Method and system for impedance matching design based on large language model
By leveraging multi-agent collaboration and the application of a large language model, impedance matching design is broken down into simple tasks. Combined with the RAG framework and intelligent retrieval, the efficiency and accuracy issues of narrowband and broadband design are resolved, enabling efficient generation and optimization of impedance matching structures.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-18
- Publication Date
- 2026-03-31
AI Technical Summary
Existing technologies in impedance matching design suffer from limitations in narrowband design and require numerous iterations in broadband design, making it difficult to efficiently design broadband matching structures.
A multi-agent collaborative approach is adopted to break down impedance matching design into multiple simple tasks. A large language model is used for semantic understanding and structured mapping. Combined with the RAG framework and intelligent retrieval mechanism, the matching structure is quickly generated and optimized.
It improves the efficiency and accuracy of impedance matching design, especially in broadband matching tasks, enabling the rapid selection of appropriate structures and parameters, reducing the number of iterations, and improving design efficiency and accuracy.
Smart Images
Figure CN121766243A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of circuit design technology, specifically relating to a method and system for impedance matching design based on a large language model. Background Technology
[0002] Existing technologies are based on Smith chart path optimization, which transforms the impedance matching problem into a coordinate optimization problem on a Smith chart. AI algorithms are used to find the optimal path from the source impedance to the target impedance, and the path shape determines the topology (series / parallel) and component parameters of the matching network. However, this approach still has the following technical problems:
[0003] (1) Matching structure designed using Smith chart is often used to implement narrowband design.
[0004] (2) The method of using genetic algorithm to select the optimal path requires repeated multi-level iterations when designing broadband structures.
[0005] Therefore, there is an urgent need for a new method that utilizes multi-agent collaboration to break down impedance matching design into several relatively simple tasks, thereby improving the efficiency and accuracy of large language models in matching structure design. Summary of the Invention
[0006] To address the aforementioned technical problems in the existing technology, the purpose of this invention is to improve the efficiency and accuracy of large language models in impedance matching structure design by utilizing a multi-agent collaborative method to break down the impedance matching design into several relatively simple tasks. The technical solution is as follows:
[0007] A method for impedance matching design based on a large language model includes the following steps:
[0008] Step 1: Collect the impedance matching design requirements information input by the user, preprocess it, and output the requirement data;
[0009] Step 2: Based on the pre-trained and fine-tuned large language model, perform semantic understanding, field extraction and structured mapping on the requirement data, and use a rule verification mechanism to unify conversion units, verify parameter ranges and fill in missing information to generate a standardized structured design task description.
[0010] Step 3: Determine the matching task type, select the appropriate design based on the matching task type, and generate the topology and parameters;
[0011] Step 4: Convert the generated topology and parameters into a simulateable netlist file, perform S-parameter simulation, generate simulation results, analyze key performance indicators, judge the performance of the initial structure according to the design objectives, generate a complete result summary report for the initial structure that meets the design objectives, and write it into the RAG knowledge base in a standardized format.
[0012] Step 5: For initial structures that do not meet the design goals, continue parameter optimization. For initial structures that achieve the design goals through parameter optimization, output the matched structure with the designed parameters and a standardized design report, and write it into the RAG knowledge base. For initial structures whose optimization results do not meet the design goals, analyze the reasons for failure based on the optimization trajectory, generate structural modification suggestions, and return to Step 3 to redesign.
[0013] Furthermore, in step 1, the requirement information includes: operating frequency band, center frequency, bandwidth range, source or load impedance requirements, target reflection coefficient, and design constraint information.
[0014] Furthermore, in step 3, the matching task types include: narrowband matching task and wideband matching task;
[0015] The specific design process for selecting the appropriate design based on the matching task type is as follows:
[0016] (1) For narrowband matching tasks: use analytical and closed-loop computational models to solve the component parameters of L-type, π-type or T-type matching networks, and verify their feasibility and physical rationality;
[0017] (2) For broadband matching tasks: RAG scheduling similar topology and historical design cases are used to generate matching topology and design parameterized initial solutions.
[0018] Furthermore, the specific steps for constructing the RAG are as follows:
[0019] (1) Data Acquisition and Preprocessing: Collect data from the demand side, design side, simulation and verification, and historical engineering cases, and perform engineering processing on the raw data, specifically including:
[0020] ① Semantic level
[0021] By employing text parsing and information extraction techniques, natural language descriptions are uniformly mapped to standardized requirement fields, and unit normalization, parameter range correction, and missing data completion are performed.
[0022] ② Structural level
[0023] The circuit topology and formula derivation are decomposed into templates, breaking them down into the smallest knowledge units: topology type, applicable frequency band, key parameters, and performance boundaries.
[0024] ③ Simulation data level
[0025] Feature extraction and index summarization are performed on continuous simulation curves to filter out abnormal and non-converging data.
[0026] ④ Index and Interface Level
[0027] The cleaned text, parameter tables, and topology descriptions are encoded into searchable structured and vectorized representations;
[0028] (2) Constructing the database: Constructing a document database, a parameter database, and a topology or model database;
[0029] (3) Vectorization and index structure establishment: Convert multi-source heterogeneous data in the document library and parameter library into vector representations that can be semantically computed and efficiently retrieved, and establish an index structure that is highly matched with the design task.
[0030] (4) Construct an intelligent processing engine layer: granularize knowledge units, segment text according to key features of radio frequency circuit design, use image-text joint encoding, and use multimodal Transformer model to synchronously parse circuit diagram vector data and related text semantic descriptions to achieve image-text feature alignment; construct a dynamically combinable three-level knowledge classification system.
[0031] (5) Construct an intelligent retrieval and solution generation layer: adopt a hybrid retrieval mechanism, integrate user keyword query and semantic vector similarity for dual-path recall; set dynamic constraint filters to allow engineers to directly use engineering expressions to accurately filter design options; introduce a dual-engine performance optimization.
[0032] Furthermore, the three-level classification system includes: ① setting a first-level label: topology; ② setting a second-level label: frequency attribute; ③ setting a third-level label: performance characteristics; to form a multi-dimensional structured index.
[0033] Furthermore, the performance optimization dual-engine specifically includes:
[0034] ① Real-time incremental index
[0035] When a user selects a specific structure, the system automatically retrieves similar electrical properties with a Q value of ±10% and constructs an auxiliary decision support network.
[0036] ② Feedback closed-loop mechanism
[0037] The feature labels are updated based on the feedback and modification suggestions, forming a knowledge loop for continuous optimization.
[0038] Furthermore, in step 4, the key performance indicators include: S-parameters, return loss, and operating bandwidth.
[0039] The results summary report includes: the optimized circuit structure, final parameter values, simulation performance analysis results, a description of the optimization process, and the results of judging the performance of the initial structure based on the design objectives.
[0040] A system for impedance matching design based on a large language model includes:
[0041] (1) Requirement Analysis Intelligent Agent
[0042] The system converts user natural language requirements into a unified Design Task JSON, obtains basic semantic understanding in the radio frequency field through pre-training, automatically determines whether the task is narrowband or wideband based on the bandwidth ratio, and controls the output format through structured templates. A rule-based validation mechanism is introduced to automatically detect and correct conversion units, numerical ranges, physical rationality, and format consistency, and outputs structured tasks and design task types.
[0043] (2) Matching design of intelligent agents
[0044] The system receives structured tasks generated by a demand parsing agent, automatically selects a design strategy based on the task type, and designs accordingly. ① For narrowband tasks, it uses analytical and closed-loop computational models to solve for the component parameters of L-type, π-type, or T-type matching networks, and verifies their feasibility and physical rationality. ② For broadband tasks, it uses RAG scheduling of similar topologies and historical design cases to generate matching topologies and design parameterized initial solutions. Finally, it outputs the topology and parameters.
[0045] (3) Verify and optimize the intelligent agent
[0046] The topology and parameters generated by the matching design agent are converted into a simulateable netlist file, and S-parameter simulation is performed to generate simulation results. The simulation results are then used to verify the automatic parsing of key performance indicators by the optimization agent. Based on the design objectives, the performance of the initial structure is judged. For initial structures that meet the design objectives, a complete result summary report is generated and written into the RAG knowledge base in a standardized format. For initial structures that do not meet the design objectives, parameter optimization is continued. For initial structures whose parameter optimization achieves the design objectives, the matched structure with the designed parameters and a standardized design report are output and written into the RAG knowledge base. For initial structures whose optimization results do not meet the design objectives, the reasons for failure are analyzed based on the optimization trajectory, structural modification suggestions are generated, and the process is returned to the matching design agent for redesign.
[0047] The parsed data and performance metrics will then be returned to [the relevant entity / company].
[0048] Furthermore, in (1), the demand parsing agent includes the following modules:
[0049] ① Input Comprehension Module
[0050] Deep semantic analysis of natural language input by users is performed using a large language model;
[0051] ② Structured data generation module
[0052] It consists of a large language model and rule templates. The large language model automatically generates field content based on the agreed JSON Schema, fills in the parsed parameters, and supplements fields when they are missing or ambiguous. The rule templates constrain the order, name and format of fields.
[0053] ③ Design a task judgment module
[0054] Using the Qwen3-8B model, the system automatically determines whether a task belongs to narrowband design or broadband design by understanding the semantic description of user needs and inferring engineering intent.
[0055] ④ Verification and Normalization Module
[0056] Perform unit conversion, range checking, physical rationality verification, field completion, and format consistency checks on the output of the large language model.
[0057] Furthermore, the deep semantic parsing of natural language includes: identifying center frequency, bandwidth, target impedance, power, VSWR, and efficiency parameters; processing fuzzy expressions and inferring implicit intentions.
[0058] Beneficial effects: (1) By using the multi-agent collaboration method, the impedance matching design is broken down into several relatively simple tasks, which improves the efficiency and accuracy of the large language model in the matching structure design. (2) By distinguishing between narrowband and broadband structure design, and by introducing RAG, a special structure database is established, which enables the trained large language model to quickly select the structure and determine the parameters during the design process. Attached Figure Description
[0059] Figure 1 This is a flowchart of the impedance matching design method based on a large language model according to the present invention;
[0060] Figure 2 This is a flowchart illustrating the construction process of the RAG framework of the present invention.
[0061] Figure 3 This is the overall system flowchart of the present invention. Detailed Implementation
[0062] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0063] like Figure 1As shown, the impedance matching design method based on a large language model of the present invention first takes the impedance matching design requirements input by the user (including natural language description, design documents, or parameter forms) as initial input data. This data is then uniformly formatted by the requirement acquisition and preprocessing module, which collects key information such as the operating frequency band, center frequency, bandwidth range, source / load impedance requirements, target reflection coefficient, and design constraints. The preprocessed requirement data is then sent to the requirement parsing agent through the model interface module. Based on a pre-trained and fine-tuned large language model, the requirement parsing agent performs semantic understanding, field extraction, and structured mapping on the input requirements. Combined with a rule-based verification mechanism, it completes unit unification, parameter range verification, and missing information completion, ultimately generating a standardized structured design task description. Based on this description, the matching task type (narrowband or broadband) is determined and used as input for subsequent matching design.
[0064] After receiving the structured design task output by the requirement parsing agent, the matching design agent selects the appropriate design process according to the task type: For narrowband matching tasks, the agent calls analytical and closed-loop computational models to solve the component parameters of L-type, π-type, or T-type matching networks and performs feasibility and physical rationality verification; For broadband matching tasks, the agent uses the RAG framework to schedule similar topologies and historical design cases from the knowledge base to generate the matching structure and design the parameterized initial solution, and submits the generated results to the verification and optimization agent through the model interface module.
[0065] The verification and optimization agent is responsible for converting the topology and parameters generated by the matching design agent into a simulateable netlist file, performing S-parameter simulation, generating simulation results, automatically parsing key performance indicators, judging the performance of the initial structure based on the design objectives, generating a complete result summary report for initial structures that meet the design objectives, and writing it into the RAG knowledge base in a standardized format; for initial structures that do not meet the design objectives, parameter optimization is performed; for initial structures whose parameter optimization achieves the design objectives, the matching structure with the designed parameters and a standardized design report are output and written into the RAG knowledge base; for initial structures whose optimization results do not meet the design objectives, the reasons for failure are analyzed based on the optimization trajectory, structural modification suggestions are generated, and the process is returned to the matching design agent for redesign.
[0066] like Figure 2 As shown, the specific process of building the RAG framework is as follows:
[0067] 1. Data Acquisition and Processing
[0068] The collected data mainly includes four categories: First, demand-side data, such as impedance matching requirement descriptions in textbooks, papers, engineering documents, and design reports, covering natural language expressions such as operating frequency, bandwidth, source / load impedance, and performance indicators; second, design-side data, including the derivation of classic narrowband matching formulas such as L / π / T, the design process of broadband multi-section transmission lines and filtered topologies, parameter ranges, and structural selection criteria; third, simulation and verification data, including S-parameters, bandwidth, group delay, and other results output from tools such as ADS, HFSS, AWR, and Qucs; and fourth, historical engineering case data, including successful design and iterative optimization processes, used to reflect non-ideal effects and empirical rules in real-world engineering. These multi-source, heterogeneous data collectively form the knowledge foundation for subsequent RAG and inference models.
[0069] In the preprocessing stage, the raw data needs to be engineered: at the semantic level, text parsing and information extraction techniques are used to uniformly map natural language descriptions into standardized requirement fields, and unit normalization, parameter range correction, and missing data completion are performed; at the structural level, circuit topology and formula derivation are decomposed into templates and broken down into the smallest knowledge units such as "topology type - applicable frequency band - key parameters - performance boundaries"; at the simulation data level, feature extraction and index summarization are performed on continuous simulation curves, and abnormal and non-convergent data are filtered out; at the indexing and interface level, the cleaned text, parameter tables, and topology descriptions are encoded into searchable structured and vectorized representations, and sent to the subsequent RAG retrieval and generation module through a unified interface.
[0070] 2. Build the database
[0071] Build a document database, a parameter database, and a topology or model database;
[0072] The document database stores data related to knowledge representation and design logic, including unstructured or semi-structured text data such as impedance matching theory descriptions, topology explanations, design criteria, applicable frequency band analysis, and simulation software documentation from textbooks, papers, patents, and engineering manuals. Through semantic segmentation and knowledge unit granularization in the preprocessing stage, the database breaks down the original long text into the smallest independently searchable design knowledge units. Each unit contains clear information on topology type, applicable scenarios, performance boundaries, and engineering constraints.
[0073] The parameter library is geared towards the storage of numerical and computable information, mainly including typical parameter configurations for broadband matching, the initial parameter ranges for multi-section matching structures, and the corresponding simulation results and performance indicators.
[0074] The topology / model database is mainly used to store impedance matching structure templates that can be directly used for design and simulation. This database mainly contains circuit netlists in a unified format for broadband matching topologies such as multi-section transmission line type, filter type, and transformer type.
[0075] 3. Vectorization and Index Structure Establishment
[0076] Transform multi-source heterogeneous data in the document library and parameter library into vector representations that are semantically computational and efficiently searchable, and establish an index structure that is highly matched with the design task;
[0077] In terms of vectorization processing, a differentiated encoding strategy is adopted for different types of data: (1) For text data in the document database, the design principles, topology descriptions and engineering experience are encoded into semantic vectors using a domain-adaptive embedding model. Before encoding, structural hints are introduced to explicitly embed key information such as topology type, frequency band range and performance indicators into the vector representation; (2) For structured data in the parameter database, the parameter combination, topology configuration and corresponding performance indicators are expanded and normalized in terms of features, and then jointly encoded with the vector representation of the text description so that the parameter-performance relationship can be perceived and aligned in the vector space.
[0078] In terms of index structure establishment, a hierarchical and hybrid index method is constructed to support retrieval needs under different precision and constraint conditions; (1) In the vector index layer, an approximate nearest neighbor search structure is adopted to achieve efficient recall based on semantic similarity; (2) In the structure index layer, an inverted index and rule filter based on metadata are introduced to quickly screen engineering constraints such as topology type, frequency band interval, bandwidth ratio and number of components; through the joint scheduling of vector similarity retrieval and structure condition filtering, the recall of candidate knowledge items that meet engineering constraints and have high semantic relevance is realized.
[0079] 4. Construct an intelligent processing engine layer
[0080] A triple-deep processing strategy is adopted to improve the organization and retrieval efficiency of RF design knowledge: First, knowledge units are granularized, and text is segmented according to the key features of RF circuit design to ensure that each information block clearly includes the structural type, operating frequency band characteristics, and core performance boundaries; Second, through graph-text joint encoding, the multimodal Transformer model is used to simultaneously parse the circuit diagram vector data and related text semantic descriptions to achieve graph-text feature alignment; Finally, a dynamically combinable three-level knowledge classification system is constructed, with first-level labels: topology architecture, second-level labels: frequency attributes, and third-level labels: performance characteristics, forming a multi-dimensional structured index.
[0081] 5. Intelligent Retrieval and Solution Generation Layer
[0082] A hybrid retrieval mechanism is adopted, integrating user keyword queries and semantic vector similarity for dual-path recall; a dynamic constraint filter is set up, allowing engineers to directly use engineering expressions to accurately filter design options; and a performance optimization dual engine is introduced—the first is a real-time incremental index, which automatically retrieves those with similar electrical performance (such as Q value ±10%) when the user selects a specific structure, and builds an auxiliary decision support network; the second is a feedback closed-loop mechanism, which organizes the context based on the retrieval fragments returned by the intelligent processing engine layer, calls the large model to generate impedance matching design schemes (recommending topology, providing parameter values, explaining impedance trajectory, and explaining design basis), and returns the final generated scheme to the matching design agent to form a knowledge closed loop of continuous optimization.
[0083] like Figure 3 As shown, the system for impedance matching design based on a large language model according to the present invention includes:
[0084] (1) Requirement Analysis Intelligent Agent
[0085] The user's natural language requirements are converted into a unified Design Task JSON. The model obtains basic semantic understanding in the RF field through pre-training. By fine-tuning the learning of requirement extraction, intent recognition, field filling and task type determination, it stably identifies center frequency, bandwidth, source / load impedance, power target and design constraint information. It automatically determines whether the task is narrowband or wideband based on the bandwidth ratio. The output format is controlled by structured templates during the generation stage.
[0086] A rule-based validation mechanism is added to automatically detect and correct units, numerical ranges, physical rationality, and format consistency. The output includes structured task and design task types, which are directly passed to the matching design agent through an interface.
[0087] The requirement analysis intelligent agent consists of four core modules: input understanding module, structured data generation module, design task judgment module, and verification and normalization module;
[0088] Input understanding module: performs deep semantic parsing of natural language, identifies key parameters such as center frequency, bandwidth, target impedance, power, VSWR, and efficiency, and processes fuzzy expressions and infers implicit intentions.
[0089] The structured data generation module consists of an LLM and rule templates. The model automatically generates field content based on the agreed JSON Schema, fills in the parsed parameters, and makes reasonable additions when fields are missing or ambiguous. The rule templates constrain the order, name, and format of fields to ensure that the output meets the structured interface requirements of the matched intelligent agent.
[0090] Design task judgment module: Adopt the Qwen3-8B model to automatically judge whether the task belongs to narrowband design or broadband design by understanding the semantics of user requirements and inferring engineering intentions.
[0091] Verification and normalization module: Perform unit conversion (MHz / GHz), range check (such as f_low < f_high), physical rationality verification (VSWR cannot be negative), field completion and format consistency check on the output of the large language model.
[0092] (2)Matching design agent
[0093] Receive the structured task generated by the requirement analysis agent, and automatically select a design strategy for design according to the task type;
[0094] ① For narrowband tasks, based on the analytical formula and topology selection mode learned during training, automatically complete the calculation of L / π / T network matching parameters and quickly generate an available design scheme;
[0095] ② For broadband tasks, combine RAG technology for topology selection and initial parameter design. Retrieve similar design instances, topology templates, optimization strategies and engineering constraints from the knowledge base, obtain reference information based on the generation ability of the large language model, generate an initial design based on the retrieved structure and parameter samples, and predict feasibility using the parameter-performance mode learned during training;
[0096] After the design is completed, output the topology structure and parameters.
[0097] (3)Verification and optimization agent
[0098] The verification and optimization agent is responsible for converting the topology and parameters generated by the matching design agent into a simulateable netlist file, and calling the corresponding tools to perform S-parameter simulation, generating simulation results. The verification and optimization agent automatically parses key performance indicators, including S-parameters, return loss, and operating bandwidth. Based on the design objectives, it judges the performance of the initial structure. For initial structures that meet the design objectives, it generates a complete result summary report, including: optimized circuit structure, final parameter values, simulation performance analysis results, optimization process description, and the results of judging the performance of the initial structure based on the design objectives, and writes it into the RAG knowledge base in a standardized format. For initial structures that do not meet the design objectives, the agent initiates a parameter optimization process, performing multiple rounds of parameter search through external optimization tools. The optimization objective is usually to maximize in-band reflection performance or minimize the full bandwidth S11. The optimization engine automatically adjusts the parameter combination based on the simulation results until the constraints are met or the iteration limit is reached. For the initial structure whose parameters have been optimized to meet the design goal, the engine outputs the matched structure with the designed parameters and a standardized design report, and writes it into the RAG knowledge base. The engine continues to optimize the parameters of the initial structure that has not met the design goal. For the initial structure whose optimization results have not met the design goal, the engine analyzes the reasons for failure based on the optimization trajectory, generates structural modification suggestions, and returns them to the matching design agent for redesign.
Claims
1. A method for impedance matching design based on a large language model, characterized in that, The method comprises the following steps: Step 1: collecting user input impedance matching design requirement information and preprocessing, outputting requirement data; Step 2: based on pre-training and fine-tuning of large language model, semantic understanding, field extraction and structured mapping of requirement data, using rule checking mechanism for unit conversion, parameter range verification and missing information completion, generating standardized structured design task description; Step 3: determining the matching task type, selecting the corresponding design according to the matching task type, and generating the topology structure and parameters; Step 4: converting the generated topology structure and parameters into a netlist file that can be simulated, performing S parameter simulation to generate simulation results, analyzing key performance indicators, and judging the performance of the initial structure according to the design target. For the initial structure that meets the design target, a complete result summary report is generated and written into the RAG knowledge base in a standardized format; Step 5: for the initial structure that does not meet the design target, continue to optimize the parameters. For the initial structure that meets the design target after parameter optimization, output the matching structure with optimized parameters and the standardized design report, and write them into the RAG knowledge base. For the initial structure that does not meet the design target after optimization, analyze the failure reason according to the optimization trajectory, generate structure-level modification suggestions, and return to step 3 for redesign.
2. The method for impedance matching design based on large language model according to claim 1, characterized in that, In step 1, the requirement information includes: working frequency band, center frequency, bandwidth range, source or load impedance requirement, target reflection coefficient and design constraint information.
3. The method of claim 1, wherein the method is characterized by: In step 3, the matching task type includes: narrowband matching task and wideband matching task; The specific design process according to the matching task type is as follows: (1) for narrowband matching task: use analytical expression and closed-form calculation model to solve the element parameters of L-type, π-type or T-type matching network, and perform realizability and physical rationality check; (2) for wideband matching task: use RAG to schedule similar topology structure and historical design cases to generate matching topology structure and parameterized initial solution design.
4. The method of claim 3, wherein, The specific steps of constructing RAG are as follows: (1) data collection and preprocessing: collect demand side, design side, simulation and verification and historical engineering case data, and perform engineering processing on the original data, including: ① semantic level Use text analysis and information extraction technology to map natural language description to standardized requirement field, and perform unit normalization, parameter range correction and missing completion; ② structure level Template decomposition of circuit topology and formula derivation, including: topology type-applicable frequency band-key parameter-performance boundary minimum knowledge unit; ③ simulation data level Extract features and index summary from continuous simulation curves, filter abnormal and non-convergent data; ④ index and interface level Encode the cleaned text, parameter table and topology description into structured and vectorized representation for retrieval; (2) database construction: construct document database, parameter database and topology or model database; (3) Vectorization and index structure establishment: Convert multi-source heterogeneous data in the document library and parameter library into vector representations that can be semantically calculated and efficiently retrieved, and establish an index structure that highly matches the design task. (4) Build an intelligent processing engine layer: Granulate knowledge units, divide text according to key features of RF circuit design, use multi-modal Transformer model to analyze circuit diagram vector data and related text semantic description, realize alignment of text and diagram features, and build a dynamically combinable three-level knowledge classification system. (5) Build an intelligent retrieval and scheme generation layer: Use a hybrid retrieval mechanism to combine user keyword queries and semantic vector similarity for double-path recall; set dynamic constraint filters to allow engineers to use engineering expressions for accurate filtering of design options; introduce a performance optimization dual-engine.
5. The method of claim 4, wherein, The three-level classification system includes: ① Set the first-level label: topological architecture; ② Set the second-level label: frequency attribute; ③ Set the third-level label: performance characteristics; to form a multi-dimensional structured index.
6. The method for impedance matching design based on large language model according to claim 4, characterized in that, The performance optimization dual-engine specifically includes: ① Real-time incremental indexing When the user selects a specific structure, automatically retrieve similar electrical performance with Q value ±10%, and build an auxiliary decision support network; ② Feedback closed-loop mechanism Update feature labels based on feedback suggestions to form a continuously optimized knowledge closed loop.
7. The method for impedance matching design based on large language model according to claim 1, characterized in that, In step 4, the key performance indicators include: S-parameters, return loss, and operating bandwidth key performance indicators. The result summary report includes: optimized circuit structure, final parameter values, simulation performance analysis results, optimization process description, and results of judging the performance of the initial structure according to the design target. 8.A system for impedance matching design based on a large language model, characterized in that, Including: (1) Demand analysis agent Convert user natural language requirements into a unified Design Task JSON, obtain basic semantic understanding in the RF field through pre-training, automatically determine the task as narrowband or wideband based on bandwidth ratio, control output format through structured templates; introduce a rule checking mechanism to automatically detect and correct conversion units, value ranges, physical reasonableness, and format consistency, and output structured tasks and design task types; (2) Matching design agent Receive structured tasks generated by the demand analysis agent, automatically select design strategies for design according to task type, ① For narrowband tasks, use analytical and closed-form computational models to solve element parameters of L-type, π-type, or T-type matching networks, and perform realizability and physical reasonableness checks; ② For wideband tasks, use RAG to schedule similar topological structures and historical design cases to generate and parameterize initial solution design for matching topological structures; finally output topological structures and parameters; (3) Verification and optimization agent Convert the topological structure and parameters generated by the matching design agent into a simulative netlist file, perform S-parameter simulation, generate simulation results, verify the automatic analysis of key performance indicators by the optimization agent, judge the performance of the initial structure according to the design target, and generate a complete result summary report for the initial structure that meets the design target, and write it into the RAG knowledge base in a standardized format; For the initial structure that does not meet the design target, parameter optimization is continued, for the initial structure that meets the design target after parameter optimization, a matching structure with a standardized design report is output, and is written into the RAG knowledge base; For the initial structure that does not meet the design target after optimization, the failure reason is analyzed according to the optimization track, a structure-level modification suggestion is generated, and is returned to the matching design intelligent agent for re-design. And the parsed data and performance indicators are returned to.
9. The system for impedance matching design based on large language model according to claim 8, wherein, In the (1), the requirement analysis intelligent agent comprises the following modules: ① Input understanding module Deep semantic analysis of natural language input by the user through a large language model; ② Structured data generation module Composed of a large language model and a rule template, the large language model automatically generates field content according to the agreed JSON Schema, fills in the parsed parameters, and supplements when the field is missing or the expression is ambiguous, and the rule template constrains the field order, name and format; ③ Design task judgment module Understand the semantic description of the user's demand and infer the engineering intention by using the Qwen3-8B model, and automatically judge whether the task belongs to narrowband design or wideband design; ④ Verification and standardization module Unit conversion, range check, physical reasonableness check, field completion and format consistency check are performed on the output of the large language model.
10. The system for impedance matching design based on large language model according to claim 9, wherein, The deep semantic analysis of natural language includes: identifying the center frequency, bandwidth, target impedance, power, VSWR, efficiency parameters; processing ambiguous expressions and inferring implicit intentions.