Search

CN-122021520-A - Automatic semiconductor device parameter extraction method based on AI intelligent agent

CN122021520ACN 122021520 ACN122021520 ACN 122021520ACN-122021520-A

Abstract

The invention discloses an automatic semiconductor device parameter extraction method based on an AI intelligent agent, which comprises the steps of receiving a parameter extraction task input by a user, searching task related information through a RAG module and inputting a large language model, then carrying out structural decomposition on the task according to the searched information through the large language model, generating a plurality of candidate parameter extraction paths based on a task decomposition result and the physical dependency relationship of device parameters, grading, selecting an optimal path, executing the optimal path, simultaneously carrying out device model simulation and parameter optimization to obtain a parameter extraction result, then calculating fitting errors of the simulation result and preset parameters, carrying out comprehensive evaluation on the current parameter extraction result by combining with preset physical constraints, outputting finally extracted model parameters and evaluation results, otherwise, adjusting a parameter optimization strategy and returning to execute the optimal path step for iterative optimization. The invention can realize the efficient and stable parameter extraction of the physical model of the device.

Inventors

  • REN KUN
  • LIU TIAN
  • GUO SIJIA
  • GU HUI

Assignees

  • 杭州电子科技大学

Dates

Publication Date
20260512
Application Date
20260415

Claims (9)

  1. 1. An automatic semiconductor device parameter extraction method based on an AI intelligent agent is characterized by comprising the following steps: Step 1, receiving a parameter extraction task input by a user, wherein the information of the task comprises a device type, a name of a used model and a target parameter set; Step 2, searching knowledge information related to the current parameter extraction task from a pre-constructed domain knowledge base through a search enhancement generation module, and inputting a search result as context information into a large language model; step 3, the large language model carries out structural decomposition on the parameter extraction task according to the task description and the retrieved knowledge information to obtain a plurality of subtasks; Step 4, the large language model generates a plurality of candidate parameter extraction paths based on the task decomposition result and the physical dependency relationship of the device parameters; step 5, scoring the generated multiple candidate parameter extraction paths by the large language model through a preset evaluation rule, and selecting an optimal path; executing the optimal path, and carrying out model simulation and parameter optimization of the device by calling an external tool to obtain a current parameter extraction result; Step 7, calculating fitting errors between the simulation result of the current model and test data provided by a user; step 8, comprehensively evaluating the current parameter extraction result by the large language model, if the evaluation passes, entering step 9, otherwise, adjusting the parameter optimization strategy and returning to step6 for iterative optimization; and 9, outputting the finally extracted model parameters and the evaluation result.
  2. 2. The method of claim 1, wherein the domain knowledge base in step 2 comprises a text vector database for storing document fragments and semantic vector representations thereof, and a device knowledge graph for representing structured knowledge of device types, model parameters, and their physical relationships, and establishing associations with the text vector database.
  3. 3. The method according to claim 2, wherein the device knowledge graph is designed with an incremental update mechanism for continuously absorbing new device modeling knowledge, and a specific update process of the incremental update mechanism includes: step a1, receiving and formatting new device model related data; step a2, performing text cleaning and semantic segmentation on input data; Step a3, extracting a prompt template by using a large language model and through preset information, extracting text information after semantic segmentation, and identifying knowledge entities, attributes and relations thereof in the text; Step a4, carrying out quality assessment of consistency and integrity analysis on the extraction result of the step a3 by utilizing a large language model and constructing an assessment prompt word, and determining whether to re-execute the step a3 according to the assessment result, wherein when the possible missing information is detected, the step a3 is re-executed; step a5, calculating semantic similarity between the newly extracted knowledge entity node and the existing node in the domain knowledge base, and judging attribute conflict and newly adding the node of the device knowledge graph according to a similarity calculation result; And a6, adding the newly added nodes, the attributes and the relations thereof to the device knowledge graph, or updating or merging the attributes of the existing nodes.
  4. 4. The method of claim 3, wherein in step a5, the node for performing attribute conflict determination and device knowledge graph is newly added, specifically comprising: when the semantic similarity exceeds a preset threshold, judging that the new node and the existing node belong to the same knowledge entity or have high relevance, and entering an attribute conflict judging or updating flow, wherein when the attribute conflict is judged, semantic analysis and unified processing are performed by using a large language model, and the attribute expression which is consistent is generated by updating is updated; And when the semantic similarity is lower than a preset threshold, judging the node as a new knowledge entity, and adding the new knowledge entity as a new node to the domain knowledge base.
  5. 5. The method of claim 1, wherein each candidate parameter extraction path contains a parameter extraction strategy comprising a parameter extraction order, a parameter set interval, an optimization function convergence requirement.
  6. 6. The method of claim 1, wherein the evaluation rule in step 5 is implemented by designing a scoring function that is a combination of a measure of how well the path solves the core problem, an evaluation of the path's step integrity, a determination of the path's logical confidence, and a weighting based on source.
  7. 7. The method according to claim 1, wherein the performing of the optimal path in step 6 specifically includes: step b1, inputting an optimal path and analyzing a corresponding task flow; Initializing a state machine, mapping task nodes in a task flow to state nodes, and loading corresponding state migration rules; step b3, judging whether an external tool needs to be called or not through a few sample learning mode based on the current state node and the task context; Step b4, if an external tool needs to be called, generating a tool calling request according to the current state; Step b5, calling an external tool and recording an execution state, returning a result and an execution time stamp; Step b6, updating the state of the state machine according to the execution result and the state transition rule; and b7, judging whether the termination state is reached, if so, entering the step 7, otherwise, returning to the step b3 to continue execution.
  8. 8. The method according to claim 1, wherein the fitting error in step 7 is calculated by a root mean square error function.
  9. 9. The method of claim 1, wherein the comprehensive evaluation in the step 8 considers both the fitting error and the preset physical constraint condition, and when the fitting error is still higher than the preset threshold or the extracted parameter does not satisfy the preset physical constraint condition, the model re-programs the optimization strategy according to the parameter that may need to be adjusted by the error analysis, and performs the step 6 to enter the next round of optimization iteration, and when the fitting error satisfies the preset threshold and the parameter satisfies the physical constraint condition, the parameter extraction process is considered to be completed.

Description

Automatic semiconductor device parameter extraction method based on AI intelligent agent Technical Field The invention relates to the technical field of semiconductor device parameter extraction, in particular to an automatic semiconductor device parameter extraction method based on an AI intelligent agent. Background Device models refer to end characteristics describing the device by current-voltage (I-V), capacitance-voltage (C-V) characteristics, and the transport of carriers in the device, which should be able to reflect the characteristics of the device in all operating regions. The semiconductor device model serves as a bridge between the fabrication of the connection circuit and the circuit designer, and the performance of the device determines the performance of the designed circuit. With the development of wireless communication technology and the like, the demand for circuit performance is continuously increasing, so that the feature size of a semiconductor device is continuously reduced, and the device model is also becoming more and more complex. In order to characterize a correct device physical model, in particular a compact model (the compact model is an analytic mathematical model established based on methods such as approximation and simplification of the physical model, the model has a simpler mathematical expression, and the model is realized at a faster calculation speed on the premise of ensuring calculation accuracy), a large number of parameters are often required to be introduced, and a numerical method is adopted for solving. However, the process of solving these parameters has a few challenges: 1. The traditional parameter extraction process is highly dependent on manual experience, and requires engineers to manually select parameters, configure optimization strategies, set search spaces and adjust fitting procedures, so that modeling efficiency is low and period is long. 2. In the traditional parameter extraction process, the lack of physical constraint or physical consistency check easily leads to that the parameters obtained by optimization are not in line with the actual physical behaviors of the device although being well fitted in numerical values, so that the reliability and popularization of the model are affected. Model parameter extraction is a technical process used for acquiring key parameters of mathematical models of semiconductor devices and circuit elements in the field of electronics. With the development of Large Language Models (LLM), they exhibit strong capabilities in the fields of natural language processing, information induction compression, and understanding decision generation. Researchers interact with external environments through LLM intelligent agents using tools to develop efficient device model parameter extraction tools. While LLM presents great potential in device parameter extraction automation, there are still some challenges: 1. the general large language model lacks the expertise and engineering constraint of a semiconductor device, and can be directly used for parameter extraction to generate illusion or error reasoning, so that the engineering precision requirement can not be met. 2. Model parameter extraction usually requires multi-step verification, while decision planning capability of a large language model is limited, and finally, parameter extraction results can be wrong. 3. The prior knowledge base can not dynamically integrate private test data and process knowledge of users, so that the prior knowledge base is difficult to adapt to different processes and different device types. Disclosure of Invention In order to solve the technical problems in the prior art, the invention provides an automatic extraction method for parameters of a semiconductor device based on fusion of an AI (advanced technology) agent and a knowledge graph, which can automatically understand a modeling task, call EDA (electronic design automation) simulation and optimization tools and complete high-precision parameter extraction through knowledge enhancement reasoning on the premise of ensuring data safety, and the specific technical scheme is as follows: An automatic semiconductor device parameter extraction method based on an AI intelligent agent comprises the following steps: Step 1, receiving a parameter extraction task input by a user, wherein the information of the task comprises a device type, a name of a used model and a target parameter set; Step 2, searching knowledge information related to the current parameter extraction task from a pre-constructed domain knowledge base through a search enhancement generation module, and inputting a search result as context information into a large language model; step 3, the large language model carries out structural decomposition on the parameter extraction task according to the task description and the retrieved knowledge information to obtain a plurality of subtasks; Step 4, the large language model generates a p