An agent-based product design early-stage scheme generation and evaluation method

Abstract

The application discloses a product design early scheme generation and evaluation method based on an agent, which can call a suitable agent to speak based on a new message proposed by a user by using a speaker selection method, the agent can generate a corresponding tool call chain based on the new message proposed by the user, execute tools on the tool call chain based on the demand, thereby obtaining updated global information and an external file set, summarize the updated global information and the external file set by a large language model to obtain new global information and an external file set, and update a message pool, global information and an external file set to realize product design early scheme generation and evaluation. Therefore, the above method provided by the application can automatically match a suitable agent only by requiring the user to propose a demand, and the agent can directly generate product design early scheme generation and evaluation results based on the demand of the user.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method for generating and evaluating early-stage product design solutions based on intelligent agents. Background Technology

[0002] The development of large-scale models has enabled artificial intelligence to generate multiple views or 3D models from text, helping designers transform vague design ideas into high-quality images or 3D models in the early stages of design. However, existing 3D compositing methods suffer from drawbacks such as long compositing times, poor 3D consistency, and difficulty in fine-grained control and editing of the compositing results, making their application in product design challenging. The early stages of product design (conceptual design stage) play a decisive role in the product's performance, cost, and sustainability. However, in traditional design processes, engineers can only perform high-fidelity structural verification using Computer-Aided Design (CAD) and Computer-Aided Engineering (CAE) tools when the design enters the detailed modeling and simulation stage. If problems are discovered at this stage, rework and modifications often require significant costs. Existing computer-aided methods, such as morphological grammars, parametric modeling, and topology optimization, while capable of generating or evaluating design structures for local tasks, still struggle to achieve a complete closed-loop support from requirements analysis to solution conception and comprehensive evaluation.

[0003] In recent years, the development of generative artificial intelligence (GenAI), such as Large Language Models (LLM) and multimodal generative models, has provided new opportunities for design automation. These generative models can not only analyze designers' design needs and intentions, and efficiently generate creative concepts and multimodal visual design solutions; they can also evaluate and optimize design solutions based on existing multimodal design solutions. However, multimodal design solutions generated by GenAI often remain at the visual level, generally lacking physical and structural feasibility, and it is difficult to automatically incorporate engineering simulation feedback such as material simulation, stress simulation, and finite element analysis (FEA) into design evaluation and iteration.

[0004] Existing work has begun to input GenAI's generated results into downstream industrial software for physical simulation analysis to determine their accessibility. However, existing toolchains are often fragmented, with a lack of organic linkage between CAD, CAE, and the generated modules. Designers need to manually transfer requirements, set up simulations, and analyze results between different systems, resulting in slow feedback loops and low iteration efficiency. Summary of the Invention

[0005] This invention provides a method for generating and evaluating early-stage product design solutions based on intelligent agents. This method can efficiently convey user requirements, set up simulations, and analyze results, with high iteration efficiency.

[0006] This invention provides a method for generating and evaluating early-stage product design solutions based on intelligent agents, comprising:

[0007] Construct a multi-agent set, which includes multiple agents based on a large language model in sequence according to the design process. Each agent communicates and shares data through a unified message pool.

[0008] Based on the new messages input by the user into the message pool, a speaker is selected from the set of multiple agents using a speaker selection method, including: checking whether the current agent is executing a tool call process in the action flow; if so, specifying the current agent to continue speaking; otherwise, selecting the next agent to speak according to the design process.

[0009] The designated agent, based on the new message, uses a large language model to plan and generate a tool call chain and its corresponding input parameters. It executes the tools on the tool call chain and obtains the tool execution results, while updating the global information and external file set. Then, using the large language model, it generates reply information based on historical messages in the message pool, new messages input by the user, and tool execution results, while updating the message pool. The large language model then summarizes the updated global information and external file set to obtain new global information and external file set. By updating the message pool, global information, and external file set, the early-stage solutions for product design can be generated and evaluated.

[0010] Preferably, the multi-agent set includes, in sequence according to the design process, a requirement analysis agent, a concept generation agent, a simulation analysis agent, and a design evaluation agent;

[0011] The requirement parsing agent is used to parse user requirements based on received natural language input, invoke tools based on user requirements, and extract design goals, constraints and key parameters.

[0012] The concept generation agent is used to generate product concept design schemes by parsing design goals, constraints and key parameters, call tools and generate input parameters according to user design requirements, and execute the called tools to generate multimodal design schemes based on the input parameters.

[0013] The simulation analysis agent is used to decompose user messages to obtain design requirements, retrieve multimodal design schemes based on design requirements, and then provide simulation results through finite element analysis.

[0014] The design evaluation agent is used to evaluate multimodal design schemes, retrieve simulation results, design objectives and constraints from the message pool, evaluate the simulation results to obtain the overall evaluation result, and update the modeling parameters in the global information to provide the next design optimization scheme and parameter optimization scheme.

[0015] Preferably, the tools invoked by the demand parsing agent include image generation tools, image redrawing tools, or dialogue summarization tools;

[0016] The tools invoked by the concept-generating agent include one or more of the following: material selection tools, model selection tools, or 3D modeling tools;

[0017] The tools invoked by the simulation analysis agent include dialogue summarization tools and finite element analysis tools;

[0018] The tools used to design and evaluate the agent's calls include text-based dialogue tools, image-based dialogue tools, and call parameter update tools.

[0019] Preferably, the dialogue summary tool calls the language model API and uses specific prompt words to summarize design requirements and design solutions;

[0020] The image generation tool and image redrawing tool use Stable Diffusion XL Base to generate images and repair and redraw images based on prompts and reverse prompts;

[0021] The parameter update tool writes the list and dictionary data structures generated by the agent into a JSON file through the JSON overwrite interface, thereby realizing the initialization and updating of parameters;

[0022] The text-based dialogue tool and the image-based dialogue tool implement text-based question-and-answer with images by calling the Visual Language Model API.

[0023] Preferably, the material selection tool is used to convert the user's design requirements and the constraints of the dialogue context in the message pool into a structured query language through a large language model.

[0024] Based on a structured query language, candidate materials that meet the constraints are retrieved from the material database of the material selection tool. The candidate materials, the user's design requirements, and the constraints of the dialogue context are used to construct prompt words. Based on the prompt layer, the optimal material is selected from the candidate materials through a large language model, and an explanation of the selection is given.

[0025] Preferably, the 3D modeling tool is used to call the model category, structured parameters, and modeling script and JSON parameter file that match the model category and structured parameters in the current global memory, generate a 3D model based on the modeling script and JSON parameter file, and update the external file set at the same time.

[0026] Preferably, the finite element analysis tool is used to call SolidWorks, record the general simulation workflow through SolidWorks' macro recording function to obtain macro code, extract, reconstruct and parameterize the macro code, and define the key variables in the simulation workflow as input parameters to support reuse, thereby obtaining SolidWorks macros;

[0027] When invoked, the simulation analysis agent calls the win32com library to start and connect to the SolidWorks session, converts the SolidWorks macros into Python scripts, injects the values ​​of the specified key variables provided by the user into the Python scripts, executes the finite element analysis task, and inputs the simulation results into the simulation analysis agent.

[0028] Preferably, the intelligent agent includes a planning module, a behavior module, an attribute module, and a memory module;

[0029] The planning module is used to receive natural language instructions from the user, and based on the natural language instructions and the context association from the memory module, it parses the task objective, task constraints, tool call chain, and input and output parameters for tool calls through a large language model.

[0030] The behavior module is used to call external tools and perform operations based on the task objectives, task constraints and attribute information obtained from the attribute module, according to the toolchain planned by the planning module and its parameters.

[0031] The memory module is used to store current global information, messages from the message pool, material database, dialogue history information, and parameter model documents, thereby forming contextual associations.

[0032] The attribute module is used to define the role of the agent in the design process, the tasks undertaken by the role, and the tools that can be invoked to accomplish the tasks.

[0033] Preferably, the designated speaking agent plans and generates a tool call chain and its corresponding input parameters based on the new message through a large language model, including: the new message that the designated speaking agent will receive, prompts for step-by-step thinking in the form of a thought chain, the name of the available tool, a brief description of the tool's purpose and the tool parameters, and prompts for calling the tool. The large language model obtains the tool call chain and its corresponding input parameters.

[0034] Preferably, by setting specific prompt words, the large language model can summarize the updated global information and external file set to obtain new global information and external file set. The specific prompt words include: summarizing the tool call result information and interaction information in the message pool, and summarizing and obtaining new global information and external files based on the existing global information and external file set.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] This invention utilizes a speaker selection method to retrieve a suitable agent to speak based on a new message submitted by the user. This agent generates a corresponding tool call chain based on the new message, executes tools on the call chain according to the request, and obtains updated global information and external file sets. A large language model is then used to summarize the updated global information and external file sets to obtain new global information and external file sets. By updating the message pool, global information, and external file sets, early-stage product design solutions are generated and evaluated. Therefore, the method provided by this invention automatically matches a suitable agent simply by requiring the user to submit a request. This agent can directly generate and evaluate early-stage product design solutions based on the user's request. Attached Figure Description

[0037] Figure 1 A flowchart of a multi-agent set provided for a specific embodiment of the present invention;

[0038] Figure 2 A flowchart illustrating a material selection tool provided in a specific embodiment of the present invention;

[0039] Figure 3 A flowchart illustrating the 3D modeling tool provided in a specific embodiment of the present invention;

[0040] Figure 4 A flowchart of the finite element analysis tool provided in a specific embodiment of the present invention;

[0041] Figure 5 A flowchart of an intelligent agent provided for a specific embodiment of the present invention;

[0042] Figure 6 A flowchart illustrating an early-stage product design scheme generation and evaluation method based on intelligent agents, provided for a specific embodiment of the present invention. Detailed Implementation

[0043] This invention aims to explore an intelligent and automated process that integrates requirements analysis, solution generation, 3D modeling, and structural simulation in the early conceptual design stage of product design. The invention proposes a multi-agent collaborative framework centered on a large language model. Through task planning, external tool invocation, automated manipulation of industrial software, and simulation result feedback, it achieves a closed-loop iteration from design intent to feasible solutions, thus balancing the breadth of creativity with engineering feasibility at the conceptual stage and providing a new path for intelligent design.

[0044] To address the problems of insufficient engineering feasibility, fragmented design and simulation toolchains, and low efficiency of design feedback loops in the early conceptual design phase of existing product design, this invention proposes a multi-agent collaborative product concept design and evaluation method and system based on a large language model. This method constructs multiple agents with specific functions within a large language model framework, achieving a fully automated and intelligent closed-loop process from design requirement analysis, scheme generation, 3D modeling to structural simulation and evaluation. This invention mainly comprises three parts: 1) a multi-agent collaborative product concept design and evaluation system; 2) a single-agent construction method suitable for product concept design and evaluation; and 3) the construction of an external toolset suitable for product concept design and evaluation.

[0045] This invention provides a method for generating and evaluating early-stage product design solutions based on intelligent agents, comprising:

[0046] like Figure 1 As shown, a specific embodiment of the present invention constructs a multi-agent set, which includes multiple agents based on a large language model in sequence according to the design process, namely, a requirement analysis agent (Problem Framer), a concept generation agent (Concept Synthesizer), a simulation analysis agent (Simulation Analyst), and a design evaluation agent (Design Evaluator). Each agent communicates and shares data through a unified message pool.

[0047] In one specific embodiment, the Problem Framer provided in this embodiment is used to automatically parse user design requirements based on received natural language input, combined with contextual information in the message pool, and utilize its text understanding and generation capabilities. It performs semantic decomposition on natural language, calls image generation tools, image redrawing tools, and dialogue summarization tools as appropriate, and uses the large language model of the above tools to extract design goals, constraints, and key parameters.

[0048] In one specific embodiment, the concept-generating agent provided in this embodiment is used to generate a product concept design scheme by parsing the design goals, constraints, and key parameters obtained in the previous step, or to generate a concept design scheme diagram through image generation and image redrawing tools. Finally, based on the user's design requirements information, it calls one or more of the material selection tools, model selection tools, or 3D modeling tools, and generates input parameters. Based on the input parameters, it executes the called tools to generate a complete multimodal design scheme for the product concept stage. The complete multimodal design scheme for the product concept stage includes a product creative concept design scheme, a design scheme in text form, a 2D product effect image (obtained through an image generation tool), and a 3D model.

[0049] In one specific embodiment, the simulation analysis agent provided in this embodiment is used to decompose user messages to obtain design requirements through dialogue summarization tools, retrieve information on design schemes in the product concept stage from upstream agents based on design requirements, i.e., multimodal design schemes, and provide simulation results based on the information on design schemes in the product concept stage through the invoked finite element analysis tools, as well as analysis and interpretation of the simulation results.

[0050] In one specific embodiment, the Design Evaluator provided in this embodiment is used to intelligently evaluate multimodal design schemes by calling text dialogue tools and image dialogue tools. The text dialogue tools and image dialogue tools provided in this embodiment are both backed by language models. The evaluation is carried out using a prompt based on a language model and a prompt language model. The evaluation agent retrieves simulation results, design goals, and constraints from the upstream simulation analysis agent and requirement parsing agent in the message pool, evaluates the simulation results to obtain the overall evaluation result, determines whether the design goals and constraints are met, and updates the modeling parameters in the global information by calling the parameter update tool to provide the next design optimization scheme and parameter optimization scheme.

[0051] The design evaluation agent provided in this embodiment of the invention is responsible for the overall evaluation of the design scheme, receiving the simulation evaluation results, comprehensively analyzing the design quality, constraint satisfaction conditions and other design indicators of the current product concept, and providing the next design optimization scheme and parameter optimization scheme.

[0052] This invention specifically addresses a method for multi-agent collaborative product concept design and evaluation based on a large language model. It constructs an external toolset composed of multiple sub-tools to automatically execute key tasks such as material selection, geometric modeling, and finite element simulation analysis during the design process through the invocation of these agents. These tools work collaboratively under the task planning and data coordination of the large language model, achieving a fully automated design closed loop from requirements analysis to simulation feedback. This invention specifically designs 11 tools, whose names and functional descriptions are shown in Table 1. The three most core tools are the material selection tool, the parametric modeling tool, and the finite element analysis tool, which correspond to the three key stages of material simulation, geometric modeling, and simulation analysis, respectively.

[0053] Table 1 lists the names and functional descriptions of 11 tools designed in specific embodiments of the present invention.

[0054]

[0055] The dialogue summarization tool called by the intelligent agent in the specific embodiment of the present invention calls the language model API and uses specific prompt words to summarize the design requirements and design schemes.

[0056] The image generation tool and image redrawing tool invoked by the intelligent agent in the specific embodiments of the present invention generate images and repair and redraw images by using Stable Diffusion XL Base1.0 based on prompt words and reverse prompt words.

[0057] The parameter update tool called by the agent in the specific embodiment of the present invention writes the list and dictionary data structure generated by the agent into a JSON file through the JSON overwrite interface, thereby realizing the initialization and updating of parameters.

[0058] The text dialogue tool and image dialogue tool invoked by the intelligent agent in the specific embodiments of the present invention realize text question answering with images by calling the visual language model API, so as to meet the evaluation requirements of multimodal design schemes.

[0059] The model selection tool provided in this specific embodiment of the invention utilizes an LLM guided by prompt words to select the category of a parametric model based on a pre-built model document and input user requirements.

[0060] The material selection tool provided in this specific embodiment of the invention is used to combine the user's design requirements with the constraints of the dialogue context in the message pool. Through a large language model, it transforms the user's design requirements and the constraints of the dialogue context into a structured query language, such as... Figure 2As shown, in one specific embodiment, the user's requirement is "I want the landing gear to be made of lightweight alloy." In another specific embodiment, the prompt for converting this into a structured query language is: "You are an SQL statement creation expert. You need to create an SQL statement based on my requirements. I have now created a table named 'materials' with the structure: ... Return only the SQL query and explanation, no additional answers. Now my request is {req}, please give me the SQL query." Based on this prompt, a structured query statement is obtained through a large language model.

[0061] This invention provides a method for retrieving candidate materials that meet constraints from a material selection tool's material database based on structured query statements. It constructs prompts by combining the candidate materials, the user's design needs, and the constraints of the dialogue context. Based on these prompts, a large language model is used to select the optimal material from the candidate materials and provide an explanation for the selection. The prompts used are: "You are a material selection expert. According to my requirements, you need to select the most suitable material ID from the preliminarily selected materials. My current requirement is {req}. Please select a material ID and provide an explanation, including your reasons for choosing this material."

[0062] like Figure 3 and Figure 4 As shown in the illustration, the 3D modeling tool provided in this specific embodiment of the invention is used to call the model category, structured parameters, and modeling scripts and JSON parameter files that match the model category and structured parameters in the current global memory. Based on the modeling scripts and JSON parameter files, a 3D model is generated, and the external file set is updated simultaneously. The JSON parameter file provided by this invention is a value modeling parameter; different 3D models can be obtained by modifying the modeling parameters. Initially, initial parameters are used, which are predefined. The agent can then continuously modify these parameters to achieve modification and iteration of the 3D model.

[0063] The finite element analysis tool provided in the specific embodiment of the present invention is used to call SolidWorks, record the general simulation workflow through the macro recording function of SolidWorks to obtain macro code, extract, reconstruct and parameterize the macro code, and define the key variables in the simulation workflow as input parameters to support reuse, thereby obtaining the SolidWorks macro;

[0064] When invoked, the simulation analysis agent calls the win32com library to start and connect to the SolidWorks session, converts SolidWorks macros into Python scripts, injects the values ​​of user-provided key variables into the Python scripts, executes the finite element analysis task, and inputs the simulation results into the simulation analysis agent.

[0065] The intelligent agent provided in the specific embodiments of the present invention, such as Figure 5 As shown, each module includes a planning module, a behavior module, an attribute module, and a memory module.

[0066] The planning module provided in a specific embodiment of the present invention is used to receive natural language instructions input from user interaction, and decompose the top-level task through a large language model based on the natural language instructions and the context association from the memory module, and infer the task objective, task constraints, tool call chain and input / output parameters for tool call, which include the thought chain.

[0067] The behavior module provided in this specific embodiment of the invention is used to complete termination checks, call external tools and execute operations according to the toolchain planned by the planning module and its parameters, generate responses and summarize text, based on task objectives, task constraints and attribute information obtained from the attribute module. This module also defines a tool library, including some Python scripts, AI model APIs, online APIs and directly callable functions.

[0068] The memory module provided in this specific embodiment of the invention is used to store short-term memory and long-term memory. Short-term memory includes current global information, agent context cache, and messages from the shared chat message pool; long-term memory includes a material database (used for material selection; without this material database, the language model may fabricate materials and their numerical properties), the dialogue history between the agent and the user, and a parameter model document. This parameter model document allows the language model to know the structure corresponding to each parameter in our parameter model, thereby generating correct parameters and forming contextual associations.

[0069] The attribute module provided in this specific embodiment of the invention is used to define the role of the intelligent agent in the design process, the tasks and requirements undertaken by the role, the tools that can be invoked to achieve the tasks, and a demonstration case.

[0070] like Figure 6 As shown, in a specific embodiment of the present invention, based on a new message input by the user into the message pool, such as "How is the stress performance of this product?", a speaker selection method is used to select a speaking agent from the multi-agent set to speak at time t+1. This includes: checking whether the current agent is currently executing a tool call process in the action flow; if so, specifying the current agent to continue speaking; otherwise, selecting the next agent to speak according to the designed process. The agent's action flow includes: termination check, toolchain call, reply, and summary.

[0071] The designated agent, based on new messages, plans and generates tool call chains and their corresponding input parameters through a large language model. It executes the tools on the tool call chain and obtains the tool execution results. This tool call chain is a set of work steps obtained by the agent through planning, which can be executed directly. At the same time, it updates global information and external file sets. Then, it generates reply information based on historical messages in the message pool, new messages input by the user, and tool execution results through the large language model. It also updates the message pool and summarizes the updated global information and external file sets through the large language model to obtain new global information and external file sets. By updating the message pool, global information, and external file sets, it achieves the generation and evaluation of early-stage product design solutions.

[0072] In one specific embodiment, at time State using quadruples It indicates. Among them. Indicates time t The agent that is speaking. Represents the set of intelligent agents in the system; Indicates in t The dialogue records in the information pool at any given time, among which They represent in Messages from a collection of users and agents at any given moment; Indicates in t The global memory information at any given moment, which is a global variable stored in the program; This represents the set of external files at time t, such as 3D models, 2D images, and databases.

[0073] When the user is at time Enter new message The message pool management system will utilize the speaker selection method, namely the speaker selection function. From the collection of intelligent agents This function selects the next agent to speak. Its working principle is to check if the current agent is currently executing a tool call process within its workflow. If so, it designates the next agent to speak based on the sequence rules within the design flow. The update formula is as follows: .

[0074] Once the agent to speak is designated, the agent's LLM will plan and generate a tool call chain and the corresponding parameters. .in These represent the tools to be invoked, which are external functions that the agent can call, along with their input parameters. The agent executes these tools sequentially and receives feedback, obtaining the tool execution result at that moment. Meanwhile, during the execution of the tool, global information and external file collection It is also constantly being updated. This update process can be expressed as a formula:

[0075]

[0076]

[0077] Ultimately, the AI ​​system generates a response based on historical information, tool execution results, and the user's message at that moment. At the same time, update the message pool. A tool that uses a large language model with specific cue words to summarize the context. Update global memory and external files After the above process, the system is updated to real-time. status This state represents the design solution required by the designer, and the designer's design solution will be continuously updated and iterated in the system. Global information includes design solutions, design problems, and design evaluations. The final form of this information needs to be summarized in conjunction with tool call results, external files, etc. Directly using update will result in inconsistencies in context information.

Claims

1. A method for generating and evaluating early-stage product design solutions based on intelligent agents, characterized in that, include: Construct a multi-agent set, which includes multiple agents based on a large language model in sequence according to the design process. Each agent communicates and shares data through a unified message pool. Based on the new messages input by the user into the message pool, a speaker is selected from the set of multiple agents using a speaker selection method, including: checking whether the current agent is executing a tool call process in the action flow; if so, specifying the current agent to continue speaking; otherwise, selecting the next agent to speak according to the design process. The designated agent, based on the new message, plans and generates a tool call chain and its corresponding input parameters through a large language model. It executes the tools on the tool call chain and obtains the tool execution results. At the same time, it updates the global information and the external file set. Then, it generates reply information based on the historical messages in the message pool, the new messages input by the user, and the tool execution results through the large language model. It also updates the message pool and summarizes the updated global information and external file set through the large language model to obtain new global information and external file set. By updating the message pool, global information, and external file set, the early-stage solutions for product design can be generated and evaluated. The designated speaking agent plans and generates a tool call chain and its corresponding input parameters based on the new message through a large language model. This includes: the new message that the designated speaking agent will receive, prompts for step-by-step thinking in the form of a thought chain, the names of available tools, a brief description of the tool's purpose and tool parameters, and prompts for tool calls. The large language model then obtains the tool call chain and its corresponding input parameters. The multi-agent set, in accordance with the design process, includes, in sequence, a requirement analysis agent, a concept generation agent, a simulation analysis agent, and a design evaluation agent; The requirement parsing agent is used to parse user requirements based on received natural language input, invoke tools based on user requirements, and extract design goals, constraints and key parameters. The concept generation agent is used to generate product concept design schemes by parsing design goals, constraints and key parameters, call tools and generate input parameters according to user design requirements, and execute the called tools to generate multimodal design schemes based on the input parameters. The simulation analysis agent is used to decompose user messages to obtain design requirements, retrieve multimodal design schemes based on design requirements, and then provide simulation results through finite element analysis. The design evaluation agent is used to evaluate multimodal design schemes, retrieve simulation results, design objectives and constraints from the message pool, evaluate the simulation results to obtain the overall evaluation result, and update the modeling parameters in the global information to provide the next design optimization scheme and parameter optimization scheme.

2. The method for generating and evaluating early-stage product design solutions based on intelligent agents according to claim 1, characterized in that, The tools invoked by the demand parsing agent include image generation tools, image redrawing tools, or dialogue summarization tools. The tools invoked by the concept-generating agent include one or more of the following: material selection tools, model selection tools, or 3D modeling tools; The tools invoked by the simulation analysis agent include dialogue summarization tools and finite element analysis tools; The tools used to design and evaluate the agent's calls include text-based dialogue tools, image-based dialogue tools, and call parameter update tools.

3. The method for generating and evaluating early-stage product design solutions based on intelligent agents according to claim 2, characterized in that, The dialogue summary tool calls the language model API and uses specific prompt words to summarize design requirements and design solutions. The image generation and redrawing tools utilize Stable Diffusion XL Base to generate, repair, and redraw images based on prompts and reverse prompts. The parameter update tool writes the list and dictionary data structures generated by the agent into a JSON file through the JSON overwrite interface, thereby realizing the initialization and updating of parameters; The text-based dialogue tool and the image-based dialogue tool implement text-based question-and-answer with images by calling the Visual Language Model API.

4. The method for generating and evaluating early-stage product design solutions based on intelligent agents according to claim 2, characterized in that, The material selection tool is used to combine the user's design requirements with the constraints of the dialogue context in the message pool, and then transforms the user's design requirements and the constraints of the dialogue context into a structured query language through a large language model. Based on a structured query language, candidate materials that meet the constraints are retrieved from the material database of the material selection tool. The candidate materials, the user's design requirements, and the constraints of the dialogue context are used to construct prompt words. Based on the prompt words, a large language model is used to select the optimal material from the candidate materials and provide an explanation of the selection.

5. The method for generating and evaluating early-stage product design solutions based on intelligent agents according to claim 2, characterized in that, The 3D modeling tool is used to call the model category, structured parameters, and modeling scripts and JSON parameter files that match the model category and structured parameters in the current global memory, generate a 3D model based on the modeling scripts and JSON parameter files, and update the external file set at the same time.

6. The method for generating and evaluating early-stage product design solutions based on intelligent agents according to claim 2, characterized in that, The finite element analysis tool is used to call SolidWorks, record the general simulation workflow through SolidWorks' macro recording function to obtain macro code, extract, reconstruct and parameterize the macro code, and define the key variables in the simulation workflow as input parameters to support reuse, thus obtaining SolidWorks macros. When invoked, the simulation analysis agent calls the win32com library to start and connect to the SolidWorks session, converts SolidWorks macros into Python scripts, injects the values ​​of user-provided key variables into the Python scripts, executes the finite element analysis task, and inputs the simulation results into the simulation analysis agent.

7. The method for generating and evaluating early-stage product design solutions based on intelligent agents according to claim 1, characterized in that, The intelligent agent includes a planning module, a behavior module, an attribute module, and a memory module; The planning module is used to receive natural language instructions from the user, and based on the natural language instructions and the context association from the memory module, it parses the task objective, task constraints, tool call chain, and input and output parameters for tool calls through a large language model. The behavior module is used to call external tools and perform operations based on the task objectives, task constraints and attribute information obtained from the attribute module, according to the toolchain planned by the planning module and its parameters. The memory module is used to store current global information, messages from the message pool, material database, dialogue history information, and parameter model documents, thereby forming contextual associations. The attribute module is used to define the role of the agent in the design process, the tasks undertaken by the role, and the tools that can be invoked to accomplish the tasks.

8. The method for generating and evaluating early-stage product design solutions based on intelligent agents according to claim 1, characterized in that, By setting specific prompt words, the large language model can summarize the updated global information and external file set to obtain new global information and external file set. The specific prompt words include: summarizing the tool call result information and interaction information in the message pool, and summarizing and obtaining new global information and external files based on the existing global information and external file set.

Images