A microwave device design method, system application

Abstract

The application provides a microwave device design method, system and application, through double-layer sample space construction and three-level progressive sample screening, the most valuable key samples for model optimization are accurately selected, the sample information value is maximized, and the redundant simulation is greatly reduced. In addition, the dynamic proportional splitting of the training set and the validation set, the adaptive weight fusion sample screening strategy can flexibly adapt to microwave device design variables of different dimensions (from low-dimensional simple structure to high-dimensional complex integrated device), without manual adjustment of parameters, and has wider adaptability. The method has good universality and practicality, and can effectively improve the precision and efficiency of microwave device design. This method not only reduces the calculation cost, but also improves the reliability and stability of the design process, and is not only suitable for the field of microwave device design, but also can be popularized to other technical fields which need complex simulation and optimization calculation.

Description

Technical Field

[0001] This invention relates to the field of microwave device design technology, and more specifically, to a microwave device design method, system, and application. Background Technology

[0002] With the rapid development of modern communications, radar, satellites, and other fields, microwave devices are being used more and more widely, and the demand for miniaturization, wide bandwidth, and high performance is becoming increasingly urgent. Traditional microwave device design mainly relies on engineers' experience, equivalent circuit analysis, or full-wave electromagnetic simulation iteration, which is time-consuming. Therefore, there is an urgent need for an intelligent optimization technique that balances efficiency, accuracy, and reliability. Surrogate model-assisted multi-objective optimization algorithms are one of the effective methods for solving expensive multi-objective optimization problems.

[0003] Proxy models replace complex physical simulations by constructing a mathematical model that approximates the true objective function, thereby significantly reducing computational costs and improving design efficiency. However, as the number of variables and objectives in the optimization problem increases, higher demands are placed on improving the efficiency of multi-objective optimization algorithms in searching for the optimal solution set. The sample update strategy is a crucial step in improving model performance and finding the optimal solution set; the selected sample points largely determine the accuracy level of the Proxy model. Summary of the Invention

[0004] The purpose of this invention is to reduce the design cost of microwave devices and shorten the design cycle.

[0005] The present invention solves the above-mentioned technical problems through the following technical means: A microwave device design method, comprising: S1. Construct the candidate sample pool SP: N A number of candidate samples constitute The matrix, the candidate samples include microwave devices k Dimensional design variables x Constructing the initial dataset DAS: Calculating the true performance parameters corresponding to the design variables of each initial sample obtained by Latin hypercube sampling through simulation; S2. Divide the initial dataset DAS into a training set TRS and a validation set TES; S3. Construct a target function forest model for the design objective on the training set TRS, and output the effect importance measure EIM; and use this model to calculate the basic data of the prediction error on the validation set TES, and construct a misjudgment forest model for the design objective. S4. Select effective samples from the candidate sample pool SP, that is, select the first samples with the smallest response values ​​using the target function forest model; select confidence differences using the misclassification forest model. The largest number of second samples; based on the effect importance metric EIM, select several third samples that can effectively supplement the blind spots of the current dataset DAS coverage; after selecting the top Q samples from the first, second, and third samples, the key samples are formed. S5: After removing redundant samples from the key samples using the cosine similarity metric, calculate the true response value and add it to the initial dataset. Simultaneously update TRS and TES, return to step S3, retrain the target function forest model and the misjudgment forest model, and repeat the closed-loop process of S3-S5 until the preset stopping condition is met; thus obtaining the target dual forest model. S6: The target amphibious forest model is used to assist in the design of microwave devices.

[0006] Furthermore, the process of constructing the target function forest model in step S3 is as follows: adopting the diversity forest algorithm framework, setting the number of decision trees, and the feature sampling ratio of each tree is the square root of the total number of features; adopting a bivariate splitting strategy, that is, each node finds the optimal split point based on feature pairs that have not appeared in the current path.

[0007] Furthermore, the process of constructing the misjudgment forest model in step S3 is as follows: Based on TES, the prediction error baseline data of the target function forest model is calculated, including the absolute difference between the true response value and the predicted response value, and the variance of the prediction results; using sample features as input, and dynamic confidence difference... To output, construct a misclassification forest model. ; The calculation uses an adaptive weighting formula. ,in It is the variance of the prediction results. This is the actual response value. It is a predicted response value. , For dynamic weights, Based on the error distribution characteristics in TES, the misjudgment forest model is adaptively adjusted, and during training, the same hyperparameter settings as the target function forest model are used, such as the number of nodes and the maximum depth of the tree.

[0008] Furthermore, the method for screening key samples in step S4 is as follows: First-level screening: The first sample is selected using the target function forest model. Predict the response value of the candidate sample pool SP, select the Top-K samples with the smallest response values, and focus on the target optimization direction; Secondary screening: The second sample is selected through a misjudgment forest model. Confidence difference in predicting candidate sample pool SP Select The largest Top-M samples accurately locate regions of high uncertainty; Three-level screening: The third sample is based on the EIM output by the target function forest model to identify high-importance feature pairs. The top-P samples in the candidate sample pool SP whose value combinations of the feature pairs do not fully cover the dataset DAS are selected to supplement key information on feature interactions. An adaptive weight fusion strategy is adopted to prioritize the three-level screening samples: the weight of the third sample is fixed at 0.2; the weight of the first sample is 0.4 and the weight of the second sample is 0.4 in the initial iteration; if the accuracy is insufficient, the weight of the second sample is increased to 0.6; if the efficiency is insufficient, the weight of the first sample is increased to 0.6, so as to ensure that the sample screening is accurately matched with the model requirements.

[0009] The present invention also provides a microwave device design system, comprising: Two-layer sample construction module: Constructing the candidate sample pool SP: N Each candidate sample constitutes The matrix, the candidate samples include microwave devices k Dimensional design variables x Constructing the initial dataset DAS: Calculating the true performance parameters corresponding to the design variables of each initial sample obtained by Latin hypercube sampling through simulation; Dataset partitioning module: Partitions the initial dataset DAS into a training set TRS and a validation set TES; Model building module: Constructs a target function forest model for the design objective on the training set TRS, outputs the effect importance measure EIM; and uses this model to calculate the basic data of prediction error on the validation set TES, and constructs a misjudgment forest model for the design objective. Key sample selection module: Selects valid samples from the candidate sample pool SP, i.e., selects the first few samples with the smallest response values ​​using the target function forest model; and selects samples using the misclassification forest model. The largest number of second samples; based on the effect importance metric EIM, select several third samples that can effectively supplement the blind spots of the current dataset DAS coverage; after selecting the top Q samples from the first, second, and third samples, the key samples are formed. Dataset update module: After removing redundant samples from the key samples using the cosine similarity metric, the true response value is calculated and added to the initial dataset. The TRS and TES are updated synchronously. The module returns to the model building module and retrains the target forest model and the misjudgment forest model until the preset stopping condition is met; thus obtaining the target dual forest model. Microwave Device Design Module: Used to assist in microwave device design by employing the target amphibious forest model.

[0010] Furthermore, the process of constructing the target function forest model in the model construction module is as follows: adopting the diversity forest algorithm framework, setting the number of decision trees, and the feature sampling ratio of each tree is the square root of the total number of features; adopting a bivariate splitting strategy, that is, each node finds the optimal split point based on feature pairs that have not appeared in the current path.

[0011] Furthermore, the process of constructing the misjudgment forest model in the model construction module is as follows: Based on TES, the prediction error baseline data of the target function forest model is calculated, including the absolute difference between the true response value and the predicted response value, and the variance of the prediction results; using sample features as input, and dynamic confidence difference... To output, construct a misclassification forest model. ; The calculation uses an adaptive weighting formula. ,in It is the variance of the prediction results. This is the actual response value. It is a predicted response value. , For dynamic weights, Based on the error distribution characteristics in TES, the misjudgment forest model is adaptively adjusted, and during training, the same hyperparameter settings as the target function forest model are used, such as the number of nodes and the maximum depth of the tree.

[0012] Furthermore, the key sample screening method in the key sample screening module is as follows: First-level screening: The first sample is selected using the target function forest model. Predict the response value of the candidate sample pool SP, select the Top-K samples with the smallest response values, and focus on the target optimization direction; Secondary screening: The second sample is selected through a misjudgment forest model. Confidence difference in predicting candidate sample pool SP Select The largest Top-M samples accurately locate regions of high uncertainty; Three-level screening: The third sample is based on the EIM output by the target function forest model to identify high-importance feature pairs. The top-P samples in the candidate sample pool SP whose value combinations of the feature pairs do not fully cover the dataset DAS are selected to supplement key information on feature interactions. An adaptive weight fusion strategy is adopted to prioritize the three-level screening samples: the weight of the third sample is fixed at 0.2; the weight of the first sample is 0.4 and the weight of the second sample is 0.4 in the initial iteration; if the accuracy is insufficient, the weight of the second sample is increased to 0.6; if the efficiency is insufficient, the weight of the first sample is increased to 0.6, so as to ensure that the sample screening is accurately matched with the model requirements.

[0013] This invention also provides a dual-mode forest model training method. Utilizing the aforementioned microwave device design method, after each round of key sample screening, the datasets DAS, TRS, and TES are dynamically updated. Based on the updated TRS and TES, the target function forest model and the misjudgment forest model are retrained respectively, thereby achieving collaborative optimization of the dual-mode forest model.

[0014] The present invention also provides a method for manufacturing a microwave device, wherein the design parameter values ​​corresponding to the ideal design target of the microwave device are obtained by using the above method and the microwave device is manufactured accordingly.

[0015] The advantages of this invention are: This invention constructs a collaborative update framework of "dual-model driven, multi-strategy fusion, and parameter adaptation." A dual-forest model (target function forest + misjudgment forest) is built, achieving both high-precision approximation of the design objective function and quantifying prediction uncertainties. Combined with Effect Importance Metric (EIM) to capture feature interaction effects, the optimization results are closer to the global optimum. Secondly, through a two-layer sample space construction and a three-level progressive sample selection, the most valuable key samples for model optimization are accurately selected, maximizing sample information value and significantly reducing redundant simulations. Furthermore, the dynamic proportional splitting of the training and validation sets and the adaptive weight fusion sample selection strategy can flexibly adapt to microwave device design variables of different dimensions (from low-dimensional simple structures to high-dimensional complex integrated devices) without manual parameter adjustment, thus broadening its adaptability. This method has good versatility and practicality, effectively improving the accuracy and efficiency of microwave device design. This method not only reduces computational costs but also improves the reliability and stability of the design process. It is not only applicable to the field of microwave device design but can also be extended to other technical fields requiring complex simulation and optimization calculations. Attached Figure Description

[0016] Figure 1 This is a flowchart of the microwave device design method in an embodiment of the present invention; Figure 2 This is the CST simulation model of the exponential four-ridged trumpet in the embodiments of the present invention; Figure 3 This is the target optimization result for the reflection coefficient of the exponential four-ridged horn in this embodiment of the invention; Figure 4 This is the target optimization result for the isolation of the exponential four-ridged horn in this embodiment of the invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This invention discloses a microwave device design method, the overall calculation process of which is as follows: Figure 1 As shown: a two-layer sample space is constructed; the initial dataset is dynamically split according to feature dimensions to build a dual-layer forest model; several samples are selected from the candidate sample pool in a three-level progressive manner; after supplementing samples and removing redundancy, the true response value is calculated, added to the DAS, and the TRS and TES are updated; the model is retrained until the design goal is achieved or the maximum number of simulations is reached. In this embodiment, the dataset consists of the device's structural material parameters X (design variables) and the corresponding performance parameters Y (design goals); the design goals are the device's performance parameters, such as (reflection coefficient, efficiency, isolation, etc.). The specific steps of this embodiment are as follows: Step 1: Construct a two-layer sample space; Candidate sample pool SP: N random samples are generated in the design variable space using a dynamic hierarchical sampling iteration method.

[0019] Randomly set the initial sample state Let the number of iterations N be set. From the initially set Gaussian distribution Randomly select a candidate sample ; Calculate the acceptance rate Generate uniform random numbers ;like ,but ,otherwise, ,make This process continues until N iterations are completed, at which point only the converged samples are retained as valid samples.

[0020] (1) Each sample is one dimensional vector ( X represents a set of design parameter values ​​(such as geometric dimensions, material properties, etc.) for microwave devices. The candidate sample pool SP is a normalized matrix containing only the design parameter information X. These samples approximately follow a Gaussian distribution, forming a high-quality candidate sample pool used as the initial population for subsequent surrogate model updates or optimization algorithm point additions.

[0021] Initial dataset DAS: Latin hypercube sampling (LHS) was used to ensure uniform coverage of the sample space, and cosine similarity was used to remove redundant samples. indivual( The initial sample (which represents the dimensions of the design variables) is used to calculate the actual performance parameters corresponding to each set of design parameters in the initial sample through simulation, forming a design parameter matrix and a performance parameter matrix. These two matrices construct the initial dataset (DAS). That is, the initial dataset DAS contains two matrices, one of which is of size... It includes the structural material parameters X (design variables) of the device, and another size is , where Y is the corresponding performance parameter (design target), and m is the number of design targets.

[0022] The design parameters of microwave devices typically depend on the specific device type (such as filters, antennas, couplers, resonators, etc.), but can be summarized into the following categories of general design parameters. These parameters collectively determine the electromagnetic performance of the device (i.e., the aforementioned design objective Y; such as operating frequency, bandwidth, insertion loss, return loss, isolation, etc.).

[0023] The following are common design parameter categories and specific examples in microwave device design: 1. Geometric Dimensions This is the most direct physical design variable, directly affecting the distribution and transmission characteristics of the electromagnetic field.

[0024] Length: Physical length of microstrip lines, striplines, resonators, and coupling segments. Width: Width of the transmission line.

[0025] Gap / Spacing: The width of the gap between coupled structures.

[0026] Radius / diameter: The dimensions of a circular resonant cavity, cylindrical conductor, or via.

[0027] Thickness: Thickness of the dielectric substrate.

[0028] Groove dimensions: the length and width of the groove in the defective structure.

[0029] Step dimensions: The length and width of different segments in an impedance transformer.

[0030] 2. Material property parameters The material properties of the substrate and conductors form the hardware foundation of the design.

[0031] Relative permittivity.

[0032] Loss angle tangent.

[0033] Medium loss.

[0034] Substrate thickness.

[0035] Electrical conductivity of metals.

[0036] Metal thickness.

[0037] 3. Electrical / Circuit Parameters These are usually design goals or intermediate variables, but in some optimization frameworks they may also be used as indirect design variables.

[0038] Characteristic impedance.

[0039] Electrical length.

[0040] Coupling coefficient.

[0041] External quality factors.

[0042] Resonant frequency.

[0043] 4. Topology / Structure Parameters These parameters determine the connection method and spatial layout of the devices.

[0044] Number of branch lines: such as the number of branches in a branch line coupler.

[0045] Resonator order: The number of resonators in the filter (determines the filter order).

[0046] Via location: The location and array distribution of grounding vias (affecting grounding performance and parasitic parameters).

[0047] Bending radius: The bending radius at the corner of the transmission line (used to reduce discontinuity reflections).

[0048] Patch size: The length and width of the radiating patch in a microstrip antenna.

[0049] 5. Boundary and Port Parameters Port location: The location of the feed point (affects impedance matching, such as the offset of the antenna feed point).

[0050] Port width: The width of the transmission line at the port (must be compatible with an external testing system).

[0051] Shielding box dimensions: length, width, and height of the encapsulation cavity (affecting cavity resonance and coupling).

[0052] 6. Special structural parameters (for specific devices) Filter: hairpin wire dimensions, diameter / height of dielectric resonator, finger length of interdigital structure.

[0053] Antenna: Patch chamfer size (circular polarization), slot length, parasitic element spacing.

[0054] Power divider: The value of the isolation resistor (Wilkinson power divider).

[0055] Ferrite devices: bias magnetic field strength, ferrite size.

[0056] In designing microwave devices, geometry is the most important optimization variable, while material parameters are often fixed after the initial design. The optimization process typically involves adjusting the geometry to meet specific electrical performance specifications, given the material and topology.

[0057] Step 2: Determine the initial dataset and divide it accordingly; The initial dataset DAS is used to determine if a design target that meets the requirements exists. If it does, the optimal solution is output in the DAS dataset; otherwise, a dynamic scaling strategy is used to construct the target training set (TRS) and the error validation set (TES). Both the TRS and TES contain two matrices: one containing the device's structural material parameters X (design variables), and the other containing the corresponding performance parameters Y (design targets). When the feature dimension is ≤10, the matrix is ​​split in an 8:2 ratio, meaning the TRS contains X (…). ) and Y ( Two matrices, TES contains X ( ) and Y ( Two matrices; when the feature dimension is between 10 and 20, they are split in a 7:3 ratio, meaning the TRS contains X( ) and Y ( Two matrices, TES contains X ( ) and Y ( Two matrices; when the feature dimension > 20, they are split in a 6:4 ratio, that is, TRS contains X ( ) and Y ( Two matrices, TES contains X ( ) and Y ( Two matrices. This dynamic scaling balances the representativeness of the training set with the evaluation reliability of the validation set in high-dimensional data.

[0058] The optimized four-ridged trumpet model has two expected goals: (1) reflection coefficient S1(1)S1(1)≤-10dB (2) isolation S2(1)S1(1)≤-30dB. If there is already a set of X in the initial dataset that makes the corresponding Y meet the expected goal, the set of X is directly output; if not, the database is randomly split according to the preset ratio.

[0059] Step 3: Construct an amphibious forest model; Constructing a target forest model based on TRS ( To approximate each objective function, a diversity forest algorithm framework is employed, with 150-250 decision trees. The feature sampling ratio for each tree is the square root of the total number of features. A bivariate splitting strategy is used, where each node searches for the optimal split point based on feature pairs not yet present in the current path. This significantly reduces the number of candidate split point evaluations for each node, enabling the modeling of quantitative and qualitative interaction effects between features. During training, the Effect Importance Metric (EIM) is output to identify feature pairs with high predictive relevance, providing a basis for subsequent sample selection.

[0060] The prediction error base data for the target function forest model is calculated based on TES, including the absolute difference between the true and predicted response values, and the variance of the prediction results (reflecting the prediction confidence). Sample features are used as input, and dynamic confidence differencing is employed. For output, construct a misclassification forest model ( ). The calculation uses an adaptive weighting formula (2), where It is the variance of the prediction results. This is the actual response value. It is a predicted response value. , For dynamic weights ( (Adaptively adjusts according to the error distribution characteristics in TES, when the error distribution is discrete) When the error distribution is concentrated During training, the misjudgment forest model uses the same hyperparameter settings as the target function forest model, such as the number of nodes and the maximum tree depth, to ensure the collaborative compatibility of the dual forest model.

[0061] (2) Step 4: Three-level progressive sample screening; Select several samples from the candidate sample pool SP that show promise for improving the quality of the surrogate model: Primary screening (goal-oriented): using the target function forest model ( Predict the response value of the candidate sample pool SP, select the Top-K samples with the smallest response values ​​(K=5-8, adaptively adjusted according to the current sample size), focus on the target optimization direction, and define them as the first sample; Secondary screening (error-oriented): This involves filtering through misjudgments in the forest model (…). ) Confidence difference in predicting candidate sample pool SP Select The top M largest samples (M=3-5) are used to accurately locate the high uncertainty region and are defined as the second sample. Three-level screening (interaction-oriented): Based on the EIM output of the target function forest model, identify high-importance feature pairs, select the top-P (P=2-4) samples in the candidate sample pool SP whose feature pair value combinations do not fully cover the dataset DAS, supplement the key information of feature interaction, and define them as third samples.

[0062] An adaptive weight fusion strategy is adopted to prioritize the three-level screening samples: the weight of the third sample is fixed at 0.2; the weight of the first sample is 0.4 and the weight of the second sample is 0.4 in the initial iteration; if the accuracy is insufficient, the weight of the second sample is increased to 0.6; if the efficiency is insufficient, the weight of the first sample is increased to 0.6, so as to ensure that the sample screening is accurately matched with the model requirements.

[0063] Step 5: After removing redundant samples from the top Q priority key samples obtained through screening using the cosine similarity metric, calculate the true response values ​​and supplement them to the initial dataset (DAS). Simultaneously update TRS and TES, return to step 3, retrain the target function forest model and the misclassification forest model, and repeat the closed-loop process of "training-judgment-screening-supplementation" until the preset stopping condition is met (there is a design goal that meets the requirements or the maximum number of simulations is reached).

[0064] To better understand the above method, the invention will be further described in detail below with reference to the accompanying drawings and a specific embodiment—the application of the amphibious forest model sample update method to the exponential quadrilateral trumpet design. The described embodiment is only one embodiment of the invention, and not all embodiments.

[0065] The exponential quadrilateral horn design has 12 design variables. The optimization design objectives are reflection coefficient S1(1)S1(1)≤-10dB and isolation S2(1)S1(1)≤-30dB in the 2-10GHz range.

[0066] The specific implementation steps are as follows: Step 1: Generate 10 variables in the design variable space using the dynamic hierarchical sampling iteration method. 6 A random sample; generated by combining LHS and cosine similarity metrics. 1 sample point; calculate the true response to obtain the initial dataset DAS.

[0067] Step 2: Determine if there is a solution in the initial dataset DAS that meets the requirements. If it does, output the current optimal solution. Otherwise, split the samples in DAS into a training set TRS (25 samples) and a validation set TES (11 samples) in a 7:3 ratio.

[0068] Step 3: Construct target function forest models for the two design objectives, reflection coefficient and isolation, on the training set TRS, setting the number of decision trees to 150 and outputting the effect importance metric EIM; and use this model to calculate the basic data of prediction error on the validation set TES, and construct misclassification forest models for the two objectives.

[0069] Step 4: Select several samples from the candidate sample pool SP that are expected to improve model quality, and select the 6 samples with the smallest response values ​​using the target function forest model; select the samples with the smallest response values ​​using the misclassification forest model. The four largest samples; based on the effect importance metric EIM, two samples that can effectively supplement the blind spots of the current dataset DAS coverage are selected.

[0070] Step 5: After filtering key samples using the adaptive weight fusion strategy and removing redundant samples using the cosine similarity metric, calculate the true response values ​​and add them to the initial dataset (DAS). Simultaneously update TRS and TES, return to Step 3, retrain the target function forest model and the misclassification forest model, and repeat the closed-loop process of "training-judgment-filtering-supplementation" until the preset stopping condition is met (there is a design goal that meets the requirements or the maximum number of simulations is reached).

[0071] The final optimization process involved 5 iterations, adding points 20 times based on the initial design points to find the design target that met the requirements. At this point, the reflection coefficient S1(1)S1(1)≤-10dB and the isolation S2(1)S1(1)≤-45dB in the 2-10GHz range were found. Figures 3-4 .

[0072] When using the amphibious forest model sample update method to assist in the design of an exponential quadrilateral horn, a total of 36+20 electromagnetic full-wave simulations were performed for the initial design points and the points added during optimization. In contrast, when using a proxy model without screening to assist in the design of an exponential quadrilateral horn, a total of 36+53 electromagnetic full-wave simulations were performed for the initial design points and the points added during optimization. The amphibious forest model sample update method effectively screens the sample set and candidate solutions, reduces unnecessary calculations, and improves the efficiency of parameter optimization.

[0073] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A microwave device design method, characterized in that, include: S1. Construct the candidate sample pool SP: N A number of candidate samples constitute The matrix, the candidate samples include microwave devices k Dimensional design variables x Constructing the initial dataset DAS: Calculating the true performance parameters corresponding to the design variables of each initial sample obtained by Latin hypercube sampling through simulation; S2. Divide the initial dataset DAS into a training set TRS and a validation set TES; S3. Construct a target function forest model for the design objective on the training set TRS, and output the effect importance measure EIM; and use this model to calculate the basic data of the prediction error on the validation set TES, and construct a misjudgment forest model for the design objective. S4. Select effective samples from the candidate sample pool SP, that is, select the first samples with the smallest response values ​​using the target function forest model; select confidence differences using the misclassification forest model. The largest number of second samples; based on the effect importance metric EIM, select several third samples that can effectively supplement the blind spots of the current dataset DAS coverage; after selecting the top Q samples from the first, second, and third samples, the key samples are formed. S5: After removing redundant samples from the key samples using the cosine similarity metric, calculate the true response value and add it to the initial dataset. Simultaneously update TRS and TES, return to step S3, retrain the target function forest model and the misjudgment forest model, and repeat the closed-loop process of S3-S5 until the preset stopping condition is met. Obtain the target amphibious forest model; S6: The target amphibious forest model is used to assist in the design of microwave devices.

2. The microwave device design method according to claim 1, characterized in that, The process of constructing the target function forest model in step S3 is as follows: adopt the diversity forest algorithm framework, set the number of decision trees, and the feature sampling ratio of each tree is the square root of the total number of features; A bivariate splitting strategy is adopted, in which each node finds the optimal split point based on feature pairs that have not appeared in the current path.

3. The microwave device design method according to claim 1, characterized in that, The process of constructing the misjudgment forest model in step S3 is as follows: Based on TES, the prediction error baseline data of the target function forest model is calculated, including the absolute difference between the true response value and the predicted response value, and the variance of the prediction results; using sample features as input, and dynamic confidence difference... To output, construct a misclassification forest model. ; The calculation uses an adaptive weighting formula. ,in It is the variance of the prediction results. This is the actual response value. It is a predicted response value. , For dynamic weights, Based on the error distribution characteristics in TES, the misjudgment forest model is adaptively adjusted, and the same hyperparameter settings as the target function forest model are used during training for the number of nodes and the maximum depth of the tree.

4. The microwave device design method according to any one of claims 1 to 3, characterized in that, The method for screening key samples in step S4 is as follows: First-level screening: The first sample is selected using the target function forest model. Predict the response value of the candidate sample pool SP, select the Top-K samples with the smallest response values, and focus on the target optimization direction; Secondary screening: The second sample is selected through a misjudgment forest model. Confidence difference in predicting candidate sample pool SP Select The largest Top-M samples accurately locate regions of high uncertainty; Three-level screening: The third sample is based on the EIM output by the target function forest model to identify high-importance feature pairs. The top-P samples in the candidate sample pool SP whose value combinations of the feature pairs do not fully cover the dataset DAS are selected to supplement key information on feature interactions. An adaptive weight fusion strategy is adopted to prioritize the three-level screening samples: the weight of the third sample is fixed at 0.2; the weight of the first sample is 0.4 and the weight of the second sample is 0.4 in the initial iteration; if the accuracy is insufficient, the weight of the second sample is increased to 0.6; if the efficiency is insufficient, the weight of the first sample is increased to 0.6, so as to ensure that the sample screening is accurately matched with the model requirements.

5. A microwave device design system, characterized in that, include: Two-layer sample construction module: Constructing the candidate sample pool SP: N A number of candidate samples constitute The matrix, the candidate samples include microwave devices k Dimensional design variables x Constructing the initial dataset DAS: Calculating the true performance parameters corresponding to the design variables of each initial sample obtained by Latin hypercube sampling through simulation; Dataset partitioning module: Partitions the initial dataset DAS into a training set TRS and a validation set TES; Model building module: Constructs a target function forest model for the design objective on the training set TRS, outputs the effect importance measure EIM; and uses this model to calculate the basic data of prediction error on the validation set TES, and constructs a misjudgment forest model for the design objective. Key sample selection module: Selects valid samples from the candidate sample pool SP, i.e., selects the first few samples with the smallest response values ​​using the target function forest model; and selects samples using the misclassification forest model. The largest number of second samples; based on the effect importance metric EIM, select several third samples that can effectively supplement the blind spots of the current dataset DAS coverage; after selecting the top Q samples from the first, second, and third samples, the key samples are formed. Dataset update module: After removing redundant samples from the key samples using the cosine similarity metric, calculate the true response value and add it to the initial dataset. Synchronously update TRS and TES, return to the model building module, and retrain the target function forest model and the misjudgment forest model until the preset stopping condition is met. Obtain the target amphibious forest model; Microwave Device Design Module: Used to assist in microwave device design by employing a target amphibious forest model.

6. The microwave device design system according to claim 5, characterized in that, The process of constructing the target function forest model in the model construction module is as follows: adopt the diversity forest algorithm framework, set the number of decision trees, and the feature sampling ratio of each tree is the square root of the total number of features; A bivariate splitting strategy is adopted, in which each node finds the optimal split point based on feature pairs that have not appeared in the current path.

7. The microwave device design system according to claim 5, characterized in that, The process of constructing the misjudgment forest model in the model construction module is as follows: Based on TES, the basic prediction error data of the target function forest model is calculated, including the absolute difference between the true response value and the predicted response value, and the variance of the prediction results; using sample features as input, and dynamic confidence difference... To output, construct a misclassification forest model. ; The calculation uses an adaptive weighting formula. ,in It is the variance of the prediction results. This is the actual response value. It is a predicted response value. , For dynamic weights, Based on the error distribution characteristics in TES, the misjudgment forest model is adaptively adjusted, and the same hyperparameter settings as the target function forest model are used during training for the number of nodes and the maximum depth of the tree.

8. The microwave device design system according to any one of claims 5 to 7, characterized in that, The key sample screening method in the key sample screening module is as follows: First-level screening: The first sample is selected using the target function forest model. Predict the response value of the candidate sample pool SP, select the Top-K samples with the smallest response values, and focus on the target optimization direction; Secondary screening: The second sample is selected through a misjudgment forest model. Confidence difference in predicting candidate sample pool SP Select The largest Top-M samples accurately locate regions of high uncertainty; Three-level screening: The third sample is based on the EIM output by the target function forest model to identify high-importance feature pairs. The top-P samples in the candidate sample pool SP whose value combinations of the feature pairs do not fully cover the dataset DAS are selected to supplement key information on feature interactions. An adaptive weight fusion strategy is adopted to prioritize the three-level screening samples: the weight of the third sample is fixed at 0.2; the weight of the first sample is 0.4 and the weight of the second sample is 0.4 in the initial iteration; if the accuracy is insufficient, the weight of the second sample is increased to 0.6; if the efficiency is insufficient, the weight of the first sample is increased to 0.6, so as to ensure that the sample screening is accurately matched with the model requirements.

9. A training method for an amphibious forest model, characterized in that, Using the microwave device design method described in any one of claims 1 to 4, after each round of key sample screening, the datasets DAS, TRS, and TES are dynamically updated; based on the updated TRS and TES, the target function forest model and the misjudgment forest model are retrained respectively to achieve collaborative optimization of the dual forest model.

10. A method for fabricating a microwave device, characterized in that, The design parameter values ​​corresponding to the ideal design target of the microwave device are obtained by using the method described in any one of claims 1 to 4, and the microwave device is fabricated accordingly.

Images