An industrial design automation method based on large artificial intelligence models and agents

Abstract

The present invention discloses an industrial design automation method based on an artificial intelligence large model and an intelligent agent, which relates to the technical field of industrial design and solves the technical problem that it is difficult to achieve industrial design automation at present. The method includes: automatically matching a large model according to the input industrial product-related information; extracting design optimization problems through a problem analysis intelligent agent based on the large model and decomposing them into optimization tasks; generating an objective function or simulation code through an objective quantification intelligent agent and providing an interface; constructing an intelligent agent through a surrogate model to fit the mapping from simulation input to output and providing an interface; packaging the above interfaces through an optimization encapsulation intelligent agent and executing the optimization tasks; and generalizing according to the industrial design automation embodiment to obtain a hierarchical intelligent agent architecture for industrial design using an artificial intelligence large model for industrial design automation. The present invention can achieve a high degree of automation in industrial design.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial design, and in particular to an industrial design automation method based on an artificial intelligence large model and an intelligent agent. Background Art

[0002] In recent years, large language models (LLMs) have rapidly emerged and demonstrated unprecedented capabilities in natural language processing tasks. These models are trained on large corpora of text and have the ability to understand and generate human language, opening up new avenues for applications in various industries. LLMs have also been widely applied to the field of industrial design.

[0003] The design cycle of industrial products usually includes several parts such as market research, requirements analysis, conceptual design, and detailed design. The development of LLMs has brought great development opportunities to industrial design. At the market research level, some studies have shown that various LLMs respond to classic behavioral economics experiments in an intuitive and empirical way. For example, in a currently adopted method, market researchers can enhance the response of LLMs by fine-tuning combined with previous survey data from similar backgrounds; this method improves the consistency between the responses of LLMs to existing product features and more importantly, new product features, and human responses. Additionally, at the requirements analysis level, some people have explored the potential of LLMs in promoting the requirements engineering (RE) process, and proposed key directions and SWOT analyses for the research and development of using LLMs for requirements engineering, improving the efficiency and accuracy of requirements-related tasks. In terms of conceptual design, some people have also proposed an artificial intelligence-enhanced multimodal collaborative design framework, using LLMs to establish an iterative prompting mechanism to generate precise visual solutions. In addition, some people have proposed a method for morphological analysis of conceptual design enhanced by LLMs, by refining the design process into three main stages: decomposition, generation, and combination, providing targeted guidance and support for designers when applying morphological analysis. In terms of detailed design, some people have made LLMs more conducive to API calls of industrial software through fine-tuning, thus realizing automated detailed design using LLMs; in addition, there is also a method that has achieved the optimized design of shear wall structures by converting the simulation intentions of CAE engineers into code.

[0004] In industrial design, the design optimization of products is an important way to improve product quality and performance, which is not only related to product functionality, production efficiency, and cost, but also directly affects market competitiveness and sustainable development. The industrial design process usually involves multiple interrelated and coupled processes. To achieve global optimal operation, it is necessary to comprehensively consider multiple aspects such as the complexity, multi-objectives, multi-scale, and dynamic optimization of problems in industrial design. Currently, various heuristic algorithms and deep learning algorithms have been applied to industrial design, and the application scope is continuously expanding, covering fields such as mechanical design and circuit design.

[0005] Specifically, in the paper "Large language models as evolutionary optimizers" (Liu S, Chen C, Qu X, et al. Large language models as evolutionary optimizers[C] / / 2024 IEEE Congress on Evolutionary Computation (CEC). IEEE, 2024: 1-8.), using large language models as zero-shot learning evolutionary operators shows that large language models can solve complex NP (nondeterministic polynomial) hard combinatorial optimization problems (NP-hard problem). However, for small-scale problem execution, the complete optimization process of this method is very expensive and time-consuming (about half a day). In addition, in industrial design, whether it is mechanical design or circuit design, it is necessary to perform mathematical modeling on physical world products to simulate the entity models of products. However, the optimizers used in this method lack mathematical modeling techniques for the physical world, making it difficult to quantify the quality of the design and thus difficult to apply to industrial problems.

[0006] In addition, in the paper titled "A large language model-based computing platform for automated building energy modeling" (Jiang G, Ma Z, Zhang L, et al. EPlus-LLM: A large language model-based computing platform for automated building energy modeling[J]. Applied Energy, 2024, 367: 123431.), fine-tuning technology is used to adjust a general large language model into a code model suitable for Application Programming Interface (API) calls, thus achieving highly automated modeling and simulation. This method enables the large language model to establish an understanding of the feedback from the physical world through the API and supports the automation of industrial design through API calls. However, the design of this method is in an open-loop state. Given the relatively weak mathematical capabilities of large language models, this open-loop design method is unreliable and an optimization process needs to be introduced to ensure the stability of the design results. Therefore, in the paper titled "Intelligent design and optimization system for shear wall structures based on large language models and generative artificial intelligence" (Qin S, Guan H, Liao W, et al. Intelligent design and optimization system for shear wall structures based on large language models and generative artificial intelligence[J]. Journal of Building Engineering, 2024, 95: 109996.), the large language model is used as the core controller, interacting with engineers to interpret their language descriptions and convert them into executable computer code. Subsequently, the system autonomously completes intelligent design tasks using appropriate structure generation and optimization methods. However, in this method, the interaction between humans and the large language model is continuous and the design needs to be adjusted constantly, making it difficult to achieve a highly automated system.

[0007] In the process of implementing the present invention, the inventors found that there are at least the following problems in the prior art:

[0008] At present, it is difficult to achieve a high degree of automation in industrial design. Summary of the Invention

[0009] The purpose of the present invention is to provide an industrial design automation method based on an artificial intelligence large model and an intelligent agent to solve the technical problem that it is currently difficult to achieve a high degree of automation in industrial design in the prior art. The many technical effects that can be produced by the preferred technical solutions provided by the present invention are described in detail below.

[0010] To achieve the above object, the present invention provides the following technical solutions:

[0011] An industrial design automation method based on an artificial intelligence large model and an intelligent agent provided by the present invention includes the following steps:

[0012] According to the relevant information of the industrial product input, extract the design optimization problem, and automatically match the large language model according to the design optimization problem;

[0013] Decompose the design optimization problem into a target quantification task, an agent model establishment task, and an optimization initialization task through the large language model;

[0014] Generate an objective function according to the target quantification task, and convert the objective function into a target quantification interface;

[0015] According to the agent model establishment task and the target quantification interface, use a neural network to fit the mapping from the emulator input to the output to establish an agent model, and convert the agent model into an agent model interface;

[0016] According to the optimization initialization task, the target quantification interface, and the agent model interface, perform target optimization on the agent model to obtain an optimization result, and output the optimal design scheme of the industrial product according to the optimization result.

[0017] Optionally, the automatically matching the large language model according to the design optimization problem includes:

[0018] Obtain the evaluation information of various different large language models in the historical call record of the large language model;

[0019] Automatically match the optimal large language model for the design optimization problem according to the task type of the design optimization problem and the evaluation information of the large language model.

[0020] Optionally, generating the objective function according to the target quantization task includes: calling a simulation software to generate the objective function according to the complexity of the target quantization task, or directly having the large language model generate the objective function.

[0021] Optionally, calling the simulation software to generate the objective function includes:

[0022] Performing preprocessing on the target quantization task using the finite element analysis model in the simulation software;

[0023] According to the preprocessing results, combining the chain of thought and retrieval enhancement to generate and call the simulator to simulate and generate the objective function.

[0024] Optionally, the method further includes: using the large language model as an analyzer to determine the complexity of the target quantization task.

[0025] Optionally, establishing the task and the target quantization interface according to the surrogate model, and using a neural network to fit the mapping from the simulator input to the output to establish the surrogate model includes:

[0026] Obtaining the solution space of the surrogate model establishment task through the large language model;

[0027] Performing uniform sampling in the solution space, and constructing the structural dataset of the industrial product through the target quantization interface and the sampling data;

[0028] Training the neural network using the structural dataset to obtain the surrogate model.

[0029] Optionally, after training the neural network using the structural dataset, the method further includes: evaluating the convergence curve and fitting effect of the neural network through the large language model; if the convergence curve diverges or the fitting effect is unqualified, the large language model adjusts the learning rate and the number of training rounds of the neural network to retrain.

[0030] Optionally, before training the neural network, the method further includes: normalizing the structural dataset through the large language model.

[0031] Optionally, optimizing the surrogate model according to the optimization initialization task, the target quantization interface, and the surrogate model interface includes:

[0032] Generating through retrieval enhancement, and selecting an optimization algorithm that matches the optimization initialization task from a preset algorithm library;

[0033] Package the target quantization interface and the proxy model interface into the target function, and perform target optimization on the target function through the selected optimization algorithm.

[0034] Optionally, the performing target optimization on the proxy model according to the optimization initialization task, the target quantization interface, and the proxy model interface further includes: after performing target optimization on the target function, performing a convergence check on the optimization result obtained by the target optimization through the large language model; if it is detected that the optimization result does not converge, the large language model adjusts the hyperparameters in the target optimization process for re-optimization.

[0035] A hierarchical agent framework applicable to the industrial design field, used to execute any one of the above-mentioned industrial design automation methods based on an artificial intelligence large model and an agent to solve the industrial design automation problem, including: an L0 requirement analysis and conceptual design agent, an L1 preliminary design agent, and an L2 detailed design agent;

[0036] The L0 requirement analysis and conceptual design agent is used to extract design optimization problems according to the relevant information of industrial products, and decompose the design optimization problems into target quantization tasks, proxy model establishment tasks, and optimization initialization tasks;

[0037] The L1 preliminary design agent is used to generate a target function according to the target quantization task, convert the target function into a target quantization interface, and according to the proxy model establishment task and the target quantization interface, use a neural network to fit the mapping from the emulator input to the output to establish a proxy model, and convert the proxy model into a proxy model interface;

[0038] The L2 detailed design agent is used to perform target optimization on the proxy model according to the optimization initialization task, the target quantization interface, and the proxy model interface to obtain an optimization result, and output the optimal design scheme of the industrial product according to the optimization result.

[0039] Implementing one of the above technical solutions of the present invention has the following advantages or beneficial effects:

[0040] The present invention uses a large language model as a bridge for product simulation and optimization. All process steps in industrial design are automatically completed by the large language model, achieving full-process automation of industrial design from design to optimization. By converting the target quantification task and surrogate model establishment task into interfaces and opening them for subsequent operations to receive feedback at the complex physical world level, the dependence on the mathematical ability of the large language model in the optimization process is eliminated. Additionally, by designing an automatic surrogate model establishment, the surrogate model is automatically established through multiple rounds of conversations before optimization, reducing the repeated simulation time in the optimization process, greatly reducing the simulation time, and ensuring the model accuracy. Description of the Drawings

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings. In the drawings:

[0042] Figure 1 is the flowchart of the industrial design automation method based on the artificial intelligence large model and intelligent agent in Embodiment 1 of the present invention;

[0043] Figure 2 is the structural schematic diagram of the ten-bar truss in Embodiment 1 of the present invention;

[0044] Figure 3 is the flowchart of step S1 of the industrial design automation method based on the artificial intelligence large model and intelligent agent in Embodiment 1 of the present invention;

[0045] Figure 4 is the flowchart of step S3 of the industrial design automation method based on the artificial intelligence large model and intelligent agent in Embodiment 1 of the present invention;

[0046] Figure 5 is the flowchart of step S4 of the industrial design automation method based on the artificial intelligence large model and intelligent agent in Embodiment 1 of the present invention;

[0047] Figure 6 is the flowchart of step S5 of the industrial design automation method based on the artificial intelligence large model and intelligent agent in Embodiment 1 of the present invention;

[0048] Figure 7 is the simulation diagram of the ten-bar truss in Embodiment 1 of the present invention;

[0049] Figure 8 is the convergence curve diagram of the ten-bar truss surrogate model in Embodiment 1 of the present invention;

[0050] Figure 9 is the optimization result diagram of the ten-bar truss in Embodiment 1 of the present invention;

[0051] Figure 10 FIG. 3 is a schematic structural diagram of the hierarchical agent framework applicable to the industrial design field in the second embodiment of the present invention. Detailed implementation manners

[0052] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, various exemplary embodiments to be described below will refer to the corresponding drawings, which form a part of the exemplary embodiments and describe various exemplary embodiments that may be adopted to implement the present invention. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present disclosure. It should be understood that they are merely examples of processes, methods, devices, etc. consistent with some aspects of the present invention as detailed in the appended claims, and other embodiments may also be used, or structural and functional modifications may be made to the embodiments listed herein without departing from the scope and essence of the present invention.

[0053] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", etc. indicate the orientation or positional relationship based on the drawings shown, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the elements referred to must have a specific orientation, be constructed and operated in a specific orientation. The terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. The meaning of the term "plurality" is two or more. The terms "connected" and "coupled" should be understood in a broad sense. For example, they may be fixedly connected, detachably connected, integrally connected, mechanically connected, electrically connected, communicatively connected, directly connected, indirectly connected through an intermediate medium, and may be the internal connection or interaction relationship between two elements. The term "and / or" includes any and all combinations of one or more of the related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0054] In order to illustrate the technical solutions described in the present invention, the following will be described by specific examples, and only the parts related to the embodiments of the present invention are shown.

[0055] Embodiment 1:

[0056] As Figure 1As shown in the figure, the present invention provides an industrial design automation method based on an artificial intelligence large model and an intelligent agent, including the following steps: S1. According to the relevant information of the input industrial product, extract design optimization problems, and automatically match a large language model according to the design optimization problems; wherein, the industrial product can be household appliances, such as televisions, refrigerators, washing machines, etc.; electronic products, such as laptop computers, mobile phones, smart watches, etc.; means of transportation, such as cars, ships, drones, etc.; the extracted design optimization problems can be to optimize the mechanical structure or circuit structure of these products; S2. Decompose the design optimization problems into target quantification tasks, surrogate model establishment tasks, and optimization initialization tasks through the large language model; S3. Generate an objective function according to the target quantification tasks, and convert the objective function into a target quantification interface; S4. According to the surrogate model establishment tasks and the target quantification interface, use a neural network to fit the mapping from the emulator input to the output to establish a surrogate model, and convert the surrogate model into a surrogate model interface; S5. According to the optimization initialization tasks, the target quantification interface, and the surrogate model interface, perform target optimization on the surrogate model to obtain an optimization result, and output the optimal design scheme of the industrial product according to the optimization result.

[0057] The industrial design automation method based on the artificial intelligence large model and the intelligent agent provided in this embodiment is applied to the industrial design field. Using the large language model as the bridge for product simulation and optimization, all process steps in industrial design are automatically completed by the large language model, realizing the full-process automation of industrial products from design to optimization; by converting both the target quantification tasks and the surrogate model establishment tasks into interfaces and opening them for subsequent operations to receive feedback at the complex physical world level, the dependence on the mathematical ability of the large language model in the optimization process is eliminated; and by designing the surrogate model to automatically establish the model, the surrogate model is automatically established through multiple rounds of conversations before optimization, reducing the repeated simulation time in the optimization process, greatly reducing the simulation time, and also ensuring the model accuracy.

[0058] Next, in combination with Figures 1 to 9 to introduce in detail the specific implementation steps of the industrial design automation method based on the artificial intelligence large model and the intelligent agent provided in this embodiment:

[0059] First, execute step S1. According to the relevant information of the input industrial product, extract design optimization problems, and automatically match a large language model according to the design optimization problems; such as Figure 2As shown, for example, it is necessary to design a structure of a ten-bar truss, which is welded by ten beam units. Then its product information can be a cross-sectional view of this structure in CAD software or the PyAnsys modeling code of the 3D model, etc. The design optimization problems that can be extracted from this information are: how to set the horizontal length l of the left half, the width w1 of the left beam, and the width w2 of the right beam to make this structure optimal. Then, match a large language model that is most suitable for solving these problems according to these optimization problems.

[0060] Specifically, as Figure 3 shown, automatically matching a large language model according to the design optimization problems includes: S11. Obtain the evaluation information of various different large language models from the historical call records of the large language model. There are many types of large language models. For example, according to the model architecture, they can be divided into: recurrent neural network, convolutional neural network, Transformer model, etc.; according to the model scale, they can be divided into: small language model, large language model, extra-large model, etc.; each type is applicable to different task situations. Therefore, the processing effects of each model for different task types can be queried through the historical call information. S12. Automatically match the optimal large language model for the design optimization problem according to the task type of the design optimization problem and the evaluation information of the large language model. Matching a large language model that is most suitable for processing this task (i.e., the optimal large language model) through the historical call record can improve the task processing efficiency and the reliability of the generated results. Intelligent selection and scheduling of the calls to different large language models according to different task types can ensure that the most suitable large language model is selected when processing a specific task, improving the quality of the generated results. By integrating the advantages of multiple models, in complex industrial design tasks, various large language models can work together to effectively solve different types of problems, ensuring the accuracy and stability of the generated results.

[0061] Then, execute step S2. Decompose the design optimization problem into a target quantification task, a surrogate model establishment task, and an optimization initialization task through the large language model. The target quantification task is to convert the abstract product information into specific and measurable indicators to more clearly evaluate the accuracy of the design optimization product. Taking Figure 2 the ten-bar truss structure in Figure 2 as an example, its target can be quantified as solving the maximum volume, maximum deformation, etc. of the ten-bar truss structure. The surrogate model establishment task is to construct a surrogate model of the industrial product and simulate and verify the fitting effect between the surrogate model and the real industrial product. For the ten-bar truss structure in Figure 2In the ten-bar truss structure, its optimization initialization task is to optimize the objectives of surrogate model 1 and surrogate model 2. The large language model automatically decomposes the design optimization problem. The established objective quantification task and surrogate model task enable the large language model to receive feedback at the complex physical world level, so as to eliminate the dependence on the mathematical ability of the large language model during the optimization process.

[0062] Next, execute step S3: Generate an objective function according to the objective quantification task, and convert the objective function into an objective quantification interface. After generating the objective function, use the large language model as a controller and code generator to convert the objective function into code form, and then package the code into an objective quantification interface, and this interface is always open for subsequent operations (that is, it can be called at any time during the subsequent process). It should be noted that the objective function can be converted into Python code, C code, C++ code or Java code, which can be selected according to the actual situation. In the objective quantification task, use the large language model as an analyzer. First, use the natural language to evaluate the shape and size of the task model, identify the optimization variables and solution space, and finally determine the number and complexity of the optimization objectives, and convert the fuzzy requirements of the optimization into a quantifiable optimization model.

[0063] Furthermore, generating an objective function according to the objective quantification task includes: calling a simulation software to generate an objective function according to the complexity of the objective quantification task, or directly making the large language model generate an objective function. Further, the large language model can be used as an analyzer to determine the complexity of the objective quantification task. The large language model discriminates the complexity of the objective quantification task by analyzing whether the optimization objectives in the objective optimization task are easy. For easily modeled optimization objectives, the large language model can directly convert them into mathematical functions (i.e., objective functions), then convert them into code, and then perform dynamic verification and package them into objective quantification interfaces. For optimization objectives with greater modeling difficulty, the large language model calls simulation software to generate objective functions. For industrial products of circuit type, simulation software specialized for electronic circuits such as LTspice, Proteus, PSPICE or MultisimLive can be used; for industrial products of mechanical structure type, simulation software specialized for multi-physics fields such as Ansys or COMSOL Multiphysics can be used. Calling simulation software can convert the complex partial differential equations in the physical problems and engineering problems in the task into the form of global matrix equations, so as to simplify the solution equations and save the calculation amount.

[0064] Specifically, such as Figure 4As shown, call the simulation software to generate the objective function, including: S31. Use the finite element analysis model in the simulation software to preprocess the target quantization task. The preprocessing includes network division, constraint application, load application, etc.; S32. According to the preprocessing results, combine the chain of thought and retrieval-augmented generation to call the simulator to simulate and generate the objective function. Among them, the chain of thought (CoT) is a technology used to enhance the reasoning ability of models in the fields of artificial intelligence and machine learning. It can improve the performance and interpretability of models on complex problems by explicitly including intermediate reasoning steps in the output of the model; Retrieval-augmented Generation (RAG) is a model that combines retrieval and generation technologies. It can reference information from external knowledge bases to generate answers or content. Call the surrogate model through the API of the simulation software, obtain a small number of simulation examples and API documents to start the simulation (such as preprocessing, applying loads, and postprocessing) to generate the objective function, convert it into code through a large language model, then pass the code to the tester to verify the simulation results, and finally package it into an interface.

[0065] Next, execute step S4. Establish the task and target quantization interface according to the surrogate model, use the neural network to fit the mapping from the simulator input to the output to establish the surrogate model, and convert the surrogate model into a surrogate model interface. Among them, computer-aided engineering software (CAE) can be used as the simulator. Similarly, use the large language model as the controller and code generator to call the neural network to establish the surrogate model and convert the surrogate model into code, and then package the code as a surrogate model interface for ready-to-use calls in subsequent processes. If in step S3, the objective function is directly generated by the large language model, the surrogate model does not need to be established, and directly enter the optimization initialization task.

[0066] In an actual optimization case, the finite element analysis in the simulation software is the most time-consuming part of the whole process. Therefore, it is necessary to construct a surrogate model to approximate the distribution of the simulator. In this embodiment, the surrogate model is constructed through a neural network model. Specifically, as Figure 5 shown, establish the task and target quantization interface according to the surrogate model, and use the neural network to fit the mapping from the simulator input to the output to establish the surrogate model, including: S41. Obtain the solution space of the surrogate model establishment task through the large language model; this solution space is the set of all possible answers generated when processing the target quantization task. S42. Perform uniform sampling in the solution space, and construct the structure dataset of industrial products through the target quantization interface and the sampled data; this structure dataset includes the parameters necessary to generate each industrial product structure. Still taking Figure 2Taking the ten-bar truss structure as an example, assuming that it is necessary to balance the horizontal length l of the left half, the width w1 of the left beam, and the width w2 of the right beam so that the structure can have the largest possible volume and deformation, then its solution space is 0.05 ≤ l ≤ 0.15, 0.01 ≤ w1, w2 ≤ 0.02, and its structure dataset is all the ten-bar truss structures that can be constructed by the parameters l, w1, and w2 within the solution space. S43. Use the structure dataset to train the neural network to obtain a surrogate model. The mapping from the simulator input to the output is to input the structure parameters of the industrial product into the simulator, and the industrial product simulated and output by the simulator. The established surrogate model is to continuously simulate the process of the simulator simulating the industrial product. Before optimization, a surrogate model is automatically established by the neural network through multiple rounds of dialogue, which can reduce the repeated simulation time during the optimization process, greatly reduce the simulation time, and at the same time ensure the model accuracy. Since the simulation results of the simulator may sometimes be very small, therefore, before training the neural network, the structure dataset is normalized by the large language model to reduce the impact of the magnitude difference between the targets on the accuracy of the surrogate model and ensure the accuracy of the surrogate model.

[0067] To further ensure the accuracy of the surrogate model, according to the surrogate model, a task and target quantization interface is established. After training the neural network, the method further includes: evaluating the convergence curve and fitting effect of the neural network by the large language model; if the convergence curve diverges or the fitting effect is unqualified, then adjust the learning rate and the number of training rounds of the neural network and retrain. Use the large language model to evaluate the quality of the trained neural network. If the convergence curve is divergent or the fitting effect is not good, it means that the accuracy of the neural network is very poor, and the learning rate and the number of training rounds need to be adjusted for retraining until a high-precision surrogate model is obtained. Among them, the fitting effect of the neural network can be evaluated by the sum of squared residuals method, R² value, residual plot, or correlation coefficient r.

[0068] Finally, execute step S5. According to the optimization initialization task, target quantization interface, and surrogate model interface, optimize the target of the surrogate model to obtain an optimization result, and output the optimal design scheme of the industrial product according to the optimization result. Only need to input to the large language model once, and the large language model can automatically complete the design and optimization process of the industrial product, realizing a high degree of automation in the entire process of industrial design.

[0069] Further, as Figure 6As shown in the figure, according to the optimized initialization task, the target quantization interface, and the surrogate model interface, the target optimization of the surrogate model is carried out, including: S51. By using the retrieval enhancement algorithm, select an optimization algorithm that matches the optimized initialization task from the preset algorithm library. The preset algorithm library can be an algorithm library built by oneself or an existing algorithm library such as the Python library, the C / C++ open-source algorithm library, The Algorithms, etc. S52. Package the target quantization interface and the surrogate model interface into the objective function, and perform target optimization on the objective function through the selected optimization algorithm. Among them, the optimization algorithms include NSGA-II (Non-dominated Sorting Genetic Algorithm II), NSWOA (Non-dominated Sorting Whale Optimization Algorithm), MOPSO (Multi-Objective Particle Swarm Optimization), and SA (Simulated Annealing Algorithm), etc. The most suitable algorithm can be selected according to the actual optimization problem for optimization.

[0070] Finally, obtain the optimal solution (the optimal solution for a single objective) or the Pareto front (the set of solutions that achieve the best balance among multiple objectives) of the target optimization. According to the optimization results, the optimal design scheme of industrial products can be automatically obtained. Figure 2 The ten-bar truss structure in [reference] is a multi-objective optimization problem. The horizontal length l of the left half, the width w1 of the left beam, and the width w2 of the right beam are the optimization variables. Input the relevant information of this structure into the method of this embodiment for automated design, such as Figure 7 shown in the figure is the automatic simulation result of this structure; Figure 8 is the convergence situation of the surrogate model of this structure; Figure 9 is the multi-objective optimization result of this structure, that is, the Pareto front; Figures 7 to 9 The process is completely automated and no longer requires human participation.

[0071] To ensure the accuracy of the target optimization, according to the optimized initialization task, the target quantization interface, and the surrogate model interface, the target optimization of the surrogate model also includes: after performing the target optimization on the objective function, perform a convergence check on the optimization results obtained by the target optimization through a large language model; if it is detected that the optimization results have not converged, adjust the hyperparameters in the target optimization process for re-optimization to ensure the best optimization results are obtained.

[0072] Embodiment 2:

[0073] Such as Figure 10As shown in the figure, the present invention also provides a hierarchical agent framework applicable to the field of industrial design, which is used to execute an industrial design automation method based on an artificial intelligence large model and an agent described in Embodiment 1 to solve the problem of industrial design automation, including: an L0 requirement analysis and concept design agent, an L1 preliminary design agent, and an L2 detailed design agent; the L0 requirement analysis and concept design agent is used to extract design optimization problems according to relevant information of industrial products, and decompose the design optimization problems into target quantification tasks, surrogate model establishment tasks, and optimization initialization tasks; the L1 preliminary design agent is used to generate an objective function according to the target quantification tasks, convert the objective function into a target quantification interface, and according to the surrogate model establishment tasks and the target quantification interface, use a neural network to fit the mapping from the emulator input to the output to establish a surrogate model, and convert the surrogate model into a surrogate model interface; the L2 detailed design agent is used to perform target optimization on the surrogate model according to the optimization initialization tasks, the target quantification interface, and the surrogate model interface to obtain an optimization result, and output the optimal design scheme of the industrial product according to the optimization result.

[0074] As Figure 10 shown, the L0-level agent can perform requirement analysis and concept design. The agent based on the black-box multimodal large model communicates with the user to obtain clear requirements. This layer obtains preliminary requirements based on the multimodal capabilities and shallow domain knowledge of the large model itself, conducts preliminary concept design on the product, and hands it over to the L1-level agent. The L0-level agent understands the real world through multimodality and accurately identifies long sequences, constructs an efficient model through evolutionary computation, and adopts various talent books scales to correspond to different scenarios. The L0 requirement analysis and concept design agent in this embodiment is the L0-level agent.

[0075] The L1-level agent can perform preliminary design on the product. By integrating a knowledge base and a knowledge graph containing professional domain knowledge, it imposes constraints based on first principles on the results of concept design. By integrating professional domain knowledge, the preliminary design can obtain a product that roughly meets the current requirements, reducing the search space for subsequent simulation optimization. At the same time, the operation knowledge of industrial software is also included in the L1-level agent, which is the basis for the L2-level agent to perform detailed design. The L1-level agent is a large world model and a preliminary design based on a gray-box world model. The L1 preliminary design agent in this embodiment is the L1-level agent.

[0076] The L2-level intelligent agent can conduct detailed design of products based on the white-box knowledge of industrial software, and call industrial software (such as CAD and CAE, etc.) through the code generation ability of large language models. For example, first convert the conceptual design into a CAD model, then use the API of the CAE software to perform finite element analysis on the CAD model and optimize the product at the performance level using the simulation results, and finally obtain the industrially designed product with the optimal design. The L2-level intelligent agent can improve the design and implementation of the intelligent agent based on the white-box design iteration of existing tools. The L2 detailed design intelligent agent in this embodiment is the L2-level intelligent agent.

[0077] The embodiment is only a special case and does not indicate that the present invention has only such an implementation.

[0078] The above are only the preferred embodiments of the present invention. Those skilled in the art know that without departing from the spirit and scope of the present invention, various changes or equivalent replacements can be made to these features and embodiments. In addition, under the teaching of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the protection scope of the present invention.

Claims

1. An industrial design automation method based on artificial intelligence large model and intelligent agent, characterized in that: The following steps are involved: Extract design optimization problems based on input relevant information of industrial products, and automatically match the large language model based on the design optimization problems; Decomposing the design optimization problem into a target quantification task, a proxy model establishment task and an optimization initialization task through the large language model; Generate an objective function according to the target quantization task, and convert the objective function into a target quantization interface; Establishing a task and target quantization interface according to the proxy model, using a neural network to fit a mapping of simulator input to output to establish a proxy model, and converting the proxy model to a proxy model interface; According to the optimization initialization task, the target quantization interface and the proxy model interface, the proxy model is optimized to obtain an optimization result, and the optimal design solution of the industrial product is output according to the optimization result; The step of establishing a task and target quantization interface according to the proxy model and using a neural network to fit the mapping of simulator input to output to establish the proxy model includes: Acquire a solution space of the proxy model establishment task through the large language model; Perform uniform sampling in the solution space, and construct a structural data set of the industrial product through the target quantization interface and the sampled data; Using the structured data set to train a neural network to obtain the proxy model; The step of performing target optimization on the proxy model according to the optimization initialization task, the target quantization interface and the proxy model interface includes: By searching for enhanced generation, an optimization algorithm matching the optimization initialization task is selected from a preset algorithm library; The target quantization interface and the proxy model interface are packaged into the target function, and the target function is optimized by the selected optimization algorithm.

2. The industrial design automation method based on artificial intelligence big model and intelligent agent according to claim 1, characterized in that: The automatically matching the large language model according to the design optimization problem includes: Acquire evaluation information of various different large language models from the historical call records of the large language model; According to the task type of the design optimization problem and the evaluation information of the large language model, the optimal large language model is automatically matched to the design optimization problem.

3. The industrial design automation method based on artificial intelligence big model and intelligent agent according to claim 1, characterized in that: The generating the target function according to the target quantization task includes: calling simulation software to generate the target function according to the complexity of the target quantization task, or directly using the large language model to generate the target function.

4. The industrial design automation method based on artificial intelligence big model and intelligent agent according to claim 3 is characterized in that: The calling simulation software to generate the target function comprises: Using the finite element analysis model in the simulation software to pre-process the target quantification task; According to the pre-processing results, the target function is generated by combining the thinking chain and retrieval enhancement generation and calling the simulator to simulate.

5. The industrial design automation method based on artificial intelligence big model and intelligent agent according to claim 3 is characterized in that: The method further includes: using the large language model as an analyzer to determine the complexity of the target quantization task.

6. The industrial design automation method based on artificial intelligence big model and intelligent agent according to claim 1, characterized in that: After the neural network is trained using the structured data set, the method further includes: evaluating the convergence curve and fitting effect of the neural network through the large language model; if the convergence curve diverges or the fitting effect is unsatisfactory, adjusting the learning rate and the number of training rounds of the neural network through the large language model to retrain.

7. The industrial design automation method based on artificial intelligence big model and intelligent agent according to claim 1, characterized in that: Before training the neural network, the method further includes: normalizing the structured data set using the large language model.

8. The industrial design automation method based on artificial intelligence big model and intelligent agent according to claim 1, characterized in that: The target optimization of the proxy model according to the optimization initialization task, the target quantization interface and the proxy model interface also includes: after the target function is optimized, the optimization result obtained by the target optimization is checked for convergence by the large language model; if it is checked that the optimization result has not converged, the large language model adjusts the hyperparameters in the target optimization process for re-optimization.

9. A hierarchical agent framework suitable for the field of industrial design, characterized in that: An industrial design automation method based on an artificial intelligence large model and an agent for executing any one of claims 1 to 8 to solve industrial design automation problems, comprising: an L0 demand analysis and conceptual design agent, an L1 preliminary design agent, and an L2 detailed design agent; The L0 demand analysis and concept design agent is used to extract design optimization problems based on relevant information of industrial products, and decompose the design optimization problems into target quantification tasks, agent model establishment tasks and optimization initialization tasks; The L1 preliminary design agent is used to generate an objective function according to the target quantization task, convert the objective function into a target quantization interface, establish a task and a target quantization interface according to the proxy model, use a neural network to fit the mapping of simulator input to output to establish a proxy model, and convert the proxy model into a proxy model interface; The L2 detailed design agent is used to perform target optimization on the proxy model according to the optimization initialization task, target quantization interface and proxy model interface, obtain optimization results, and output the optimal design solution for the industrial product according to the optimization results.

Images