Passive device automatic parameter extraction method based on large language model and related equipment

Abstract

The embodiment of the invention discloses an automatic passive device parameter extraction method based on a large language model and related equipment, and the method achieves the automation of a parameter extraction process of a whole passive device through the interactive connection of a model optimizer and the large language model. Therefore, an automatic parameter extraction technology which is more efficient, more accurate and more in line with an actual application scene is provided; the executable optimization codes are provided for the model optimizer, so that optimization strategy information output by the large language model is fused with optimization operation characteristics in the model optimizer, rapid and accurate conversion and transmission of the optimization strategy information are achieved, automatic interaction connection between the model optimizer and the large language model is completed, and the optimization strategy information is optimized. High-degree automation is achieved, manual follow-up is not needed, and the passive device parameter extraction efficiency is improved. The invention is widely applied to the technical field of electronic circuits.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of electronic circuit technology, and in particular to an automated parameter extraction method and related equipment for passive devices based on a large language model. Background Technology

[0002] With the rapid development of technologies such as 5G / 6G communication, the Internet of Things, automotive millimeter-wave radar, and satellite internet, the operating frequencies of electronic systems have fully entered the radio frequency (RF) and even millimeter-wave domains. In these ultra-high frequency bands, the performance of every tiny passive component on a chip plays a decisive role in the success or failure of the entire system. Therefore, the market demand for high-precision RF device models is unprecedentedly urgent.

[0003] In addition, the chip industry is fiercely competitive with extremely rapid product iteration. Moore's Law, which traditionally improves performance by shrinking transistor size, has gradually slowed down, and the industry has shifted to a "beyond Moore" approach, integrating dies with different functions through heterogeneous integration and advanced packaging technologies. In this high-density packaging, not only does the number of passive components increase rapidly, but the electromagnetic coupling between them also becomes exceptionally complex, requiring increasingly complex models. The success of a design heavily depends on the accurate modeling of each passive component and its interactions.

[0004] As the application frequency bands of passive devices continue to rise, the models become increasingly complex, and the requirements for model accuracy also increase. Traditional parameter extraction methods require a large number of subjective judgments and manual operations by human experts. This is not only time-consuming, labor-intensive, and costly, but also severely slows down the design process and hinders the inheritance and scaling of model results, becoming a key bottleneck in the entire chip development cycle. Therefore, automating parameter extraction is a rigid requirement for shortening the R&D cycle, reducing labor costs, and seizing market opportunities. Summary of the Invention

[0005] To address the technical challenges in the field of electronic technology, such as the need for automated parameter extraction of passive devices, the present invention aims to provide a method and related equipment for automated parameter extraction of passive devices based on a large language model.

[0006] On one hand, embodiments of the present invention include an automated parameter extraction method for passive devices based on a large language model, wherein the automated parameter extraction method for passive devices based on a large language model includes:

[0007] Obtain the initial file; the initial file includes netlist data and measured data for passive devices;

[0008] Obtain the first prompt word; the first prompt word represents the target parameters that the user requires the passive device to achieve.

[0009] Perform at least one round of iterative process; each round of iterative process includes the following steps:

[0010] The data to be simulated is input into the model optimizer for simulation processing, and the characteristic response image and model parameter information output by the model optimizer are obtained; wherein, when the current iteration process is the first iteration process, the data to be simulated is the initial file, otherwise, the data to be simulated is the data obtained by optimizing the initial file according to the executable optimization code obtained in the previous iteration process;

[0011] Based on the characteristic response image and the model parameter information, obtain the second prompt word;

[0012] The first prompt word and the second prompt word are input into the large language model for processing; the first judgment value and optimization strategy information output by the large language model, or the second judgment value output by the large language model, are obtained.

[0013] Once the first judgment value is obtained, the next round of iteration is triggered based on the optimization strategy information;

[0014] Once the second judgment value is obtained, the execution of all the above iteration processes ends, and the optimization result is output.

[0015] Further, obtaining the second prompt word based on the characteristic response image and the model parameter information includes:

[0016] Obtain historical optimization records for passive devices; the historical optimization records include optimization results obtained from iterative processes that have been executed and model parameter information of passive devices;

[0017] Error data of the response image based on the aforementioned characteristics;

[0018] Sensitivity analysis was performed on the model parameter information;

[0019] The second prompt word is determined based on the model parameter information, the error data, and the sensitivity analysis results.

[0020] Further, obtaining the second prompt word based on the characteristic response image and the model parameter information includes:

[0021] Obtain historical optimization records for passive devices; the historical optimization records include optimization results obtained from iterative processes that have been executed and model parameter information of passive devices;

[0022] Error data of the response image based on the aforementioned characteristics;

[0023] Obtain historical optimization strategies for passive devices; the historical optimization strategies include optimization strategy information obtained from iterative processes that have already been executed;

[0024] Sensitivity analysis was performed on the model parameter information;

[0025] The second prompt word is determined based on the model parameter information, the error data, the optimization strategy information, and the sensitivity analysis results.

[0026] Furthermore, triggering the next iteration process based on the optimization strategy information includes:

[0027] Update the historical optimization strategy according to the optimization strategy information;

[0028] The executable optimized code is obtained by converting the optimization strategy information.

[0029] Execute the next iteration.

[0030] Further, the step of inputting the first prompt word and the second prompt word into a large language model for processing includes:

[0031] The first prompt word and the second prompt word are constrained to obtain constrained prompt words;

[0032] The output rules of the optimization strategy information of the large language model are constrained;

[0033] The constraint prompts are input into the large language model.

[0034] Further, the constraint processing of the first prompt word and the second prompt word to obtain constrained prompt words includes:

[0035] Set constraints for user requirement analysis, manual optimization experience tips, and optimization strategy output;

[0036] Based on the user demand parsing constraint rules, the manual optimization experience suggestion rules, and the optimization strategy output constraint rules, the first suggestion word and the second suggestion word are filtered to obtain the constraint suggestion word.

[0037] Furthermore, the automated parameter extraction method for passive devices based on a large language model also includes:

[0038] Set a first count threshold and multiple second count thresholds; each second count threshold is less than the first count threshold.

[0039] When the number of iterations executed reaches the first threshold, the execution of all iterations ends and the optimization result is output.

[0040] When the number of iterations has reached any of the second thresholds, based on the optimization strategy information obtained from the iterations, prompt word modification guidance information is generated, and the first prompt word is modified according to the prompt word modification guidance information.

[0041] Furthermore, the step of generating prompt word modification guidance information based on the optimization strategy information obtained from the executed iteration process includes:

[0042] Obtain the average value of the optimization strategy information obtained from the executed iteration process;

[0043] Obtain the deviation information of the average value relative to the first prompt word;

[0044] Based on the deviation information and the second threshold number already reached, the modification range for the first prompt word is determined; wherein the modification range is positively correlated with both the deviation information and the second threshold number already reached.

[0045] Based on the stated modification range, the prompt word modification guidance information is generated.

[0046] On the other hand, embodiments of the present invention also include a computer device, including a memory and a processor, the memory for storing at least one program, and the processor for loading at least one program to execute the passive device automated parameter extraction method based on a large language model in the embodiments.

[0047] On the other hand, embodiments of the present invention also include a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the passive device automated parameter extraction method based on a large language model in the embodiments.

[0048] The beneficial effects of the embodiments of the present invention are as follows: The automated parameter extraction method for passive devices based on a large language model in the embodiments automates the entire parameter extraction process for passive devices through the interactive connection between the model optimizer and the large language model, thereby providing a more efficient, accurate, and practically applicable automated parameter extraction technology; by providing executable optimization code to the model optimizer, the optimization strategy information output by the large language model integrates the optimization operation characteristics within the model optimizer, realizing the rapid and accurate conversion and transmission of optimization strategy information, completing the automated interactive connection between the model optimizer and the large language model, achieving a high degree of automation, eliminating the need for manual follow-up, and improving the efficiency of passive device parameter extraction. Attached Figure Description

[0049] Figure 1This is a schematic diagram illustrating the steps of the automated parameter extraction method for passive devices based on a large language model in the embodiment.

[0050] Figure 2 This is a schematic diagram illustrating the process and principle of the automated parameter extraction method for passive devices based on a large language model in the embodiment.

[0051] Figure 3 This is a schematic diagram illustrating the principle of constraining prompt words and the large language model in the embodiment;

[0052] Figure 4 This is a schematic diagram showing the response images and errors of various characteristics of the passive device before optimization in the embodiment.

[0053] Figure 5 This is a schematic diagram showing the response images and errors of various characteristics of the passive device after optimization in the embodiment.

[0054] Figure 6 This is a schematic diagram illustrating the historical errors of the multi-round iterative optimization process performed on the passive device in the embodiment. Detailed Implementation

[0055] Terminology Explanation:

[0056] Parameter extraction, also known as model parameter extraction, is a technical process in the field of electronics used to obtain key parameters of mathematical models of semiconductor devices and circuit elements. Its core objective is to transform the physical characteristics of devices into quantifiable parameters through analytical calculation, numerical optimization, or intelligent algorithms, so as to ensure that the circuit simulation results are consistent with the measured data. There are application requirements for parameter extraction in fields such as integrated circuit design, power electronic device modeling, and semiconductor process development.

[0057] SPICE netlist: SPICE netlist is a plain text-based, standardized language that follows a specific syntax. It is used to describe all the information of a circuit to circuit simulators (such as SPICE, HSPICE, Spectre), and serves as the "blueprint" or "script" for circuit simulation.

[0058] Model Optimizer: The model optimizer is a core algorithm engine embedded in the parameter extraction software. It can automatically and intelligently adjust a large number of parameters in the SPICE model through the operation settings of engineers, so that the error between the simulation results of the model (such as IV and CV curves) and the actual hardware measurement data is minimized, thereby realizing parameter extraction.

[0059] Radio frequency (RF): Radio frequency (RF) generally refers to the electromagnetic spectrum with frequencies ranging from 3 kHz to 300 GHz. In the field of circuits, it specifically refers to circuits, systems, and related technologies that operate in this frequency band. Its core characteristic is that the wavelength of the signal is on the same order of magnitude as the physical size of the circuit.

[0060] Moore's Law: An industry observation and prediction put forward by Intel co-founder Gordon Moore in 1965: The number of transistors that can be placed on an integrated circuit (IC) will roughly double every 18 to 24 months.

[0061] Beyond Moore: "Beyond Moore" is a semiconductor technology development path that complements "Moore's Law." It doesn't simply aim to increase digital computing density by shrinking transistor size (More Moore), but rather to enhance the overall value of the system by diversifying functionality and integrating non-digital functions based on different physical principles (such as radio frequency, power management, and sensors) into the chip system.

[0062] Bayesian optimization is an efficient global optimization algorithm specifically designed to solve "black box" problems where the objective function has extremely high evaluation costs and unknown form. Instead of optimizing the expensive "black box" function itself, it first builds an inexpensive, probabilistic approximation model of the function (most commonly a Gaussian process). This model not only predicts the function value at any point but also provides information about the uncertainty of the prediction.

[0063] "Black box" model: In mathematical and engineering optimization, a "black box" problem refers to a class of problems in which we can give an input (a set of parameters) and get an output (an evaluation score, such as cost or error), but we have no idea about the functional relationship between the input and the output, nor can we obtain the internal structure of the function or its derivative (gradient) information.

[0064] The Levenberg-Marquardt (LM) algorithm is an iterative optimization algorithm specifically designed for solving nonlinear least squares problems. It is one of the most commonly used and classic methods for curve fitting tasks. The LM algorithm adaptively integrates two classic optimization methods: gradient descent and Gauss-Newton's method. The LM algorithm introduces a damping factor λ. When λ is large, the algorithm behaves similarly to robust gradient descent; when λ is small, the algorithm behaves similarly to the fast Gauss-Newton method. The algorithm dynamically adjusts the value of λ based on the fitting performance at each step, thus combining the advantages of both methods: maintaining stability when far from the optimal solution and achieving fast convergence when approaching the optimal solution.

[0065] Random algorithm: The simplest global optimization algorithm. Its strategy is to generate a series of candidate solutions completely randomly within the feasible region of the parameters, then evaluate the quality of each solution, and finally return the best one.

[0066] Genetic Algorithm: A global optimization heuristic algorithm that simulates the evolutionary process of "survival of the fittest" in the biological world to find the optimal solution.

[0067] Currently, parameter extraction for passive devices generally employs optimization fitting methods. Specifically, parameter extraction is treated as an optimization problem, where the model parameters of the equivalent circuit model of the passive device are continuously adjusted to minimize the error function between the simulated response and the actual measured data. The core drawback of traditional passive device parameter extraction lies in its heavy reliance on the expertise and experience of human specialists. This requires manually selecting model optimization parameters, setting parameter optimization ranges, choosing optimization functions, and adjusting simulation response optimization weights. The entire process is time-consuming and labor-intensive, and difficult to automate or scale up, significantly reducing efficiency. More importantly, as operating frequencies extend to the radio frequency (RF) or millimeter-wave bands, passive devices need to consider various parasitic effects, necessitating more complex equivalent circuit topologies for description. This drastically increases the number of model parameters and the complexity of the simulated output response, further increasing the optimization difficulty. Moreover, with the continuous development of integrated circuit manufacturing and packaging technologies, the size of passive devices is shrinking, their density is increasing, and their operating frequencies are rising. The rapid modeling and parameter extraction of passive devices has become a key factor affecting chip design efficiency and cost-effectiveness.

[0068] Regarding parameter extraction for passive devices, some related technologies and their shortcomings are as follows:

[0069] 1. Traditional Machine Learning / Data-Driven Approaches: These approaches treat parameter extraction as a regression or classification problem, training a machine learning model (such as a neural network, support vector machine, or decision tree) to learn from the device's measurement data (e.g., IV, CV, or S-parameters). Traditional machine learning methods require creating a large dataset, and the resulting models are end-to-end specialized models with limited functionality, poor interpretability, and low interactivity.

[0070] 2. Expert Systems / Knowledge Base Systems: These systems simulate the decision-making process of a model engineer by establishing a comprehensive "IF-THEN" rule base. The knowledge base stores facts and rules, and the inference engine applies these rules to draw conclusions or recommend actions based on the current situation. However, the knowledge in expert systems is explicit and manually entered, its boundaries are strictly defined, making maintenance and expansion extremely difficult, and it cannot handle any new situations outside the rule base.

[0071] 3. Bayesian Optimization: A global optimization algorithm focused on solving "expensive black-box" optimization problems, well-suited for parameter extraction scenarios requiring extensive and time-consuming simulations (such as SPICE or EM simulations). Bayesian optimization is a purely mathematical method, utilizing no physical knowledge or historical experience about the device. It is a highly specialized optimization algorithm. The LLM scheme is a more comprehensive automation framework, where optimization is merely one tool that can be invoked; the large language model can determine "when" and "how" the optimizer is used.

[0072] 4. A "Conversational Collaborative Modeling Assistant" based on a large language model: Utilizing the powerful natural language understanding and code generation capabilities of LLM, this assistant translates the vague, high-level instructions given by engineers in natural language into low-level, precise, and executable modeling software operation commands or scripts. While it still requires waiting for the engineer's instructions, it is essentially a powerful interactive assistant that understands semiconductor expertise, but it is not yet fully automated.

[0073] Large language models have shown remarkable potential and prospects in the field of device model parameter extraction, enabling the parameter extraction process to move from "mechanization" to "intelligence" and "automation." Large language models not only significantly lower the technical barriers to device modeling and parameter extraction, but also integrate massive amounts of domain knowledge, possessing global reasoning capabilities, dynamic error correction capabilities, and adaptive adjustment capabilities. This can greatly reduce the heavy reliance on human experts in the field of modeling and parameter extraction.

[0074] However, the application of large language models in the field of passive device parameter extraction is still quite immature. For example, as the operating frequency and application areas of passive devices continue to expand, the complexity of passive device models is increasing, and the number of models varies greatly. Large language models lack much professional knowledge and experience in device modeling and parameter extraction. Relying solely on large language models cannot transform complex "black box" problems into "white box" problems, resulting in output results with a certain degree of randomness and uncertainty. In addition, large language models lack a mechanism for deep reflection on historically generated optimization strategies, often getting stuck in long optimization stagnation during parameter extraction. When facing parameter extraction of similar or identical devices, the consistency and scalability of the output strategies of large language models are also poor.

[0075] Based on the above principles, this invention addresses the shortcomings of existing parameter extraction methods and the potential of large language models in parameter extraction. It proposes an innovative automated parameter extraction method for passive devices based on large language models, aiming to provide a more accurate and efficient parameter extraction method.

[0076] This embodiment provides an automated parameter extraction method for passive devices based on a large language model. (Refer to...) Figure 1The automated parameter extraction method for passive devices based on a large language model includes the following steps:

[0077] S1. Obtain the initial file;

[0078] S2. Obtain the first prompt word;

[0079] S3. Perform at least one round of iteration.

[0080] Any round of iteration includes the following steps:

[0081] S301. Input the data to be simulated into the model optimizer for simulation processing, and obtain the characteristic response image and model parameter information output by the model optimizer;

[0082] S302. Obtain the second prompt word based on the characteristic response image and model parameter information;

[0083] S303. Input the first prompt word and the second prompt word into the large language model for processing; obtain the first judgment value and optimization strategy information output by the large language model, or the second judgment value output by the large language model;

[0084] S304. Once the first judgment value is obtained, the next round of iteration is triggered based on the optimization strategy information;

[0085] S305. When the second judgment value is obtained, the entire iteration process ends and the optimization result is output.

[0086] In this embodiment, steps S1-S3 (including S301-S305) can be executed by a computer. The specific process and principle of steps S1-S3 are as follows: Figure 2 As shown. Figure 2 The red box in the diagram illustrates the process and principle of a single iteration. Figure 2 In the flowchart shown, rounded rectangles represent new modules created in the process by the present invention, including a data processing module, a large language model interaction module, a strategy conversion and transmission module, a prompt word module, and a successful optimization device classification knowledge base; right-angled rectangles represent user input and output data or existing modules, such as the initial SPICE netlist, measured data, user requirements, and model optimizer; and ellipses represent iterative convergence conditions.

[0087] In step S1, the acquired initial file includes the netlist data and measured data of passive devices. Specifically, the initial file includes the initial SPICE netlist of one or more passive devices and the corresponding actual measurement data. The user does not need to specify the type of passive device (e.g., capacitor, inductor, or resistor). If multiple passive devices are input simultaneously, the netlist data and measured data of each passive device can be sequentially placed into the initial file. This embodiment uses one passive device as an example for illustration.

[0088] In step S1, the measured data in the initial file is the basis for measuring the accuracy of the device model. For the extraction of parameters of RF passive devices, the measured S-parameters (including S11, S12, S21 and S22) of the passive device in a certain RF range can generally be provided as measured data.

[0089] In step S2, the user can edit the first prompt word based on the target parameters they want the passive device to achieve. Specifically, the user determines a final indicator of the passive device parameter extraction results based on their own needs, such as the root mean square error between the model's characteristic simulation data and the measured data being below 5%, and edits the first prompt word. The first prompt word can be used as the basis for the large language model to determine whether to generate an optimization strategy. The parameter extraction requirements are stored in a text file in the form of natural language commands to obtain the first prompt word. The large language model can automatically read the file of the first prompt word to parse the requirements and generate optimization strategies.

[0090] Reference Figure 1 and Figure 2 After completing steps S1-S2, at least one round of iteration is executed, i.e., at least one round of steps S301-S305. Generally, multiple rounds of iteration are required. In this embodiment, one round of iteration, i.e., steps S301-S305, is used as an example for explanation.

[0091] If this iteration is the first iteration, then in step S301, the initial file obtained in step S1 is used as the simulation data, and the simulation data is input into the model optimizer for simulation processing; if this iteration is not the first iteration (i.e., subsequent iterations after the first iteration), then in step S301, the executable optimization code obtained by executing steps S301-S305 in the previous iteration is input into the model optimizer to optimize the initial file, and the resulting data is used as the simulation data.

[0092] In step S301, the model optimizer automatically calls the input simulation data to perform the initial simulation. After simulation or optimization, it outputs model parameter information and several characteristic response images. Each characteristic response image contains the curves of the measured data and the simulated data, as well as their root mean square error (i.e., the error data corresponding to the characteristic response image). The model optimizer outputs these optimization results to the data processing module, which then executes step S302.

[0093] When executing step S302, the data processing module can acquire historical optimization records, specifically including optimization results obtained from iterations performed before the current iteration (e.g., optimized passive device model parameter information, characteristic response images, corresponding error data, and historical optimization strategies; if this iteration is the first iteration, i.e., no previous iterations exist, then the model's initial netlist information and initial error data are used as historical optimization records). The data processing module can further perform sensitivity analysis on model-related parameters such as model parameter information and then store the results. The model parameter information, error data, and sensitivity analysis results obtained in this iteration can be used as one of the prompt words (i.e., the second prompt word) when interacting with the large language model, assisting the large language model in generating optimization strategies.

[0094] In this embodiment, refer to Figure 2 When the data processing module generates the second prompt word in step S302, if the current iteration process is not the first iteration process, that is, if the iteration process has been executed before, it can also obtain the optimization strategy information obtained by the already executed iteration process, i.e., the historical optimization strategy. Based on the model parameter information, error data, sensitivity analysis results obtained in the current iteration process, and the historical optimization strategies obtained in previous iteration processes, the second prompt word is generated.

[0095] In this embodiment, step S303 is executed, whereby the first prompt word obtained in step S2 and the second prompt word obtained in step S302 of the current iteration process are input into the large language model for processing. Specifically, refer to... Figure 2 The API of the large language model can be called by the large language model interaction module, thereby realizing interaction with the large language model.

[0096] In this embodiment, before interacting with the large language model in step S303, strict constraints can be applied to the first and second prompt words input to the large language model to obtain the constraint prompt words. That is, the prompt words actually input to the large language model in step S303 are the constraint prompt words.

[0097] Specifically, refer to Figure 3First, constraints were imposed on the role and task definition of the large language model, enabling it to clearly define the parameter extraction task for passive devices, reduce the search and consideration of invalid information, and selectively call external or specific knowledge bases to help complete the task. Second, constraints were also imposed on the rules for the large language model to parse user requirements, including how to compare user parameter extraction files with real-time error files, when to output corresponding Boolean algebra instructions, and the action instructions after outputting Boolean algebra instructions. Simultaneously, constraint prompts were generated based on prompting rules derived from human optimization experience. For example, in the parameter extraction task for devices with certain types of netlist information, prompts were given to the large language model to select which parameters to optimize, which optimization algorithm is effective, and the setting range of optimization weight values. Constraint prompts were also generated based on several data files that assist the large language model in generating optimization strategies. These files include model SPICE netlist information, historical changes in model parameter values ​​and errors, sensitivity analysis results, and historical optimization strategy results. The large language model adaptively adjusts its optimization strategies based on these files generated by the auxiliary optimization strategies, that is, it performs multiple in-depth reflections on the optimization strategy in conjunction with historical data files, dynamically adjusting the optimization strategy output each time.

[0098] In this embodiment, before interacting with the large language model in step S303, the rules for the output of the large language model's optimization strategy can be constrained. Specifically, the update direction, generated prompts, and output format of the large language model's optimization strategy can be constrained. In this embodiment, such as... Figure 3As shown, the update of the optimization strategy for the constrained large language model output comes from six aspects, including selecting the feature response images to be optimized, selecting the model parameters for this optimization, setting the optimization range of the model parameters, selecting the optimization algorithm, determining whether to reset the values ​​of certain optimization parameters, and setting the optimization weight of each feature response image. First, setting the optimization order of feature response images: The model optimizer outputs several feature response images each time it optimizes. Different images may contain different physical meanings, and as the number of images gradually increases, optimizing multiple images together will reduce optimization efficiency. The large language model needs to determine and select the optimization order of these images. Second, selecting model optimization parameters: Through its own thinking and reasoning, combined with historical optimization results and sensitivity analysis, the large language model provides two sets of parameters: those participating in the current optimization and those not participating in the optimization, named Used and Unused, respectively. Third, the optimization range of model parameters: For parameters participating in optimization (Used), the large language model needs to set their optimization range to prevent some parameters from deviating from their physical meaning during optimization. Setting the optimization range also helps the optimization quickly find the optimal solution. Each model parameter has an initial optimization range [Min, Max]. During each optimization, the large language model needs to consider whether to reset the optimization range. Fourth, select the optimization algorithm: the large language model selects algorithms already existing within the model optimizer, including the Levenberg-Marquardt algorithm, the Random algorithm, and genetic algorithms. Extensive optimization has shown that the Levenberg-Marquardt algorithm, combined with sensitivity analysis, is significantly effective for parameter extraction of passive devices. These can serve as empirical indicators for the large language model to learn and as strategic guidance. Fifth, determine whether to reset the value of a certain optimization parameter: The following situations require prompting the large language model to determine whether to reset certain optimization parameters. The parameters are adjusted. For example, if the model parameters do not fall within the optimization range, do not conform to physical meaning, or the fit between the model and the data suddenly deteriorates after a certain optimization, it may be that some parameter values ​​have fallen into unreasonable values. The large language model can refer to the stored historical parameter values ​​to reset the unbalanced parameter values ​​back to their original values ​​or more reasonable values. The parameters that need to be reset are output in the format of [full parameter name, reset parameter value]. Sixth, set the optimization weight of each characteristic response image result: The optimization weight is an important parameter factor controlling the model optimization within the model optimizer. For two characteristic response images with large differences in root mean square error, if they are optimized at the same time, the optimization weight of the image with larger error can be appropriately increased, and the optimization weight of the other image can be appropriately decreased.For parameter extraction of RF passive devices, the characteristic response mainly refers to the S-parameters: S11, S12, S21, and S22. For these four S-parameters, the large language model needs to output weight strategies in a format similar to [1,4,4,1] in sequence. The large language model needs to update the optimization strategy according to the above six aspects and output the optimization strategy in a specific format according to the rules. Then, through the optimization strategy conversion and transmission module, the above optimization strategy is output as script code that can be executed inside the model optimizer. The macro automatically executes the code, thereby realizing the full automation of parameter extraction.

[0099] After completing such Figure 3 After constraining the prompt words and the large language model as shown, step S303 can be executed. In this embodiment, the constraint prompt words obtained after constraints, and the first and second prompt words before constraints, can be distinguished.

[0100] In this embodiment, when executing step S303, the large language model will automatically read the parameter request file (first prompt word) input by the user and parse the user's requirements. The parsing mainly checks whether the content of the second prompt word (such as real-time error information) meets the user's requirements.

[0101] In this embodiment, when executing step S303, if the large language model determines that the content represented by the second prompt word does not match the first prompt word (for example, the current real-time error status of passive devices does not meet the user's needs), then the large language model will output the first judgment value of the Boolean instruction "FALSE" to the data processing module, indicating that the optimization is not completed and the iteration continues. At the same time, the large language model will generate a specific optimization strategy information, which will be sent to the optimization strategy transmission module to trigger the execution of step S304, that is, to trigger the execution of the next round of iteration based on the optimization strategy information.

[0102] In this embodiment, during step S304, the optimization strategy conversion module automatically receives and stores the fixed-format optimization strategy file (i.e., optimization strategy information) generated by the large language model. The optimization strategy conversion module then parses the optimization strategy information, converting it into executable optimization code that the model optimizer can run. This allows the model optimizer to automatically call the code internally to complete the corresponding optimization steps, thereby achieving complete automation of parameter extraction. The optimization strategy conversion module in this embodiment includes some code languages ​​supported by the internal model optimizer of the parameter extraction software, ensuring that the optimization strategy output by the large language model can be completely and accurately converted into code supported by each model optimizer. The optimization strategy conversion module sends the executable optimization code to the model optimizer, enabling the model optimizer to run it during the next iteration.

[0103] In this embodiment, when performing step S304, refer to Figure 2 Furthermore, the optimization strategy conversion module can update historical optimization strategies based on optimization strategy information. Specifically, the optimization strategy conversion module adds the optimization strategy information obtained in the current iteration process to the historical optimization strategy (including optimization strategy information obtained in each of the previous iteration processes), thereby updating the historical optimization strategy. The updated historical optimization strategy is then used in the next iteration process to generate the second prompt word.

[0104] During each iteration, the model optimizer can automatically execute optimization steps based on the code transformed by the optimization strategy, displaying the real-time optimization progress. After each optimization, it outputs a new characteristic response image and new error information, and obtains the latest model parameter information. These optimized results are automatically sent to the data processing module.

[0105] In this embodiment, when executing step S303, if the large language model determines that the content represented by the second prompt word matches the first prompt word (e.g., the current real-time error status of passive devices meets user requirements), the large language model will output a second judgment value of the Boolean instruction "TRUE" to the data processing module, indicating that the optimization is complete. Simultaneously, a large language model optimization report will be generated, triggering step S305. Upon receiving the second judgment value, i.e., the "TRUE" instruction, the data processing module will stop the optimization iteration and output the optimization result. Specifically, refer to... Figure 2 After completing the parameter extraction tasks for all passive devices in sequence, the historical optimization information of successfully optimized passive devices is stored in a categorized knowledge base as a reference for subsequent large language model generation optimization strategies. The categorized knowledge base organizes and stores SPICE netlist information, historical optimization strategies, and optimization log files of different passive devices, which helps the large language model to deeply acquire parameter extraction experience and quickly provide optimization strategies when faced with devices with the same or similar netlist information. Then, the optimized SPICE netlists and optimization logs of all devices are output as optimization results.

[0106] The automated parameter extraction method for passive devices based on a large language model in this embodiment automates the entire parameter extraction process by connecting the model optimizer and the large language model. This provides a more efficient, accurate, and practically applicable automated parameter extraction technology. To achieve this goal, the method creates input constraint rules for prompt words and output optimization strategies from the large language model, and designs an optimization strategy conversion and transmission mechanism. It also creates an automated module that automatically converts the optimization strategies output by the large language model into executable optimization scripts within the model optimizer. Furthermore, it innovatively integrates model parameter sensitivity analysis, adaptive optimization strategy adjustment, and historical optimization information classification, storage, and transmission design, and extends the operating frequency band for passive device parameter extraction to the radio frequency band. By constraining the prompt words input to the large language model and incorporating human experience in passive device parameter extraction, the large language model can accurately grasp optimization requirements based on the parameter extraction task of passive devices, conduct in-depth thinking and knowledge retrieval in the professional parameter extraction field, and greatly reduce the high dependence on human experts. By constraining the optimization strategy output by the large language model, the large language model can accurately grasp the adjustable optimization settings within the model optimizer. This includes how to set the optimization order of the characteristic response image, filter the model's optimization parameters, set the parameter optimization range, select the optimization algorithm, reset the optimization parameter values, and adjust the optimization weights of the simulation response. This allows for more focused and targeted in-depth thinking and the generation of specific, concise optimization strategy text. Through an optimization strategy conversion and transmission mechanism, an automated module was created to convert the large language model's output optimization strategy into an executable optimization script within the model optimizer. This module automatically converts specific strategy text into script code that the model optimizer supports running. Macros automatically execute the code and automatically store and transmit the optimization results to the large language model, thus automating the entire passive device parameter extraction process. Furthermore, sensitivity analysis of model parameters helps identify redundancy and correlation among model parameters, revealing the key parameters that have the greatest impact on the total error of the characteristic response. This facilitates the large language model's strategic selection of optimization components, reduces ineffective optimizations often occurring in complex circuit topologies, and significantly improves the efficiency of parameter extraction. Adaptive optimization strategy adjustments help large language models summarize and reflect on optimization pitfalls, providing more efficient optimization strategies as optimization iterations progress. The design of classifying, storing, and transmitting historical optimization information, by storing historical optimization strategies for successfully extracted passive devices and optimization log files, establishes a categorized knowledge base for the large language model to reference and access. This helps the large language model deeply acquire parameter extraction experience and provide optimization strategies more quickly and accurately according to the classification of passive devices.

[0107] In summary, the automated parameter extraction method for passive devices based on a large language model in this embodiment automates the entire parameter extraction process by constraining the input prompts and output strategies of the large language model, performing rigorous optimization strategy conversion and transmission, and facilitating automated interaction between the model optimizer and the large language model. It integrates multiple parameter extraction strategies to provide a more efficient, accurate, and practically applicable parameter extraction technology for passive devices. By constraining the input prompts of the large language model, it incorporates human experience in passive device parameter extraction, enabling the large language model to clearly understand the parameter extraction task, accurately grasp optimization requirements, and engage in in-depth thinking and knowledge retrieval within the professional parameter extraction domain. This significantly reduces reliance on human experts and shortens the large language model's processing time. By constraining the optimization strategies output by the large language model, it enables the model to accurately grasp the adjustable optimization operation settings within the model optimizer, allowing for more focused and targeted in-depth thinking and the generation of specific, concise optimization strategy text, reducing invalid and redundant information in the large language model's output. Furthermore, this invention creates an automated module that converts the output strategy of a large language model into an executable optimization script within the model optimizer by optimizing the strategy conversion and transmission mechanism. This module can automatically convert specific strategy text into script code that can be run within the model optimizer. The code is automatically executed with the help of macros, and the optimization results are automatically stored and transmitted to the large language model. This solves the problem that the model optimizer cannot automatically read the output strategy of the large language model and therefore cannot automate the parameter extraction process. Furthermore, model parameter sensitivity analysis is introduced as a strategy reference for the large language model to select and optimize components. This provides navigation for the internal selection of optimization parameters in the model optimizer, transforming the "black box" problem of complex circuit topologies into a "white box" problem, reducing ineffective optimization, and greatly improving the efficiency of parameter extraction. Moreover, an adaptive optimization strategy adjustment mechanism for the large language model is created, which can deeply reflect on historically generated optimization strategies multiple times. Combining the hint rules of human optimization experience with its own deep reasoning, it avoids getting stuck in multiple iterations of optimization stagnation during parameter extraction. In addition, a design for the classification, storage, and transmission of historical optimization information is created. By storing the historical optimization strategies of successfully extracted passive devices and optimization log files, a classified knowledge base is established for the large language model to refer to and call. This helps the large language model to deeply acquire parameter extraction experience and provide optimization strategies more quickly and accurately when facing passive devices of the same or similar categories, reducing search and thinking time and maintaining the consistency and scalability of output strategies.

[0108] In this embodiment, when performing step S3, which is to perform at least one round of iteration, the following steps may also be performed:

[0109] S306. Set the first count threshold and multiple second count thresholds;

[0110] S307. When the number of iterations executed reaches the first threshold, end the execution of all iterations and output the optimization results;

[0111] S308. When the number of iterations has reached any second threshold, based on the optimization strategy information obtained from the iterations, generate prompt word modification guidance information, and modify the first prompt word according to the prompt word modification guidance information.

[0112] In step S306, the threshold values ​​K1 and K can be set for the first count. 21 K 22 K 23 There are multiple second-order thresholds, and each second-order threshold is less than the first-order threshold, for example, satisfying K. 21 <K 22 <K 23 <K1. In this embodiment, the first number threshold and each of the second number thresholds are constants.

[0113] In step S307, the first count threshold K1 is used as the upper limit of the number of rounds of the iterative process (steps S301-S305). That is, after each round of the iterative process (steps S301-S305) is completed, the number of iterations k that have been executed is counted, for example, k = k + 1. If it is determined that the number of iterations has reached the first count threshold, for example, k = K1, then the execution of all iterations ends and the optimization result is output.

[0114] By executing step S307, an upper limit on the number of iterations can be set to avoid executing too many iterations, thereby saving computational resources.

[0115] In step S308, similar to step S307, after each iteration (steps S301-S305), the number of iterations k has been counted, for example, k = k + 1, and the number of iterations k has been compared with K. 21 K 22 K 23 The relationship between the magnitudes of the second thresholds, etc. The number of iterations k already performed will successively reach K. 21 K 22 K 23 Each of the second threshold numbers satisfies k = K sequentially. 21 k = K 22 k = K 23 .

[0116] When the number of iterations k has been performed reaches any second threshold, for example, k = K 21 Then we can obtain the executed K. 21The optimization strategy information obtained in each round of iteration (which represents the specific parameters of each passive device) is averaged, for example, the average of the K rounds already executed. 21 The optimization strategy information obtained in each iteration is added together and divided by K. 21 Obtain the average value; calculate the deviation information Δ between the average value and the first prompt word set in step S2 (which represents the target parameter of each passive device), and determine the deviation information Δ based on the second threshold K already reached. 21 The modification magnitude of the first prompt word is determined in a positive correlation; that is, the larger the deviation information Δ, the larger the modification magnitude of the first prompt word. The second threshold K is then used. 21 The larger the value, the greater the scope of modification required for the primary prompt word. Based on the scope of modification, prompt word modification guidance information is generated, which suggests modifying the primary prompt word according to the required modification amount.

[0117] After generating the prompt word modification guidance information, the computer can display the prompt word modification guidance information through a screen or other means, thereby guiding the user to modify the first prompt word. When the user modifies the first prompt word, then in each subsequent iteration process, i.e., steps S301-S305, the last modified first prompt word (instead of the first prompt word set in step S2) will be used.

[0118] Regardless of whether the user actually modifies the first prompt word, as the iteration process continues, the number of iterations, k, continuously increases. Eventually, the number of iterations, k, will reach the next threshold for the second iteration, for example, k = K. 22 At this point, step S308 is triggered again, generating new prompt word modification guidance information to remind the user to modify the first prompt word.

[0119] In this embodiment, the principle of executing step S308 is as follows: Before the number of iterations k reached the upper limit of the number of rounds, i.e., the first number threshold K1, multiple second number thresholds are set. In this case, when the number of iterations k reached the second number thresholds, it indicates that a sufficient number of iterations have been executed but the optimization extraction of the passive device parameters has not yet been completed. Therefore, prompt word modification guidance information is generated to prompt the user to modify the first prompt word used for passive device parameter extraction. Specifically, the modification range corresponding to the prompt word modification guidance information is positively correlated with the deviation information, thereby guiding the user to modify the requirement represented by the first prompt word to the optimization limit determined by the executed iteration process, which is conducive to completing the parameter optimization extraction faster. The modification range corresponding to the prompt word modification guidance information is also positively correlated with the second number thresholds already reached, thereby guiding the user to modify the requirement represented by the first prompt word by a larger margin in the later iteration process, which is conducive to completing the parameter optimization extraction faster. Therefore, by executing step S308, it is beneficial to extract the parameters of the passive device that meet the user's needs and improve the success rate of passive device parameter extraction.

[0120] In this embodiment, the effect of running steps S1-S3 (including multiple rounds of iterative processes with steps S301-S305) is as follows: Figures 4-6 As shown.

[0121] Figure 4 The graph shows the characteristic response of a passive device obtained from the initial simulation in the model optimizer. The red curve represents the measured data, and the blue curve represents the simulation data from the model optimizer. The simulation frequency band for this passive device is a certain radio frequency band, and the output characteristic responses are S11, S12, S21, and S22, respectively. The error in the graph is the root mean square error. The automated parameter extraction device based on a large language model of this invention is used to... Figure 4 The passive device shown undergoes automated parameter extraction. An example of user requirements is that the root mean square error (RMSE) of all S-parameters is less than 5%, and the absolute value of the difference between the RMSEs of all S-parameters is less than 2%. After fully automated iterative optimization, the final optimized result is as follows: Figure 5 As shown, Figure 5 The measured data and model simulation data of the optimized device show a good fit. Figure 6 The historical error data of the optimization iteration process (including steps S301-S305) performed on the passive device in multiple rounds of iteration are obtained by... Figure 6 It can be seen that the passive device can meet the user's needs in just 5 automatic iterations, with a total time of less than 10 minutes.

[0122] according to Figures 4-6It is evident that the automated parameter extraction method for passive devices based on a large language model provided by this invention possesses excellent efficiency, accuracy, and practicality. This invention proposes an innovative automated parameter extraction method and apparatus for passive devices based on a large language model. The core innovation of this invention lies in achieving efficient automated interaction between the model optimizer and the large language model. This automated connection is achieved through strict constraints on the input prompts and strategy outputs of the large language model, as well as strict optimization strategy conversion and transmission, enabling complete automation of the entire parameter extraction process for passive devices. Furthermore, it integrates multiple parameter extraction strategies to provide a more efficient, accurate, and practically applicable passive device parameter extraction technology. The core technological innovations of this invention are mainly reflected in the following interrelated aspects:

[0123] The core of this invention is to create an automated module that automatically converts the optimization strategy output of a large language model into executable optimization code within the model optimizer by designing a strict optimization strategy conversion and transmission mechanism. This fully automates the parameter extraction process for passive devices. To achieve this design, the invention strictly constrains the input prompts and strategy outputs of the large language model, converting the manually executed optimization buttons within the model optimizer into automated code modules. This ensures that the optimization strategy output by the large language model not only incorporates the optimization operation characteristics within the model optimizer but also includes prompt rules based on extensive human optimization experience. Ultimately, this achieves rapid and accurate conversion and transmission of optimization strategies, completing the automated interactive connection between the model optimizer and the large language model. The advantages of this mechanism lie not only in its high degree of automation, eliminating the need for manual intervention and completely freeing up human hands, but also in its strict strategy output constraints, which significantly reduce invalid content output by the large language model, aiming to complete point-to-point tasks and greatly improving the efficiency of parameter extraction. Furthermore, this invention integrates sensitivity analysis, adaptive optimization strategy adjustment, and the design for the classification, storage, and transmission of historical optimization information. Sensitivity analysis reduces the randomness of the output strategy of the large language model, improving the efficiency of parameter selection. An adaptive optimization strategy adjustment mechanism uses optimization strategies from previous iterations as user prompts for the large language model, allowing it to deeply reflect on historically generated strategies and dynamically update and adjust them quickly by combining human optimization experience with its own deep reasoning. This avoids stagnation during parameter extraction due to multiple iterations. The design of classifying, storing, and transmitting historical optimization information stores the historical optimization strategies and optimization log files of successfully extracted passive devices, establishing a categorized knowledge base for the large language model to reference and access. This helps the large language model deeply acquire parameter extraction experience, enabling it to provide optimization strategies more quickly and accurately when facing similar or identical passive devices, reducing search and thinking time, and maintaining the consistency and scalability of the output strategy.

[0124] In summary, this invention designs an optimization strategy conversion and transmission mechanism, creates an automated device for rapid interconnection between the model optimizer and the large language model, and incorporates sensitivity analysis, adaptive optimization strategy adjustment, and historical optimization information classification, storage, and transmission design. This scheme not only fully automates the parameter extraction process for passive devices, greatly reducing manual dependence and time costs, but also effectively improves the efficiency and accuracy of parameter extraction, demonstrating versatility and scalability across various passive devices, model optimizers, and large language models.

[0125] Compared with existing technologies, the automated parameter extraction method and apparatus for passive devices based on large language models proposed in this invention have significant advantages. Traditional parameter extraction assistants based on large language models, while translating engineer-issued instructions into executable commands for optimization software, are essentially still interactive tools and cannot achieve full automation, resulting in low efficiency. Furthermore, the input instructions and output results of large language models are not strictly constrained, leading to prolonged knowledge search and deep thinking times, and a tendency to output random and redundant information. In addition, when optimization tasks continue, large language models may stagnate, meaning that multiple rounds of optimization fail to significantly reduce the error level.

[0126] This invention, by constraining the input prompts of a large language model and incorporating human experience in extracting parameters for passive components, enables the large language model to clearly understand the parameter extraction task for passive components, accurately grasp optimization requirements, and significantly reduce its reliance on human experts, thus shortening the large language model's thinking time. By constraining the optimization strategies output by the large language model, it allows the model to accurately grasp the adjustable optimization operation settings within the model optimizer, enabling more focused and targeted in-depth thinking and the generation of specific, concise optimization strategy text, reducing invalid and redundant information in the large language model's output. Furthermore, this invention creates an automated module for converting the large language model's output strategy into an executable optimization script within the model optimizer through an optimization strategy conversion and transmission mechanism, fully automating the entire parameter extraction process and eliminating significant mental and physical effort. Moreover, this invention introduces model parameter sensitivity analysis as a strategy reference for the large language model to select optimization components, providing navigation for the internal selection of optimization parameters within the model optimizer, reducing invalid optimization, and greatly improving parameter extraction efficiency. Furthermore, this invention creates an adaptive optimization strategy adjustment mechanism for large language models, which can deeply reflect on historically generated optimization strategies multiple times. Combining prompts from human optimization experience with its own deep reasoning, it avoids getting stuck in multiple iterations of optimization stagnation during parameter extraction. In addition, this invention creates a design for the categorized storage and transmission of historical optimization information, which can provide optimization strategies more quickly and accurately when faced with passive devices of the same or similar categories, reducing search and thinking time and maintaining the consistency and scalability of the output strategy.

[0127] In summary, this invention designs an optimization strategy conversion and transmission mechanism, creates an automated device for rapid interconnection between the model optimizer and the large language model, and incorporates sensitivity analysis, adaptive optimization strategy adjustment, and historical optimization information classification, storage, and transmission design. This solution not only overcomes the limitations of existing technologies, fully automating the passive device parameter extraction process and significantly reducing manual reliance and time costs, but also effectively improves the efficiency and accuracy of parameter extraction, exhibiting high versatility and scalability.

[0128] A computer program can be written to execute the automated parameter extraction method for passive devices based on a large language model as described in this embodiment. This computer program can be written into a computer device or storage medium. When the computer program is read out and run, the automated parameter extraction method for passive devices based on a large language model as described in this embodiment can be executed, thereby achieving the same technical effect as the automated parameter extraction method for passive devices based on a large language model as described in this embodiment.

[0129] It should be noted that, unless otherwise specified, when a feature is referred to as "fixed" or "connected" to another feature, it can be directly fixed or connected to the other feature, or indirectly fixed or connected to the other feature. Furthermore, the descriptions of "upper," "lower," "left," and "right" used in this disclosure are only relative to the relative positional relationships of the components of this disclosure in the accompanying drawings. The singular forms "a," "an," and "the" used in this disclosure are also intended to include the plural forms, unless the context clearly indicates otherwise. Moreover, unless otherwise defined, all technical and scientific terms used in this embodiment have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this embodiment specification is only for describing specific embodiments and is not intended to limit the embodiments of the invention. The term "and / or" as used in this embodiment includes any combination of one or more of the associated listed items.

[0130] It should be understood that although the terms first, second, third, etc., may be used to describe various elements in this disclosure, these elements should not be limited to these terms. These terms are only used to distinguish elements of the same type from each other. For example, a first element may also be referred to as a second element without departing from the scope of this disclosure, and similarly, a second element may also be referred to as a first element. The use of any and all instances or exemplary language (“e.g.,” “such as,” etc.) provided in this embodiment is intended only to better illustrate embodiments of the invention and, unless otherwise required, does not impose a limitation on the scope of embodiments of the invention.

[0131] It should be recognized that embodiments of the present invention can be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium. The method can be implemented using standard programming techniques—including a non-transitory computer-readable storage medium configured with a computer program, wherein such a storage medium causes the computer to operate in a specific and predefined manner—according to the methods and drawings described in the specific embodiments. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, if desired, the program can be implemented in assembly or machine language. In any case, the language can be a compiled or interpreted language. Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit (ASIC).

[0132] Furthermore, the procedures described in this embodiment can be performed in any suitable order, unless otherwise indicated by this embodiment or otherwise obviously contradictory to the context. The procedures (or variations and / or combinations thereof) described in this embodiment can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. A computer program includes multiple instructions executable by one or more processors.

[0133] Furthermore, the method can be implemented in any suitable type of computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices, etc. Aspects of embodiments of the invention can be implemented as machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it is readable by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein. Furthermore, the machine-readable code, or portions thereof, can be transmitted via wired or wireless networks. The invention of this embodiment includes these and other different types of non-transitory computer-readable storage media when such media comprises instructions or programs that implement the steps above in conjunction with a microprocessor or other data processor. Embodiments of the invention also include the computer itself when programmed according to the methods and techniques of embodiments of the invention.

[0134] A computer program can be applied to input data to perform the functions of this embodiment, thereby transforming the input data to generate output data stored in non-volatile memory. The output information can also be applied to one or more output devices, such as a display. In a preferred embodiment of the invention, the transformed data represents physical and tangible objects, including a specific visual depiction of physical and tangible objects generated on the display.

[0135] The above are merely preferred embodiments of the present invention. The embodiments of the present invention are not limited to the above-described implementations. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the embodiments of the present invention, as long as they achieve the same technical effects, should be included within the scope of protection of the embodiments of the present invention. Within the scope of protection of the embodiments of the present invention, the technical solutions and / or implementation methods can have various modifications and variations.

Claims

1. A method for automated parameter extraction of passive devices based on a large language model, characterized in that, The automated parameter extraction method for passive devices based on a large language model includes: Obtain the initial file; the initial file includes netlist data and measured data for passive devices; Obtain the first prompt word; the first prompt word represents the target parameters that the user requires the passive device to achieve. Perform at least one round of iterative process; each round of iterative process includes the following steps: The data to be simulated is input into the model optimizer for simulation processing, and the characteristic response image and model parameter information output by the model optimizer are obtained; wherein, when the current iteration process is the first iteration process, the data to be simulated is the initial file, otherwise, the data to be simulated is the data obtained by optimizing the initial file according to the executable optimization code obtained in the previous iteration process; Based on the characteristic response image and the model parameter information, obtain the second prompt word; The first prompt word and the second prompt word are input into the large language model for processing; the first judgment value and optimization strategy information output by the large language model, or the second judgment value output by the large language model, are obtained. Once the first judgment value is obtained, the next round of iteration is triggered based on the optimization strategy information; Once the second judgment value is obtained, the execution of all the above iteration processes ends, and the optimization result is output.

2. The method for automated parameter extraction of passive devices based on a large language model according to claim 1, characterized in that, The step of obtaining the second prompt word based on the characteristic response image and the model parameter information includes: Obtain historical optimization records for passive devices; the historical optimization records include optimization results obtained from iterative processes that have been executed and model parameter information of passive devices; Error data of the response image based on the aforementioned characteristics; Sensitivity analysis was performed on the model parameter information; The second prompt word is determined based on the model parameter information, the error data, and the sensitivity analysis results.

3. The method for automated parameter extraction of passive devices based on a large language model according to claim 1, characterized in that, The step of obtaining the second prompt word based on the characteristic response image and the model parameter information includes: Obtain historical optimization records for passive devices; the historical optimization records include optimization results obtained from iterative processes that have been executed and model parameter information of passive devices; Error data of the response image based on the aforementioned characteristics; Obtain historical optimization strategies for passive devices; the historical optimization strategies include optimization strategy information obtained from iterative processes that have already been executed; Sensitivity analysis was performed on the model parameter information; The second prompt word is determined based on the model parameter information, the error data, the optimization strategy information, and the sensitivity analysis results.

4. The method for automated parameter extraction of passive devices based on a large language model according to claim 3, characterized in that, The step of triggering the next iteration process based on the optimization strategy information includes: Update the historical optimization strategy according to the optimization strategy information; The executable optimized code is obtained by converting the optimization strategy information. Execute the next iteration.

5. The method for automated parameter extraction of passive devices based on a large language model according to claim 1, characterized in that, The step of inputting the first prompt word and the second prompt word into a large language model for processing includes: The first prompt word and the second prompt word are constrained to obtain constrained prompt words; The output rules of the optimization strategy information of the large language model are constrained; The constraint prompts are input into the large language model.

6. The method for automated parameter extraction of passive devices based on a large language model according to claim 5, characterized in that, The step of constraining the first prompt word and the second prompt word to obtain constrained prompt words includes: Set constraints for user requirement analysis, manual optimization experience tips, and optimization strategy output; Based on the user demand parsing constraint rules, the manual optimization experience suggestion rules, and the optimization strategy output constraint rules, the first suggestion word and the second suggestion word are filtered to obtain the constraint suggestion word.

7. The method for automated parameter extraction of passive devices based on a large language model according to any one of claims 1-6, characterized in that, The automated parameter extraction method for passive devices based on a large language model also includes: Set a first count threshold and multiple second count thresholds; each second count threshold is less than the first count threshold. When the number of iterations executed reaches the first threshold, the execution of all iterations ends and the optimization result is output. When the number of iterations has reached any of the second thresholds, based on the optimization strategy information obtained from the iterations, prompt word modification guidance information is generated, and the first prompt word is modified according to the prompt word modification guidance information.

8. The method for automated parameter extraction of passive devices based on a large language model according to claim 7, characterized in that, The step of generating prompt word modification guidance information based on the optimization strategy information obtained from the executed iteration process includes: Obtain the average value of the optimization strategy information obtained from the executed iteration process; Obtain the deviation information of the average value relative to the first prompt word; Based on the deviation information and the second threshold number already reached, the modification range for the first prompt word is determined; wherein the modification range is positively correlated with both the deviation information and the second threshold number already reached. Based on the stated modification range, the prompt word modification guidance information is generated.

9. A computer device, characterized in that, It includes a memory and a processor, wherein the memory is used to store at least one program, and the processor is used to load at least one program to execute the method for automatic parameter extraction of passive devices based on a large language model as described in any one of claims 1-8.

10. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the passive device automated parameter extraction method based on a large language model as described in any one of claims 1-8.

Images