Device automated design system, method, and electronic device

Abstract

This invention provides an automated device design system, method, and electronic device, belonging to the field of computer technology. The system includes: a requirements analysis agent, used to analyze device design requirements using a large language model to obtain target design parameters; a structure generation agent, used to generate an initial device structure based on the target design parameters using parametric modeling tools; a simulation verification agent, used to simulate the initial device structure using simulation software to obtain performance indicators of the initial device structure; and an optimization decision agent, used to iteratively adjust the target design parameters using a multi-objective optimization algorithm when performance indicators do not meet set requirements, to obtain optimized design parameters for the device. All the above agents work collaboratively based on an agent orchestration framework to form an automated design process. This invention solves the bottlenecks in efficiency and cost of traditional methods.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically to an automated device design system, an automated device design method, an electronic device, a machine-readable storage medium, and a computer program product. Background Technology

[0002] Traditional device design flows typically begin with manual requirements analysis, where engineers rely on experience to break down design goals into specific parameters. This process is susceptible to subjective factors and suffers from information loss. The subsequent modeling and simulation phase requires engineers to manually construct physical models using TCAD tools and repeatedly adjust parameters for multiphysics simulation verification. During this process, fragmented toolchains lead to inefficient data flow. In the multi-objective optimization stage, engineers must manually weigh competing metrics such as performance, cost, and reliability, often falling into local optima due to a lack of systematic algorithmic support. Finally, the verification phase relies on physical prototype testing, resulting in high trial-and-error costs and lengthy iteration cycles.

[0003] It is evident that as device design evolves towards high performance, miniaturization, and multifunctionality, the bottlenecks of traditional device design methods in terms of efficiency and cost are becoming increasingly prominent. Summary of the Invention

[0004] The purpose of this invention is to provide an automated device design system, method, and electronic device to address the increasingly prominent bottlenecks in efficiency and cost of traditional device design methods.

[0005] To achieve the above objectives, embodiments of the present invention provide an automated device design system, comprising: A requirement parsing agent is used to parse the device design requirements using a large language model to obtain the target design parameters. A structure-generating intelligent agent is used to generate the initial structure of the device based on the target design parameters using a parametric modeling tool; The simulation verification agent is used to simulate the initial structure of the device using simulation software to obtain the performance indicators of the initial structure of the device. An optimization decision-making agent is used to iteratively adjust the target design parameters using a multi-objective optimization algorithm when the performance indicators do not meet the set requirements, so as to obtain the optimized design parameters of the device. The requirement analysis agent, the structure generation agent, the simulation verification agent, and the optimization decision agent work together based on an agent orchestration framework to form an automated design process.

[0006] Optionally, the intelligent agent orchestration framework is the LangChain framework.

[0007] Optionally, the step of using a large language model to analyze the device design requirements and obtain the target design parameters includes: The design intent corresponding to the device design requirements is identified using the large language model. The design intent is associated with at least one of the following: domain knowledge graph, data triples, and process rule base, to obtain the requirement constraints corresponding to the design intent. The large language model is used to generate structured requirements based on the design intent and the requirement constraints; The target design parameters are generated based on the structured requirements using the large language model.

[0008] Optionally, generating target design parameters based on the structured requirements using the large language model includes: Receive the device target parameter requirements corresponding to the structured requirements; The large language model is used to generate device quantization requirements based on the device target parameter requirements and structured requirements; The target design parameters are generated based on the device quantization requirements using the large language model.

[0009] Optionally, the iterative adjustment of the target design parameters using a multi-objective optimization algorithm to obtain the optimized design parameters of the device includes: An initial population is generated based on the target design parameters; the initial population includes multiple individuals, each of which represents a set of device design parameters. Repeat the following steps until the set termination condition is met: Determine the performance metrics for each individual in the current population; A multi-objective optimization algorithm is used to perform population evolution on the current population to obtain an updated population; the updated population includes newly born individuals and some individuals in the current population.

[0010] Optionally, the optimization decision-making agent is further used for: If the number of iterations reaches a set threshold and the performance indicators of the optimized design parameters do not meet the set requirements, a dilemma report is generated. Based on the aforementioned dilemma report, a search is conducted in the knowledge base to obtain device modification suggestions; the knowledge base includes at least one of the following: a historical design case library, a domain knowledge base, and a physical model and empirical formula library; Based on the device modification suggestions, at least one of the constraints, reward function, and initial population of the multi-objective optimization algorithm is modified. The target design parameters are iteratively adjusted again based on the modified multi-objective optimization algorithm to obtain the updated optimized design parameters of the device. Add the updated and optimized design parameters to the knowledge base.

[0011] Optionally, the current population is subjected to population evolution using a multi-objective optimization algorithm to obtain an updated population; the updated population includes newly born individuals and a portion of the individuals in the current population, including: Perform a non-dominated sort on each individual to determine the sorting result for each individual; Calculate the crowding degree for each individual to obtain the crowding degree calculation result for each individual; Genetic operations are performed on each individual based on the sorting results and the crowding calculation results to generate new individuals; The newly created individuals are merged with some individuals in the current population to obtain the updated population.

[0012] Optionally, determining the performance index of each individual in the current population includes: The structure-generating agent is used to generate the device structure corresponding to each individual in the current population. The simulation verification agent is used to simulate the device structure corresponding to each individual, and the performance index corresponding to each individual is obtained.

[0013] On the other hand, embodiments of the present invention also provide an automated device design method, applied to the aforementioned automated device design system, the method comprising: Step 100: Use a large language model to analyze the device design requirements and obtain the target design parameters; Step 200: Generate the initial structure of the device based on the target design parameters using a parametric modeling tool; Step 300: Simulate the initial structure of the device using simulation software to obtain the performance indicators of the initial structure of the device; Step 400: If the performance indicators do not meet the set requirements, use a multi-objective optimization algorithm to iteratively adjust the target design parameters to obtain the optimized design parameters of the device. Steps 100, 200, 300, and 400 are scheduled and coordinated by the intelligent agent orchestration framework to form an automated design process.

[0014] Optionally, the intelligent agent orchestration framework is the LangChain framework.

[0015] Optionally, the step of using a large language model to analyze the device design requirements and obtain the target design parameters includes: The design intent corresponding to the device design requirements is identified using the large language model. The design intent is associated with at least one of the following: domain knowledge graph, data triples, and process rule base, to obtain the requirement constraints corresponding to the design intent. The large language model is used to generate structured requirements based on the design intent and the requirement constraints; The target design parameters are generated based on the structured requirements using the large language model.

[0016] Optionally, generating target design parameters based on the structured requirements using the large language model includes: Receive the device target parameter requirements corresponding to the structured requirements; The large language model is used to generate device quantization requirements based on the device target parameter requirements and structured requirements; The target design parameters are generated based on the device quantization requirements using the large language model.

[0017] Optionally, the iterative adjustment of the target design parameters using a multi-objective optimization algorithm to obtain the optimized design parameters of the device includes: An initial population is generated based on the target design parameters; the initial population includes multiple individuals, each of which represents a set of device design parameters. Repeat the following steps until the set termination condition is met: Determine the performance metrics for each individual in the current population; A multi-objective optimization algorithm is used to perform population evolution on the current population to obtain an updated population; the updated population includes newly born individuals and some individuals in the current population.

[0018] Optionally, the method further includes: If the number of iterations reaches a set threshold and the performance indicators of the optimized design parameters do not meet the set requirements, a dilemma report is generated. Based on the aforementioned dilemma report, a search is conducted in the knowledge base to obtain device modification suggestions; the knowledge base includes at least one of the following: a historical design case library, a domain knowledge base, and a physical model and empirical formula library; Based on the device modification suggestions, at least one of the constraints, reward function, and initial population of the multi-objective optimization algorithm is modified. The target design parameters are iteratively adjusted again based on the modified multi-objective optimization algorithm to obtain the updated optimized design parameters of the device. Add the updated and optimized design parameters to the knowledge base.

[0019] Optionally, the current population is subjected to population evolution using a multi-objective optimization algorithm to obtain an updated population; the updated population includes newly born individuals and a portion of the individuals in the current population, including: Perform a non-dominated sort on each individual to determine the sorting result for each individual; Calculate the crowding degree for each individual to obtain the crowding degree calculation result for each individual; Genetic operations are performed on each individual based on the sorting results and the crowding calculation results to generate new individuals; The newly created individuals are merged with some individuals in the current population to obtain the updated population.

[0020] Optionally, determining the performance index of each individual in the current population includes: The structure-generating agent is used to generate the device structure corresponding to each individual in the current population. The simulation verification agent is used to simulate the device structure corresponding to each individual, and the performance index corresponding to each individual is obtained.

[0021] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described automated device design method.

[0022] On the other hand, the present invention also provides a machine-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described automated device design method.

[0023] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-mentioned automated device design method.

[0024] Through the above technical solutions, the automated design flow formed by the requirement analysis intelligent agent, simulation verification intelligent agent, and optimization decision intelligent agent in the embodiments of the present invention is obtained by orchestration using an intelligent agent orchestration framework. This enables full-process automation in the embodiments of the present invention, intelligently scheduling the design flow using the intelligent agent orchestration framework, eliminating gaps in manual intervention, and improving the efficiency of automated device design. The embodiments of the present invention combine the simulation verification intelligent agent and the optimization decision intelligent agent to achieve iterative adjustment of target design parameters, obtaining optimized design parameters for the device. Through virtual-real fusion iteration, and by using multi-physics simulation data to drive the self-correction of device design parameters, the cost of physical verification is significantly reduced. Therefore, the embodiments of the present invention solve the problems of low efficiency and high cost in traditional device design methods.

[0025] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0026] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the structure of the automated device design system provided by the present invention; Figure 2 This is a flowchart illustrating the automated device design method provided by the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0027] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0028] In recent years, although some research has attempted to automate single steps (such as parameter optimization or modeling scripts), the following limitations still exist: as device design develops towards high performance, miniaturization, and multifunctionality, the bottlenecks of traditional device design methods in terms of efficiency and cost are becoming increasingly prominent; design inputs need to be highly structured and are difficult to understand natural language requirements; various automation tools are isolated from each other, lacking global collaboration and decision-making; the optimization process relies on fixed algorithms and cannot introduce domain knowledge for dynamic adjustments when encountering difficulties. Therefore, there is an urgent need for a device automation design scheme that can understand natural language intent, connect the entire process, and make dynamic decisions based on knowledge.

[0029] Therefore, the purpose of this invention is to provide an automated device design system, method, and electronic device to solve the above-mentioned problems.

[0030] System Implementation Examples Please refer to Figure 1 On the other hand, embodiments of the present invention also provide an automated device design system, including: a requirement analysis agent 10, a structure generation agent 20, a simulation verification agent 30, and an optimization decision agent 40.

[0031] In this invention, the intelligent agent refers to a software module that encapsulates specific domain knowledge, tool invocation capabilities, and decision-making logic. Each intelligent agent can receive tasks, invoke corresponding internal or external tools for processing, and return results. Simultaneously, it can communicate and collaborate with other intelligent agents according to a predetermined process.

[0032] The requirement parsing agent 10 (or requirement parsing agent) is used to parse the device design requirements using a Large Language Model (LLM) to obtain target design parameters. The requirement parsing agent 10 can obtain the device design requirements input by the user in natural language through a natural language interface. It then uses a pre-trained LLM to parse the user-inputted device design requirements, generating structured target design parameters (e.g., including performance parameters and constraints). In one embodiment, the user's device design requirement is: "I want to design a power MOSFET for a server power module, using a 380V input LLC topology, with a target peak efficiency >98%, preferably in a TOLL or DFN8*8 package." The requirement parsing agent 10 inputs the above device design requirements into the LLM, and then uses the pre-trained LLM to parse the device design requirements, obtaining the target design parameters as follows: Device model: Infineon IPT65R033G7 (650V SiC MOSFET); Core parameter (directly affecting efficiency): Drain-source breakdown voltage (V... ( DS )max ): 650V (meets the 20% margin requirement of 380V input); On-resistance (R) ( DS(on) )max Ω: 33mΩ (extremely low, significantly reducing conduction losses); Gate charge (Q9): 110nC (low charge, reducing gate losses during switching); Package: TOLL (HSOF-8, surface mount package, suitable for high power density designs). Efficiency verification: In an LLC circuit with 380V input and 1kW output, peak efficiency > 98.5% (measured data, meeting the stringent efficiency requirements of server power supplies). It should be noted that target design parameters can be stored as files. For example, target design parameters can be saved as a JSON format design target file.

[0033] Structure generation agent 20 (or device structure generation agent) is used to generate the initial device structure based on the target design parameters using a parametric modeling tool. Structure generation agent 20 generates the initial device structure based on the target design parameters of requirement parsing agent 10 by calling (e.g., a Python script, CAD API). The parametric modeling tool can be implemented by integrating modeling software through LangChain's Tools module. For example, exemplary code for integrating modeling software through LangChain's Tools module is as follows: from langchain.tools import Tool cad_tool = Tool(name="CAD modeling", func=generate_structure).

[0034] In one embodiment, the specific process by which the structure generation agent 20 calls the parametric modeling tool to generate a specific structure is illustrated by the following exemplary code: class StructureGenerationAgent: def generate_initial_structure(self, design_targets, llm_guidance): "Generating the initial structure based on target design parameters and LLM guidance" # Step 1: Select the basic structure from the template library template = self.select_template("RESURF_LDMOS") # Step 2: Apply the parameter range recommended by LLM params = { "L_drift": self.sample_param(llm_guidance["Ld_range"], "mean"), "N_drift": self.sample_param(llm_guidance["Nd_range"], "mean"), "T_ox": 14e-9,# 14nm "X_j": 0.4e-6,# 0.4μm "L_gate": 0.35e-6,# 0.35μm "STI_depth": 0.5e-6, # Shallow trench isolation depth } # Step 3: Call the parametric modeling engine structure_script = self.build_tcad_script(template, params) # Step 4: Add monitoring points and comments structure_script += """ # Key Design Parameters (from LLM guidance) # BV target: 120V # Rsp target: 30 mΩ·mm² # Current parameters: Ld={:.1f}μm, Nd={:.1e}cm -3 Tox={:.1f}nm """.format(params["L_drift"]*1e6, params["N_drift"]*1e-6, params["T_ox"]*1e9) return structure_script, params Example of a generated TCAD structural description file (Sentaurus TCL format): #Automatically generated LDMOS initial structure - Based on the target: BV≥120V, Rsp≤30mΩ·mm² #Generation time: 2024-01-01 10:00:00 #Strategy source: LLM-recommended single RESURF structure # Define Materials (sdegeo:create-rectangle (position 0 0 0) (position 50e-6 10e-6 0) "Silicon" "region_1") # Define the drift zone (sdegeo:create-rectangle (position 2e-6 0 0) (position 7.5e-6 0.5e-6 0); Ld=5.5μm, Xj=0.5μm "Silicon" "DriftRegion") # Set drift region doping (sdedr:define-constant-profile "Doping" "PhosphorusActiveConcentration" 2.0e16; Nd=2e16 cm -3 (Median of the LLM recommended range) "DriftRegion") # Define gate oxide layer (sdegeo:create-rectangle (position 0.5e-6 0 0) (position 1.5e-6 14e-9 0); gate length 1.0 μm, oxide layer thickness 14 nm "SiO2" "GateOxide") # Define source / leakage area (sdegeo:create-rectangle (position 0 0 0) (position 1e-6 0.3e-6 0) "Silicon" "Source") (sdegeo:create-rectangle (position 8e-6 0 0) (position 9e-6 0.3e-6 0)"Silicon" "Drain") # Configure source and drain heavy doping (sdedr:define-constant-profile "Doping" "ArsenicActiveConcentration"1e20 "Source") (sdedr:define-constant-profile "Doping" "ArsenicActiveConcentration"1e20 "Drain") # Add metal contact (sdegeo:define-contact-set "source" 4.0 (color:rgb 1 0 0) "##") (sdegeo:define-contact-set "drain" 4.0 (color:rgb 0 0 1) "##") (sdegeo:define-contact-set "gate" 4.0 (color:rgb 0 1 0) "##") # Set the mesh: Refine key areas (drift areas, grid edges) (sdedr:define-multibox-size 0.1e-6 0.05e-6 0.0 "DriftRegion") (sdedr:define-multibox-size 0.05e-6 0.01e-6 0.0 "GateOxide" "DriftRegion" 1e-6) # Save the structure file (sde:save-model "initial_ldmos_structure.tdr").

[0035] The simulation verification agent 30 is used to simulate the initial structure of the device using simulation software to obtain the performance indicators of the initial structure. The simulation software is integrated through the Tools module of LangChain. Example code for integrating the simulation software through the Tools module of LangChain is as follows: from langchain.tools import Tool simulation_tool = Tool(name="Simulation Analysis", func=run_simulation).

[0036] The simulation verification agent 30 can use simulation software such as Sentaurus TCAD, ANSYS, and COMSOL to simulate the initial structure of the device, automatically run multiphysics analysis, and evaluate performance indicators. These performance indicators include various parameters, such as turn-off loss (Eoff), short-circuit withstand time (SCWT), on-resistance, junction temperature, and thermal stress. For example, in one embodiment, the core of simulating the initial structure of the device and obtaining its performance indicators is "geometry → materials → physics → mesh → solution → post-processing." The specific process includes: 1. Geometry: Using the "COMSOL CAD" module to draw the MOSFET structure; 2. Materials: Selecting materials such as silicon and copper from the "material library"; 3. Physics: Adding the "semiconductor module" (electric field), "heat transfer module" (thermal field), and "structural mechanics module" (structural field), and setting coupling; 4. Mesh: Using the "automatic mesh" function; 5. Solution: Clicking "calculate," the software automatically performs multiphysics coupling solution; 6. Post-processing: Using the "results" module to extract indicators such as on-resistance, junction temperature, and thermal stress. This invention enables intelligent access to the CAD / TCAD / EDA toolchain through the multi-tool collaborative scheduling mechanism of the LangChain framework.

[0037] The optimization decision-making agent 40 is used to iteratively adjust the target design parameters using a multi-objective optimization algorithm when the performance indicators do not meet the set requirements, thereby obtaining the optimized design parameters of the device. When the performance indicators of the initial device structure obtained by the simulation verification agent 30 do not meet the device design requirements, the optimization decision-making agent 40 can use multi-objective optimization algorithms such as Non-Dominated Sorting Genetic Algorithm Generation 1 (NSGA-I), Non-Dominated Sorting Genetic Algorithm Generation 2 (NSGA-II), and decomposition-based multi-objective evolutionary algorithm to iteratively adjust the target design parameters, thereby obtaining the optimized design parameters of the device. Therefore, this embodiment of the invention, by combining the simulation verification agent 30 and the optimization decision-making agent 40, significantly reduces the cost of physical verification by using multi-physics simulation data to drive the self-correction of device design parameters.

[0038] The requirement analysis agent, structure generation agent, simulation verification agent, and optimization decision agent collaborate based on an agent orchestration framework to form an automated design process. In one embodiment, the automated design process formed by the requirement analysis agent 10, structure generation agent 20, simulation verification agent 30, and optimization decision agent 40 is orchestrated using the LangChain framework. The LangChain framework assigns a specific role to each agent (e.g., structure generation agent 20 corresponds to "structural engineer," simulation verification agent 30 corresponds to "simulation expert"), and coordinates tasks through a standardized operating procedure (SOP). In one embodiment, the LangChain-driven task orchestration is a chained process. The design stages are linked together using SequentialChain, for example: design_chain = SequentialChain([ demand_analysis_agent, structure_generation_agent, simulation_agent, optimization_agent ]) This invention, based on the LangChain ReAct framework (core technologies including the dynamic decision tree construction method: SequentialChain & AgentExecutor), implements an automated process generation technology for device design based on a multi-agent framework. This includes a design requirement parsing agent 10, a device structure generation agent 20, a simulation verification agent 30, and an optimization decision agent 40. This embodiment utilizes a natural language-driven multi-agent collaborative architecture, based on an LLM (Large Language Model) and LangChain framework multi-agent interaction system, allowing users to input complete device design requirements via natural language. The requirement parsing agent 10 generates target design parameters. The target design parameters are iteratively adjusted through the combination of the structure generation agent 20, the simulation verification agent 30, and the optimization decision agent 40 to obtain the optimized design parameters for the device.

[0039] The automated design flow formed by the requirement analysis agent 10, simulation verification agent 30, and optimization decision agent 40 in this embodiment of the invention is orchestrated using an agent orchestration framework. This enables fully automated design throughout the entire process. The intelligent scheduling of the design flow using the agent orchestration framework eliminates gaps in manual intervention and improves the efficiency of automated device design. This embodiment of the invention combines the simulation verification agent 30 and the optimization decision agent 40 to iteratively adjust the target design parameters, obtaining the optimized design parameters for the device. Through virtual-real fusion iteration and multi-physics simulation data-driven self-correction of device design parameters, the cost of physical verification is significantly reduced. Therefore, this embodiment of the invention solves the problems of low efficiency and high cost in traditional device design methods.

[0040] In other aspects of the embodiments of the present invention, the step of using a large language model to parse the device design requirements and obtain target design parameters includes: using the large language model to identify the design intent corresponding to the device design requirements; associating the design intent with at least one of a domain knowledge graph, data triples, and a process rule base to obtain the requirement constraints corresponding to the design intent; using the large language model to generate structured requirements based on the design intent and the requirement constraints; and using the large language model to generate target design parameters based on the structured requirements.

[0041] In some embodiments, a user's device design requirements may be expressed in vague and colloquial terms. For example, a user's device design requirement might be, "I want to design a power MOSFET for a server power module, hoping it will be highly efficient, small in size, and able to operate stably for a long time." This requirement includes general requirements in three dimensions: performance (efficiency), size (volume), and reliability (stability). Therefore, it is necessary to use a large language model to analyze and associate the device design requirements with relevant knowledge to obtain structured requirements. For example, firstly, the large language model is used to identify key entities and intents, that is, to identify the design intent corresponding to the device design requirement. For example, the design intent identified by the large language model for the above device design requirement, "I want to design a power MOSFET for a server power module, hoping it will be highly efficient, small in size, and able to operate stably for a long time," is as follows: Application scenario: "Server power module" → belongs to the category of data center power supply.

[0042] Device type: "Power MOSFET" → The technology route is determined to be silicon-based power devices (possibly high-voltage superjunction MOSFETs or medium-to-low-voltage Trench MOSFETs).

[0043] Performance target: "High efficiency" → corresponding to low conduction loss and low switching loss.

[0044] Physical goal: "Small size" → corresponding to high power density, which needs to be achieved by optimizing chip area and packaging.

[0045] Reliability target: "Long-term stable operation" → corresponding to reliability at high temperatures (such as HTRB, H3TRB) and robustness (such as avalanche capability).

[0046] Next, the design intent is associated with at least one of the following: a domain knowledge graph, data triples, and a process rule base, to obtain the corresponding requirement constraints. That is, the design intent is associated with one or more of these methods to obtain the corresponding requirement constraints. For example, in one embodiment, the knowledge association of the design intent is performed as follows: Querying the domain knowledge graph and / or the process rule base: Retrieving typical design specifications (such as efficiency requirements for 80 PLUS Titanium certification), common topologies (such as LLC resonant converters), and operating conditions (such as input voltage 380Vdc, switching frequency 100-300kHz) related to "server power supply" and "power MOSFET". Matching data triples: From the "materials-process-performance" library, searching for typical performance boundaries (such as Rdson*A, FOM) of MOSFETs with different structures (planar, trench, superjunction) under similar applications (such as data center PSUs). This yields the requirement constraints corresponding to the design intent. For example, requirement constraints could include low on-resistance (Rds(on)) and excellent switching characteristics (low Qg, Qgd). Requirement could also include optimized chip specific on-resistance (Rsp) and the use of highly integrated packaging. Furthermore, it could require the device to meet high temperature and high humidity reverse bias (H3TRB), high temperature gate bias (HTGB), and high avalanche energy (EAS) requirements. A large language model is then used to generate structured requirements based on the design intent and requirement constraints. In one embodiment, the structured requirements are described as follows: "Based on your description, you need a power MOSFET for server power supplies (AC-DC or DC-DC level). We break this requirement down and relate it to the following quantifiable design dimensions:" 1. Electrical performance: The core objective is to improve the overall efficiency, which requires the device to have low on-resistance (Rds(on)) and excellent switching characteristics (low Qg, Qgd).

[0047] 2. Physical size: The core objective is to increase power density, which requires optimizing the chip's specific on-resistance (Rsp) and using highly integrated packaging.

[0048] 3. Reliability: The core objective is to ensure long-term operation in harsh environments, which requires the device to meet the requirements of high temperature and high humidity reverse bias (H3TRB), high temperature gate bias (HTGB) and high avalanche energy (EAS).

[0049] Please confirm that the above understanding is correct, or do you have more specific metrics (such as target efficiency value, expected package size)? If the user provides no further details regarding the structured requirements described above, the target design parameters will be generated based on these requirements using the large language model. This process transforms the vague notions of "good" in the user's device design requirements into professional, structured problem definitions, which are then manually verified to prevent subsequent deviations and significantly improve the success rate of device design.

[0050] In other aspects of the embodiments of the present invention, the step of generating target design parameters based on the structured requirements using the large language model includes: receiving device target parameter requirements corresponding to the structured requirements; generating device quantization requirements based on the device target parameter requirements and the structured requirements using the large language model; and generating target design parameters based on the device quantization requirements using the large language model.

[0051] If the user has specific device target parameter requirements for the above structured requirements description, for example, the user confirms and adds: "Yes, for an LLC topology with 380V input, target peak efficiency >98%, preferably TOLL or DFN8*8 package," then it is necessary to receive the device target parameter requirements corresponding to the structured requirements; and use the large language model to generate device quantization requirements based on the device target parameter requirements and structured requirements. For example, in one embodiment, the large language model is used to perform target quantization and parameter mapping based on the device target parameter requirements and structured requirements to obtain the following device quantization requirements: Efficiency target (98%) → By combining the LLC topology model and known controller and transformer losses, the budget for MOSFET losses is calculated backwards. For example, the total loss per MOSFET allowed by the system is ≤1.5W.

[0052] Loss allocation: Based on the switching frequency (assuming 150kHz), the loss budget is decomposed into conduction loss and switching loss.

[0053] Conduction loss budget: ~1.0W → The maximum allowable Rds(on) is derived by reversing the current at a specific current (e.g., 20A RMS).

[0054] Switching loss budget: ~0.5W → deduced from the constraints on gate charge (Qg, Qgd) and output charge (Qoss).

[0055] Packaging constraints (TOLL / DFN8*8) determine the maximum allowable area of ​​the chip, thus defining the core challenge of "achieving the target Rds(on) within a given area".

[0056] Reliability targets are directly related to industry standards, such as H3TRB@125°C, 1000 hours, and EAS requirements of ≥200mJ.

[0057] Based on the above design target list for device quantization requirements, the large language model will ultimately output a machine-readable and easily reviewable design target file (e.g., in JSON format) as follows: json { "design_request": { "user_input": "A power MOSFET designed for server power supplies, characterized by high efficiency, small size, and long-term stability." "clarified_input": "LLC resonant converter for 380V input, target peak efficiency >98%, TOLL or DFN8x8 package preferred." }, "quantified_targets": { "application_context": { "topology": "LLC_Resonant_Converter", "input_voltage": "380V_dc", "switching_frequency": "150kHz (estimated)", "cooling_condition": "Forced_Air_Cooling" }, "electrical_performance": { "efficiency_target": ">98% (peak)", "total_power_loss_budget_per_fet": "1.5W", "conduction_loss_budget": "1.0W @ 20A_rms", "max_rdson_at_25c": "5.0 mΩ (derived)", "switching_loss_budget": "0.5W", "max_total_gate_charge_qg": "120 nC (derived)", "max_gate_drain_charge_qgd": "25 nC (derived)", "breakdown_voltage_bv": "≥ 650 V" }, "physical_constraints": { "preferred_packages": ["TOLL", "DFN8x8"], "max_die_area": ​​"25 mm² (estimated based on package)" }, "reliability_requirements": { "htrb": "1000h @ 150°C, Vds = 520V", "h3trb": "1000h @ 125°C, 85% RH, Vds = 520V", "avalanche_ruggedness_eas": "≥ 200 mJ (single pulse)", "operating_junction_temp": "≤ 175°C" }, "compliance_standards": ["AEC-Q101", "RoHS"] } } The target design parameters represented by the above example code are as follows: Application scenarios: Topology: LLC resonant converter; Input voltage: 380V DC; Switching frequency: 150kHz (estimated); Heat dissipation: forced air cooling.

[0058] Electrical performance indicators: Efficiency targets: Peak efficiency > 98%; Total power loss budget (single tube): 1.5 watts; On-state loss budget: ≤ 1.0 watt at 20A RMS current; Maximum on-resistance at 25°C (R0). DS(on) : 5.0 milliohms (derived value); Switching loss budget: 0.5 watts; Total gate charge (Q G : ≤120 nanocoulombs; gate drain charge (Q) GD ≤25 nanocoulombs; breakdown voltage (V) BR ): ≥650 volts.

[0059] physical constraints Recommended package: TOLL (HSOF-8), DFN8x8; Maximum chip area: 25 square millimeters (based on package estimation).

[0060] Reliability requirements High Temperature Reverse Bias Test (HTRB): 1000 hours @ 150°C, V DS =520V; High temperature and high humidity reverse bias test (H3TRB): 1000 hours @ 125°C, 85% humidity, V DS =520V; Avalanche tolerance (E AS ): Single pulse ≥200 mJ; Operating junction temperature: ≤175°C.

[0061] Compliance Standards AEC-Q101 (Automotive Electronics Reliability Standard); RoHS (Lead-Free / Hazardous Substances Restriction Directive).

[0062] Therefore, this embodiment of the invention utilizes a large language model to perform requirement analysis and knowledge association on the device design requirements input by the user in natural language, as well as to perform target quantification and parameter mapping to obtain target design parameters, thereby realizing the mapping from performance indicators to physical parameters and improving the comprehensiveness and accuracy of device design requirement analysis.

[0063] In other aspects of the embodiments of the present invention, the step of iteratively adjusting the target design parameters using a multi-objective optimization algorithm to obtain optimized design parameters for the device includes: generating an initial population based on the target design parameters; the initial population includes multiple individuals, each individual representing a set of device design parameters; repeating the following steps until a set termination condition is met: determining the performance index of each individual in the current population; performing population evolution on the current population using a multi-objective optimization algorithm to obtain an updated population; the updated population includes newly generated individuals and some individuals in the current population.

[0064] This invention can utilize various multi-objective optimization algorithms to iteratively adjust the target design parameters, thereby obtaining the optimized design parameters of the device. This invention uses the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) as an example. It should be noted that the optimization decision-making agent 40 in this embodiment does not directly operate the TCAD software, but rather collaborates with simulation management agents (such as simulation verification agent 30 and structure generation agent 20) to construct an "optimization loop." The steps of the Non-Dominated Sorting Genetic Algorithm II are as follows: Step 1: Initialization and target setting; Step 2: Generating the initial population; Step 3: Evaluating the population; Step 4: The core algorithm of the Non-Dominated Sorting Genetic Algorithm II evolves; Step 5: Iterative loop; Step 6: Convergence and output.

[0065] The first step involves initialization and target setting. Initialization can involve setting termination conditions. For example, the termination condition could be that optimization terminates when any of the following conditions are met: 1. The maximum number of iterations is reached (e.g., 100 iterations). 2. There is no significant improvement in the Pareto front for N consecutive iterations. 3. A preset performance target is achieved (e.g., breakdown voltage BV > 700V and specific on-resistance Rsp < 20). Target setting could be, for example, finding the optimal device design parameters. The second step involves randomly generating an initial population within the parameter range of the target design parameters of the requirement analysis agent 10. The initial population includes multiple individuals, each representing a set of device design parameters. For example, in one embodiment, the initial population includes 100 sets of device design parameters. The third step requires determining the performance index of each individual in the current population. In one embodiment, determining the performance index of each individual in the current population includes: generating the device structure corresponding to each individual in the current population using the structure generation agent 20; and simulating the device structure corresponding to each individual using the simulation verification agent 30 to obtain the performance index corresponding to each individual. For example, the optimization decision agent 40 submits 100 sets of device design parameters in batches to the simulation management agent. The simulation management agent receives these 100 tasks. It calls the structure generation agent 20 to generate a corresponding TCAD structure file for each set of parameters. Then it calls the simulation verification agent 30 to run the TCAD simulation in parallel and obtain the performance indicators (breakdown voltage BV, specific on-resistance Rsp, total gate charge Qg) for each individual. The simulation management agent summarizes all the results (100 rows of data) and returns them to the optimization decision agent 40. In the fourth step, a multi-objective optimization algorithm is used to perform population evolution on the current population to obtain an updated population. The updated population includes new individuals and some individuals in the current population, including: performing non-dominated sorting on each individual to determine the sorting result of each individual; calculating the crowding degree for each individual to obtain the crowding degree calculation result of each individual; performing genetic operations based on the sorting result and the crowding degree calculation result of each individual to generate new individuals; merging the new individuals with some individuals in the current population to obtain the updated population. For example, after the optimization decision-making agent 40 receives the evaluation results, it begins to execute the core steps of the second generation of the non-dominated sorting genetic algorithm: 1. Fast non-dominated sorting; 2. Crowding calculation; 3. Genetic operation (generating the next generation). The goal of fast non-dominated sorting is to find the "Pareto optimal solution" in the population. This is done by comparing the objective function values ​​between individuals. The result is that individuals 2 (minimum Rsp) and 3 (maximum BV) are likely to be ranked at the first frontier (Rank 1), as they are not mutually dominant and dominate other individuals. Individual 1 may be ranked at the second frontier (Rank 2). The goal of crowding calculation is to maintain the diversity of solutions within the same non-dominated level (e.g., Rank 1), avoiding all solutions from crowding into a single local region.The method involves calculating the distance of each solution to its neighboring solutions in the target space (BV-Rsp-Qg 3D space). The result shows that individuals 2 and 3, being far apart in the target space and having high crowding, are retained. Genetic operations include selection, crossover, and mutation (generating the next generation). Selection chooses superior "parents" based on Rank (lower is better) and crowding (higher is better). Crossover randomly selects two parents, swaps some design variables, and produces "offspring." For example, offspring might inherit P_width from individual 2 and Epitaxy_Thickness from individual 3. Mutation involves applying small random perturbations to some variables of the offspring to explore new design spaces. For example, the offspring's Doping_P might be randomly changed from 2.0e16 to 2.1e16. In the fifth step, the newly generated offspring are merged with some superior parents to form the second generation population (also 100 individuals). Then, the process returns to the third step to evaluate the second generation population. Finally, the process returns to the fourth step to perform non-dominated sorting, crowding calculation, and genetic operations on the second generation population. This process is repeated, typically for 50-200 generations. In step six, optimization terminates when any of the following conditions are met: 1. The maximum number of iterations is reached (e.g., 100 generations). 2. There is no significant improvement in the Pareto front for N consecutive generations. 3. The preset performance target is achieved (e.g., BV > 700V and Rsp < 20). It should be noted that embodiments of the present invention can also utilize the LangChain framework's AgentExecutor to dynamically select optimization strategies. For example, when the current multi-objective optimization algorithm's optimization effect is unsatisfactory, the LangChain framework's AgentExecutor can be used to switch multi-objective optimization algorithms to iteratively adjust the target design parameters.

[0066] It should be noted that the optimization decision-making agent 40 can also use ConversationBufferMemory to record historical device design parameters to ensure the continuity of multiple iterations.

[0067] Traditional device design methods rely on empirical trial and error for multi-objective optimization, making it difficult to systematically quantify the complex coupling relationships between indicators and prone to getting trapped in local optima. This invention addresses this by employing a multi-objective Pareto front reinforcement learning search strategy (NSGA-II and LLM) and iterative path optimization (path reorganization based on simulation feedback) within the LangChain framework to achieve global optimal solution search under complex constraints.

[0068] In other aspects of this invention, the optimization decision agent 40 is further configured to generate a dilemma report when the number of iterations reaches a set threshold and the performance indicators of the optimized design parameters do not meet set requirements; retrieve device modification suggestions from a knowledge base based on the dilemma report; the knowledge base includes at least one of a historical design case library, a domain knowledge base, and a physical model and empirical formula library; modify at least one of the constraints, reward function, and initial population of the multi-objective optimization algorithm based on the device modification suggestions; iteratively adjust the target design parameters again based on the modified multi-objective optimization algorithm to obtain updated optimized design parameters for the device; and add the updated optimized design parameters to the knowledge base.

[0069] Suppose we are optimizing a silicon-based IGBT (Insulated Gate Bipolar Transistor) with the goal of reducing turn-off loss (Eoff) and maintaining sufficient short-circuit withstand time (SCWT). However, a typical contradiction arises during the optimization process: "Whenever the parameters are adjusted to reduce the turn-off loss Eoff, the short-circuit withstand time (SCWT) drops sharply, and vice versa." This is a well-known trade-off dilemma in power device design. Multi-objective optimization algorithms may oscillate repeatedly around this "cliff," making it difficult to find a satisfactory solution. In this case, embodiments of the present invention can use a knowledge enhancement system to assist the multi-objective optimization algorithm in iteratively adjusting the target design parameters to obtain updated optimized design parameters for the device. Embodiments of the present invention can pre-establish one or more of the following: a historical design case library, a domain knowledge library, and a physical model and empirical formula library. In one embodiment, a historical design case library, a domain knowledge library, and a physical model and empirical formula library for semiconductors are pre-established. This includes a pre-established historical design case library containing thousands of similar cases related to the trade-off dilemma of reducing turn-off loss (Eoff) and short-circuit withstand time (SCWT), and the historical design cases are stored in a vectorized manner. In one embodiment, the vectorized storage of historical design cases is implemented through the following exemplary code: Python # Example: Vectorized storage of historical design cases historical_cases = [ { "id": "IGBT_Case_2023_001", "design_targets": {"BV": 1200, "Eoff": 2.0, "SCWT": 10}, "key_parameters": { "buffer_layer_doping": 5e16, "field_stop_thickness": 60e-6, "gate_oxide_thickness": 110e-9, "p_base_doping": 2e17, "lifetime_killing_level": "medium" }, "performance_results": {"Eoff_actual": 1.95, "SCWT_actual": 9.8, "Vce_sat": 1.85}, "lessons_learned": "A good balance is achieved between Eoff and SCWT by introducing a stepped-doped field stop layer", "designer_notes": "Note: p-base doping should not be lower than 1.8e17, otherwise the risk of latch-up increases." "tags": ["IGBT", "1200V", "tradeoff", "field_stop"] }, # ... Thousands of similar cases ] The domain knowledge base can include at least one of the following: textbook chapters, technical papers, patent documents, design rule manuals, and failure analysis reports. Textbook chapters can be discussions on the "Eoff-SCWT tradeoff," such as in *Power Semiconductor Device Physics*. Technical papers can be papers such as "A Novel Carrier Stored Layer for IGBT Trade-off Improvement." Patent documents can be patent documents such as the US Patent "IGBT with improved short-circuit capability." Design rule manuals can be process constraints and reliability rules provided by the fab manufacturer. Failure analysis reports can be historical failure cases caused by insufficient short-circuit withstand time (SCWT).

[0070] The vectorized storage of the physical model and empirical formula library is implemented through the following exemplary code: # Vectorization of Empirical Formulas empirical_knowledge = [ { "formula": "SCWT ∝ (W_n-)^2 / (μ_n * E_C)", "explanation": "Short-circuit withstand time is directly proportional to the square of the N-region width and inversely proportional to electron mobility and critical electric field." "application_context": "Preliminary estimation of IGBT short-circuit capability", "confidence": 0.85 }, { "formula": "Eoff ≈ Q_rr * V_dc + (I_c * t_f * V_dc) / 6", "explanation": "Turn-off losses are mainly determined by reverse recovery charge and current tailing." "application_context": "IGBT turn-off loss analysis", "confidence": 0.9 } ] All these documents, including the historical design case library, domain knowledge base, and physical model and empirical formula library, are converted into vector embeddings and stored in a vector database (such as ChromaDB) to form a multimodal knowledge graph.

[0071] When the optimization decision agent 40 finds that 10 consecutive generations (a set threshold of number of optimizations) cannot simultaneously improve the shutdown loss Eoff and short-circuit withstand time SCWT, it generates a dilemma report and reports the dilemma to the AgentExecutor in parallel. Example code for the optimization decision agent 40 to generate the dilemma report is as follows: # Report on the challenges of optimization_agent dilemma_report = { "problem_type": "TRADEOFF_STAGNATION", "conflicting_objectives": ["Eoff", "SCWT"], "current_status": { "best_Eoff": 1.8,# mJ (but the corresponding SCWT is only 6μs) "best_SCWT": 12,# μs (but the corresponding Eoff is as high as 3.2 mJ) "pareto_front_stagnant": True, "generations_stagnant": 10 }, "current_design_region": { "buffer_doping_range": [3e16, 8e16], "field_stop_range": [40e-6, 80e-6], "key_observations": "Lowering Eoff will inevitably lead to SCWT < 8 μs, while increasing SCWT will inevitably lead to Eoff > 3 mJ" } } Upon receiving the report, AgentExecutor immediately initiates a knowledge retrieval. AgentExecutor searches the dilemma report based on one or more of the following: a historical design case library, a domain knowledge library, and a physical model and empirical formula library, to obtain device modification suggestions. In one embodiment, AgentExecutor searches the dilemma report based on the historical design case library to obtain device modification suggestions. An example code for searching the dilemma report based on the historical design case library to obtain device modification suggestions is as follows: # Constructing a search query retrieval_query =""" Seek historical design examples that successfully balance IGBT turn-off loss (Eoff) and short-circuit capability (SCWT).

[0072] Current challenge: It seems impossible to achieve Eoff < 2.0 mJ and SCWT > 10 μs simultaneously.

[0073] Special attention should be paid to: field cutoff layer design, carrier lifetime control, and buffer layer optimization techniques.

[0074] """ Example of search results: # Vector similarity retrieval (performed in ChromaDB) similar_cases = vector_db.similarity_search( query=retrieval_query, filter={"device_type": "IGBT", "voltage_class": "1200V"}, k=5# Return the 5 most similar cases ) [ { "case_id": "IGBT_Case_2022_045", "similarity_score": 0.92, Summary: "Using a 'dual gradient field cutoff layer' design, SCWT=11μs was achieved at Eoff=1.9mJ". "key_insight": "Instead of uniformly changing the field stop layer concentration, it is designed as a gradient structure with highly doped inner layers and lightly doped outer layers." }, { "case_id": "Patent_US20210057123", "similarity_score": 0.88, "summary": "Improving the trade-off through local carrier lifetime control", "key_insight": "Light electron irradiation is applied only to the collector side to maintain the integrity of the carrier lifetime in the P-base region." } ] The following is an example code for retrieving device modification suggestions from the domain knowledge base based on the aforementioned dilemma report: # Searching for physical principles physics_knowledge = vector_db.similarity_search( query="The physical trade-off mechanism between IGBT Eoff and SCWT, and how to break through the traditional trade-off curve", filter={"doc_type": "textbook_chapter"}, k=3 ) Returned physics knowledge: 1. **The fundamental contradiction mechanism:** - Reducing Eoff requires decreasing the stored charge during turn-off → This necessitates reducing minority carrier lifetime. - Improving SCWT requires enhancing current saturation capability → Maintaining sufficient carrier concentration and mobility is necessary. 2. **Breakthrough Thinking** (from textbook): - **Non-uniform lifetime control:** Introducing composite centers only in specific regions. - **Vertical Doping Optimization**: Fine-tuning the distribution of buffer and field stop layers - **Auxiliary emitter design:** Adding a local P+ region to enhance current saturation.

[0075] The following is an example code demonstrating the device modification suggestions obtained by retrieving the dilemma report based on the physical model and empirical formula library: # Retrieve relevant empirical formulas empirical_formulas = vector_db.similarity_search( query="Quantitative Relationship Formula between SCWT and Structural Parameters", filter={"knowledge_type": "empirical_formula"}, k=2 ) The returned empirical formula: { "formula": "SCWT ≈ (q * N_D * W^2) / (2 * ε_s * v_sat * E_C)", "parameters": { "W": "N - Drift zone width", "N_D": "Drift region doping concentration", "v_sat": "Carrier saturation velocity (~1e7 cm / s)" }, "implication": "SCWT is directly proportional to W^2! A small increase in W can significantly increase SCWT." } Upon receiving device modification suggestions, AgentExecutor will retrieve the following example code for the knowledge enhancement suggestion package: # Generated knowledge enhancement suggestions knowledge_enhanced_suggestion = { "problem_diagnosis": "You have encountered the classic IGBT Eoff-SCWT tradeoff". "historical_solutions": [ { "approach": "Dual gradient field cutoff layer", "success_case": "IGBT_Case_2022_045", "implementation": "The field cutoff layer is divided into two layers: the 50nm area near the collector is doped with 8e16, and the remaining part is doped with 3e16". "expected_improvement": "SCWT>10μs can be obtained at Eoff<2.0mJ" }, { "approach": "Selective lifetime control" "success_case": "Patent_US20210057123", "implementation": "Irradiation with 1e12 / cm² electrons is performed only within a 50μm depth on the collector side". "expected_improvement": "Eoff decreased by 20%, while SCWT only decreased by 5%" } ], "physical_insights": { "key_mechanism": "Eoff is mainly determined by the minority carrier extraction rate at turn-off, while SCWT is determined by the electric field distribution at current saturation." "breakthrough_point": "Through vertical non-uniform design, the turn-off process and the saturation process can be optimized separately." }, "quantitative_guidance": { "sensitivity_analysis": "According to empirical formulas, SCWT is most sensitive to the drift region width W (quadratic relationship)". "recommended_action": "Prioritize increasing W from 120μm to 130μm, expected to improve SCWT by ~18%, and Eoff by only ~5%", "parameter_adjustment_range": { "W_drift": [125e-6, 135e-6], # Original range [110e-6, 125e-6] "buffer_doping": [4e16, 6e16], # Original range [3e16, 8e16] "lifetime_killing": ["localized", "graded"]# New design dimension } }, "warnings_and_constraints": [ Warning: Increasing W will cause Vce_sat to rise; other parameters need to be optimized simultaneously to compensate. "Constraint: According to Fab design rules, the maximum W cannot exceed 140μm (lithographic alignment limit)". "Reliability Note: The thermal resistance at the gradient doping interface needs to be reassessed." ], "recommended_next_steps": [ 1. Add a 'gradient design flag' to the optimization variables to allow the algorithm to explore non-uniform structures. 2. Extend the upper limit of the search range for W to 135μm. 3. Prioritize exploring the 'dual-gradient field cutoff layer' approach (historical success rate 68%). 4. Configure monitoring: Focus on changes in Vce_sat, ensuring it does not exceed 2.1V. ] } After receiving the knowledge enhancement suggestion package (i.e., device modification suggestion), the optimization decision agent 40 does not blindly follow it, but intelligently integrates it into the optimization process. For example, the optimization decision agent 40 modifies one or more of the constraints, reward function, and initial population of the multi-objective optimization algorithm based on the device modification suggestion. In one embodiment, the optimization decision agent 40 modifies the constraints, reward function, and initial population of the multi-objective optimization algorithm based on the device modification suggestion. Specifically, modifying the constraints of the multi-objective optimization algorithm based on the device modification suggestion can involve expanding the numerical range of the constraints. For example, the upper limit of the parameter range of W_drift (drift layer width) can be expanded from 110 μm to 135 μm. In other embodiments, modifying the constraints of the multi-objective optimization algorithm based on the device modification suggestion can also involve expanding the categorical variables in the constraints. For example, new field stop layer design types (including single_gradient and double_gradient), new carrier lifetime control types (including none, uniform, and localized_collector), and a new gradient ratio (a conditional variable that only takes effect when the carrier lifetime control type field_stop_design ≠ uniform, with a range of (0.3, 0.7)). In one embodiment, an exemplary code for the optimization decision agent 40 to modify the constraints of the multi-objective optimization algorithm based on the device modification suggestions is as follows: # Expanding and optimizing variable space new_design_variables = { # ... Existing variables "W_drift": (110e-6, 135e-6), # The upper limit is extended from 110 to 135μm "field_stop_design": ["uniform", "single_gradient", "double_gradient"], # Add a categorical variable "lifetime_control_type": ["none", "uniform", "localized_collector"], "gradient_ratio": (0.3, 0.7) if field_stop_design != "uniform" elseNone } The optimization decision-making agent 40 modifies the reward function of the multi-objective optimization algorithm based on the device modification suggestions. This can be done by adding knowledge constraints, i.e., adding a knowledge-guided term to the original objective function. The modified reward function rewards designs that conform to historical success patterns, adopt recommended dual-gradient designs, and have parameter combinations close to historical successes, while penalizing designs that violate empirical rules. The modified reward function's reward can be obtained based on the sum of the base score and the knowledge bonus score. The formula for the base score is: base_score = alpha*Eoff + beta*(1 / SCWT); the original objective is to minimize Eoff and maximize SCWT. The knowledge bonus score for designs conforming to historical success patterns is knowledge_bonus = 0. The knowledge bonus score for designs using recommended dual-gradient designs is knowledge_bonus -= 0.1. The knowledge bonus score for parameter combinations close to historical successes is knowledge_bonus -= 0.05. The knowledge bonus score for designs that violate empirical rules is knowledge_bonus += 10.0. In one embodiment, the following is an example code showing how the optimization decision agent 40 modifies the reward function of the multi-objective optimization algorithm based on the device modification suggestions: # Add knowledge-guided items to the original objective function def enhanced_objective_function(design): # Original objective: Minimize Eoff, maximize SCWT base_score = alpha * Eoff + beta * (1 / SCWT) # Knowledge Enhancement: Rewards designed to align with historical success patterns knowledge_bonus = 0 # If the recommended dual-gradient design is adopted, a reward will be given. if design["field_stop_design"] == "double_gradient": `knowledge_bonus -= 0.1` # The negative sign indicates that the optimization algorithm aims to minimize this value, so the reward is a deduction. # Penalties will be imposed for violating the rules of experience. if design["W_drift"]>140e-6:# Violation of Fab rules knowledge_bonus += 10.0 # Major penalty # If the parameter combination is close to the historical success rate, a reward will be given. if is_close_to_successful_case(design, "IGBT_Case_2022_045"): knowledge_bonus -= 0.05 return base_score + knowledge_bonus The optimization decision-making agent 40 can modify the initial population of the multi-objective optimization algorithm based on the device modification suggestions by generating a "knowledge-guided" initial population. This means it's no longer completely randomized initialization, but rather based on historical success cases. For example, 30% of the individuals in the initial population can be derived from mutations of historical success cases, 40% can be randomly generated within the numerical range of the device modification suggestions, and 30% can be completely randomized (maintaining exploratory nature). In one embodiment, exemplary code for the optimization decision-making agent 40 modifying the initial population of the multi-objective optimization algorithm based on the device modification suggestions is as follows: # It's not completely random initialization, but based on historical success stories. initial_population = [] # 30% of individuals are variations of historical success stories. for case in top_3_similar_cases: for _ in range(10): individual = mutate_historical_design(case["parameters"]) initial_population.append(individual) # 40% of individuals are randomly generated within the knowledge recommendation subspace. for _ in range(40): individual = random_sample_within(knowledge_enhanced_suggestion["parameter_adjustment_range"]) initial_population.append(individual) # 30% of individuals are completely randomized (to maintain an exploratory approach) for _ in range(30): individual = completely_random_sample() initial_population.append(individual) The optimization decision-making agent 40 iteratively adjusts the target design parameters again based on the modified multi-objective optimization algorithm to obtain updated optimized design parameters for the device; these updated optimized design parameters are then added to the knowledge base. Example code for continuous monitoring by the AgentExecutor is as follows: monitoring_indicators = { "knowledge_utilization_rate": 0.75, # 75% of the optimized individuals adopted a knowledge recommendation structure. "improvement_vs_history": "The current best design (Eoff=1.88, SCWT=10.5) surpasses 85% of historical cases". "new_pareto_frontier": "A new design point not found in the historical database" "insight_generation": "New discovery: Eoff-SCWT achieves optimal tradeoff when gradient ratio = 0.55" } In this embodiment of the invention, after optimization, regardless of success or failure, the new design experience (updated and optimized design parameters) will be fed back into the knowledge base, achieving a closed-loop update of the knowledge base. An example code for adding the updated and optimized design parameters to the knowledge base is as follows: # New Knowledge Added to the Database new_knowledge_entry = { "problem_context": "IGBT 1200V Eoff-SCWT tradeoff optimization", "solution_attempted": "Dual gradient field cutoff layer + selective lifetime control", "result": { "achieved": {"Eoff": 1.88, "SCWT": 10.5, "Vce_sat": 2.05}, "key_learning": "The gradient ratio is optimal between 0.5 and 0.6; Vce_sat deteriorates significantly beyond 0.7." }, "generated_insights": [ "A 'gradient ratio optimal interval' has been discovered that was not mentioned in traditional textbooks." "The protective effect of local lifetime control on SCWT has been verified." ], "tags": ["IGBT", "tradeoff_breakthrough", "gradient_design", "new_insight"] } # Vectorize and store in database vector_db.add_documents( documents=[json.dumps(new_knowledge_entry)], metadatas=[{"type": "optimization_experience", "confidence": 0.9}] ) In other embodiments, error self-correction can also be achieved. If the simulation results of the simulation verification agent 30 are unsatisfactory, a backtracking mechanism is automatically triggered, returning to the structure generation agent 20 to generate updated design parameters.

[0076] This invention utilizes a knowledge base-based contextual retrieval enhancement mechanism. Through knowledge enhancement (RetrievalQA) and combined with a vector database (such as ChromaDB), it retrieves historical design cases, domain knowledge bases, and physical models and empirical formula libraries to assist the optimization decision agent 40 in iteratively adjusting the target design parameters, thereby achieving dynamic multi-objective optimization. By combining reinforcement learning and the domain knowledge base, it enables the search for the global optimal solution under complex constraints.

[0077] Therefore, this invention proposes a device automation design system based on a multi-agent framework, aiming to achieve three major breakthroughs through AI agent collaboration and intelligent orchestration technology: 1) end-to-end automation, utilizing toolchains such as LangChain to intelligently schedule the design process and eliminate gaps in manual intervention; 2) dynamic multi-objective optimization, combining reinforcement learning and domain knowledge bases to achieve global optimal solution search under complex constraints; 3) virtual-real fusion iteration, driving parameter self-correction through multi-physics simulation data, significantly reducing physical verification costs. By reconstructing the "human-machine collaboration" paradigm, it addresses the technical shortcomings of traditional methods, such as low efficiency, weak innovation, and high cost.

[0078] This invention provides an automated device design system based on a multi-agent framework. Breakthrough developments in large language models (LLMs) (such as Deepseek, GPT-4, and MetaGPT) have provided fundamental capabilities for automating complex tasks. The multi-agent framework, through role division, task decomposition, and collaboration, significantly enhances the engineering application potential of AI systems. In the field of device design, traditional methods rely on engineer experience, resulting in low efficiency, long iteration cycles, and difficulties in multi-objective optimization. This invention utilizes open-source frameworks such as LangChain to achieve efficient orchestration of LLMs through modular components (such as Agents, Chains, and Tools), supporting multi-tool collaboration and knowledge integration, thus providing a technological foundation for automated device design. This invention achieves automated device design that understands natural language intent, connects the entire process, and enables dynamic decision-making based on knowledge.

[0079] The device automated design system includes a processor and a memory. The aforementioned requirement analysis agent 10, structure generation agent 20, simulation verification agent 30, and optimization decision agent 40 are all stored as program units in the memory. The processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0080] A processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured.

[0081] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0082] Method Implementation Examples On the other hand, please refer to Figure 2 The present invention also provides an automated device design method, applied to the aforementioned automated device design system, the method comprising: Step 100: Use a large language model to analyze the device design requirements and obtain the target design parameters; Step 200: Generate the initial structure of the device based on the target design parameters using a parametric modeling tool; Step 300: Simulate the initial structure of the device using simulation software to obtain the performance indicators of the initial structure of the device; Step 400: If the performance indicators do not meet the set requirements, use a multi-objective optimization algorithm to iteratively adjust the target design parameters to obtain the optimized design parameters of the device. Steps 100, 200, 300, and 400 are scheduled and coordinated by the intelligent agent orchestration framework to form an automated design process.

[0083] The automated design flow formed by steps 100, 200, 300, and 400 in this embodiment of the invention is orchestrated using an intelligent agent orchestration framework. This achieves full-process automation, intelligently scheduling the design flow using the intelligent agent orchestration framework to eliminate gaps in manual intervention and improve the efficiency of automated device design. Furthermore, when performance indicators do not meet set requirements, this embodiment of the invention utilizes a multi-objective optimization algorithm to iteratively adjust the target design parameters. Through virtual-real fusion iteration and multi-physics simulation data-driven self-correction of device design parameters, the cost of physical verification is significantly reduced. Therefore, this embodiment of the invention solves the problems of low efficiency and high cost in traditional device design methods.

[0084] Optionally, the intelligent agent orchestration framework is the LangChain framework.

[0085] Optionally, the step of using a large language model to analyze the device design requirements and obtain the target design parameters includes: The design intent corresponding to the device design requirements is identified using the large language model. The design intent is associated with at least one of the following: domain knowledge graph, data triples, and process rule base, to obtain the requirement constraints corresponding to the design intent. The large language model is used to generate structured requirements based on the design intent and the requirement constraints; The target design parameters are generated based on the structured requirements using the large language model.

[0086] Optionally, generating target design parameters based on the structured requirements using the large language model includes: Receive the device target parameter requirements corresponding to the structured requirements; The large language model is used to generate device quantization requirements based on the device target parameter requirements and structured requirements; The target design parameters are generated based on the device quantization requirements using the large language model.

[0087] Optionally, the iterative adjustment of the target design parameters using a multi-objective optimization algorithm to obtain the optimized design parameters of the device includes: An initial population is generated based on the target design parameters; the initial population includes multiple individuals, each of which represents a set of device design parameters. Repeat the following steps until the set termination condition is met: Determine the performance metrics for each individual in the current population; A multi-objective optimization algorithm is used to perform population evolution on the current population to obtain an updated population; the updated population includes newly born individuals and some individuals in the current population.

[0088] Optionally, the method further includes: If the number of iterations reaches a set threshold and the performance indicators of the optimized design parameters do not meet the set requirements, a dilemma report is generated. Based on the aforementioned dilemma report, a search is conducted in the knowledge base to obtain device modification suggestions; the knowledge base includes at least one of the following: a historical design case library, a domain knowledge base, and a physical model and empirical formula library; Based on the device modification suggestions, at least one of the constraints, reward function, and initial population of the multi-objective optimization algorithm is modified. The target design parameters are iteratively adjusted again based on the modified multi-objective optimization algorithm to obtain the updated optimized design parameters of the device; the updated optimized design parameters are then added to the knowledge base.

[0089] Optionally, the current population is subjected to population evolution using a multi-objective optimization algorithm to obtain an updated population; the updated population includes newly born individuals and a portion of the individuals in the current population, including: Perform a non-dominated sort on each individual to determine the sorting result for each individual; Calculate the crowding degree for each individual to obtain the crowding degree calculation result for each individual; Genetic operations are performed on each individual based on the sorting results and the crowding calculation results to generate new individuals; The newly created individuals are merged with some individuals in the current population to obtain the updated population.

[0090] Optionally, determining the performance index of each individual in the current population includes: The structure-generating agent is used to generate the device structure corresponding to each individual in the current population. The simulation verification agent is used to simulate the device structure corresponding to each individual, and the performance index corresponding to each individual is obtained.

[0091] Equipment Implementation Examples Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3As shown, the electronic device may include: a processor 310, a communication interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communication interface 320, and the memory 330 communicate with each other through the communication bus 340. The processor 310 can call logic instructions in the memory 330 to execute an automated device design method, which includes: step 100, using a large language model to parse the device design requirements and obtain target design parameters; step 200, using a parametric modeling tool to generate an initial device structure based on the target design parameters; step 300, using simulation software to simulate the initial device structure and obtain the performance indicators of the initial device structure; step 400, if the performance indicators do not meet the set requirements, using a multi-objective optimization algorithm to iteratively adjust the target design parameters to obtain optimized design parameters for the device; wherein steps 100, 200, 300, and 400 are scheduled and coordinated by an intelligent agent orchestration framework to form an automated design process.

[0092] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0093] Program Product Examples On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a machine-readable storage medium. When the computer program is executed by a processor, the computer can execute an automated device design method. The method includes: step 100, using a large language model to parse the device design requirements and obtain target design parameters; step 200, using a parametric modeling tool to generate an initial device structure based on the target design parameters; step 300, using simulation software to simulate the initial device structure and obtain performance indicators of the initial device structure; step 400, if the performance indicators do not meet the set requirements, using a multi-objective optimization algorithm to iteratively adjust the target design parameters and obtain optimized design parameters for the device; wherein steps 100, 200, 300, and 400 are scheduled and coordinated by an intelligent agent orchestration framework to form an automated design process.

[0094] Storage Media Examples In another aspect, the present invention also provides a machine-readable storage medium storing a computer program thereon, which, when executed by a processor, implements an automated device design method. This method includes: step 100, using a large language model to parse the device design requirements and obtain target design parameters; step 200, using a parametric modeling tool to generate an initial device structure based on the target design parameters; step 300, using simulation software to simulate the initial device structure and obtain performance indicators of the initial device structure; and step 400, if the performance indicators do not meet the set requirements, using a multi-objective optimization algorithm to iteratively adjust the target design parameters to obtain optimized design parameters for the device. Steps 100, 200, 300, and 400 are scheduled and coordinated by an intelligent agent orchestration framework to form an automated design process.

[0095] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0096] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0097] 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 device automation design system, characterized by, include: A requirement parsing agent is used to parse the device design requirements using a large language model to obtain the target design parameters. A structure-generating intelligent agent is used to generate the initial structure of the device based on the target design parameters using a parametric modeling tool; A simulation verification agent is used to simulate the initial structure of the device using simulation software to obtain the performance indicators of the initial structure of the device. An optimization decision-making agent is used to iteratively adjust the target design parameters using a multi-objective optimization algorithm when the performance indicators do not meet the set requirements, so as to obtain the optimized design parameters of the device. The requirement analysis agent, the structure generation agent, the simulation verification agent, and the optimization decision agent work together based on an agent orchestration framework to form an automated design process.

2. The device automation design system of claim 1, wherein, The intelligent agent orchestration framework uses the LangChain framework.

3. The device automation design system of claim 1, wherein, The process of using a large language model to analyze the device design requirements and obtain the target design parameters includes: The design intent corresponding to the device design requirements is identified using the large language model. The design intent is associated with at least one of the following: domain knowledge graph, data triples, and process rule base, to obtain the requirement constraints corresponding to the design intent. The large language model is used to generate structured requirements based on the design intent and the requirement constraints; The target design parameters are generated based on the structured requirements using the large language model.

4. The device automation design system of claim 3, wherein, The process of generating target design parameters based on the structured requirements using the large language model includes: Receive the device target parameter requirements corresponding to the structured requirements; The large language model is used to generate device quantization requirements based on the device target parameter requirements and structured requirements; The target design parameters are generated based on the device quantization requirements using the large language model.

5. The automated device design system according to claim 1, characterized in that, The iterative adjustment of the target design parameters using a multi-objective optimization algorithm to obtain the optimized design parameters of the device includes: An initial population is generated based on the target design parameters; the initial population includes multiple individuals, each of which represents a set of device design parameters. Repeat the following steps until the set termination condition is met: Determine the performance metrics for each individual in the current population; A multi-objective optimization algorithm is used to perform population evolution on the current population to obtain an updated population; the updated population includes newly born individuals and some individuals in the current population.

6. The automated device design system according to claim 5, characterized in that, The optimization decision-making agent is also used for: If the number of iterations reaches a set threshold and the performance indicators of the optimized design parameters do not meet the set requirements, a dilemma report is generated. Based on the aforementioned dilemma report, a search is performed in the knowledge base to obtain device modification suggestions; the knowledge base includes at least one of the following: a historical design case library, a domain knowledge base, and a physical model and empirical formula library. Based on the device modification suggestions, at least one of the constraints, reward function, and initial population of the multi-objective optimization algorithm is modified. The target design parameters are iteratively adjusted again based on the modified multi-objective optimization algorithm to obtain the updated optimized design parameters of the device. Add the updated and optimized design parameters to the knowledge base.

7. The automated device design system according to claim 5, characterized in that, The current population is subjected to population evolution using a multi-objective optimization algorithm to obtain an updated population. The updated population includes newly created individuals and a portion of the current population, including: Perform a non-dominated sort on each individual to determine the sorting result for each individual; Calculate the crowding degree for each individual to obtain the crowding degree calculation result for each individual; Genetic operations are performed on each individual based on the sorting results and the crowding calculation results to generate new individuals; The newly created individuals are merged with some individuals in the current population to obtain the updated population.

8. The automated device design system according to claim 5, characterized in that, Determining the performance index of each individual in the current population includes: The structure-generating agent generates the device structure corresponding to each individual in the current population. The simulation verification agent is used to simulate the device structure corresponding to each individual, and the performance index corresponding to each individual is obtained.

9. An automated device design method, characterized in that, The method, applied to the automated design system for the device according to any one of claims 1 to 8, comprises: Step 100: Use a large language model to analyze the device design requirements and obtain the target design parameters; Step 200: Generate the initial structure of the device based on the target design parameters using a parametric modeling tool; Step 300: Simulate the initial structure of the device using simulation software to obtain the performance indicators of the initial structure of the device; Step 400: If the performance indicators do not meet the set requirements, use a multi-objective optimization algorithm to iteratively adjust the target design parameters to obtain the optimized design parameters of the device. Steps 100, 200, 300, and 400 are scheduled and coordinated by the intelligent agent orchestration framework to form an automated design process.

10. The automated device design method according to claim 9, characterized in that, The intelligent agent orchestration framework uses the LangChain framework.

11. The automated device design method according to claim 9, characterized in that, The process of using a large language model to analyze the device design requirements and obtain the target design parameters includes: The design intent corresponding to the device design requirements is identified using the large language model. The design intent is associated with at least one of the following: domain knowledge graph, data triples, and process rule base, to obtain the requirement constraints corresponding to the design intent. The large language model is used to generate structured requirements based on the design intent and the requirement constraints; The target design parameters are generated based on the structured requirements using the large language model.

12. The automated device design method according to claim 11, characterized in that, The process of generating target design parameters based on the structured requirements using the large language model includes: Receive the device target parameter requirements corresponding to the structured requirements; The large language model is used to generate device quantization requirements based on the device target parameter requirements and structured requirements; The target design parameters are generated based on the device quantization requirements using the large language model.

13. The automated device design method according to claim 9, characterized in that, The iterative adjustment of the target design parameters using a multi-objective optimization algorithm to obtain the optimized design parameters of the device includes: An initial population is generated based on the target design parameters; the initial population includes multiple individuals, each of which represents a set of device design parameters. Repeat the following steps until the set termination condition is met: Determine the performance metrics for each individual in the current population; A multi-objective optimization algorithm is used to perform population evolution on the current population to obtain an updated population; the updated population includes newly born individuals and some individuals in the current population.

14. The automated device design method according to claim 13, characterized in that, The method further includes: If the number of iterations reaches a set threshold and the performance indicators of the optimized design parameters do not meet the set requirements, a dilemma report is generated. Based on the aforementioned dilemma report, a search is performed in the knowledge base to obtain device modification suggestions; the knowledge base includes at least one of the following: a historical design case library, a domain knowledge base, and a physical model and empirical formula library. Based on the device modification suggestions, at least one of the constraints, reward function, and initial population of the multi-objective optimization algorithm is modified. The target design parameters are iteratively adjusted again based on the modified multi-objective optimization algorithm to obtain the updated optimized design parameters of the device. Add the updated and optimized design parameters to the knowledge base.

15. The automated device design method according to claim 13, characterized in that, The current population is subjected to population evolution using a multi-objective optimization algorithm to obtain an updated population. The updated population includes newly created individuals and a portion of the current population, including: Perform a non-dominated sort on each individual to determine the sorting result for each individual; Calculate the crowding degree for each individual to obtain the crowding degree calculation result for each individual; Genetic operations are performed on each individual based on the sorting results and the crowding calculation results to generate new individuals; The newly created individuals are merged with some individuals in the current population to obtain the updated population.

16. The automated device design method according to claim 13, characterized in that, Determining the performance index of each individual in the current population includes: The structure-generating agent generates the device structure corresponding to each individual in the current population. The simulation verification agent is used to simulate the device structure corresponding to each individual, and the performance index corresponding to each individual is obtained.

17. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the device automated design method according to any one of claims 1 to 8.

18. A machine-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the device automated design method according to any one of claims 1 to 8.

19. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the device automated design method according to any one of claims 1 to 8.

Images