Radio frequency relay optimization design method fusing prior knowledge of large language model and Kriging-IAGA
By integrating large language models with the Kriging-IAGA method, the problems of long simulation cycles and nonlinear coupling in the optimization design of RF relays were solved, achieving efficient and accurate RF relay design, reducing insertion loss and improving product consistency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies in the optimization design of RF relays suffer from problems such as long simulation cycles and difficulty in handling nonlinear coupling in orthogonal experiments, resulting in large fluctuations in insertion loss and difficulty in finding the global optimal solution.
An optimization design method integrating prior knowledge from a large language model and Kriging-IAGA is adopted. A sample library is constructed through single-factor analysis and orthogonal experiments, and a Kriging prediction model is established. An improved adaptive genetic algorithm (IAGA) is combined for global optimization, and a large language model is introduced for process feasibility assessment. The model is dynamically corrected to improve prediction accuracy and optimization efficiency.
It significantly shortened the design cycle, reduced insertion loss in the high-frequency band, improved product consistency and RF performance, and ensured process feasibility.
Smart Images
Figure CN121960020A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electromagnetic relay design technology, and relates to a radio frequency relay structure design method. Specifically, it relates to a radio frequency relay material-structure optimization design method that aims to minimize insertion loss and integrates prior knowledge of large language models and Kriging-IAGA. Background Technology
[0002] Radio frequency (RF) relays are widely used in aerospace, radar, and high-frequency communication systems. Insertion loss is a key electromagnetic compatibility parameter for evaluating signal transmission quality. Due to the compact internal structure of miniature relays such as TO-5, even small changes in structural and material parameters can cause drastic fluctuations in insertion loss under high-frequency signals. Existing technologies typically employ the following two optimization methods:
[0003] 1. Pure Finite Element Simulation (HFSS): This method involves trial and error through parameter scanning. However, a single full-wave simulation can take several hours, and the complex nonlinear coupling between parameters results in extremely high computational costs, making it difficult to traverse the entire solution space.
[0004] 2. Orthogonal Experimentation: This method selects representative levels for combination. While it reduces the number of experiments, it is limited to the analysis of discrete points, making it difficult to find the global optimum and unable to establish a continuous prediction model.
[0005] Therefore, there is an urgent need for an intelligent design method that can both ensure prediction accuracy and significantly improve optimization efficiency. Summary of the Invention
[0006] This invention provides a radio frequency relay optimization design method that integrates prior knowledge from large language models with Kriging-IAGA. This method solves the problems of long simulation cycles and difficulty in handling nonlinear coupling in orthogonal experiments in traditional relay optimization design. While significantly shortening the design cycle, it effectively reduces insertion loss in high-frequency bands and improves product consistency and radio frequency performance.
[0007] The objective of this invention is achieved through the following technical solution:
[0008] A method for optimizing the design of radio frequency relays that integrates prior knowledge from large language models with Kriging-IAGA includes the following steps:
[0009] Step 1: Key Parameter Screening: Establish a full-wave simulation model of the RF relay. Use single-factor analysis to screen mechanical dimension parameters, process assembly parameters, and software simulation parameters to determine structural and material parameters that significantly affect insertion loss as design variables. The structural parameters include continuous variables such as spring width, contact bending angle, contact spacing, and spring end distance. The material parameters include discrete variables such as plating material scheme (contact plating material and plating thickness). The optimization scheme of the contact plating material and plating structure uses discrete variable encoding for global optimization. The total plating thickness is set to be greater than the skin depth corresponding to the operating frequency.
[0010] Step 2: Sample Database Construction: Sample points are selected using a combination of single-factor analysis and orthogonal experiments. The insertion loss response value of each sample point in the target frequency band is calculated using finite element simulation software (HFSS) to construct the initial training sample set. The specific steps are as follows:
[0011] Step 2: Obtain response data under changes in a single parameter through single-factor analysis;
[0012] Step 22: Use orthogonal experimental methods to obtain cross-response data under simultaneous changes of two or more parameters;
[0013] Steps 2 and 3: Use HFSS simulation to obtain the insertion loss response value of each sample in the target frequency band;
[0014] The response data, cross-response data, and insertion loss response values obtained in steps 24, 21, and 23 are used together to form the initial training sample set for training the Kriging prediction model.
[0015] Step 3: Proxy Model Construction: Based on the initial training sample set, a Kriging prediction model between the design variables and the insertion loss is established using the Kriging interpolation method. The Kriging prediction model includes a multinomial regression part and a random distribution part.
[0016] Step 4, Global Optimization: Employ an improved adaptive genetic algorithm (IAGA) with a variation of the insertion loss (IL) calculated using the Kriging prediction model. The fitness function is used to perform a global optimization search on the design variables; the improved adaptive genetic algorithm adopts a nonlinear normalized geometric ranking selection strategy and automatically adjusts the crossover probability based on the population fitness. and mutation probability When individual fitness greater than the population average fitness When, crossover probability and mutation probability Calculate using the following formula:
[0017]
[0018]
[0019] in, This represents the maximum fitness of the current population. For the two individuals to be crossed, the greater fitness, , , , It is a constant and can be adjusted according to the actual scenario;
[0020] Step 5, Dynamic Correction: During the iteration of the genetic algorithm, the best individual in the current generation is extracted at each preset generation interval for HFSS simulation verification. The real simulation data is added to the training sample set and the Kriging prediction model is dynamically updated. The dynamic correction strategy is as follows: every 20 generations, the parameter combination corresponding to the best individual in the current population is selected, the HFSS software is called to perform accurate simulation, and the newly generated simulation data is used to correct the surrogate model.
[0021] Step Six: Large Language Model Correction: During the search process, a large language model is introduced to semantically evaluate the manufacturability of candidate individuals, and the evaluation results are used as a penalty factor and integrated into the fitness function to achieve synergistic optimization of electromagnetic performance and process performance. The specific steps are as follows:
[0022] Step 61: Introduce a large language model preprocessed using retrieval enhancement generation technology to construct an online process evaluation mechanism. Its semantic evaluation specifically includes: mapping the structural parameters and material parameters generated by the genetic algorithm into natural language description text, and generating engineering query instructions;
[0023] Step 62: Vectorize the relevant documents such as the manufacturing process specifications, test reports, historical failure cases, and material property manuals of the RF relays and build a relay manufacturing knowledge base;
[0024] Step 63: Input the engineering inquiry command into the large language model pre-loaded with the relay manufacturing knowledge base, obtain the process feasibility score output by the large language model, and use this score to make a weighted adjustment to the final fitness value of the individual: ,in, This is the electromagnetic insertion loss value. This represents the process constraint weighting coefficient. It is the process feasibility score for LLM;
[0025] Step 7: Parameter Output and Prototype Preparation: When the algorithm converges or reaches the maximum number of iterations, the optimal combination of design variables is output, and the prototype of the RF relay is prepared and tested based on these parameters.
[0026] Compared with the prior art, the present invention has the following advantages:
[0027] 1. High optimization efficiency: The Kriging surrogate model replaces the frequent HFSS full-wave simulations, and the high-precision model built using single-factor and orthogonal experimental data significantly reduces the number of simulations.
[0028] 2. Strong global search capability: The improved adaptive genetic algorithm (IAGA) overcomes the shortcomings of traditional algorithms that are prone to premature convergence. It effectively prevents evolutionary stagnation through a non-zero lower bound probability strategy and introduces a large language model to increase the weight of process feasibility.
[0029] 3. Reliable prediction accuracy: Through the "dynamic correction" strategy, high-value simulation data is continuously added to the model, ensuring the prediction accuracy of the model near the optimal solution. Attached Figure Description
[0030] Figure 1 This is a flowchart of the radio frequency relay optimization design method of the present invention;
[0031] Figure 2 This is a schematic diagram of the insertion loss of a two-port network for an RF relay.
[0032] Figure 3 This is the prediction graph of the Kriging surrogate model in this invention;
[0033] Figure 4 This is a performance diagram of the Kriging proxy model in this invention;
[0034] Figure 5 This is a graph showing the measured insertion loss data of the optimized RF relay. Detailed Implementation
[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.
[0036] This invention provides an optimization design method for RF relays that integrates prior knowledge from a large language model (LLM) with Kriging-IAGA. First, key structural and material parameters affecting insertion loss are screened through HFSS simulation and single-factor analysis. Then, a sample library is constructed using a combination of single-factor analysis and orthogonal experiments to establish a high-precision Kriging surrogate model to replace time-consuming finite element simulation. Next, an improved adaptive genetic algorithm (IAGA) is introduced for multi-parameter collaborative global optimization. The algorithm incorporates dynamic crossover mutation probability adjustment and an elite retention strategy, and uses HFSS to periodically verify and correct the surrogate model. Finally, a large language model (LLM) is introduced as a process constraint to construct a dual evaluation system of "electromagnetic performance-process implementation." The LLM is used to semantically score the manufacturability of the optimal solution and correct the fitness function. Figure 1 As shown, the specific steps are as follows:
[0037] Step 1: Key Parameter Screening: Establish the HFSS parametric model of the relay. Through single-factor analysis, identify the variables that have the greatest impact on insertion loss from three categories of parameters: mechanical dimensions, process assembly, and software simulation. These variables include reed width, contact bending angle, contact spacing, reed end distance, and material coating.
[0038] Step 2: Constructing the training sample library: Sample points are generated using a combination of univariate analysis and orthogonal experiments. First, basic trend data is obtained through univariate analysis. Then, response data of multi-parameter cross-variation is obtained through orthogonal experiments. HFSS simulation is used to obtain the insertion loss value corresponding to each sample, ensuring that the sample data can accurately reflect the nonlinear coupling effect between parameters.
[0039] Step 3: Establish the Kriging surrogate model: Train the Kriging model using the training sample library to construct the mathematical mapping relationship between design variables and insertion loss. The model includes a multinomial regression part and a random distribution part, which can accurately predict the insertion loss trend under different structural and material parameters.
[0040] Step 4, IAGA Global Optimization: An improved adaptive genetic algorithm (IAGA) is used for optimization, with a variation of the insertion loss (IL) value predicted by the Kriging model as the fitness function. Based on the difference between individual fitness and the average fitness of the population, the crossover probability and mutation probability are dynamically and nonlinearly adjusted to avoid the algorithm getting trapped in local optima and to protect superior individuals.
[0041] Step 5, Dynamic Model Correction: To prevent the surrogate model from distorting in unsampled regions, during the genetic algorithm iteration process (e.g., every 20 generations), the current optimal solution is verified using HFSS entity simulation. Real data is fed back to the sample library and the Kriging model is updated, forming a closed loop of "prediction-optimization-correction".
[0042] Step Six: LLM-Based Optimization and Correction: To avoid the theoretically optimal solution generated by the algorithm exceeding actual processing capabilities, this invention introduces a large language model preprocessed using Retrieval Enhanced Generation (RAG) technology to construct an online process evaluation mechanism. Structured data such as relay manufacturing process specifications, historical failure case libraries, and material property manuals, as well as experimental reports, are converted into vector embeddings to build a knowledge base. When the LLM receives process evaluation prompts, it automatically extracts keywords and retrieves relevant document fragments from the library using cosine similarity, ensuring that the scoring basis originates from an accurate knowledge base. The structural and material parameters generated by the genetic algorithm are mapped to natural language description text, and the process feasibility score output by the large language model is obtained. The final fitness value of each individual is then weighted and corrected as follows: ,in This is the electromagnetic insertion loss value. These are process constraint weighting coefficients, which can vary depending on the actual application scenario. It is the process feasibility score of LLM. The correction makes individuals with low process feasibility eliminated in the population evolution and outputs feasible optimal parameters.
[0043] Step 7: Prototype Verification: Prototype fabrication and testing based on the optimized parameters.
[0044] Example:
[0045] This embodiment takes a certain model of TO-5 radio frequency relay as an example, and the goal is to minimize its insertion loss in the DC-5GHz frequency band.
[0046] Step 1: Establish a parametric model using Ansys HFSS. Determine key optimization parameters and their ranges: reed width. (0.60~0.85mm), contact bending angle (75°~85°), contact spacing (1.90~2.40mm), distance from the end of the spring (0.15-0.35mm) and coating scheme (For example =1 indicates a 3μm gold plating. =2 indicates a 3μm silver plating. =3 represents a 3μm silver plating layer plus a thin gold layer.
[0047] Step 2: Constructing the training sample library. Sample points are generated using a combination of univariate analysis and orthogonal experiments. First, basic trend data is obtained through univariate analysis. Then, response data of multi-parameter cross-variation is obtained through orthogonal experiments. HFSS simulation is used to obtain the insertion loss value corresponding to each sample, ensuring that the sample data accurately reflects the nonlinear coupling effect between parameters.
[0048] Step 3: Train the Kriging model. In this embodiment, the Kriging surrogate model is used to construct the functional relationship between design variables and insertion loss. The model assumes that the response function y(x) consists of a global trend model g(x) and a local random deviation z(x). Considering the continuity of the RF relay insertion loss variation, the covariance matrix of z(x) in this invention is described using the Gaussian Correlation Function. The hyperparameters in the model are determined using maximum likelihood estimation, thereby achieving high-precision fitting of the nonlinear response surface. Figure 4 As shown, the coefficient of determination R of the trained Kriging model is... 2 =0.997, RMSE=0.049 dB, which has the ability to replace simulation software.
[0049] Step 4: Set IAGA parameters. Set the population size N=50 and the maximum number of iterations to 100. To overcome the problem of standard genetic algorithms easily getting trapped in local optima, this invention adopts an improved adaptive genetic algorithm (IAGA). Its core improvement lies in the crossover probability. and mutation probability It can automatically adjust according to the fitness of an individual. That is, the algorithm sets a non-zero lower probability to prevent evolutionary stagnation. When the individual fitness... greater than the population average fitness At this time, a nonlinear adjustment strategy is adopted. The specific adaptive calculation formula is as follows:
[0050]
[0051]
[0052] in, This represents the maximum fitness of the current population. This represents the average fitness value of the current population. The higher fitness of the two individuals to be crossed is taken as the coefficient value. In this embodiment, to ensure that the superior individual still has a certain evolutionary potential, the coefficients are set as follows: =0.9, =0.6, =0.1, =0.01. Furthermore, this invention employs an elite preservation strategy, directly replicating the individual with the highest fitness in each generation to the next, ensuring that the optimal solution is not lost due to crossover and mutation operations.
[0053] Step 5: Iterative Optimization and Correction. In this embodiment, the method of this invention (Kriging-IAGA) converges to the optimal solution (minimum insertion loss of 0.73dB at 4 GHz) in the 43rd generation, without getting trapped in local optima. During the iteration process, HFSS is called every 20 generations to verify and correct surrogate model errors.
[0054] Step Six: Collaborative Global Optimization Based on LLM and IAGA. An improved adaptive genetic algorithm is used for optimization. To avoid the theoretically optimal solution generated by the algorithm exceeding the actual manufacturing capability, this invention introduces a large language model preprocessed using retrieval enhancement generation technology to construct an online process evaluation mechanism: 1. Prompt word construction: After the genetic algorithm generates each generation of individuals, the numerical genes of the individuals are mapped to natural language description text. The following structured prompt words are constructed: "The current design scheme is a TO-5 RF relay: the spring material is beryllium copper, the width is 0.72mm, the contact spacing is 2.10mm, and the process of first plating nickel and then silver is adopted. Based on the relay manufacturing process specifications, please evaluate the forming risk and plating adhesion risk of this structure, and output a feasibility confidence score of 0 to 1."; 2. LLM reasoning and scoring: The finely tuned large language model (preloaded with relay design manuals, standards, and historical failure case libraries) is called to reason about the above prompt words, and the LLM outputs the process feasibility score of the design scheme. Where 1 represents fully mature technology and 0 represents unmanufacturable technology; 3. Dual fitness evaluation: Define the comprehensive fitness function. The process score is calculated by combining the electromagnetic insertion loss IL predicted by the Kriging model with the LLM output: In the formula, This represents the process constraint weighting coefficient. The formula automatically eliminates pseudo-optimal solutions with "excellent electromagnetic performance but extremely low process scores" through a penalty term mechanism; 4. Evolutionary Iteration: Based on comprehensive fitness. The aforementioned IAGA is used for optimization until the termination condition is met.
[0055] Step 7: Final Parameter Output and Verification. The optimized parameters are: reed width 0.82mm, angle 81.5°, spacing 2.15mm, and end distance 0.25mm. (For example...) Figure 5 As shown, the measured insertion loss of the prototype within 5GHz is less than 1.8dB, and the curve is smooth, eliminating the severe fluctuations of the original product at 3~4GHz.
[0056] Prototype fabrication was conducted. Due to unavoidable manufacturing errors, 20 prototypes with parameters within an 8% error range were selected for insertion loss testing. An Agilent E5071C was used for testing. A total of 25 contact points were measured. The RF characteristics showed good consistency, with insertion loss less than 1.8dB within 5GHz. The fluctuation frequency point shifted from 3-4GHz to approximately 5.5GHz. Specific insertion losses before and after optimization are shown in Table 1. Test data indicates that the optimized prototype exhibits good RF characteristic consistency, with fluctuations within the operating frequency band largely eliminated and a smooth decrease within the 1-5GHz band.
[0057]
Claims
1. A method for optimizing the design of radio frequency relays by integrating prior knowledge from large language models and Kriging-IAGA, characterized in that... The method includes the following steps: Step 1: Key Parameter Screening: Establish a full-wave simulation model of the RF relay, and screen mechanical dimension parameters, process assembly parameters, and software simulation parameters through single-factor analysis to determine the structural and material parameters that have a significant impact on insertion loss as design variables; Step 2: Sample Database Construction: Sample points are selected using a combination of single-factor analysis and orthogonal experiments. The insertion loss response value of each sample point in the target frequency band is calculated using the finite element simulation software HFSS to construct the initial training sample set. Step 3: Proxy Model Construction: Based on the initial training sample set, a Kriging prediction model between design variables and insertion loss is established using the Kriging interpolation method; Step 4, Global Optimization: Employ an improved adaptive genetic algorithm (IAGA) with a variation of the insertion loss value IL calculated using the Kriging prediction model. The fitness function is used to perform a global optimization search on the design variables; Step 5, Dynamic Correction: During the iteration of the genetic algorithm, the best individual in the current generation is extracted at each preset interval for HFSS simulation verification. The real simulation data is added to the training sample set and the Kriging prediction model is dynamically updated. Step 6, Large Language Model Correction: During the search process, a large language model is introduced to semantically evaluate the manufacturability of candidate individuals, and the evaluation results are used as a penalty factor and fused into the fitness function to achieve synergistic optimization of electromagnetic performance and process performance. Step 7: Parameter Output and Prototype Preparation: When the algorithm converges or reaches the maximum number of iterations, the optimal combination of design variables is output, and the prototype of the RF relay is prepared and tested based on these parameters.
2. The RF relay optimization design method integrating prior knowledge of a large language model and Kriging-IAGA as described in claim 1, characterized in that... In step one, the structural parameters include the spring width, contact bending angle, contact spacing, and spring end distance; the material parameters include the contact plating material and plating thickness.
3. The RF relay optimization design method integrating prior knowledge of a large language model and Kriging-IAGA as described in claim 2, characterized in that... The optimization scheme for the contact coating material and coating structure adopts discrete variable coding for global optimization, wherein the total coating thickness is set to be greater than the skin depth corresponding to the working frequency.
4. The RF relay optimization design method integrating prior knowledge of large language models and Kriging-IAGA as described in claim 1, characterized in that... The specific steps of step two are as follows: Step 2: Obtain response data under changes in a single parameter through single-factor analysis; Step 22: Use orthogonal experimental methods to obtain cross-response data under simultaneous changes of two or more parameters; Steps 2 and 3: Use HFSS simulation to obtain the insertion loss response value of each sample in the target frequency band; The response data, cross-response data, and insertion loss response values obtained in steps 2-4 and 2-3 are used to construct the initial training sample set for training the Kriging prediction model.
5. The RF relay optimization design method integrating prior knowledge of large language models and Kriging-IAGA as described in claim 1, characterized in that... In step three, the Kriging prediction model includes a multinomial regression component and a random distribution component.
6. The RF relay optimization design method integrating prior knowledge of large language models and Kriging-IAGA as described in claim 1, characterized in that... In step four, the improved adaptive genetic algorithm adopts a nonlinear normalized geometric ranking selection strategy and automatically adjusts the crossover probability based on the population fitness. and mutation probability When individual fitness greater than the population average fitness When, crossover probability and mutation probability Calculate using the following formula: in, This represents the maximum fitness of the current population. For the two individuals to be crossed, the greater fitness, , , , It is a constant.
7. The RF relay optimization design method integrating prior knowledge of a large language model and Kriging-IAGA as described in claim 1, characterized in that... In step five, the dynamic correction strategy is as follows: every 20 generations, select the parameter combination corresponding to the best individual in the current population, call HFSS software to perform accurate simulation, and use the newly generated simulation data to correct the surrogate model.
8. The RF relay optimization design method integrating prior knowledge of large language models and Kriging-IAGA as described in claim 1, characterized in that... The specific steps of step six are as follows: Step 61: Introduce a large language model preprocessed using retrieval enhancement generation technology to construct an online process evaluation mechanism. Its semantic evaluation specifically includes: mapping the structural parameters and material parameters generated by the genetic algorithm into natural language description text, and generating engineering query instructions; Step 62: Vectorize the relevant documents such as the manufacturing process specifications, test reports, historical failure cases, and material property manuals of the RF relays and build a relay manufacturing knowledge base; Step 63: Input the engineering inquiry command into the large language model pre-loaded with the relay manufacturing knowledge base, obtain the process feasibility score output by the large language model, and use this score to make a weighted adjustment to the final fitness value of the individual: ,in, This is the electromagnetic insertion loss value. This represents the process constraint weighting coefficient. This is the process feasibility score for LLM.