An aircraft CFD simulation automation method and system based on a large language model and multi-agent collaboration

Abstract

The application discloses an aircraft CFD simulation automation method and system based on a large language model and multi-agent cooperation, which adopts a double-layer architecture of an arrangement agent and a CFD professional sub-agent cluster. The arrangement agent is driven by a large language model, communicates with the sub-agent cluster through a JSON-RPC 2.0 standardized protocol bridge layer, provides tool self-discovery, parameter self-inspection and a structured tool calling interface, is responsible for task decomposition, workflow planning, cross-session persistent memory and automatic convergence monitoring and multi-level quality decision-making. The sub-agent cluster includes a mesh generation sub-agent, a CFD solving sub-agent, a convergence monitoring sub-agent, a post-processing analysis sub-agent and an aerodynamic optimization sub-agent. The application realizes end-to-end automation of an aircraft CFD simulation whole process, shortens a simulation cycle, improves a solving success rate, and realizes knowledge cross-session accumulation and unattended operation.

Description

Technical Field

[0001] This application belongs to the field of automated aerodynamic simulation technology for aircraft, specifically relating to an automated CFD simulation method and system for aircraft based on a large language model and multi-agent collaboration. Background Technology

[0002] In aircraft aerodynamic design and analysis, CFD simulation is a complex process with multiple stages and a long cycle. A typical process includes geometric parameterization, computational mesh generation, flow field solution, convergence assessment, post-processing and visualization, result analysis, and optimization decision-making. In traditional CFD simulation processes, each stage is usually manually performed by different engineers using different software tools. Taking airfoil aerodynamic analysis as an example, engineers first parameterize the airfoil shape in OpenVSP, then generate the computational mesh in Pointwise or SnappyHexMesh, then configure the solution parameters and submit the calculation in OpenFOAM, perform post-processing in ParaView, and finally assess convergence based on the residual curves and force coefficient curves. The processes rely on file input / output and manual judgment, and a single complete iteration typically takes several hours.

[0003] Traditional CFD simulation workflows suffer from several shortcomings. The manual, sequential execution of multiple steps leads to inefficiency. Mesh generation tools, CFD solvers, and post-processing tools operate independently, lacking a unified, automated orchestration interface. In multidisciplinary design optimization scenarios, after each design variable update, the entire loop of mesh generation, CFD solving, post-processing, and gradient calculation must be re-executed, resulting in optimization cycles lasting days to weeks. Convergence judgment relies on engineers observing residual and force coefficient curves, leading to subjective and inconsistent criteria. For complex conditions such as transonic flows, residuals may exhibit periodic oscillations rather than monotonically decreasing, making it difficult for novice engineers to accurately determine convergence and when to terminate the calculation. Aerodynamic shape optimization requires iterative searches within the design space, with each iteration requiring a complete loop of shape update, mesh regeneration, CFD solving, objective function evaluation, and gradient or surrogate model update. The manual breakpoints at each stage of the existing toolchain make fully automated optimization loop closure extremely difficult.

[0004] In recent years, large language models have made significant progress in code generation and task planning, but their application in the field of full-process automation of CFD simulation is still in its infancy. Some studies have attempted to use large language models to generate OpenFOAM configuration files or solver scripts. However, the output text of large language models still requires manual review, modification, and execution, and cannot directly drive the solver engine for reliable computation. The large language model only acts as a text generator and does not have the ability to directly manipulate the tool. Existing simulation tools based on large language models adopt a single-agent architecture, where a single large language model instance interacts directly with the simulation tool. The context window of a single agent is limited, making it difficult to handle multiple sub-tasks such as mesh generation, solver configuration, convergence monitoring, and result analysis simultaneously. Single agents lack task decomposition and parallel execution capabilities, and cannot maintain long-term consistency in multi-step processes. In existing solutions, the interaction between large language models and simulation tools is through generating text commands or non-standardized function calls, lacking structured parameter passing, type checking, and error propagation mechanisms. When adding new simulation tools or simulation types, the prompts or code of the large language model must be manually modified, and it cannot automatically discover and adapt to new tools. Existing solutions start from scratch in each session and cannot remember user preferences, environment configurations, and past simulation experience, requiring users to repeatedly provide the same information in each interaction. CFD solving is a lengthy process, and existing solutions lack the ability to monitor the solution process midway. Judgments such as whether the solution has converged, diverged, or needs to be adjusted to continue iteration still require manual intervention, making unattended automation impossible.

[0005] Therefore, existing technologies have not yet solved the problem of using a large language model as the top-level orchestration agent and scheduling multiple professional CFD sub-agents to work together through standardized protocols to achieve full automation of the aircraft CFD simulation process. This results in low efficiency of aircraft CFD simulation, many manual intervention steps, inability to accumulate cross-session knowledge, and a lack of automatic convergence monitoring and quality decision-making in the solution process. Summary of the Invention

[0006] This application aims to provide an automated method and system for aircraft CFD simulation based on a large language model and multi-agent collaboration. By orchestrating the hierarchical collaboration of intelligent agents and specialized sub-agent clusters, combined with standardized protocol bridging, cross-session persistent memory, and automatic convergence monitoring and quality decision-making, the end-to-end automation of the entire aircraft CFD simulation process can be achieved.

[0007] To achieve the above technical objectives, this application specifically employs the following technical solution: One aspect of this application provides an automated method for CFD simulation of aircraft based on a large language model and multi-agent cooperation, comprising the following steps: S1. Construct a hierarchical architecture including an orchestration agent and a cluster of CFD professional sub-agents. The orchestration agent is driven by a large language model, and the cluster of CFD professional sub-agents includes multiple professional CFD simulation sub-agents. S2. The orchestration agent receives the aircraft CFD simulation requirements input by the user in natural language, extracts physical parameters from the natural language text using a large language model, and decomposes the aircraft CFD simulation requirements into an ordered sequence of simulation sub-tasks. S3. The orchestration agent communicates with the CFD professional sub-agent cluster through a standardized protocol bridging layer. The standardized protocol bridging layer adopts the JSON-RPC 2.0 protocol and provides a tool self-discovery interface, a parameter self-introspection interface, and a structured tool call interface. S4. The orchestration agent queries available professional CFD simulation sub-agents and input parameter patterns through the tool self-discovery interface, verifies the input parameters corresponding to the simulation sub-task sequence through the parameter introspection interface, and issues simulation sub-tasks to the corresponding professional CFD simulation sub-agents through the structured tool call interface. S5. The professional CFD simulation sub-agent executes the simulation sub-task and returns the structured execution result to the orchestration agent; S6. During the execution of the CFD solution task by the professional CFD simulation sub-agent, the orchestration agent continuously receives the intermediate status information returned by the professional CFD simulation sub-agent through the standardized protocol bridging layer, evaluates the intermediate status information in real time according to the preset multi-level thresholds, and automatically makes quality decisions such as continuing iteration, extending the number of iteration steps, adjusting parameters and resubmitting, or terminating the solution. S7. The orchestration agent integrates the structured execution results and outputs the CFD simulation results of the aircraft.

[0008] In one implementation, the tool discovery interface is the tools / list interface, the parameter introspection interface is the get_skill_input_info interface, and the structured tool invocation interface is the tools / call interface. The orchestration agent queries the currently available professional CFD simulation sub-agents and their corresponding input parameter modes through the tools / list interface. When adding or replacing a professional CFD simulation sub-agent, the orchestration agent automatically discovers the added or replaced professional CFD simulation sub-agent through the tools / list interface.

[0009] In one implementation, the orchestration agent queries the professional CFD simulation sub-agent for the parameter names, data types, default values, and value ranges required by the get_skill_input_info interface; and performs at least one operation based on the query results: missing parameter inference, out-of-range parameter prompting, and contradictory parameter consistency check.

[0010] In one implementation, the missing parameter is inferred as: when the user does not provide the target dimensionless wall distance y + At that time, the target y is determined based on the turbulence model and near-wall treatment method used. + The range is determined, and the height of the first layer of mesh is estimated by combining the incoming flow conditions, fluid property parameters, characteristic length, and empirical relationships of the turbulent boundary layer of a flat plate. When the user does not provide a turbulence model, the default turbulence model is determined according to the simulation conditions and preset model selection rules.

[0011] In one implementation, the orchestration agent plans a workflow topology based on the type and complexity of the aircraft CFD simulation task. The workflow topology includes linear serial workflows, parallel execution workflows, iterative workflows, or hybrid workflows. In a linear serial workflow, the output of the previous simulation subtask serves as the input of the next simulation subtask. In a parallel execution workflow, the orchestration agent simultaneously schedules multiple specialized CFD simulation sub-agents of the same type to execute in parallel. In a iterative workflow, the orchestration agent performs a loop of shape parameterization, mesh generation, CFD solving, objective function and gradient evaluation, and shape updating until the objective function converges or the maximum number of iterations is reached.

[0012] In one implementation, the orchestration agent maintains a persistent memory across sessions, which stores user preference information, environment configuration information, and simulation experience information; at the start of each new session, the orchestration agent automatically loads persistent memory entries related to the current aircraft CFD simulation task.

[0013] In one implementation, user preference information includes the commonly used Mach number range, preferred turbulence model, default angle of attack range, and preferred output format; environment configuration information includes the number of available CPU cores, the installed solver version, the working directory path, and the list of currently available professional CFD simulation sub-agents; simulation experience information includes the quality performance of past mesh parameter combinations, the convergence characteristics of specific airfoils under specific working conditions, and known solution divergence boundary conditions.

[0014] In one implementation, the preset multi-level thresholds include a convergence residual threshold, a force coefficient fluctuation threshold, a residual oscillation rate of change threshold, a residual oscillation amplitude threshold, a residual growth threshold, a force coefficient divergence threshold, and a maximum iteration step threshold; the quality decision is performed in the following order: When all residual components fall below the convergence residual threshold, and the fluctuation range of the lift coefficient and drag coefficient within the recent iteration window is less than the force coefficient fluctuation threshold, the solution is determined to pass, the solution is terminated, and post-processing begins. When each residual component fails to fall below the convergent residual threshold, and the absolute value of the rate of change of each residual component is less than the residual oscillation rate of change threshold and the peak-to-valley difference of each residual component is less than the residual oscillation amplitude threshold within a consecutive preset number of iterations, a warning is issued, the maximum number of iterations threshold is increased, and observation continues. When each residual component continues to grow within a preset number of iterations and the growth slope is greater than the residual growth threshold, or when the fluctuation range of the lift coefficient and drag coefficient within the recent iteration window is greater than the force coefficient divergence threshold, it is judged as a failure, the current solution is terminated, the cause of divergence is analyzed, the parameters are adjusted and the solution task is resubmitted. When the current iteration step count reaches the maximum iteration step count threshold and each residual component has not decreased below the convergence residual threshold, it is determined to be a timeout. The current residual component levels, lift coefficient, and drag coefficient calculation results are summarized and the user's decision request is output.

[0015] In one implementation, the CFD specialized sub-agent cluster is organized according to a two-level Agent+Skill structure, where Agent represents a specialized domain and Skill represents a specific operation within that domain; the CFD specialized sub-agent cluster includes at least two of the following: a mesh generation sub-agent, a CFD solving sub-agent, a convergence monitoring sub-agent, a post-processing analysis sub-agent, and an aerodynamic optimization sub-agent.

[0016] In one implementation, the backend of the professional CFD simulation sub-agent employs a general solution framework based on the finite volume method, a gradient optimization framework based on the adjoint method, commercial CFD software, or open-source CFD software; the standardized protocol bridging layer is transparent to the internal implementation of the professional CFD simulation sub-agent.

[0017] In one implementation, the orchestration agent supports at least one interactive platform among WeChat, Lark, command-line interface, and web interface; the CFD professional sub-agent cluster is deployed on a local workstation or cloud server; the aircraft CFD simulation results include at least one of the following: calculation results, residual curves, force coefficients, flow field cloud diagrams, and analysis reports.

[0018] In another aspect of this application, an automated CFD simulation system for aircraft based on a large language model and multi-agent cooperation is provided, comprising: The orchestration agent, driven by a large language model, is used to receive the aircraft CFD simulation requirements input by the user in natural language, extract physical parameters from the natural language text, decompose the aircraft CFD simulation requirements into an ordered sequence of simulation sub-tasks, and schedule professional CFD simulation sub-agents through a standardized protocol bridging layer. The CFD professional sub-agent cluster includes multiple professional CFD simulation sub-agents, which are used to receive simulation sub-tasks issued by orchestration agents through a standardized protocol bridging layer, execute simulation links in the aircraft CFD simulation process, and return structured execution results. A standardized protocol bridging layer, using the JSON-RPC 2.0 protocol, is used to provide tool self-discovery interfaces, parameter introspection interfaces, and structured tool call interfaces to enable communication between orchestration agents and CFD professional sub-agent clusters; The orchestration agent is also used to continuously receive mid-course state information returned by professional CFD simulation sub-agents through a standardized protocol bridging layer during the CFD solution task, automatically make convergence monitoring and quality decisions based on preset multi-level thresholds, and integrate structured execution results to output the aircraft CFD simulation results.

[0019] The beneficial effects of this application are as follows: (1) Achieve end-to-end automation of the entire CFD simulation process for aircraft, significantly shortening the simulation cycle. In the traditional CFD simulation process for aircraft, multiple engineers need to manually execute the steps of geometric parameterization, mesh generation, solution configuration, flow field solution, convergence judgment, and post-processing analysis in different tools. A single complete iteration usually takes several hours, and aerodynamic optimization takes several days to several weeks. This application shortens the entire process of a single steady-state CFD simulation of an airfoil (including mesh generation, CFD solution, and result analysis) to several minutes by orchestrating intelligent agents to automatically decompose simulation sub-tasks and scheduling a cluster of CFD-specialized sub-intelligent agents to execute collaboratively, and no manual intervention is required throughout the process. In the mesh independence research scenario, the parallel generation and solution of multiple sets of meshes makes the total time only about 1.3 times that of a single set; in the aerodynamic optimization scenario, the optimization cycle can run automatically overnight, compressing the optimization cycle from days to hours.

[0020] (2) Overcoming the context and capability bottlenecks of single agents through a multi-agent hierarchical collaborative architecture. Existing simulation tools based on large language models all adopt a single agent architecture. The context window and attention resources of a single instance are limited, making it difficult to handle multiple professional simulation sub-tasks such as mesh generation, solution monitoring, and result analysis simultaneously. This application adopts a two-layer hierarchical architecture of orchestration agent and CFD professional sub-agent cluster. The orchestration agent focuses on high-level task planning and scheduling, while the professional CFD simulation sub-agents encapsulate the complete capabilities of specific simulation links. The orchestration agent and the professional CFD simulation sub-agents work collaboratively through a standardized protocol bridging layer. The above architecture decomposes the entire process of complex aircraft CFD simulation into multiple independently executable simulation sub-tasks, effectively breaking through the capability limit of the single agent architecture, while maintaining a clear hierarchical structure and high scalability.

[0021] (3) Establish a standardized protocol bridge to achieve tool self-discovery independent of the solver. This application uses the JSON-RPC2.0 standardized protocol to bridge the orchestration agent and the CFD professional sub-agent cluster, providing three core interfaces: tool self-discovery interface (tools / list), parameter introspection interface (get_skill_input_info), and structured tool call interface (tools / call). Adding or replacing a professional CFD simulation sub-agent only requires exposing the tool interface that conforms to the protocol specification, without modifying the orchestration agent code or prompts. The backend of the professional CFD simulation sub-agent can adopt any CFD solving framework. The standardized protocol bridge layer is completely transparent to the internal implementation of the professional CFD simulation sub-agent, ensuring that this application is applicable to various CFD toolchains and that the scope of protection is not limited by specific software platforms.

[0022] (4) Lowering the parameter setting threshold and improving the success rate of the solution through the parameter introspection mechanism. Before each call to the structured tool, the orchestration agent automatically queries the input parameter pattern information through the parameter introspection interface, and performs missing parameter inference, out-of-range parameter prompts, and consistency checks of contradictory parameters accordingly. The above mechanism reduces the probability of solution divergence caused by improper parameter settings from the source, enabling researchers with CFD theoretical foundation but lacking practical experience to reliably obtain high-quality aircraft CFD simulation results.

[0023] (5) Simulation knowledge is accumulated and reused across sessions through a cross-session persistent memory mechanism. The orchestration agent maintains the cross-session persistent memory library, automatically saving user preference information, environment configuration information, and simulation experience information. As the number of uses increases, the system automatically learns the user's work habits, convergence characteristics under specific working conditions, and reliable parameter combinations, making subsequent aircraft CFD simulation tasks more efficient and accurate, and eliminating the burden of users repeatedly providing the same information in each interaction.

[0024] (6) Unattended operation of CFD solution is achieved through automated convergence monitoring and multi-level quality decision-making. During the CFD solution process, the orchestration agent continuously receives mid-process status information of residuals and force coefficients, and automatically makes judgments based on four threshold levels: pass / warning / fail / timeout, such as continuing the solution, extending the number of iterations, adjusting parameters and resubmitting, or outputting user decision requests. This allows the entire CFD solution process to automatically complete convergence determination and anomaly handling without manual monitoring of the residual curve. Attached Figure Description

[0025] Figure 1 This is a system framework diagram of the automated CFD simulation method for aircraft used in this application; Figure 2 This is the computational mesh for the NACA0012 airfoil in the embodiment of this application; Figure 3 This is the convergence history curve of the steady-state CFD residuals in embodiment NACA0012 of this application; Figure 4 This is the convergence history of lift / drag as the iteration step changes in the embodiments of this application; where a is the convergence history of the lift coefficient and b is the convergence history of the drag coefficient. Detailed Implementation

[0026] The technical solution of this application will be clearly and completely described below with reference to specific embodiments. However, those skilled in the art will understand that the embodiments described below are only some embodiments of this application, not all embodiments, and are only used to illustrate this application, and should not be regarded as limiting the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] To address the problems in existing CFD simulation workflows, such as low efficiency due to multiple manual sequential steps, fragmented CFD toolchains lacking unified orchestration, lack of standardized protocol bridging layers and multi-agent collaboration mechanisms between large language models and specialized CFD simulation tools, and lack of cross-session persistent memory and automated convergence monitoring capabilities, the purpose of this application is: (1) A method for automating CFD simulation of aircraft based on large language model and multi-agent collaboration is provided. The method adopts a hierarchical architecture of orchestration agent - CFD professional sub-agent cluster, which automatically decomposes the user's natural language simulation requirements into an ordered sequence of simulation sub-tasks and schedules professional CFD simulation sub-agents to perform collaboratively, thereby realizing end-to-end automation of the entire process from geometric parameterization, mesh generation, CFD solution to result analysis, and eliminating manual breakpoints between each link. (2) Establish a standardized protocol bridge layer between the orchestration agent and the professional CFD simulation sub-agent. The standardized protocol bridge layer adopts the JSON-RPC 2.0 protocol and provides a tool self-discovery interface (tools / list), a parameter self-introduction interface (get_skill_input_info), and a structured tool call interface (tools / call). This makes the addition and deletion of professional CFD simulation sub-agents transparent to the orchestration agent. Any CFD tool and large language model agent that supports the standardized protocol bridge layer can be flexibly connected, realizing the complete decoupling of the standardized protocol bridge layer from the internal implementation of the professional CFD simulation sub-agent. (3) Construct a cross-session persistent memory library for orchestration agents, automatically save user preference information, environment configuration information and simulation experience information, so that users can understand the working environment and historical experience without having to provide the same information repeatedly in subsequent sessions, and realize the continuous accumulation and cross-session reuse of simulation knowledge; (4) Establish an automated convergence monitoring and multi-level quality decision mechanism for the CFD solution process. During the CFD solution process, continuously receive residual and force coefficient mid-way status information, and automatically make quality decisions based on preset multi-level thresholds, such as continuing iteration, extending the number of iteration steps, adjusting parameters and resubmitting, or outputting user decision requests, so as to realize unattended automatic operation of the entire CFD solution process.

[0028] This application proposes an automated CFD simulation method for aircraft based on a large language model and multi-agent collaboration. The method employs a hierarchical architecture of orchestration agents and CFD-specific sub-agent clusters. The upper-layer orchestration agent, driven by a large language model, is responsible for natural language understanding, task decomposition, workflow planning, and result integration. The lower-layer CFD-specific sub-agent cluster consists of multiple specialized CFD simulation sub-agents, which receive instructions from the upper layer and independently execute specific simulation operations through a standardized protocol bridging layer. The standardized protocol bridging layer adopts the JSON-RPC 2.0 protocol and provides tool discovery, parameter description, and structured tool invocation interfaces compatible with the MCP tool calling specification. Communication between the two layers is achieved through the standardized protocol bridging layer, realizing complete decoupling between the orchestration layer and the execution layer.

[0029] like Figure 1As shown, the overall architecture of the natural language-driven CFD intelligent simulation system unifies the scheduling of a cluster of specialized CFD sub-agents through an orchestration agent, achieving full automation from natural language requirements to simulation result output. The natural language input layer receives unstructured simulation requirements described by the user in natural language. The orchestration agent, as the core scheduling hub, sequentially completes task understanding, task decomposition, scheduling orchestration, and result integration. Task understanding parses the semantics and intent of the natural language input; task decomposition breaks down the overall simulation task into executable simulation sub-tasks; scheduling orchestration coordinates the execution order and dependencies of various specialized CFD simulation sub-agents; and result integration summarizes the output results of each specialized CFD simulation sub-agent. The orchestration agent achieves context state and historical information access through bidirectional data interaction across a session persistent memory, supporting long-range context management. The standardized protocol bridging layer adopts the JSON-RPC 2.0 protocol and an MCP-compatible protocol to achieve standardized communication and tool calls between the orchestration agent and the cluster of specialized CFD sub-agents. The CFD professional sub-agent cluster consists of a mesh generation sub-agent, a CFD solving sub-agent, a convergence monitoring sub-agent, a post-processing analysis sub-agent, and an aerodynamic optimization sub-agent. Each professional CFD simulation sub-agent is uniformly scheduled by the orchestrating agents through a standardized protocol bridging layer. The results output layer returns calculation results, residual curves, force coefficients, flow field contour maps, and comprehensive analysis reports to the user.

[0030] First layer: Arrangement of intelligent agents The orchestration agent layer, driven by a large language model, serves as the core scheduling hub of the automated CFD simulation system for aircraft. The orchestration agent can be driven by one or more large language models possessing natural language understanding, task planning, and structured tool invocation capabilities. As the sole interaction point between the user and the simulation system, the orchestration agent integrates four functions: natural language understanding and task decomposition, workflow planning, cross-session persistent memory, automated convergence monitoring, and multi-level quality decision-making.

[0031] (1) Natural Language Understanding and Task Decomposition Users input their CFD simulation requirements for aircraft into the orchestration agent using natural language. The orchestration agent leverages the semantic understanding capabilities of a large language model to extract key physical parameters from unstructured natural language text and decomposes the complex simulation requirements into an ordered sequence of simulation subtasks.

[0032] For example, the input could be "Perform steady-state CFD simulation of airfoil NACA0012, Mach number 0.3, Reynolds number 5 million, angle of attack 3 degrees, 6000 mesh elements". The programming agent would then analyze and obtain the airfoil designation NACA0012, Mach number 0.3, and Reynolds number 5 × 10⁻⁶. 6The target parameters include an angle of attack of 3° and a grid cell count of 6000. The orchestration agent decomposes the simulation requirements into three simulation sub-tasks: Sub-task 1, airfoil geometry parameterization and computational mesh generation, with target parameters including airfoil model NACA0012, grid cell count of 6000, and y-axis. + The values ​​are automatically derived from the Mach number and Reynolds number; Subtask 2, steady CFD flow field solution, with target parameters including incoming Mach number 0.3 and Reynolds number 5 × 10⁻⁶. 6 1. Angle of attack 3°; Subtask 3: Extraction and analysis of aerodynamic coefficients, outputting lift coefficient, drag coefficient and convergence status report.

[0033] During task decomposition, the orchestration agent automatically identifies missing parameters that are not explicitly provided by the user but can be inferred from physical relationships. For example, the user provides the Mach number and Reynolds number but not y. + When the value is set, the orchestration agent automatically calculates the recommended first-layer mesh height and y-value based on the empirical relationship of the turbulent boundary layer of the flat plate. + Value. When the user does not specify a turbulence model, the orchestration agent automatically selects the applicable default turbulence model based on the Mach number and Reynolds number range.

[0034] (2) Workflow planning The orchestration agent automatically plans the workflow topology based on the type and complexity of the aircraft CFD simulation task. The workflow topology includes linear serial workflows, parallel execution workflows, iterative workflows, and hybrid workflows.

[0035] Linear cascaded workflows are suitable for single-condition steady-state CFD simulation tasks. In a linear cascaded workflow, simulation subtasks are executed sequentially in the order of "mesh generation → CFD solution → result analysis," with the output of the previous simulation subtask serving as the input of the next simulation subtask.

[0036] Parallel execution workflows are suitable for tasks requiring multiple independent computations, such as mesh independence studies. In a parallel execution workflow, the orchestration agent simultaneously schedules multiple sub-agents specializing in CFD simulations of the same type to execute in parallel; for example, generating three sets of meshes of different sizes in parallel or performing three sets of CFD solutions in parallel. After all parallel simulation sub-tasks are completed, the orchestration agent summarizes and compares the results.

[0037] The iterative workflow is suitable for tasks such as aerodynamic shape optimization that require iterative searching within the design space. In this workflow, the orchestration agent executes a loop of "shape parameterization → mesh generation → CFD solution → objective function and gradient evaluation → shape update" until the objective function converges or the preset maximum number of iterations is reached. In each iteration, the orchestration agent automatically adjusts the optimization strategy parameters based on the convergence status and step size of the previous iteration.

[0038] The hybrid workflow is a combination of linear serial workflow, parallel execution workflow, and iterative cyclic workflow. In each iteration of the aerodynamic optimization task, the mesh generation step after the shape update triggers the parallel workflow to evaluate multiple mesh schemes simultaneously.

[0039] (3) Cross-session persistent memory mechanism The orchestration agent maintains a cross-session persistent memory. This memory stores user preference information, environment configuration information, and simulation experience information. User preference information includes commonly used Mach number ranges, preferred turbulence models, default angle of attack ranges, and preferred output formats. Environment configuration information includes the number of available CPU cores, installed solver versions, working directory paths, and a list of currently available professional CFD simulation sub-agents. Simulation experience information includes the quality performance of past mesh parameter combinations, convergence characteristics of specific airfoils under specific conditions, and known solution divergence boundary conditions.

[0040] The cross-session persistent memory stores information in compact factual entries, without recording temporary task progress or intermediate artifact file paths, to avoid memory bloat and contamination by outdated information. At the start of each new session, the orchestration agent automatically loads persistent memory entries related to the current aircraft CFD simulation task. The orchestration agent understands the working environment and historical experience without requiring users to repeatedly provide preference settings and environmental information, enabling continuous accumulation and cross-session reuse of simulation knowledge.

[0041] (4) Automated convergence monitoring and multi-level quality decision making During the execution of the CFD solution task by the professional CFD simulation sub-agent, the orchestration agent continuously receives mid-process state information returned by the professional CFD simulation sub-agent through a standardized protocol bridging layer. This mid-process state information includes the current iteration step number, the current value and historical sequence of each residual component, the convergence history of the lift coefficient, and the convergence history of the drag coefficient. Each residual component includes velocity residuals, pressure residuals, and turbulence variable residuals.

[0042] The convergence monitoring sub-agent evaluates the intermediate state information in real time based on preset multi-level thresholds and generates the corresponding convergence state and quality evaluation results; the orchestration agent makes four types of quality decisions based on the quality evaluation results.

[0043] First-class quality decision: If all components of the residual have fallen below the set convergence criterion and the fluctuation of the force coefficient within the recent iteration window is less than the preset threshold, it is determined that the convergence criterion has been met (pass), the solution is automatically terminated and the post-processing process begins.

[0044] The second type of quality decision: If the residuals exhibit periodic oscillations near the target value but no longer show a monotonically decreasing trend, it is judged as a convergence warning. The preset maximum number of iterations is automatically extended and observation continues. At the same time, a continuing solution instruction is sent to the professional CFD simulation sub-agent.

[0045] The third type of quality decision: If the residual shows a continuous increasing trend or the force coefficient exhibits violent oscillations, it is determined as a solution divergence (fail). The current solution is automatically terminated, the possible causes of divergence are analyzed, and the corresponding parameters are adjusted before the solution task is resubmitted to the professional CFD simulation sub-agent. Possible causes of divergence include insufficient mesh quality, improper boundary condition settings, and excessively large time steps.

[0046] The fourth type of quality decision: If the solution has reached the preset maximum number of iterations but the residual has not yet fallen below the convergence criterion, it is judged as a timeout. The current residual level and force coefficient calculation results are automatically summarized and reported to the user, who can then decide whether to accept the current results or adjust the parameters and recalculate.

[0047] Table 1 Multi-level decision table for automated convergence monitoring of CFD solutions.

[0048] Table 1 presents a multi-level decision table for automated convergence monitoring in CFD solutions. When the decision condition is "pass," the automatic decision is to terminate the solution and proceed to post-processing. When the decision condition is "warning," the automatic decision is to extend the iteration steps and continue observation. When the decision condition is "fail," the automatic decision is to terminate the current solution, analyze the cause of divergence, adjust parameters, and resubmit. When the decision condition is "timeout," the automatic decision is to report the current results and solicit user feedback.

[0049] The decision engine enables fully automated, unattended operation of the CFD solution process. From submitting the calculation to obtaining convergence results or handling anomalies, no manual monitoring of residual curves or force coefficient changes is required.

[0050] Second layer: Standardized protocol bridging layer A standardized protocol bridging layer is positioned between the orchestration agent and the CFD specialized sub-agent cluster. Communication between the orchestration agent and the CFD specialized sub-agent cluster uses the JSON-RPC 2.0 protocol. The standardized protocol bridging layer provides tool discovery interfaces, parameter introspection interfaces, and structured tool invocation interfaces.

[0051] (1) Tool self-discovery interface (tools / list) When the orchestration agent starts up or when a new specialized CFD simulation sub-agent is added to the CFD specialized sub-agent cluster, it calls the tool discovery interface `tools / list`. The `tools / list` interface queries the currently available specialized CFD simulation sub-agents and their corresponding input parameter schemas. Adding or deleting specialized CFD simulation sub-agents is transparent to the orchestration agent. When adding a new machine-wide CFD specialized sub-agent, there is no need to modify the orchestration agent's code or prompts. The orchestration agent obtains the list of available tools and the input parameter schema through the `tools / list` interface.

[0052] (2) Parameter introspection interface (get_skill_input_info) Before invoking a specific professional CFD simulation sub-agent, the orchestration agent calls the parameter introspection interface `get_skill_input_info`. The `get_skill_input_info` interface queries the names, data types, default values, and reasonable value ranges of the parameters required by the professional CFD simulation sub-agent. Based on the parameter schema returned by the `get_skill_input_info` interface, the orchestration agent automatically performs validation and completion operations.

[0053] When user input parameters are incomplete, the orchestration agent uses default values ​​or infers the missing parameters based on physical relationships. Physical relationships include determining the target y based on the turbulence model, near-wall treatment, and incoming flow conditions. + The system determines the range and further estimates the height of the first mesh layer, as well as the number of boundary layers based on the target mesh size and preset mesh generation rules. When user input parameters exceed a reasonable range, the orchestration agent proactively sends confirmation or correction prompts to the user. When there are physical contradictions among multiple parameters provided by the user, the orchestration agent performs consistency checks based on physical constraints and provides correction suggestions.

[0054] Parameter consistency anomalies include: the deviation between the Reynolds number calculated based on the incoming Mach number, characteristic length, temperature, and fluid properties and the user-specified Reynolds number exceeding a preset threshold; and the target mesh size exceeding the preset computational scale range corresponding to the currently available computing resources. The parameter introspection mechanism ensures that every tool call initiated by the orchestration agent to the specialized CFD simulation sub-agent has complete parameters and physical rationality, reducing the probability of solution divergence due to improper parameter settings from the outset.

[0055] (3) Structured tool call interface (tools / call) The orchestration agent initiates standardized JSON-RPC 2.0 requests to specialized CFD simulation sub-agents via the structured tool call interface `tools / call`. The request body includes the tool name and a strongly typed parameter dictionary that has undergone parameter introspection validation. Upon completion, the specialized CFD simulation sub-agent returns a structured response. This response includes execution status flags, quality assessment metrics, and intermediate result data. This standardized calling mechanism achieves complete decoupling between the orchestration layer and the execution layer. Any large language model proxy supporting the JSON-RPC 2.0 protocol can be integrated into the aircraft CFD simulation automation system as an orchestration agent, eliminating the need for custom development for specific large language models.

[0056] Third layer: CFD specialized sub-agent cluster A CFD-specific sub-agent cluster consists of multiple specialized CFD simulation sub-agents. Each specialized CFD simulation sub-agent encapsulates the complete capabilities of a specific simulation stage. These specialized CFD simulation sub-agents are organized using a two-tier Agent+Skill structure. The Agent represents the specialized domain, and the Skill represents the specific operation within that domain. For example, Table 2 shows the Agent-Skill functional mapping relationship for the specialized CFD simulation sub-agent cluster.

[0057] Table 2 Agent-Skill Function Mapping Table for Sub-Agent Cluster

[0058] Each specialized CFD simulation sub-agent, after receiving a JSON-RPC call from the orchestration agent, independently completes its specialized task and returns a structured execution result. Specialized CFD simulation sub-agents do not communicate directly with each other. All coordination and scheduling are handled by the orchestration agent, ensuring a clear, layered architecture and scalability.

[0059] The professional CFD simulation sub-agent is an independently packaged professional CFD functional module that does not depend on a specific CFD solver or commercial / open source software platform.

[0060] This application's method interacts with a professional CFD simulation sub-agent via a standardized JSON-RPC 2.0 protocol. The standardized protocol bridging layer is completely transparent to the internal implementation of the professional CFD simulation sub-agent. The backend of the professional CFD simulation sub-agent can employ any CFD solution framework, including but not limited to general solution frameworks based on the finite volume method, gradient optimization frameworks based on adjoint methods, commercial CFD software, or open-source CFD software. The professional CFD simulation sub-agent exposes its services externally through a tool interface conforming to the protocol specification. The solver-independent architecture ensures that this application's method is applicable to various toolchains in the field of aircraft CFD simulation, and its scope of protection is not limited by any specific software implementation.

[0061] Example: NACA0012 Airfoil Steady CFD Simulation Full Process Automation (1) User input and parameter extraction The user inputs the following into the orchestration agent via the interactive platform: "Perform steady-state CFD simulation of the NACA0012 airfoil, Mach number 0.3, Reynolds number 5 million, angle of attack 3 degrees, 6000 mesh elements, y+ is set to 50." The orchestration agent then resolves and obtains the airfoil type, Mach number, Reynolds number, angle of attack, number of mesh elements, and y+. + Key parameters, etc.

[0062] Table 3 Natural Language Demand Parameter Extraction

[0063] In this case, the user did not explicitly provide the turbulence model parameters. The orchestration agent retrieved the user's preference for using the SA one-equation turbulence model in past simulation tasks from the persistent memory and automatically added it to the parameter configuration.

[0064] Based on the extracted parameters, the orchestration agent automatically decomposes the overall simulation task into three ordered subtasks: Subtask 1, calls the mesh generation sub-agent (generate-cfd-mesh) to generate the computational mesh (6000 elements, y) for the NACA0012 airfoil. + =50); Subtask 2, call the CFD solving sub-agent (run-cfd-simulation) to perform steady flow field solution (Ma=0.3, Re=5×10). 6 (α=3.0°, SA turbulence model); Subtask 3: Extract aerodynamic coefficients and convergence state, and generate a results report.

[0065] (2) Tool self-discovery and parameter verification Before scheduling the sub-agents, the orchestration agent queries the currently available toolchains through the tools / list interface of the protocol bridging layer to confirm that both the airfoil agent's mesh generation tool and CFD solving tool are available. Then, it queries the input parameter schemas of the two tools through the get_skill_input_info interface. Verification results show that the four required parameters for the mesh generation tool (airfoil type, number of mesh cells, y...) are... + The three required parameters for the CFD solution tool (angle of attack, Mach number, Reynolds number, and turbulence model, etc.) have been provided by the user or automatically completed from persistent memory and preset rules, and all parameter values ​​are within a reasonable range, requiring no addition or correction by the user.

[0066] (3) Mesh generation and quality assessment After the parameters pass validation, the orchestration agent initiates a structured call request to the mesh generation sub-agent via the tools / call interface, passing in the airfoil type, target mesh count, and y-axis. + Parameters such as Mach number. After the mesh-generating sub-agent completes its execution, it returns a structured response, including the actual number of meshes generated and the detection values ​​of various quality indicators: Table 4. Mesh quality assessment results for NACA0012 airfoil CFD simulation

[0067] The quality decision engine of the orchestration agent evaluates the above indicators: if all three quality indicators are at the pass level, the mesh quality meets the solver requirements, there is no need to trigger closed-loop feedback to regenerate the mesh, and the subsequent CFD solving tasks can continue to be executed.

[0068] NACA0012 airfoil computational mesh as follows Figure 2 As shown, the near-field boundary layer mesh and leading and trailing edge mesh of the airfoil are displayed.

[0069] (4) CFD solution and convergence monitoring The orchestration agent initiates a structured call request to the CFD solver agent via the tools / call interface, passing in the angle of attack, Mach number, Reynolds number, and mesh data references. The solver agent employs the SIMPLE algorithm for pressure-velocity coupling and uses the Spalart-Allmaras (SA) one-equation turbulence model for turbulent closure, performing iterative solutions for the steady flow field. During the solution process, the orchestration agent continuously receives and evaluates mid-process state information, including pressure residuals, velocity residuals, turbulence variable residuals, and the iteration history of the lift coefficient CL and drag coefficient CD.

[0070] like Figure 3 and 4 As shown, by the 578th iteration, both the pressure residual and the velocity residual had decreased to 10. -6 Below the order of magnitude, the fluctuation range of the lift coefficient CL and drag coefficient CD in the last 200 steps is less than 0.1%. The programming agent determines that the convergence condition is met, automatically terminates the solution, and enters the post-processing process. The final aerodynamic coefficients are: CD=0.0106183, CL=0.324372.

[0071] (5) Results summary and output After the solution is completed, the programming agent integrates the mesh quality assessment results, the solution convergence process, and the final aerodynamic coefficients to automatically generate a structured simulation report. The report includes: mesh quality (6000 elements, non-orthogonality 46.65°, aspect ratio 440.96, skewness 0.87, all three are pass), and solution status (convergence in 578 steps, pressure and velocity residuals both reduced to 10).-6 The simulation parameters include: magnitude, final aerodynamic performance (CD=0.0106183, CL=0.324372), and simulation summary (Ma=0.3, Re=5×10⁻⁶). 6 (α=3.0°). Simultaneously, the orchestration agent writes key experiences from this simulation (the number of convergence iterations for this airfoil under this condition, mesh quality level, etc.) into persistent memory for reference in subsequent similar tasks. Throughout the entire implementation process, from user input of natural language commands to obtaining a complete simulation report including mesh quality, residual curves, aerodynamic coefficients, and flow field contour maps, the orchestration agent automatically completes eight stages: task parsing, tool discovery, parameter verification, mesh generation, quality assessment, CFD solving, convergence monitoring, and result summarization. No manual intervention is required during the normal convergence process in this embodiment.

[0072] Although the embodiments of this application have been described above in conjunction with the accompanying drawings, this application is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of this application, and these are all within the scope of protection of this application.

Claims

1. An automated method for CFD simulation of aircraft based on large language models and multi-agent cooperation, characterized in that, Includes the following steps: S1. Construct a hierarchical architecture including an orchestration agent and a cluster of CFD professional sub-agents. The orchestration agent is driven by a large language model, and the cluster of CFD professional sub-agents includes multiple professional CFD simulation sub-agents. S2. The orchestration agent receives the aircraft CFD simulation requirements input by the user in natural language, extracts physical parameters from the natural language text using a large language model, and decomposes the aircraft CFD simulation requirements into an ordered sequence of simulation sub-tasks. S3. The orchestration agent communicates with the CFD professional sub-agent cluster through a standardized protocol bridging layer. The standardized protocol bridging layer adopts the JSON-RPC 2.0 protocol and provides a tool self-discovery interface, a parameter self-introspection interface, and a structured tool call interface. S4. The orchestration agent queries available professional CFD simulation sub-agents and input parameter patterns through the tool self-discovery interface, verifies the input parameters corresponding to the simulation sub-task sequence through the parameter introspection interface, and issues simulation sub-tasks to the corresponding professional CFD simulation sub-agents through the structured tool call interface. S5. The professional CFD simulation sub-agent executes the simulation sub-task and returns the structured execution result to the orchestration agent; S6. During the execution of the CFD solution task by the professional CFD simulation sub-agent, the orchestration agent continuously receives the intermediate status information returned by the professional CFD simulation sub-agent through the standardized protocol bridging layer, evaluates the intermediate status information in real time according to the preset multi-level thresholds, and automatically makes quality decisions such as continuing iteration, extending the number of iteration steps, adjusting parameters and resubmitting, or terminating the solution. S7. The orchestration agent integrates the structured execution results and outputs the CFD simulation results of the aircraft.

2. The automated CFD simulation method for aircraft based on large language models and multi-agent cooperation as described in claim 1, characterized in that, The tool self-discovery interface is the tools / list interface, the parameter introspection interface is the get_skill_input_info interface, and the structured tool invocation interface is the tools / call interface. The orchestration agent queries the currently available professional CFD simulation sub-agents and their corresponding input parameter modes through the tools / list interface. When adding or replacing a professional CFD simulation sub-agent, the orchestration agent automatically discovers the added or replaced professional CFD simulation sub-agent through the tools / list interface.

3. The automated CFD simulation method for aircraft based on a large language model and multi-agent cooperation as described in claim 2, characterized in that, The orchestration agent queries the professional CFD simulation sub-agent for the parameter names, data types, default values, and value ranges required by the get_skill_input_info interface; and performs at least one operation based on the query results, including missing parameter inference, out-of-range parameter prompting, and contradictory parameter consistency verification.

4. The automated CFD simulation method for aircraft based on large language models and multi-agent cooperation as described in claim 1, characterized in that, The orchestration agent plans the workflow topology based on the type and complexity of the aircraft CFD simulation task. The workflow topology includes linear serial workflow, parallel execution workflow, iterative workflow, or hybrid workflow. In a linear serial workflow, the output of the previous simulation subtask serves as the input of the next simulation subtask; in a parallel execution workflow, the orchestration agent simultaneously schedules multiple CFD simulation sub-agents of the same type to execute in parallel; in a cyclic iterative workflow, the orchestration agent executes a loop of shape parameterization, mesh generation, CFD solution, objective function and gradient evaluation, and shape update until the objective function converges or the maximum number of iterations is reached.

5. The automated CFD simulation method for aircraft based on a large language model and multi-agent cooperation as described in claim 1, characterized in that, The orchestration agent maintains a persistent memory across sessions, which stores user preference information, environment configuration information, and simulation experience information. At the start of each new session, the orchestration agent automatically loads persistent memory entries related to the current aircraft CFD simulation task.

6. The automated CFD simulation method for aircraft based on a large language model and multi-agent cooperation as described in claim 1, characterized in that, The preset multi-level thresholds include a convergence residual threshold, a force coefficient fluctuation threshold, a residual oscillation rate threshold, a residual oscillation amplitude threshold, a residual growth threshold, a force coefficient divergence threshold, and a maximum number of iterations threshold; the quality decision is executed in the following order: When all residual components fall below the convergence residual threshold, and the fluctuation range of the lift coefficient and drag coefficient within the recent iteration window is less than the force coefficient fluctuation threshold, the solution is determined to pass, the solution is terminated, and post-processing begins. When each residual component fails to fall below the convergent residual threshold, and the absolute value of the rate of change of each residual component is less than the residual oscillation rate of change threshold and the peak-to-valley difference of each residual component is less than the residual oscillation amplitude threshold within a consecutive preset number of iterations, a warning is issued, the maximum number of iterations threshold is increased, and observation continues. When each residual component continues to grow within a preset number of iterations and the growth slope is greater than the residual growth threshold, or when the fluctuation range of the lift coefficient and drag coefficient within the recent iteration window is greater than the force coefficient divergence threshold, it is judged as a failure, the current solution is terminated, the cause of divergence is analyzed, the parameters are adjusted and the solution task is resubmitted. When the current iteration step count reaches the maximum iteration step count threshold and each residual component has not decreased below the convergence residual threshold, it is determined to be a timeout. The current residual component levels, lift coefficient, and drag coefficient calculation results are summarized and the user's decision request is output.

7. The automated CFD simulation method for aircraft based on large language models and multi-agent cooperation as described in claim 1, characterized in that, The CFD professional sub-agent cluster is organized according to a two-level structure of Agent+Skill, where Agent represents the professional domain and Skill represents the specific operation within the professional domain; the CFD professional sub-agent cluster includes at least two of the following: mesh generation sub-agent, CFD solving sub-agent, convergence monitoring sub-agent, post-processing analysis sub-agent, and aerodynamic optimization sub-agent.

8. The automated CFD simulation method for aircraft based on large language models and multi-agent cooperation as described in claim 1, characterized in that, The backend of the professional CFD simulation sub-agent adopts a general solution framework of finite volume method, a gradient optimization framework based on adjoint method, commercial CFD software or open source CFD software; the standardized protocol bridging layer is transparent to the internal implementation of the professional CFD simulation sub-agent.

9. The automated CFD simulation method for aircraft based on large language models and multi-agent cooperation as described in claim 1, characterized in that, The orchestration agent supports at least one interactive platform among WeChat, Lark, command line interface and web interface; the CFD professional sub-agent cluster is deployed on a local workstation or cloud server. The CFD simulation results of the aircraft include at least one of the following: calculation results, residual curves, force coefficients, flow field cloud diagrams, and analysis reports.

10. An automated CFD simulation system for aircraft based on a large language model and multi-agent collaboration, characterized in that, include: The orchestration agent, driven by a large language model, is used to receive the aircraft CFD simulation requirements input by the user in natural language, extract physical parameters from the natural language text, decompose the aircraft CFD simulation requirements into an ordered sequence of simulation sub-tasks, and schedule professional CFD simulation sub-agents through a standardized protocol bridging layer. The CFD professional sub-agent cluster includes multiple professional CFD simulation sub-agents, which are used to receive simulation sub-tasks issued by orchestration agents through a standardized protocol bridging layer, execute simulation links in the aircraft CFD simulation process, and return structured execution results. A standardized protocol bridging layer, using the JSON-RPC 2.0 protocol, is used to provide tool self-discovery interfaces, parameter introspection interfaces, and structured tool call interfaces to enable communication between orchestration agents and CFD professional sub-agent clusters; The orchestration agent is also used to continuously receive mid-course state information returned by professional CFD simulation sub-agents through a standardized protocol bridging layer during the CFD solution task, automatically make convergence monitoring and quality decisions based on preset multi-level thresholds, and integrate structured execution results to output the aircraft CFD simulation results.

Images