a high-performance LNA matching network topology is automatically synthesized through reverse engineering. Based on this method, a low-noise amplifier with an operating frequency range of 10-16 GHz was designed, and its circuit schematic can be found in [reference needed]. Figure 6 , Figure 6 The diagram shows the basic structure of an LNA amplifier. The left side shows the input matching network, whose left end is connected to the input port. The output of the input matching network is connected to the input of the LNA via a DC blocking capacitor. The LNA input is also connected to a resistor. The bias signal is used to configure the DC operating point of the LNA. The middle unit is the LNA (Low Noise Amplifier), which serves as the core amplification module, performing low-noise gain amplification of the input signal. On the right is the output matching network, whose right end connects to the output port; the LNA's output is connected to this output matching network, and the output matching network is also connected to... The power supply provides power to the system; the output of the output matching network is connected to the output port through another DC blocking capacitor to achieve the output of the amplified signal.

[0053] See simulation results Figure 7 The characteristic curves of key performance parameters of low-noise amplifiers (LNAs) as a function of frequency in the 10~16GHz frequency band. Figure 7 In (a), the horizontal axis is labeled "freq", representing frequency in GHz, with a range of 10~16 GHz; the vertical axis is labeled "dB (S (2,1))", representing the forward transmission coefficient S. 21 The decibel value is expressed in dB; it shows the forward transmission gain (S) of the LNA in the 10~16GHz frequency band. 21 The gain varies with frequency. The curve shows that as the frequency increases from 10 GHz to 16 GHz, the gain gradually decreases from about 26.8 dB to about 24.7 dB.

[0054] Figure 7In (b), the horizontal axis is marked "freq", representing frequency in "GHz", with a range of 10~16GHz; the vertical axis is marked "NFmin", representing minimum noise figure in "dB"; it shows the variation of minimum noise figure of LNA with frequency in the 10~16GHz band, and the noise figure of LNA is less than 1.9dB. nf(2) represents the actual noise figure of port 2, that is, the noise figure of the signal input from port 1 and output from port 2.

[0055] Figure 7 In (c), the horizontal axis, labeled "freq," represents frequency in GHz, ranging from 10 to 16 GHz; the vertical axis, labeled "MaxGain1," represents maximum gain in dB. This shows the maximum gain of the LNA as a function of frequency within the 10-16 GHz band. Gain flatness is good, between 26 dB and 27.6 dB.

[0056] Figure 7 In (d), the horizontal axis is labeled "freq", representing frequency in GHz, with a range of 10~16 GHz; the vertical axis is labeled "dB (S (1,1))", representing the input reflection coefficient S. 11 The value is in decibels (dB); it shows the impedance matching characteristics of the LNA input port within the 10~16GHz frequency band. 11 The smaller the value, the better the matching performance. 11 The parameter is less than -8dB. Figure 7 This indicates that within the 10-16GHz frequency range, the noise figure is less than 1.9dB, S21 is higher than 24.5dB, the gain flatness is good, and S11 is less than -8dB.

[0057] Those skilled in the art should understand that, without departing from the core principles of the present invention, there are various alternative solutions to the above-described specific embodiments of the present invention. These alternative solutions can also achieve the purpose of the invention and should be covered within the protection scope of the present invention.

[0058] In this embodiment, the convolutional neural network used to construct the mapping relationship between the pixelated matrix and S-parameters can have its specific architecture adjusted in various ways. For example, the number of convolutional layers, the size and number of convolutional kernels, the number of fully connected layers, and the number of neurons, among other hyperparameters, can all be changed according to different accuracy and efficiency requirements. Furthermore, the prediction model formed by this convolutional neural network can also be replaced with other deep learning models suitable for processing image-based inputs and completing regression tasks, as long as they can achieve high-precision fitting from the topology to the S-parameters.

[0059] The genetic algorithm used in this embodiment for global search in the pixelated matrix space can be replaced by any other algorithm with global optimization capabilities to achieve the same inverse design objective. These algorithms include, but are not limited to, Empire Competition, Particle Swarm Optimization, Simulated Annealing, Greedy Algorithms, and Neighborhood Search Algorithms.

[0060] In this embodiment, the trained model can be quickly adapted to new LNA design scenarios (such as different semiconductor processes). Transfer learning techniques can be used. In specific implementation, some network layers can be frozen on the basis of the pre-trained model and the top network can be fine-tuned with only a small amount of new data. Other transfer learning strategies such as feature extraction can also be used to achieve rapid iteration and deployment of the model with extremely low data cost.

[0061] Besides the average absolute error form used in the cost function of this embodiment, other forms such as mean square error, error based on logarithms or decibels, or weighting factors at specific frequency points can also be used. As long as the core components of the cost function can reflect the difference between the predicted S-parameters and the target S-parameters in the real and imaginary parts, or are directly related to derived performance indicators such as the noise figure and gain of the LNA, they are all equivalent substitutions of this invention.

[0062] like Figure 2 As shown, this embodiment provides a wide-area intelligent design system for ultra-wideband low-noise amplifiers, including: The topology representation module is used to decompose the topology of the LNA matching network into a pixelated matrix. The performance prediction module is used to process the pixelated matrix using a pre-trained deep learning model to obtain the predicted S-parameters corresponding to the topology. The intelligent optimization module is used to preset the target S-parameters and use a global optimization algorithm to iteratively optimize the pixelated matrix. In each iteration, the error between the predicted S-parameters and the target S-parameters in the real and imaginary parts is minimized, and the optimal pixelated matrix is ​​obtained after the iteration is completed. The reverse synthesis module is used to reverse synthesize the optimal pixelated matrix to obtain the optimal topology of the LNA matching network.

[0063] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0064] This embodiment also provides a computer device, which includes a processor and a memory. The memory is used to store a computer program (in this embodiment, the computer program includes computational components and iterative components, capable of model calculation and model updating). The computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to realize the corresponding method flow or corresponding function. The processor described in this embodiment can be used for the operation of the ultra-wideband low-noise amplifier wide-area intelligent design method.

[0065] This embodiment also provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the ultra-wideband low-noise amplifier wide-area intelligent design method in the above embodiment.

[0066] This embodiment also provides a computer program product, which includes a computer program that, when executed by a processor, implements the corresponding steps of the ultra-wideband low-noise amplifier wide-area intelligent design method described in the above embodiment.

[0067] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0068] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0069] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0070] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A wide-area intelligent design method for an ultra-wideband low-noise amplifier, characterized in that, Includes the following steps: The topology of the LNA matching network is decomposed into pixels to obtain a pixelated matrix; The pixelated matrix is ​​processed using a pre-trained deep learning model to obtain the predicted S-parameters corresponding to the topological structure. A target S-parameter is preset, and a global optimization algorithm is used to iteratively optimize the pixelated matrix. In each iteration, the error between the predicted S-parameter and the target S-parameter in the real and imaginary parts is minimized. After the iteration is completed, the optimal pixelated matrix is ​​obtained. By inversely synthesizing the optimal pixelation matrix, the optimal topology of the LNA matching network is obtained.

2. The wide-area intelligent design method for an ultra-wideband low-noise amplifier according to claim 1, characterized in that, The pre-trained deep learning model is a convolutional neural network, which includes: an input layer, several convolutional layer modules, several fully connected layer modules, and an output layer connected in sequence. The convolutional layer module includes a convolutional layer, a normalization layer, and an activation layer connected in sequence. The fully connected layer module includes a fully connected layer, a normalization layer, an activation layer, and a random deactivation layer connected in sequence.

3. The wide-area intelligent design method for an ultra-wideband low-noise amplifier according to claim 2, characterized in that, The training process of the pre-trained deep learning model includes the following steps: The pixelated matrix is ​​filtered using an electrical connectivity judgment algorithm, and electromagnetic simulation is performed to obtain an S-parameter dataset, which includes an S11 parameter dataset, an S21 parameter dataset, an S12 parameter dataset, and an S22 parameter dataset. The convolutional neural network is trained based on the S11 parameter dataset to obtain a pre-trained model; The parameters of the convolutional layers of the pre-trained model are frozen using transfer learning, and the parameters of the fully connected layers of the pre-trained model are fine-tuned based on the S21 parameter dataset, S12 parameter dataset, and S22 parameter dataset, respectively. The pre-trained model is then trained to obtain a pre-trained deep learning model.

4. The wide-area intelligent design method for an ultra-wideband low-noise amplifier according to claim 1, characterized in that, The preset target S-parameter specifically includes: For the input of the LNA matching network, the first target S-parameters corresponding to the best noise matching are preset; Alternatively, for the output of the LNA matching network, preset the second target S-parameter corresponding to the maximum power matching.

5. The wide-area intelligent design method for an ultra-wideband low-noise amplifier according to claim 4, characterized in that, When the first target S-parameters corresponding to the optimal noise matching are preset, the pixelation matrix is ​​iteratively optimized using a global optimization algorithm. In each iteration, the error between the predicted S-parameters and the target S-parameters in the real and imaginary parts is minimized. After iteration, the optimal pixelation matrix is ​​obtained, specifically including: Calculate the mean absolute error between the real and imaginary parts of the first target S-parameters and the predicted S-parameters, and use it as the input matching cost function; The input matching cost function is minimized, and the pixelated matrix is ​​iteratively optimized using a genetic algorithm until a preset maximum number of iterations is reached, resulting in the best noise-matching two-dimensional matrix, which is then used as the optimal pixelated matrix.

6. The wide-area intelligent design method for an ultra-wideband low-noise amplifier according to claim 4, characterized in that, When the second target S-parameters corresponding to the preset maximum power matching are used, the pixelation matrix is ​​iteratively optimized using a global optimization algorithm. In each iteration, the error between the predicted S-parameters and the target S-parameters in the real and imaginary parts is minimized. After the iteration is completed, the optimal pixelation matrix is ​​obtained, specifically including: The mean absolute error of the real and imaginary parts of the second target S-parameters and the predicted S-parameters is calculated and used as the output matching cost function; The output matching cost function is minimized, and the pixelation matrix is ​​iteratively optimized using a genetic algorithm until a preset maximum number of iterations is reached, resulting in a two-dimensional matrix with maximum power matching, which is taken as the optimal pixelation matrix.

7. A wide-area intelligent design system for an ultra-wideband low-noise amplifier, characterized in that, include: The topology representation module is used to decompose the topology of the LNA matching network into a pixelated matrix. The performance prediction module is used to process the pixelated matrix using a pre-trained deep learning model to obtain the predicted S-parameters corresponding to the topology. The intelligent optimization module is used to preset the target S-parameters and use a global optimization algorithm to iteratively optimize the pixelated matrix. In each iteration, the error between the predicted S-parameters and the target S-parameters in the real and imaginary parts is minimized, and the optimal pixelated matrix is ​​obtained after the iteration is completed. The reverse synthesis module is used to reverse synthesize the optimal pixelated matrix to obtain the optimal topology of the LNA matching network.

8. An electronic device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable in the processor, wherein the processor, when executing the computer program, implements the wide-area intelligent design method for an ultra-wideband low-noise amplifier as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the wide-area intelligent design method for an ultra-wideband low-noise amplifier as described in any one of claims 1 to 6.

10. A computer program product, the computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the wide-area intelligent design method for an ultra-wideband low-noise amplifier as described in any one of claims 1 to 6.

Images