System development method, device, equipment, medium and program product based on large model

Abstract

The application relates to a large model-based system development method, device, equipment, medium and program product, which comprises the following steps: receiving product demand information input in natural language through an application service layer; obtaining each physical mathematical model based on the product demand information through a target agent of an intelligent enhancement layer, wherein the physical mathematical model is used for describing design parameters and constraint conditions of each subsystem; calling each simulation tool of a physical base layer through a simulation verification agent of the intelligent enhancement layer, performing simulation calculation on each physical mathematical model in parallel, and generating simulation results; performing multi-objective optimization on multiple objectives corresponding to each simulation result and the product demand information through a total work agent of the intelligent enhancement layer, obtaining optimized design parameters, and performing simulation calculation again to generate updated simulation results until the multi-objective optimization meets a convergence condition, obtaining target design parameters of each physical mathematical model, and generating a target development scheme. The method can improve the research and development efficiency.

Description

Technical Field

[0001] This application relates to the field of digital design and simulation technology for industrial products, and in particular to a system development method, apparatus, equipment, medium and program product based on a large model. Background Technology

[0002] The traditional V-shaped development process is a widely adopted framework in industries such as automotive. Its characteristic is that the development process is divided into two aspects: the left side represents the development phase, progressing step-by-step from requirements analysis to goal setting and decomposition; the right side represents the testing phase, progressing step-by-step from component and system testing to whole-machine testing and acceptance testing. This systematic approach ensures comprehensive verification at each development stage, but for critical systems such as safety, it faces dual challenges in terms of development efficiency and safety certification.

[0003] In traditional technologies, the following approach is typically used: product requirements are presented in the form of natural language documents, which are then manually interpreted and transformed into design specifications by domain experts. Each professional team uses its own simulation tools (such as fluid simulation tools, structural simulation tools, etc.) to independently model and simulate. The adjustment and optimization of design parameters rely on the experience of engineers for manual judgment and modification. After optimization, the code development team manually writes code based on the design documents, and finally, testing and verification are conducted.

[0004] However, the development process requires a significant amount of manual coordination. From understanding and transforming requirements, adjusting design parameters, scheduling simulation tasks, summarizing multidisciplinary simulation results, to iterative optimization of design parameters, each stage relies heavily on manual communication and judgment by engineers, resulting in long development cycles and low efficiency. Summary of the Invention

[0005] Therefore, it is necessary to provide a system development method, apparatus, equipment, medium, and program product based on a large model that can improve R&D efficiency in response to the above-mentioned technical problems.

[0006] Firstly, this application provides a system development method based on a large model, the method comprising:

[0007] The application service layer receives product requirement information input in natural language.

[0008] The target agent of the intelligent enhancement layer obtains various physical and mathematical models based on the product requirement information. The physical and mathematical models are used to describe the design parameters and constraints of each subsystem.

[0009] The simulation verification agent of the intelligent enhancement layer calls the simulation tools of the physical base layer to perform simulation calculations on each physical mathematical model in parallel and generate simulation results.

[0010] Through the chief engineer agent of the intelligent enhancement layer, multi-objective optimization is performed based on the simulation results and the product requirement information to obtain optimized design parameters. The optimized design parameters are then fed back to the simulation verification agent, which triggers the simulation verification agent to update the design parameters in each physical mathematical model and re-perform simulation calculations to generate updated simulation results. This process continues until the multi-objective optimization meets the convergence condition, at which point the target design parameters of each physical mathematical model are obtained. Based on each physical mathematical model and the target design parameters, a target development plan is generated.

[0011] In some optional embodiments, the chief engineer agent through the intelligent enhancement layer performs multi-objective optimization based on the simulation results and the product requirement information to obtain optimized design parameters, including:

[0012] Based on the physics discipline indicators in each simulation result, reference target values ​​are generated for multiple objectives corresponding to the product demand information, and each non-compliant objective is determined based on the multiple objectives corresponding to the product demand information and the reference target values.

[0013] Based on each of the non-compliance targets, the penalty value for each non-compliance target is calculated using the penalty function method, and the optimization direction is determined based on the penalty value;

[0014] Based on the optimization direction, the design parameters are adjusted in multiple directions using a multi-objective optimization algorithm to generate optimized design parameters.

[0015] In some optional embodiments, the step of obtaining the target design parameters of each physical mathematical model after the multi-objective optimization satisfies the convergence condition includes:

[0016] Based on the physics discipline indicators in the updated simulation results, new reference target values ​​are generated for the multiple objectives corresponding to the product demand information.

[0017] If the new reference target values ​​all meet the multiple targets corresponding to the product requirement information, or if the change in the physics subject index in the adjacent updated simulation results is less than the preset convergence threshold, the optimized design parameters determined in this study will be used as the target design parameters for each physics mathematical model.

[0018] In some alternative embodiments, the method further includes:

[0019] During the multi-objective optimization iteration process, the chief engineer agent performs optimization layer risk identification, which includes at least one of parameter sensitivity risk identification, constraint out-of-bounds risk identification, and robustness risk identification.

[0020] Risk registration items are generated based on the results of the risk identification at the optimization layer.

[0021] In some optional embodiments, before the simulation verification agent of the intelligent enhancement layer calls the various simulation tools of the physical base layer to perform simulation calculations on each of the physical mathematical models in parallel and generate simulation results, the method further includes:

[0022] The simulation verification agent of the intelligent enhancement layer generates corresponding simulation tasks based on each of the physical and mathematical models, and classifies each of the simulation tasks.

[0023] The computing resources of the physical base layer are divided into multiple resource levels;

[0024] The simulation tasks after classification are matched with the resource levels after hierarchical division, and the optimization algorithm is used to solve the problem with at least two of the total task completion delay, computing cost and simulation quality as optimization objectives to generate resource allocation schemes corresponding to each simulation task.

[0025] The computing resources corresponding to each simulation task are determined based on the resource allocation scheme.

[0026] In some optional embodiments, the resource hierarchy includes GPU clusters, CPU clusters, and edge computing nodes; the process of matching the classified simulation tasks with the hierarchically divided resource hierarchy, and using an optimization algorithm to solve for at least two of the following as optimization objectives—total task completion latency, computing cost, and simulation quality—to generate a resource allocation scheme corresponding to each simulation task, includes:

[0027] Based on the subject type and fidelity of each simulation task, fluid simulation and neural network training tasks are matched to the GPU cluster, structural simulation and electromagnetic simulation tasks are matched to the CPU cluster, and software-in-the-loop testing and hardware-in-the-loop testing tasks are matched to the edge computing node to generate an initial resource allocation scheme. The fidelity of the simulation task is the credibility of the physical mathematical model.

[0028] Using at least two of the following as optimization objectives—total task completion delay, computing cost, and simulation quality—an optimization algorithm is used to solve the initial resource allocation scheme and output the resource allocation scheme corresponding to each simulation task.

[0029] In some optional embodiments, the simulation verification agent through the intelligent enhancement layer calls various simulation tools of the physical base layer to perform simulation calculations on each of the physical mathematical models in parallel and generate simulation results, including:

[0030] Determine the simulation requirements corresponding to each of the physical and mathematical models;

[0031] The type of simulation tool will be determined based on the aforementioned simulation requirements;

[0032] The simulation verification agent of the intelligent enhancement layer calls the simulation tool corresponding to the simulation tool type in the physical base layer, and performs simulation calculations on the physical mathematical model through the called simulation tool to generate simulation results.

[0033] In some optional embodiments, the step of calling a simulation tool corresponding to the simulation tool type in the physical base layer through the simulation verification proxy of the intelligent enhancement layer, and performing simulation calculations on the physical mathematical model through the called simulation tool to generate simulation results includes:

[0034] When the simulation tool type is a physics solver, the simulation verification agent of the intelligent enhancement layer calls the physics solver of the physical base layer, and performs simulation calculations on the physical mathematical model through the scheduled physics solver to generate simulation results;

[0035] When the simulation tool type is a neural network, the simulation verification agent of the intelligent enhancement layer calls the neural network of the physical base layer, and performs simulation calculations on the physical mathematical model through the neural network to generate simulation results.

[0036] When the simulation tool type is a hybrid type, the simulation verification agent of the intelligent enhancement layer calls the neural network of the physical base layer, and performs simulation calculations on the physical mathematical model through the neural network to generate candidate simulation results. Then, the simulation verification agent of the intelligent enhancement layer calls the physical solver of the physical base layer, and performs simulation calculations on the physical mathematical model corresponding to the candidate simulation results through the scheduled physical solver to generate simulation results.

[0037] In some alternative embodiments, after receiving product demand information input in natural language through the application service layer, the process further includes:

[0038] The product requirement information is encapsulated by the application service layer to obtain a target requirement package described by a preset physical description language;

[0039] The target agent through the intelligent enhancement layer obtains various physical and mathematical models based on the product demand information, including:

[0040] The chief engineer agent of the intelligent enhancement layer calls the requirements analysis agent to parse the target requirements package and generate an initial system model.

[0041] The chief engineer agent of the intelligent enhancement layer calls the architecture design agent to process the initial system model and generate various physical and mathematical models.

[0042] In some optional embodiments, the step of invoking the requirements analysis agent through the chief engineer's agent of the intelligent enhancement layer to parse the target requirements package and generate an initial system model includes:

[0043] The chief engineer agent of the intelligent enhancement layer calls the demand analysis agent to perform intent recognition and entity extraction on the target demand package, and obtain demand indicators and demand constraints.

[0044] The chief engineer agent of the intelligent enhancement layer calls the requirements analysis agent to obtain the requirements indicator mapping template, and generates quantitative design indicators based on the requirements indicators and requirements constraints using the requirements indicator mapping template.

[0045] Based on the quantitative design metrics, an initial system model is generated, which includes at least one of the following: a requirement diagram, a use case diagram, a block definition diagram, an internal block diagram, a parameter diagram, and a state machine.

[0046] In some alternative embodiments, the method further includes:

[0047] The chief engineer agent of the intelligent enhancement layer calls the demand analysis agent to perform demand layer risk identification. The demand layer risk identification includes identifying at least one of demand conflicts, unachievable indicators, regulatory gaps, and interface inconsistencies.

[0048] If the risk identification result at the demand layer indicates the existence of a risk, an alarm message will be output.

[0049] If the risk identification result at the demand layer indicates that there is no risk, the process continues by calling the architecture design agent through the chief engineer agent of the intelligent enhancement layer to process the initial system model and generate various physical and mathematical models.

[0050] In some optional embodiments, the process of invoking the architecture design agent through the chief engineer's agent of the intelligent enhancement layer to process the initial system model and generate various physical and mathematical models includes:

[0051] The chief engineer agent calls the architecture design agent to parse the block definition diagram and parameter diagram in the initial system model and determine the physical field type corresponding to the product requirement information.

[0052] Obtain the governing equations that match the physical field type from the physical base layer;

[0053] Based on the constraint equations, geometric parameters, and material parameters of each subsystem in the initial system model, the boundary conditions, initial conditions, and design parameters of each physical and mathematical model are determined.

[0054] Based on the data interaction relationships between the subsystems, establish multiphysics coupling relationships;

[0055] Based on the governing equations, boundary conditions, initial conditions, design parameters, and multiphysics coupling relationships, various physical mathematical models are generated.

[0056] In some optional embodiments, after generating the target development scheme based on each of the physical mathematical models and the target design parameters, the method further includes:

[0057] The code generation agent of the intelligent enhancement layer reads the model information and architecture configuration in the target development scheme and generates executable code.

[0058] The executable code is tested by the test verification agent of the intelligent enhancement layer, and a digital twin is obtained after the test is passed.

[0059] In some optional embodiments, the step of reading model information and architecture configuration from the target development scheme through the code generation agent of the intelligent enhancement layer and generating executable code includes:

[0060] The code generation agent of the intelligent enhancement layer reads the initial system model in the target development scheme;

[0061] Based on the initial system model, a control logic skeleton and software component code framework are generated.

[0062] The control algorithm code and initial values ​​of calibration parameters are generated by using a large language model based on the control strategy in the physical mathematical model.

[0063] Perform a consistency check on the generated code and output the executable code after the check passes.

[0064] In some optional embodiments, the testing of the executable code via the test verification proxy of the smart enhancement layer includes:

[0065] Based on the executable code, software-in-the-loop testing, hardware-in-the-loop testing, and physical testing are performed sequentially.

[0066] After the test passes, save the executable code.

[0067] In some alternative embodiments, the method further includes:

[0068] The chief engineer agent schedules each intelligent agent; the intelligent agent includes at least one of the target agent, the simulation verification agent, the code generation agent, and the test verification agent.

[0069] The chief engineer agent communicates with each of the intelligent agents via the gRPC protocol, the intelligent agents communicate with each other via the A2A protocol, the simulation verification agent calls the simulation tool of the physical base layer via the MCP protocol, and compensates for the latency in gRPC communication and / or MCP communication via an FIR filter.

[0070] In some alternative embodiments, the method further includes:

[0071] The V-model R&D efficiency is calculated through the chief engineer's agent, and the V-model R&D efficiency is used to quantify the closed-loop quality from product requirement information to testing and verification.

[0072]

[0073] in, The value represents the R&D efficiency, where N is the total number of R&D stages. For the first Verification functions for each stage of requirements and specifications. For the first The weight coefficients for each stage, where the total number of R&D stages is the number of stages included in the entire V-model process from product requirement information to testing and verification.

[0074] In some alternative embodiments, the method further includes:

[0075] Acquire source domain experience data and extract transferable domain-invariant features from the source domain experience data;

[0076] Based on the domain-invariant features, transfer learning is performed on samples in the target domain to generate a target domain model corresponding to the target domain.

[0077] In some optional embodiments, the step of performing transfer learning on samples in the target domain based on the domain-invariant features to generate a target domain model corresponding to the target domain includes:

[0078] Use the empirical data from the source domain as the source domain sample, and obtain a small amount of data from the target domain as the target domain sample.

[0079] Minimize the difference in feature distribution between the source neighborhood samples and the target neighborhood samples in the regenerating kernel Hilbert space by maximizing the mean difference;

[0080] Based on the task loss in the target domain and the maximum mean difference, a migration loss function is constructed;

[0081] With the goal of minimizing the transfer loss, the task model of the target domain is trained to obtain the transferred target domain model.

[0082] In some alternative embodiments, the method further includes:

[0083] The target design parameters in the target development scheme are sent to the physical product through the digital twin mapping interface, and the operation data of the physical product is collected and sent back to the corresponding digital twin.

[0084] The digital twin is corrected based on the operational data to generate a corrected digital twin.

[0085] In some optional embodiments, the step of correcting the digital twin based on the operational data to generate a corrected digital twin includes:

[0086] The running data is compared with the predicted values ​​of the digital twin, and the residuals are calculated.

[0087] If the residual exceeds a preset threshold, the digital twin is corrected, and based on the model corresponding to the corrected digital twin and the running data, the process of multi-objective optimization based on the simulation results and the product requirement information is continued through the chief engineer agent of the intelligent enhancement layer.

[0088] In some alternative embodiments, the method further includes:

[0089] The chief engineer's agent obtains information on changes to product demand;

[0090] The affected downstream entries are determined based on the change information;

[0091] Based on the affected downstream entries, the corresponding smart agent is triggered to perform a collaborative update.

[0092] Secondly, this application also provides a system development apparatus based on a large model, the apparatus comprising:

[0093] The requirement information acquisition module is used to receive product requirement information input in natural language through the application service layer;

[0094] The physical mathematical model generation module is used to obtain various physical mathematical models based on the product requirement information through the target agent of the intelligent enhancement layer. The physical mathematical models are used to describe the design parameters and constraints of each subsystem.

[0095] The simulation module is used to call various simulation tools in the physical base layer through the simulation verification agent of the intelligent enhancement layer, perform simulation calculations on each of the physical mathematical models in parallel, and generate simulation results.

[0096] The multi-objective optimization module is used to perform multi-objective optimization based on the simulation results and product requirement information through the chief engineer agent of the intelligent enhancement layer, obtain optimized design parameters, and feed the optimized design parameters back to the simulation verification agent, triggering the simulation verification agent to update the design parameters in each physical mathematical model and re-perform simulation calculations to generate updated simulation results. This process continues until the multi-objective optimization meets the convergence condition, obtaining the target design parameters of each physical mathematical model, and generating a target development plan based on each physical mathematical model and the target design parameters.

[0097] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the above embodiments.

[0098] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method in any of the above embodiments.

[0099] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in any of the above embodiments.

[0100] The aforementioned system development method, apparatus, equipment, medium, and program products based on large models receive product requirement information input in natural language through the application service layer; the target agent of the intelligent enhancement layer obtains various physical mathematical models based on the product requirement information, which are used to describe the design parameters and constraints of each subsystem; the simulation verification agent of the intelligent enhancement layer calls various simulation tools of the physical base layer to perform simulation calculations on each physical mathematical model in parallel and generate simulation results; the chief engineer agent of the intelligent enhancement layer performs multi-objective optimization based on the simulation results and the multiple objectives corresponding to the product requirement information to obtain optimized design parameters, and feeds the optimized design parameters back to the simulation verification agent, triggering the simulation verification agent to update the design parameters in each physical mathematical model and re-perform simulation calculations to generate updated simulation results, until the multi-objective optimization meets the convergence condition, and obtains the target design parameters of each physical mathematical model. Based on each physical mathematical model and the target design parameters, a target development plan is generated, realizing full-process automation from requirement to simulation, supporting multi-disciplinary parallel simulation and closed-loop iterative optimization of design parameters, significantly shortening the development cycle, reducing R&D costs, and improving R&D efficiency. Attached Figure Description

[0101] 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.

[0102] Figure 1 This is a block diagram of a virtual physical system for industrial product development based on a large model, as shown in one embodiment.

[0103] Figure 2 This is a flowchart illustrating a system development method based on a large model in one embodiment;

[0104] Figure 3 This is a flowchart of the closed-loop iterative optimization process in one embodiment;

[0105] Figure 4 Here is a flowchart of the resource configuration steps in one embodiment;

[0106] Figure 5 This is a schematic diagram of a V-shaped process in one embodiment;

[0107] Figure 6 This is a schematic diagram of an AI agent matrix framework in one embodiment;

[0108] Figure 7 This is a knowledge transfer mechanism from the source domain (automotive) to the target domain (photovoltaics / robotics / aerospace) in one embodiment;

[0109] Figure 8 This is a structural block diagram of a system development apparatus based on a large model in one embodiment;

[0110] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0111] 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.

[0112] 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.

[0113] The system development method based on a large model provided in this application can be applied to, for example... Figure 1 The diagram shows a virtual physical system based on a large model for industrial product development. This virtual physical system for industrial product development includes an application service layer, an intelligent enhancement layer, and a physical base layer.

[0114] The physics foundation layer integrates fundamental physical laws from continuum mechanics, electromagnetism, thermodynamics, and fluid mechanics, constructing a unified differential equation description system that supports cross-scale modeling from small molecular scales to macroscopic systems. For example, in the context of automotive development, this layer includes the Navier-Stokes equations describing fluid motion and the mechanical equilibrium equations describing the deformation of solid structures.

[0115] The multiphysics coupling control equations for the physical base layer are as follows:

[0116] Formula (7-1) Continuity equation (mass conservation):

[0117]

[0118] in, Fluid density (unit: ); It is a velocity vector; For time; This is the divergence operator. The equation represents the change in control volume mass per unit time equal to the net outflow mass, and is fundamental to automotive CFD analysis.

[0119] Momentum conservation equation (Navier-Stokes equation):

[0120]

[0121] in, For pressure ( ); Dynamic viscosity ( ); For volume force ( ); This is the Laplace operator. The equation describes the conservation of fluid momentum and is used in automotive aerodynamics and thermal management simulations.

[0122] Energy conservation equation:

[0123]

[0124] in, Total energy per unit mass ( ); Thermal conductivity ( ); For temperature ( ); For heat source item ( This equation describes energy conservation and is used in simulations of battery thermal management, engine thermal load, etc.

[0125] Structural mechanics equilibrium equations:

[0126]

[0127] in, For Cauchy stress tensor; It is a volume force; Let be the displacement vector. This equation describes the dynamic response of the structure under load and is used for vehicle body durability, crash, and NVH analysis.

[0128] The intelligent enhancement layer integrates physical information artificial intelligence (PIA) with multi-objective optimization algorithms. This layer receives physical constraints from the physical base layer and task instructions from the application service layer. Guided by physical laws, it performs rapid prediction or simulation using AI algorithms, while simultaneously employing multi-objective optimization algorithms (such as NSGA-III) to find the optimal balance among conflicting design objectives (e.g., performance, cost, weight). The AI ​​algorithm can be a PINN network. The PINN network structure includes an input layer, hidden layers, and an output layer. The input layer dimension represents the design parameters, the hidden layers use a multilayer perceptron with swish or tanh activation functions, and the output layer represents the physical field distribution. The loss function includes data loss and physical constraint loss, with weights λ ranging from 0.1 to 10. The NSGA-III algorithm includes a reference point generation module employing a hyperplane uniform distribution strategy with Δ = 0.1–0.2; the population size is 200–500, and the number of iterations is 1000–3000; constraints are handled using a penalty function method with penalty function weights ranging from 0.01 to 10.

[0129] Furthermore, this system employs a Physical Information Neural Network (PINN) to achieve deep fusion of mechanisms and data, with its core loss function being:

[0130] PINN total loss function:

[0131]

[0132] in, Data-driven loss measures the deviation between predicted and measured values. The physical constraint loss measures whether the prediction satisfies the governing equations. For boundary condition loss; , , These are weighting coefficients, and the weights are... The range of values ​​is The three loss contributions are adjusted adaptively to balance them.

[0133] Physical constraint loss:

[0134]

[0135] in, To coordinate the control equations at the points residuals at the location; For neural network parameters; The number of points is used for matching. This loss forces the neural network predictions to satisfy physical laws, ensuring physical consistency.

[0136] The application service layer serves as the interface between the system and users, as well as other external systems. This layer provides a series of standardized microservice interfaces, including at least design, simulation, verification, optimization, and knowledge interfaces. Users can submit design requirements, initiate simulation tasks, view optimization results, and access the knowledge base through these interfaces, without needing to concern themselves with the underlying complex physical models and algorithm implementations. Furthermore, this application service layer defines a unified Physical Description Language (PDL), including tags such as ``, `<body>`, and `<body>`, supporting data exchange and model integration across different CAD / CAE systems. It also provides a digital twin mapping interface, supporting real-time data synchronization and continuous optimization between virtual models and physical products.

[0137] In this system, the three-tier architecture communicates via a standardized data protocol (such as the MCP protocol). The application service layer translates user requirements into standard instructions and passes them to the intelligent enhancement layer; the intelligent enhancement layer calls the physical model of the physical base layer for calculations and feeds the results back to the user through the application service layer. Furthermore, by constructing a three-tier architecture of physical base layer, intelligent enhancement layer, and application service layer, clear decoupling and organic integration are achieved from physical law description and AI fusion optimization to user interaction. This architecture provides a solid, scalable, and intelligent foundation for the refactoring of the V-shaped development process, effectively solving the problems of toolchain silos and data fragmentation in traditional development, and enabling physical consistency and multi-objective optimization, thereby significantly improving development efficiency and design quality.

[0138] In one exemplary embodiment, such as Figure 2 As shown, a system development method based on a large model is provided, which can be applied to... Figure 1 The following explanation uses a virtual physical system for industrial product development as an example, including steps 202 to 208. Wherein:

[0139] S202: Receive product requirement information input in natural language through the application service layer.

[0140] The application service layer, as the top-level interface facing users, is responsible for receiving externally inputted product requirement information. Product requirement information refers to the user's expectations regarding the performance, functionality, cost, size, weight, and other aspects of the target product. This product requirement information can be text entered by R&D engineers through a graphical user interface (GUI), or structured or unstructured data imported from a product lifecycle management (PLM) system or requirement management tools (such as Polarion or DOORS).

[0141] Users submit requirements (e.g., design an electric SUV with a range of 800km and a drag coefficient Cd≤0.22) through the design / simulation microservice interface of the application service layer. The application service layer encapsulates unstructured requirements into standard requirement packages using a unified Physical Description Language (PDL) and uploads them to the virtual physical chief engineer agent via the MCP protocol. The chief engineer agent creates a requirement work order, initializes the project context and traceable baseline, and dispatches the requirement parsing subtask to the requirement analysis agent (i.e., one of the target agents).

[0142] S204: The target agent of the intelligent enhancement layer obtains the physical and mathematical models based on the product requirement information. The physical and mathematical models are used to describe the design parameters and constraints of each subsystem.

[0143] The target agent of the intelligent enhancement layer is used to parse product demand information and transform it into computable physical and mathematical models.

[0144] A physical mathematical model is a set of digital expressions used to describe the behavior, performance, and design constraints of a target product or its subsystems under specific physical fields (such as structural mechanics, fluid mechanics, thermodynamics, and electromagnetism). It provides a rigorous and executable mathematical description for simulation calculations. This physical mathematical model includes at least: governing equations (such as partial differential equations) describing the physical laws, boundary and initial conditions defining the solution range, key design parameters (variables) that designers can adjust to optimize performance, and design constraints that these parameters must satisfy (such as upper stress limits and lower temperature limits). These physical mathematical models can be for the entire vehicle or for individual subsystems (such as the body, battery pack, motor, and chassis); no specific limitations are imposed here.

[0145] In one embodiment, the target agent integrates a physical model library containing pre-built template models for different product types (such as sedans and SUVs) and subsystems. The target agent calls the corresponding template by matching keywords (such as SUV and electric vehicle) in the product requirement information, and assigns and adjusts the design parameters and constraints in the template according to the specific indicators in the requirements (such as 800km range), thereby generating an initial physical mathematical model.

[0146] In other embodiments, the target agent initiates a cascading process: First, it invokes a requirements analysis agent to transform the natural language requirements into a structured system modeling language (such as SysML) model; then, it invokes an architecture design agent, which, based on the SysML model, retrieves the corresponding governing equations, material properties, etc., from the physical base layer, and automatically assembles and generates various physical mathematical models through a series of inferences and mappings. During this process, if multiple subsystems are involved, multiple corresponding physical mathematical models will be generated.

[0147] S206: The simulation verification agent of the intelligent enhancement layer calls various simulation tools of the physical base layer to perform simulation calculations on each physical mathematical model in parallel and generate simulation results.

[0148] The simulation verification agent in the intelligent enhancement layer is an AI agent specifically responsible for managing and executing simulation tasks.

[0149] The physical base layer is the underlying support of the system, integrating various proven solvers (i.e., simulation tools), such as SU2 or OpenFOAM for computational fluid dynamics, CalculiX or Abaqus for structural finite element analysis, and Adams for multibody dynamics. These tools are encapsulated and have standardized calling interfaces.

[0150] The simulation verification agent receives multiple physical mathematical models from the target agent and submits them to the most suitable simulation tool in the physical base layer. Optionally, this simulation verification agent supports parallel computing mechanisms, initiating simulation tasks for multiple independent physical mathematical models simultaneously, rather than sequentially waiting for one to complete before starting the next. This parallel processing maximizes the utilization of high-performance computing clusters (HPC) or cloud computing resources. Upon completion of each simulation task, corresponding simulation result data is generated; for example, CFD simulations may output pressure contour maps and velocity vector fields, while structural simulations may output stress distributions and deformations.

[0151] In one embodiment, the simulation verification agent automatically selects the optimal solver based on the subject type (e.g., fluid, structure) and accuracy requirements of each physical mathematical model. In another embodiment, the simulation verification agent first assesses the urgency and computational complexity of the task, then requests the most economical computing resources (e.g., GPU or CPU nodes) from the resource scheduling module, and finally starts the simulation on those resources.

[0152] S208: Through the chief engineer agent of the intelligent enhancement layer, multi-objective optimization is performed based on the simulation results and product requirement information to obtain the optimized design parameters. The optimized design parameters are then fed back to the simulation verification agent, which triggers the simulation verification agent to update the design parameters in each physical and mathematical model and re-perform simulation calculations to generate updated simulation results. This process continues until the multi-objective optimization meets the convergence condition, at which point the target design parameters of each physical and mathematical model are obtained. Based on each physical and mathematical model and the target design parameters, a target development plan is generated.

[0153] The chief agent in the intelligent enhancement layer acts as the central coordinator of the entire multi-agent system, responsible for allocating tasks and executing core optimization decision-making tasks.

[0154] Multi-objective optimization refers to the process of simultaneously considering multiple potentially conflicting design objectives (e.g., maximizing driving range and minimizing manufacturing cost) and seeking the optimal balance when designing a complex system. In this embodiment, these objectives are derived from product requirement information.

[0155] The chief engineer agent compiles the first-round simulation results from various physical and mathematical models and compares these results with multiple objectives in the product requirements information. A multi-objective optimization algorithm (e.g., the non-dominated sorting genetic algorithm NSGA-III or the decomposition-based multi-objective evolutionary algorithm MOEA / D) is used to evaluate the current set of design parameters. The evaluation results in a new set of optimized design parameters that aim to achieve a better trade-off among the objectives. The chief engineer agent sends these new design parameters to the simulation verification agent through a feedback loop.

[0156] After receiving the new design parameters, the simulation verification agent updates the corresponding design parameter values ​​in the previous physical and mathematical models (e.g., updating the body panel thickness from 1.0 mm to 0.9 mm) and calls the simulation tool to perform simulation calculations, thus forming a closed-loop iteration. This closed loop will continue to run until one or more convergence conditions are met. The convergence condition is a pre-defined stopping criterion, such as after several consecutive iterations, the changes in the values ​​of all objective functions are very small (less than a preset threshold), or a feasible solution that satisfies all objective thresholds has been found. When the iteration terminates, the finally generated optimal design parameters are determined as the target design parameters.

[0157] Finally, the chief engineer will integrate all physical and mathematical models and their corresponding target design parameters, along with the optimization process and key decision-making rationale, to form a complete target development plan. This plan can be directly used as input for subsequent detailed design, engineering drawing, or code generation.

[0158] The aforementioned system development method based on large models receives product requirement information input in natural language through the application service layer; the target agent of the intelligent enhancement layer obtains various physical and mathematical models based on the product requirement information, which are used to describe the design parameters and constraints of each subsystem; the simulation verification agent of the intelligent enhancement layer calls various simulation tools in the physical base layer to perform simulation calculations on each physical and mathematical model in parallel and generate simulation results; the chief engineer agent of the intelligent enhancement layer performs multi-objective optimization based on the simulation results and the multiple objectives corresponding to the product requirement information to obtain optimized design parameters, and feeds the optimized design parameters back to the simulation verification agent, triggering the simulation verification agent to update the design parameters in each physical and mathematical model and re-perform simulation calculations to generate updated simulation results, until the multi-objective optimization meets the convergence condition, and obtains the target design parameters of each physical and mathematical model. Based on each physical and mathematical model and the target design parameters, a target development plan is generated, realizing full-process automation from requirements to simulation, supporting multi-disciplinary parallel simulation and closed-loop iterative optimization of design parameters, significantly shortening the development cycle, reducing the R&D costs caused by repeated manual coordination and trial and error, and significantly improving the performance and quality of the final design plan.

[0159] In some optional embodiments, the chief engineer agent of the intelligent enhancement layer performs multi-objective optimization based on the simulation results and product requirement information to obtain optimized design parameters. This includes: generating reference target values ​​for the multiple objectives corresponding to the product requirement information based on the physics indicators in the simulation results, and determining each non-compliant objective based on the multiple objectives corresponding to the product requirement information and the reference target values; calculating the penalty value for each non-compliant objective using the penalty function method, and determining the optimization direction based on the penalty value; and adjusting the design parameters in multiple directions using a multi-objective optimization algorithm based on the optimization direction to generate optimized design parameters.

[0160] Combination Figure 3 As shown, Figure 3 This is a flowchart of a closed-loop iterative optimization process in one embodiment. In this embodiment, the entire process includes: Construction Phase: Generative AI based on physical constraints automatically generates an initial design scheme. For example, after the requirements analysis agent generates a SysML model, the architecture design agent automatically generates an initial architecture design based on the model. Computation Phase: A distributed physics engine performs parallel computation of multidisciplinary performance indicators. For example, the physics engine agent calls SU2 for fluid simulation and CalculiX for structural simulation, exchanging data through the PDL interface. Critique Phase: A multi-dimensional evaluation system is established, including performance, cost, reliability, and manufacturability. For example, the AI ​​agent evaluates the computation results using the NSGA-III algorithm and generates optimization suggestions. Repetition Phase: Intelligent iterative optimization is performed based on the NSGA-III algorithm to find the best feasible solution within a limited interaction budget. For example, the code generation agent generates updated C / C++ code based on the optimization suggestions, and the testing agent verifies the code functionality.

[0161] When performing multi-objective optimization, the chief engineer's agent first needs to evaluate the current design scheme (i.e., the current set of design parameters). The agent extracts quantified values ​​of key performance indicators from the simulation results; these values ​​are called physical performance indicators (PMIs), such as the maximum stress value obtained from structural simulation (unit: MPa), the drag coefficient (Cd) obtained from CFD simulation, or the energy consumption per 100km obtained from vehicle simulation (unit: kWh / 100km). Then, the agent compares the current values ​​of these PMIs (i.e., reference target values) with the expected values ​​of each target explicitly defined in the product requirements information. For example, the requirement might be a drag coefficient Cd ≤ 0.22. If the current simulation yields a Cd value of 0.25, then the low drag target is not met. Through this step-by-step comparison, the agent can identify all targets that fail to meet the requirements and mark them as non-compliant targets.

[0162] After identifying non-compliant targets, the chief engineer agent uses a penalty function method to quantify the severity of these non-compliance situations. Specifically, for solutions that violate constraints (i.e., targets are not met), a penalty term is added to their fitness function (in this scenario, the evaluation function of the optimization algorithm) to reduce the score of the solution. The size of the penalty term (i.e., the penalty value) is usually proportional to the degree of constraint violation (i.e., the difference between the non-compliant target and the target value). For example, if the target value of Cd is 0.22 and the current value is 0.25, the difference is 0.03, resulting in a penalty value positively correlated with 0.03; if the current value is 0.30, the difference is larger, and the penalty value is also larger. The chief engineer agent calculates the penalty value for each non-compliant target based on its type (e.g., upper limit constraint, lower limit constraint) and a preset penalty coefficient. Then, the optimization algorithm determines the optimization direction for the next round of search based on the distribution of these penalty values, i.e., in which direction adjusting the design parameters can most effectively reduce the total penalty value, thereby bringing the design solution closer to the compliant state from the non-compliant state.

[0163] Finally, the chief engineer combines the search direction of its built-in multi-objective optimization algorithm (such as NSGA-III) with the optimization direction guided by the penalty value. Based on this combined direction, the algorithm makes multi-dimensional and systematic adjustments to the current design parameters. This adjustment is multi-directional, meaning the algorithm simultaneously attempts to modify multiple design parameters (such as changing plate thickness, cross-sectional shape, and material grade) to explore different design possibilities. After one round of adjustments, the algorithm generates a new set of optimized design parameters that aim to reduce penalties for non-compliant targets while maintaining or improving the performance of other compliant targets. This new set of parameters is then output to initiate the next round of simulation iterations.

[0164] NSGA-III uses a strategy of uniformly distributing reference points on the hyperplane to generate reference directions: the hyperplane distribution of reference points is as follows:

[0165]

[0166] in, To divide parameters, Pick ; For the first The strategy guarantees that the reference point is on each target axis. The target species are evenly distributed in the spatial dimension, and the population size is 200-500.

[0167] Target normalization:

[0168]

[0169] in, For the first Original values ​​for each target; For the ideal point; The zenith is the highest point. After normalization, all targets are on a comparable scale, which facilitates the correlation of reference points and the selection of niches.

[0170] Constraint handling (penalty function method):

[0171]

[0172] in, For the first One constraint function; For the penalty function weight, take The constrained optimization is transformed into unconstrained optimization by using a penalty function, with 1000-3000 iterations until the Pareto front converges.

[0173] This embodiment introduces a penalty function-based objective evaluation mechanism, transforming the fuzzy problem of failing to meet objectives into a clear, quantifiable problem of penalty values. This allows multi-objective optimization algorithms to efficiently drive the search process towards satisfying all constraints within a unified mathematical framework. This approach avoids manually setting complex priorities or weights, automatically and quickly finding feasible optimal solutions, significantly improving the robustness and efficiency of multi-objective optimization.

[0174] In some optional embodiments, the target design parameters of each physical mathematical model are obtained after the multi-objective optimization meets the convergence condition. This includes: generating new reference target values ​​for the multiple objectives corresponding to the product demand information based on the physics discipline indicators in the updated simulation results; and using the optimized design parameters determined in this instance as the target design parameters of each physical mathematical model when all the new reference target values ​​meet the multiple objectives corresponding to the product demand information, or when the change in the physics discipline indicators in adjacent updated simulation results is less than a preset convergence threshold.

[0175] The convergence condition consists of two logical conditions; the iteration terminates when either one is satisfied.

[0176] Condition 1: All targets met. This means that all new reference target values ​​satisfy the multiple targets corresponding to the product requirements. In other words, in this iteration, after the chief engineer evaluated all simulation results, it was found that every physical metric (such as Cd value, maximum stress, and total weight) had reached the preset target thresholds in the product requirements (e.g., Cd ≤ 0.22, maximum stress ≤ 250 MPa, weight ≤ 1800 kg). In this case, it indicates that a feasible solution that fully meets all design requirements has been found, and the optimization process can naturally end.

[0177] Condition 2: Minimal Change Condition. This means that the change in the physics subject indicators in adjacent updated simulation results is less than a preset convergence threshold. In other words, although all objectives may not have been fully achieved (e.g., Cd is 0.221, slightly higher than 0.22), the absolute values ​​of the changes in all key physics subject indicators output in two consecutive iterations (i.e., the current iteration and the previous iteration) are already less than a pre-set, very small positive threshold (e.g., the change in Cd value < 0.001, the change in weight < 0.1 kg). This indicates that no matter how the parameters are adjusted, the simulation results have almost stopped changing, indicating that the optimization algorithm has reached the limit of the problem (i.e., the Pareto front), and continuing the calculation will not make substantial progress. To save computational resources, the iteration should be terminated at this point.

[0178] Once any of the above conditions are met, the chief engineer agent determines that the optimization has converged and identifies the set of design parameters used in this iteration that produced the current optimal result as the final target design parameters.

[0179] Specifically, design parameters (such as cross-sectional dimensions, material grade, control gain, and heat dissipation structure) are modified based on the optimization results to generate a new generation of design parameter sets. The modified design parameters are then fed back into step five for resimulation, forming a simulation-evaluation-modification-resimulation closed loop until the Pareto front converges or the index meets the target threshold.

[0180]

[0181] in, For the first Substitute the target vector, The convergence threshold, To satisfy the constraints, For the first There are several target thresholds. The optimal design parameter set is output when any termination condition is met.

[0182] This embodiment provides a clear, practical, and efficient standard for the automatic termination of the optimization process by setting two complementary convergence conditions: achieving all criteria and minimizing the amount of change. The first condition ensures the validity of the output solution, while the second condition ensures the efficiency of the computation process, avoiding wasting resources on invalid searches.

[0183] In some optional embodiments, the method further includes: performing optimization layer risk identification by a chief engineer agent during the multi-objective optimization iteration process, wherein the optimization layer risk identification includes at least one of parameter sensitivity risk identification, constraint out-of-bounds risk identification, and robustness risk identification; and generating risk registration items based on the results of the optimization layer risk identification.

[0184] Each time design parameters are evaluated or adjusted, the chief engineer's representative activates a risk analysis engine to probe the current candidate solutions and their surrounding areas. This risk analysis engine supports one or more of the following risk identification modes:

[0185] Parameter sensitivity risk identification: This model applies a small perturbation (e.g., ±1%) to the value of each current design parameter and then uses a Physical Information Neural Network (PINN) or a fast surrogate model to predict the impact of this perturbation on key performance indicators (such as maximum stress). If a small parameter change leads to drastic fluctuations in performance indicators (i.e., high gradients), it indicates a high sensitivity risk to the design, meaning that there is a high probability of performance failure due to material tolerances or operating condition fluctuations during manufacturing or actual use.

[0186] Constraint Boundary Exceedance Risk Identification: This mode monitors design parameter values ​​and simulation result values ​​in real time during the optimization process, checking whether they are approaching or have exceeded their physical limits or safety boundaries. For example, if the stress value of a design scheme has reached 98% of the material's yield strength, it indicates an extremely high risk of constraint breach, with almost no safety margin.

[0187] Robustness Risk Identification: This model uses methods such as Monte Carlo simulation or range analysis to assess whether the design scheme can maintain its key performance when there are expected fluctuations in design parameters (such as manufacturing tolerances) or uncertainties in the external environment (such as changes in wind speed or temperature). If the performance distribution range is too wide or a large proportion falls into the substandard area, it indicates that the scheme is not robust enough and there is a risk.

[0188] The chief engineer will generate a standardized risk register for any risks identified using the patterns described above. This register can be a structured data object, including at least the following fields: risk type (e.g., high parameter sensitivity), risk description (e.g., peak impact force changes by more than 15% when the front bumper beam thickness varies between 1.8mm and 2.0mm), risk level (high, medium, low), associated design parameters or objectives, and the time of discovery (iteration round). These risk registers will be continuously accumulated and serve as part of the project's traceability baseline, providing crucial decision-making references for engineers when finalizing the design.

[0189] In some alternative embodiments, combined with Figure 4 As shown, Figure 4 This is a flowchart of a resource configuration step in one embodiment. Before the simulation verification agent of the intelligent enhancement layer calls various simulation tools in the physical base layer to perform parallel simulation calculations on each physical mathematical model and generate simulation results, this resource configuration step includes:

[0190] S402: The simulation verification agent of the intelligent enhancement layer generates corresponding simulation tasks based on each physical and mathematical model and classifies each simulation task.

[0191] Upon receiving each physical mathematical model, the simulation verification agent does not immediately submit it to the simulation tool for execution. Instead, it first creates a corresponding simulation task object for each model. This task object is an abstract encapsulation of a simulation calculation execution, including all metadata required for that calculation, such as: the file path of the physical mathematical model to be solved, the required computational accuracy, the estimated computation time, and the type of software license required. Subsequently, the simulation verification agent systematically classifies these simulation tasks based on this metadata, especially according to the computational characteristics of the task (such as whether it belongs to fluid dynamics calculation, structural finite element calculation, heat conduction calculation, or electromagnetic field calculation) and resource requirements (such as whether GPU acceleration is supported, the required memory size, and whether it is latency-sensitive). For example, all tasks requiring large-scale parallel matrix operations can be classified as GPU-accelerated, and all tasks requiring frequent I / O operations can be classified as memory-intensive.

[0192] S404: Divide the computing resources of the physical base layer into multiple resource levels.

[0193] The system abstracts and hierarchizes all hardware computing resources managed by the physical infrastructure layer. These resources typically comprise a heterogeneous high-performance computing cluster, including different types of computing nodes. The system divides these resources into multiple logical resource tiers. For example, they can be divided as follows: the first tier consists of nodes equipped with high-performance GPUs (Graphics Processing Units), suitable for highly parallel computing; the second tier consists of general-purpose computing nodes equipped with high-frequency CPUs (Central Processing Units), suitable for complex logic control and serial computing; and the third tier consists of edge computing nodes deployed in R&D labs or production sites, suitable for real-time simulation tasks requiring low-latency responses.

[0194] S406: Match each classified simulation task with each hierarchical resource level, and use an optimization algorithm to solve for at least two of the following as optimization objectives: total task completion delay, computing cost, and simulation quality, to generate a resource allocation scheme for each simulation task.

[0195] After completing task classification and resource tiering, the simulation verification agent initiates a resource scheduling optimizer. This optimizer constructs a mathematical model defining the allocation of tasks to different resource tiers as an optimization problem. The optimization objective of this problem includes at least two or more of the following three: 1) minimizing the total completion latency of all simulation tasks (i.e., the total time from the start of the first task to the end of the last task); 2) minimizing computing costs (i.e., the economic costs of using these computing resources, such as electricity and cloud service fees); 3) maximizing simulation quality (i.e., ensuring that tasks with high accuracy requirements are allocated to resources with better performance). The optimizer employs a multi-objective optimization algorithm (such as NSGA-III or MOEA / D) to solve this combinatorial optimization problem, generating a set of Pareto optimal solutions that achieve the best trade-offs among the selected objectives. The optimal solution is then selected as the resource allocation scheme for each simulation task.

[0196] S408: Determine the computing resources corresponding to each simulation task based on the resource allocation scheme.

[0197] Finally, the simulation verification agent binds each simulation task to a specific computing resource (e.g., scheduling task A to GPU node 3 and task B to CPU node 7) based on the resource allocation scheme generated in the previous step, and then starts the actual simulation calculation process.

[0198] This embodiment solves the problems of wasted computing resources and low simulation efficiency caused by improper resource scheduling in traditional development by introducing an intelligent resource allocation mechanism that includes task classification, resource grading, optimized matching, and solution implementation. This mechanism can perform optimal matching based on task characteristics and resource capabilities, minimizing overall simulation time and reducing computational costs while ensuring simulation quality.

[0199] In some optional embodiments, the resource hierarchy includes GPU clusters, CPU clusters, and edge computing nodes. Each classified simulation task is matched with each hierarchically divided resource hierarchy, and an optimization algorithm is used to solve for at least two of the following as optimization objectives: total task completion latency, computing cost, and simulation quality. This generates a resource allocation scheme for each simulation task, including: matching fluid simulation and neural network training tasks to GPU clusters, structural simulation and electromagnetic simulation tasks to CPU clusters, and software-in-the-loop testing and hardware-in-the-loop testing tasks to edge computing nodes based on the subject type and fidelity of each simulation task, generating an initial resource allocation scheme. The fidelity of the simulation task is the credibility of the physical mathematical model. The initial resource allocation scheme is solved using an optimization algorithm with at least two of the following as optimization objectives: total task completion latency, computing cost, and simulation quality, outputting the resource allocation scheme for each simulation task.

[0200] The discipline type of the simulation task refers to the physics discipline to which the simulation belongs, such as fluid mechanics, structural mechanics, electromagnetism, or thermodynamics. Fidelity is a measure of the credibility or accuracy level of a physical mathematical model. High-fidelity models typically use finer meshes and smaller computational steps, thus requiring more computation and taking longer. In this step, the simulation verification agent generates an initial allocation scheme based on a pre-defined rule base: fluid simulations and neural network training, due to their highly parallel matrix operation characteristics, are best suited for running on GPU clusters with a large number of computing cores; structural simulations and electromagnetic simulations typically involve solving large sparse matrices, making them more suitable for running on general-purpose CPU clusters; while software-in-the-loop testing and hardware-in-the-loop testing have strict time requirements for response latency, therefore they need to be deployed on edge computing nodes closest to the controlled object.

[0201] After generating the initial allocation scheme, the system does not adopt it directly. Instead, it inputs it as an initial solution into a multi-objective optimization algorithm. This algorithm iteratively optimizes the initial scheme with the objectives of minimizing the total task completion latency, minimizing computational cost, and maximizing simulation quality. During optimization, the algorithm may make some counterintuitive adjustments. For example, it might temporarily schedule a structural simulation task originally matched to a CPU cluster to a GPU cluster if it detects that the GPU cluster is idle and the task can benefit from GPU acceleration, thereby shortening the total completion latency. After a finite number of iterations, the algorithm outputs a final resource allocation scheme that achieves the best balance among the multiple optimization objectives.

[0202] For ease of understanding, this application categorizes each simulation task by discipline (CFD / structural / electromagnetic / thermal), fidelity (low / medium / high), and time criticality (online / offline) to obtain a task profile. A GPU cluster is configured for PINN training and high-fidelity CFD solving; a CPU cluster is used for structural / electromagnetic finite element solving; and edge real-time nodes are used for SIL / HIL online tasks, thus obtaining a resource profile. Based on the matching of task profiles and resource profiles, and combined with NSGA-III to optimize among throughput, latency, and cost objectives, a resource allocation scheme is proposed.

[0203]

[0204] in, For resource allocation schemes, For the total completion time of the task, For computing power cost, For simulation quality (fidelity). , , As weight.

[0205] Finally, resources are dynamically reallocated based on simulation execution feedback (uneven load, convergence anomalies) to ensure optimal overall throughput and latency.

[0206] In some optional embodiments, the simulation verification agent of the intelligent enhancement layer calls various simulation tools in the physical base layer to perform simulation calculations on each physical mathematical model in parallel and generate simulation results. This includes: determining the simulation requirements corresponding to each physical mathematical model; determining the simulation tool type based on the simulation requirements; calling the simulation tool in the physical base layer that corresponds to the simulation tool type through the simulation verification agent of the intelligent enhancement layer, and performing simulation calculations on the physical mathematical model through the called simulation tool to generate simulation results.

[0207] Upon receiving the physical mathematical model to be simulated, the simulation verification agent first analyzes it to determine the simulation requirements. Simulation requirements are a complete description of the functional and performance requirements of a simulation task, including at least the following information: the type of physical field to be solved (e.g., steady-state flow field or transient temperature field), the required computational accuracy (e.g., fast estimation for concept screening or high-fidelity calculation for final verification), the acceptable upper limit of computation time, and whether a specific format of intermediate or final results is required. By clearly defining the simulation requirements, the agent can provide an accurate basis for subsequent tool selection.

[0208] After clarifying the simulation requirements, the simulation verification agent determines the type of simulation tool to be invoked for this simulation based on a pre-defined requirement-tool type mapping rule base. The simulation tool type is a classification abstraction of different solver tools in the physics base layer. It does not refer to a specific commercial software, but rather describes the capabilities and applicable scenarios of a class of tools. For example, if the simulation requirement is high-precision transient compressible fluid flow, the corresponding simulation tool type is a high-fidelity computational fluid dynamics solver; if the simulation requirement is to quickly estimate the maximum stress of a structure under static load, the corresponding type is a fast structural analysis solver; if the simulation requirement is to predict the temperature field distribution within 1 second, the corresponding type might be a neural network-based proxy model.

[0209] Finally, the simulation verification agent, based on the simulation tool type determined in the previous step, searches for and invokes a specific simulation tool instance in the physical base layer that matches that type. For example, if the type is a high-fidelity computational fluid dynamics solver, the agent might choose to invoke the SU2 or OpenFOAM solver already encapsulated in the physical base layer; if the type is a fast structural analysis solver, it might choose to invoke a lightweight solver based on a reduced-order model. Then, the agent inputs the data from the physical mathematical model into the tool, drives it to perform calculations, and ultimately obtains and outputs the simulation results.

[0210] In some optional embodiments, the simulation verification agent of the intelligent enhancement layer calls the simulation tool in the physical base layer corresponding to the simulation tool type, and performs simulation calculations on the physical mathematical model through the called simulation tool to generate simulation results. This includes: when the simulation tool type is a physical solver, the simulation verification agent of the intelligent enhancement layer calls the physical solver of the physical base layer, and performs simulation calculations on the physical mathematical model through the scheduled physical solver to generate simulation results; when the simulation tool type is a neural network, the simulation verification agent of the intelligent enhancement layer calls the neural network of the physical base layer, and performs simulation calculations on the physical mathematical model through the neural network to generate simulation results; when the simulation tool type is a hybrid type, the simulation verification agent of the intelligent enhancement layer calls the neural network of the physical base layer, and performs simulation calculations on the physical mathematical model through the neural network to generate candidate simulation results, and the simulation verification agent of the intelligent enhancement layer calls the physical solver of the physical base layer, and performs simulation calculations on the physical mathematical model corresponding to the candidate simulation results through the scheduled physical solver to generate simulation results.

[0211] A physics solver is a traditional simulation software that directly solves a system of partial differential equations based on first principles (such as the laws of conservation of mass, momentum, and energy) using numerical discretization methods (such as the finite element method and the finite volume method). When the simulation tool type is determined to be a physics solver, the simulation verification agent will directly call this type of tool in the physics base layer. This approach is suitable for situations requiring precise quantification of key performance indicators or for providing high-confidence simulation data to meet safety certification requirements (such as ISO 26262).

[0212] In this context, "neural network" specifically refers to a pre-trained deep learning model used to approximate the distribution of a specific physical field, such as the Physical Information Neural Network (PINN). When the simulation tool type is neural network, the agent invokes this network model. The input to this model is the parameters of the physical mathematical model (such as the values ​​of design variables and boundary conditions), and the output is the directly predicted physical field results (such as maximum stress values ​​and average temperatures). This approach belongs to the data-driven agent model method, which is extremely fast, typically several orders of magnitude faster than traditional physical solvers. It is suitable for concept screening and early exploration phases that require rapidly traversing a large number of design schemes.

[0213] Hybrid type is a cascaded mode designed to combine the speed advantage of neural networks with the accuracy advantage of physics solvers. In this mode, the simulation verification agent first invokes the neural network to perform one or more rapid inferences on the original physical mathematical model, generating a set of candidate simulation results. Then, the agent filters these candidates to select the most valuable or most precisely validated design solutions (e.g., solutions near the Pareto front). Finally, the agent invokes the high-fidelity physics solver to perform precise calculations only on these selected solutions, thereby significantly reducing the number of costly high-fidelity simulations while ensuring the accuracy of the final results.

[0214] In the above embodiments, PINN (fast but approximate) is selected for the concept / solution exploration stage, which requires rapid traversal of a large design space, inverse problem solving, and high real-time requirements; traditional solvers (slow but accurate) are selected for the detailed verification stage, which requires high-fidelity full-field results and safety-critical certification; a hybrid approach can also be adopted, using PINN to quickly screen candidate solutions and traditional solvers to accurately verify shortlisted solutions, balancing efficiency and accuracy (corresponding to the calculation stage mentioned above).

[0215] In some optional embodiments, after receiving product requirement information input in natural language through the application service layer, the method further includes: encapsulating the product requirement information through the application service layer to obtain a target requirement package described by a preset physical description language; obtaining various physical mathematical models based on the product requirement information through the target agent of the intelligent enhancement layer, including: calling the requirement analysis agent through the chief engineer agent of the intelligent enhancement layer to parse the target requirement package and generate an initial system model; and calling the architecture design agent through the chief engineer agent of the intelligent enhancement layer to process the initial system model and generate various physical mathematical models.

[0216] Combination Figure 5 As shown, Figure 5 This is a schematic diagram of a V-shaped process in one embodiment, wherein the V-shaped process includes:

[0217] Left-side decomposition: This involves automatically generating executable physical and mathematical models from product requirements, and using generative AI to automatically construct the system architecture, achieving intelligent mapping between requirements and models. For example, the requirement parsing agent uses natural language processing technology to parse user requirements, automatically generating quantifiable design metrics and constraints, and outputting a SysML v2 model.

[0218] Bottom-level verification: Establish a high-fidelity physical simulation environment that supports hybrid verification of Hardware-in-the-Loop (HIL) and Software-in-the-Loop (SIL) to ensure consistency between the digital model and the physical world. For example, the simulation verification agent calls CarMaker / SUMO via the MCP protocol to perform SIL / HIL tests and feeds the results back to the overall AI agent.

[0219] Right-side integration: The NSGA-III multi-objective optimization algorithm automatically finds the Pareto optimal solution, achieving intelligent integration from subsystems to the entire vehicle / machine. For example, the chief engineer's intelligent agent uses the NSGA-III algorithm to find the optimal balance among multiple objectives such as demand, performance, cost, and reliability.

[0220] The Physical Description Language (PDL) is a structured, machine-readable data description specification used to uniformly express various types of information related to physical systems. It is similar to common data exchange formats in the computer field (such as XML or JSON), but its semantics are specifically defined for the engineering physics domain. PDL includes a set of predefined tags and data structures used to describe key elements in product requirements, such as product name, performance targets, design constraints, and operating environment conditions. Upon receiving a natural language requirement, the application service layer invokes a PDL encapsulation engine to convert the unstructured or semi-structured requirement information into a target requirement package file conforming to the PDL specification. This encapsulation process ensures the integrity, parsability, and unambiguity of the requirement information in subsequent processes.

[0221] The Requirements Analysis Agent is an AI agent specifically designed to process and analyze product requirements. The Chief Engineer Agent dispatches target requirement packages in PDL format to the Requirements Analysis Agent. This agent parses the PDL file, extracting requirement items, metrics, and constraints, and uses natural language processing and systems engineering modeling techniques to transform it into a structured initial system model. This initial system model is typically expressed using a systems modeling language (such as SysML), describing the product's functional architecture, logical structure, and the constraints between key parameters at the system level. Specifically, the Requirements Analysis Agent first performs intent recognition and entity extraction: it calls a large language model to perform intent recognition and entity extraction on the natural language requirements, combining this with automotive domain knowledge bases (NVH, durability, safety, regulations) to complete implicit constraints (such as collision regulations, weight limits, and cost targets). Then, it performs requirement-metric mapping: using a requirement-metric mapping template, qualitative requirements are transformed into quantifiable design metrics (range, Cd, torsional stiffness, temperature rise, etc.) and constraints, forming a metric-constraint table. Finally, SysML modeling: Based on the SysMLv2 modeling engine, the indicator-constraint table and functional decomposition results are automatically converted into SysML models and output after consistency verification.

[0222] SysML modeling includes: Requirement Diagram: Functional / non-functional requirements and their derivation and satisfaction relationships; Use Case Diagram: Interaction scenarios between the main actors and the system; Block Definition Diagram (BDD): System functional blocks, subsystem blocks, and their hierarchical decomposition; Internal Block Diagram (IBD): Inter-block ports, connectors, and data / energy flows; Parametric Diagram: Mathematical constraint equations for design metrics and constraints (e.g., range = f(battery energy, Cd, mass)); Interfaces and State Machines: Signal interface definitions and operational state transitions. Optionally, the requirements analysis agent simultaneously performs requirements-level risk identification: identifying requirement conflicts, unattainable metrics, regulatory gaps, interface inconsistencies, etc., and outputting risk register items to be incorporated into the traceable baseline.

[0223] The architecture design agent is an AI agent specifically responsible for system architecture design and physical model generation. The chief engineer agent dispatches the initial system model (SysML model) generated by the requirements analysis agent to the architecture design agent. This agent parses the system decomposition, functional definitions, interface descriptions, and parameter constraints in the model, then maps and converts them into a physical mathematical model that can be directly executed by the simulation solver. Specifically, the architecture design agent first performs physical field identification: it parses the SysML parameter diagram and block definition diagram to identify the relevant physical fields (fluid, structural, electromagnetic, thermal). Then, it performs governing equation matching: matching corresponding governing equations from the physical mechanism library of the physical base layer—continuity / Navier-Stokes equations (aerodynamics), structural mechanics equilibrium equations (durability / collision), heat conduction / energy equations (thermal management), and Maxwell equations (electromagnetism). Finally, it determines the boundary conditions and design variables: based on the SysML constraint equations, geometric parameters, and material parameters, it determines the boundary conditions, initial conditions, design variable set (independent variables), and response set (dependent variables). The following steps are required: establishing multi-physics coupling relationships (such as unidirectional electro-thermal coupling and bidirectional fluid-structure interaction coupling), and configuring discretization and solver parameters. Finally, the output is packaged: the above content is encapsulated into a physical mathematical model using PDL and then uniformly arranged by the chief engineer agent. The physical mathematical model includes the governing equations (partial differential equation form), boundary and initial conditions, design parameter sets (independent variables) and response sets (dependent variables), multi-physics coupling relationships (unidirectional / bidirectional coupling), discretization and solver configuration (mesh, time step, solver type), and physical consistency constraints (conservation laws, constitutive relations). The architecture design agent generates multiple physical mathematical models by subsystem / discipline—such as a vehicle aerodynamic model, body structure model, battery thermal model, and motor electromagnetic model. Each model corresponds to an independent simulation verification task, which is uniformly arranged and resource-allocated by the chief engineer agent.

[0224] By introducing PDL for requirement encapsulation and through the division of labor between requirement analysis agents and architecture design agents, a fully automated and traceable conversion chain is achieved, from unstructured natural language to structured system models, and then to executable physical and mathematical models. This process ensures the integrity and consistency of requirements during transmission, laying a solid foundation for subsequent simulation and optimization, and is a key link in realizing the full automation of the requirement-model-simulation process.

[0225] In some optional embodiments, the chief engineer's agent of the intelligent enhancement layer calls the requirements analysis agent to parse the target requirements package and generate an initial system model. This includes: calling the requirements analysis agent through the chief engineer's agent of the intelligent enhancement layer to perform intent recognition and entity extraction on the target requirements package to obtain requirements indicators and requirements constraints; calling the requirements analysis agent through the chief engineer's agent of the intelligent enhancement layer to obtain a requirements indicator mapping template, and generating quantitative design indicators based on the requirements indicators and requirements constraints using the requirements indicator mapping template; and generating an initial system model based on the quantitative design indicators. The initial system model includes at least one of the following: requirements diagram, use case diagram, block definition diagram, internal block diagram, parameter diagram, and state machine.

[0226] The intelligent enhancement layer of this application includes an AI intelligent agent matrix framework, combined with... Figure 6 As shown, Figure 6 This is a schematic diagram of an AI agent matrix architecture in one embodiment. Each agent accesses the simulation toolchain via the MCP protocol, performs task scheduling through a communication protocol with gRPC / FIR filters to compensate for latency, and achieves inter-agent collaboration and information sharing through the A2A (Agent-to-Agent) protocol, forming a hybrid architecture combining decentralized and centralized scheduling. This AI agent matrix architecture includes:

[0227] The chief engineer agent acts as a central coordinator, integrating multiple agent tasks and balancing conflicting objectives through the NSGA-III optimization algorithm to maintain compliance with standards such as ISO 26262. For example, in the automotive NVH optimization case, the chief engineer agent receives NVH error data from the SysML model of the demand parsing agent and the simulation verification agent, and uses the NSGA-III algorithm to find the optimal solution among the three objectives of noise, vibration, and comfort.

[0228] The professional intelligent agent matrix includes a requirement parsing agent, an architecture design agent, a simulation verification agent, and a code generation agent. Each agent interacts with simulation tools via the MCP protocol and uses a communication protocol with gRPC / FIR filters to compensate for latency for task scheduling. For example: The requirement parsing agent takes natural language requirements as input, outputs a SysML v2 model, and generates test cases. For instance, if the input is that the in-vehicle noise should be below 60dB when the vehicle is traveling at high speed, the requirement parsing agent can automatically generate the corresponding SysML requirement model and associate it with test cases. The architecture design agent generates AUTOSAR component architectures based on the optimization goals of the overall intelligent agent. For instance, based on the optimization goals of the overall intelligent agent (such as lowest cost and optimal performance), the agent can automatically generate the corresponding AUTOSAR component architectures. The simulation verification agent calls CarMaker / SUMO via the MCP protocol to perform SIL / HIL tests and feeds the results back to the overall intelligent agent. For instance, the agent can call CarMaker's vehicle dynamics model to perform simulations and output vibration and noise data for the overall intelligent agent to evaluate. The code generation agent generates C / C++ code based on the architecture design and interacts with the test agent via gRPC. For example, the agent can automatically generate C code that meets functional safety requirements based on the AUTOSAR component architecture and communicate with the test agent through the gRPC interface.

[0229] The agent employs a value-based reinforcement learning decision-making model, and the agent's decision-making Bellman optimal equation is as follows:

[0230]

[0231] in, For state The optimal value function; For immediate returns; Discount factor ( ); Let be the state transition probability. This equation enables the agent to autonomously explore the optimal path in a complex R&D space.

[0232] Agent matrix cooperative model:

[0233]

[0234] in, This represents the total output of the agent matrix. The number of agents; For the first Individual agent weights; Output it; The coefficient of coordination; This is a collaborative operator. The formula describes the aggregation and collaborative effects between the overall engineering agent matrix and the specialized agent matrix.

[0235] Specifically, the requirements analysis agent first performs deep semantic parsing on the target requirements package in PDL format. This process includes two core tasks: intent identification and entity extraction. Intent identification refers to determining the functional purpose of each requirement statement, such as whether it is a performance requirement, a functional requirement, or a security constraint. Entity extraction refers to extracting key objects, parameters, and values ​​from the requirement statements. For example, the entities extracted from the requirement of a minimum driving range of 800 kilometers include the object "driving range," the parameter "minimum," and the value "800 kilometers." Through these two tasks, the requirements analysis agent transforms the original vague requirements into a set of structured requirement indicators and requirements constraints.

[0236] The demand indicator mapping template is a pre-built knowledge base within the system, storing rules for transforming qualitative or semi-quantitative requirements for a specific domain into quantifiable engineering indicators. For example, in the automotive field, a mapping template might define good handling as a demand indicator, which should be translated into a roll gradient ≤ 4.5 deg / g and yaw rate gain ≥ 0.12s. -1 These are two quantifiable design metrics. The requirements analysis agent will call a mapping template that matches the product domain, and based on the requirements metrics and constraints extracted in the first step, it will generate a set of clear and measurable quantifiable design metrics through template matching and calculation.

[0237] Based on the generated quantitative design metrics, the requirements analysis agent invokes a system modeling engine to automatically create an initial system model. This model is a digital blueprint of the product system, and it includes at least one or more of the following views: a requirements diagram to show the hierarchical structure and traceability relationships of requirements; a use case diagram to describe the interaction scenarios between the system and external stakeholders; a block definition diagram to define the top-level structure and subsystem partitioning of the system; an internal block diagram to show the connection relationships between subsystems; a parameter diagram to define the mathematical constraints between key performance parameters; and a state machine diagram to describe the system's operational state transition logic.

[0238] This embodiment uses a standardized process of intent recognition, entity extraction, template mapping, and automated modeling to accurately and efficiently transform ambiguous natural language requirements into a structured system model. This automated modeling approach eliminates the manual translation and modeling steps in traditional requirements analysis, greatly reducing errors and omissions, and ensuring that all design decisions from the outset of the project are based on an accurate, traceable, and verifiable model.

[0239] In some optional embodiments, the method further includes: invoking a requirements analysis agent through the chief engineer agent of the intelligent enhancement layer to perform requirements layer risk identification, wherein requirements layer risk identification includes identifying at least one of requirements conflicts, unreachable indicators, regulatory gaps, and interface inconsistencies; outputting alarm information if the results of requirements layer risk identification indicate the existence of risks; and continuing to execute the step of invoking an architecture design agent through the chief engineer agent of the intelligent enhancement layer to process the initial system model and generate various physical and mathematical models if the results of requirements layer risk identification indicate the absence of risks.

[0240] While parsing requirements, generating metrics, and building models, the requirements analysis agent activates a built-in risk inspection engine. This engine comprehensively scans the current requirement set and generated metrics from multiple dimensions to identify potential problems. Specifically, the engine performs at least one or more of the following identification tasks:

[0241] Identifying conflicting requirements: Detecting whether there are logical or physical contradictions between different requirements. For example, one requirement stipulates that the vehicle weight should not exceed 1500kg, while another requirement stipulates that it should be equipped with a 100kWh battery pack. Based on the current battery energy density, these two requirements may not be able to be met simultaneously.

[0242] Identify unattainable targets: Assess whether the generated quantitative design targets are practically achievable under the constraints of current technology and physical laws. For example, requiring a drag coefficient of 0.18 for a mid-size SUV is extremely difficult to achieve with current engineering capabilities.

[0243] Identify regulatory gaps: Check whether the set of requirements is missing any mandatory laws, regulations or industry standards related to the product application area, such as functional safety (ISO 26262) or collision safety regulations for automobiles.

[0244] Identify interface inconsistencies: Check whether there are mismatches at the requirements and model levels in the defined interaction interfaces (such as mechanical size interfaces, electrical signal interfaces, and communication protocol interfaces) between different subsystems or components.

[0245] If the risk inspection engine detects any of the above-mentioned risks, the requirements analysis agent will immediately generate a structured alert message and output it through the chief engineer agent or directly to the user interface. This alert message will detail the risk type, risk description, involved requirement items or model elements, and potential impact.

[0246] Only if the requirements-level risk check passes completely (i.e., no of the aforementioned risks are found) will the chief engineer agent continue to schedule the architecture design agent to execute the subsequent physical and mathematical model generation steps. If a risk is found, the process will be paused, awaiting manual intervention and correction. This proactive quality assurance mechanism avoids discovering fundamental problems in the later stages of the project (such as the simulation or testing phase), thereby saving significant rework costs and time.

[0247] In some optional embodiments, the chief engineer agent of the intelligent enhancement layer calls the architecture design agent to process the initial system model and generate various physical mathematical models. This includes: calling the architecture design agent through the chief engineer agent to parse the block definition diagram and parameter diagram in the initial system model to determine the physical field type corresponding to the product requirement information; obtaining the control equations matching the physical field type from the physical base layer; determining the boundary conditions, initial conditions, and design parameters of each physical mathematical model based on the constraint equations, geometric parameters, and material parameters of each subsystem in the initial system model; establishing multi-physics coupling relationships based on the data interaction relationships between each subsystem; and generating various physical mathematical models based on the control equations, boundary conditions, initial conditions, design parameters, and multi-physics coupling relationships.

[0248] The architecture design agent first reads the initial system model generated by the requirements analysis agent, focusing on the Block Definition Diagram (BDD) and Parametric Diagram. The BDD describes the system's composition and hierarchical relationships, while the Parametric Diagram defines the mathematical constraints between key performance parameters. By analyzing the functional descriptions of each subsystem (block) in the BDD and the types of physical variables (such as force, velocity, temperature, and voltage) involved in the Parametric Diagram, the architecture design agent can determine the main physical field types involved in each subsystem. For example, a block describing the vehicle body structure, whose relevant parameters include stress and deformation, corresponds to a solid mechanics field; a block describing the cooling system, whose relevant parameters include temperature and heat flux, corresponds to a thermodynamic field.

[0249] After determining the type of physics field, the architecture design agent initiates a query to the physics base layer. The physics base layer contains a built-in physics mechanism library that stores the fundamental governing equations corresponding to various physics fields. Based on the type of physics field determined in the previous step (such as a fluid dynamics field), the agent retrieves and obtains the matching set of governing equations (such as the Navier-Stokes equations) from this library.

[0250] The architecture design agent extracts the geometric parameters (such as length, width, and thickness), material parameters (such as elastic modulus, Poisson's ratio, and density), and internal constraint equations of each subsystem from the parametric diagram and constraints of the initial system model. Based on this information, the agent determines for each physical mathematical model: specific boundary conditions (i.e., known quantities applied to the boundary of the solution domain, such as fixed constraints, applied loads, and given heat flux density), initial conditions (i.e., the field distribution at the start of the solution, such as the initial temperature field), and design parameters (such as plate thickness and cross-sectional dimensions) that can be adjusted for subsequent optimization.

[0251] Industrial products are often systems with coupled multi-physics fields. The architecture design agent identifies the physical field pairs that require data exchange or energy transfer based on the interfaces and connections between subsystem blocks in the initial system model (defined by the Internal Block Graph, IBD). For example, if there is heat transfer between a motor block and a cooling jacket block, a unidirectional electro-thermal coupling needs to be established between the electromagnetic field model and the thermodynamic model; if there is interaction between a body block and an external flow field block, a bidirectional fluid-structure interaction needs to be established between the structural mechanics model and the fluid dynamics model.

[0252] Finally, the architecture design agent integrates and encapsulates all the above information—the governing equations obtained from the physical base layer, the boundary and initial conditions derived from the model, the design parameters available for optimization, and the multiphysics coupling relationships—to generate one or more complete physical mathematical model files for each subsystem that can be directly called by simulation tools.

[0253] In some optional embodiments, after generating the target development scheme based on each physical mathematical model and target design parameters, the method further includes: reading the model information and architecture configuration in the target development scheme through the code generation agent of the intelligent enhancement layer, generating executable code; testing the executable code through the test verification agent of the intelligent enhancement layer, and obtaining a digital twin after the test is passed.

[0254] The code generation agent is an AI agent specifically responsible for transforming design models into executable software code. After the target development plan is generated, the chief engineer agent invokes this agent. The code generation agent first reads the model: it reads the SysML model (state machine / activity diagram) and AUTOSAR configuration (SWC interface definition) from the development plan. Then, it performs model-driven skeleton generation: based on Model-Driven Engineering (MDE), it transforms the SysML state machine / activity diagram into a control logic skeleton and the AUTOSAR SWC interface into a software component code framework. Third, it generates the LLM algorithm: the large language model generates control algorithm code and initial values ​​for calibration parameters based on the control strategy in the physical mathematical model. Finally, it performs consistency verification and output: after OCL consistency verification, it outputs compilable code (C / C++) and integrates it with the simulation environment via PDL / MCP.

[0255] The Test Verification Agent is an AI agent specifically designed to perform automated testing. It is activated after the code generation agent completes code generation. It retrieves the generated executable code and deploys it to the appropriate testing environment for systematic testing. The testing includes at least functional testing (verifying whether the code implements the intended design functionality) and performance testing (verifying whether the code meets non-functional requirements such as real-time performance and memory usage). Once all preset test cases have successfully passed, the Test Verification Agent encapsulates and saves the executable code that passed the tests, its corresponding physical and mathematical model, and related simulation verification data as a whole. This integrated entity, encompassing the model, code, and verification data, constitutes the digital twin of the physical product. This digital twin can serve as an executable digital asset for subsequent virtual verification, performance prediction, and continuous optimization.

[0256] Specifically, the test verification agent performs step-by-step testing: based on the generated code, it sequentially performs SIL (Software-in-the-Loop) testing, HIL (Hardware-in-the-Loop) testing, bench testing, and vehicle testing, progressively improving fidelity; HIL uses FIR filters to compensate for communication latency. Then, compliance and fault verification are performed: the chief engineer agent maintains ISO 26262 functional safety and ISO 21434 cybersecurity compliance; the fault simulation agent injects potential fault verification and fault tolerance capabilities. Finally, after passing the tests, the software version is frozen, enters mass production, and the final product is produced; the runtime data is fed back to the knowledge accumulation agent, entering the next round of the R&D cycle.

[0257] This embodiment extends the development process from simulation and optimization to code generation and automated testing, forming a fully automated closed loop from requirements to deployable software. By automatically generating code and performing systematic automated testing, it not only significantly reduces the time spent on manual code writing and debugging but also ensures strict consistency between the code and the validated model. The resulting digital twin provides a valuable digital foundation for subsequent product iterations and maintenance.

[0258] In some optional embodiments, the code generation agent of the intelligent enhancement layer reads the model information and architecture configuration in the target development scheme and generates executable code, including: reading the initial system model in the target development scheme through the code generation agent of the intelligent enhancement layer; generating a control logic skeleton and software component code framework based on the initial system model; generating control algorithm code and initial values ​​of calibration parameters based on the control strategy in the physical and mathematical model through the large language model; performing consistency verification on the generated code and outputting the executable code after the verification passes.

[0259] The code generation agent first extracts the initial system model (i.e., the SysML model) generated by the requirements analysis agent in the early stages from the target development plan. This model includes the system's static structure, dynamic behavior, and constraints.

[0260] Based on the read SysML model, the code generation agent uses Model-Driven Engineering (MDE) technology to automatically generate the main framework of the code. Specifically, it converts the state machine diagrams and activity diagrams in the model into corresponding control logic skeleton code; simultaneously, based on the software component types defined in the model (such as application layer software components conforming to the AUTOSAR standard) and port interfaces, it generates communication framework code between components. This generated code provides a structured skeleton for the entire software.

[0261] The code generation agent also integrates a Large Language Model (LLM). The LLM is a deep learning model trained on massive amounts of text data, possessing powerful natural language understanding and code generation capabilities. The agent inputs the control strategy described in the physical mathematical model (e.g., a PID control algorithm, a Kalman filter state update equation) into the LLM in the form of structured prompts. Based on these prompts, the LLM automatically generates the core algorithm code to implement the control strategy and generates a set of reasonable initial values ​​for relevant control parameters (such as PID gain).

[0262] Finally, the code generation agent performs consistency checks on all automatically generated code (including skeleton code and LLM-generated algorithm code). The checks include, but are not limited to: whether the code conforms to predefined coding standards, whether the interface definitions between modules match, whether the syntax is correct, and whether there are potential memory leaks or data race issues. After passing the checks, the agent compiles and links the code, ultimately outputting a complete executable file that can be deployed and run on the target hardware.

[0263] This embodiment organically combines model-driven engineering and large language model code generation technologies, which greatly reduces the tedious and error-prone manual coding work in software development and improves the automation level of code generation, the quality of generated code, and development efficiency.

[0264] In some optional embodiments, the executable code is tested through a test verification agent of the intelligent enhancement layer, including: performing software-in-the-loop testing, hardware-in-the-loop testing, and entity testing on the executable code in sequence; and saving the executable code after the tests pass.

[0265] The test verification agent systematically tests the executable code using a V-shaped verification strategy that progresses from virtual to real-world verification. This strategy specifically includes:

[0266] First, Software-in-the-Loop (SIL) testing is performed. In this test, the code, as a pure software module, is simulated in a closed loop within a host environment along with a virtual controlled object model (i.e., a physical-mathematical model). This stage primarily verifies the code's logical functionality, mathematical algorithms, and basic numerical stability.

[0267] After passing the SIL test, a Hardware-in-the-Loop (HIL) test is performed. In this test, code is burned onto the target hardware (such as an Electronic Control Unit (ECU)) and connected to a high-fidelity real-time emulator. This emulator simulates a complete physical environment, including sensor signals and actuator responses. The HIL test verifies the code's real-time performance, interface compatibility, and electromagnetic interference immunity on real hardware.

[0268] After passing the HIL test, entity testing is conducted. In this test, the code is finally deployed to an actual physical product prototype and tested in a real or near-real working environment to ultimately verify the product's performance and reliability under real-world usage conditions.

[0269] Only after all three levels of testing (SIL, HIL, and entity testing) have passed will the test verification agent mark the executable code of that version as verified and officially save it to the code repository as a release or production-ready version.

[0270] In some optional embodiments, the method further includes: scheduling each intelligent agent through a chief engineer agent; the intelligent agents include at least one of a target agent, a simulation verification agent, a code generation agent, and a test verification agent; wherein the chief engineer agent communicates with each intelligent agent through the gRPC protocol, the intelligent agents communicate with each other through the A2A protocol, the simulation verification agent calls the simulation tool of the physical base layer through the MCP protocol, and the latency in gRPC communication and / or MCP communication is compensated through an FIR filter.

[0271] The Chief Engineer Agent, acting as the central scheduler in the system, is responsible for managing and coordinating the work of all specialized intelligent agents. Scheduling methods include, but are not limited to: task decomposition, task delegation, result aggregation, and exception handling. The Chief Engineer Agent dynamically activates or suspends the corresponding specialized agents based on the needs of the current development phase.

[0272] gRPC protocol: A high-performance, open-source, and general-purpose Remote Procedure Call (RPC) framework. The chief engineer agent and various specialized agents use gRPC for bidirectional communication, transmitting control commands, status reports, and structured data. Based on HTTP / 2 and Protocol Buffers, gRPC provides efficient and reliable binary transport.

[0273] The A2A (Agent-to-Agent) protocol is a protocol for direct communication between specialized intelligent agents. It allows an agent to request information or delegate subtasks directly to another agent without going through a chief agent. For example, a code generation agent can directly request the latest optimization parameters from a simulation verification agent via the A2A protocol.

[0274] The MCP (Model Context Protocol) is a dedicated protocol for simulation verification agents to interact with various simulation tools in the physical base layer. The MCP protocol defines a standard set of interfaces for the agent to discover tool capabilities, initiate computation tasks, monitor progress, and obtain results, thereby shielding the proprietary interface differences between different simulation tools.

[0275] In distributed systems, network transmission latency is unavoidable, especially in scenarios with high real-time requirements such as HIL testing. Therefore, the system integrates an FIR (Finite Impulse Response) filter module into all communication links based on gRPC and MCP. This module monitors the round-trip time of network data packets in real time and utilizes the predictive compensation capability of the FIR filter to correct the phase and delay of the received signal.

[0276] Specifically, the gRPC / FIR delay compensation formula for proxy communication is as follows. For distributed proxy communication delay, an FIR filter is used for compensation: The FIR filter delay compensation model is as follows:

[0277]

[0278] in, For input signals; The output is after compensation; These are the coefficients of the FIR filter; The filter order is used. This model compensates for constant delay in gRPC communication. This ensures the real-time performance and consistency of the Hybrid Real-Virtual Prototype (HIL).

[0279] In some optional embodiments, the method further includes: calculating V-model development efficiency via a chief engineer agent, whereby V-model development efficiency is used to quantify the closed-loop quality from product requirement information to test and verification.

[0280]

[0281] in, This represents the R&D efficiency value. A higher value indicates a more efficient and higher-quality R&D process. N represents the total number of R&D stages, specifically the number of clearly defined stages included in the entire V-model process, from the input of product requirements information to the completion of final testing and verification. These stages may include, for example, requirements analysis, system design, detailed design, unit testing, integration testing, and system testing. For the first This function validates the requirements and specifications for each stage. It measures the extent to which the actual outputs of that stage (such as requirements specifications, design models, and code) meet its input requirements (i.e., the output of the previous stage or the target specifications of this stage). The function can return a logical value (0 for validation failure, 1 for validation success) or a continuous value between 0 and 1, representing the percentage of validation success. For the first The weighting coefficients for each stage are used to reflect the importance of different stages to the final quality and success of the entire project. For example, the weighting coefficient for the requirements analysis stage can be set higher than that for the code writing stage. The total number of development stages N is the number of stages included in the entire V-model process from product requirements information to testing and verification.

[0282] In some optional embodiments, the method further includes: acquiring source domain experience data and extracting transferable domain-invariant features from the source domain experience data; and performing transfer learning on samples in the target domain based on the domain-invariant features to generate a target domain model corresponding to the target domain.

[0283] Combination Figure 7 As shown, Figure 7 This is a knowledge transfer mechanism from a source domain (automotive) to a target domain (photovoltaics / robotics / aerospace) in one embodiment.

[0284] The source domain refers to a specific industrial sector (such as the automotive industry) where rich design experience, simulation data, and optimization cases have been accumulated through this method or traditional methods. Experience data refers to these accumulated digital assets that can guide future designs, including but not limited to: historical simulation result datasets, validated physical and mathematical models, successful Pareto front solutions for multi-objective optimization, and mature design principles and engineering experience. The system extracts transferable domain-invariant features from this source domain experience data using feature extraction algorithms. Domain-invariant features refer to physical laws, design patterns, or optimization strategies that have similarities or commonalities across different industrial sectors, such as vortex shedding characteristics of fluid-borne objects, stress distribution patterns of structures under load, or trade-off methods in multi-objective optimization. Combined with... Figure 7 Domain-specific knowledge bases: These include knowledge in specialized fields such as automotive NVH (noise, vibration, and harshness), durability, and safety; photovoltaic conversion efficiency and thermal management; robot dynamics and control; and aerospace aerodynamics and structural strength. For example, the automotive NVH knowledge base contains a correlation model between vehicle vibration and noise, while the photovoltaic thermal management knowledge base contains a model showing the relationship between photovoltaic panel temperature and conversion efficiency.

[0285] The target domain refers to a new industrial field (such as photovoltaics, robotics, or aerospace) where this method needs to be applied but where data accumulation is relatively limited. The system acquires a small but representative set of data samples from the target domain (such as a preliminary design scheme and simulation results). Then, using the domain-invariant features extracted in the first step as prior knowledge, and combining these limited target domain samples, the model is trained or fine-tuned using transfer learning algorithms (such as parameter-based transfer, feature-based transfer, or instance-based transfer). Finally, a new model suitable for the target domain is generated, namely the target domain model. This model inherits knowledge from the source domain, thus exhibiting good prediction and optimization performance even when target domain data is scarce.

[0286] In this embodiment, the adaptive interface utilizes Domain Adaptive Transfer Learning (DATL) technology to achieve knowledge transfer between different industrial domains. It minimizes the feature distribution difference between the source and target domains using MMD (Maximum Mean Difference), with a weight parameter λ ranging from 0.5 to 2.0. For example, when transferring automotive NVH knowledge to robot dynamics simulation, the feature distributions of the two domains can be aligned using the MMD loss function.

[0287] Furthermore, a unified Physical Description Language (PDL) and MCP protocol are defined to support data exchange and model integration between different CAD / CAE systems. For example, the PDL definition can be used to describe finite element models of automobile bodies and thermal management models of photovoltaic panels, while the MCP protocol can be used to proxy communication with simulation tools such as CarSim and SU2.

[0288] This embodiment extracts and reuses common knowledge across different fields, enabling the system and method of this invention to be rapidly and cost-effectively extended to entirely new industrial sectors. This transfer learning capability significantly lowers the barrier to entry for intelligent R&D in new fields, reducing the time and resource investment required to accumulate data and train models from scratch.

[0289] In some optional embodiments, based on domain-invariant features, transfer learning is performed on samples from the target domain to generate a target domain model corresponding to the target domain. This includes: using empirical data from the source domain as source domain samples and obtaining a small amount of data from the target domain as target domain samples; minimizing the feature distribution difference between the source domain samples and the target domain samples in the regenerating kernel Hilbert space by using the maximum mean difference; constructing a transfer loss function based on the task loss of the target domain and the maximum mean difference; and training the task model of the target domain with the goal of minimizing the transfer loss to obtain the transferred target domain model.

[0290] First, the system extracts high-quality historical simulation data or design scheme data from the experience database of the source domain, serving as source domain samples X_s. Simultaneously, it obtains a small number of data points from the target domain that have undergone physical verification or high-fidelity simulation, serving as target domain samples X_t. Here, "small number" refers to the size of the source domain samples, typically only 5% to 10%.

[0291] The system uses Maximum Mean Discrepancy (MMD) as a metric to measure the difference in feature distribution between source neighborhood samples X_s and target neighborhood samples X_t in the Reproducing Kernel Hilbert Space (RKHS). In this embodiment, the MMD is used to minimize the feature distribution difference between the source and target neighborhoods. The Maximum Mean Discrepancy (MMD) is:

[0292]

[0293] in, Map the kernel function to the regenerating kernel Hilbert space (RKHS); , The distribution of the source and target domains; , The number of samples; This is the RKHS norm. The smaller the MMD, the closer the distributions of the two domains are.

[0294] The smaller the MMD value, the closer the distributions of the two domains in the feature space. One of the goals of transfer learning is to use a feature transformation matrix to make the distribution of the transformed source domain features as aligned as possible with the distribution of the target domain features, i.e., to minimize this MMD value.

[0295] To ensure both the model's prediction accuracy in the target domain and effective utilization of source domain knowledge during the transfer process, the system constructs a composite transfer loss function, L_DATL. This function is obtained by a weighted sum of two parts; specifically, the DATL transfer loss is:

[0296]

[0297] Formula explanation: Losses due to mission failures in the target domain; For the transfer weights, take This loss function aligns the feature distributions of the two domains while minimizing the task error, enabling cross-domain transfer of physical models and engineering experience.

[0298] Finally, the system uses the transfer loss function L_DATL as the training objective to train a task model in the target domain (e.g., a PINN model for predicting photovoltaic panel temperature). During training, the optimizer simultaneously minimizes the task loss and the distribution difference loss, enabling the model to learn not only from the limited labeled data in the target domain but also to transfer common knowledge learned from the source domain (such as the basic physical laws of heat conduction). The resulting model after training is the transferred target domain model.

[0299] In the above embodiments, by minimizing the distribution difference between the source and target domains in the RKHS space, this method can effectively transfer knowledge from the source domain to the target domain, even if the target domain has only a small number of samples. This enables the rapid establishment of high-performance prediction and optimization models in emerging or data-scarce industrial sectors, significantly accelerating the diffusion and application of the technology across different industries.

[0300] In some optional embodiments, the method further includes: sending the target design parameters in the target development scheme to the physical product through a digital twin mapping interface, and collecting the operating data of the physical product and sending it back to the corresponding digital twin; correcting the digital twin based on the operating data to generate a corrected digital twin.

[0301] After the target development plan is generated, the system encapsulates the target design parameters (e.g., optimized suspension stiffness values, motor control parameters, or battery thermal management strategies) determined in the plan into a digital twin mapping interface (PDL), and then sends them to the controller of the physical product (such as the vehicle ECU) or the manufacturing execution system (MES) through a secure network connection. This process realizes the digital instruction mapping from virtual design to physical entity.

[0302] During the operation of physical products (such as vehicle operation, photovoltaic power plant generation, and robot operation), sensors distributed in key parts of the product (such as vibration sensors, temperature sensors, current sensors, and strain gauges) continuously collect real-time operational data. This data is transmitted back to the cloud or local server via an edge gateway or vehicle communication module, and then written into a digital twin corresponding to the physical product through a digital twin mapping interface. The digital twin is a complete digital image of the physical product in virtual space; it includes not only the product's design model but also integrates its operational data throughout its entire lifecycle.

[0303] The returned operational data is used to compare and analyze the prediction output of the digital twin. If a discrepancy is found between the digital twin's predictions and the physical entity's measured values, the system will use this new measured data as a calibration basis to adjust or retrain the model parameters within the digital twin (such as the neural network weights in PINN) online, thereby correcting the model and making its predictions more closely resemble the actual performance of the physical entity. This corrected digital twin has improved fidelity and can more accurately reflect the current state and performance of the physical product.

[0304] In some optional embodiments, the digital twin is corrected based on the running data to generate a corrected digital twin, including: comparing the running data with the predicted value of the digital twin and calculating the residual; if the residual exceeds a preset threshold, the digital twin is corrected, and based on the model and running data corresponding to the corrected digital twin, the steps of multi-objective optimization based on multiple objectives corresponding to various simulation results and product requirement information are continued through the chief engineer agent of the intelligent enhancement layer.

[0305] The digital twin mapping interface, deployed at the application service layer, serves as the "last mile" bridge connecting the NSGA-III optimization results (sixth step output) of the intelligent enhancement layer with the physical product, and is also a unified channel for bidirectional interaction between the digital and physical worlds. This interface reuses the multiphysics equations of the physical base layer and the PINN model of the intelligent enhancement layer downwards, and provides standardized microservice calls to the application service layer upwards, following the hierarchical division of the three-layer architecture.

[0306] The interface consists of four core components:

[0307] PDL data encapsulation layer: It encapsulates the design parameters and model status of the virtual side, as well as the sensor readings of the physical side, into a unified PDL format, shielding the protocol differences of different industrial products (such as automotive CAN / Ethernet, photovoltaic Modbus, and robot ROS).

[0308] MCP Protocol Channel: Serves as a communication bus between the agent and the physical product / edge gateway, reusing the MCP protocol mechanism to achieve unified access across vendor devices.

[0309] Model consistency verification module: compares the predicted values ​​of the digital twin with the measured values ​​of the physical twin, quantifies model drift, and determines whether to trigger a model correction or re-optimization cycle.

[0310] Delay compensation module: Based on FIR filter, it compensates for remote communication delay and ensures the causal consistency of real-time mapping.

[0311] The system continuously collects physical product operation data (measured values) and compares them one by one with the simulation predictions of the digital twin under the same operating conditions, calculating the difference between the two, i.e., residual e = |y_predicted - y_measured|. This residual reflects the accuracy of the current digital twin model and the degree of deviation from the actual physical world.

[0312] Furthermore, the digital twin mapping interface supports bidirectional data synchronization, forming a closed-loop interaction between the virtual and physical worlds.

[0313] (1) Virtual to physical downlink synchronization: The optimal design parameter set obtained by the overall engineering intelligence agent through NSGA-III optimization is encapsulated in PDL and sent to the controller or manufacturing execution system of the physical product to realize the direct mapping of the design optimization results to the physical product.

[0314] (2) Physical to Virtual Uplink Synchronization: Collect sensor data (vehicle vibration / temperature / stress, photovoltaic power generation / module temperature, aircraft load, robot joint torque, etc.) during the operation of physical products and feed it back to the digital twin as a real benchmark for model correction and continuous optimization.

[0315] Specifically, after generating the optimal R&D plan, the interface maps the optimal parameters to the physical prototype; after the test is passed and the product is mass-produced, the interface continuously feeds back the running data to the knowledge accumulation agent.

[0316] Triggering modes can include: periodic synchronization: milliseconds to seconds, used for online monitoring and real-time control feedback; event synchronization: triggered by abnormal events (such as vibration exceeding the threshold or sudden temperature rise), used for fault early warning; change synchronization: triggered by design iteration, corresponding to the parameter distribution after the convergence of multi-objective optimization in step 6.

[0317] The system presets an acceptable residual threshold for key performance indicators (such as maximum stress, temperature, and vibration amplitude). If the calculated residual continuously or intermittently exceeds this threshold, it indicates that the predictive ability of the digital twin has undergone significant model drift. In this case, the system automatically triggers a model correction process: using newly collected operational data as new training samples, incrementally learning or fine-tuning the internal model of the digital twin (such as PINN), thereby generating a corrected digital twin.

[0318] More importantly, this embodiment also includes a mechanism to trigger re-optimization. After the model is corrected, the system will take the corrected digital twin model and its latest operating data, and then continue to execute the steps of multi-objective optimization based on the simulation results and product requirement information through the chief engineer agent of the intelligent enhancement layer. This means that, based on the real feedback from the physical world, the system will initiate a new round of virtual simulation and multi-objective optimization to seek a design solution with better performance and more robustness under real-world conditions.

[0319] The core value of a digital twin lies in maintaining high-fidelity consistency with the physical product. This interface uses PINN online learning to achieve model drift compensation.

[0320] (1) Residual calculation: Based on the measured data of the physical product, calculate the deviation between the predicted value and the measured value of the digital twin. This deviation is used as the new collocation input for the data loss L_data.

[0321] (2) Online correction: Utilizing PINN’s small-sample online learning capability, the measured data is used as the new collocation point x_i and substituted into the physical constraint loss. Under the premise of preserving the physical law constraint, the neural network parameter θ is adaptively corrected to compensate for the drift of the model accumulated over time.

[0322] (3) Re-optimization trigger: When the model residual exceeds the preset threshold, the interface automatically sends back the corrected model and the latest running data, triggering NSGA-III re-optimization, forming a continuous improvement closed loop of running-correction-re-optimization.

[0323] In some optional embodiments, the method further includes: obtaining change information on product demand information through a chief engineer agent; determining the affected downstream items based on the change information; and triggering the corresponding smart agent to perform collaborative updates based on the affected downstream items.

[0324] During development, product requirements may change due to market feedback, regulatory updates, or technological shifts. Change information refers to any additions, deletions, or modifications made to the original requirement items. The chief engineer's agent will acquire this change information in real time through its subscribed requirement change event bus (such as an event listener integrated with PLM or requirement management tools). Change information includes at least the changed item, the type of change (addition, modification, deletion), and the new content after the change.

[0325] Once the affected downstream items are identified, the chief engineer agent automatically initiates a coordinated update process. Based on the type of affected item, it triggers the corresponding specialized intelligent agents to perform collaborative updates. For example, if the change affects the system architecture, the architecture design agent updates the SysML model and regenerates the physical mathematical model. If the change affects design parameters or constraints, the simulation verification agent and optimization agent restart or partially re-optimize based on existing optimization results. If the change affects control logic, the code generation agent regenerates the affected code. Finally, the test verification agent performs regression testing on the updated parts to ensure that the change has not introduced new defects.

[0326] The entire change propagation and update process was automated under the overall coordination of the chief engineer, minimizing manual coordination and rework.

[0327] Specifically, demand change coordination is one of the core capabilities of the virtual physical AI chief engineer in replacing human coordination. Its complete mechanism consists of a closed loop of five links, which is deeply coupled with the nine-step process:

[0328] (1) Change awareness: The chief engineer's agent subscribes to the requirement document version management system (such as Polarion, DOORS) through the MCP protocol. Based on the large language model, it performs differential recognition on the semantic changes of requirement items, not only capturing literal modifications, but also identifying the associated changes of implicit indicators. For example, when the range changes from 800km to 850km, the agent can automatically identify the associated implicit indicators such as battery energy, vehicle weight, and Cd constraints.

[0329] (2) Change impact analysis: Based on the traceability matrix automatically constructed by the satisfy / verify / derive relationship in the SysML v2 model, the chief engineer agent performs reachability analysis along the five-layer traceability chain of requirements-design indicators-physical and mathematical models-code-test cases, identifies all affected downstream items, and outputs a list of impact scope and risk assessment.

[0330] (3) Change propagation and collaborative update: Based on the list of affected areas, the chief engineer intelligent agent automatically triggers collaborative updates of the architecture design agent (updating the physical and mathematical model), the simulation verification agent (re-executing the fifth step of simulation), and the code generation agent (regenerating the affected code) according to the nine-step process. The entire process is completed through the A2A protocol to complete the task delegation between agents.

[0331] (4) Regression verification: After the change propagation is completed, the chief engineer agent automatically schedules the ninth step of test verification (SIL-HIL-bench-real vehicle) to ensure that the change does not introduce new defects; if the regression test fails, the rollback mechanism is triggered to restore to the previous stable baseline.

[0332] (5) Baseline update and version management: After all verifications are passed, the updated project can be traced back to the baseline, retaining a complete change history (the person making the change, the time of the change, the affected items, and the verification results), and supporting version rollback and auditing at any time.

[0333] The above five steps constitute a closed loop of demand change linkage, compressing the manual coordination process that would take days or even weeks in the traditional V model to hours or even minutes.

[0334] 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.

[0335] Based on the same inventive concept, this application also provides a large-model-based system development apparatus for implementing the large-model-based system development method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations in one or more embodiments of the large-model-based system development apparatus provided below can be found in the limitations of the large-model-based system development method described above, and will not be repeated here.

[0336] In one exemplary embodiment, such as Figure 8 As shown, a system development device based on a large model is provided, including: a requirement information acquisition module 801, a physical-mathematical model generation module 802, a simulation module 803, and a multi-objective optimization module 804, wherein:

[0337] The demand information acquisition module 801 is used to receive product demand information input in natural language through the application service layer.

[0338] The physical mathematical model generation module 802 is used to obtain various physical mathematical models based on product requirement information through the target agent of the intelligent enhancement layer. The physical mathematical models are used to describe the design parameters and constraints of each subsystem.

[0339] Simulation module 803 is used to call various simulation tools in the physical base layer through the simulation verification agent of the intelligent enhancement layer, perform simulation calculations on various physical mathematical models in parallel, and generate simulation results.

[0340] The multi-objective optimization module 804 is used to perform multi-objective optimization based on various simulation results and product requirement information through the chief engineer agent of the intelligent enhancement layer, obtain optimized design parameters, and feed back the optimized design parameters to the simulation verification agent, triggering the simulation verification agent to update the design parameters in each physical mathematical model and re-perform simulation calculations to generate updated simulation results. This process continues until the multi-objective optimization meets the convergence condition, and the target design parameters of each physical mathematical model are obtained. Based on each physical mathematical model and the target design parameters, a target development plan is generated.

[0341] In some optional embodiments, the multi-objective optimization module 804 is specifically used to generate reference target values ​​for multiple objectives corresponding to product demand information based on the physics indicators in each simulation result, and to determine each non-compliant objective based on the multiple objectives corresponding to product demand information and the reference target values; to calculate the penalty value of each non-compliant objective using the penalty function method, and to determine the optimization direction based on the penalty value; and to adjust the design parameters in multiple directions using a multi-objective optimization algorithm based on the optimization direction to generate optimized design parameters.

[0342] In some optional embodiments, the multi-objective optimization module 804 is specifically used to generate new reference target values ​​corresponding to multiple objectives of product demand information based on the physics discipline indicators in the updated simulation results; when all the new reference target values ​​meet the multiple objectives corresponding to product demand information, or when the change in the physics discipline indicators in adjacent updated simulation results is less than a preset convergence threshold, the optimized design parameters determined this time are used as the target design parameters of each physical mathematical model.

[0343] In some optional embodiments, the multi-objective optimization module 804 is specifically used to perform optimization layer risk identification through the chief engineer agent during the multi-objective optimization iteration process. The optimization layer risk identification includes at least one of parameter sensitivity risk identification, constraint out-of-bounds risk identification, and robustness risk identification; and generate risk registration items based on the results of the optimization layer risk identification.

[0344] In some optional embodiments, the above apparatus further includes: a resource allocation module, configured to generate corresponding simulation tasks based on each physical mathematical model through the simulation verification agent of the intelligent enhancement layer, and classify each simulation task; divide the computing resources of the physical base layer into multiple resource levels; match each classified simulation task with each resource level after hierarchical division, and solve the optimization algorithm with at least two of the total task completion delay, computing cost and simulation quality as optimization objectives to generate a resource allocation scheme corresponding to each simulation task; and determine the computing resources corresponding to each simulation task based on the resource allocation scheme.

[0345] In some optional embodiments, the resource hierarchy includes GPU clusters, CPU clusters, and edge computing nodes. Specifically, the resource allocation module is used to match fluid simulation and neural network training tasks to GPU clusters, structural simulation and electromagnetic simulation tasks to CPU clusters, and software-in-the-loop testing and hardware-in-the-loop testing tasks to edge computing nodes based on the subject type and fidelity of each simulation task, generating an initial resource allocation scheme. The fidelity of the simulation task is the credibility of the physical mathematical model. Using at least two of the following as optimization objectives—total task completion delay, computing cost, and simulation quality—the initial resource allocation scheme is solved through an optimization algorithm, outputting the resource allocation scheme corresponding to each simulation task.

[0346] In some optional embodiments, the simulation module 803 is specifically used to determine the simulation requirements corresponding to each physical mathematical model; determine the simulation tool type based on the simulation requirements; call the simulation tool corresponding to the simulation tool type in the physical base layer through the simulation verification agent of the intelligent enhancement layer, and perform simulation calculations on the physical mathematical model through the called simulation tool to generate simulation results.

[0347] In some optional embodiments, the simulation module 803 is specifically used to: when the simulation tool type is a physics solver, call the physics solver of the physics base layer through the simulation verification agent of the intelligent enhancement layer, and perform simulation calculations on the physics mathematical model through the scheduled physics solver to generate simulation results; when the simulation tool type is a neural network, call the neural network of the physics base layer through the simulation verification agent of the intelligent enhancement layer, and perform simulation calculations on the physics mathematical model through the neural network to generate simulation results; when the simulation tool type is a hybrid type, call the neural network of the physics base layer through the simulation verification agent of the intelligent enhancement layer, and perform simulation calculations on the physics mathematical model through the neural network to generate candidate simulation results, and call the physics solver of the physics base layer through the simulation verification agent of the intelligent enhancement layer, and perform simulation calculations on the physics mathematical model corresponding to the candidate simulation results through the scheduled physics solver to generate simulation results.

[0348] In some optional embodiments, the above-mentioned requirement information acquisition module 801 is specifically used to encapsulate product requirement information through the application service layer to obtain a target requirement package described by a preset physical description language.

[0349] The physical and mathematical model generation module 802 is specifically used to call the requirements analysis agent through the chief engineer agent of the intelligent enhancement layer to parse the target requirements package and generate an initial system model; and to call the architecture design agent through the chief engineer agent of the intelligent enhancement layer to process the initial system model and generate various physical and mathematical models.

[0350] In some optional embodiments, the aforementioned requirement information acquisition module 801 is specifically used to call the requirement analysis agent through the chief engineer agent of the intelligent enhancement layer to perform intent recognition and entity extraction on the target requirement package to obtain requirement indicators and requirement constraints; to call the requirement analysis agent through the chief engineer agent of the intelligent enhancement layer to obtain the requirement indicator mapping template, and to generate quantitative design indicators based on the requirement indicators and requirement constraints using the requirement indicator mapping template; and to generate an initial system model based on the quantitative design indicators, wherein the initial system model includes at least one of requirement diagram, use case diagram, block definition diagram, internal block diagram, parameter diagram, and state machine.

[0351] In some optional embodiments, the aforementioned requirement information acquisition module 801 is specifically used to call the requirement analysis agent through the chief engineer agent of the intelligent enhancement layer to perform requirement layer risk identification. Requirement layer risk identification includes identifying at least one of requirement conflicts, unreachable indicators, regulatory gaps, and interface inconsistencies. If the result of requirement layer risk identification indicates the existence of a risk, an alarm message is output. If the result of requirement layer risk identification indicates the absence of a risk, the process continues to execute the step of calling the architecture design agent through the chief engineer agent of the intelligent enhancement layer to process the initial system model and generate various physical and mathematical models.

[0352] In some optional embodiments, the aforementioned physical mathematical model generation module 802 is specifically used to call the architecture design agent through the chief engineer agent, parse the block definition diagram and parameter diagram in the initial system model, and determine the physical field type corresponding to the product requirement information; obtain the control equations matching the physical field type from the physical base layer; determine the boundary conditions, initial conditions and design parameters of each physical mathematical model based on the constraint equations, geometric parameters and material parameters of each subsystem in the initial system model; establish multi-physics coupling relationships based on the data interaction relationships between each subsystem; and generate each physical mathematical model based on the control equations, boundary conditions, initial conditions, design parameters and multi-physics coupling relationships.

[0353] In some optional embodiments, the above apparatus further includes: a code generation module, configured to read model information and architecture configuration in the target development scheme through the code generation agent of the intelligent enhancement layer, generate executable code; test the executable code through the test verification agent of the intelligent enhancement layer, and obtain a digital twin after the test is passed.

[0354] In some optional embodiments, the code generation module is specifically used to read the initial system model in the target development scheme through the code generation agent of the intelligent enhancement layer; generate the control logic skeleton and software component code framework based on the initial system model; generate control algorithm code and initial values ​​of calibration parameters based on the control strategy in the physical and mathematical model through the large language model; perform consistency verification on the generated code, and output the executable code after the verification is passed.

[0355] In some optional embodiments, the code generation module is specifically used to perform software-in-the-loop testing, hardware-in-the-loop testing, and physical testing sequentially based on the executable code; after the tests are passed, the executable code is saved.

[0356] In some optional embodiments, the above apparatus further includes: a scheduling module, specifically configured to schedule each intelligent agent through a chief engineer agent; the intelligent agents include at least one of a target agent, a simulation verification agent, a code generation agent, and a test verification agent; wherein the chief engineer agent communicates with each intelligent agent through the gRPC protocol, the intelligent agents communicate with each other through the A2A protocol, the simulation verification agent calls the simulation tool of the physical base layer through the MCP protocol, and the latency in gRPC communication and / or MCP communication is compensated through an FIR filter.

[0357] In some alternative embodiments, the above-described apparatus is further configured to calculate V-model development efficiency via a chief engineer agent, wherein V-model development efficiency is used to quantify the closed-loop quality from product requirement information to testing and verification.

[0358]

[0359] in, The value represents the R&D efficiency, where N is the total number of R&D stages. For the first Verification functions for each stage of requirements and specifications. For the first The weight coefficients for each stage, and the total number of R&D stages N is the number of stages included in the entire V-model process from product requirement information to testing and verification.

[0360] In some optional embodiments, the above apparatus further includes: a transfer module, configured to acquire source domain experience data and extract transferable domain-invariant features from the source domain experience data; and based on the domain-invariant features, perform transfer learning on samples in the target domain to generate a target domain model corresponding to the target domain.

[0361] In some optional embodiments, the above-mentioned transfer module is specifically used to take the empirical data of the source domain as the source domain sample and obtain a small amount of data of the target domain as the target domain sample; minimize the feature distribution difference between the source domain sample and the target domain sample in the regenerating kernel Hilbert space by the maximum mean difference; construct a transfer loss function based on the task loss of the target domain and the maximum mean difference; train the task model of the target domain with the goal of minimizing the transfer loss to obtain the transferred target domain model.

[0362] In some optional embodiments, the above apparatus further includes: a mapping module, used to send the target design parameters in the target development scheme to the physical product through a digital twin mapping interface, and collect the operation data of the physical product and send it back to the corresponding digital twin; and to correct the digital twin based on the operation data to generate a corrected digital twin.

[0363] In some optional embodiments, the mapping module is specifically used to compare the running data with the predicted value of the digital twin and calculate the residual; if the residual exceeds a preset threshold, the digital twin is corrected, and based on the model and running data corresponding to the corrected digital twin, the process of multi-objective optimization based on multiple objectives corresponding to various simulation results and product demand information is continued through the chief engineer agent of the intelligent enhancement layer.

[0364] In some optional embodiments, the above apparatus further includes: an update module, configured to obtain change information of product demand information through the chief engineer agent; determine the affected downstream items based on the change information; and trigger the corresponding smart agent to perform collaborative updates based on the affected downstream items.

[0365] Each module in the aforementioned system development device based on a large model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0366] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, 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 an 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 any of the above embodiments. 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. When the computer program is executed by the processor, it implements a system development method based on a large model.

[0367] Those skilled in the art will understand that Figure 9 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.

[0368] 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 method embodiments.

[0369] 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 method embodiments.

[0370] 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 method embodiments.

[0371] 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.

[0372] 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.

[0373] 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.

[0374] 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 system development method based on a large model, characterized in that, The method includes: The application service layer receives product requirement information input in natural language. The target agent of the intelligent enhancement layer obtains various physical and mathematical models based on the product requirement information. The physical and mathematical models are used to describe the design parameters and constraints of each subsystem. The simulation verification agent of the intelligent enhancement layer calls the simulation tools of the physical base layer to perform simulation calculations on each physical mathematical model in parallel and generate simulation results. Through the chief engineer agent of the intelligent enhancement layer, multi-objective optimization is performed based on the simulation results and the product requirement information to obtain optimized design parameters. The optimized design parameters are then fed back to the simulation verification agent, which triggers the simulation verification agent to update the design parameters in each physical mathematical model and re-perform simulation calculations to generate updated simulation results. This process continues until the multi-objective optimization meets the convergence condition, at which point the target design parameters of each physical mathematical model are obtained. Based on each physical mathematical model and the target design parameters, a target development plan is generated. The chief engineer agent, through the intelligent enhancement layer, performs multi-objective optimization based on the simulation results and product requirement information to obtain optimized design parameters, including: Based on the physics discipline indicators in each simulation result, reference target values ​​are generated for multiple objectives corresponding to the product demand information, and each non-compliant objective is determined based on the multiple objectives corresponding to the product demand information and the reference target values. Based on each of the non-compliance targets, the penalty value for each non-compliance target is calculated using the penalty function method, and the optimization direction is determined based on the penalty value; Based on the optimization direction, the design parameters are adjusted in multiple directions using a multi-objective optimization algorithm to generate optimized design parameters.

2. The method according to claim 1, characterized in that, After the multi-objective optimization satisfies the convergence condition, the target design parameters of each physical mathematical model are obtained, including: Based on the physics discipline indicators in the updated simulation results, new reference target values ​​are generated for the multiple objectives corresponding to the product demand information. If the new reference target values ​​all meet the multiple targets corresponding to the product requirement information, or if the change in the physics subject index in the adjacent updated simulation results is less than the preset convergence threshold, the optimized design parameters determined in this study will be used as the target design parameters for each physics mathematical model.

3. The method according to claim 1, characterized in that, The method further includes: During the multi-objective optimization iteration process, the chief engineer agent performs optimization layer risk identification, which includes at least one of parameter sensitivity risk identification, constraint out-of-bounds risk identification, and robustness risk identification. Risk registration items are generated based on the results of the risk identification at the optimization layer.

4. The method according to any one of claims 1 to 3, characterized in that, Before the simulation verification agent of the intelligent enhancement layer calls the various simulation tools of the physical base layer to perform simulation calculations on each of the physical mathematical models in parallel and generate simulation results, the process also includes: The simulation verification agent of the intelligent enhancement layer generates corresponding simulation tasks based on each of the physical and mathematical models, and classifies each of the simulation tasks. The computing resources of the physical base layer are divided into multiple resource levels; The simulation tasks after classification are matched with the resource levels after hierarchical division, and the optimization algorithm is used to solve the problem with at least two of the total task completion delay, computing cost and simulation quality as optimization objectives to generate the resource allocation scheme corresponding to each simulation task. The computing resources corresponding to each simulation task are determined based on the resource allocation scheme.

5. The method according to claim 4, characterized in that, The resource hierarchy includes GPU clusters, CPU clusters, and edge computing nodes; the process of matching each classified simulation task with each hierarchically divided resource hierarchy, and using an optimization algorithm to solve for at least two of the following as optimization objectives: total task completion latency, computing cost, and simulation quality, generates a resource allocation scheme corresponding to each simulation task, including: Based on the subject type and fidelity of each simulation task, fluid simulation and neural network training tasks are matched to the GPU cluster, structural simulation and electromagnetic simulation tasks are matched to the CPU cluster, and software-in-the-loop testing and hardware-in-the-loop testing tasks are matched to the edge computing node to generate an initial resource allocation scheme. The fidelity of the simulation task is the credibility of the physical mathematical model. Using at least two of the following as optimization objectives—total task completion delay, computing cost, and simulation quality—an optimization algorithm is used to solve the initial resource allocation scheme and output the resource allocation scheme corresponding to each simulation task.

6. The method according to any one of claims 1 to 3, characterized in that, The simulation verification agent of the intelligent enhancement layer calls various simulation tools of the physical base layer to perform simulation calculations on each of the physical mathematical models in parallel and generate simulation results, including: Determine the simulation requirements corresponding to each of the physical and mathematical models; The type of simulation tool will be determined based on the aforementioned simulation requirements; The simulation verification agent of the intelligent enhancement layer calls the simulation tool corresponding to the simulation tool type in the physical base layer, and performs simulation calculations on the physical mathematical model through the called simulation tool to generate simulation results.

7. The method according to claim 6, characterized in that, The process of calling a simulation tool corresponding to the simulation tool type in the physical base layer through the simulation verification agent of the intelligent enhancement layer, and performing simulation calculations on the physical mathematical model using the called simulation tool to generate simulation results includes: When the simulation tool type is a physics solver, the simulation verification agent of the intelligent enhancement layer calls the physics solver of the physical base layer, and performs simulation calculations on the physical mathematical model through the scheduled physics solver to generate simulation results; When the simulation tool type is a neural network, the simulation verification agent of the intelligent enhancement layer calls the neural network of the physical base layer, and performs simulation calculations on the physical mathematical model through the neural network to generate simulation results. When the simulation tool type is a hybrid type, the simulation verification agent of the intelligent enhancement layer calls the neural network of the physical base layer, and performs simulation calculations on the physical mathematical model through the neural network to generate candidate simulation results. Then, the simulation verification agent of the intelligent enhancement layer calls the physical solver of the physical base layer, and performs simulation calculations on the physical mathematical model corresponding to the candidate simulation results through the scheduled physical solver to generate simulation results.

8. The method according to any one of claims 1 to 3, characterized in that, After receiving product demand information input in natural language through the application service layer, the process further includes: The product requirement information is encapsulated by the application service layer to obtain a target requirement package described by a preset physical description language; The target agent through the intelligent enhancement layer obtains various physical and mathematical models based on the product demand information, including: The chief engineer agent of the intelligent enhancement layer calls the requirements analysis agent to parse the target requirements package and generate an initial system model. The chief engineer agent of the intelligent enhancement layer calls the architecture design agent to process the initial system model and generate various physical and mathematical models.

9. The method according to claim 8, characterized in that, The process of invoking the requirements analysis agent through the chief engineer's agent in the intelligent enhancement layer to parse the target requirements package and generate an initial system model includes: The chief engineer agent of the intelligent enhancement layer calls the demand analysis agent to perform intent recognition and entity extraction on the target demand package, and obtain demand indicators and demand constraints. The chief engineer agent of the intelligent enhancement layer calls the requirements analysis agent to obtain the requirements indicator mapping template, and generates quantitative design indicators based on the requirements indicators and requirements constraints using the requirements indicator mapping template. Based on the quantitative design metrics, an initial system model is generated, which includes at least one of the following: a requirement diagram, a use case diagram, a block definition diagram, an internal block diagram, a parameter diagram, and a state machine.

10. The method according to claim 9, characterized in that, The method further includes: The chief engineer agent of the intelligent enhancement layer calls the demand analysis agent to perform demand layer risk identification. The demand layer risk identification includes identifying at least one of demand conflicts, unachievable indicators, regulatory gaps, and interface inconsistencies. If the risk identification result at the demand layer indicates the existence of a risk, an alarm message will be output. If the risk identification result at the demand layer indicates that there is no risk, the process continues by calling the architecture design agent through the chief engineer agent of the intelligent enhancement layer to process the initial system model and generate various physical and mathematical models.

11. The method according to claim 8, characterized in that, The process of calling the architecture design agent through the chief engineer's agent in the intelligent enhancement layer to process the initial system model and generate various physical and mathematical models includes: The chief engineer agent calls the architecture design agent to parse the block definition diagram and parameter diagram in the initial system model and determine the physical field type corresponding to the product requirement information. Obtain the governing equations that match the physical field type from the physical base layer; Based on the constraint equations, geometric parameters, and material parameters of each subsystem in the initial system model, the boundary conditions, initial conditions, and design parameters of each physical and mathematical model are determined. Based on the data interaction relationships between the subsystems, establish multiphysics coupling relationships; Based on the governing equations, boundary conditions, initial conditions, design parameters, and multiphysics coupling relationships, various physical mathematical models are generated.

12. The method according to any one of claims 1 to 3, characterized in that, After generating the target development scheme based on each of the aforementioned physical mathematical models and the target design parameters, the method further includes: The code generation agent of the intelligent enhancement layer reads the model information and architecture configuration in the target development scheme and generates executable code. The executable code is tested by the test verification agent of the intelligent enhancement layer, and a digital twin is obtained after the test is passed.

13. The method according to claim 12, characterized in that, The code generation agent of the intelligent enhancement layer reads the model information and architecture configuration in the target development scheme and generates executable code, including: The code generation agent of the intelligent enhancement layer reads the initial system model in the target development scheme; Based on the initial system model, a control logic skeleton and software component code framework are generated. The control algorithm code and initial values ​​of calibration parameters are generated by using a large language model based on the control strategy in the physical mathematical model. Perform a consistency check on the generated code and output the executable code after the check passes.

14. The method according to claim 12, characterized in that, The test verification proxy through the intelligent enhancement layer tests the executable code, including: Based on the executable code, software-in-the-loop testing, hardware-in-the-loop testing, and physical testing are performed sequentially. After the test passes, save the executable code.

15. The method according to any one of claims 1 to 3, characterized in that, The method further includes: The chief engineer agent schedules each intelligent agent; the intelligent agent includes at least one of the target agent, the simulation verification agent, the code generation agent, and the test verification agent. The chief engineer agent communicates with each of the intelligent agents via the gRPC protocol, the intelligent agents communicate with each other via the A2A protocol, the simulation verification agent calls the simulation tool of the physical base layer via the MCP protocol, and compensates for the latency in gRPC communication and / or MCP communication via an FIR filter.

16. The method according to any one of claims 1 to 3, characterized in that, The method further includes: The V-model R&D efficiency is calculated through the chief engineer's agent, and the V-model R&D efficiency is used to quantify the closed-loop quality from product requirement information to testing and verification. in, The value represents the R&D efficiency, where N is the total number of R&D stages. For the first Verification functions for each stage of requirements and specifications. For the first The weight coefficients for each stage, where the total number of R&D stages is the number of stages included in the entire V-model process from product requirement information to testing and verification.

17. The method according to any one of claims 1 to 3, characterized in that, The method further includes: Acquire source domain experience data and extract transferable domain-invariant features from the source domain experience data; Based on the domain-invariant features, transfer learning is performed on samples in the target domain to generate a target domain model corresponding to the target domain.

18. The method according to claim 17, characterized in that, Based on the domain-invariant features, Transfer learning is performed on samples from the target domain to generate a target domain model corresponding to the target domain, including: Use the empirical data from the source domain as the source domain sample, and obtain a small amount of data from the target domain as the target domain sample. Minimize the difference in feature distribution between the source neighborhood samples and the target neighborhood samples in the regenerating kernel Hilbert space by maximizing the mean difference; Based on the task loss in the target domain and the maximum mean difference, a migration loss function is constructed; With the goal of minimizing the transfer loss, the task model of the target domain is trained to obtain the transferred target domain model.

19. The method according to any one of claims 1 to 3, characterized in that, The method further includes: The target design parameters in the target development scheme are sent to the physical product through the digital twin mapping interface, and the operation data of the physical product is collected and sent back to the corresponding digital twin. The digital twin is corrected based on the operational data to generate a corrected digital twin.

20. The method according to claim 19, characterized in that, The step of correcting the digital twin based on the operational data to generate a corrected digital twin includes: The running data is compared with the predicted values ​​of the digital twin, and the residuals are calculated. If the residual exceeds a preset threshold, the digital twin is corrected, and based on the model corresponding to the corrected digital twin and the running data, the process of multi-objective optimization is continued through the chief engineer agent of the intelligent enhancement layer, based on the simulation results and the product demand information corresponding to multiple objectives.

21. The method according to any one of claims 1 to 3, characterized in that, The method further includes: The chief engineer's agent obtains information on changes to product demand; The affected downstream entries are determined based on the change information; Based on the affected downstream entries, the corresponding smart agent is triggered to perform a collaborative update.

22. A system development apparatus based on a large model, characterized in that, The device includes: The requirement information acquisition module is used to receive product requirement information input in natural language through the application service layer; The physical mathematical model generation module is used to obtain various physical mathematical models based on the product requirement information through the target agent of the intelligent enhancement layer. The physical mathematical models are used to describe the design parameters and constraints of each subsystem. The simulation module is used to call various simulation tools in the physical base layer through the simulation verification agent of the intelligent enhancement layer, perform simulation calculations on each of the physical mathematical models in parallel, and generate simulation results. The multi-objective optimization module is used to perform multi-objective optimization based on the simulation results and product requirement information through the chief engineer agent of the intelligent enhancement layer, obtain optimized design parameters, and feed the optimized design parameters back to the simulation verification agent, triggering the simulation verification agent to update the design parameters in each physical mathematical model and re-perform simulation calculations to generate updated simulation results. This process continues until the multi-objective optimization meets the convergence condition, and the target design parameters of each physical mathematical model are obtained. Based on each physical mathematical model and the target design parameters, a target development plan is generated. The multi-objective optimization module is further configured to generate reference target values ​​for the multiple objectives corresponding to the product demand information based on the physics indicators in each simulation result, and determine each non-compliant objective based on the multiple objectives corresponding to the product demand information and the reference target values; calculate the penalty value for each non-compliant objective using the penalty function method, and determine the optimization direction based on the penalty value; and adjust the design parameters in multiple directions using a multi-objective optimization algorithm based on the optimization direction to generate optimized design parameters.

23. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 21.

24. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 21.

25. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 21.

Images