Circuit device model parameter extraction and optimization method based on artificial intelligence

Abstract

The invention relates to the field of integrated circuit layout design, provides a circuit device model parameter extraction and optimization method based on artificial intelligence, and aims at solving the technical problems that a traditional EDA tool is high in artificial dependence, low in optimization efficiency, poor in algorithm adaptability and split in tool. Automatic data interaction among the modules is realized through a standardized API; preprocessing and intention recognition are carried out on the multi-modal input, and user requirements are converted into formatted data; searching similar circuit historical parameters based on the knowledge base and generating a recommendation list; extracting a target curve through an image recognition model and carrying out standardization processing; dynamically matching an optimization algorithm according to circuit characteristics and adjusting hyper-parameters in real time; multi-tool collaborative optimization control is realized, and the parameter search progress is dynamically updated; and outputting the optimal parameter combination and carrying out verification iteration. The method is used for high-precision automatic optimization and full-process closed-loop verification of circuit parameters.

Description

Technical Field

[0001] This invention relates to the field of integrated circuit layout and provides a method for extracting and optimizing circuit device model parameters based on artificial intelligence. Background Technology

[0002] Optimization analysis has long been a core technology in electronic design automation (EDA). It aims to automatically adjust circuit parameters or design variables through algorithms to meet constraints such as performance, power consumption, and area (PPA). It plays an irreplaceable role compared to other types of analysis in addressing the explosive growth of design complexity, resolving conflicts and trade-offs among multiple objectives, and improving design reliability and robustness.

[0003] However, current circuit simulation optimization analysis requires manual parameter configuration, algorithm selection, and target curve generation under objective constraints. Specific problems include:

[0004] High reliance on manual intervention: Engineers need to manually set the parameters to be optimized (such as transistor size and model parameters) and their ranges in EDA tools based on experience, which is easily affected by subjective factors and the parameter selection may not be comprehensive.

[0005] Low target generation efficiency: If optimizing the target involves curve fitting (such as Ic-Vbe characteristic curve), it is necessary to manually plot points to extract target data, which is time-consuming and prone to introducing errors. In particular, it has insufficient ability to process complex curves (such as logarithmic coordinates and noisy curves).

[0006] Poor algorithm adaptability: Optimization algorithms (such as genetic algorithms and Bayesian optimization) require manual selection and lack automatic adaptation to circuit characteristics (linear / nonlinear, multi-objective / single-objective), resulting in slow convergence speed or getting stuck in local optima.

[0007] Tool fragmentation: Tools such as schematic design, simulation, and waveform analysis run independently, and data transfer relies on manual operation, making it difficult to automate the entire optimization process.

[0008] While existing technologies (such as traditional EDA tools) offer optimization modules, their level of intelligence is low, failing to meet the demands for efficient and accurate automation, especially in highly complex circuit designs. Summary of the Invention

[0009] The purpose of this invention is to provide an automated extraction and optimization method for circuit device model parameters based on artificial intelligence. By using multimodal intent recognition, knowledge base parameter recommendation, intelligent image extraction of target curves, dynamic algorithm matching, and multi-tool collaborative control, this method solves the technical pain points of traditional EDA tools, such as strong reliance on manual intervention, low optimization efficiency, poor algorithm adaptability, and fragmented tools, thereby achieving high-precision automatic optimization of circuit parameters and closed-loop verification throughout the entire process.

[0010] To achieve the above objectives, the present invention employs the following technical means:

[0011] This invention provides a method for extracting and optimizing circuit device model parameters based on artificial intelligence, comprising the following steps:

[0012] Step (1): System communication mechanism construction: Taking AI agent as the core module, establish data communication links between front-end user interface module, circuit schematic design module, large language model (LLM) module and circuit simulator module through standardized API to realize automated data interaction between modules;

[0013] Step 1.1: Abstract and standardize the interfaces of the front-end user interface module, circuit schematic design module, large language model module, and circuit simulator module to form a unified API calling specification and realize the preliminary definition of the communication protocol between modules;

[0014] Step 1.2: For the communication protocol defined in Step 1.1, the AI ​​agent acts as the central scheduler, and the asynchronous message queue is used to realize the automatic routing and load balancing of data requests and responses between modules, ensuring communication efficiency under high concurrency.

[0015] Step 1.3: For the communication link established in Step 1.2, introduce data encryption and authentication mechanisms to ensure the security of sensitive circuit design data during transmission.

[0016] This step, through standardized APIs and AI agent scheduling, solves the problems of module fragmentation and data transmission dependence in traditional EDA tools, realizes full-process automated collaboration, and improves system integration and reliability.

[0017] Step (2): Multimodal input preprocessing and intent recognition: Receive text, voice or image data input by the user, convert it into structured text through OCR or speech recognition technology, use a pre-trained language model to parse the design constraints and objectives in the natural language, and generate formatted requirement data that can be recognized within the system.

[0018] Step 2.1: Use OCR or speech recognition engine to perform preliminary conversion on the text, voice or image data input by the user to generate an intermediate text representation;

[0019] Step 2.2: Use a pre-trained language model to perform semantic parsing and entity recognition on the intermediate text obtained in Step 2.1, and extract key design constraints and performance indicators;

[0020] Step 2.3: Perform logical formatting on the semantic structure extracted in Step 2.2 to generate a JSON format instruction set that the system can execute.

[0021] This step, through multimodal recognition and semantic parsing, solves the problems of traditional input methods relying on manual interpretation and being prone to errors, and achieves automated and accurate understanding of design intent.

[0022] Step (3): Parameter recommendation and range setting: Based on the internal knowledge base, retrieve the historical design parameters of similar circuits to generate an initial recommendation list, and combine the physical rules of the process design kit (PDK) to verify and correct the legality of the parameter range;

[0023] Step 3.1: For historical circuit design cases stored in the internal knowledge base, use a similarity retrieval method based on graph neural networks to find the case set that best matches the current design requirements. Step 3.2: For the case parameters retrieved in Step 3.1, combine them with the current process design kit rules to perform parameter validity verification and range correction.

[0024] Step 3.3: Generate an initial recommendation list and sort it for use in subsequent optimization processes based on the parameter set corrected in Step 3.2.

[0025] This step, through knowledge base retrieval and PDK rule verification, solves the problem of parameter settings relying on experience and easily exceeding process limitations, thus improving the rationality and reliability of parameter initialization.

[0026] Step (4): Automated extraction and standardization of target curve: The coordinate axis is located and data points are extracted from the measurement curve image uploaded by the user through the image recognition model, and a digital target curve is generated by fitting, and noise filtering and coordinate normalization are performed.

[0027] Step 4.1: For the curve image uploaded by the user, use a convolutional neural network to perform coordinate axis detection and curve region segmentation, and extract the original pixel set;

[0028] Step 4.2: Perform coordinate inverse calculation and data point mapping on the pixel set extracted in Step 4.1 to generate a preliminary digital curve;

[0029] Step 4.3: Perform noise filtering and coordinate normalization on the curve data generated in Step 4.2, and output the target curve data in standard format.

[0030] This step, through image recognition and data post-processing, solves the problems of low efficiency and large errors in traditional manual point plotting, and achieves rapid and accurate digitization of the target curve.

[0031] Step (5): Dynamic matching of optimization algorithm: Based on the dimension of circuit characteristic parameters, nonlinearity and number of multiple targets, the appropriate optimization algorithm is automatically selected from the preset algorithm library, and the residual is monitored in real time to dynamically adjust hyperparameters or switch algorithms.

[0032] Step 5.1: For the current circuit design problem, extract its feature vector, including parameter dimension, nonlinearity, and number of multi-objectives;

[0033] Step 5.2: For the feature vectors extracted in Step 5.1, use a reinforcement learning-based algorithm selector to match the optimal optimization algorithm from a pre-set algorithm library;

[0034] Step 5.3: For the algorithm selected in Step 5.2, monitor the convergence status in real time during the operation, and dynamically adjust the hyperparameters or trigger algorithm switching.

[0035] This step, through feature-driven and dynamic adjustment, solves the problems of blind selection and slow convergence in traditional optimization algorithms, thereby improving optimization efficiency and adaptability.

[0036] Step (6): Multi-tool collaborative optimization control: The AI ​​agent calls the circuit simulator through standardized API to generate a netlist and submit the simulation, analyzes the simulation results to extract key indicators, and dynamically updates the parameter search progress and residual curve;

[0037] Step 6.1: For the current parameter combination, the AI ​​agent calls the schematic editor to generate a standard netlist file and submits it to the simulator via API;

[0038] Step 6.2: Monitor the simulation status in real time for the simulation task submitted in Step 6.1, and parse the output file to extract key performance indicators upon completion;

[0039] Step 6.3: Calculate the residuals between the indicators extracted in Step 6.2 and the target curve, and dynamically update the optimization progress and convergence curve.

[0040] This step, through automated scheduling and result analysis using multiple tools, solves the problems of frequent manual operations and data transmission breaks in traditional processes, and achieves closed-loop optimization control.

[0041] Step (7): Optimization result verification and iteration: Output the optimal parameter combination and simulation verification report. If the residual does not reach the preset threshold, automatically trigger the parameter search range expansion or multi-fidelity optimization iteration, and generate a design conflict diagnosis report.

[0042] Step 7.1: Perform simulation verification on the parameter combination output by the optimization process and generate a verification report, including performance comparison and residual analysis;

[0043] Step 7.2: For the residual results in Step 7.1, determine whether they meet the preset threshold. If they do not meet the threshold, trigger the multi-strategy iteration mechanism.

[0044] Step 7.3: For the iterative process triggered in step 7.2, support parameter range expansion, multi-fidelity model switching, and generate conflict diagnosis reports.

[0045] Creative Explanation: This step solves the problems of traditional optimization results relying on manual judgment and low iteration efficiency through automated verification and intelligent iteration, and achieves a high-precision, adaptive optimization closed loop.

[0046] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0047] 1. By using multimodal intent recognition and parameter recommendation, the technical problems of traditional EDA tools, such as strong reliance on manual intervention and subjective error in parameter setting, are solved. This achieves fully automated configuration, reduces manual intervention, avoids reliance on experience, and improves the efficiency of target curve generation by more than 80%.

[0048] 2. By using dynamic algorithm matching and hyperparameter adjustment, the technical problems of blind selection and slow convergence in traditional optimization algorithms are solved, thus improving optimization efficiency and shortening the optimization cycle by 50% compared with traditional methods.

[0049] 3. By using AI agent collaborative control and standardized API integration, the technical problems of fragmentation and reliance on manual data transmission in traditional EDA tools have been solved, achieving seamless integration of the entire process and supporting one-stop optimization from requirement input to result verification.

[0050] 4. By employing multi-objective optimization and nonlinear curve recognition techniques, the technical problem of insufficient support for complex scenarios by traditional methods has been solved, achieving effective support for complex scenarios such as high-frequency circuits and radio frequency devices.

[0051] 5. This invention organically combines six core technologies: multimodal intent recognition (step 2), knowledge base / PDK joint parameter initialization (step 3), image-driven target curve digitization (step 4), feature-aware dynamic algorithm matching (step 5), AI agent-led multi-tool collaborative control (steps 1 and 6), and automatic verification and intelligent iteration (step 7). This results in a system-level synergistic effect that transcends the superposition of single modules: the entire optimization process can complete the end-to-end closed loop from fuzzy requirement input to precise parameter output without relying on any manual nodes. Moreover, a positive reinforcement loop of "perception-decision-execution-feedback" is formed between each subsystem. For example, the curve shape obtained from image recognition directly affects the algorithm selection strategy, while the simulation residual drives the redefinition of parameter range and knowledge base update. At the same time, the design constraints resolved by LLM directly guide the direction of GNN parameter recommendation, and the PDK rule verification results are also used to build a more reliable initial population. This multi-level, multi-dimensional information linkage and dynamic coupling enables the system to have self-evolution capabilities and scenario generalization capabilities, fundamentally breaking through the limitations of traditional EDA tools such as "static configuration, single-point optimization, and manual connection," and realizing truly intelligent and adaptive extraction and optimization of circuit device model parameters. Attached Figure Description

[0052] Figure 1 This is a simplified flowchart of the present invention;

[0053] Figure 2 This is an example of a failed image curve display. The black curve is the original curve, and the green hash points are data points automatically extracted through step 4.1 below. The left graph's X-axis is a linear coordinate, and the right graph's X-axis is a logarithmic coordinate.

[0054] Figure 3 This is used to extract the model parameters of transistor Q1 in a circuit designed in a schematic tool.

[0055] Figure 4 Users can communicate with AI via natural language to propose optimization needs for the current model. The AI ​​then provides suggestions on the parameters to be optimized and their ranges in a table format. Users can make adjustments within the table.

[0056] Figure 5 Set the optimization target for the AI ​​prompt. For example, if you need to fit the IV curve of the device model, please upload an image and the system will automatically recognize the curve example in the image.

[0057] Figure 6 To facilitate optimization analysis, the error trends of the current simulation IV curve and the user-uploaded IV curve are dynamically displayed. Under normal circumstances, the error curve will gradually approach 0 as the simulation progresses;

[0058] Figure 7 After optimization, the AI ​​extracts the simulation results and displays the final values ​​of the model optimization parameters in a table.

[0059] Figure 8 To verify the results, the IV curve fitting effect is shown, with red representing the target and blue representing the result. The two curves basically overlap, indicating... Figure 7 The optimized parameters of the model shown match the actual parameters of the model. In other words, this set of parameter values ​​can be used to accurately simulate the device model. Detailed Implementation

[0060] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0061] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0062] This invention provides a method for extracting and optimizing circuit device model parameters based on artificial intelligence, comprising the following steps:

[0063] Step (1): System communication mechanism construction: Taking AI agent as the core module, establish data communication links between front-end user interface module, circuit schematic design module, large language model (LLM) module and circuit simulator module through standardized API to realize automated data interaction between modules;

[0064] Step 1.1: Abstract and standardize the interfaces of the front-end user interface module, circuit schematic design module, large language model module, and circuit simulator module to form a unified API calling specification and achieve a preliminary definition of the communication protocol between modules. The specific logic is as follows: encapsulate API interfaces for human-computer interaction for the user interface; encapsulate API interfaces for schematic creation, modification, and netlist generation for the circuit schematic; encapsulate API interfaces for circuit input and simulation result output for the circuit simulation module; and then the AI ​​agent uses the reasoning capabilities of the large language model to call the corresponding API interfaces and automatically complete the relevant functions.

[0065] Step 1.2: For the communication protocol defined in Step 1.1, the AI ​​agent acts as the central scheduler, and the asynchronous message queue is used to realize the automatic routing and load balancing of data requests and responses between modules, ensuring communication efficiency under high concurrency.

[0066] Step 1.3: For the communication link established in Step 1.2, introduce data encryption and authentication mechanisms to ensure the security of sensitive circuit design data during transmission.

[0067] This step, through standardized APIs and AI agent scheduling, solves the problems of module fragmentation and data transmission dependence in traditional EDA tools, realizes full-process automated collaboration, and improves system integration and reliability.

[0068] Step (2): Multimodal input preprocessing and intent recognition: Receive text, voice or image data input by the user, convert it into structured text through OCR or speech recognition technology, use a pre-trained language model to parse the design constraints and objectives in the natural language, and generate formatted requirement data that can be recognized within the system.

[0069] Step 2.1: Use OCR or speech recognition engine to perform preliminary conversion on the text, voice or image data input by the user to generate an intermediate text representation;

[0070] Step 2.2: Use a pre-trained language model to perform semantic parsing and entity recognition on the intermediate text obtained in Step 2.1, and extract key design constraints and performance indicators;

[0071] Step 2.3: Perform logical formatting on the semantic structure extracted in Step 2.2 to generate a JSON format instruction set that the system can execute.

[0072] This step, through multimodal recognition and semantic parsing, solves the problems of traditional input methods relying on manual interpretation and being prone to errors, and achieves automated and accurate understanding of design intent.

[0073] Step (3): Parameter recommendation and range setting: Based on the internal knowledge base, retrieve the historical design parameters of similar circuits to generate an initial recommendation list, and combine the physical rules of the process design kit (PDK) to verify and correct the legality of the parameter range;

[0074] Step 3.1: For historical circuit design cases stored in the internal knowledge base, a similarity retrieval method based on graph neural networks is used to find the case set that best matches the current design requirements. Compared with traditional methods such as manual search and keyword search, this method greatly improves efficiency and accuracy.

[0075] Step 3.2: For the case parameters retrieved in Step 3.1, perform parameter validity verification and range correction in conjunction with the current process design kit rules; specifically, deeply integrate and verify the historical case parameters with the current process design kit (PDK) rules, and on this basis, realize intelligent correction of parameter range to ensure that the parameters not only meet process constraints, but also have the potential for circuit performance optimization.

[0076] This step employs the following innovative technology combination:

[0077] 1. PDK rule parser + parameter constraint graph modeling: Develop a PDK rule parsing engine to automatically parse the rule descriptions in the PDK file, convert them into machine-executable constraint functions, construct a parameter constraint graph, where nodes are parameters and edges represent constraint relationships, and use graph traversal algorithms to perform multi-parameter coupling verification;

[0078] 2. Parameter Coupling Modeling and Correction Based on GNN: Graph Neural Network (GNN) is used to model parameter dependencies. A parameter correction network is trained with parameter vectors and PDK rules as inputs and the corrected parameter vectors as outputs.

[0079] 3. Historical case-driven safety margin learning: The system learns from the knowledge base cases where "boundary parameters lead to performance degradation" and automatically introduces dynamic safety margins for certain parameters to avoid the design being on the edge of a performance cliff.

[0080] 4. Explainable Correction Report Generation: Combining Natural Language Generation (NLG) technology, the correction process is transformed into a structured report. This includes the triggered PDK rule clauses; parameter dependency analysis; references to historical similar cases; and suggested further simulation verification items.

[0081] Step 3.3: Generate an initial recommendation list and sort it for use in subsequent optimization processes based on the parameter set corrected in Step 3.2.

[0082] This step, through knowledge base retrieval and PDK rule verification, solves the problem of parameter settings relying on experience and easily exceeding process limitations, thus improving the rationality and reliability of parameter initialization.

[0083] Step (4): Automated extraction and standardization of target curve: The coordinate axis is located and data points are extracted from the measurement curve image uploaded by the user through the image recognition model, and a digital target curve is generated by fitting, and noise filtering and coordinate normalization are performed.

[0084] Step 4.1: For the curve images uploaded by users, a convolutional neural network is used for coordinate axis detection and curve region segmentation to extract the original pixel set. Considering the diversity of user-uploaded images, which may come from datasheet scans, experimental measurement screenshots, hand-drawn curves, etc., the quality varies. The system needs to be compatible with grayscale images, color images, and low-resolution images. In addition, for high-frequency circuits, radio frequency devices, etc., the curves may have nonlinear, multi-peak, or steep change characteristics. During the mapping process, the original curve shape is preserved to avoid over-smoothing, and a piecewise interpolation method is used to handle abrupt change regions. Coordinate inversion: The OpenCV library is used for image geometric correction and precise coordinate axis positioning; the Tesseract OCR engine is integrated to read the scale labels and automatically switch between linear or logarithmic mapping according to the scale type, supporting mixed coordinates (such as linear x-axis and logarithmic y-axis). Data point mapping and fitting: Bilinear interpolation is used to map pixels to physical values ​​to reduce quantization errors; for curves with a lot of noise, a Savitzky-Golay filter is first used for smoothing before mapping.

[0085] Step 4.2: Perform coordinate inverse calculation and data point mapping on the pixel set extracted in Step 4.1 to generate a preliminary digital curve;

[0086] Step 4.3: Perform noise filtering and coordinate normalization on the curve data generated in Step 4.2, and output the target curve data in standard format.

[0087] This step, through image recognition and data post-processing, solves the problems of low efficiency and large errors in traditional manual point plotting, and achieves rapid and accurate digitization of the target curve.

[0088] Step (5): Dynamic matching of optimization algorithm: Based on the dimension of circuit characteristic parameters, nonlinearity and number of multiple targets, the appropriate optimization algorithm is automatically selected from the preset algorithm library, and the residual is monitored in real time to dynamically adjust hyperparameters or switch algorithms.

[0089] Step 5.1: For the current circuit design problem, extract its feature vector, including parameter dimension, nonlinearity, and number of objectives; for low-dimensional (≤5), low-nonlinear, single-objective circuits, preset the simplex method optimization algorithm; for high-dimensional (>10), high-nonlinear, single-objective circuits, preset the genetic optimization algorithm; for variable-dimensional, high-nonlinear, multi-objective circuits, preset the differential evolution algorithm; for nonlinear, single-objective circuits with provided initial values, preset the Levenberg-Marquardt algorithm.

[0090] Step 5.2: For the feature vectors extracted in Step 5.1, use a reinforcement learning-based algorithm selector to match the optimal optimization algorithm from a pre-set algorithm library;

[0091] Step 5.3: For the algorithm selected in Step 5.2, monitor the convergence status in real time during the operation, and dynamically adjust the hyperparameters or trigger algorithm switching.

[0092] This step, through feature-driven and dynamic adjustment, solves the problems of blind selection and slow convergence in traditional optimization algorithms, thereby improving optimization efficiency and adaptability.

[0093] Step (6): Multi-tool collaborative optimization control: The AI ​​agent calls the circuit simulator through standardized API to generate a netlist and submit the simulation, analyzes the simulation results to extract key indicators, and dynamically updates the parameter search progress and residual curve;

[0094] Step 6.1: For the current parameter combination, the AI ​​agent calls the schematic editor to generate a standard netlist file and submits it to the simulator via API;

[0095] Step 6.2: Monitor the simulation status in real time for the simulation task submitted in Step 6.1, and parse the output file to extract key performance indicators upon completion;

[0096] Step 6.3: Calculate the residuals between the indicators extracted in Step 6.2 and the target curve, and dynamically update the optimization progress and convergence curve.

[0097] This step, through automated scheduling and result analysis using multiple tools, solves the problems of frequent manual operations and data transmission breaks in traditional processes, and achieves closed-loop optimization control.

[0098] Step (7): Optimization result verification and iteration: Output the optimal parameter combination and simulation verification report. If the residual does not reach the preset threshold, automatically trigger the parameter search range expansion or multi-fidelity optimization iteration, and generate a design conflict diagnosis report.

[0099] Step 7.1: Perform simulation verification on the parameter combination output by the optimization process and generate a verification report, including performance comparison and residual analysis;

[0100] Step 7.2: For the residual results in Step 7.1, determine whether they meet the preset threshold. If they do not meet the threshold, trigger the multi-strategy iteration mechanism.

[0101] Step 7.3: For the iterative process triggered in step 7.2, support parameter range expansion, multi-fidelity model switching, and generate conflict diagnosis reports.

[0102] Creative Explanation: This step solves the problems of traditional optimization results relying on manual judgment and low iteration efficiency through automated verification and intelligent iteration, and achieves a high-precision, adaptive optimization closed loop.

[0103] Example 1

[0104] This invention provides a method for extracting and optimizing circuit device model parameters based on artificial intelligence, comprising the following steps:

[0105] (1) Establish a communication mechanism between various modules of the system. The Agent serves as the core, connecting the front-end UI module, circuit schematic design module, Large Model Language (LLM), and circuit simulator module. Modules exchange data efficiently and securely through corresponding APIs. The UI module interacts with the user via Copilot dialogue.

[0106] (2) Input preprocessing and intent recognition.

[0107] 2.1 Multimodal input support: Accepts text, voice, or image input (such as scanned copies of user-written requirement documents) and converts them into structured text via OCR or speech recognition.

[0108] 2.2 Natural Language Semantic Parsing: A pre-trained language model is used to extract key constraints (such as "gain > 60dB", "power consumption < 10mW") and design goals (such as "low-noise amplifier"), supporting both Chinese and English parsing. Finally, formatted data is generated internally by the system.

[0109] (3) Parameter Recommendation and Range Setting

[0110] 3.1 Knowledge-based reasoning: Retrieve similar circuits (such as "two-stage op amp") from the internal knowledge base, extract historical design parameters (such as input pair transistor dimensions and compensation capacitor values), and generate an initial recommendation list.

[0111] 3.2 Physical rule verification: Combine the physical rules of the process design kit (PDK) (such as minimum channel length and current density limits) to filter illegal parameter combinations and limit the parameter range.

[0112] (4) Automated extraction and standardization of target curves

[0113] 4.1 Users upload images of measured curves (such as grayscale images or logarithmic coordinate graphs in a datasheet). The system uses an image recognition model (such as CNN) to locate the coordinate axes, extract data points, and fit them into a digital curve.

[0114] 4.2 Supports post-processing such as noise filtering and coordinate normalization (e.g., logarithmic to linear conversion) to ensure compatibility between the target curve and the simulation environment.

[0115] (5) Optimize the algorithm for dynamic matching

[0116] 5.1 Based on circuit characteristics (such as nonlinearity, parameter dimension, number of multiple objectives), the system automatically selects an appropriate algorithm from the algorithm library (including Bayesian optimization, genetic algorithm, particle swarm optimization, differential evolution, etc.).

[0117] The system dynamically adjusts hyperparameters (such as population size and learning rate) and monitors residuals in real time. If the convergence speed is too slow, the algorithm is switched. The system has multiple built-in residual calculation methods.

[0118] (6) Multi-tool collaboration and AI agent control

[0119] AI agents automatically complete the following processes by integrating schematic editors (such as BTD-ABS, BTD-BCS, etc.), simulators (BTD-Sim), and waveform analysis tools (BTD-Wave, etc.) through standardized APIs:

[0120] a. Generate a netlist based on the optimized parameters and submit it for simulation.

[0121] b. Analyze the simulation results and extract key indicators (such as gain and phase margin).

[0122] c. Dynamically display the residual curve in the interactive interface and update the parameter search progress.

[0123] (7) Optimization result verification and iteration

[0124] 7.1 Outputs the optimal parameter combination and simulation verification report, supporting export to CSV / JSON format.

[0125] 7.2 If the residual does not reach the threshold (e.g., residual > 10⁻³), the system automatically triggers the following iterations:

[0126] a. Expanding the parameter search range

[0127] b. Introduce multi-fidelity optimization (e.g., combining fast approximation models with high-precision simulations).

[0128] c. Generate diagnostic reports that highlight potential design conflicts (such as power consumption versus bandwidth discrepancies).

Claims

1. A method for extracting and optimizing circuit device model parameters based on artificial intelligence, characterized in that, Includes the following steps: Step (1): System communication mechanism construction: Taking AI agent as the core module, establish data communication links between front-end user interface module, circuit schematic design module, large language model module and circuit simulator module through standardized API to realize automated data interaction between modules; Step (2): Multimodal input preprocessing and intent recognition: Receive text, voice or image data input by the user, convert it into structured text through OCR or speech recognition technology, use a pre-trained language model to parse the design constraints and objectives in the natural language, and generate formatted requirement data that can be recognized within the system. Step (3): Parameter recommendation and range setting: Based on the internal knowledge base, retrieve the historical design parameters of similar circuits, generate an initial recommendation list, and combine the physical rules of the process design kit to verify and correct the legality of the parameter range; Step (4): Automated extraction and standardization of target curve: The coordinate axis is located and data points are extracted from the measurement curve image uploaded by the user through the image recognition model, and a digital target curve is generated by fitting, and noise filtering and coordinate normalization are performed. Step (5): Dynamic matching of optimization algorithm: Based on the dimension of circuit characteristic parameters, nonlinearity and number of multiple targets, the appropriate optimization algorithm is automatically selected from the preset algorithm library, and the residual is monitored in real time to dynamically adjust hyperparameters or switch algorithms. Step (6): Multi-tool collaborative optimization control: The AI ​​agent calls the circuit simulator through standardized API to generate a netlist and submit the simulation, analyzes the simulation results to extract key indicators, and dynamically updates the parameter search progress and residual curve; Step (7): Optimization result verification and iteration: Output the optimal parameter combination and simulation verification report. If the residual does not reach the preset threshold, automatically trigger the parameter search range expansion or multi-fidelity optimization iteration, and generate a design conflict diagnosis report.

2. The method according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1: Abstract and standardize the interfaces of the front-end user interface module, circuit schematic design module, large language model module, and circuit simulator module to form a unified API calling specification and realize the preliminary definition of the communication protocol between modules; Step 1.2: For the communication protocol defined in Step 1.1, the AI ​​agent acts as the central scheduler, and the asynchronous message queue is used to realize the automatic routing and load balancing of data requests and responses between modules, ensuring communication efficiency under high concurrency. Step 1.3: For the communication link established in Step 1.2, introduce data encryption and authentication mechanisms to ensure the security of sensitive circuit design data during transmission.

3. The method according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Use OCR or speech recognition engine to perform preliminary conversion on the text, voice or image data input by the user to generate an intermediate text representation; Step 2.2: Use a pre-trained language model to perform semantic parsing and entity recognition on the intermediate text obtained in Step 2.1, and extract key design constraints and performance indicators; Step 2.3: Perform logical formatting on the results extracted in Step 2.2 to generate a JSON format instruction set that the system can execute.

4. The method according to claim 1, characterized in that, Step 3 includes the following steps: Step 3.1: For historical circuit design cases stored in the internal knowledge base, use a similarity retrieval method based on graph neural networks to find the set of cases that best match the current design requirements; Step 3.2: For the case parameters retrieved in Step 3.1, perform parameter validity verification and range correction in accordance with the current process design suite rules. This includes the following steps: PDK rule parser and parameter constraint graph modeling: The PDK rule parsing engine automatically parses the rule descriptions in the PDK file, converts them into machine-executable constraint functions, constructs a parameter constraint graph, where nodes are parameters and edges represent constraint relationships, and uses graph traversal algorithms to perform multi-parameter coupling verification. Parameter coupling modeling and correction based on GNN: Graph Neural Network (GNN) is used to model parameter dependencies. A parameter correction network is trained with parameter vectors and PDK rules as inputs and the corrected parameter vectors as outputs. Historical case-driven safety margin learning: The system learns cases from the knowledge base where "boundary parameters lead to performance degradation" and automatically introduces dynamic safety margins for certain parameters; Interpretable Correction Report Generation: Combining Natural Language Generation (NLG) technology, the correction process is transformed into a structured report, including: triggered PDK rule clauses; parameter dependency analysis; references to historical similar cases; and suggested further simulation verification items. Step 3.3: Generate an initial recommendation list and sort it for use in subsequent optimization processes based on the parameter set corrected in Step 3.

2.

5. The method according to claim 1, characterized in that, Step 4 includes the following steps: Step 4.1: For the curve image uploaded by the user, use a convolutional neural network to perform coordinate axis detection and curve region segmentation, and extract the original pixel set; Step 4.2: Perform coordinate inverse calculation and data point mapping on the pixel set extracted in Step 4.1 to generate a preliminary digital curve; Step 4.3: Perform noise filtering and coordinate normalization on the curve data generated in Step 4.2, and output the target curve data in standard format.

6. The method according to claim 1, characterized in that, Step 5 includes the following steps: Step 5.1: For the current circuit design problem, extract its feature vector, including parameter dimension, nonlinearity, and number of multi-objectives; Step 5.2: For the feature vectors extracted in Step 5.1, use a reinforcement learning-based algorithm selector to match the optimal optimization algorithm from a pre-set algorithm library; Step 5.3: For the algorithm selected in Step 5.2, monitor the convergence status in real time during the operation, and dynamically adjust the hyperparameters or trigger algorithm switching.

7. The method according to claim 1, characterized in that, Step 6 includes the following steps: Step 6.1: For the current parameter combination, the AI ​​agent uses the large language model to infer and calls the API for generating netlists registered for the schematic editor in Step 1.1 to generate a standard netlist file, and submits it to the simulator through the API; Step 6.2: Monitor the simulation status in real time for the simulation task submitted in Step 6.1, and parse the output file to extract key performance indicators upon completion; Step 6.3: Calculate the residuals between the indicators extracted in Step 6.2 and the target curve, and dynamically update the optimization progress and convergence curve.

8. The method according to claim 1, characterized in that, Step 7 includes the following steps: Step 7.1: Perform simulation verification on the parameter combination output by the optimization process and generate a verification report, including performance comparison and residual analysis; Step 7.2: For the residual results in Step 7.1, determine whether they meet the preset threshold. If they do not meet the threshold, trigger the multi-strategy iteration mechanism. Step 7.3: For the iterative process triggered in step 7.2, support parameter range expansion, multi-fidelity model switching, and generate conflict diagnosis reports.

Images