Multi-agent computer-aided engineering and computer-aided design method

Abstract

The invention discloses a multi-agent computer-aided engineering and a computer-aided design method. CAE simulation and design optimization for the design of a complex product are high in threshold and difficulty. According to the method, autonomous planning and automatic execution of CAE simulation and CAD design optimization tasks are realized by using a multi-agent technology based on a large language model. A user interacts with the multi-agent system in the modes of text input, voice, a graphical user interface and the like, the multi-agent system autonomously plans workflow steps according to the requirements of the user, all agents cooperatively execute the working steps in a labor division mode, CAE and CAD software is called to execute simulation and design change, CAE and CAD result data are read and analyzed, the task completion condition is evaluated, and the task completion condition is evaluated. And the task can be re-planned if necessary. The system can automatically process error information, and can notify a user to process when user intervention is needed. The performance of the system can be improved through fine tuning of a large language model and retrieval enhancement generation technology.

Description

Technical Field

[0001] The present invention relates to computer-aided engineering and computer-aided design, and more particularly to a multi-agent computer-aided engineering and computer-aided design method. Background Art

[0002] Computer-aided engineering (CAE) is a technology that uses computer software for engineering simulation analysis and verifies and optimizes product design and performance based on the analysis results. Computer-aided design (CAD) is a technology that uses computer software to create accurate two-dimensional or three-dimensional models for product design and drawing. CAE software and CAD software are indispensable analysis and design tools in modern product design and engineering. The model created in CAD software is a basic input for CAE analysis, and the analysis results of CAE are used to verify the design performance of the product, discover potential design problems, and guide the improvement and optimization of CAD design. In complex product and engineering designs, it is usually necessary to perform multiple loop iterations between CAD software and CAE software to complete the design optimization.

[0003] With the increasing complexity of modern product design and the continuous improvement of performance and reliability requirements, the functions and integration degrees of CAD and CAE software are also continuously improving, the software operation is becoming more and more complex, and the usage threshold is getting higher and higher. Especially in the field of CAE, due to the increasingly complex working environment of products, the simulation analysis of product design performance often involves the interaction and coupling of multiple disciplines and complex physical models, which poses high requirements on the theoretical knowledge, engineering experience, and software operation ability of software users. The cultivation of high-end talents is difficult and time-consuming. There is still a large gap in high-end simulation analysis and design talents in China, which restricts the application and popularization of CAE and CAD technologies.

[0004] In recent years, artificial intelligence large language models have achieved breakthrough development. These large language models integrate a vast amount of knowledge, perform excellently in tasks such as text generation, question answering, translation, and code writing, and have certain logical reasoning capabilities. This provides conditions and opportunities for the development of CAE and CAD technologies towards intelligence, automation, and collaboration.

[0005] Currently, the applications of artificial intelligence large language models and their related technologies in the fields of CAE and CAD are basically some simple application scenarios. For example, using large language models to develop software operation wizards, or directly generating program codes for CAE or CAD software. However, these program codes can rarely run directly and require manual modification, and cannot be directly applied to actual product design analysis. There has been no publicly released intelligent agent (Agent, or also known as intelligent proxy) product based on large language models that can automate the CAE simulation analysis and CAD design optimization processes through natural language human-computer interaction.

[0006] To solve the above problems, the present invention proposes a computer-aided engineering and computer-aided design method for a multi-agent system (MAS) based on an artificial intelligence large language model. The user conducts dialogue interaction with the multi-agent system, puts forward the purposes and requirements of product design and simulation analysis. Multiple agents in the multi-agent system cooperate with each other in division of labor, autonomously plan and automatically execute tasks, call CAE and CAD software for simulation analysis and design optimization, and realize the intelligence and automation of product design simulation analysis and optimization. Summary of the Invention

[0007] Object of the Invention: To realize the intelligence and automation of CAE simulation analysis and CAD design optimization by using agent technology based on an artificial intelligence large language model, lower the usage threshold of CAE and CAD technologies, and improve work efficiency. The user conducts dialogue interaction with the multi-agent system, puts forward design purposes and simulation analysis requirements and goals. The multi-agent system autonomously plans and automatically executes the work processes and steps, and completes the simulation analysis and design optimization tasks. A system composed of multiple agents can autonomously plan the work processes and steps required to complete the tasks. Each agent cooperates with each other in division of labor according to its own design function, calls the corresponding CAE and CAD software to complete the simulation analysis and design optimization tasks, reads the output data and information of the CAE and CAD software and conducts analysis, evaluates whether the task goals are achieved and ends the task. When necessary, the multi-agent system can re-plan the work content and steps. The multi-agent system can automatically handle errors and require the user to intervene for handling when the errors or problems encountered exceed the system's automatic processing ability.

[0008] Technical solution: All types of CAE simulation analysis methods have strict processes and steps, and complex CAD designs are assembled from simple geometric bodies according to certain rules. In view of these characteristics of CAE and CAD, by leveraging the reasoning and generation capabilities of artificial intelligence large language models, combined with the enterprise's exclusive knowledge database and Retrieval-Augmented Generation (RAG) technology, an intelligent agent that can autonomously plan, decompose, and schedule CAE and CAD tasks, and intelligent agents that execute various specialized tasks are developed to form a multi-agent system, which autonomously plans and automatically executes CAE simulation analysis and CAD optimization design tasks according to the user's requirements. The user interacts with the multi-agent system through dialogue, putting forward the goals and requirements of product design and simulation analysis. The multi-agent system autonomously conducts reasoning, makes decisions, formulates, and coordinates each intelligent agent to automatically execute the work plan and steps. The intelligent agents developed for CAE and CAD tasks can use large language models to generate program codes executable by CAE software or CAD software, and call the corresponding CAE or CAD software to run the program codes for simulation analysis and design optimization. The multi-agent system can read the running status and output information and data of CAE and CAD software, analyze and evaluate the task execution and completion situation, re-plan the remaining work when necessary, automatically handle errors and problems, or request the user to intervene when it cannot handle them automatically, so as to realize the cyclic iteration of automatically conducting simulation analysis and design optimization until the goals and requirements put forward by the user are achieved. The organizational structure of the multi-agent system can be designed according to the complexity of the task. In a multi-agent system with a centralized structure, there is a supervisor intelligent agent and multiple worker intelligent agents. The supervisor intelligent agent, as the central node of the system, is responsible for dispatching and coordinating specific work according to the functions of each intelligent agent, monitoring and evaluating the work execution situation, correcting errors, and managing the communication between each intelligent agent. The worker intelligent agents are designed to automatically complete various specific tasks. For more complex tasks, the multi-agent system can also be designed into a hierarchical structure. In this structure, the supervisor intelligent agent, as the central node, manages multiple subordinate intelligent agents, and the subordinate intelligent agents can have their own subordinate intelligent agents, forming a tree-like hierarchy. The tasks are decomposed and refined layer by layer, with the superior managing the subordinate and the subordinate executing specific work. Separate multi-agent subsystems that can operate independently can also be developed according to the characteristics of different CAE software or CAD software. Multiple multi-agent subsystems cooperate with each other to complete complex multi-physical layer coupling simulation analysis and design optimization tasks that require calling multiple CAE software and CAD software in different professional fields. This structure can be compatible with heterogeneous CAE and CAD software systems, with better flexibility and scalability. The specific steps are as follows: (1)Configure software and hardware environment: Software configuration: operating system, programming language environment, deep learning framework, GPU support tools, model loading and inference tools, large language model fine-tuning tools, vector database tools, other auxiliary tools and libraries; Hardware configuration: If local configuration is required, the corresponding GPU video memory, memory capacity and CPU core number need to be configured according to the parameter scale of the selected large model. Cloud service solutions can also be considered and rented on demand; (2)Select and fine-tune large language models: Select open-source large language models that support inference and tool calls. When the inference and generation capabilities of large language models in answering professional questions in the CAE and CAD fields cannot meet the needs, fine-tune the large language models with professional knowledge data in the CAE and CAD fields to improve the accuracy and reliability of model inference and generation in these professional fields; (3)Establish an enterprise-specific knowledge base: When an enterprise has special requirements for product design and simulation analysis, the enterprise can use its own proprietary documents, data or program codes related to product design and simulation analysis standards and specifications to establish a vector database, and use retrieval-augmented generation (RAG) technology to assist large language models in generating more accurate, reliable and targeted answers. Common open-source vector database tools in the Python language include DeepLake, FAISS, Chroma, LlamaIndex, etc.; (4) Develop a multi-agent system: The main development tasks and steps include: Based on the actual needs of the enterprise, as well as the CAE and CAD software systems used, product design, simulation analysis methods, standards, and specifications, etc., clarify the development goals, functions, performance, and interaction methods of the multi-agent system; Select a suitable organizational structure for the multi-agent system, such as a centralized structure (one supervisor agent is responsible for managing and monitoring multiple worker agents), a hierarchical structure (a tree-like hierarchical structure where the upper level manages the lower level and the lower level executes specific tasks), a subsystem structure (composed of multiple multi-agent subsystems, and the subsystems can operate independently or cooperate with each other), or a hybrid structure, etc.; Determine the implementation methods for communication and interaction between agents, such as using shared memory, message passing, etc.; Determine the design scheme, decompose tasks, and define the roles of agents, such as analysis, planning, decision-making, execution agents, etc.; Select a suitable development tool framework for multi-agent system development and programming. The main functions of the multi-agent system include: The interaction methods between users and the multi-agent system, such as inputting natural language instructions, voice, graphical user interface (GUI), reading text files written by users in advance, or multi-modal interaction methods, etc.; The supervisor agent, as the central node of the system, manages and monitors the division of labor and cooperation of each worker agent, autonomously plans and automatically executes work processes and steps; Automatically handle errors, and when problems or errors beyond the system's automatic processing ability occur, require users to intervene for processing; The agents designed for CAD and CAE tasks can use large language models to generate program codes executable by CAE and CAD software, and call CAE and CAD software to run the program codes to automatically complete simulation analysis and design optimization tasks; Read the running status and output information and data of CAE and CAD software, evaluate the task completion situation, and re-plan the task if necessary; When necessary, first retrieve the vector database and use RAG technology to assist the large language model to generate more accurate, reliable, and targeted answers. After the development of the agent system is completed, test the system functions and make adjustments and optimizations if necessary; (5) Deploy and apply the multi-agent system: Deploy the multi-agent system to the actual application environment, and conduct continuous monitoring and maintenance. Improve the system according to the actual application situation and problems that occur.

[0009] Beneficial effects: (1) By leveraging the inference and generation capabilities of artificial intelligence large language models, combined with the knowledge base of the enterprise's unique product design, analysis, and evaluation standard specifications and RAG technology, the multi-agent system autonomously plans and automatically completes product design and simulation optimization tasks according to the user's requirements; (2) Engineers only need to interact with the agent system using simple operations such as natural language, voice, or GUI, reducing the threshold for using CAE and CAD technologies; (3) Engineers do not have to switch between multiple CAE and CAD software and perform data exchange operations. The agent system automatically calls each software according to the plan to complete the corresponding tasks, improving work efficiency; (4) Frees engineers from delving into complex software operations and theoretical background knowledge, allowing them to focus on product design and innovation; (5) The multi-agent system is easy to expand and maintain. Description of the Drawings

[0010] Figure 1 is the method flowchart of the present invention. Embodiment

[0011] The technical solution of the present invention will be further described below in conjunction with the embodiments. A multi-agent computer-aided engineering and computer-aided design method described in this embodiment includes the steps: (1) Configure the software and hardware environment: Taking the example of fine-tuning a large model with 70 billion (70B) parameters using cloud services and then locally deploying the fine-tuned large model, using cloud services to fine-tune the large model can greatly save the cost of purchasing hardware locally. Select a cloud service provider and purchase the instance type (hardware resource configuration) recommended by the cloud service provider according to the model parameter scale. Generally, an instance type with a configuration of 8×H100 80GB or 8×A100 80GB GPU is required. Install the libraries required for fine-tuning the model: python, transformers, peft, torch, datasets, tokenizers, etc. The hardware configuration required for locally deploying a large model with 70 billion parameters: The recommended configuration for FP16 precision is as follows: GPU: 2×A100 80GB, Memory: 512GB DDR4, Storage: 2TB NVMe SSD, CPU: Above 64 cores. Taking the Ollama tool as an example for locally deploying the large model, Ollama needs to be installed locally. Taking the DeepLake tool as an example for locally deploying the vector database, deeplake needs to be installed; (2)Select and fine-tune the large language model: Select an open-source large language model that supports reasoning and tool calls. Taking the llama-3.1-70b model as an example, the large language model can be fine-tuned with data on professional knowledge in the CAE and CAD fields to improve the accuracy and reliability of the model's reasoning and generation in these professional fields. The main steps include: loading the pre-trained model, loading the dataset, data processing, configuring PEFT, configuring training parameters, starting training, saving the model, and model evaluation. After the large language model is fine-tuned, download the model file from the cloud service to the local. Use Ollama to create and run the model to deploy the model locally; (3)Establish an enterprise-specific knowledge base: Enterprises can use their own unique documents such as standards and specifications for product design and simulation analysis to establish an exclusive knowledge vector database. Taking the creation of a local vector database using LangChain and DeepLake as an example, the main steps include: loading local knowledge files (supported formats: PDF, txt, Word, Markdown), text chunking, selecting an embedding model (i.e., the fine-tuned large model deployed locally), creating a DeepLake vector database, adjusting the chunk size to obtain the best retrieval effect. Then, according to the user's question or input, relevant information can be retrieved from the local vector library first, and the retrieved information can be used as additional context or knowledge to input to the large model to assist the large model in generating more accurate and targeted answers; (4) Developing a multi-agent system: The multi-agent system for simulation analysis and design optimization is a system with a clear task process and mainly relies on the capabilities of large language models. It is a pure software agent system that does not require real-time interaction and belongs to a system of medium to low complexity. The LangChain and LangGraph frameworks based on the Python language can be selected for programming to develop the agent system. This combined framework is good at natural language processing, process orchestration, text and code generation, and has a rich external toolset that can be utilized. Agents transfer data by sharing the state of the graph and messages by the edges of the graph. When user intervention is required, the system can pause execution, wait for the user's confirmation or feedback, and then continue execution based on the user's input. Taking the design optimization of a mechanical part as an example, the enterprise's design specification standards require that the part meet the requirements of mechanical strength and vibration frequency. The CAD software and CAE software adopted by the enterprise both provide Python APIs, and Python scripts can be used to automate modeling and simulation analysis. The enterprise's design specification standards, historical design simulation cases, and relevant technical data have been established in a vector database for RAG. Develop a multi-agent system using the LangChain and LangGraph frameworks based on the Python language. The multi-agent system is designed as a centralized structure containing 7 agents, including 1 supervisor agent and 6 worker agents. Among them, the supervisor agent, as the central node of the system, is responsible for dispatching and coordinating specific tasks according to the functions of each agent, monitoring and evaluating the execution of tasks, correcting errors, and managing the communication between agents. The supervisor agent also serves as the interface for interacting with users. The 6 worker agents are respectively: task planning agent, information retrieval agent, geometric modeling agent, strength analysis agent, dynamics analysis agent, and report generation agent. LangGraph represents the structure of multi-agents through a graph (Graph). Each agent is a node in the graph. Agents transfer data by sharing the state of the graph and messages by the edges of the graph.The tasks and workflows executed by each agent mainly include: (1) The user interacts with the supervisor agent that also serves as an interaction interface, putting forward the goals and requirements of the design and analysis tasks. The supervisor agent, as the central node of the system, is responsible for dispatching and coordinating specific work according to the functions of each agent, monitoring and evaluating the work execution situation, correcting errors, and managing the communication between each agent, etc.; (2) The information retrieval agent is responsible for retrieving the enterprise-specific knowledge vector database, evaluating and self-correcting the retrieved information to ensure accurate and reliable relevant information is retrieved from the vector database. The retrieved information will be used by other agents to assist the large language model in generating more accurate answers; (3) The task planning agent uses the large language model to generate specific workflow steps. When necessary, it can first retrieve relevant information from the vector database through the information retrieval agent to assist the large language model in generating more accurate and targeted answers; (4) The geometric modeling agent uses the large language model to generate a program executable by CAD software based on the design parameters. When necessary, it can first retrieve relevant information from the vector database through the information retrieval agent to assist the large language model in generating more accurate and targeted answers, and then call the CAD software to run the program to establish the geometric model of the part; (5) The strength analysis agent uses the large language model to generate a program executable by CAE software. When necessary, it can first retrieve relevant information from the vector database through the information retrieval agent to assist the large language model in generating more accurate and targeted answers, and then call the CAE software to run the program for strength analysis to check whether the part design meets the strength requirements. If the analysis result meets the strength requirements, it enters the dynamics analysis step; if it does not meet the strength requirements, it returns to the geometric modeling step to modify the geometric model according to the rules; (6) The dynamics analysis agent uses the large language model to generate a program executable by CAE software. When necessary, it can first retrieve relevant information from the vector database through the information retrieval agent to assist the large language model in generating more accurate and targeted answers, and then call the CAE software to run the program for dynamics analysis to check whether the part meets the vibration frequency requirements. If the analysis result meets the vibration frequency requirements, it meets all the design requirements and enters the next step; if it does not meet the vibration frequency requirements, it returns to the geometric modeling step to modify the geometric model according to the rules. After the geometric model is modified, the subsequent steps are executed in sequence; (7) After all the analyses are completed and the part meets all the design requirements, the report generation agent reads the output information and data of the CAD and CAE software, generates a design analysis report, and returns the conclusion, report, and design files to the user. When user intervention is required during the task execution process, it can be paused at any node where user intervention is required, waiting for the user to input, and then continue to execute after the user has completed the input. This multi-agent organizational structure design can facilitate function expansion, such as adding a new durability analysis agent, etc.; (5) Deploy and apply the multi-agent system: Deploy the multi-agent system to the production environment, and conduct continuous monitoring and maintenance. Improve the system according to the actual application situation and the problems that occur.

[0012] The embodiments described above are only specific implementation examples of the present invention, and do not limit the scope of implementation of the present invention. Changes made according to the principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A multi-agent computer-aided engineering and computer-aided design method, characterized in that: Including the steps: (1) Configure the software and hardware environment for locally tuning and deploying large language models, as well as developing and deploying multi-agent systems and vector databases, either locally or by leveraging cloud services; (2) Select an open-source artificial intelligence large language model and fine-tune it with professional knowledge data in the CAE and CAD fields to improve the accuracy and reliability of the inference and generation of the large language model in the CAE and CAD professional fields; (3) Establish a vector database with enterprise-specific documents, data, or program codes such as standards and specifications for product design and simulation analysis, and use retrieval-augmented generation (RAG) technology to assist the large language model in generating more accurate, reliable, and targeted answers; (4) For the CAE simulation analysis and CAD design optimization tasks of product design, develop a multi-agent system based on the large language model. The user interacts with the multi-agent system, proposes the goals and requirements for product design and simulation optimization, and each agent collaborates, autonomously plans, and automatically executes the workflow and steps. Use the large language model to generate program codes executable by CAE or CAD software, and call the CAE or CAD software to run the program codes to automatically complete the product design simulation analysis and optimization; (5) Deploy the developed multi-agent to the production environment of actual applications and conduct continuous monitoring and maintenance in actual applications.

2. The multi-agent computer-aided engineering and computer-aided design method according to claim 1, characterized in that: In step (4), the interaction mode between the user and the multi-agent system is one of the input natural language instruction mode, voice mode, graphical user interface (GUI) mode, reading a text file written by the user in advance, etc., or a multi-modal interaction mode that combines them.

3. A multi-agent computer-aided engineering and computer-aided design method according to claim 1, characterized in that: In step (4), the multi-agent system can read the data and information output by CAE and CAD software, automatically evaluate the task completion status, and re-plan the task when necessary.

4. A multi-agent computer-aided engineering and computer-aided design method according to claim 1, characterized in that: In step (4), the multi-agent system can automatically handle errors or problems. When the errors or problems exceed the system's automatic handling capacity, the system can request the user to intervene for handling.

Images