Multimodal large model and dual-agent collaborative circuit transient multi-objective optimization method

By employing a multi-modal large model and a dual-agent collaborative method for transient multi-objective optimization of circuits, this approach solves the problems of difficult feature extraction and high-cost simulation of transient parameters in nonlinear circuits in existing technologies, and achieves stable and efficient circuit optimization.

CN123070047APending Publication Date: 2026-11-10UNIV OF ELECTRONICS SCI & TECH OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611552357.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-09
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing multi-objective optimization methods for circuits suffer from low feature extraction success rate and insufficient robustness when dealing with transient parameters of nonlinear circuits. High-precision transient simulation is costly, and traditional surrogate models lack physical perception, resulting in insufficient optimization efficiency.

Method used

A multimodal large model and dual-agent collaborative circuit transient multi-objective optimization method is adopted. By constructing a low-level mathematical agent model and a high-level multimodal large language model, combined with circuit topology and transient waveform images, cross-modal feature extraction and semantic confidence fusion are achieved. Simulation strategies are dynamically switched to reduce computational costs and improve optimization efficiency.

Benefits of technology

It improves the stability of transient feature extraction for nonlinear circuits, reduces the cost of high-precision simulation calculations, improves global optimization performance, and enhances the engineering practicality of the Pareto front.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070047A_ABST
    Figure CN123070047A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-modal large-scale model and dual-agent collaborative method for circuit transient multi-objective optimization, belonging to the field of integrated circuit and electronic design automation (EDA) optimization design. This invention constructs a dual-agent collaborative evaluation framework consisting of a low-level mathematical agent and a top-level multimodal large-scale language model agent. In the multi-objective optimization iteration, a dynamic switching mechanism based on the prediction variance of the low-level model is used to actively screen and trigger realistic circuit transient simulations. Then, through physically inspired structured prompts, the large model is driven to perform thought chain reasoning, stably extracting complex nonlinear transient feature values ​​and outputting semantic confidence. Finally, this confidence is used as a probability weight and fused into the expected improvement function to guide the low-level evolutionary engine to avoid the physical failure parameter region. This invention effectively improves the stability of complex waveform feature extraction, significantly reduces the computational cost of expensive optimization simulations, and enhances the efficiency and engineering practicality of global optimization for nonlinear circuits.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of integrated circuits and electronic design automation (EDA) optimization design, specifically involving a multimodal large model and dual-agent collaborative circuit transient multi-objective optimization method. Background Technology

[0002] In modern integrated circuit design, the optimization of circuit parameters, such as transistor aspect ratio and resistor / capacitor values, is a crucial factor determining the final performance of the chip. With the shrinking of process nodes and the increase in system complexity, data-driven circuit design methods based on multi-objective optimization algorithms, such as evolutionary algorithms, are gradually becoming a research direction in the EDA field.

[0003] However, existing multi-objective optimization techniques for circuits are mostly applied to optimizing the DC or AC characteristics of simple linear modules, such as the gain and phase margin of a single-stage operational amplifier. These metrics can usually be directly obtained through numerical extraction. But in the design of system-level nonlinear circuits such as low-dropout linear regulators (LDOs) and switching DC-DC converters, the core performance metrics are often reflected in transient time-domain simulation waveforms. For example, the power-on soft-start time of an LDO, the transient response time (recovery time) of a load current jump, and the dead time of a DC-DC converter. Existing extraction methods for these complex transient parameters face the following problems:

[0004] First, transient feature extraction has a low success rate and insufficient robustness. The time-domain waveforms of nonlinear circuits are often accompanied by damped oscillations, non-monotonic changes, glitches, and noise. Existing methods typically rely on writing code scripts in a simulator to capture specific threshold crossing points. This rule-based approach is prone to errors or capturing incorrect feature values ​​when dealing with complex waveforms, thus feeding back incorrect fitness values ​​to the optimization algorithm, leading to abnormal interruptions in the optimization process or getting trapped in local optima.

[0005] Second, high-precision transient simulation is computationally expensive. Transient simulation of system-level circuits typically takes a long time. Evolutionary algorithms require numerous population evaluation iterations; if a real transient simulation is called for each evaluation, it becomes an expensive multi-objective optimization problem, resulting in high computational time costs.

[0006] Third, existing surrogate models lack physical perception capabilities. To reduce the number of simulations required, current technologies typically introduce mathematical surrogate models (such as Gaussian processes and neural networks). However, these traditional surrogate models primarily map circuit parameters to final performance, making it difficult to analyze them in conjunction with circuit topology and waveforms containing physical information. This results in significant prediction errors when dealing with system-level circuits with nonlinear physical constraints, easily misleading the optimization direction.

[0007] Therefore, existing data-driven circuit optimization methods suffer from problems such as difficulty in feature extraction, high simulation costs, and insufficient optimization efficiency due to the lack of physical feature correlation when dealing with transient parameters of nonlinear circuits. Summary of the Invention

[0008] The purpose of this invention is to provide a multi-modal large model and dual-agent collaborative circuit transient multi-objective optimization method to solve the problems of script extraction of transient features being prone to errors, high cost of high-precision transient simulation, and lack of physical perception in traditional mathematical agent models in existing circuit optimization techniques.

[0009] The technical solution adopted in this invention is as follows:

[0010] A multimodal large model and dual-agent collaborative transient multi-objective optimization method for circuits includes the following steps:

[0011] Step 1: Determine the nonlinear circuit topology based on domain knowledge and requirements, and determine the circuit decision variables to be optimized and the transient electrical performance indicators as optimization targets.

[0012] Step 2: Initialize N individuals representing different circuit parameter configurations based on circuit decision variables to form an initial population;

[0013] Step 3: Construct a dual-agent collaborative evaluation framework: Initialize the underlying mathematical agent model A to quickly predict the transient indicators and prediction variance of an individual; introduce the top-level multimodal large language model agent B to extract transient feature values ​​from transient waveform images and output semantic confidence.

[0014] Step 4: Use a multi-objective evolutionary algorithm heuristic operator to generate a new circuit parameterized offspring population;

[0015] Step 5: For each individual in the offspring population, use the underlying mathematical proxy model A to make predictions and output the prediction variance. Based on whether the variance exceeds the dynamically set threshold, decide whether to directly accept the prediction result of the underlying mathematical proxy model A or trigger a real circuit transient simulation.

[0016] Step 6: If a real circuit transient simulation is triggered in step 5, then the transient waveform image data is obtained. Using the constructed physical heuristic structured prompt words, the top-level multimodal large language model agent B is driven to perform chain reasoning to extract the target transient feature value and the corresponding semantic confidence.

[0017] Step 7: Store the data containing the target transient feature value and semantic confidence obtained in Step 6 into the database as anchor points, update the underlying mathematical proxy model A, and integrate the semantic confidence into the calculation of the next generation expectation boosting function;

[0018] Step 8: Update the best individual in the current file based on Pareto dominance;

[0019] Step 9: If the optimal circuit parameter scheme in the current file does not meet the performance index specified by the user and has not reached the maximum number of iterations, repeat steps 4-9; if the iteration termination condition is met, output the final Pareto front scheme of nonlinear circuit parameters.

[0020] Furthermore, step 5 is detailed as follows:

[0021] Calculate the variance of the predictions made by the underlying mathematical proxy model A for the current individual's decision variables;

[0022] When the variance is less than or equal to the threshold, the predicted mean of the underlying mathematical proxy model A is used directly. This serves as the result of the individual's fitness assessment;

[0023] When the variance value exceeds the threshold, the agent prediction is stopped, and the intelligent simulation tool is called to perform a real transient simulation. After the simulation is completed, the top-level multimodal large language model agent B is called to perform feature extraction, and the feature extraction results are used as new samples to repair the cognitive blind spots of the bottom-level mathematical agent model A.

[0024] Furthermore, step 6 is detailed as follows:

[0025] Step 6.1: Construct multimodal input: Convert the circuit topology information, process node parameters, and test excitation conditions of the current individual into text modal data; convert the transient waveforms output by the intelligent simulation tool into visual modal image data;

[0026] Step 6.2: Constructing a Physics-Inspired Thinking Chain: Based on circuit principles, design a segmented reasoning template to guide the top-level multimodal large language model agent B to deduce the following logic: First, identify the transient operating state of the current circuit by combining the test excitation conditions of the text modality; second, observe the visual modality image and locate the key nonlinear feature points in the waveform; finally, combine the physical prior knowledge of the process node parameters to determine the legality of the damped oscillations in the waveform recovery process, and finally calculate the transient feature value, while outputting a semantic confidence score that characterizes the physical rationality of the extraction result. .

[0027] Furthermore, in step 7, the specific process of incorporating semantic confidence into the calculation of the next-generation expectation boosting function is as follows:

[0028] Semantic confidence of proxy B output from top-level multimodal large language model in historical anchor data Construct an auxiliary confidence probability prediction model This is used to predict the compliance probability of any unknown solution at the physical and semantic level; this compliance probability is then incorporated into the traditional expectation boosting function as a penalty term to reduce dimensionality, thus constructing a semantically enhanced expectation boosting function: in, The standard expected improvement value is calculated for the underlying mathematical surrogate model A based on the predicted mean and predicted variance; in subsequent evolution, it utilizes... The most promising offspring individuals will be selected for the next round of real-world simulation evaluation.

[0029] Furthermore, the circuit decision variables include the aspect ratio, resistance value, capacitance value and / or bias current value of transistors in nonlinear circuits; the transient electrical performance indicators include, but are not limited to, the power-on soft-start time of low-dropout linear regulators, the load jump transient response time, and the dead time of switching DC-DC converters.

[0030] Furthermore, the underlying mathematical proxy model A is a Gaussian process regression model or a Kriging model.

[0031] Furthermore, the top-level multimodal large language model agent B has cross-modal processing capabilities for joint understanding of images and text, and achieves online data interaction with multi-objective evolutionary algorithms through the application programming interface (API).

[0032] Furthermore, the Pareto dominance-based approach includes the classical Pareto dominance method, the ε-dominance method, or any high-dimensional non-dominant ranking method suitable for multi-objective optimization environments.

[0033] Furthermore, the method of generating a new circuit parameterized offspring population using a multi-objective evolutionary algorithm heuristic operator specifically includes recombining and mutating the decision variables of the parent individuals at the real-number encoding level using simulated binary crossover, polynomial mutation, or differential evolution mutation operations.

[0034] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0035] 1. Improved stability of nonlinear waveform feature extraction: Unlike methods that rely on scripts to extract threshold points, this invention introduces a multimodal large language model to assist in feature extraction. Through structured prompts, the model can combine the overall trend of the waveform (such as identifying damped oscillations and filtering glitches), thereby stably extracting system-level electrical parameters such as soft-start time and recovery time when facing changing transient responses of nonlinear circuits, reducing the number of times the optimization process is interrupted due to extraction failures.

[0036] 2. Effectively reduces the computational cost of high-precision simulation: This invention employs a dynamic proxy switching mechanism based on prediction variance. The system only triggers high-precision transient simulation when the prediction variance of the underlying mathematical proxy model exceeds a set threshold. This mechanism allocates computing power to the solution space with higher uncertainty, reducing ineffective calls to real simulations, thereby improving overall optimization efficiency within a limited simulation time budget.

[0037] 3. A cross-modal physical perception mechanism is introduced to improve the global optimization effect: Traditional mathematical proxy models mainly rely on numerical mapping. This invention designs a semantic confidence fusion mechanism, which transforms the understanding of circuit legality (confidence score) by the large language model into a probability distribution, and combines it with the underlying multi-objective optimization acquisition function (…). This mechanism enables the underlying optimization process to identify and avoid parameter regions that are prone to causing circuit non-convergence or failure, thus improving the engineering practicality of the Pareto front being optimized. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a multimodal large model and dual-agent collaborative circuit transient multi-objective optimization method provided in an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0040] This invention provides a method for transient multi-objective optimization of circuits using a multimodal large model and dual-agent collaboration, comprising the following steps:

[0041] Step 1: Determine the nonlinear circuit topology based on domain knowledge and requirements, and determine the circuit decision variables to be optimized and the transient electrical performance indicators as optimization targets; the circuit decision variables include the aspect ratio, resistance value, capacitance value and / or bias current value of transistors in the nonlinear circuit; the transient electrical performance indicators include the power-on soft-start time of the low dropout linear regulator, the load jump transient response time, and the dead time of the switching DC-DC converter.

[0042] Step 2: Initialize N individuals representing different circuit parameter configurations based on circuit decision variables to form an initial population;

[0043] Step 3: Construct a dual-agent collaborative evaluation framework: Initialize the underlying mathematical agent model A to quickly predict the transient indicators and prediction variance of an individual; introduce the top-level multimodal large language model agent B to extract complex transient feature values ​​from transient waveform images and output semantic confidence.

[0044] Step 4: Use a multi-objective evolutionary algorithm heuristic operator to generate a new circuit parameterized offspring population;

[0045] Step 5: Dynamic Agent Switching and Active Learning Based on Predicted Variance: For each individual in the offspring population, the underlying mathematical agent model A is used to predict and output the predicted variance. The switching is then determined based on whether the variance exceeds a dynamically set threshold. The decision is whether to directly accept the prediction results of the underlying mathematical proxy model A, or to trigger a real circuit transient simulation.

[0046] The specific rules for dynamic agent switching and active learning based on prediction variance are as follows: Calculate the variance of the predictions made by the underlying mathematical agent model A for the current individual's decision variables. ;

[0047] when When an individual is determined to be within the known confidence region of the underlying mathematical surrogate model A, the predicted mean of the underlying mathematical surrogate model A is directly used. This serves as the result of the individual's fitness assessment;

[0048] when When the underlying mathematical agent model A encounters a cognitive "blind spot", the system stops agent prediction and calls an intelligent simulation tool to perform a real transient simulation. After the simulation is completed, the top-level multimodal large language model agent B is called to perform feature extraction, and the result is used as a new high-quality sample to repair the cognitive blind spot of the underlying mathematical agent model A.

[0049] Step 6, Physics-inspired waveform feature extraction: If a real circuit transient simulation is triggered in step 5, the transient waveform image data is obtained. The constructed physics-inspired structured cue words are used to drive the top-level multimodal large language model agent B to perform chain reasoning and extract the target transient feature value and the corresponding semantic confidence.

[0050] The physical-inspired waveform feature extraction specifically includes the following steps:

[0051] Step 6.1: Construct multimodal input: Convert the circuit topology information, process node parameters, and test stimulus conditions of the current individual into text modal data; convert the transient waveforms output by the SPICE simulator into visual modal image data;

[0052] Step 6.2: Constructing a Physics-Inspired Thinking Chain: Based on circuit principles, design a segmented reasoning template to guide the top-level multimodal large language model agent B to deduce the following logic: First, identify the transient operating state of the current circuit by combining the test excitation conditions of the text modality; second, observe the visual modality image to locate key nonlinear feature points in the waveform, such as the lowest voltage drop point or the rising edge control point; finally, combine the physical prior knowledge of the process node parameters to determine the legality of the damped oscillations in the waveform recovery process, and finally calculate the transient feature values, while outputting a semantic confidence score that characterizes the physical rationality of the extraction result. .

[0053] Step 7: Update the proxy model based on semantic dimensionality reduction feedback: Store the high-quality data containing feature values ​​and semantic confidence obtained in Step 6 as "anchor points" in the database, update the underlying mathematical proxy model A, and integrate the semantic confidence into the calculation of the next generation expectation boosting function to guide the algorithm to avoid physically unstable regions.

[0054] The proxy model update based on semantic dimensionality reduction feedback is achieved through a modified expectation boosting function. The specific process is as follows: the semantic confidence of the top-level multimodal large language model proxy B output from historical "anchor" data is used. Construct an auxiliary confidence probability prediction model This is used to predict the compliance probability of any unknown solution at the physical and semantic level; this compliance probability is then incorporated into the traditional expectation boosting function as a penalty term to reduce dimensionality, thus constructing a semantically enhanced expectation boosting function: in, The standard expected improvement value is calculated for the underlying mathematical surrogate model A based on the predicted mean and predicted variance; in subsequent evolution, it utilizes... The most promising offspring individuals will be selected for the next round of real-world simulation evaluation.

[0055] Step 8: Update the best individual in the current file based on Pareto dominance;

[0056] Step 9: If the optimal circuit parameter scheme in the current file does not meet the performance index specified by the user and has not reached the maximum number of iterations, repeat steps 4-9; if the iteration termination condition is met, output the final Pareto front scheme of nonlinear circuit parameters.

[0057] Furthermore, the underlying mathematical proxy model A is a Gaussian process regression model or a Kriging model, to support the simultaneous output of the predicted mean and the predicted variance.

[0058] Furthermore, the top-level multimodal large language model agent B possesses cross-modal processing capabilities for joint understanding of images and text, and can achieve online data interaction with the underlying evolutionary algorithm framework through an application programming interface (API).

[0059] Furthermore, the Pareto dominance-based approach includes the classic Pareto dominance method, the ε-dominance method, or any high-dimensional non-dominant ranking method suitable for multi-objective optimization environments.

[0060] Furthermore, the generation of offspring using a multi-objective evolutionary algorithm heuristic operator specifically includes recombining and mutating the decision variables of the parent individuals at the real-number encoding level using simulated binary crossover, polynomial mutation, or differential evolution mutation operations.

[0061] Furthermore, this embodiment provides a multi-modal large model and dual-agent collaborative transient multi-objective optimization method for circuits. Unlike traditional linear circuits (such as operational amplifiers) that mainly focus on the frequency domain or static operating point, this embodiment specifically optimizes system-level circuits with extreme piecewise nonlinear characteristics (such as switching power supply controllers containing pulse width modulation (PWM) comparators and dead-time control logic).

[0062] As a specific application scenario implementation method, see Figure 1 This embodiment takes the parameter optimization of a voltage-mode Buck DC-DC buck converter as an example to detail the processing procedure of the present invention:

[0063] Step 1: Determine the nonlinear circuit topology based on domain knowledge and requirements, and identify the decision variables for the circuit to be optimized, as well as the transient electrical performance indicators that serve as the optimization target. In this embodiment, the circuit to be optimized is a Buck switching power supply controller, which internally includes an error amplifier, a PWM comparator, an oscillator, and dead-time control logic, exhibiting significant state transition and nonlinear characteristics. Specifically, this embodiment extracts the dimensions (width-to-length ratio W / L) of the underlying power transistor (MOSFET) and the preceding driver MOSFET, the delay capacitance value in the dead-time control logic, the device parameters in the Miller plateau compensation circuit, and the compensation RC parameters of the error amplifier, etc., to construct a decision variable vector. Simultaneously, the objectives of multi-objective optimization are set as minimizing power consumption, minimizing chip area, and optimizing the transient response time under load jumps. ) and dead time ( These are the four core transient indicators.

[0064] Step 2: Initialize N individuals representing different circuit parameter configurations to form an initial population. Latin hypercube sampling is used to generate the initial N individual parameter combinations within the design space of the circuit decision variables, ensuring a uniform distribution of the initial samples in the multidimensional space.

[0065] Step 3: Construct a dual-agent collaborative evaluation framework. First, initialize a Gaussian process regression model as the underlying mathematical agent model A, which receives real-number encoded circuit parameters and outputs the predicted mean of transient indices. and the prediction variance that characterizes uncertainty Simultaneously, a large model with vision-language multimodal processing capabilities (such as a fine-tuned model based on the Transformer architecture) is deployed as the top-level multimodal large language model agent B, and API communication is established with the main program of the optimization algorithm to extract complex transient feature values ​​from transient waveform images and output semantic confidence.

[0066] Step 4: Generate a new circuit-parameterized offspring population using a multi-objective evolutionary algorithm heuristic operator. Randomly select parent individuals and perform simulated binary crossover and polynomial mutation to generate offspring individuals with the new parameter configuration.

[0067] Step 5: Dynamic Agent Switching and Active Learning Based on Prediction Variance. For each offspring individual, the underlying mathematical agent model A is first input for evaluation. If the prediction variance evaluated by model A... This indicates that the current parameter combination is similar to historical experience, and the predicted mean is directly adopted as the fitness of the individual, skipping SPICE simulation to save computing power; if This indicates that the optimization algorithm has explored an unknown nonlinear mutation region. At this point, the mathematical proxy evaluation is terminated, and the EDA tool is automatically invoked to perform a full-chip transient simulation.

[0068] Step 6: Physics-Inspired Waveform Feature Extraction. For the individuals that triggered the realistic simulation in Step 5, cross-modal transformation and feature extraction are performed. The core of this step lies in guiding the top-level multimodal large language model agent B to perform data parsing and feature extraction step by step according to expert-level logic through a pre-set structured prompt word project. The specific process is as follows:

[0069] Step 6.1: Constructing Multimodal Context Input: First, data extraction and encapsulation are performed. The waveform data of node voltage / current changing over time from the SPICE output is rendered into a high-resolution image with redundant meshes removed, serving as the visual modal input. Simultaneously, the circuit simulation text information of the current individual is extracted to construct the semantic modal input, for example, generating the text: "Current simulation conditions: Voltage-mode Buck converter, input voltage 12V, target output voltage 5V, in..." "At this time, the load current jumps from 10mA to 500mA."

[0070] Step 6.2: Constructing a Physics-Inspired Thinking Chain Prompt Template: The system concatenates the multimodal input into a pre-designed structured thinking chain prompt template. This template aims to guide the top-level multimodal large language model agent B to execute logical reasoning sequentially to extract the "load step transient response time" and determine the "system stability." The specific steps are described below:

[0071] 6.2.1 Physical State Prediction: By providing a prompt containing the message "Analyze what physical phenomenon should occur in the current circuit based on the text excitation conditions", the top-level multimodal large language model agent B is guided to identify the load increase phenomenon that occurs at a specific time point, and based on this, infer the physical process of the output voltage rising again after a transient drop.

[0072] 6.2.2 Visual Feature Anchoring: Guide the top-level multimodal big language model agent B to observe the waveform image to locate the lowest point of voltage drop and the coordinate area that crosses the target voltage error band. Use the visual understanding ability of the top-level multimodal big language model agent B to lock the coordinates of the trough and the coordinates of the time point when the waveform enters the error band and no longer exceeds it.

[0073] 6.2.3 Physical legality and stability test: By including the instruction "Observe the process of voltage rise to steady state, determine whether there is damped oscillation and evaluate its convergence", the model is guided to identify the envelope trend of the waveform, thereby determining whether the system is in a convergent state or a divergent and unstable state.

[0074] 6.2.4 Feature Calculation and Confidence Measurement: The guided model performs time difference calculation based on the aforementioned analysis to obtain the specific value of the transient response time as an accurate feature value, and outputs a semantic confidence score C representing physical rationality in combination with the stability assessment results; if the waveform is determined to be smooth and convergent, C is assigned a value of 1.0; if there are multiple oscillations but eventually converge, C is assigned a value of 0.7; if there are severe oscillations and tends to be unstable, C is assigned a value of 0.1.

[0075] Step 7: Agent Model Update Based on Semantic Dimensionality Reduction Feedback. Using the accurate feature values ​​obtained in Step 6, incrementally train and update the underlying mathematical agent model A to correct its cognitive blind spots. Simultaneously, construct the semantically enhanced expectation boosting function. ,in, For semantic enhancement, the hybrid expectation boost value, The standard expected improvement value calculated for the underlying mathematical proxy model A. It is a probability surface fitted by the semantic confidence feedback from the large language model. In the subsequent offspring selection, the algorithm not only pursues the mathematically optimal goal, but also actively avoids those physically dangerous parameter areas that the large language model judges to be prone to PWM waveform oscillation or dead zone shoot-through.

[0076] Step 8: Update the optimal archive. Compare the evaluated individuals with the individuals in the current archive, and update the Pareto front solution set using non-dominated sorting (such as the fast non-dominated sorting method in the non-dominated sorting genetic algorithm with elitist strategy (NSGA-II) framework).

[0077] Step 9: Algorithm Termination Check. Check if the preset maximum number of SPICE simulation calls or maximum number of generations has been reached. If not, return to Step 4 to continue iteration; if reached, stop optimization and output Pareto optimal nonlinear circuit parameter schemes containing different power consumption, area, and transient response trade-offs for designers to make decisions.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for transient multi-objective optimization of circuits using a multimodal large model and dual-agent collaboration, characterized in that, Includes the following steps: Step 1: Determine the nonlinear circuit topology based on domain knowledge and requirements, and determine the circuit decision variables to be optimized and the transient electrical performance indicators as optimization targets. Step 2: Initialize N individuals representing different circuit parameter configurations based on circuit decision variables to form an initial population; Step 3: Construct a dual-agent collaborative evaluation framework: Initialize the underlying mathematical agent model A to quickly predict the transient indicators and prediction variance of an individual; introduce the top-level multimodal large language model agent B to extract transient feature values ​​from transient waveform images and output semantic confidence. Step 4: Use a multi-objective evolutionary algorithm heuristic operator to generate a new circuit parameterized offspring population; Step 5: For each individual in the offspring population, use the underlying mathematical proxy model A to make predictions and output the prediction variance. Based on whether the variance exceeds the dynamically set threshold, decide whether to directly accept the prediction result of the underlying mathematical proxy model A or trigger a real circuit transient simulation. Step 6: If a real circuit transient simulation is triggered in step 5, then the transient waveform image data is obtained. Using the constructed physical heuristic structured cue words, the top-level multimodal large language model agent B is driven to perform chain reasoning to extract the target transient feature value and the corresponding semantic confidence. Step 7: Store the data containing the target transient feature value and semantic confidence obtained in Step 6 into the database as anchor points, update the underlying mathematical proxy model A, and integrate the semantic confidence into the calculation of the next generation expectation boosting function; Step 8: Update the best individual in the current file based on Pareto dominance; Step 9: If the optimal circuit parameter scheme in the current file does not meet the performance indicators specified by the user and has not reached the maximum number of iterations, repeat steps 4-9; if the iteration termination condition is met, output the final Pareto front scheme of nonlinear circuit parameters.

2. The method for transient multi-objective optimization of circuits using a multimodal large model and dual-agent collaboration as described in claim 1, characterized in that, Step 5 is described in detail below: Calculate the variance of the predictions made by the underlying mathematical proxy model A for the current individual's decision variables; When the variance is less than or equal to the threshold, the predicted mean of the underlying mathematical proxy model A is used directly. This serves as the result of the individual's fitness assessment; When the variance value exceeds the threshold, the proxy prediction is stopped, and the intelligent simulation tool is invoked to perform a real transient simulation. After the simulation is completed, the top-level multimodal large language model agent B is called to perform feature extraction, and the feature extraction results are used as new samples to repair the cognitive blind spots of the bottom-level mathematical agent model A.

3. The method for transient multi-objective optimization of circuits using a multimodal large model and dual-agent collaboration as described in claim 1, characterized in that, Step 6 is as follows: Step 6.1: Construct multimodal input: Convert the circuit topology information, process node parameters, and test excitation conditions of the current individual into text modal data; convert the transient waveforms output by the intelligent simulation tool into visual modal image data; Step 6.2: Constructing a Physics-Inspired Thinking Chain: Based on circuit principles, design a segmented reasoning template to guide the top-level multimodal large language model agent B to deduce the following logic: First, identify the transient operating state of the current circuit by combining the test excitation conditions of the text modality; second, observe the visual modality image and locate the key nonlinear feature points in the waveform; finally, combine the physical prior knowledge of the process node parameters to determine the legality of the damped oscillations in the waveform recovery process, and finally calculate the transient feature value, while outputting a semantic confidence score that characterizes the physical rationality of the extraction result. .

4. A method for transient multi-objective optimization of circuits using a multimodal large model and dual-agent collaboration as described in claim 1 or 3, characterized in that, In step 7, the specific process of incorporating semantic confidence into the calculation of the next-generation expectation boosting function is as follows: Semantic confidence of proxy B output from top-level multimodal large language model in historical anchor data Construct an auxiliary confidence probability prediction model This is used to predict the compliance probability of any unknown solution at the physical and semantic level; this compliance probability is then incorporated into the traditional expectation boosting function as a penalty term to reduce dimensionality, thus constructing a semantically enhanced expectation boosting function: in, The standard expected improvement value is calculated for the underlying mathematical surrogate model A based on the predicted mean and predicted variance; in subsequent evolution, it utilizes... The most promising offspring individuals will be selected for the next round of real-world simulation evaluation.

5. The method for transient multi-objective optimization of circuits using a multimodal large model and dual-agent collaboration as described in claim 1, characterized in that, The circuit decision variables include the aspect ratio, resistance, capacitance, and / or bias current of transistors in nonlinear circuits; the transient electrical performance indicators include the power-on soft-start time of low-dropout linear regulators, the load jump transient response time, and the dead time of switching DC-DC converters.

6. The method for transient multi-objective optimization of circuits using a multimodal large model and dual-agent collaboration as described in claim 1, characterized in that, The underlying mathematical proxy model A is a Gaussian process regression model or a Kriging model.

7. The method for transient multi-objective optimization of circuits using a multimodal large model and dual-agent collaboration as described in claim 1, characterized in that, The top-level multimodal large language model agent B has cross-modal processing capabilities for joint understanding of images and text, and achieves online data interaction with multi-objective evolutionary algorithms through the application programming interface (API).

8. The method for transient multi-objective optimization of circuits using a multimodal large model and dual-agent collaboration as described in claim 1, characterized in that, The Pareto dominance-based approach includes the classic Pareto dominance method, the ε-dominance method, or any high-dimensional non-dominance sorting method suitable for multi-objective optimization environments.

9. The method for transient multi-objective optimization of circuits using a multimodal large model and dual-agent collaboration as described in claim 1, characterized in that, The method of generating a new circuit parameterized offspring population using a multi-objective evolutionary algorithm heuristic operator specifically includes recombining and mutating the decision variables of the parent individuals at the real-number encoding level using simulated binary crossover, polynomial mutation, or differential evolution mutation operations.