An AI agent-based semiconductor device parameter automatic extraction method

Abstract

This invention discloses an automatic parameter extraction method for semiconductor devices based on an AI agent, comprising: receiving a parameter extraction task input by a user; retrieving task-related information through a RAG module and inputting it into a large language model; then, the large language model performs a structured decomposition of the task based on the retrieved information; generating multiple candidate parameter extraction paths and scoring them based on the task decomposition results and the physical dependencies of the device parameters, and selecting the optimal path; executing the optimal path while simultaneously performing device model simulation and parameter optimization to obtain the parameter extraction results; then calculating the fitting error between the simulation results and preset parameters, and then comprehensively evaluating the current parameter extraction results in conjunction with preset physical constraints; if the evaluation is successful, the final extracted model parameters and evaluation results are output; otherwise, the parameter optimization strategy is adjusted and the process returns to the optimal path execution step for iterative optimization. This invention enables efficient and stable parameter extraction from device physical models.

Description

Technical Field

[0001] This invention relates to the field of semiconductor device parameter extraction technology, specifically a method for automatic extraction of semiconductor device parameters based on an AI agent. Background Technology

[0002] Device models describe the terminal characteristics of a device through current-voltage (IV), capacitance-voltage (CV) characteristics, and the transport processes of charge carriers within the device. These models should reflect the device's characteristics across all operating regions. Semiconductor device models serve as a bridge between circuit fabrication and circuit design; the performance of the device determines the performance of the designed circuit. With the development of technologies such as wireless communication, the demand for circuit performance continues to grow, leading to a continuous shrinking of semiconductor device feature sizes and increasingly complex device models. To characterize accurate physical models of devices, especially compact models (analytical mathematical models established based on approximations and simplifications of physical models, possessing simpler mathematical expressions and enabling faster computation while maintaining accuracy), a large number of parameters often need to be introduced and solved numerically. However, solving these parameters presents some challenges:

[0003] 1. Traditional parameter extraction processes rely heavily on human experience, requiring engineers to manually select parameters, configure optimization strategies, set search spaces, and adjust fitting processes, resulting in low modeling efficiency and long cycles.

[0004] 2. In the traditional parameter extraction process, the lack of physical constraints or physical consistency checks can easily lead to the optimized parameters having a good numerical fit but not conforming to the actual physical behavior of the device, thus affecting the reliability and generalizability of the model.

[0005] Model parameter extraction is a technique used in electronics to obtain key parameters of mathematical models of semiconductor devices and circuit elements. With the development of Large Language Models (LLMs), they have demonstrated powerful capabilities in natural language processing, information induction and compression, and understanding-based decision generation. Researchers have developed efficient device model parameter extraction tools by leveraging LLM intelligent agents to interact with the external environment. Despite the significant potential of LLMs in automating device parameter extraction, several challenges remain:

[0006] 1. General-purpose large language models lack semiconductor device expertise and engineering constraints. Directly using them for parameter extraction can lead to illusions or erroneous reasoning, and cannot meet engineering accuracy requirements.

[0007] 2. Model parameter extraction usually requires multiple verification steps, but the decision-making and planning capabilities of large language models are limited, which may ultimately lead to errors in the parameter extraction results.

[0008] 3. The existing knowledge base cannot dynamically integrate users' private test data and process knowledge, making it difficult to adapt to different processes and different device types. Summary of the Invention

[0009] To address the aforementioned technical problems in existing technologies, this invention proposes an automatic parameter extraction method for semiconductor devices based on the fusion of AI agents and knowledge graphs. This method, while ensuring data security, automatically understands modeling tasks, invokes EDA simulation and optimization tools, and completes high-precision parameter extraction through knowledge-enhanced reasoning. The specific technical solution is as follows:

[0010] An automatic parameter extraction method for semiconductor devices based on AI agents includes:

[0011] Step 1: Receive the parameter extraction task input by the user. The task information includes the device type, the name of the model used, and the target parameter set.

[0012] Step 2: Through the retrieval enhancement generation module, retrieve knowledge information related to the current parameter extraction task from the pre-built domain knowledge base, and input the retrieval results as context information into the large language model;

[0013] Step 3: The large language model performs a structured decomposition of the parameter extraction task based on the task description and retrieved knowledge information, resulting in several sub-tasks;

[0014] Step 4: The large language model generates multiple candidate parameter extraction paths based on the task decomposition results and the physical dependencies of device parameters;

[0015] Step 5: The large language model scores the generated multiple candidate parameter extraction paths using preset evaluation rules and selects the optimal path;

[0016] Step 6: Execute the optimal path, and use external tools to perform device model simulation and parameter optimization to obtain the current parameter extraction results;

[0017] Step 7: Calculate the fitting error between the current model simulation results and the user-provided test data;

[0018] Step 8: The large language model performs a comprehensive evaluation of the current parameter extraction results; if the evaluation passes, proceed to Step 9; otherwise, adjust the parameter optimization strategy and return to Step 6 for iterative optimization.

[0019] Step 9: Output the final extracted model parameters and evaluation results.

[0020] Furthermore, the domain knowledge base in step 2 includes a text vector database and a device knowledge graph; the text vector database is used to store document fragments and their semantic vector representations; the device knowledge graph is used to represent structured knowledge of device types, model parameters and their physical relationships, and to establish an association with the text vector database.

[0021] Furthermore, the device knowledge graph is designed with an incremental update mechanism to continuously absorb new device modeling knowledge. The specific update process of the incremental update mechanism includes:

[0022] Step a1: Receive and format the new device model-related data;

[0023] Step a2: Perform text cleaning and semantic segmentation on the input data;

[0024] Step a3: Using a large language model and a preset information extraction prompt template, extract the semantically segmented text information and identify the knowledge entities, attributes and their relationships in the text;

[0025] Step a4: Using a large language model and by constructing evaluation prompt words, perform a quality assessment of the consistency and completeness analysis of the extraction results in step a3, and decide whether to re-execute step a3 based on the evaluation results. If missing information is detected, step a3 is re-executed.

[0026] Step a5: Calculate the semantic similarity between the newly extracted knowledge entity nodes and the existing nodes in the domain knowledge base, and based on the similarity calculation results, perform attribute conflict judgment and add nodes to the device knowledge graph;

[0027] Step a6: Add the new node and its attributes and relationships to the device knowledge graph, or update or merge the attributes of existing nodes.

[0028] Furthermore, in step a5, the process of determining attribute conflicts and adding nodes to the device knowledge graph specifically includes:

[0029] When the semantic similarity exceeds a preset threshold, it is determined that the new node and the existing node belong to the same knowledge entity or have a high correlation, and enters the attribute conflict judgment or update process: when it is determined that there is an attribute conflict, the large language model is used to perform semantic analysis and unified processing and update to generate a consistent attribute expression, that is, update the node attribute; if it is determined that there is no attribute conflict, the new attribute is merged into the corresponding existing node.

[0030] When the semantic similarity is lower than a preset threshold, the node is determined to be a new knowledge entity and added as a new node to the domain knowledge base.

[0031] Furthermore, each candidate parameter extraction path includes a parameter extraction strategy, which includes parameter extraction order, parameter setting range, and convergence requirements of the optimization function.

[0032] Furthermore, the evaluation rule in step 5 is implemented by designing a scoring function, which is obtained by a weighted combination of measuring the degree to which the path solves the core problem, evaluating the completeness of the path's steps, judging the logical confidence of the path, and the source of the basis.

[0033] Furthermore, the execution process of the optimal path in step 6 specifically includes:

[0034] Step b1: Input the optimal path and parse the corresponding task flow;

[0035] Step b2: Initialize the state machine, map the task nodes in the task flow to state nodes, and load the corresponding state transition rules;

[0036] Step b3: Based on the current state node and task context, determine whether it is necessary to call external tools through few-shot learning;

[0037] Step b4: If it is necessary to call external tools, generate a tool call request based on the current state;

[0038] Step b5: Call an external tool and record the execution status, return results, and execution timestamp;

[0039] Step b6: Update the state machine state based on the execution result and state transition rules;

[0040] Step b7: Determine if the termination state has been reached. If it has, proceed to step 7; otherwise, return to step b3 to continue execution.

[0041] Furthermore, the fitting error in step 7 is calculated using the root mean square error function.

[0042] Furthermore, the comprehensive evaluation in step 8 considers both the fitting error and the preset physical constraints. When the fitting error is still higher than the preset threshold or the extracted parameters do not meet the preset physical constraints, the model re-plans the optimization strategy based on the parameters that may need to be adjusted according to the error analysis, and executes step 6 to enter the next round of optimization iteration. When the fitting error meets the preset threshold and the parameters meet the physical constraints, the parameter extraction process is considered complete.

[0043] Compared with existing technologies, the method of this invention can achieve efficient and stable parameter extraction of device physical models under complex parameter coupling conditions, reduce the degree of manual dependence in the parameter extraction process, and provide reliable model parameter support for RF device modeling and circuit design. Attached Figure Description

[0044] Figure 1 This is a main flowchart of an automatic extraction method for semiconductor device parameters based on an AI agent according to this embodiment;

[0045] Figure 2 This is a schematic diagram of the device knowledge graph update process in this embodiment;

[0046] Figure 3 This is a schematic diagram of the execution process of the optimal path in this embodiment. Detailed Implementation

[0047] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0048] like Figure 1 As shown, this embodiment provides a method for automatically extracting semiconductor device parameters based on an AI agent, including:

[0049] Step 1: The user inputs the model from which the device parameters need to be extracted.

[0050] Users input parameters through the system interface to extract task information, including the type of device to be modeled, the name of the model used, and the set of target parameters to be extracted. After receiving the above information, the AI ​​agent system structures and organizes the task content, generates a standardized task description, and uses it as input for subsequent AI agent planning and decision-making.

[0051] Step 2: Retrieve model information using the RAG module.

[0052] The system utilizes the Retrieval Enhancement Generation (RAG) module to retrieve relevant knowledge information from the knowledge base. Specifically, based on the user-input task description, it performs keyword retrieval, semantic vector retrieval, and structured knowledge graph retrieval to obtain relevant document fragments, semantically similar knowledge, and device parameter relationship information, respectively. The multi-source retrieval results are then fused and filtered, and the final retrieval results are input as contextual information into the large language model to enhance the model's domain understanding of the current parameter extraction task.

[0053] The RAG module comprises a vector knowledge base and a knowledge graph. The system first parses semiconductor device model documents, parameter specification documents, and experimental data, and then constructs a domain knowledge base. This domain knowledge base includes a text vector database and a device knowledge graph. The text vector database stores original document fragments and their semantic vector representations. The device knowledge graph represents structured knowledge such as device types, model parameters, and their physical relationships, and records the index information of corresponding text fragments in the graph nodes to establish a connection between the device knowledge graph and the text vector database.

[0054] Furthermore, to ensure the system can continuously absorb new device modeling knowledge, this invention designs an incremental update mechanism for the device knowledge graph, such as... Figure 2 As shown, its update process includes the following steps:

[0055] Step a1: Input relevant information about the device model.

[0056] The system receives semiconductor device model-related materials used to build a domain knowledge base. These materials include, but are not limited to, semiconductor device model documentation, device physical model formulas, parameter definition manuals, technical papers, device test reports, and process specification documents. The system first performs standardized formatting on the input data and converts it into processable text data, providing the foundation for subsequent text processing and knowledge extraction.

[0057] Step a2: Perform text cleaning and semantic segmentation on the data.

[0058] Input data undergoes text preprocessing to improve the accuracy of subsequent information extraction. Preprocessing specifically includes removing irrelevant symbols, headers and footers, table of contents information, repeated paragraphs, and non-text characters from the document, and standardizing the text. Subsequently, the text is segmented based on semantic information, dividing long texts into multiple independent semantic units according to semantic relevance, and using a sliding window strategy to maintain the continuity of contextual information.

[0059] Step a3: The large model automatically extracts nodes, attributes, and relationships.

[0060] The semantically segmented text is input into the large language model one by one. The model is guided to identify knowledge entities, attributes and their relationships in the text by using a preset information extraction prompt template.

[0061] Step a4: The large model evaluates the completeness of the extracted information.

[0062] The system utilizes a large language model to evaluate the quality of the information extraction results obtained in step a3. Specifically, it constructs evaluation prompts to guide the model in performing consistency and completeness analysis on the current extraction results. When potential missing information is detected, the model re-executes step a3, thereby improving the coverage and accuracy of knowledge extraction.

[0063] Step a5: Calculate the semantic similarity with existing nodes.

[0064] For the nodes of the knowledge entities obtained in step a4, the system generates corresponding semantic vectors using a text embedding model and calculates their similarity with the vector representations of existing nodes in the knowledge base. By calculating cosine similarity, the system determines whether the newly extracted node represents the same or highly related knowledge entity as an existing node. The calculation function is as follows:

[0065]

[0066] A and B represent two different sets of nodes.

[0067] When the similarity exceeds a preset threshold, the node is considered to belong to the same knowledge entity or have a high degree of correlation with existing nodes, and enters the subsequent conflict judgment or attribute update process.

[0068] Step a5.1: Determine attribute conflicts.

[0069] When step a5 determines that a new node has a semantic correspondence with an existing node in the knowledge base, the system further compares the attribute information of the two. If the newly extracted attribute differs significantly from the existing attribute in terms of numerical range, unit, or semantic meaning, it is determined to be an attribute conflict. The system uses a large language model and relevant rules to make the judgment to determine whether there is a real conflict between the attributes or simply different expressions, specifically including:

[0070] Step a5.1.1: Update attributes.

[0071] When step a5.1 determines that there is an attribute conflict, the system uses a large language model to perform semantic analysis and unified processing through attribute update prompts, comprehensively judges different attribute values, updates and generates consistent attribute expressions, and retains the corresponding data source information.

[0072] Step a5.1.2: Attribute merging.

[0073] When step a5.1 determines that the attributes do not conflict, that is, when the newly extracted node attributes are considered to be consistent with the existing node attributes or to be supplementary information, the system uses the large language model to semantically integrate the attribute information according to the preset attribute merging prompt words, and merges and supplements it into the corresponding knowledge graph node.

[0074] Step a5.2: Add nodes to the knowledge graph.

[0075] When the semantic similarity between the new node calculated in step a5 and the existing nodes in the knowledge base is lower than a preset threshold, the system determines that the node is a new knowledge entity and adds it to the knowledge base as a new node. At the same time, the relevant attribute information and relationship information extracted in step a3 are written into the knowledge graph structure, thereby expanding the node set of the domain knowledge base.

[0076] Step a6: Update the knowledge graph.

[0077] After adding nodes, updating attributes, or merging attributes, the system writes the processed knowledge information into the knowledge graph database and updates the corresponding vector indexes to support subsequent semantic retrieval and retrieval enhancement generation tasks. By continuously receiving new device data and repeating the above process, the knowledge base can be continuously expanded and improved, thereby providing more accurate and richer domain knowledge support for subsequent parameter extraction tasks.

[0078] Step 3: Decompose the parameter extraction task.

[0079] The large language model performs a structured decomposition of the overall parameter extraction task based on the task description and the domain knowledge information retrieved in step 2. Specifically, the model divides the parameter extraction task into several sub-tasks based on the parameter structure and physical dependencies of the device model. In this way, the complex parameter extraction problem is broken down into multiple executable small-scale optimization problems, thereby reducing the coupling between parameters and improving extraction efficiency.

[0080] Step 4: Generate candidate extraction paths.

[0081] Based on the task decomposition results obtained in step 3, the large language model generates multiple candidate parameter extraction paths by combining the physical dependencies between device parameters and existing domain knowledge. Each path contains a complete parameter extraction strategy, including the parameter extraction order, parameter setting range, and convergence requirements of the optimization function.

[0082] Step 5: Evaluate and select the optimal path.

[0083] The large language model scores the candidate parameter extraction paths obtained in step 4 using evaluation rules. The scoring function is as follows:

[0084]

[0085] C(P) measures the degree to which the core problem is solved, F(P) assesses the completeness of the steps, L(P) judges the logical confidence and the source of the basis, and W is the weight of each item. The system finally selects the path with the highest score as the optimal path to execute.

[0086] Step 6: Execute the optimal path extraction task.

[0087] Based on the optimal parameter extraction path selected in step 5, the system generates the corresponding parameter extraction execution flow and configures the script for executing the extraction task. Simultaneously, it uses external tools to simulate the electrical characteristics of the device model and combines this with optimization algorithms to search and update the model parameters, thus completing the current parameter extraction task.

[0088] like Figure 3 As shown, the execution process of the optimal path in step 6 is implemented as follows:

[0089] Step b1: Input the optimal path.

[0090] The system reads the optimal parameter extraction path determined in step 5 and parses the corresponding task flow according to the path, including task nodes such as parameter initialization and parameter update, to provide a process basis for subsequent execution.

[0091] Step b2: Initialize the state machine.

[0092] The system executes according to a pre-built state machine, maps each task node in the parameter extraction process to a state node, and loads the corresponding state transition rules to describe the execution order and dependencies between each task step, while initializing state variables and execution record information.

[0093] Step b3: Learn whether to invoke the tool using a small number of samples.

[0094] Based on the current state node and task context, the system provides hints to the large language model through few-shot learning, enabling the model to determine whether the current task step requires calling an external tool module. If it determines that a tool needs to be called, step b4 is executed; otherwise, step b6 is executed.

[0095] Step b4: Generate a tool invocation request based on the current state.

[0096] The system generates a tool call request based on the task type corresponding to the current status node, and fills in the relevant call parameters according to the predefined parameter template, including information such as model type, parameters to be optimized, and input data.

[0097] Step b5: Call the external tool and record it.

[0098] The system sends call requests to external tools through a unified interface protocol to execute corresponding computational tasks. Simultaneously, the system records information such as the execution status of the tool calls, return results, and execution timestamps for subsequent process control and result tracking.

[0099] Step b6: Update the state according to the state transition rules.

[0100] The system updates the state machine state based on the execution result of the current step and the preset state transition rules, and determines the next executable task node.

[0101] Step b7: Determine whether the termination state has been reached.

[0102] The system determines whether the current state machine has reached the preset termination state. If the termination state has been reached, the system proceeds to step 7 to calculate the error and evaluate the result; if the termination state has not been reached, the system returns to step b3 to continue executing the next task step.

[0103] Step 7: Calculate the parameter error value RMS.

[0104] The system calculates the parameter fitting error based on simulation results and user-provided test data. Root mean square error is used to ensure high fitting accuracy of the model within the critical operating range.

[0105] First, define the relative error function ej(i), where Tj(i) is the value of the j-th target dataset at the i-th point, and Sj(i) is the value of the j-th simulation dataset at the i-th point.

[0106]

[0107] Define the root mean square error function as RMS. Here, P is the number of input data points with indices j = 0, 1, ..., p-1, nj is the number of data points in the j-th input, Wj is the weight coefficient of the j-th input, and ej(i) is the relative error of the j-th dataset at the i-th point.

[0108]

[0109] The calculated error value is used to evaluate the fitting effect of the current parameter combination.

[0110] Step 8: Evaluate whether the parameters and errors are reasonable.

[0111] The large language model combines the error and simulation curve characteristics obtained in step 7 to comprehensively evaluate the current parameter extraction results. When the error is still higher than the preset threshold or the extracted parameters do not meet the preset physical constraints, the model re-plans the optimization strategy based on the parameters that may need to be adjusted according to the error analysis, and executes step 6 to enter the next round of optimization iteration; when the error meets the preset accuracy requirements and the parameters meet the physical constraints, the parameter extraction process is considered complete.

[0112] Step 9: Output the parameter extraction results.

[0113] The system outputs the final model parameters and related evaluation results, which may include extracted parameter values, error indices, and comparison graphs of fitted curves.

[0114] In summary, this invention constructs a domain knowledge base comprising a text vector database and a device knowledge graph, and utilizes Retrieval Enhanced Generation (RAG) technology to provide domain knowledge support for large language models. This enables the model to understand the device model structure, parameter meanings, and their physical relationships, thereby achieving automated planning and task decomposition of the extraction process during parameter extraction. This reduces the reliance on engineer experience in traditional parameter extraction methods and improves the automation level and modeling efficiency of the parameter extraction process. Furthermore, by constructing an iterative parameter extraction mechanism of simulation-evaluation-optimization, and simultaneously introducing error analysis and physical constraint verification during the evaluation process, parameter optimization not only reduces the error between simulation results and measurement data but also ensures that the extracted parameters meet the physical laws of the device, thereby improving the physical rationality and modeling stability of the model parameters.

[0115] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatically extracting semiconductor device parameters based on an AI intelligent agent, characterized in that, include: Step 1: Receive the parameter extraction task input by the user. The task information includes the device type, the name of the model used, and the target parameter set. Step 2: Through the retrieval enhancement generation module, retrieve knowledge information related to the current parameter extraction task from the pre-built domain knowledge base, and input the retrieval results as context information into the large language model; Step 3: The large language model performs a structured decomposition of the parameter extraction task based on the task description and retrieved knowledge information, resulting in several sub-tasks; Step 4: The large language model generates multiple candidate parameter extraction paths based on the task decomposition results and the physical dependencies of device parameters; Step 5: The large language model scores the generated multiple candidate parameter extraction paths according to the preset evaluation rules and selects the optimal path; the evaluation rules are implemented by designing a scoring function, which is obtained by a weighted combination of measuring the degree to which the path solves the core problem, evaluating the completeness of the steps of the path, judging the logical confidence of the path, and the source of the basis; Step 6: Execute the optimal path, and use external tools to perform device model simulation and parameter optimization to obtain the current parameter extraction results; the execution process of the optimal path specifically includes: Step b1: Input the optimal path and parse the corresponding task flow; Step b2: Initialize the state machine, map the task nodes in the task flow to state nodes, and load the corresponding state transition rules; Step b3: Based on the current state node and task context, determine whether it is necessary to call external tools through few-shot learning; Step b4: If it is necessary to call external tools, generate a tool call request based on the current state; Step b5: Call an external tool and record the execution status, return results, and execution timestamp; Step b6: Update the state machine state based on the execution result and state transition rules; Step b7: Determine if the termination state has been reached. If it has, proceed to step 7; otherwise, return to step b3 and continue execution. Step 7: Calculate the fitting error between the current model simulation results and the user-provided test data; Step 8: The large language model performs a comprehensive evaluation of the current parameter extraction results, taking into account both the fitting error and the preset physical constraints. If the evaluation passes, proceed to Step 9; otherwise, adjust the parameter optimization strategy and return to Step 6 for iterative optimization. Step 9: Output the final extracted model parameters and evaluation results.

2. The method according to claim 1, characterized in that, The domain knowledge base in step 2 includes a text vector database and a device knowledge graph; the text vector database is used to store document fragments and their semantic vector representations; the device knowledge graph is used to represent structured knowledge of device types, model parameters and their physical relationships, and to establish an association with the text vector database.

3. The method according to claim 2, characterized in that, The device knowledge graph is designed with an incremental update mechanism to continuously absorb new device modeling knowledge. The specific update process of the incremental update mechanism includes: Step a1: Receive and format the new device model-related data; Step a2: Perform text cleaning and semantic segmentation on the input data; Step a3: Using a large language model and a preset information extraction prompt template, extract the semantically segmented text information and identify the knowledge entities, attributes and their relationships in the text; Step a4: Using a large language model and by constructing evaluation prompt words, perform a quality assessment of the consistency and completeness analysis of the extraction results in step a3, and decide whether to re-execute step a3 based on the evaluation results. If missing information is detected, step a3 is re-executed. Step a5: Calculate the semantic similarity between the newly extracted knowledge entity nodes and the existing nodes in the domain knowledge base, and based on the similarity calculation results, perform attribute conflict judgment and add nodes to the device knowledge graph; Step a6: Add the new node and its attributes and relationships to the device knowledge graph, or update or merge the attributes of existing nodes.

4. The method according to claim 3, characterized in that, In step a5, the process of determining attribute conflicts and adding nodes to the device knowledge graph specifically includes: When the semantic similarity exceeds a preset threshold, it is determined that the new node and the existing node belong to the same knowledge entity or have a high correlation, and enters the attribute conflict judgment or update process: when it is determined that there is an attribute conflict, the large language model is used to perform semantic analysis and unified processing and update to generate a consistent attribute expression, that is, update the node attribute; if it is determined that there is no attribute conflict, the new attribute is merged into the corresponding existing node. When the semantic similarity is lower than a preset threshold, the node is determined to be a new knowledge entity and added as a new node to the domain knowledge base.

5. The method according to claim 1, characterized in that, Each candidate parameter extraction path includes a parameter extraction strategy, which includes parameter extraction order, parameter setting range, and convergence requirements of the optimization function.

6. The method according to claim 1, characterized in that, The fitting error in step 7 is calculated using the root mean square error function.

7. The method according to claim 1, characterized in that, The comprehensive evaluation in step 8 considers both the fitting error and the preset physical constraints. When the fitting error is still higher than the preset threshold or the extracted parameters do not meet the preset physical constraints, the model re-plans the optimization strategy according to the parameters that need to be adjusted based on the error analysis, and executes step 6 to enter the next round of optimization iteration. When the fitting error meets the preset threshold and the parameters meet the physical constraints, the parameter extraction process is considered complete.

Images