Radio frequency device hyper-parameter automatic optimization algorithm library

Abstract

The invention provides a radio frequency device hyper-parameter automatic optimization algorithm library. The radio frequency device hyper-parameter automatic optimization algorithm library comprises an interface module, a directional optimization algorithm module and a multi-modal optimization engine module, according to the optimization algorithm library, the automation level and the performance optimization efficiency of radio frequency device design are remarkably improved by constructing a domain knowledge graph covering six core devices including a radio frequency filter, a power amplifier, a numerical control attenuator and a low-noise amplifier and combining a customized optimization algorithm and a cross-level cooperation strategy. On the algorithm design level, aiming at the specific characteristics of the device, different hybrid optimization strategies are adopted, the global search capability of a genetic algorithm and a particle swarm optimization algorithm and the local optimization advantage of a gradient descent method are combined, and dynamic weight adjustment is matched, so that the targeted radio frequency device optimization process design is realized. According to the optimization algorithm library, through an independent calling interface for packaging the automatic design algorithm, the use difficulty of the automatic design algorithm is simplified, and the optimization process is normalized.

Description

Technical Field

[0001] This invention relates to the field of electronic design automation technology, and in particular to a packaged automatic hyperparameter optimization algorithm library for radio frequency devices based on machine learning technology. Background Technology

[0002] Electronic Design Automation (EDA) is a technology that uses Computer-Aided Design (CAD) to build a complete solution from architecture design to wafer fabrication. EDA is a crucial research area in electronic design today. Compared to traditional circuit design flows, EDA technology greatly frees up developers' time, standardizes the circuit design process, and significantly accelerates circuit design efficiency. Computer-Aided Engineering (CAE), on the other hand, focuses on physical reliability analysis, identifying design flaws early through thermal / mechanical / electromagnetic simulations. EDA and CAE form a dual-engine paradigm: EDA generates circuit entities (such as RTL synthesis and placement / routing), while CAE verifies physical characteristics such as thermal distribution and electromagnetic compatibility. Both improve efficiency through optimization algorithms, which undertake key tasks such as multi-objective trade-offs and placement / routing optimization.

[0003] However, existing optimization algorithms face three major bottlenecks:

[0004] 1. The contradiction between accuracy and efficiency: Full-wave simulation requires several hours for a single iteration, while fast models will have an error of more than 20% in the millimeter-wave band;

[0005] 2. Curse of Dimensionality: High-frequency parasitic effects and multi-parameter constraints make the upper limit of variables in commercial EDA tools insufficient. For example, multi-band PA optimization needs to handle 20+ variables, but the Cadence-Cerebrus tool only supports 10 dimensions.

[0006] 3. Cross-level gap: The correlation between schematic and layout parameters is less than 0.65, resulting in over 50% of schematic designs needing rework, and the existing joint optimization module coverage is less than 40%. More importantly, AI technology presents a "prediction-optimization" disconnect, with machine learning only used for performance prediction, and the optimization engine still relying on traditional processes. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides an automatic optimization algorithm library for hyperparameters of radio frequency devices.

[0008] To achieve the above objectives, this invention provides an automatic hyperparameter optimization algorithm library for radio frequency devices, comprising:

[0009] The interface module is used to configure the target characteristic parameters and optimization objects of different RF devices;

[0010] The targeted optimization algorithm module includes: multiple automatic optimization algorithm units for functional devices, which, based on the target feature parameters and the optimization object, call the automatic optimization algorithm unit that matches the optimization object to execute the automatic optimization process; the automatic optimization algorithm unit optimizes the device parameters and finally outputs the optimal design parameters.

[0011] The multimodal optimization engine module is embedded in each automatic optimization algorithm unit of the directional optimization algorithm module. It includes multiple optimization algorithm units that perform optimization operations based on the target feature parameters of the input RF device through the corresponding automatic optimization algorithm unit for the functional device.

[0012] Preferably, the multimodal optimization engine module includes: a gradient descent algorithm subunit, a genetic algorithm subunit, and a particle swarm optimization algorithm subunit; based on the target feature parameters of the input RF device, the automatic optimization algorithm unit for functional devices calls the corresponding subunit to perform optimization operations during optimization.

[0013] Preferably, the multimodal optimization engine module also includes: a dynamic weight adjustment algorithm and a sensitivity analysis algorithm; the dynamic weight adjustment algorithm is used to adjust the weights of different optimization objectives; the sensitivity analysis algorithm is used to analyze the sensitivity of design parameters; the optimization algorithm library also includes: a simulation engine, used to simulate key circuits in the optimization process.

[0014] Preferably, the automatic optimization algorithm unit for functional devices optimizes target feature parameters at both the schematic layer level and the board layer level. During the optimization process, the results of the schematic layer-level optimization are input into the board layer-level optimization, and the optimized GDSII format file is output as the final output result.

[0015] Preferably, when this optimization algorithm library performs principle layer-level optimization, the steps include:

[0016] Obtain the netlist file of the RF device, verify the integrity of the netlist, and prompt the user when the verification is complete;

[0017] Initialize the optimization parameters and netlist objects, and discretize the parameter space;

[0018] For the design parameter information, ADS co-simulation is used, and optimization algorithms are employed for multi-objective analysis and optimization.

[0019] The optimization convergence result is evaluated. If it meets the design requirements, the optimal design scheme is output. If it does not meet the design requirements, the optimization algorithm optimization space is updated, and the simulation and optimization process is re-executed.

[0020] Preferably, the directional optimization algorithm module, depending on the optimization object, includes the following automatic optimization algorithm units for functional devices: an automatic optimization algorithm unit for RF filters, an automatic optimization algorithm unit for RF switching circuits, an automatic optimization algorithm unit for RF power amplifiers, an automatic optimization algorithm unit for RF low-noise amplifiers, an automatic optimization algorithm unit for RF drive amplifiers, and an automatic optimization algorithm unit for RF digitally controlled attenuators.

[0021] Preferably, the automatic optimization algorithm unit for functional devices in the directional optimization algorithm module includes the following steps when performing the optimization operation:

[0022] Initialize the target characteristic parameters of the RF device and resolve the netlist;

[0023] The principle layer of the radio frequency device is optimized. During the optimization, the target characteristic parameters are adjusted in multiple stages according to the design focus of the radio frequency device to generate an initial design scheme.

[0024] Layout-level optimization of RF devices: The initial design scheme is imported into the layout design as the initial value, the corresponding target feature parameters are initialized, and the corresponding optimization algorithm unit in the multi-modal optimization engine module is called for optimization.

[0025] The final design result is output after verification.

[0026] Preferably, in the directional optimization algorithm module, the automatic optimization algorithm unit for RF filters uses the PSO algorithm for optimization search when performing principle layer-level optimization and layout layer-level optimization. In layout layer-level optimization, particle movement velocity correction is assisted based on gradient descent algorithm.

[0027] Preferably, in the directional optimization algorithm module, the automatic optimization algorithm unit for RF driver amplifiers, the automatic optimization algorithm unit for RF low noise amplifiers, and the automatic optimization algorithm unit for RF digitally controlled attenuators, when performing principle layer-level optimization, use the Morris global sensitivity analysis algorithm to identify the dominant parameters and dynamically compress the design range of variable design parameters; and optimize the design focus through dynamic weight allocation.

[0028] Preferably, in the directional optimization algorithm module, after obtaining the initial design scheme, the automatic optimization algorithm unit for RF filters uses a simulation engine to simulate the current design scheme, uses an evaluation function to calculate the difference between the simulation results and the expected values ​​of the optimization target parameters, and adds the weights as the score of the current scheme.

[0029] The automatic hyperparameter optimization algorithm library for radio frequency devices provided by this invention has the following advantages compared with the prior art:

[0030] This invention significantly improves the automation level and performance optimization efficiency of RF device design by constructing a domain knowledge graph covering six core devices: RF filters, power amplifiers, digitally controlled attenuators, and low-noise amplifiers, and by combining customized optimization algorithms and cross-level collaborative strategies.

[0031] This invention analyzes the limitations of the optimization process for radio frequency devices, proposes the necessity of a targeted design process, and designs a targeted optimization framework based on the characteristics of different devices / chips / circuits.

[0032] At the algorithm design level, different hybrid optimization strategies are adopted based on the specific characteristics of the devices. These strategies combine the global search capabilities of genetic algorithms and particle swarm optimization with the local optimization advantages of gradient descent, along with dynamic weight adjustment, to achieve targeted optimization process design for RF devices. This library simplifies the use of automated design algorithms and standardizes the optimization process by providing independent calling interfaces that encapsulate the automated design algorithms. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the structure of the automatic hyperparameter optimization algorithm library for radio frequency devices provided by the present invention;

[0034] Figure 2 A flowchart illustrating the algorithm technology of the automatic hyperparameter optimization algorithm library for radio frequency devices provided by this invention.

[0035] Figure 3 The principle layer-level optimization logic flowchart of the RF device hyperparameter automatic optimization algorithm library provided by this invention (the layer-level logic is the same);

[0036] Figure 4 The schematic diagram of the automatic optimization algorithm library for RF device hyperparameters provided by this invention is used to determine the output result of the optimization (the logic is the same at the layout level).

[0037] Figure 5 This is a schematic diagram illustrating the classification of optimization objects and optimization targets in the automatic hyperparameter optimization algorithm library for radio frequency devices provided by this invention.

[0038] Figure 6 A schematic diagram of the RF filter automatic design interface calling logic of the RF device hyperparameter automatic optimization algorithm library provided by the present invention;

[0039] Figure 7 Pseudocode diagram for adaptive adjustment of parameter variation range in the automatic optimization algorithm library for RF device hyperparameters provided by this invention;

[0040] Figure 8 The filter schematic topology diagram optimized in Example 1 is provided in the RF device hyperparameter automatic optimization algorithm library provided by the present invention.

[0041] Figure 9 The layout design diagram of the optimization result in Example 1 of the RF device hyperparameter automatic optimization algorithm library provided by the present invention;

[0042] Figure 10 The topology diagram of the optimized switching circuit schematic in Example 2 is provided in the automatic optimization algorithm library for RF device hyperparameters provided by the present invention.

[0043] Figure 11 The layout design diagram of the optimization result in Example 2 of the RF device hyperparameter automatic optimization algorithm library provided by the present invention;

[0044] Figure 12 The optimized RF power amplifier schematic topology diagram in Example 3 is provided in the RF device hyperparameter automatic optimization algorithm library provided by the present invention.

[0045] Figure 13 The layout design diagram of the optimization result in Example 3 of the RF device hyperparameter automatic optimization algorithm library provided by the present invention;

[0046] Figure 14 The topology diagram of the optimized radio frequency driver amplifier schematic in Example 4 is provided in the radio frequency device hyperparameter automatic optimization algorithm library provided by the present invention.

[0047] Figure 15 The layout design diagram of the optimization result in Example 4 of the automatic optimization algorithm library for RF device hyperparameters provided by the present invention;

[0048] Figure 16 The topology diagram of the optimized radio frequency low noise amplifier schematic in Example 5 is provided in the radio frequency device hyperparameter automatic optimization algorithm library provided by the present invention.

[0049] Figure 17 The layout design diagram of the optimization result in Example 5 of the RF device hyperparameter automatic optimization algorithm library provided by the present invention;

[0050] Figure 18 The schematic topology diagram of the optimized radio frequency digitally controlled attenuator in Example 6 is provided in the radio frequency device hyperparameter automatic optimization algorithm library provided by the present invention.

[0051] Figure 19 The layout design diagram of the optimization result in Example 5 of the automatic optimization algorithm library for RF device hyperparameters provided by the present invention. Detailed Implementation

[0052] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0053] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0054] like Figure 1 , 2 As shown, the RF device hyperparameter automatic optimization algorithm library (hereinafter referred to as the "optimization algorithm library") provided by this invention includes: an interface module, a directional optimization algorithm module, and a multimodal optimization engine module. The interface module is used to configure the target feature parameters and optimization objects of different RF devices. The directional optimization algorithm module includes: multiple automatic optimization algorithm units for functional devices, which, based on the target feature parameters and optimization objects, call the automatic optimization algorithm unit matching the optimization object to execute the automatic optimization process; the automatic optimization algorithm unit optimizes the device parameters and ultimately outputs the optimal design parameters. The multimodal optimization engine module is embedded in each automatic optimization algorithm unit of the directional optimization algorithm module, and includes: multiple optimization algorithm units that perform optimization operations through the corresponding automatic optimization algorithm units for functional devices based on the input target feature parameters of the RF device.

[0055] Specifically, the optimization algorithm library provides an interface module (107) for users to call for optimization. The input information (130) of the interface includes: the object to be optimized, the optimization parameters, and the parameter scan range. By inputting the device type information in the parameter information, the algorithm interface will automatically call the matching customized optimization process algorithm (101-106). After the optimization process, the optimization algorithm library (100) will output the final optimized design result (131). During the optimization process, based on the different optimization objects, the optimization algorithm sub-units in the multimodal optimization engine module are called to complete the final optimization. If the optimization fails, the optimization algorithm library will output the reason for the failure. The reasons for failure include: unreasonable input parameters, netlist simulation failure, automatic layout splicing failure, and unreasonable final optimization result.

[0056] In short, the optimization algorithm library significantly improves the automation level and performance optimization efficiency of RF device design by constructing a domain knowledge graph covering six core RF devices (i.e., optimization objects), including RF filters, power amplifiers, digitally controlled attenuators, and low-noise amplifiers. This is achieved through customized optimization algorithms and cross-level collaborative strategies. The library analyzes the limitations of the RF device optimization process, proposes the necessity of a targeted design process, and designs a targeted optimization framework based on the characteristics of different devices / chips / circuits. At the algorithm design level, different hybrid strategies are adopted for the specific characteristics of RF devices. These strategies combine the global search capabilities of genetic algorithms and particle swarm optimization with the local optimization advantages of gradient descent based on the target feature parameters of the RF devices, along with dynamic weight adjustment, to achieve targeted RF device optimization process design. The library simplifies the use of automated design algorithms and standardizes the optimization process by providing independent calling interfaces for the encapsulated automated design algorithms.

[0057] For nonlinear parameter spaces (such as power amplifier gain-efficiency curves), this invention employs a GA / PSO global search strategy (110), setting the parameter range with a crossover probability of 0.6-0.8 and a mutation probability of 0.01-0.05. For continuously differentiable parameters (such as filter LC parameters), an improved adaptive gradient descent method (113) is used, dynamically adjusting the learning rate to a range of 1e-5 to 1e-3. A parameter sensitivity analysis algorithm based on Monte Carlo sampling is established, and key influencing factors are determined through 500-1000 rapid simulations. A multi-objective Pareto optimization framework is developed to support automatic trade-offs between gain, noise, and power consumption.

[0058] In this embodiment, the multimodal optimization engine module includes a gradient descent algorithm subunit, a genetic algorithm subunit, and a particle swarm optimization algorithm subunit. Based on the target feature parameters of the input RF device, the automatic optimization algorithm unit for functional devices calls the corresponding subunit to perform optimization operations during optimization.

[0059] The multimodal optimization engine module also includes: a dynamic weight adjustment algorithm and a sensitivity analysis algorithm; the dynamic weight adjustment algorithm is used to adjust the weights of different optimization objectives; the sensitivity analysis algorithm is used to analyze the sensitivity of design parameters; the optimization algorithm library also includes: a simulation engine, used to simulate key circuits in the optimization process. For example, through dynamic weight adjustment, the optimization algorithm can adjust the optimization focus at different stages.

[0060] Sensitivity analysis algorithms can identify sensitive and insensitive variables. The optimization process focuses on adjusting sensitive variables, reducing the dimensionality of the adjustable parameter space, lowering the difficulty of optimization, and improving optimization efficiency. The simulation engine transforms design parameters into circuit characteristic information. The optimization goal of the circuit is its performance, and the optimization space corresponds to the circuit's design parameters. Simulation is necessary to transform design parameters into circuit performance, allowing the design parameters to be scored based on performance and the target.

[0061] Among them, the optimization algorithm library (100) is a fully encapsulated Python code library that can be called as a third-party library in other Python code environments.

[0062] The automatic optimization algorithm unit for functional devices optimizes target feature parameters at both the schematic and layout levels. During optimization, the results from the schematic-level optimization are input into the layout-level optimization, and the final output is a GDSII format file. At the schematic level, the optimization algorithm library performs design optimization based on the characteristics of the device's variable design parameters, and netlist simulation is used for schematic simulation. At the layout level, the optimization algorithm library performs variable design parameter optimization on the layout design results obtained from automatic placement and routing based on the schematic topology.

[0063] like Figure 3 As shown, the optimization algorithm library performs the following steps when performing principle layer-level optimization:

[0064] Obtain the netlist file of the RF device, verify the integrity of the netlist, and prompt the user when the verification is complete;

[0065] Initialize the optimization parameters and netlist objects, and discretize the parameter space;

[0066] For the design parameter information, ADS co-simulation is used, and optimization algorithms are employed for multi-objective analysis and optimization.

[0067] The optimization convergence result is evaluated. If it meets the design requirements, the optimal design scheme is output. If it does not meet the design requirements, the optimization algorithm optimization space is updated, and the simulation and optimization process is re-executed.

[0068] like Figure 4 As shown, the specific steps for judging the output result of schematic optimization (with the same logic at the layout level) include:

[0069] The schematic layer-level optimization results are evaluated. If the improvement is stalled or resources are exhausted according to the evaluation criteria, manual design review is conducted. Based on the design indicators, it is determined whether layout-level design is feasible. If approved, the process enters the layout optimization workflow entry point; otherwise, it returns to the schematic for redesign. If the optimization result error meets the evaluation criteria, an optimization workflow report is generated, and the process enters the layout design workflow entry point.

[0070] In this embodiment, different optimization algorithm interfaces in the interface module (107) have different input parameters. The output of individual optimization algorithms is uniformly the save path of the optimization design results. That is, each optimization algorithm module will output a lot of information and files during the optimization process, such as: optimization design results, optimization process data, GDSII files, netlist files, etc. These files are all saved in a specified folder. The optimization result of the module is the path of this folder. If an abnormal problem occurs at the end of the optimization process, the optimization algorithm library will end the optimization process in advance, output the current optimization progress and report an error. The circuit object simulation of the optimization algorithm library is implemented by calling a third-party simulation engine (120). The optimization algorithm library itself does not have circuit simulation capabilities.

[0071] like Figure 5 As shown, this optimization algorithm library focuses on six commonly used RF functional devices: RF filters, power amplifiers, digitally controlled attenuators, and low-noise amplifiers. It develops customized automatic optimization processes and dedicated optimization workflows. The automatic optimization algorithm units corresponding to the functional devices include: automatic optimization algorithm units for RF filters, RF switching circuits, RF power amplifiers, RF low-noise amplifiers, RF driver amplifiers, and RF digitally controlled attenuators.

[0072] The automatic optimization algorithm unit for functional devices in the directional optimization algorithm module includes the following steps during the optimization operation:

[0073] Initialize the target characteristic parameters of the RF device and resolve the netlist;

[0074] The principle layer of the radio frequency device is optimized. During the optimization, the target characteristic parameters are adjusted in multiple stages according to the design focus of the radio frequency device to generate an initial design scheme.

[0075] Layout-level optimization of RF devices: The initial design scheme is imported into the layout design as the initial value, the corresponding target feature parameters are initialized, and the corresponding optimization algorithm unit in the multi-modal optimization engine module is called for optimization.

[0076] The final design result is output after verification.

[0077] like Figure 6 As shown, taking an RF filter as an example, its interface call logic is as follows: the optimization indicators related to the RF filter, the topology-related parameters, and the modifiable optimization algorithm hyperparameters are used as input parameters and input into the RF filter optimization module through the corresponding algorithm interface for data optimization. After that, the output includes the circuit topology, layout design GDSII file, and related test reports and optimization process records containing the design results. Figure 7 Code example for adaptively adjusting for parameter variations.

[0078] In the optimization algorithm library, the automatic optimization algorithm unit (101) for RF filters supports automatic integrated optimization of three filter types: Butterworth, Chebyshev, and Elliptic. In the optimization process, the first stage uses a genetic algorithm to optimize the order / cutoff frequency, and the second stage uses the gradient descent method to fine-tune the LC parameters and integrate the Q-value constraints to ensure that the component parameters meet the manufacturing capabilities.

[0079] Specifically, the optimization module takes the radio frequency filter as the optimization object, and takes the six optimization target parameters of the radio frequency filter (passband start frequency, passband cutoff frequency, passband gain, stopband attenuation, group delay, and impedance matching) as the main optimization objects, and quantizes them into a six-dimensional parameter space.

[0080] Impedance matching is a characteristic of circuit devices, derived from circuit performance. Appropriate impedance matching allows devices to perform at their designated levels within the system (50-ohm impedance matching is commonly used). However, this is not a practical design specification. Therefore, impedance matching is treated as a constraint; if the impedance matching corresponding to a design parameter is inappropriate, that design parameter is abandoned.

[0081] At the level of RF filter optimization algorithm design, this invention proposes a hybrid optimization strategy that integrates a multi-objective Pareto front search algorithm with a dynamic weight adjustment strategy. This scheme constructs a four-dimensional objective function that includes insertion loss, in-band ripple, out-of-band suppression, and size constraints. It then employs a multi-objective optimization algorithm for global Pareto front search, followed by local optimization of candidate solutions using a dynamic weight adjustment strategy, ultimately achieving a comprehensive balance of performance indicators.

[0082] During the optimization process, the corresponding evaluation function is used to calculate the difference between the optimization target parameter and the expected value in the simulation results under the current scheme. Then, the weights are added together as the score of the current scheme, and the feasibility of the current design scheme is judged based on the simulation results.

[0083] In the layer-level optimization algorithm, a gradient descent optimization algorithm has been added. The gradient information guides the movement direction of the particle swarm, effectively accelerating the convergence efficiency of PSO particles, reducing the number of simulation calls, and thus saving overall optimization time.

[0084] The automatic optimization algorithm unit (102) for RF switching circuits develops a charge injection compensation algorithm by establishing a multi-objective optimization framework of insertion loss, isolation and switching time to suppress switching transient interference.

[0085] Specifically, the optimization algorithm combines the four optimization objectives of the RF switching circuit (1dB compression point, insertion loss, isolation, and return loss) into a four-dimensional design space.

[0086] This optimization algorithm utilizes a genetic algorithm to simultaneously address nonlinear coupling metrics in switching circuit design. Depending on the design stage, it can optimize designs with different focuses, and after weight adjustments, effectively inherit the "genes" from the previous stage into the next generation of designs. Furthermore, the unique crossover and mutation operations of the genetic algorithm can effectively guide the design system to escape multiple local optima in the design parameter space and perform a global optimum search by combining the current optimal design "genes."

[0087] The optimization stage for RF switch circuits is determined by efficiency. When there is no significant improvement in efficiency after five consecutive optimization iterations, the current stage is considered nearing completion, and the next design stage can begin. The two-level automated design flow for RF switch circuits must complete three stages, from optimizing insertion loss and return loss to achieving the final 1dB compression point, before algorithm convergence verification and exiting the current optimization phase. In addition to the three exit conditions of the basic optimization flow, the RF switch circuit optimization flow can also terminate if the efficiency improvement of the optimal solution is less than 0.5% for 20 consecutive iterations after completing the three optimization stages, outputting the current optimal solution. Finally, hard constraints are used in the design scheme evaluation function to isolate the impact of poor design schemes on the overall optimization.

[0088] For the automatic optimization algorithm unit (103) for RF power amplifiers, it focuses on the integration of unit gain and in-band stability by constructing nonlinear behavior description equations to carry out overall optimization design.

[0089] Specifically, the RF power amplifier (PA) optimization unit focuses on the collaborative optimization of multiple objectives, constructing a six-dimensional parameter space that includes the operating frequency band, saturated output power, power-added efficiency (PAE), and return loss. Addressing the nonlinear relationships such as the positive correlation between PAE and output power being constrained by transistor saturation effects and the negative coupling between efficiency and linearity, a multi-objective particle swarm optimization (MOPSO) algorithm is employed to perform Pareto front search. Through dynamic inertia weight adjustment and crowding ranking mechanisms, conflicts between objectives are balanced, ensuring comprehensive optimization of core indicators such as efficiency, bandwidth, and linearity.

[0090] The technical process employs a phased optimization strategy: at the schematic layer level, MOPSO is used for global optimization to quickly generate preliminary solutions that meet RF scattering parameters and voltage / current characteristics; at the layout layer level, a hybrid optimization framework is introduced, combining the global search capability of MOPSO with the local tuning advantages of the genetic algorithm (NSGA-II) to address high-dimensional parameter spaces and electromagnetic coupling effects. For constraint handling, a stability factor K (K>1.2) is used to enforce circuit stability, and process rules such as metal density are embedded as adaptive penalty functions to ensure the manufacturability of the design. Furthermore, this optimization unit uses process specifications for metal density and system stability to strengthen constraints, accelerating overall optimization efficiency.

[0091] For the automatic optimization algorithm unit (104) for RF low noise amplifiers, it integrates an automatic calculation and constraint mechanism for the stability factor K, combined with a dynamic weight adjustment algorithm, to accelerate the optimization efficiency.

[0092] Specifically, the RF low-noise amplifier (LNA) optimization unit employs an improved non-dominated sorting genetic algorithm (NSGA-II) framework to collaboratively optimize multi-objective coupling characteristics such as noise figure (NF), gain, and linearity (P1dB). By constructing a weighted fitness function that includes noise normalization scores, gain scores, in-band flatness, and stability penalties, transistor size, bias voltage, and matching network parameters are used as core optimization variables. An arithmetic crossover strategy is employed, particularly for high-sensitivity parameters (such as gate inductance), while uniform mutation is applied to discrete parameters (number of transistors in parallel), effectively avoiding optimization biases caused by differences in parameter sensitivity.

[0093] The technical process adopts a four-stage optimization architecture: First, the range of key component parameters is identified through netlist analysis, and the dominant variables (input matching inductance, source degradation inductance, etc.) are screened using Morris sensitivity analysis, and low-sensitivity parameters (<0.001) are fixed to reduce dimensionality; then, a phased dynamic weight adjustment strategy is implemented, initially focusing on noise optimization (weight 0.6), and later balancing gain (0.3) and linearity indicators; finally, the noise figure improvement rate (<0.1% for 10 consecutive generations) is used as the convergence criterion to ensure that the overall performance of gain enhancement and input matching meeting design specifications is achieved while maintaining noise control.

[0094] For the automatic optimization algorithm unit (105) for RF driver amplifiers, a gain flatness optimization function is constructed and ±0.5dB in-band ripple constraint is added.

[0095] Specifically, the driver amplifier (DA) optimization unit employs a hierarchical hybrid optimization strategy, integrating the global search capability of a multi-objective genetic algorithm with the local fine-tuning capability of gradient descent. The technical framework consists of two phases: First, NSGA-II is used to perform combined optimization of the matching network topology (equal-Q LC structure), transistor size, and negative feedback network parameters to achieve global design space exploration; then, based on the adjoint network method, the sensitivity of S-parameters to component values ​​is calculated, and gradient optimization is implemented for return loss at key frequency points to directionally adjust the parameter change trajectory.

[0096] The technical process employs a dynamic parameter control mechanism: in the initial stage, an optimization baseline is established through netlist analysis and variable parameter detection, and key variables are screened using sensitivity analysis; during the optimization process, the variation probability (in the range of 0.05-0.15) of highly sensitive parameters (such as negative feedback resistors) is dynamically adjusted based on the convergence state, breaking through local optima traps through parameter perturbation. In the layout-level optimization stage, a gradient direction correction strategy is specifically introduced, combining electromagnetic field simulation data to compensate for parasitic parameters in the layout, ensuring consistency between high-frequency characteristics and the schematic design.

[0097] Morris's algorithm can analyze the sensitivity of different parameters; the more sensitive the parameter, the greater its impact on circuit performance. Therefore, insensitive parameters can be ignored to some extent in optimization, while sensitive parameters are given more attention. Increasing the exploration of highly sensitive parameters (here, a genetic algorithm is used; mutation means that only this parameter changes more significantly, and the mutation rate is the probability of mutation occurring. By increasing these two parameters, sensitive parameters can explore more of the optimization space, thus escaping local optima and improving search efficiency).

[0098] Among them, the automatic optimization algorithm unit (106) for radio frequency digitally controlled attenuators establishes a joint optimization objective function of attenuation accuracy and linearity, and the weight coefficients can be dynamically configured.

[0099] Specifically, the RF digitally controlled attenuator optimization module adopts an improved non-dominated sorting genetic algorithm (NSGA-II) framework to collaboratively optimize the multi-objective coupling characteristics of insertion loss, attenuation range, return loss, and attenuation accuracy. This framework achieves efficient Pareto front search through non-dominated sorting and crowding calculation, while introducing an adaptive weighting strategy to balance the conflicting relationships between the indices.

[0100] The dynamic weighting mechanism for RF digitally controlled attenuators is as follows: initially, the focus is on optimizing insertion loss and return loss (60% weighting), and later, the focus shifts to fine-tuning attenuation accuracy (RMSA) and phase error (RMSP) (70% weighting). Simultaneously, during the iteration process, the weighting of different target parameters is adaptively and dynamically adjusted.

[0101] The optimization algorithm improvement strategy for attenuators is mainly reflected in two aspects: one is the elite preservation enhancement strategy, which adopts a dual elite library mechanism. The main library retains the globally non-dominated optimal solution, while the secondary library stores local optimal solutions. Cross-validation is performed in each generation to prevent premature convergence. The other improvement involves using simulated binary cross-validation for continuous variables (e.g., resistance and capacitance values) and single-point cross-validation for discrete variables (size design parameters) to improve search efficiency. The dual elite library uses the main library to store all design parameter information, while the secondary library stores some locally optimal design parameter information. Under normal circumstances, genetic cross-validation only occurs based on the main library. The secondary library introduces a certain degree of artificial selection, improving the score of the generated results in each generation (the introduction of the secondary library increases the proportion of locally optimal genes).

[0102] The dynamic change of weights is not limited to just one or two stages. Instead, while ensuring the focus, when the optimization gets stuck (local optimum), the weights are adaptively adjusted (slightly adjusted up or down to try to break out of the local optimum) to make the optimization more efficient.

[0103] The following section will explain the optimization process for six different types of radio frequency device modules.

[0104] Example 1 (Optimization of Passive Devices):

[0105] Application: Radio frequency bandpass filter;

[0106] Optimize the process:

[0107] Call the optimization interface of the input library, and select the RF filter netlist as the netlist file in the input parameters, such as... Figure 8As shown. The filter's optimization metrics include: passband start frequency, passband cutoff frequency, passband gain, impedance attenuation, group delay, and impedance matching. Hyperparameter information for the optimization algorithm is left blank and remains at the default value. The optimization algorithm library calls and executes the automatic optimization algorithm module for RF filter modules based on the device type. This example uses three different optimization objectives to test the optimization algorithm library, and uses an unmodified genetic algorithm and a particle swarm optimization algorithm as comparison algorithms.

[0108] The main optimization parameters to be measured in the design of RF filters include: passband start frequency, passband cutoff frequency, passband gain, impedance attenuation, group delay, and impedance matching. These six optimization objectives can be quantified into a six-dimensional parameter space, as shown in the following formula:

[0109] F = [f start ,f stop G pass A stop ,τ group Z match ]

[0110] Among these, there are significant correlations and mutual exclusions between the various objectives, such as the passband gain G. pass With stopband attenuation A stop They exhibit a nonlinear negative correlation; group delay τ group There is an inverse relationship between flatness and insertion loss; impedance matching Z match Parameter optimization affects the executable range of other parameters. Since there are certain mutually exclusive relationships among the optimization objectives of RF filters, the weighting relationships between different objectives need to be set during multi-objective optimization. In the equation, F represents the performance parameter space of the RF filter. f represents the possible expansion space of the six performance parameters: start frequency, passband cutoff frequency, passband gain, impedance attenuation, group delay, and impedance matching. In this embodiment, the circuit characteristics corresponding to different design parameters only concern the performance of the above six parameters.

[0111] Therefore, this example optimization algorithm library prioritizes the PSO optimization algorithm as the highest-priority optimization algorithm when optimizing RF filters, forming a framework for an automatic RF filter optimization algorithm based on the PSO optimization algorithm. The specific steps are as follows:

[0112] Step 1: Parameter Initialization and Netlist Resolution. Input parameters are validated, and preprocessing is performed using the netlist. The reliability of the initial variable values ​​is verified through zero-iteration simulation.

[0113] Step 2: Explore design controls using the PSO algorithm to search for the optimal design result. DRC checks filter and eliminate design schemes that violate the rules. Through a multi-condition termination mechanism, the final principle-level automatic optimization process will generate a preliminary design scheme.

[0114] Step 3: Simulate the current design using a simulation engine. Calculate the difference between the simulation results and the expected values ​​of the optimization target parameters using an evaluation function. Weighted sums are then used as the score for the current design. The evaluation function for the RF filter design is as follows:

[0115]

[0116] Here, "Score" represents the score of the current design scheme; This indicates that the scores for the six performance parameters are summed; ω i This represents the weight of the i-th performance parameter; The score for a single performance characteristic is calculated by dividing the difference between the simulated value and the target value by the maximum value of that performance characteristic, and this error is used as the score. The closer the simulation result is to the target, the lower the score. The optimization algorithm aims to find the design parameters that result in the lowest score.

[0117] For the weighting of different optimization objectives in RF filters, the weighting logic reflects the highest importance of passband performance, followed by stopband characteristics. To prevent extreme design schemes, during the evaluation of schemes, if the impedance matching deviates from the design specifications by more than 15%, the entire design is eliminated. If the group delay ripple of the scheme exceeds 100 picoseconds, a penalty is added, increasing the final score by 0.5.

[0118] Step 4: After obtaining the schematic prototype design, the results are evaluated and used as the initial optimized values ​​for the RF filter version design at the layout level. By checking the imported schematic prototype optimized design, the layout of key components within the initial layout framework is preserved, and the parameters are updated in the layout design file via a parameter setting script. The RF filter layout design with completed initial value settings requires an initial simulation, and the feasibility of the current design scheme is determined based on the simulation results.

[0119] Furthermore, since page-level simulation is extremely time-consuming, the automatic optimization algorithm library adds an auxiliary particle movement velocity correction based on gradient descent to the page-level PSO optimization algorithm. For example, the formula is:

[0120]

[0121] in, Let be the gradient (first derivative) at parameter x, and Δx be the perturbation step size of parameter x.

[0122] By using gradient information to guide the particle swarm's movement direction, convergence efficiency can be effectively accelerated, the number of simulation calls can be reduced, and thus the overall optimization time can be saved.

[0123] Step 5: The optimization algorithm library will output the final optimization results, including the GDSII file and the design scheme.

[0124] After testing, compared with traditional genetic algorithms and particle swarm optimization algorithms, the automatic optimization algorithm library in this embodiment can converge to the target faster and achieve better optimization results to a certain extent for RF filter modules. The optimization effect for impedance matching is particularly good, while there are still shortcomings in stopband attenuation. The optimization target settings and optimization results for the three rounds tested in the example are shown in Table 1. The layout-completed GDSII file output by the optimization results is shown in... Figure 9 As shown.

[0125]

[0126] Table 1. RF Filter Optimization Test Table (Example 1) (A represents the algorithm in this paper, N represents the comparative genetic algorithm)

[0127] P represents the comparison particle swarm algorithm)

[0128] Example 2 (Optimization of Passive Components):

[0129] Application: Radio frequency switching circuits;

[0130] Optimize the process:

[0131] Call the optimized interface of the input library, and select the RF switch circuit netlist file in the input parameters, such as... Figure 10 As shown. The hyperparameter information for the optimization algorithm is left blank and remains at the default value. Based on the device type, the optimization algorithm library calls and executes the automatic optimization algorithm module for RF switch circuit modules. This example uses four different optimization objectives to test the optimization algorithm library, and uses unmodified genetic algorithms and particle swarm optimization algorithms as comparison algorithms.

[0132] The interface module of the input library is called, and the netlist file in the input parameters is selected as the RF filter netlist. The optimization objectives of the RF switch circuit include four core indicators: 1dB compression point (P1dB), insertion loss (IL), isolation (ISO), and return loss (S11 / S22), forming a four-dimensional parameter space. Design constraints include impedance matching, process rules, and thermal stability. The weights are set using a possible method for adaptive dynamic adjustment logic. The automatic optimization algorithm library calls a genetic algorithm at different stages to optimize the design with different focuses, and after the weights are modified, it retains the "genes" of the previous stage as effectively as possible in the next generation of design schemes, forming multiple local optima and performing a global optimum search. The specific steps are as follows:

[0133] Step 1: Parameter Space Construction and Genetic Algorithm Encoding and Initialization. In this embodiment, after constructing the parameter space, binary encoding is used to simulate the "gene sequence," covering continuous parameters such as transistor size and matching inductance / capacitance values. Based on the netlist parsing results, 50-100 initial schemes are generated. The feasibility of the parameters is verified through zero-run simulations, and invalid individuals are screened out.

[0134] Step 2: The automatic optimization process for the two levels of the RF switching circuit must go through three stages. During the convergence optimization process, if the improvement rate of the optimal solution is <1% for 5 consecutive generations, the process switches to the next stage. After completing the three-stage optimization, if the improvement rate is <0.5% for 20 consecutive generations, the Pareto optimal solution set is output as the final solution. Stage 1 focuses on insertion loss and return loss to quickly converge to the low-loss region; Stage 2 focuses on isolation and the 1dB compression point to improve linearity and signal integrity; Stage 3 balances the four indicators and inherits the advantages of the first two stages through an elite retention strategy. Normalization is used in the overall process, and each indicator is normalized using Min-Max normalization to eliminate dimensional differences.

[0135] Step 3: Add constraint penalties to the evaluation function for optimization control. If a design violates impedance matching or process rules, the design is directly eliminated. Simultaneously, to prevent the loss of superior genes, the top 5% of the best individuals are retained in each generation.

[0136] Step 4: After obtaining the optimal solution of the schematic, map the optimal solution of the schematic to the layout and initialize the geometric parameters (transmission line width / spacing); the automatic optimization algorithm library uses a genetic algorithm to perform optimization design and performs DRC detection and thermal coupling simulation on the layout design results to ensure the reliability of the design results.

[0137] Step 5: The final output is a design result that meets the testing standards.

[0138] After testing, compared to genetic algorithms and particle swarm optimization algorithms, the automatic optimization algorithm library in this embodiment achieves the design specifications for isolation and return loss in RF switching circuits, and shows better optimization performance in return loss and 1dB compression point. The optimization target settings and results of the four rounds tested in the example are as follows:

[0139] Table 2 shows the optimized layout output GDSII file, as follows: Figure 11 As shown.

[0140]

[0141]

[0142] Table 2 Example 2 RF Switch Circuit Optimization Test Table (A represents the algorithm in this paper, N represents the comparative genetic algorithm)

[0143] P represents the comparison particle swarm algorithm)

[0144] Example 3 (Active Device Optimization):

[0145] Application: Radio frequency power amplifier;

[0146] Optimize the process:

[0147] Call the optimization interface of the input library, and select the power amplifier netlist file in the input parameters, such as... Figure 12 As shown. The hyperparameter information for the optimization algorithm is left blank and remains at the default value. Based on the device type, the optimization algorithm library calls and executes the automatic optimization algorithm module for the power amplifier module. This example uses three different optimization objectives to test the optimization algorithm library, and uses an unmodified genetic algorithm and a particle swarm optimization algorithm as comparison algorithms.

[0148] The optimization targets for RF power amplifiers include: operating start frequency, operating cutoff frequency, saturated output power, in-band flatness, power-added efficiency, and return loss. These six optimization targets can be quantified into a six-dimensional parameter space, as shown in the following formula:

[0149] F = [f start ,f stop ,P out ,PAE,P add ,S 11 ]

[0150] The RF PA module exhibits significant nonlinearity between its various targets, resulting in a difference between power-added efficiency (PAE) and output power (P). out The power-added efficiency (P0) is positively correlated but limited by transistor saturation; addThere is a negative coupling between linearity and return loss (S). 11 The optimization of the input matching network directly impacts its design. A performance balance (efficiency-linearity-bandwidth tradeoff) needs to be achieved through multi-objective optimization. For the design requirements of the RF power amplifier, the automatic optimization algorithm library selects the MOPSO algorithm as the optimal optimization algorithm in the RF power amplifier optimization process. The specific process is as follows:

[0151] Step 1: After constructing the parameter space and initializing the parameters and constraints, the Morris sensitivity analysis algorithm is used to screen the dominant parameters and fix the low-sensitivity parameters to reduce the optimization dimensionality and the number of optimizations. Simultaneously, by using an initial variation range of ±30% for transistor size parameters, linear probing is used to compress the range to the process-feasible region, with adaptive range adjustment.

[0152] Step 2: Principle layer-level optimization is divided into two stages. Stage 1 focuses on efficiency and output power, enabling the design to quickly approach the saturation region. Stage 2 balances in-band flatness and return loss, optimizing linearity. During optimization, the top 10% of non-dominated solutions are retained in each generation, implementing an elitist strategy. When the Pareto front update rate of the optimized solution is <2% for 8 consecutive generations, the optimization stage is switched, and the weights are adjusted as needed to balance global exploration and local development for better optimization results. After the two-stage optimization process is completed, if the improvement rate of the optimal solution is <1% for 15 consecutive generations, the solution is output.

[0153] Step 3: Import the optimal solution from the schematic into the layout for initial value mapping, initializing transmission line width and spacing parameters; the automatic optimization algorithm library uses the MOPSO optimization algorithm for global optimization design flow. Simultaneously, the NSGA-II optimization algorithm is used for local fine-tuning. DRC checks are performed on layout-level design schemes to prevent the output of non-compliant solutions.

[0154] Step 4: Output the final design result after verification.

[0155] After testing, the automatic optimization algorithm library in this example outperforms traditional genetic algorithms and particle swarm optimization algorithms in optimizing several final results, including saturated output power, for RF power amplifiers. The optimization objective settings and results for the three rounds tested in this example are shown in Table 3. The optimized layout output, a completed GDSII file, is shown below. Figure 13 As shown.

[0156]

[0157]

[0158] Table 3 Example 3 RF Power Amplifier Optimization Test Table (A represents the algorithm in this paper, N represents the comparison genetic algorithm, and P represents the comparison particle swarm algorithm)

[0159] Example 4 (Active Device Optimization):

[0160] Application: Radio frequency driver amplifier;

[0161] Optimize the process:

[0162] Call the optimization interface of the input library, and select the RF driver amplifier circuit netlist file in the input parameters, such as... Figure 14 As shown. The hyperparameter information for the optimization algorithm is left blank and remains at the default value. Based on the device type, the optimization algorithm library calls and executes the automatic optimization algorithm library for the RF driver amplifier module. This example uses four different optimization objectives to test the optimization algorithm library, and uses an unmodified genetic algorithm and particle swarm optimization algorithm as comparison algorithms.

[0163] The optimization objectives of the RF driver amplifier consist of a four-dimensional design space comprised of four search objectives: lower gain limit, upper in-band flatness limit, lower input return loss limit, and upper output return loss limit. The automatic optimization algorithm library uses NSGA-II as the optimal algorithm for global optimum search. The specific steps are as follows:

[0164] Step 1: Perform dynamic netlist construction and parameter preprocessing.

[0165] Step 2: In the principle layer-level optimization stage, the dominant parameters are identified by the Morris global sensitivity analysis algorithm, and the design range of variable design parameters is dynamically compressed to eliminate invalid design areas.

[0166] Step 3: Through a two-stage dynamic weight allocation, different optimization focuses are achieved. The first stage focuses on gain and bandwidth, allowing the optimization algorithm to quickly expand the frequency band. The second stage balances return loss and stability, suppressing oscillation risk. A constraint penalty mechanism is added to eliminate design schemes with poor stability. After completing the two-stage optimization, if the update rate of the optimal solution is <0.5% for 20 consecutive generations, the solution set is output.

[0167] Step 4: In the layout-level optimization stage, the design scheme from the schematic layer is imported as the initial value into the layout design, and the relevant parameters of the microstrip lines are initialized. The automatic optimization algorithm library, while calling the genetic algorithm for global optimization, introduces a gradient descent algorithm to adjust the direction of parameter changes and performs DRC checks on the layout-level scheme to avoid outputting non-compliant design schemes.

[0168] Step 5: Output the final design result after verification.

[0169] Testing showed that the automatic optimization algorithm library in this example achieves better optimization results for RF driver amplifiers compared to traditional genetic algorithms and particle swarm optimization algorithms. The optimization objective settings and results for the four rounds tested in this example are shown in Table 4. The completed layout GDSII file output by the optimization results is shown in the figure below. Figure 15 As shown.

[0170]

[0171] Table 4 Example 4 RF Driver Amplifier Optimization Test Table (A represents the algorithm in this paper, N represents the comparison genetic algorithm, P represents the comparison particle swarm algorithm)

[0172] Example 5 (Active Device Optimization):

[0173] Application: Radio frequency low noise amplifier;

[0174] Optimize the process:

[0175] Call the optimization interface of the input library, and select the RF low-noise amplifier circuit netlist file in the input parameters, such as... Figure 16 As shown, the hyperparameter information for the optimization algorithm is left blank and remains at the default value. Based on the device type, the optimization algorithm library calls and executes the automatic optimization algorithm library for the RF low-noise amplifier module. This example uses four different optimization objectives to test the optimization algorithm library, and uses unmodified genetic algorithms and particle swarm optimization algorithms as comparison algorithms.

[0176] The optimization objectives of the RF low-noise amplifier consist of a six-dimensional design space comprised of six optimization goals: noise figure, lower gain limit, output 1dB compression point, upper limit of in-band flatness, lower limit of input return loss, and upper limit of output return loss. The automatic optimization algorithm library uses NSGA-II as the optimal algorithm for global optimum search, and the specific steps are as follows:

[0177] Step 1: Perform format validation on the input LNA topology netlist file, variable parameter list, and their initial value range. Also, preset hard parameters.

[0178] Step 2: In the principle layer-level optimization stage, the dominant parameters are identified by the Morris global sensitivity analysis algorithm, and the design range of variable design parameters is dynamically compressed to eliminate invalid design areas.

[0179] Step 3: Through a two-stage dynamic weight allocation, optimization is guided by different emphases. The first stage focuses on noise figure and gain; the second stage balances linearity and in-band flatness. A constraint penalty mechanism is added to eliminate design schemes with poor stability. After completing the two-stage optimization, if the noise figure improvement rate of the optimal solution for 20 consecutive generations is <0.1% and the S21 fluctuation is <0.5dB, then the solution set is output.

[0180] Step 4: In the layout-level optimization stage, the design scheme from the schematic layer is imported into the layout design as the initial value, and the relevant parameters of the microstrip lines are initialized. The automatic optimization algorithm library also uses a genetic algorithm for optimization design and performs DRC checks on the layout-level schemes to avoid the output of non-compliant design schemes.

[0181] Step 5: Output the final design result after verification.

[0182] After testing, compared to genetic algorithms and particle swarm optimization algorithms, the automatic optimization algorithm library in this example achieves better optimization results for RF low-noise amplifiers. The optimization objective settings and results for the four rounds tested in this example are shown in Table 5. The completed layout GDSII file output by the optimization results is shown in... Figure 17 As shown.

[0183]

[0184]

[0185] Table 5 Example 5 RF Low Noise Amplifier Optimization Test Table (A represents the algorithm in this paper, N represents the comparison genetic algorithm, P represents the comparison particle swarm algorithm)

[0186] Example 6 (Active Device Optimization)

[0187] Application: Radio frequency digitally controlled attenuator

[0188] Optimize the process:

[0189] Call the optimized interface of the input library, and select the RF digitally controlled attenuator circuit netlist file in the input parameters, such as... Figure 18 As shown. The hyperparameter information for the optimization algorithm is left blank and remains at the default value. Based on the device type, the optimization algorithm library calls and executes the automatic optimization algorithm library for the RF digitally controlled attenuator module. This example uses a set of different optimization objectives to optimize and test the optimization algorithm library, and uses unmodified genetic algorithms and particle swarm optimization algorithms as comparison algorithms.

[0190] The optimization objectives of the RF digitally controlled attenuator consist of a four-dimensional design space comprised of four optimization goals: upper limit of insertion loss, attenuation range, lower limit of return loss, and upper limit of attenuation accuracy. To address the design requirements of the RF digitally controlled attenuator, this embodiment employs the SGA-II algorithm to construct a phased optimization architecture, combining a dynamic weight adjustment mechanism with impedance matching constraint processing technology to build a hierarchical framework. The specific steps are as follows:

[0191] Step 1: Dynamic construction of netlist and parameter preprocessing.

[0192] Step 2: In the principle layer-level optimization stage, the dominant parameters are identified by the Morris global sensitivity analysis algorithm, and the design range of variable design parameters is dynamically compressed to eliminate invalid design areas. The automatic optimization algorithm library uses an improved genetic algorithm for global optimization.

[0193] Step 3: A two-stage dynamic weight allocation is used to achieve optimization with different focuses. The first stage focuses on optimizing insertion loss and return loss, while the later stages focus on fine-tuning attenuation accuracy and phase error. During the iteration process, the proportional weights of different target parameters are adaptively and dynamically adjusted. Impedance matching constraints are embedded in the constraints to improve the impedance matching degree of the design scheme; a constraint penalty mechanism is added. The optimization process incorporates variable asynchronous length and variability rate for highly sensitive parameters; a dual elite library mechanism is adopted, with the main library retaining globally non-dominated solutions and the secondary library storing locally optimal solutions, and cross-validation updates are performed every 5 generations. After completing the two-stage optimization, if the attenuation accuracy fluctuation is <0.1dB for 25 consecutive generations, a scheme set is output.

[0194] Step 4: In the layout-level optimization stage, the design scheme from the schematic layer is imported into the layout design as the initial value, and the relevant parameters of the microstrip lines are initialized. The automatic optimization algorithm library also uses a genetic algorithm for optimization design and performs DRC checks on the layout-level schemes to avoid the output of non-compliant design schemes.

[0195] Step 5: Output the final design result after verification.

[0196] Testing showed that, compared to traditional algorithms, the automatic optimization algorithm library in this example performs better in optimizing return loss for RF digitally controlled attenuation modules. The optimization target settings and results tested in this example are as follows:

[0197] Table 6 shows the optimized layout output, specifically the completed GDSII file. Figure 19 As shown.

[0198]

[0199] Table 6 Example 6 RF Low Noise Amplifier Optimization Test Table (A represents the algorithm in this paper, N represents the comparison genetic algorithm, P represents the comparison particle swarm algorithm)

[0200] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. An automatic hyperparameter optimization algorithm library for radio frequency devices, characterized in that, include: The interface module is used to configure the target characteristic parameters and optimization objects of different RF devices; The targeted optimization algorithm module includes: multiple automatic optimization algorithm units for functional devices, which, based on the target feature parameters and the optimization object, call the automatic optimization algorithm unit that matches the optimization object to execute the automatic optimization process; the automatic optimization algorithm unit optimizes the device parameters and finally outputs the optimal design parameters. The multimodal optimization engine module is embedded in each automatic optimization algorithm unit of the directional optimization algorithm module. It includes multiple optimization algorithm units that perform optimization operations based on the target feature parameters of the input RF device through the corresponding automatic optimization algorithm unit for the functional device.

2. The RF device hyperparameter automatic optimization algorithm library according to claim 1, characterized in that, The multimodal optimization engine module includes: a gradient descent algorithm subunit, a genetic algorithm subunit, and a particle swarm optimization algorithm subunit; based on the target feature parameters of the input RF device, the automatic optimization algorithm unit for functional devices calls the corresponding subunit to perform optimization operations during optimization.

3. The RF device hyperparameter automatic optimization algorithm library according to claim 2, characterized in that, The multimodal optimization engine module also includes: a dynamic weight adjustment algorithm and a sensitivity analysis algorithm; the dynamic weight adjustment algorithm is used to adjust the weights of different optimization objectives; The sensitivity analysis algorithm is used to analyze the sensitivity of design parameters; the optimization algorithm library also includes a simulation engine for simulating key circuits in the optimization process.

4. The RF device hyperparameter automatic optimization algorithm library according to claim 1, characterized in that, The automatic optimization algorithm unit for functional devices optimizes target feature parameters at both the schematic layer level and the layout layer level. During the optimization process, the results of the principle layer-level optimization are input into the layout layer-level optimization, and the optimized GDSII format file is output as the final output result.

5. The RF device hyperparameter automatic optimization algorithm library according to claim 4, characterized in that, When the optimization algorithm library performs principle layer-level optimization, the steps include: Obtain the netlist file of the RF device, verify the integrity of the netlist, and prompt the user when the verification is complete; Initialize the optimization parameters and netlist objects, and discretize the parameter space; For the design parameter information, ADS co-simulation is used, and optimization algorithms are employed for multi-objective analysis and optimization. The optimization convergence result is evaluated. If it meets the design requirements, the optimal design scheme is output. If it does not meet the design requirements, the optimization algorithm optimization space is updated, and the simulation and optimization process is re-executed.

6. The RF device hyperparameter automatic optimization algorithm library according to claim 1, characterized in that, The directional optimization algorithm module, based on the optimization object, includes automatic optimization algorithm units for functional devices, such as: automatic optimization algorithm unit for RF filters, automatic optimization algorithm unit for RF switching circuits, automatic optimization algorithm unit for RF power amplifiers, automatic optimization algorithm unit for RF low-noise amplifiers, automatic optimization algorithm unit for RF drive amplifiers, and automatic optimization algorithm unit for RF digitally controlled attenuators.

7. The RF device hyperparameter automatic optimization algorithm library according to claim 6, characterized in that, The automatic optimization algorithm unit for functional devices in the directional optimization algorithm module includes the following steps when performing the optimization operation: Initialize the target characteristic parameters of the RF device and resolve the netlist; The principle layer of the radio frequency device is optimized. During the optimization, the target characteristic parameters are adjusted in multiple stages according to the design focus of the radio frequency device to generate an initial design scheme. Layout-level optimization of RF devices: The initial design scheme is imported into the layout design as the initial value, the corresponding target feature parameters are initialized, and the corresponding optimization algorithm unit in the multi-modal optimization engine module is called for optimization. The final design result is output after verification.

8. The RF device hyperparameter automatic optimization algorithm library according to claim 7, characterized in that, In the directional optimization algorithm module, the automatic optimization algorithm unit for radio frequency filters uses the PSO algorithm for optimization search when performing principle layer-level optimization and layout layer-level optimization. In layout layer-level optimization, particle movement velocity correction is assisted based on gradient descent algorithm.

9. The RF device hyperparameter automatic optimization algorithm library according to claim 7, characterized in that, In the directional optimization algorithm module, the automatic optimization algorithm unit for RF driver amplifiers, the automatic optimization algorithm unit for RF low noise amplifiers, and the automatic optimization algorithm unit for RF digitally controlled attenuators use the Morris global sensitivity analysis algorithm to identify dominant parameters and dynamically compress the design range of variable design parameters when performing principle layer-level optimization. Optimize design focus through dynamic weight allocation.

10. The RF device hyperparameter automatic optimization algorithm library according to claim 6, characterized in that, In the directional optimization algorithm module, after obtaining the initial design scheme, the automatic optimization algorithm unit for RF filters uses a simulation engine to simulate the current design scheme, and uses an evaluation function to calculate the difference between the simulation results and the expected values ​​of the optimization target parameters. The weighted sums are then used as the score of the current scheme.

Images