Virtual physical system for industrial product development

Abstract

The application relates to a virtual physical system for industrial product research and development, which comprises a physical basic layer, a basic support layer, a research and development logic model layer, an intelligent agent cooperation layer, a product engineering layer and an application scenario layer. The system comprises a downlink data flow transfer mode and an uplink data flow transfer mode. In the downlink data flow transfer mode, product demand information is transferred from the application scenario layer to the physical basic layer in stages to obtain a physical calculation result. In the uplink data flow transfer mode, the physical calculation result is transferred from the physical basic layer to the application scenario layer in stages and is converted into a target achievement. By adopting the system, a complete closed loop from demand to verification is realized in a digital space through a six-layer architecture and bidirectional data flow transfer, the research and development cycle is shortened, and the research and development efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of research and development technology, and in particular to a virtual physical system for industrial product research and development. Background Technology

[0002] With the rapid development of digital simulation technology, technologies such as computer-aided engineering, multiphysics simulation, and digital twins have been widely used in industrial product development. These technologies can simulate and verify product performance to a certain extent in digital space, thereby partially replacing physical prototype testing and shortening the development cycle.

[0003] In traditional technologies, R&D personnel typically use separate software tools to perform different stages of R&D tasks. For example, they use 3D modeling tools for structural design, separate finite element analysis tools for mechanical simulation, computational fluid dynamics tools for thermal flow analysis, and specialized testing software for verification. Data transfer between these tools relies on manual export and import, and the various stages of the R&D process (conceptual design, detailed design, simulation verification, process manufacturing, and operation monitoring) are disconnected from each other at the information level.

[0004] However, the above-mentioned traditional R&D methods have the following technical problems: there is a lack of a unified data flow mechanism between the various separate simulation tools, and design parameters need to be manually converted when transferred between different tools, which makes it difficult to guarantee data consistency and traceability; the information fragmentation between different stages of R&D makes it difficult to obtain full-process data support for design decisions, and after each design change, modeling and settings need to be repeated in multiple tools, resulting in low R&D efficiency. Summary of the Invention

[0005] Therefore, it is necessary to provide a virtual physical system for industrial product development that can realize a complete closed loop from demand to verification in the digital space and improve R&D efficiency, in response to the above-mentioned technical problems.

[0006] In a first aspect, this application provides a virtual physical system for industrial product research and development, the system comprising:

[0007] The physical foundation layer stores a knowledge base that digitally represents the physical laws of the real physical world;

[0008] The basic support layer is deployed on top of the physical infrastructure layer and is used to provide data processing capabilities based on the physical infrastructure layer.

[0009] The R&D logic model layer, deployed on top of the basic support layer, is used to provide a standardized product R&D process template. The process template defines multiple R&D stages from requirements to testing and verification, as well as the mapping relationship between each R&D stage.

[0010] The intelligent agent collaboration layer, deployed on top of the basic support layer, includes a chief intelligent agent and multiple specialized intelligent agents. The chief intelligent agent is used to schedule the specialized intelligent agents to collaboratively execute R&D sub-tasks.

[0011] The product engineering layer, deployed above the intelligent agent collaboration layer, is used to transform the output of the intelligent agent collaboration layer into target results that conform to engineering specifications. The target results are represented by a three-dimensional digital model.

[0012] The application scenario layer, deployed above the product engineering layer, is used to receive product requirement information and display the target results.

[0013] The system includes a downlink data flow mode and an uplink data flow mode. In the downlink data flow mode, the product demand information flows from the application scenario layer to the physical infrastructure layer to obtain physical calculation results. In the uplink data flow mode, the physical calculation results flow from the physical infrastructure layer to the application scenario layer and are transformed into the target result.

[0014] In some optional embodiments, the downlink data transfer mode includes:

[0015] The application scenario layer receives product requirement information and then transmits it to the product engineering layer.

[0016] The product engineering layer breaks down the product requirement information into engineering tasks and then transmits them to the intelligent agent collaboration layer.

[0017] The intelligent agent collaboration layer, after scheduling the corresponding professional intelligent agents to execute the engineering task through the chief engineer intelligent agent, transmits the processing results to the R&D logic model layer.

[0018] The R&D logic model layer organizes the R&D sequence according to the process template and then transmits it to the basic support layer;

[0019] The basic support layer allocates computing resources and calls the knowledge base in the physical foundation layer to perform physical field numerical calculations to obtain physical calculation results.

[0020] In some optional embodiments, the uplink data transfer mode includes:

[0021] The physical foundation layer transmits the physical calculation results to the basic support layer for data integration.

[0022] The basic support layer transmits the integrated data to the R&D logic model layer for phased verification and optimization.

[0023] The R&D logic model layer transmits the verified and optimized data to the agent collaboration layer for multi-agent joint inference;

[0024] The intelligent agent collaboration layer transmits the inference results to the product engineering layer to generate the target results;

[0025] The product engineering layer transmits the target result to the application scenario layer, and the application scenario layer outputs the target result.

[0026] In some optional embodiments, the knowledge base of the physical foundation layer includes: basic laws describing the real physical world based on the principle of least action, governing equations describing the laws of motion of matter, and different physical field models for multi-field coupling calculations; the governing equations include at least one of the continuity equation, the momentum conservation equation, and the energy conservation equation.

[0027] In some optional embodiments, the basic support layer includes: a large language model, a data engineering laboratory, a cloud-native architecture, and a physical solver interface; the basic support layer is used to integrate physical laws with AI algorithms, providing computing power, data processing power, and model reasoning power to each agent in the agent collaboration layer, and the large language model includes an attention mechanism.

[0028] In some optional embodiments, the process template in the R&D logic model layer is a V-shaped R&D process template. The V-shaped R&D process template includes a decomposition stage sequence and an integration verification stage sequence. Each decomposition stage forms a bidirectional traceability link with its corresponding integration verification stage. The bidirectional traceability link is used to backtrack to the corresponding decomposition stage for design correction when verification fails.

[0029] In some optional embodiments, the R&D efficiency of the R&D logic model layer is evaluated using the following function:

[0030]

[0031] in, To improve the R&D efficiency of the V-shaped R&D process template. This represents the total number of research and development stages. Let be the verification function for the requirements and specifications of the i-th development phase. For the needs of the i-th research and development stage, For the specifications of the i-th research and development stage, is the weighting coefficient for the i-th research and development stage.

[0032] In some alternative embodiments, the plurality of specialized intelligent agents include at least one of the following:

[0033] A demand analysis intelligent agent is used to analyze the product demand information to obtain at least one initial design scheme;

[0034] A multiphysics processing agent is used to determine the physical field model corresponding to the initial design scheme, establish a multiphysics layer coupling relationship, and simulate the initial design scheme based on the multiphysics coupling relationship to obtain simulation results.

[0035] A physical information augmentation agent is used to identify problems in the simulation results and optimize the design scheme. The optimized design scheme is then fed back to the multiphysics processing agent for re-simulation until the simulation results meet the design requirements, thus obtaining the target design scheme.

[0036] In some optional embodiments, the plurality of specialized intelligent agents further include:

[0037] The testing and verification agent is used to acquire a virtual test bench, test and verify the target design scheme on the virtual test bench, record performance information under various working conditions, obtain test results based on the performance information, and, if the test results fail, revise the target design scheme and trigger the multiphysics processing agent to re-simulate the revised target design scheme, and re-test and verify the re-simulated design scheme until the test passes; the virtual test bench is consistent with the real physical testing equipment;

[0038] A target operating condition simulation agent is used to configure various target operating conditions on the virtual test bench, including extreme operating conditions and fault operating conditions.

[0039] In some optional embodiments, the plurality of specialized intelligent agents further include: a process design intelligent agent, used to generate manufacturing process parameters based on the target design scheme that has passed the test;

[0040] A manufacturing simulation intelligent agent is used to digitally simulate the manufacturing process based on the manufacturing process parameters to obtain manufacturing simulation results.

[0041] A quality pre-assessment intelligent agent is used to obtain predicted quality indicators based on the manufacturing simulation results and compare them with the target quality indicators in the product demand information to obtain quality comparison results.

[0042] A process optimization agent is used to adjust process parameters and / or process routes based on the quality comparison results until the product quality indicators obtained based on the manufacturing simulation results meet the product quality indicators in the product requirement information.

[0043] In some optional embodiments, the plurality of specialized intelligent agents further include:

[0044] A digital twin intelligent agent is used to acquire product operation data collected from physical products in the real physical world, synchronize the product operation data to the digital twin, and perform monitoring and early warning based on the status data of the digital twin.

[0045] In some optional embodiments, the plurality of specialized intelligent agents further include:

[0046] Intelligent optimization agents are used to generate product optimization suggestions based on accumulated operational data and user feedback;

[0047] A knowledge-accumulating intelligent agent is used to extract fault modes, performance degradation patterns, and optimization opportunities from operational data throughout the product's lifecycle, and to update the knowledge base.

[0048] The product iteration intelligent agent is used to trigger a new round of R&D cycle, starting from the optimization suggestions and entering the requirement definition stage to form a continuous iterative R&D closed loop.

[0049] In some optional embodiments, the decision-making strategy of each specialized intelligent agent selects the optimal action based on a value function iteration method, wherein the optimal value function of the value function iteration method is:

[0050]

[0051] in, Let be the optimal value function for state s, representing the maximum cumulative reward obtainable from state s; For an immediate reward of performing action a in state s, Discount factor ( ), used to measure the importance of future returns, Let be the state transition probability, representing the probability that after performing action a in state s, the state transitions to state s'.

[0052] In some optional embodiments, the simulation is performed using a neural network, the total loss function of which is:

[0053]

[0054] The total loss of the neural network, Data-driven loss measures the deviation between the neural network's predictions and the actual measured data. The physical constraint loss measures whether the neural network predictions satisfy the governing equations. Boundary condition loss measures whether the neural network satisfies a given condition at the boundary. , , The weighting coefficients are used to adaptively adjust the contributions of the three losses.

[0055] In some optional embodiments, the collaborative outputs of multiple specialized intelligent agents constitute an agent collaborative output matrix, and the aggregated output of the agent collaborative output matrix is ​​represented as follows:

[0056]

[0057] in, This is the total output of the agent matrix. The number of the specialized intelligent agents. For the first The weights of the aforementioned specialized intelligent agents For the first The output of the aforementioned specialized intelligent agent For the first The collaboration coefficient between the j-th specialized intelligent agent and the j-th specialized intelligent agent. For cooperative operators.

[0058] In some optional embodiments, the product engineering layer includes:

[0059] The architecture and platform definition module is used to define the general architecture of the product platform and the interface specifications between modules;

[0060] The physical integration design module is used for the joint design and interface matching of multiple physical subsystems of a product;

[0061] The system-level engineering management module is used to manage requirement items, design changes, and verification records throughout the product lifecycle.

[0062] The Quality and Manufacturing Engineering module is used to perform virtual quality assessments and digital manufacturing process design.

[0063] The supply chain and cost analysis module is used to build supply chain network models and estimate and optimize product costs.

[0064] In some alternative embodiments, the engineering performance of the product engineering layer is evaluated using the following function:

[0065]

[0066] in, For the engineering performance of product X, The output of the physical integration design module, This is the output of the simulation stage in the physical integration design module. This is the output of the verification stage in the system-level engineering management module. This is the output of the quality and manufacturing engineering module. , , , These are the weight coefficients for each corresponding module.

[0067] In some optional embodiments, the application scenario layer includes:

[0068] The demand interaction unit is used to obtain product demand information input by users in the form of natural language and / or structured data;

[0069] The visualization unit is used to output the target results in at least one of the following forms: 3D model, simulation cloud map, data chart, and scheme report.

[0070] In some optional embodiments, the application value of the application scenario layer is evaluated using the following function:

[0071]

[0072] Total value of the application layer For the number of application scenarios, Let k be the weight of the k-th scene. Let be the utility function for this scenario. Input data for product X.

[0073] In some optional embodiments, the human-machine collaboration efficiency of the system is evaluated using the following function:

[0074]

[0075] in, The overall efficiency of human-machine collaboration To contribute to the efficiency of experts. To contribute to the efficiency of AI The synergistic effect of human-computer interaction, , , These are weighting coefficients, which are dynamically adjusted according to different R&D scenarios.

[0076] The aforementioned virtual physical system for industrial product R&D includes: a physical foundation layer, storing a knowledge base that digitally represents physical laws in the real physical world; a basic support layer, deployed on top of the physical foundation layer, providing data processing capabilities based on the physical foundation layer; an R&D logic model layer, deployed on top of the basic support layer, providing standardized product R&D process templates, the process templates defining multiple R&D stages from requirements to testing and verification, and the mapping relationships between each R&D stage; an intelligent agent collaboration layer, deployed on top of the basic support layer, including a chief intelligent agent and multiple specialized intelligent agents, the chief intelligent agent being used to schedule the specialized intelligent agents to collaboratively execute R&D sub-tasks; and a product engineering layer, deployed on top of the intelligent agent collaboration layer. The system is used to transform the output of the intelligent agent collaboration layer into a target result that conforms to engineering specifications. The target result is represented by a three-dimensional digital model. The application scenario layer, deployed above the product engineering layer, is used to receive product requirement information and display the target result. The system includes a downlink data flow mode and an uplink data flow mode. In the downlink data flow mode, the product requirement information flows from the application scenario layer to the physical foundation layer to obtain physical calculation results. In the uplink data flow mode, the physical calculation results flow from the physical foundation layer to the application scenario layer and are transformed into the target result. Through a six-layer architecture and bidirectional data flow, a complete closed loop from requirement to verification is realized in the digital space, shortening the R&D cycle and improving R&D efficiency. Attached Figure Description

[0077] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0078] Figure 1 This is a block diagram of a virtual physical system used for industrial product development in one embodiment;

[0079] Figure 2 This is a schematic diagram of the interaction logic of the system in one embodiment;

[0080] Figure 3 This is a schematic diagram of the system's interaction logic in another embodiment;

[0081] Figure 4 This is a block diagram of an industrial product development process in one embodiment;

[0082] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0084] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0085] In one exemplary embodiment, such as Figure 1 As shown, this application provides a virtual physical system for industrial product research and development, which includes: a physical foundation layer, a basic support layer, a research and development logic model layer, an intelligent agent collaboration layer, a product engineering layer, and an application scenario layer.

[0086] The basic support layer is deployed on top of the physical foundation layer, the R&D logic model layer is deployed on top of the basic support layer, the intelligent agent collaboration layer is deployed on top of the basic support layer, the product engineering layer is deployed on top of the intelligent agent collaboration layer, and the application scenario layer is deployed on top of the product engineering layer.

[0087] The physics foundation layer, the lowest level of the system, stores a knowledge base that digitally represents the physical laws of the real physical world. This knowledge base includes governing equations describing the laws of motion of matter, mathematical modeling tools, and a library of basic algorithms. The physics foundation layer ensures the scientific rigor and accuracy of all calculations and deductions performed at higher levels of the system. The real physical world refers to the objective physical reality in which humans live, encompassing the three elements of matter, energy, and information. Digital representation refers to the transformation of physical laws into mathematical expressions that are recognizable and computable by computers, including but not limited to partial differential equations, algebraic equations, numerical calculation models, and simulation algorithms. The knowledge base of the physics foundation layer is the cornerstone of the entire system, ensuring that all calculations and deductions at higher levels conform to physical laws.

[0088] The virtual physical system in this application constructs a virtual digital twin system corresponding to the real physical world by mapping the three major components of the real physical world to three digital dimensions of the digital physical world. Specifically, the real physical world consists of three major elements: matter, energy, and information. Matter is the carrier of physical entities, energy is the measure of matter's motion, and information is the description of matter's state; together, they constitute the complete system of the real physical world. The virtual physical system maps the three major elements of the real physical world to three digital dimensions: the physical dimension, the carrier dimension, and the human dimension. The physical dimension corresponds to the digital expression of energy elements in the real physical world, manifested in the physical foundation layer of the digital physical world as physical models and simulation algorithms used to describe the laws of matter's motion and perform physical field simulation calculations. The carrier dimension corresponds to the digital expression of matter elements in the real physical world, manifested in the product engineering layer as a three-dimensional digital model of the target outcome, used to virtualize product entities in the real physical world. The human dimension corresponds to the digital expression of information elements in the real physical world, manifested in the intelligent agent collaboration layer as various professional intelligent agents and the human-computer interaction interface of the application scenario layer, used to perform intelligent decision-making and human-computer collaboration.

[0089] In some optional embodiments, the virtual physical system maintains a set of states of the digital physical world in the time and space dimensions. Each state element in the state set represents a combination of matter, energy, and information at a specific spatiotemporal location, used to describe the complete physical state of the system at any time and any location. The state set supports real-time updates and predictive deduction to achieve an accurate description of the real physical world in all spatiotemporal dimensions.

[0090] Furthermore, the virtual physical system and the real physical world form a continuously evolving closed-loop system through bidirectional interaction. Specifically, the application scenario layer receives physical parameter inputs from the real physical world, including one or more of temperature, stress, and velocity, which are entered into the digital physical world through sensors or manual input. The virtual physical system calculates the system state based on the physical model and data, and the system state is used to predict and guide the behavior of physical products in the real physical world. The application scenario layer receives human operation instructions (including design operations, optimization operations, and decision-making operations) in the digital physical world. The operation instructions affect physical products in the real physical world in real time, and the feedback data from the real physical world in turn corrects the model of the digital physical world, forming a human-machine symbiotic closed loop, enabling the prediction accuracy and decision-making ability of the virtual physical system to continuously evolve. "Bidirectional interaction" means that the data flow between the digital physical world and the real physical world is bidirectional: on the one hand, physical parameters and human needs from the real physical world are input into the digital physical world through the application scenario layer (upstream data flow); on the other hand, design schemes and optimization decisions generated by the digital physical world based on physical calculations and intelligent inference are output to the real physical world through the application scenario layer (downstream data flow). The data flows in both directions form a closed loop and continue to circulate.

[0091] The foundational support layer, deployed above the physical foundation layer, provides data processing capabilities based on the physical foundation layer. It acts as a bridge connecting underlying science and upper-layer applications, deeply integrating physical laws with AI algorithms to provide upper-layer intelligent agents with computing power, data processing capabilities, and model inference capabilities. Specifically, the foundational support layer includes a large language model, a data engineering lab, a cloud-native architecture, and a physics solver interface. The large language model, based on the Transformer architecture, encodes and decodes input text through an attention mechanism, performing semantic understanding and structured transformation of unstructured product requirement information, converting it into structured technical indicators and design constraints. The data engineering lab performs standardization, cleaning, alignment, and feature extraction on multi-source heterogeneous data to eliminate data silos and achieve unified management and access to multi-source data. The cloud-native architecture enables elastic scheduling of computing resources, automatic service deployment, and dynamic load balancing, ensuring stable system operation under high-concurrency computing tasks. The physics solver interface responds to upper-layer requests, invoking numerical solvers in the physical foundation layer to discretize and solve the control equations, and returning the results to the upper layer.

[0092] The R&D logic model layer, deployed above the foundational support layer, provides standardized product development process templates. These templates define multiple R&D stages from requirements to testing and verification, as well as the mapping relationships between these stages. The R&D logic model layer structures and modularizes the complex industrial R&D process, ensuring a complete closed loop from requirements to verification through a V-shaped bidirectional mapping. The mapping relationship refers to the bidirectional traceability link between the decomposition stage on the left and the integration and verification stage on the right in the R&D process. This means that the requirements and design decisions of each stage on the left can be traced back to the verification results of the corresponding stage on the right. When verification fails, the process can be traced back to the corresponding design stage for correction. The R&D logic model layer provides standardized R&D process templates, making R&D processes reusable and traceable across different industries.

[0093] The agent collaboration layer is deployed above the basic support layer and includes a chief engineer agent and multiple specialized agents. The chief engineer agent is used to schedule the various specialized agents to collaboratively execute R&D sub-tasks. The agent collaboration layer is the intelligent brain of the system. Each specialized agent has autonomous decision-making capabilities and can independently complete specific R&D tasks. An agent refers to a software entity with autonomous decision-making capabilities. It maintains its own state space, action space, and decision-making strategy, and can perceive the environmental state, autonomously select actions, and execute tasks. The decision-making strategy selects the optimal action based on the value function iteration method to maximize the cumulative reward. The chief engineer agent is the master agent in the agent collaboration layer. It is responsible for receiving R&D tasks from the upper layer, identifying the capability tags of each specialized agent, decomposing complex tasks into several sub-tasks that can be executed in parallel or sequentially, and dynamically arranging the execution order and data routing paths of each specialized agent according to the real-time load status of each specialized agent. Specialized intelligent agents are intelligent agents with specific professional capabilities, including but not limited to demand analysis intelligent agents, physical modeling intelligent agents, multiphysics simulation intelligent agents, simulation result analysis intelligent agents, testing and verification intelligent agents, and digital twin construction intelligent agents. Each specialized intelligent agent focuses on performing a specific type of R&D sub-task.

[0094] The Product Engineering layer, deployed above the Intelligent Agent Collaboration layer, transforms the output of the Intelligent Agent Collaboration layer into target deliverables that conform to engineering specifications. These target deliverables are represented by three-dimensional digital models. The Product Engineering layer directly addresses the R&D needs of specific industrial products, transforming underlying intelligent capabilities into R&D tools and methods tailored to specific products. Engineering specifications refer to the technical standards, industry norms, and internal design guidelines that must be followed in industrial product R&D, including but not limited to product dimensional tolerance standards, material selection specifications, manufacturing process standards, quality acceptance standards, and industry safety specifications. Target deliverables are the deliverables that conform to engineering specifications, output by the Product Engineering layer. Their forms include, but are not limited to, three-dimensional digital models of the product, detailed design schemes, simulation reports, manufacturing process documents, quality assurance plans, and cost analysis reports. A three-dimensional digital model is a virtual representation of a physical entity in the real physical world within digital space. It includes the product's geometric structure information, dimensional tolerance information, material property information, and the assembly relationships and interface specifications between various components.

[0095] The application scenario layer is deployed above the product engineering layer to receive product requirement information and display target results. Located at the top layer of the system, it directly faces end users and encapsulates all underlying capabilities into industry solutions. Users can access all functions of the digital-physical world through the application scenario layer's interface. Product requirement information refers to the R&D needs of the target product input by users through the application scenario layer's interface. Input formats include natural language descriptions and structured data. Natural language descriptions express product functional requirements and performance indicators in colloquial or written language, such as designing a high-efficiency photovoltaic inverter. Structured data consists of technical indicators and constraints input in the form of tables, parameter lists, or standardized documents, such as efficiency target values, cost budgets, and operating temperature ranges.

[0096] The system includes a downlink data flow mode and an uplink data flow mode. In the downlink data flow mode, product demand information flows from the application scenario layer to the physical infrastructure layer to obtain physical calculation results. In the uplink data flow mode, the physical calculation results flow from the physical infrastructure layer to the application scenario layer and are transformed into target results.

[0097] Specifically, the downlink data flow model is a demand-driven data flow direction: users submit R&D requirements through the application scenario layer, the requirements are broken down into engineering tasks by the product engineering layer, then the intelligent agent collaboration layer calls the corresponding intelligent agents to execute them, then the R&D logic model layer organizes the process according to the V-model, then the basic support layer allocates computing resources and models, and finally the physical foundation layer performs physical calculations and AI inference. The uplink data flow model is a capability-supported data flow direction: the calculation results of the physical foundation layer are integrated and processed by the basic support layer, then verified and optimized by the R&D logic model layer, then multi-agent collaborative inference is performed by the intelligent agent collaboration layer, then the product engineering layer generates engineering solutions, and finally the application scenario layer outputs the final results.

[0098] This embodiment uses a six-layer architecture for bidirectional data flow, enabling product requirement information to flow step by step from the application scenario layer to the physical foundation layer to obtain physical calculation results, and then flow step by step back to the application scenario layer and be transformed into target results. This completes the entire R&D closed loop from requirement input to result output in the digital space, solving the problems of information silos and process fragmentation in traditional R&D methods.

[0099] In one exemplary embodiment, such as Figure 2 The diagram illustrates the system's interaction logic. System interaction follows a two-way data flow pattern: top-down requirement communication and bottom-up capability support. In the downlink data flow, users submit development requirements at the application scenario layer, which are then passed down level by level. Each layer processes the requirements accordingly: the product engineering layer breaks down tasks, the intelligent agent collaboration layer schedules intelligent agents for execution, the development logic model layer organizes the development process, the basic support layer allocates computing resources, and the physical infrastructure layer performs physical calculations. In the uplink data flow, physical calculation results are passed up level by level, with each layer processing the data: the basic support layer integrates the data, the development logic model layer verifies and optimizes it, the intelligent agent collaboration layer performs joint simulations, the product engineering layer generates engineering solutions, and the application scenario layer outputs and displays the results to the user.

[0100] In some of the alternative embodiments, the combination continues. Figure 2 The downlink data flow pattern includes: the application scenario layer receives product requirement information and then transmits it to the product engineering layer; the product engineering layer breaks down the product requirement information into engineering tasks and then transmits them to the intelligent agent collaboration layer; the intelligent agent collaboration layer schedules the corresponding professional intelligent agents to execute the engineering tasks through the chief engineer intelligent agent and then transmits the processing results to the R&D logic model layer; the R&D logic model layer organizes the R&D sequence according to the process template and then transmits it to the basic support layer; the basic support layer allocates computing resources and calls the knowledge base in the physical foundation layer to perform physical field numerical calculations to obtain physical calculation results.

[0101] Specifically, users input product requirement information through the interactive interface of the application scenario layer. This information includes unstructured natural language descriptions and structured technical specifications. The application scenario layer then formats and encapsulates the received product requirement information and transmits it to the product engineering layer through standardized inter-layer data interfaces.

[0102] After receiving product requirement information, the product engineering layer calls the architecture and platform definition module to parse the information, transforming it into specific executable engineering task units. An engineering task is a concrete execution unit derived from breaking down the product requirement information according to the product development process; each engineering task corresponds to a subtask of one or more specialized intelligent agents. The product engineering layer packages these decomposed engineering tasks and transmits them to the intelligent agent collaboration layer through inter-layer interfaces.

[0103] After receiving the engineering task, the chief engineer's intelligent agent parses the engineering task and identifies the capability tags and dependencies required for each engineering task. Then, the chief engineer's intelligent agent assigns a suitable professional intelligent agent to each subtask based on the capability tags and real-time load status of each professional intelligent agent. Each professional intelligent agent executes its own subtask in parallel or serially and generates intermediate processing results. The chief engineer's intelligent agent collects the intermediate processing results of all professional intelligent agents, summarizes them, and transmits them to the R&D logic model layer through the inter-layer interface.

[0104] After receiving the processing results from the intelligent agent collaboration layer, the R&D logic model layer organizes the results into a R&D sequence that conforms to the standard R&D process, according to the stage divisions and mapping relationships between each stage in the V-shaped R&D process template. The R&D sequence is a sequence of executable steps arranged according to the order and dependencies of each R&D stage defined in the V-shaped R&D process template. The R&D logic model layer then transmits the organized R&D sequence to the basic support layer.

[0105] After receiving the R&D sequence, the resource management unit dynamically allocates corresponding computing resources based on the computational requirements of each step in the R&D sequence (including the amount of computation, whether GPU acceleration is required, whether large-scale parallel computing is required, etc.). The physics solver interface transforms the physics computation requests in the R&D sequence into solution tasks that can be executed by the physics foundation layer, and calls the control equations and numerical solvers in the physics foundation layer knowledge base to perform physical field numerical calculations. After the physics foundation layer completes the calculation, it returns the physics computation results to the basic support layer.

[0106] This embodiment uses five sequentially executed data transfer steps to transmit product demand information from the application scenario layer to the physical infrastructure layer and complete physical calculations. This achieves a complete closed loop of demand-driven downlink data flow, ensuring orderly and controllable data transfer between layers and avoiding information silos and data inconsistencies caused by human intervention.

[0107] In some of these alternative embodiments, combined with Figure 3 The diagram illustrates another interaction logic of the system. The uplink data flow pattern includes: the physical foundation layer transmits the physical calculation results to the basic support layer for data integration; the basic support layer transmits the integrated data to the R&D logic model layer for stage verification and optimization; the R&D logic model layer transmits the verified and optimized data to the intelligent agent collaboration layer for multi-agent joint inference; the intelligent agent collaboration layer transmits the inference results to the product engineering layer to generate the target result; the product engineering layer transmits the target result to the application scenario layer, and the application scenario layer outputs the target result.

[0108] Specifically, after completing the numerical calculations of the physical fields, the physics foundation layer generates raw physical calculation result data. These results include, but are not limited to: physical field distribution data (such as the numerical distribution of temperature, stress, flow, and electromagnetic fields at various points in space), performance index data (such as energy conversion efficiency, structural strength coefficient, fatigue life, etc.), and process data generated during simulation. The physics foundation layer transmits the physical calculation results to the foundation support layer through inter-layer interfaces. Upon receiving the physical calculation results, the foundation support layer integrates the multi-source heterogeneous raw calculation results through the data processing pipeline. This includes aligning timestamps of different physical field data, unifying spatial coordinates, standardizing data formats, and removing outliers, thus integrating the scattered raw calculation results into a structured, unified data format.

[0109] After the basic support layer completes data integration, it transmits the integrated data to the R&D logic model layer. Upon receiving the data, the R&D logic model layer performs stage verification on the integrated data according to the verification functions for each R&D stage in the V-shaped R&D process template. Stage verification involves checking the matching degree between the physical calculation results and the preset requirement indicators and specifications for the current R&D stage to determine whether the design scheme meets the design requirements. If the verification passes, the process proceeds to the next step; if the verification fails, optimization direction information is generated based on the verification results, the design scheme is optimized and adjusted, and the downstream data flow is triggered to re-execute the simulation calculation.

[0110] After the development logic model layer completes its phase verification and optimization, the verified data is transmitted to the agent collaboration layer. Upon receiving the data, the chief engineer agent in the collaboration layer schedules multiple related professional agents to jointly extrapolate the data. Multi-agent joint extrapolation refers to multiple professional agents collaborating on the same dataset based on their respective domain knowledge. Each professional agent processes the data in parallel and exchanges intermediate results through standardized data interfaces between agents, ultimately forming a comprehensive extrapolation conclusion.

[0111] After the intelligent agent collaboration layer completes the joint simulation, it transmits the simulation results to the product engineering layer. Upon receiving the simulation results, the product engineering layer transforms them into target deliverables that conform to engineering specifications. Transforming them into target deliverables that conform to engineering specifications means that the product engineering layer generates specific design schemes, engineering drawings, simulation reports, or process documents based on the simulation results, and performs formatting and compliance checks according to industry standards and enterprise specifications to ensure that all output deliverables meet actual engineering requirements.

[0112] After the product engineering layer generates the target deliverables, it transmits these deliverables to the application scenario layer via inter-layer interfaces. Upon receiving the target deliverables, the application scenario layer displays them to the user in a user-readable format using visualization units. These displays include, but are not limited to: graphical rendering of 3D models, color mapping of simulation cloud maps, charts of performance data, and structured presentations of solution reports.

[0113] This embodiment uses five sequentially executed data transfer steps to transmit physical calculation results from the physical infrastructure layer to the application scenario layer and transform them into readable target results. This achieves a complete closed loop for capability-supported uplink data flow, ensuring that physical calculation results can be presented to users in a form that conforms to engineering specifications.

[0114] In some optional embodiments, the knowledge base of the physics foundation layer includes: basic laws describing the real physical world based on the principle of least action, governing equations describing the laws of motion of matter, and different physical field models for multi-field coupling calculations; the governing equations include at least one of the continuity equation, the momentum conservation equation, and the energy conservation equation.

[0115] The physical foundation layer defines the basic laws of the physical world, provides mathematical modeling tools, computational frameworks, and basic algorithm libraries, and ensures the scientific rigor and accuracy of all calculations and deductions in the upper layers.

[0116] Specifically, the physics foundation layer describes the fundamental laws of the real physical world based on the principle of least action, which is as follows:

[0117]

[0118] in, For action, For Lagrange units, For generalized coordinates, For generalized speed, In terms of time, this principle is the unifying foundation of classical mechanics, quantum mechanics, and field theory, providing underlying physical laws to support the entire digital physics world. The principle of least action is the unifying foundation of classical mechanics, quantum mechanics, and field theory, providing underlying physical laws to support the entire digital physics world, ensuring that all physical simulations and calculations in the system conform to the fundamental laws of nature. Describing the fundamental laws of the real physical world based on the principle of least action means that the knowledge base uses the principle of least action as the highest-level physical law. All governing equations describing the laws of motion of matter are derived within the framework of the principle of least action, thus ensuring that the entire foundational knowledge system of physics is logically self-consistent and unified.

[0119] Governing equations are specific mathematical descriptions of the laws governing the motion of matter in the real physical world, including equations for the conservation of mass, momentum, and energy. The set of governing equations describes the physical behaviors of matter, such as motion, deformation, heat transfer, and flow, in both spatial and temporal dimensions. Each governing equation is stored in a discretized form at the physics foundation layer, and can be called by the physics solver to perform numerical calculations.

[0120] The mass conservation equation is:

[0121]

[0122] In this formula, Fluid density (unit: ), describes the mass distribution of the fluid, These are the velocity components of the fluid in the x, y, and z directions (unit: ...). ), For time (unit: This equation represents the conservation of mass during fluid motion, meaning that the mass flowing into any control volume per unit time is equal to the mass flowing out.

[0123] The momentum conservation equation is:

[0124]

[0125] In this formula, Pressure (unit: ), Fluid dynamic viscosity (unit: ), Volume force (such as gravity, electromagnetic force, etc., unit: ), For gradient operators, The Laplace operator describes the conservation of fluid momentum, meaning that fluid acceleration is determined by pressure gradient force, viscous force, and volume force.

[0126] The energy conservation equation is:

[0127]

[0128] in, The total energy per unit mass of fluid (internal energy + kinetic energy, unit: ), Thermal conductivity (unit: ), Temperature (unit: ), This is a viscous dissipation term, representing the portion of work done by viscous forces that is converted into heat energy. For heat source items (unit: For example, chemical reactions that release heat, electromagnetic heating, etc., this equation describes the conservation of fluid energy, that is, the increase in energy is equal to the sum of external work, heat conduction and internal heat sources.

[0129] Multi-field coupling refers to the mutual influence and interaction between two or more different physical fields, such as thermal convection coupling between flow and temperature fields, fluid-structure interaction between flow and structural mechanics fields, and electromagnetic-thermal coupling between electromagnetic and temperature fields. The physics foundation layer stores multiple physical field models. A physical field refers to the distribution of physical quantities in spatial and temporal dimensions, including but not limited to flow fields (describing the spatiotemporal distribution of fluid velocity, pressure, and density), temperature fields (describing the spatial and temporal distribution of temperature), electromagnetic fields (describing the spatiotemporal distribution of electric and magnetic field strengths), and structural mechanics fields (describing the spatiotemporal distribution of stress, strain, and displacement in solid structures). The physics foundation layer supports the establishment and joint simulation of coupling relationships between different physical fields.

[0130] This embodiment integrates fundamental laws, governing equations, and physical field models into the same knowledge base, ensuring that the physics foundation layer can provide unified, complete, and self-consistent physical knowledge support for all physical calculations and simulations in the upper layers.

[0131] In some optional embodiments, the basic support layer includes: a large language model, a data engineering laboratory, a cloud-native architecture, and a physical solver interface; the basic support layer is used to integrate physical laws with AI algorithms, providing computing power, data processing power, and model inference power to each agent in the agent collaboration layer, and the large language model includes an attention mechanism.

[0132] Large language models are industry-specific large language models, specifically those fine-tuned from general-purpose large language models like DeepSeek and OpenAI using industry-specific corpora. These models are used to semantically understand and structure unstructured user input (including product requirements, design intentions, optimization goals, etc., described in natural language), transforming it into structured technical indicators and design constraints. Large language models include an attention mechanism, the calculation formula for which is:

[0133]

[0134] in, For querying the matrix, The key matrix, For value matrices, The dimension of the key vector. The normalized exponential function is the core computational unit of the large language model, providing semantic understanding and reasoning capabilities for upper-level agents.

[0135] The Data Engineering Lab is used for standardization, cleaning, alignment, and feature extraction of multi-source heterogeneous data. Multi-source heterogeneous data includes data from different sources, with different formats and semantics, including but not limited to: real-time monitoring data from physical sensors (time-series data such as temperature, pressure, and vibration), offline measured datasets from historical databases, design documents and standard specifications from external systems, and intermediate output data from simulation processes. The Data Engineering Lab performs ETL (Extract-Transform-Load) processing on multi-source heterogeneous data, eliminating data silos, achieving unified management and access to multi-source data, and providing support for the data fusion capabilities of the basic support layer and the data processing needs of upper-layer intelligent agents.

[0136] Cloud-native architecture is used to achieve elastic scheduling of computing resources, automatic deployment of services, and dynamic load balancing, ensuring stable operation of the system under high-concurrency computing tasks and on-demand allocation of resources.

[0137] The physics solver interface is a standardized calling interface between the basic support layer and the physical foundation layer. It is used to respond to upper-layer calling requests, call the numerical solver in the physical foundation layer to discretize and solve the control equations, and return the calculation results to the upper layer.

[0138] The foundational support layer integrates physical mechanisms with AI algorithms, providing computational, data processing, and model inference capabilities to the agents in the agent collaboration layer. This integration means that the foundational support layer incorporates the physical laws (in the form of governing equations and physical constraints) from the physical foundation layer with AI algorithms (typically represented by large language models and data-driven algorithms) into a unified execution framework. This allows upper-layer agents to simultaneously benefit from both physically driven and data-driven computational modes when accessing the capabilities of the foundational support layer. Specific integration methods include, but are not limited to: using the residual terms of the governing equations as loss functions to constrain the training of neural networks (i.e., physically constrained neural network algorithms); expanding the training dataset with synthetic data generated from physical simulations; and using physical laws to validate the rationality of text generated by large language models.

[0139] This embodiment integrates a large language model, a data engineering laboratory, a cloud-native architecture, and a physical solver interface into a unified basic support layer, achieving a deep fusion of physical mechanisms and AI algorithms. It provides a unified service interface for upper-layer intelligent agents to access computing power, data processing capabilities, and model inference capabilities, reducing the complexity of upper-layer intelligent agents calling underlying capabilities.

[0140] In some optional embodiments, the process template in the R&D logic model layer is a V-shaped R&D process template. The V-shaped R&D process template includes a decomposition stage sequence and an integration verification stage sequence. Each decomposition stage forms a bidirectional traceability link with the corresponding integration verification stage. The bidirectional traceability link is used to backtrack to the corresponding decomposition stage for design correction when verification fails.

[0141] The V-shaped R&D process template is a standardized process template that visually represents the industrial product R&D process using a V-shaped structure. The V-shaped R&D process template includes a decomposition phase sequence and an integration and verification phase sequence. The decomposition phase sequence, located on the left side of the V, progressively breaks down product R&D requirements from top to bottom into multiple decomposition phases: system-level requirements, subsystem-level requirements, component-level requirements, and detailed design specifications. Specifically, the left-side decomposition phase sequence starts from the top left apex of the V and includes: product requirements definition phase, system architecture design phase, subsystem design phase, and component detailed design phase, with each phase progressively refining design specifications from macro to micro. The integration and verification phase sequence, located on the right side of the V, progressively integrates the design results from the decomposition phases into a complete product from bottom to top and performs multiple verification phases. Specifically, the right-side integration and verification phase sequence starts from the bottom right apex of the V and includes: unit testing phase, integration testing phase, system verification phase, and product acceptance phase, with each phase progressively verifying the functional integrity and performance compliance of the product from micro to macro.

[0142] A bidirectional traceability link is formed between each decomposition stage and its corresponding integration verification stage. This bidirectional traceability link refers to the mapping relationship established between each decomposition stage on the left and its corresponding integration verification stage on the right. This mapping relationship records the correspondence between the outputs of the decomposition stage (design specifications, parameters, etc.) and the inputs of the integration verification stage (test cases, acceptance criteria, etc.). For example, the outputs of the component detailed design stage correspond to the test cases of the unit testing stage, and the outputs of the system architecture design stage correspond to the verification plan of the system verification stage.

[0143] The bidirectional traceability link is used to backtrack to the corresponding decomposition stage for design correction when verification fails. Specifically, when the verification result of a certain integration verification stage on the right does not meet the design requirements (such as unit test failure, system verification failure, etc.), the development logic model layer automatically locates the corresponding decomposition stage on the left along the bidirectional traceability link, identifies the location of the design defect that caused the verification failure, and maps the deviation information in the verification result to specific design correction requirements, triggering the design correction process. After the correction is completed, it re-enters the corresponding verification stage along the same mapping relationship for re-verification, thus forming a closed loop of design-verification-backtracking-correction-re-verification until verification passes.

[0144] This embodiment achieves complete closed-loop management of product development from requirements to verification through the two-way traceability link of the V-shaped R&D process template. When verification fails, it can quickly trace back to the corresponding design stage for correction, which greatly reduces repetitive work and communication costs in the R&D process.

[0145] In some optional embodiments, the R&D efficiency of the R&D logic model layer is evaluated using the following function:

[0146]

[0147] in, The R&D efficiency of the V-shaped R&D process template is a dimensionless quantitative evaluation value used to measure the overall efficiency and quality level of the R&D process. This represents the total number of development stages, specifically the number of stages included in the V-shaped development process template. Let be the verification function for the requirements and specifications of the i-th R&D phase. It is used to evaluate the degree of matching between the actual design specifications and the requirement indicators in this phase. The value is between 0 and 1, where 1 represents a perfect match and 0 represents a complete mismatch. The requirements for the i-th development phase include the functional requirements, performance indicators, and technical constraints that this phase should meet. This refers to the specifications for the i-th research and development phase, including the specific design parameters, technical specifications, and implementation plan for that phase. is the weight coefficient for the i-th R&D stage, used to measure the difference in importance of different R&D stages to the overall R&D efficiency, and the sum of the weight coefficients of each stage is 1.

[0148] This function quantifies the closed-loop quality of the V-model from requirements to verification. It evaluates the efficiency of the entire R&D process by weighting and averaging the matching degree between requirements and specifications at each stage. A higher R&D efficiency value indicates a higher degree of requirement-specification matching at each stage of the R&D process, and better R&D efficiency and quality.

[0149] This embodiment quantifies the execution effect of the V-shaped R&D process template through an R&D efficiency evaluation function, providing a quantitative basis for the continuous optimization of the R&D process.

[0150] In some alternative embodiments, the multiple specialized intelligent agents include at least one of the following: a demand analysis intelligent agent, a multiphysics processing intelligent agent, and a physical information augmentation intelligent agent.

[0151] The requirement analysis agent is used to analyze product requirement information to obtain at least one initial design scheme. Specifically, the requirement analysis agent generates the initial design scheme through the following steps: First, the requirement analysis agent receives product requirement information from the product engineering layer and transforms the unstructured product requirement information into a structured technical indicator system; second, based on the industry knowledge base and historical project data, the requirement analysis agent automatically supplements the boundaries and constraints of the technical indicators (such as efficiency requirements, cost targets, environmental conditions, etc.); then, the conceptual design agent (as a component of the requirement analysis agent or an associated agent working under the scheduling of the requirement analysis agent) retrieves similar product cases from the product database and combines them with the basic physical laws in the physical foundation layer knowledge base to generate at least one conceptual design scheme as the initial design scheme. At least one scheme can be one or more, and the specific number is flexibly determined by the requirement analysis agent based on the complexity of the design space and the user-specified requirements.

[0152] A multiphysics processing agent is used to determine the physical field model corresponding to the initial design scheme and establish multiphysics layer coupling relationships. Based on these coupling relationships, the initial design scheme is simulated to obtain simulation results. Specifically, after receiving the initial design scheme, the multiphysics processing agent matches the required physical field model from the knowledge base of the physical foundation layer according to the product characteristics of the initial design scheme (such as product type, working environment, performance requirements, etc.). (Matching refers to retrieving and selecting the physical field model and corresponding control equations and solution parameters that most accurately describe the physical behavior of the product from the knowledge base based on the physical attributes of the design scheme.) It also establishes coupling relationships between different physical fields. Then, based on the established physical field model and multiphysics coupling relationships, the multiphysics processing agent performs high-precision simulation calculations on the initial design scheme. The specific content of the simulation calculations includes, but is not limited to, electromagnetic field distribution simulation, heat conduction simulation, structural mechanics simulation, and multiphysics coupling simulation.

[0153] The physical information augmentation agent is used to identify problems in the simulation results and optimize the design scheme. The optimized design scheme is then fed back to the multiphysics processing agent for re-simulation until the simulation results meet the design requirements, thus obtaining the target design scheme. Specifically, after receiving the simulation results output by the multiphysics processing agent, the physical information augmentation agent performs intelligent analysis on the simulation results (including post-processing of the multidimensional data field output by the simulation, extraction of key performance indicators, identification of weak performance areas and design bottlenecks), identifies problems in the simulation results (such as local overheating, stress concentration, low efficiency, and other issues that do not meet the design requirements), and generates specific optimization suggestions based on the identified problems (including but not limited to adjusting design parameters, modifying material selection, changing structural layout, and other optimization directions). The physical information augmentation agent feeds back the optimized design scheme to the multiphysics processing agent, triggering the multiphysics processing agent to update the design parameters in each physical mathematical model and re-perform simulation calculations, forming a design-simulation-analysis-optimization-re-simulation cycle until the simulation results meet all design indicators, thus obtaining the target design scheme.

[0154] The execution order of the above-mentioned professional intelligent agents is as follows: first, the demand analysis intelligent agent performs the task; then, the output is passed to the multiphysics processing intelligent agent to perform simulation; finally, the physical information enhancement intelligent agent performs analysis and optimization, forming a complete intelligent agent collaborative link from demand analysis to simulation optimization and then to solution output.

[0155] This embodiment achieves a complete agent chain from demand input to design output through the sequential collaboration of a demand analysis agent, a multiphysics processing agent, and a physical information enhancement agent. Intermediate results are transmitted between the agents through standardized data interfaces, automatically completing the entire process of design generation, simulation calculation, and optimization iteration without the need for manual switching between multiple tools and data conversion.

[0156] In some alternative embodiments, the multiple specialized intelligent agents also include: a test and verification intelligent agent and a target operating condition simulation intelligent agent.

[0157] The test and verification agent is used to acquire a virtual test bench, test and verify the target design scheme on the virtual test bench, record the performance information under various working conditions, obtain the test results based on the performance information, and modify the target design scheme if the test results fail. The agent then triggers a multiphysics processing agent to re-simulate the modified target design scheme and re-test and verify the re-simulated design scheme until the test passes. The virtual test bench is consistent with the real physical test equipment.

[0158] Specifically, the test verification agent first acquires a virtual test bench, a virtual testing environment built in the digital-physical world with parameters identical to those of real physical testing equipment. This virtual test bench simulates the real-world testing environment to perform various performance tests on the product design. The test verification agent then applies various preset test conditions to the target design within the virtual test bench, conducting virtual tests. During the test, it records the product's performance information (including measured values ​​of key performance indicators, response curves, failure modes, etc.) in real time under each condition. Based on the recorded performance information, it compares it with the preset performance indicators to determine whether each performance characteristic meets the design requirements, thus obtaining the test results. Each condition refers to a different combination of working conditions applied during the test verification process, including normal working conditions and extreme working conditions.

[0159] If the test results fail (i.e. certain performance indicators do not meet the design requirements), the test verification agent automatically corrects the target design scheme (specific correction methods include adjusting design parameters, modifying material selection, or improving structural layout), and triggers the multiphysics processing agent to re-simulate the corrected target design scheme, and re-tests and verifies the re-simulated design scheme until the test passes.

[0160] The target operating condition simulation agent is used to configure various target operating conditions on the virtual test bench. The target operating conditions include extreme operating conditions and failure operating conditions.

[0161] Specifically, the target operating condition simulation agent applies various preset target operating conditions in a virtual test bench to test and verify product performance. Extreme operating conditions refer to extreme working conditions that a product rarely encounters during normal use but could potentially lead to product failure if encountered. These include, but are not limited to, temperature cycling tests (cyclic changes between extremely low and extremely high temperatures), vibration tests (applying mechanical vibrations of different frequencies and amplitudes), electromagnetic compatibility tests (applying electromagnetic interference of different intensities and frequencies), load surge tests (simulating sudden increases or decreases in load), and accelerated life testing (accelerating the product aging process by increasing working stress). Fault operating conditions refer to various fault states that a product may encounter during operation, including but not limited to component failure, heat dissipation failure, and sensor drift failure. The target operating condition simulation agent applies extreme operating conditions to verify the product's reliability and safety under extreme conditions, and applies fault operating conditions to deduce the propagation path and impact range of faults, verifying the product's fault tolerance and fault recovery capabilities.

[0162] This embodiment verifies that the intelligent agent and the target working condition simulation intelligent agent are tested on a virtual test bench. This replaces the traditional physical prototype testing in a digital way, which greatly reduces the testing cost, shortens the testing cycle, and enables the safe study of extreme working conditions and failure conditions that are difficult to reproduce or have safety risks in real physical experiments in the digital space.

[0163] In some alternative embodiments, the multiple specialized intelligent agents also include: a process design intelligent agent, a manufacturing simulation intelligent agent, a quality pre-assessment intelligent agent, and a process optimization intelligent agent.

[0164] The process design agent is used to generate manufacturing process parameters based on a target design scheme that has passed testing. Specifically, the process design agent receives a target design scheme that has passed testing and verification, and automatically generates manufacturing process parameters based on the 3D model of the product and material properties in the target design scheme. Manufacturing process parameters are specific parameters that guide the product manufacturing process, including but not limited to: machining path (such as tool movement trajectory and machining sequence arrangement), process parameters (such as machining temperature, applied pressure, feed rate, cooling rate, etc.), and assembly sequence (the assembly order and mating method of each component).

[0165] Manufacturing simulation agents are used to digitally simulate the manufacturing process based on manufacturing process parameters to obtain manufacturing simulation results. Specifically, the manufacturing simulation agent receives manufacturing process parameters generated by the process design agent and digitally simulates the entire manufacturing process in digital space. The manufacturing process refers to the entire process of gradually transforming raw materials or semi-finished products into the final product according to design drawings and process requirements, including but not limited to machining processes (such as cutting, milling, grinding, etc.), welding processes (such as arc welding, laser welding, resistance welding, etc.), and assembly processes (assembly of various components). The manufacturing simulation agent digitally simulates each of the above manufacturing stages and generates manufacturing simulation results.

[0166] The quality pre-assessment agent is used to derive predicted quality indicators based on manufacturing simulation results and compare them with target quality indicators in the product requirement information to obtain quality comparison results. Specifically, the quality pre-assessment agent receives the manufacturing simulation results output by the manufacturing simulation agent and predicts the product's quality indicators based on these results. The predicted quality indicators are quantitative indicators of product quality predicted based on the manufacturing simulation results, including but not limited to dimensional accuracy (the degree of deviation between the actual and designed dimensions of the product), surface roughness (the microscopic unevenness of the product surface), and mechanical properties (such as hardness, strength, toughness, and other mechanical performance parameters). The quality pre-assessment agent compares the predicted quality indicators with the target quality indicators in the product requirement information item by item, calculates the deviation value of each indicator and the overall matching degree, and obtains the quality comparison results.

[0167] The process optimization agent is used to adjust process parameters and / or process routes based on quality comparison results until the product quality indicators obtained based on manufacturing simulation results meet the product quality indicators in the product requirement information. Specifically, after receiving the quality comparison results, if the predicted quality indicators do not meet the requirements of the target quality indicators, the process optimization agent analyzes the causes of the quality deviation and adjusts the process parameters (such as adjusting the processing temperature, feed rate, etc.) and / or process routes (such as changing the processing sequence, adding or reducing processing steps, etc.) accordingly. Then, it triggers the manufacturing simulation agent to re-perform manufacturing simulation based on the adjusted process parameters and / or process routes, and the quality pre-assessment agent performs quality assessment again, forming a cycle of simulation → assessment → optimization → re-simulation until the predicted quality indicators meet the requirements of the target quality indicators.

[0168] This embodiment achieves a complete transformation from design scheme to manufacturing process in the digital space through the sequential collaboration of process design intelligence, manufacturing simulation intelligence, quality pre-assessment intelligence, and process optimization intelligence. By using process simulation and virtual quality assessment, quality problems in the manufacturing process can be identified and resolved in advance, avoiding trial and error and rework in real manufacturing.

[0169] In some optional embodiments, the multiple specialized intelligent agents also include: a digital twin intelligent agent, used to acquire product operation data collected from physical products in the real physical world, synchronize the product operation data to the digital twin, and perform monitoring and early warning based on the status data of the digital twin.

[0170] Specifically, firstly, the digital twin agent creates a corresponding digital twin of a physical product that has already been manufactured in the real physical world, and injects all data from the design and manufacturing stages (including 3D models, material parameters, process records, etc.) into the digital twin model. The digital twin is a virtual mirror model of the physical product in the digital physical world, and a one-to-one mapping relationship is established between it and the physical product. There is a pre-established mapping relationship between the state variables of the digital twin and the physical state parameters of the corresponding measurement points on the physical product. Secondly, after the product is deployed and running, the digital twin agent acquires the product operation data collected by the physical product in real time through IoT interfaces and sensor networks. The product operation data consists of real-time state parameters collected by various sensors deployed on the physical product during actual operation, including but not limited to temperature data, vibration data, current data, and voltage data. The digital twin agent receives the above product operation data at a preset time frequency (e.g., once per second) and synchronizes the received product operation data to the corresponding state variables in the digital twin, so that the digital twin and the physical product maintain state synchronization.

[0171] The digital twin, based on a physical model and real-time input data, continuously calculates the product's real-time status and compares the results with the actual operating data of the physical product. The digital twin intelligent agent performs real-time monitoring and trend analysis based on the digital twin's status data, specifically including: real-time monitoring of the digital twin's status data to assess the product's current health; predictive analysis of status data trends to identify potential performance degradation trends; and fault warnings when abnormal states are detected. The digital twin intelligent agent also transmits synchronized status data to a visualization unit, graphically displaying the digital twin's current and predicted status in real time, and supporting remote access and simultaneous monitoring of multiple products.

[0172] This embodiment uses a digital twin intelligent agent to create a one-to-one virtual image of the physical product, and synchronizes the operating data of the physical product in real time. This enables remote real-time monitoring and health management of the physical product's operating status, and can issue early warnings before a fault occurs, thus avoiding unplanned downtime and safety accidents.

[0173] In some alternative embodiments, the multiple specialized intelligent agents also include: an intelligent optimization agent, a knowledge accumulation agent, and a product iteration agent.

[0174] The intelligent optimization agent is used to generate product optimization suggestions based on accumulated operational data and user feedback. Specifically, the intelligent optimization agent collects operational data (including fault data, maintenance records, performance degradation curves, etc.) and user feedback information throughout the product's entire lifecycle, analyzes the above data, identifies potential opportunities for product performance improvement and cost reduction, and generates optimization suggestions for the next generation of the product (such as specific optimization directions and measures such as improving efficiency, reducing costs, and extending lifespan).

[0175] The knowledge-accumulating agent is used to extract failure modes, performance degradation patterns, and optimization opportunities from operational data throughout the product's entire lifecycle, and to update the knowledge base. Specifically, the knowledge-accumulating agent extracts valuable knowledge from the operational data throughout the product's lifecycle, including Failure Mode and Effects Analysis (FMEA, i.e., a systematic analysis of possible failure modes of the product and their impact on system functionality), performance degradation patterns (the trend and rate of gradual decline in product performance indicators over time), and optimization opportunity identification (improvement directions discovered in the product's operational data). This extracted knowledge is then updated in the physical foundation layer's knowledge base, while simultaneously updating the industry knowledge base and historical project database, providing a reference for subsequent product development's conceptual design, simulation analysis, and process optimization.

[0176] The product iteration agent triggers a new round of R&D cycle, starting with optimization suggestions and moving into the requirements definition phase, forming a continuously iterative R&D closed loop. Specifically, after the intelligent optimization agent generates optimization suggestions, the product iteration agent automatically initiates a new R&D cycle, inputting these suggestions as new product requirement information into the requirements definition phase. This triggers the requirements analysis agent to conduct a new round of requirements analysis and conceptual design, thus forming a continuous closed loop of operation-analysis-optimization-iteration. The knowledge accumulated in each iteration feeds back into the knowledge base and AI models of the physical foundation layer, continuously improving the prediction accuracy and intelligent decision-making capabilities of the digital physical world.

[0177] This embodiment achieves automatic accumulation and continuous iteration of knowledge throughout the product lifecycle through the collaboration of intelligent optimization agents, knowledge accumulation agents, and product iteration agents, enabling the system to continuously evolve as operational data accumulates.

[0178] In some optional embodiments, the decision-making strategy of each specialized agent is based on a value function iteration method to select the optimal action. This method iteratively updates the value function of the state, enabling the agent to select the action sequence with the highest cumulative reward within a complex research and development space. The optimal value function of the value function iteration method is:

[0179]

[0180] in, Let be the optimal value function for state s, representing the maximum cumulative reward obtainable from state s; For the immediate reward of performing action a in state s, represent the return from state s. The maximum cumulative reward that can be obtained by starting is Discount factor ( The discount factor is used to measure the importance of future returns. The closer the discount factor is to 1, the more the agent values ​​long-term returns; the closer it is to 0, the more the agent values ​​immediate returns. Let be the state transition probability, representing the probability of transitioning to state s' after performing action a in state s. This probability characterizes the uncertainty of the environmental response. The optimal value function iteratively updates the state value function using the Bellman optimality equation until it converges to the optimal value function. The agent selects the action that maximizes the value function in each state based on the optimal value function.

[0181] In this system, each specialized agent selects its optimal action based on a value function iterative method. Specifically, each specialized agent maintains its own state space, action space, and decision strategy. The state space refers to the set of all environmental states that the specialized agent may perceive during task execution. For example, the state space of a simulation agent includes current design parameters, simulation progress, intermediate results, etc. The action space refers to the set of all actions that the specialized agent can perform in each state. For example, the actions of a simulation agent include adjusting mesh density, modifying convergence criteria, switching solvers, etc. The decision strategy refers to the mapping rules from state to action for the specialized agent, i.e., which action to perform given a state.

[0182] This embodiment utilizes the optimal decision-making strategy based on the value function iteration method for each specialized intelligent agent, enabling each agent to autonomously select the optimal action path in a complex R&D space, thereby achieving an improvement in capabilities from passively executing instructions to proactively optimizing decisions.

[0183] In some optional embodiments, the simulation is performed using a neural network, with the total loss function of the neural network being:

[0184]

[0185] Let be the total loss of the neural network, which is the objective function to be minimized during training. The data-driven loss measures the deviation between the neural network's predictions and the actual measured data. When historical measured data or experimental data is available, this loss term drives the neural network's predictions to be as close to the actual data as possible. The physical constraint loss measures whether the neural network's predictions satisfy the governing equations. This loss term is obtained by calculating the residuals of the governing equations at random sampling points within the computational domain and taking the mean square error, thus driving the neural network to learn physical laws. Boundary condition loss measures whether the neural network satisfies given conditions at the boundary (such as no-slip wall condition, isothermal boundary condition, adiabatic boundary condition, etc.). , , The weighting coefficients are used to adaptively adjust the contributions of the three losses, enabling the neural network to fit measured data while learning physical laws.

[0186] The simulation tools used by the multiphysics processing agent when simulating the initial design scheme include neural networks. The neural network is a Physically Informed Neural Network (PINN), which embeds physical constraints as part of the loss function into the training objective of the neural network, ensuring that the output of the neural network statistically satisfies the physical laws described by the governing equations. When the multiphysics processing agent calls the simulation tool to perform simulation calculations, it first receives the design parameters and boundary conditions as inputs to the neural network, and then calculates the predicted physical field distribution through the forward propagation of the neural network. During training, the total loss function of the neural network consists of three parts, as detailed above.

[0187] This embodiment embeds physical constraints into the loss function of a neural network, enabling the neural network to learn simulation results that conform to physical laws under conditions of small sample data. This overcomes the shortcomings of pure data-driven models in predicting inaccurately when data is insufficient, as well as the limitations of pure physics-driven models in terms of computational efficiency, and achieves efficient simulation calculation by fusing mechanism and data.

[0188] In some optional embodiments, the collaborative outputs of multiple specialized agents constitute an agent collaborative output matrix, and the aggregated output of the agent collaborative output matrix is ​​represented as:

[0189]

[0190] in, The total output of the agent matrix is ​​the final output value after comprehensively considering the independent contributions of each agent and the collaborative contributions among agents. This refers to the number of specialized intelligent agents, that is, the total number of specialized intelligent agents participating in the current collaborative task. For the first The weights of individual agents are used to measure their importance in the current task. The weights of the same agent can be dynamically adjusted in different task scenarios. For the first The output of a specialized intelligent agent is the intermediate result generated by that agent after independently executing sub-tasks. For the first The synergy coefficient between the j-th specialized agent and the j-th specialized agent is used to measure the magnitude of the synergy between the two agents. The synergy coefficient is positive when the combined outputs of the two agents can generate additional value, and negative otherwise. For collaborative operators, it represents the collaborative information fusion operation between the outputs of two agents, including but not limited to fusion methods such as feature concatenation, weighted fusion, and cross attention.

[0191] The collaborative outputs of multiple specialized intelligent agents constitute the agent collaborative output matrix. The agent collaborative output matrix is ​​the set of outputs from the collaborative work of multiple specialized intelligent agents in the agent collaboration layer. Each specialized intelligent agent communicates with each other through a standardized data interface, and generates an independent output after independently completing its sub-task. The chief intelligent agent then merges the independent outputs to generate a comprehensive inference conclusion.

[0192] When the cooperation coefficient between agents When the value is greater than zero, it means that the combined output of the two agents can produce a gain effect greater than the sum of their individual contributions.

[0193] This embodiment quantifies the combined effect of the independent and collaborative contributions of each professional agent through the aggregation output formula of the multi-agent collaborative output matrix, enabling the system to evaluate and optimize the collaborative configuration among agents and fully leverage the advantages of multi-agent collaborative work.

[0194] In some optional embodiments, the product engineering layer includes: an architecture and platform definition module, a physical integration design module, a system-level engineering management module, a quality and manufacturing engineering module, and a supply chain and cost analysis module.

[0195] The architecture and platform definition module defines the general architecture of the product platform and the interface specifications between modules. Specifically, in the early stages of product development, this module determines the basic platform architecture of the product (such as the vehicle platform, chip platform, etc.), defines the interface specifications, communication protocols, and data exchange formats between various functional modules, and provides a unified architectural framework for subsequent physical integration design. This module ensures that different subsystems are compatible and interconnected at the architectural level.

[0196] The Physical Integration Design module is used for the joint design and interface matching of multiple physical subsystems of a product. Specifically, the module receives simulation optimization results from the intelligent agent collaboration layer, transforms design parameters into specific physical subsystem design schemes, and performs joint design and interface matching of multiple physical subsystems to ensure that each subsystem can work collaboratively at the physical level. This module includes a simulation phase for verifying the integration scheme during the design process.

[0197] The system-level engineering management module is configured to manage requirement items, design changes and verification records throughout the full life cycle of a product; specifically, the system-level engineering management module is responsible for requirement management (recording, tracking and reviewing all requirement items), design change management (recording the reason, content and approval status of each design change) and verification record management (storing verification reports and test data of all stages) in the full life cycle of the product, so as to ensure the product development process is traceable and auditable. This module includes a verification link, which is used for performing compliance verification on the design scheme.

[0198] The quality and manufacturing engineering module is configured to perform virtual quality assessment and digital manufacturing process design; specifically, the quality and manufacturing engineering module performs virtual assessment on the manufacturability, assemblability and quality indicators of the product in the product design stage, and performs digital manufacturing process design after the product design is finalized, so as to generate process parameters and process flows.

[0199] The supply chain and cost analysis module is configured to construct a supply chain network model and perform product cost estimation and optimization. Specifically, the supply chain and cost analysis module constructs the supply chain network model (including nodes and connection relationships such as raw material suppliers, component suppliers, manufacturing factories, and logistics paths) according to the bill of materials and manufacturing process of the product, and performs product cost estimation and optimization on this basis.

[0200] In this embodiment, through the five major engineering modules of the product engineering layer, the output of the agent collaboration layer is converted into a full-range engineering scheme covering product platform definition, physical integration design, system-level management, quality and manufacturing, supply chain and cost, so that the target result output by the system has complete engineering implementability.

[0201] In some optional embodiments, the engineering effectiveness of the product engineering layer is evaluated by the following function:

[0202]

[0203] wherein, is the engineering effectiveness of product X, which is a dimensionless evaluation value for quantifying the overall working quality of the product engineering layer, and is used for measuring the quality and efficiency of the conversion process from the design scheme to a manufacturable product, is the output of the physical integration design module, including the integrated design scheme of the product, subsystem interface definitions and co-simulation results, which can be specifically quantified by indicators such as scheme completeness and interface matching degree, is the output of the simulation link in the physical integration design module, including multi-physics field simulation results and performance prediction data, which can be specifically quantified by indicators such as simulation accuracy and calculation efficiency, This refers to the output of the verification phase in the system-level engineering management module, including requirement verification results, design review records, and verification test data. Specifically, it can be quantified by metrics such as requirement coverage and defect density. The outputs of the Quality and Manufacturing Engineering module include virtual quality assessment results and manufacturing process design schemes, which can be quantified by indicators such as manufacturability score and process maturity. , , , These are the weighting coefficients for each module, used to measure the contribution of different modules to the overall engineering efficiency. All weighting coefficients are non-negative, and their relative magnitudes reflect the differences in the importance placed on each engineering stage by different product types. This formula aggregates the output capabilities of each underlying engineering module into product engineering efficiency, used to evaluate the overall performance of the product engineering layer.

[0204] This embodiment quantifies the contribution of each engineering module's output to the overall engineering performance through a product engineering performance evaluation function, providing a quantitative basis for performance evaluation and module configuration optimization at the product engineering layer.

[0205] In some optional embodiments, the application scenario layer includes: a demand interaction unit and a visualization display unit.

[0206] The Requirements Interaction Unit is used to acquire product requirements information input by users in the form of natural language and / or structured data. Specifically, the Requirements Interaction Unit provides a user interface that supports users to input product development requirements in various forms. Natural language refers to product requirements described by users in spoken or written language, such as designing a photovoltaic inverter that operates in a high-temperature environment with an efficiency of no less than 98%. The Requirements Interaction Unit uses the large language model of the underlying support layer to perform semantic understanding and structured transformation of the natural language input, converting it into a structured technical indicator system. Structured data refers to technical indicators and constraints input by users in the form of tables, parameter lists, or standardized documents. The Requirements Interaction Unit directly parses and validates the structured data. The Requirements Interaction Unit supports mixed input of both forms, allowing users to choose the most suitable input method according to their preferences.

[0207] The visualization unit is used to output the target results in at least one of the following forms: 3D model, simulation cloud map, data chart, and solution report. Specifically, after generating the target results at the product engineering level, the visualization unit visualizes the target results in a user-readable format. The 3D model is a graphical rendering of the product's 3D digital model, allowing users to view the product structure from any angle through operations such as rotation, scaling, and sectioning. The simulation cloud map displays the physical field calculation results (such as temperature distribution, stress distribution, etc.) as a color-mapped cloud map on the surface of the 3D model or within the calculation domain, allowing users to intuitively identify weak performance areas through color distribution. The data chart displays performance data, comparative analysis results, etc., in the form of line graphs, bar charts, tables, etc. The solution report organizes the structured information of the target results (including design scheme description, simulation result summary, process scheme, etc.) into a complete solution report document.

[0208] This embodiment supports both natural language and structured data input methods through the demand interaction unit, which lowers the user's barrier to entry; and presents the target results in multiple forms through the visualization display unit, which improves the user's perception efficiency of the design scheme and the quality of decision-making.

[0209] In some optional embodiments, the application value of the application scenario layer is evaluated using the following function:

[0210]

[0211] Total value of the application layer For the first The collaboration coefficient between the j-th specialized intelligent agent and the j-th specialized intelligent agent. For the number of application scenarios, The weight of the k-th scene is used to measure the weight of the k-th scene. The contribution of each scenario to the overall application value is prioritized, and the weight of different scenarios can be set according to factors such as market demand and strategic importance. The utility function for this scenario is a function that maps the product data input for this scenario to a utility value. The utility value quantifies the empowering effect of the digital physical world in this scenario (such as improved R&D efficiency, reduced costs, improved product performance, etc.). The data input for product X refers to all types of data that the target product inputs into the system during its research and development and operation, including but not limited to product requirement information, design parameters, simulation boundary conditions, test cases, process parameters, and operational status data, under different scenarios. The specific data items and formats are dynamically adapted according to the needs of the scenario. This formula expresses the empowering value of the digital physical world for different industrial application scenarios. By weighted aggregation of the utility functions of each scenario, the overall value of the system in multiple application scenarios is quantitatively evaluated.

[0212] This embodiment quantifies the enabling value of the digital physical world for different industrial scenarios by applying a value aggregation model, providing a quantitative decision-making basis for the system's scenario expansion strategy and resource allocation optimization.

[0213] In some optional embodiments, the human-machine collaboration efficiency of the system is evaluated using the following function:

[0214]

[0215] in, The overall efficiency of human-machine collaboration is a dimensionless evaluation value that quantifies the overall efficiency of human experts and AI systems working together in the digital-physical world. It is used to measure the comprehensive effectiveness of human-machine collaboration models. The contribution of experts to efficiency is a quantifiable value based on the experience, intuition, and creative thinking of human experts to the product development process. It can be specifically characterized by indicators such as the quality of design decisions, problem identification ability, and the number of innovative solutions. The contribution of AI to efficiency is a quantifiable value based on the contribution of the AI ​​system's computing power, data processing power, and pattern recognition capabilities to the product development process. Specifically, it can be characterized by indicators such as simulation calculation speed, data processing volume, and optimized search efficiency. The synergistic effect of human-computer interaction is a quantitative measure of the efficiency of information transmission, the quality of feedback loops, and the degree of goal alignment between human experts and AI systems. Specifically, it can be characterized by indicators such as interaction frequency, feedback response time, and decision consistency. , , These are weighting coefficients that are dynamically adjusted based on different R&D scenarios, such as the conceptual design phase. The (human weight) can be increased to leverage the advantages of human creative thinking during the simulation calculation phase. The AI ​​weights can be adjusted to fully leverage the computational advantages of AI. This model quantifies the gain effect of human-machine collaboration and validates its effectiveness. The synergistic effect occurs when the experiential judgment of human experts and the computational analysis of AI systems complement each other through high-quality interaction, resulting in increased synergistic efficiency. It is greater than the sum of the efficiency of humans working independently and AI working independently.

[0216] This embodiment quantifies the contributions of human experts, AI systems, and their interactions to the overall collaborative efficiency through a human-machine symbiotic collaboration efficiency model. This provides a quantitative basis for optimizing human-machine task allocation and interaction design, and verifies the gain effect of human-machine collaboration.

[0217] For ease of understanding, combined with Figure 4 As shown, Figure 4This is a flowchart illustrating an industrial product development process in one embodiment, comprising six stages: requirements definition, detailed design, virtual testing, process manufacturing, digital twin deployment, and operational optimization. Each stage involves specific agent collaboration, data flow, and physical computation. The entire process forms a V-model-driven digital closed loop, with the knowledge accumulated in each iteration feeding back into the physical mechanism library and AI model, enabling continuous evolution of R&D capabilities. These six stages mainly include:

[0218] Phase 1: Requirements Definition and Conceptual Design. This phase is executed by the architecture and platform engineering modules of the product engineering layer, as well as the requirements analysis and concept design agents in the agent collaboration layer. Users input product requirements through the application scenario layer interface. The system calls the large language model in the basic support layer to transform unstructured product requirements into structured technical indicators. The requirements analysis agent automatically supplements the technical boundaries and constraints of the requirements (such as efficiency requirements, cost targets, environmental conditions, etc.) based on the industry knowledge base and historical project data. The concept design agent retrieves similar product cases from the product database and generates 3-5 conceptual design schemes by combining them with physical laws from the physical foundation layer knowledge base. The multiphysics processing agent quickly estimates each scheme and selects the optimal 2-3 schemes to proceed to the next phase. The system outputs a conceptual design report, including: a scheme comparison analysis table, preliminary simulation results, and recommended schemes.

[0219] Phase Two: Detailed Design and Multiphysics Simulation. The execution entities are the product system engineering module in the product engineering layer, and the multiphysics processing agent and physical information enhancement agent in the agent collaboration layer. The product engineering layer receives the conceptual design scheme, automatically generates a 3D structural model (CAD model) of the product, and defines the assembly relationships and interface specifications between various components. The multiphysics processing agent matches the required physical field models (such as electric field, magnetic field, thermal field, structural mechanics field, etc.) from the physical foundation layer knowledge base according to the product characteristics and establishes multiphysics coupling relationships. Based on a physical constraint neural network algorithm, the multiphysics processing agent combines physical constraints with historical measured data to perform high-precision simulation calculations on the design scheme. The simulation content includes electromagnetic field distribution simulation, heat conduction simulation, structural mechanics simulation, and multiphysics coupling simulation. The physical information enhancement agent intelligently analyzes the simulation results, identifies bottleneck problems in the design (such as local overheating, stress concentration, low efficiency, etc.), and provides optimization suggestions. The system automatically adjusts the design parameters based on the simulation analysis results, reconstructs the model, and simulates again, forming a design-simulation-analysis-optimization cycle until the simulation results meet all design indicators. The system outputs a detailed design solution, including: a complete 3D model, a detailed simulation report (color cloud map, data charts), and optimization records.

[0220] Phase Three: Virtual Testing and Verification. The execution entities are the system quality and manufacturing module at the product engineering layer, and the test verification agent and target condition simulation agent in the agent collaboration layer. The test verification agent builds a virtual testbed in the digital physical world based on the detailed design scheme. The parameters of the virtual testbed are completely consistent with those of real physical test equipment. The target condition simulation agent applies extreme operating conditions to the virtual testbed, including temperature cycling tests, vibration tests, electromagnetic compatibility tests, load surge tests, and accelerated life tests. The test verification agent monitors the product's performance under various operating conditions, records the changing trends of key parameters, and calculates the product's reliability indicators. The test verification agent also proactively injects various potential faults (such as component failures, heat dissipation failures, etc.), deduces the propagation path and impact range of the faults, and verifies the system's fault tolerance. If the test fails, the system automatically traces back to the design phase (e.g., Phase Two) to locate the root cause of the problem and propose a corrective solution, forming a closed loop of test-analysis-correction-retest. After successful verification, the system outputs a complete test verification report, including test data for all operating conditions, reliability assessment results, and fault analysis records.

[0221] Phase Four: Process Design and Manufacturing Simulation. The execution entities are the Quality and Manufacturing Engineering module at the Product Engineering layer, and the process design agent, manufacturing simulation agent, quality pre-assessment agent, and process optimization agent in the agent collaboration layer. The process design agent automatically generates manufacturing process parameters based on the product's 3D model and material properties: processing path, process parameters (temperature, pressure, speed, etc.), assembly sequence, etc. The manufacturing simulation agent digitally simulates the entire manufacturing process, including machining simulation, welding simulation, assembly simulation, and quality simulation. The quality pre-assessment agent predicts product quality indicators (such as dimensional accuracy, surface roughness, mechanical properties, etc.) based on the manufacturing simulation results and compares them with design requirements. The process optimization agent adjusts process parameters or process routes based on the quality assessment results, forming a simulation-assessment-optimization cycle until the manufacturing quality meets design requirements. The system outputs complete manufacturing process documents, including process parameter tables, manufacturing simulation reports, and quality assurance plans.

[0222] Phase Five: Digital Twin Deployment and Real-time Monitoring. The execution entities are the digital twin intelligent agents in the intelligent agent collaboration layer and the visualization display unit in the application scenario layer. The digital twin intelligent agent creates a corresponding digital twin for a physical product that has already been manufactured in the real physical world, injecting all data from the design and manufacturing stages (3D models, material parameters, process records, etc.) into the digital twin model. After product deployment, the digital twin intelligent agent collects real-time product operating data (temperature, vibration, current, voltage, etc.) through the Internet of Things and sensor networks, synchronizing it to the digital twin at a preset time frequency. The digital twin continuously calculates the real-time status of the product based on the physical model and real-time input data, maintaining synchronization with the physical product. The digital twin intelligent agent performs real-time monitoring and trend analysis on the status data of the digital twin, including status assessment, trend prediction, and fault early warning. The visualization display unit in the application scenario layer provides a visualized digital twin monitoring interface, supporting remote access and simultaneous monitoring of multiple products.

[0223] Phase Six: Operational Optimization and Iterative Evolution. The main actors are the intelligent optimization agent, knowledge accumulation agent, and product iteration agent in the intelligent agent collaboration layer. The system continuously collects operational data throughout the product's entire lifecycle, forming a product knowledge base. The knowledge accumulation agent extracts fault modes and impact analysis, performance degradation patterns, and optimization opportunities from operational data (including fault data, maintenance records, performance degradation curves, etc.) and updates it to the knowledge base in the physical foundation layer. The intelligent optimization agent generates optimization suggestions for the next generation of the product based on accumulated knowledge and user feedback (such as improving efficiency, reducing costs, and extending lifespan). The product iteration agent initiates a new round of R&D cycle, starting from the optimization suggestions and entering Phase One, forming a continuously iterative R&D closed loop. The knowledge accumulated in each iteration feeds back into the knowledge base and AI model of the physical foundation layer, continuously improving the system's prediction accuracy and intelligent decision-making capabilities.

[0224] The above six stages constitute a complete digital R&D closed loop. Each stage is driven by a dedicated intelligent agent, achieving full lifecycle digital management from conceptual design to decommissioning and recycling through deep integration of physical mechanisms and AI algorithms. The core value of this process lies in: Time and space compression: reducing the traditional R&D cycle of months or even years to days or even hours, improving efficiency by 4 to 1000 times; Cost reduction: replacing a large number of physical prototype tests with virtual simulation, reducing trial-and-error costs by more than 60%; Extreme exploration: enabling safe research into extreme working conditions and unknown scenarios in digital space, pushing the boundaries of physical testing; Knowledge accumulation: the experience of each project is accumulated in a knowledge base, forming the company's digital assets and continuously improving R&D capabilities.

[0225] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0226] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores the data involved in the aforementioned stages. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection.

[0227] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0228] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0229] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above embodiments.

[0230] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above embodiments.

[0231] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0232] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0233] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0234] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A virtual physical system for industrial product research and development, characterized in that, The system includes: The physical foundation layer stores a knowledge base that digitally represents the physical laws of the real physical world; The basic support layer is deployed on top of the physical infrastructure layer and is used to provide data processing capabilities based on the physical infrastructure layer. The R&D logic model layer, deployed on top of the basic support layer, is used to provide a standardized product R&D process template. The process template defines multiple R&D stages from requirements to testing and verification, as well as the mapping relationship between each R&D stage. The intelligent agent collaboration layer, deployed on top of the basic support layer, includes a chief intelligent agent and multiple specialized intelligent agents. The chief intelligent agent is used to schedule the specialized intelligent agents to collaboratively execute R&D sub-tasks. The product engineering layer, deployed above the intelligent agent collaboration layer, is used to transform the output of the intelligent agent collaboration layer into target results that conform to engineering specifications. The target results are represented by a three-dimensional digital model. The application scenario layer, deployed above the product engineering layer, is used to receive product requirement information and display the target results. The system includes a downlink data flow mode and an uplink data flow mode. In the downlink data flow mode, the product demand information flows from the application scenario layer to the physical infrastructure layer to obtain physical calculation results. In the uplink data flow mode, the physical calculation results flow from the physical infrastructure layer to the application scenario layer and are transformed into the target result.

2. The system according to claim 1, characterized in that, The downlink data transfer mode includes: The application scenario layer receives product requirement information and then transmits it to the product engineering layer. The product engineering layer breaks down the product requirement information into engineering tasks and then transmits them to the intelligent agent collaboration layer. The intelligent agent collaboration layer, after scheduling the corresponding professional intelligent agents to execute the engineering task through the chief engineer intelligent agent, transmits the processing results to the R&D logic model layer. The R&D logic model layer organizes the R&D sequence according to the process template and then transmits it to the basic support layer; The basic support layer allocates computing resources and calls the knowledge base in the physical foundation layer to perform physical field numerical calculations to obtain physical calculation results.

3. The system according to claim 1, characterized in that, The uplink data transfer mode includes: The physical foundation layer transmits the physical calculation results to the basic support layer for data integration. The basic support layer transmits the integrated data to the R&D logic model layer for phased verification and optimization. The R&D logic model layer transmits the verified and optimized data to the agent collaboration layer for multi-agent joint inference; The intelligent agent collaboration layer transmits the inference results to the product engineering layer to generate the target results; The product engineering layer transmits the target result to the application scenario layer, and the application scenario layer outputs the target result.

4. The system according to any one of claims 1 to 3, characterized in that, The knowledge base of the physical foundation layer includes: basic laws describing the real physical world based on the principle of least action, governing equations describing the laws of motion of matter, and different physical field models for multi-field coupling calculations; the governing equations include at least one of the continuity equation, the momentum conservation equation, and the energy conservation equation.

5. The system according to any one of claims 1 to 3, characterized in that, The basic support layer includes: a large language model, a data engineering laboratory, a cloud-native architecture, and a physical solver interface; the basic support layer is used to integrate physical laws with AI algorithms, providing computing power, data processing power, and model reasoning power to each agent in the agent collaboration layer, and the large language model includes an attention mechanism.

6. The system according to any one of claims 1 to 3, characterized in that, The process template in the R&D logic model layer is a V-shaped R&D process template. The V-shaped R&D process template includes a decomposition stage sequence and an integration verification stage sequence. Each decomposition stage and its corresponding integration verification stage form a bidirectional traceability link. The bidirectional traceability link is used to backtrack to the corresponding decomposition stage for design correction when verification fails.

7. The system according to claim 6, characterized in that, The R&D efficiency of the R&D logic model layer is evaluated using the following function: in, To improve the R&D efficiency of the V-shaped R&D process template. This represents the total number of research and development stages. Let be the verification function for the requirements and specifications of the i-th development phase. For the needs of the i-th research and development stage, For the specifications of the i-th research and development stage, is the weighting coefficient for the i-th research and development stage.

8. The system according to any one of claims 1 to 3, characterized in that, The plurality of specialized intelligent agents includes at least one of the following: A demand analysis intelligent agent is used to analyze the product demand information to obtain at least one initial design scheme; A multiphysics processing agent is used to determine the physical field model corresponding to the initial design scheme, establish a multiphysics layer coupling relationship, and simulate the initial design scheme based on the multiphysics coupling relationship to obtain simulation results. A physical information augmentation agent is used to identify problems in the simulation results and optimize the design scheme. The optimized design scheme is then fed back to the multiphysics processing agent for re-simulation until the simulation results meet the design requirements, thus obtaining the target design scheme.

9. The system according to claim 8, characterized in that, The aforementioned specialized intelligent agents also include: The testing and verification agent is used to acquire a virtual test bench, test and verify the target design scheme on the virtual test bench, record performance information under various working conditions, obtain test results based on the performance information, and, if the test results fail, revise the target design scheme and trigger the multiphysics processing agent to re-simulate the revised target design scheme, and re-test and verify the re-simulated design scheme until the test passes; the virtual test bench is consistent with the real physical testing equipment; A target operating condition simulation agent is used to configure various target operating conditions on the virtual test bench, including extreme operating conditions and fault operating conditions.

10. The system according to claim 9, characterized in that, The plurality of specialized intelligent agents also include: a process design intelligent agent, used to generate manufacturing process parameters based on the target design scheme that has passed the test; A manufacturing simulation intelligent agent is used to digitally simulate the manufacturing process based on the manufacturing process parameters to obtain manufacturing simulation results. A quality pre-assessment intelligent agent is used to obtain predicted quality indicators based on the manufacturing simulation results and compare them with the target quality indicators in the product demand information to obtain quality comparison results. A process optimization agent is used to adjust process parameters and / or process routes based on the quality comparison results until the product quality indicators obtained based on the manufacturing simulation results meet the product quality indicators in the product requirement information.

11. The system according to claim 10, characterized in that, The aforementioned specialized intelligent agents also include: A digital twin intelligent agent is used to acquire product operation data collected from physical products in the real physical world, synchronize the product operation data to the digital twin, and perform monitoring and early warning based on the status data of the digital twin.

12. The system according to claim 11, characterized in that, The aforementioned specialized intelligent agents also include: Intelligent optimization agents are used to generate product optimization suggestions based on accumulated operational data and user feedback; A knowledge-accumulating intelligent agent is used to extract fault modes, performance degradation patterns, and optimization opportunities from operational data throughout the product's lifecycle, and to update the knowledge base. The product iteration intelligent agent is used to trigger a new round of R&D cycle, starting from the optimization suggestions and entering the requirement definition stage to form a continuous iterative R&D closed loop.

13. The system according to claim 8, characterized in that, The decision-making strategies of each specialized intelligent agent are based on a value function iteration method to select the optimal action. The optimal value function of the value function iteration method is: in, Let be the optimal value function for state s, representing the maximum cumulative reward obtainable from state s; For an immediate reward of performing action a in state s, Discount factor ( ), used to measure the importance of future returns, Let be the state transition probability, representing the probability that after performing action a in state s, the state transitions to state s'.

14. The system according to claim 8, characterized in that, The simulation is performed using a neural network, and the total loss function of the neural network is: The total loss of the neural network, Data-driven loss measures the deviation between the neural network's predictions and the actual measured data. The physical constraint loss measures whether the neural network predictions satisfy the governing equations. Boundary condition loss measures whether the neural network satisfies a given condition at the boundary. , , The weighting coefficients are used to adaptively adjust the contributions of the three losses.

15. The system according to claim 8, characterized in that, The collaborative outputs of multiple specialized intelligent agents constitute an agent collaborative output matrix, and the aggregated output of the agent collaborative output matrix is ​​represented as follows: in, This is the total output of the agent matrix. The number of the specialized intelligent agents. For the first The weights of the aforementioned specialized intelligent agents For the first The output of the aforementioned specialized intelligent agent For the first The collaboration coefficient between the j-th specialized intelligent agent and the j-th specialized intelligent agent. For cooperative operators.

16. The system according to any one of claims 1 to 3, characterized in that, The product engineering layer includes: The architecture and platform definition module is used to define the general architecture of the product platform and the interface specifications between modules; The physical integration design module is used for the joint design and interface matching of multiple physical subsystems of a product; The system-level engineering management module is used to manage requirement items, design changes, and verification records throughout the product lifecycle. The Quality and Manufacturing Engineering module is used to perform virtual quality assessments and digital manufacturing process design. The supply chain and cost analysis module is used to build supply chain network models and perform product cost estimation and optimization.

17. The system according to claim 16, characterized in that, The engineering performance of the product engineering layer is evaluated using the following function: in, For the engineering performance of product X, The output of the physical integration design module, This is the output of the simulation stage in the physical integration design module. This is the output of the verification stage in the system-level engineering management module. This is the output of the quality and manufacturing engineering module. , , , These are the weight coefficients for each corresponding module.

18. The system according to any one of claims 1 to 3, characterized in that, The application scenario layer includes: The demand interaction unit is used to obtain product demand information input by users in the form of natural language and / or structured data; The visualization unit is used to output the target results in at least one of the following forms: 3D model, simulation cloud map, data chart, and scheme report.

19. The system according to claim 18, characterized in that, The application value of the application scenario layer is evaluated using the following function: Total value of the application layer For the number of application scenarios, Let k be the weight of the k-th scene. Let be the utility function for this scenario. Input data for product X.

20. The system according to any one of claims 1 to 3, characterized in that, The human-machine collaboration efficiency of the system is evaluated using the following function: in, The overall efficiency of human-machine collaboration To contribute to the efficiency of experts. To contribute to the efficiency of AI The synergistic effect of human-computer interaction, , , These are weighting coefficients, which are dynamically adjusted according to different R&D scenarios.

Images