A CFD simulation intelligent agent system based on multimodal and knowledge enhancement and its autonomous optimization method

Abstract

This invention relates to a CFD simulation intelligent agent system based on multimodal and knowledge enhancement, and its autonomous optimization method, belonging to the field of CFD intelligent simulation technology. It addresses the problems existing in combining CFD simulation with existing AI-assisted tools. The invention adopts a six-layer decoupled architecture: the first layer is the user interaction layer; the second layer is the CFD intelligent agent large model collaboration layer; the third layer is the dedicated scenario small model layer; the fourth layer is the knowledge management and reasoning layer; the fifth layer is the CFD tool adaptation layer; and the sixth layer is the basic support and computing layer. The user interaction layer includes a natural language interface, a sketch input interface, a file upload interface, and a real-time log interface. The CFD intelligent agent large model collaboration layer includes Architect agents, InputWriter agents, Runner agents, and Reviewer agents, supporting message passing and context memory. This invention allows for more flexible expression of simulation requirements.

Description

Technical Field

[0001] This invention belongs to the field of CFD intelligent simulation technology, specifically relating to a CFD simulation intelligent agent system based on multimodal and knowledge enhancement and its autonomous optimization method. Background Technology

[0002] Computational fluid dynamics (CFD) simulations have important applications in aerospace, automotive engineering, and energy and power industries. Traditional CFD simulation workflows include preprocessing (geometric modeling, mesh generation), solver configuration, computation, and post-processing (result visualization, data extraction). Software such as OpenFOAM, Fluent, and STAR-CCM+ require users to manually write or configure numerous parameter files, such as blockMeshDict, controlDict, fvSchemes, and fvSolution.

[0003] Recently, AI-assisted CFD tools based on Large Language Models (LLM) have emerged, such as OpenFOAMGPT and Foam-Agent, attempting to automatically generate CFD configuration files through natural language understanding. However, existing technologies suffer from the following problems: First, high professional threshold and complex configuration. Traditional CFD software requires in-depth knowledge of fluid mechanics and software operation experience, with numerous configuration file parameters, requiring users to repeatedly consult documentation, resulting in high learning costs. Second, limited support for single software and poor versatility. Existing AI-assisted tools only support a single CFD software and cannot be used across platforms, requiring users to relearn when switching software. Third, limited knowledge retrieval and accuracy. Existing technologies mainly retrieve tutorial cases through vector retrieval, but lack the integration and utilization of multi-source knowledge such as command documents, execution scripts, and historical experience, leading to insufficient retrieval accuracy and knowledge coverage. Fourth, limited input methods and weak understanding capabilities. Existing systems can only accept textual descriptions and cannot understand multimodal information such as sketches, geometric images, and existing simulation results provided by users, limiting the ways users can express their needs. Fifth, limited error correction strategies and low efficiency. Existing technologies employ a simple retry strategy, treating all errors uniformly without distinguishing between simple and complex errors, resulting in high correction costs and low efficiency. Summary of the Invention

[0004] The problem this invention aims to solve is that combining traditional CFD simulation methods with existing AI-assisted tools presents challenges such as high professional barriers, single software support, limited knowledge retrieval, restricted input methods, low error correction efficiency, lack of predictive ability, and static knowledge that cannot be learned. The invention proposes a CFD simulation intelligent agent system based on multimodal and knowledge enhancement, and its autonomous optimization method.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A CFD simulation intelligent agent system based on multimodal and knowledge enhancement adopts a six-layer decoupled architecture. The first layer is the user interaction layer; the second layer is the CFD intelligent agent large model collaboration layer; the third layer is the special scenario small model layer; the fourth layer is the knowledge management and reasoning layer; the fifth layer is the CFD tool adaptation layer; and the sixth layer is the basic support and computing layer.

[0007] The user interaction layer includes a natural language interface, a sketch input interface, a file upload interface, and a real-time log interface;

[0008] The CFD agent large model collaboration layer includes Architect agent, InputWriter agent, Runner agent, and Reviewer agent, and supports message passing and context memory;

[0009] The specialized scenario mini-model layer includes a pre-processing intelligent model group, a solution intelligent model group, and a post-processing intelligent model group; the knowledge management and reasoning layer includes a knowledge document library, a historical case library, an execution script library, and a command document library.

[0010] The CFD tool adapter layer supports one or more of the following CFD software: OpenFOAM, Fluent, STAR-CCM+, and CFX.

[0011] The basic support and computing layer includes a data storage management module, a computing hardware resource module, and a computing scheduling and management module.

[0012] Furthermore, the pre-processing intelligent model group includes a geometry recognition and cleaning model, a mesh adaptive generation model, and a boundary condition classification model; the solution intelligent model group includes an initial field intelligent generation model, a flow field accelerated convergence model, a convergence recognition and diagnosis model, a physical information neural network solution model, and an intelligent turbulence modeling model; the post-processing intelligent model group includes an aerodynamic characteristic prediction model, a multi-source data fusion model, a flow field intelligent enhancement model, and a flow field super-resolution reconstruction model.

[0013] Furthermore, the Architect agent is responsible for task decomposition and subtask planning, breaking down the CFD simulation task into multiple subtasks, each corresponding to the generation of a configuration file; the InputWriter agent is responsible for automatically generating CFD configuration files, searching for similar configuration file templates for each subtask, and generating configuration file content that meets user needs based on a large language model; the Runner agent is responsible for running the CFD simulation and monitoring the execution status, executing the generated execution script, monitoring the simulation process in real time, and determining whether the simulation has been successfully completed or failed; the Reviewer agent is responsible for error detection, diagnosis, and correction, and when the Runner agent detects a simulation failure, it initiates a hierarchical error correction process.

[0014] An autonomous optimization method for a CFD simulation intelligent agent system based on multimodal and knowledge enhancement, implemented using the aforementioned CFD simulation intelligent agent system based on multimodal and knowledge enhancement, includes the following steps: S1. Multimodal task understanding and knowledge retrieval: Receive multimodal input from the user, understand and parse it through a multimodal large language model, extract key information of the CFD simulation task, and enhance retrieval through a knowledge management and reasoning layer to obtain multi-source retrieval results;

[0015] S2. Multi-agent collaborative configuration generation: The CFD agent large model collaboration layer and the special scenario small model layer use a message passing mechanism to decompose tasks, generate configuration files, execute simulations and verify results for multi-source retrieval results. If the result verification is successful, proceed to step S5; if the verification fails, proceed to step S3.

[0016] S3. Hierarchical Error Detection and Adaptive Correction: When the result verification fails, the Reviewer agent initiates a hierarchical error correction process, selects different correction strategies according to the error complexity, and obtains the corrected configuration file;

[0017] S4. Predictive error prevention and proactive optimization: Before the simulation is executed, the knowledge management and reasoning layer predicts potential problems and proactively optimizes the configuration by static analysis and historical pattern matching of the corrected configuration file, and then returns to step S2 to generate multi-agent collaborative configuration.

[0018] S5. Intelligent post-processing and result extraction: For successfully verified results, a post-processing workflow is automatically generated to extract the target physical quantity and generate visualized results;

[0019] S6. Continuous Knowledge Base Updates and Adaptive Learning: The configuration, results, and correction strategies of each simulation are recorded in the knowledge management and reasoning layer. The knowledge management and reasoning layer is trained through an adaptive improvement method until the target evaluation conditions are met.

[0020] Furthermore, the specific implementation method of step S1 includes the following steps:

[0021] S1.1. Receive multimodal input from the user, including natural language text input, sketch input, image input, and file input;

[0022] S1.2. Extract physical type, geometric parameters, state parameters, boundary conditions, and solver requirements through a multimodal large language model to obtain key information for the CFD simulation task;

[0023] S1.3. Key information for CFD simulation tasks is retrieved and enhanced through a knowledge management and reasoning layer. A multi-stage, multi-source, and cross-validation retrieval strategy is adopted. The multi-source retrieval results are weighted and fused to obtain the comprehensive similarity score, which is calculated using the following formula:

[0024]

[0025] in, To calculate the overall similarity score, For the first The weight coefficient of each knowledge source For the first The similarity scores of each knowledge source; the weight coefficients for knowledge management and reasoning layer retrieval are set as follows: knowledge document base weight 0.4, historical case base weight 0.3, command document base weight 0.2, and execution script base weight 0.1.

[0026] Furthermore, in step S2, each agent listens to specific outputs, publishes messages to the environment message bus, and the environment maintains a message queue to ensure that agents execute in the order of dependencies, supporting iterative loops. The collaboration between the CFD agent large model collaboration layer and the special scenario small model layer is achieved by automatically selecting the model through a task complexity scoring mechanism. The task complexity scoring formula is:

[0027]

[0028] in, Score for complexity. For geometric complexity, For the complexity of the physical model, For novelty, , , These are the weighting coefficients for geometric complexity, physical model complexity, and novelty, respectively; when Use a dedicated scene small model layer. When using a large model collaboration layer of CFD agents.

[0029] Furthermore, the specific implementation method of step S3 includes the following steps:

[0030] S3.1. Error Classification and Complexity Scoring: The Reviewer agent extracts error output information from the CFD solver and classifies it based on error feature keywords;

[0031] S3.2. Establish an adaptive hierarchical upgrade system, including a Level 1 rapid correction layer, applicable to syntax errors and formatting errors, which uses a small-scale language model in the dedicated scenario small model layer for correction; a Level 2 intelligent correction layer, applicable to unreasonable parameters and conflicting boundary conditions, which uses a multimodal large language model in the CFD agent large model collaboration layer combined with physical constraints for correction; and a Level 3 refactoring correction layer, applicable to solver selection errors and task comprehension deviations, which re-plans the task for correction.

[0032] S3.3. Set the adaptive hierarchical upgrade method to automatically upgrade to the next level and repeat the correction process if the current level correction fails.

[0033] Furthermore, the specific implementation method of step S4 includes the following steps:

[0034] S4.1. Configuration Pre-check: The knowledge management and reasoning layer performs physical parameter rationality checks, mesh quality estimation, file integrity checks, and parameter dependency verification on the generated configuration to obtain the pre-check results; S4.2. Historical Failure Pattern Matching: Extract historical failure cases from the historical case library, calculate the feature similarity between the current configuration and the historical failure cases, and if the similarity is greater than a threshold, issue a warning and provide historical experience; the feature similarity calculation uses cosine similarity, and the calculation formula is:

[0035]

[0036] Where, and are the feature vectors of the two configurations respectively, is the vector dot product, and is the Euclidean norm of the vector;

[0037] S4.3. Proactive Optimization Suggestions: The knowledge management and reasoning layer proposes optimization suggestions based on pre-check results and historical experience;

[0038] S4.4. User Confirmation and Automatic Application: Display warning information and optimization suggestions to the user, and automatically apply the optimization suggestions after the user makes a selection.

[0039] Furthermore, the specific implementation method of step S5 includes the following steps:

[0040] S5.1. Post-processing task identification: Extracting target physical quantities from user requirements; S5.2. Generate post-processing script: Based on the target physical quantity, generate the execution script and VTK format conversion command;

[0041] S5.3. Generate Data Extraction Script: Automatically generate script, read post-processed data files, extract target physical quantities, and save as JSON format; S5.4. Visualization and Report Generation: Use visualization tools to generate flow field cloud maps and comparison charts.

[0042] Furthermore, the specific implementation method of step S6 includes the following steps:

[0043] S6.1. Success Case Recording: Record user requirements, generated configuration files, simulation results, extract feature vectors of successful cases, and add them to the knowledge literature base; S6.2. Failure Case and Correction Strategy Record: Record the error type, error message, correction method, and correction result of the failure, and add it to the historical case library; S6.3. Knowledge Base Vector Update: Regularly vectorize newly added cases and update the FAISS vector index; S6.4. Dynamic adjustment of recommendation weights: The recommendation weights are dynamically adjusted based on the usage frequency and success rate of the cases.

[0044] The beneficial effects of this invention are:

[0045] The present invention discloses a CFD simulation intelligent agent system based on multimodal and knowledge enhancement. By supporting multimodal input methods such as natural language, sketches, and images, it enables users to express simulation needs more intuitively and flexibly without the need for precise textual descriptions or technical terms, thereby effectively reducing the threshold for use.

[0046] The present invention discloses a CFD simulation intelligent agent system based on multimodal and knowledge enhancement. By constructing a multi-source heterogeneous knowledge base (tutorial case library, command help document library, execution script library, and historical case experience library) and adopting a multi-stage fusion retrieval strategy, it can improve retrieval accuracy and generate more accurate configuration files compared with the single tutorial retrieval of the prior art.

[0047] The present invention discloses a CFD simulation intelligent agent system based on multimodal and knowledge enhancement. Through a three-level hierarchical error correction mechanism, different strategies are adopted for errors of different complexities. Compared with the unified retry strategy of existing technologies, the simulation success rate is significantly improved, thereby effectively reducing the number of repeated calculations in traditional CFD simulation. Through intelligent parameter optimization and predictive error prevention, invalid calculations are reduced, saving computing resources and time costs.

[0048] The present invention discloses a CFD simulation intelligent agent system based on multimodal and knowledge enhancement. By configuring pre-checks and historical failure mode matching, it predicts potential problems and actively optimizes before simulation execution, which can reduce invalid calculations by about 50%, save computing resources and time costs, and can also provide a reference for the aerodynamic design of other missile, aircraft and other models. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the structure of a CFD simulation intelligent agent system based on multimodal and knowledge enhancement as described in this invention;

[0050] Figure 2 This is a flowchart of an autonomous optimization method for a CFD simulation intelligent agent system based on multimodal and knowledge enhancement, as described in this invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0052] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0053] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 - Appendix Figure 2 Detailed explanation is as follows:

[0054] Example 1:

[0055] A CFD simulation intelligent agent system based on multimodal and knowledge enhancement adopts a six-layer decoupled architecture. The first layer is the user interaction layer; the second layer is the CFD intelligent agent large model collaboration layer; the third layer is the special scenario small model layer; the fourth layer is the knowledge management and reasoning layer; the fifth layer is the CFD tool adaptation layer; and the sixth layer is the basic support and computing layer.

[0056] The user interaction layer includes a natural language interface, a sketch input interface, a file upload interface, and a real-time log interface;

[0057] The CFD agent large model collaboration layer includes Architect agent, InputWriter agent, Runner agent, and Reviewer agent, and supports message passing and context memory;

[0058] The specialized scenario mini-model layer includes a pre-processing intelligent model group, a solution intelligent model group, and a post-processing intelligent model group; the knowledge management and reasoning layer includes a knowledge document library, a historical case library, an execution script library, and a command document library.

[0059] The CFD tool adapter layer supports one or more of the following CFD software: OpenFOAM, Fluent, STAR-CCM+, and CFX.

[0060] The basic support and computing layer includes a data storage management module, a computing hardware resource module, and a computing scheduling and management module.

[0061] Furthermore, the knowledge management and reasoning layer uses the FAISS vector database for storage and retrieval, and the vectorization encoding adopts the OpenAI text-embedding-ada-002 or HuggingFace Sentence-BERT model.

[0062] Furthermore, the pre-processing intelligent model group includes a geometry recognition and cleaning model, a mesh adaptive generation model, and a boundary condition classification model; the solution intelligent model group includes an initial field intelligent generation model, a flow field accelerated convergence model, a convergence recognition and diagnosis model, a physical information neural network solution model, and an intelligent turbulence modeling model; the post-processing intelligent model group includes an aerodynamic characteristic prediction model, a multi-source data fusion model, a flow field intelligent enhancement model, and a flow field super-resolution reconstruction model.

[0063] Furthermore, the model architecture of the dedicated scenario mini-model includes, but is not limited to: Transformer architecture (for sequence-to-sequence tasks), Convolutional Neural Network (CNN) (for geometric recognition tasks), Graph Neural Network (GNN) (for mesh topology analysis), and ResNet (for flow field feature extraction); the training method of the dedicated scenario mini-model includes: supervised learning using historical CFD simulation data; extracting knowledge from the general large model (DeepSeek-V2) and compressing it into the mini-model (Qwen2.5-7B) using knowledge distillation technology; and optimizing the model's performance in specific tasks through reinforcement learning (PPO, DQN algorithms).

[0064] Furthermore, the Architect agent is responsible for task decomposition and subtask planning, breaking down the CFD simulation task into multiple subtasks, each corresponding to the generation of a configuration file; the InputWriter agent is responsible for automatically generating CFD configuration files, searching for similar configuration file templates for each subtask, and generating configuration file content that meets user needs based on a large language model; the Runner agent is responsible for running the CFD simulation and monitoring the execution status, executing the generated execution script, monitoring the simulation process in real time, and determining whether the simulation has been successfully completed or failed; the Reviewer agent is responsible for error detection, diagnosis, and correction, and when the Runner agent detects a simulation failure, it initiates a hierarchical error correction process.

[0065] Example 2:

[0066] An autonomous optimization method for a CFD simulation intelligent agent system based on multimodal and knowledge enhancement, implemented based on the CFD simulation intelligent agent system based on multimodal and knowledge enhancement described in Example 1, includes the following steps: S1. Multimodal task understanding and knowledge retrieval: Receive multimodal input from the user, understand and parse it through a multimodal large language model, extract key information of the CFD simulation task, and enhance retrieval through a knowledge management and reasoning layer to obtain multi-source retrieval results;

[0067] Furthermore, the specific implementation method of step S1 includes the following steps:

[0068] S1.1. Receive multimodal input from the user, including natural language text input, sketch input, image input, and file input;

[0069] S1.2. Extract physical type, geometric parameters, state parameters, boundary conditions, and solver requirements through a multimodal large language model to obtain key information for the CFD simulation task;

[0070] Furthermore, the state parameters include one or more of the following: Reynolds number Re, Mach number Ma, and angle of attack α;

[0071] S1.3. Key information for CFD simulation tasks is retrieved and enhanced through a knowledge management and reasoning layer. A multi-stage, multi-source, and cross-validation retrieval strategy is adopted. The multi-source retrieval results are weighted and fused to obtain the comprehensive similarity score, which is calculated using the following formula:

[0072]

[0073] in, To calculate the overall similarity score, For the first The weight coefficient of each knowledge source For the first The similarity scores of each knowledge source; the weight coefficients for knowledge management and reasoning layer retrieval are set as follows: knowledge document base weight 0.4, historical case base weight 0.3, command document base weight 0.2, and execution script base weight 0.1.

[0074] S2. Multi-agent collaborative configuration generation: The CFD agent large model collaboration layer and the special scenario small model layer use a message passing mechanism to decompose tasks, generate configuration files, execute simulations and verify results for multi-source retrieval results. If the result verification is successful, proceed to step S5; if the verification fails, proceed to step S3.

[0075] Furthermore, in step S2, each agent listens to specific outputs, publishes messages to the environment message bus, and the environment maintains a message queue to ensure that agents execute in the order of dependencies, supporting iterative loops. The collaboration between the CFD agent large model collaboration layer and the special scenario small model layer is achieved by automatically selecting the model through a task complexity scoring mechanism. The task complexity scoring formula is:

[0076]

[0077] in, Score for complexity. For geometric complexity, For the complexity of the physical model, For novelty, , , These are the weighting coefficients for geometric complexity, physical model complexity, and novelty, respectively; when Use a dedicated scene small model layer. When using a large model collaboration layer of CFD agents.

[0078] Furthermore, the Architect agent analyzes the file structure of similar cases and decomposes the CFD simulation task into multiple sub-tasks.

[0079] Furthermore, the InputWriter agent retrieves similar configuration file templates for each subtask, generates configuration files based on the large language model, and generates an execution script (Allrun) containing commands such as blockMesh, checkMesh, and {solver}.

[0080] Furthermore, the Runner agent executes the Allrun script, starts the CFD solver, monitors the simulation process in real time, and captures standard output and error output;

[0081] Furthermore, during the execution of the intelligent agent, specialized scenario-specific small models can be selectively invoked to assist decision-making: in the preprocessing stage, geometric recognition and mesh generation models and boundary condition classification models are invoked; in the solution stage, initial field intelligent generation models and flow field accelerated convergence models are invoked; and in the postprocessing stage, aerodynamic performance intelligent evaluation models and flow field intelligent analysis models are invoked.

[0082] S3. Hierarchical Error Detection and Adaptive Correction: When the result verification fails, the Reviewer agent initiates a hierarchical error correction process, selects different correction strategies according to the error complexity, and obtains the corrected configuration file;

[0083] Furthermore, the specific implementation method of step S3 includes the following steps:

[0084] S3.1. Error Classification and Complexity Scoring: The Reviewer agent extracts error output information from the CFD solver and classifies it based on error feature keywords;

[0085] S3.2. Establish an adaptive hierarchical upgrade system, including a Level 1 rapid correction layer, applicable to syntax errors and formatting errors, which uses a small-scale language model in the dedicated scenario small model layer for correction; a Level 2 intelligent correction layer, applicable to unreasonable parameters and conflicting boundary conditions, which uses a multimodal large language model in the CFD agent large model collaboration layer combined with physical constraints for correction; and a Level 3 refactoring correction layer, applicable to solver selection errors and task comprehension deviations, which re-plans the task for correction.

[0086] Furthermore, the physical constraint verification includes the following conditions:

[0087] (1) CFL condition check:

[0088]

[0089] in, For time step, For flow rate, This refers to the grid size;

[0090] (2) Matching of Reynolds number with turbulence model;

[0091] (3) The Mach number matches the solver type;

[0092] (4) Boundary conditions and solver compatibility check;

[0093] (5) Grid quality assessment: aspect ratio, value;

[0094] S3.3. Set the adaptive hierarchical upgrade method to automatically upgrade to the next level and repeat the correction process if the current level correction fails.

[0095] S4. Predictive error prevention and proactive optimization: Before the simulation is executed, the knowledge management and reasoning layer predicts potential problems and proactively optimizes the configuration by static analysis and historical pattern matching of the corrected configuration file, and then returns to step S2 to generate multi-agent collaborative configuration.

[0096] Furthermore, the specific implementation method of step S4 includes the following steps:

[0097] S4.1. Configuration Pre-check: The knowledge management and reasoning layer performs physical parameter rationality checks, mesh quality estimation, file integrity checks, and parameter dependency verification on the generated configuration to obtain the pre-check results; S4.2. Historical Failure Pattern Matching: Extract historical failure cases from the historical case library, calculate the feature similarity between the current configuration and the historical failure cases, and if the similarity is greater than a threshold, issue a warning and provide historical experience; the feature similarity calculation uses cosine similarity, and the calculation formula is:

[0098]

[0099] Where, and are the feature vectors of the two configurations respectively, is the vector dot product, and is the Euclidean norm of the vector;

[0100] S4.3. Proactive Optimization Suggestions: The knowledge management and reasoning layer proposes optimization suggestions based on pre-check results and historical experience;

[0101] S4.4. User Confirmation and Automatic Application: Display warning information and optimization suggestions to the user, and automatically apply the optimization suggestions after the user makes a selection.

[0102] S5. Intelligent post-processing and result extraction: For successfully verified results, a post-processing workflow is automatically generated to extract the target physical quantity and generate visualized results;

[0103] Furthermore, the specific implementation method of step S5 includes the following steps:

[0104] S5.1. Post-processing task identification: Extracting target physical quantities from user requirements; S5.2. Generate post-processing script: Based on the target physical quantity, generate the execution script and VTK format conversion command;

[0105] S5.3. Generate Data Extraction Script: Automatically generate script, read post-processed data files, extract target physical quantities, and save as JSON format; S5.4. Visualization and Report Generation: Use visualization tools to generate flow field cloud maps and comparison charts.

[0106] Furthermore, Tecplot or other visualization tools can be used to generate flow field cloud maps;

[0107] S6. Continuous Knowledge Base Updates and Adaptive Learning: The configuration, results, and correction strategies of each simulation are recorded in the knowledge management and reasoning layer. The knowledge management and reasoning layer is trained through an adaptive improvement method until the target evaluation conditions are met.

[0108] Furthermore, the specific implementation method of step S6 includes the following steps:

[0109] S6.1. Success Case Recording: Record user requirements, generated configuration files, simulation results, extract feature vectors of successful cases, and add them to the knowledge literature base; S6.2. Failure Case and Correction Strategy Record: Record the error type, error message, correction method, and correction result of the failure, and add it to the historical case library; S6.3. Knowledge Base Vector Update: Regularly vectorize newly added cases and update the FAISS vector index; S6.4. Dynamic adjustment of recommendation weights: The recommendation weights are dynamically adjusted based on the usage frequency and success rate of the cases.

[0110] Furthermore, the target evaluation criteria compare the mean square error and the mean absolute error as prediction errors.

[0111] The formula for calculating the mean squared error (MSE) is:

[0112]

[0113] in, For the sample size, For the true value, This is a predicted value;

[0114] The formula for calculating the mean absolute error (MAE) is:

[0115] .

[0116] The specific application examples of this embodiment are as follows:

[0117] Case 1: Simulation of laminar flow-driven square cavity flow;

[0118] User input: "Simulate flow in a square cavity with a Reynolds number of 1000, with a velocity of 1 m / s on the top wall and the other three walls stationary";

[0119] The system execution flow is as follows:

[0120] Step A: The Architect agent receives user input and extracts key information through a large language model: Re=1000, the geometry type is a square cavity, the top wall is a moving wall (1m / s), and the other three walls are stationary walls;

[0121] Step B: The Architect agent searches the knowledge base and matches the cavity case of OpenFOAM (similarity 0.95), which is a classic example of square cavity driven flow.

[0122] Step C: The Architect agent decomposes the task into 5 sub-tasks: (1) Write blockMeshDict for mesh generation, (2) Write controlDict for solver control, (3) Write 0 / U boundary condition file, (4) Write 0 / p boundary condition file, and (5) Write Allrun execution script.

[0123] Step D: The InputWriter agent generates the following configuration files in sequence: blockMeshDict is set to a 20×20 grid, controlDict is set to endTime=1s, and 0 / U is set to the top wall velocity as (1 0 0) and the other walls as (0 0 0).

[0124] Step E: The Runner agent executes the Allrun script, which sequentially runs blockMesh to generate the mesh, checkMesh to check the mesh quality, and icoFoam to start the solver;

[0125] Step F: The simulation was completed successfully without errors.

[0126] Step G: The post-processing module automatically extracts the velocity and pressure fields and generates VTK format for Tecplot visualization;

[0127] This embodiment demonstrates the system's ability to process standard CFD tasks. The entire process takes about 2 minutes, which is 15 times more efficient than manual configuration (about 30 minutes).

[0128] Case 2: Multimodal Input Case;

[0129] A user uploads a hand-drawn airfoil sketch and enters the text: "Simulate the aerodynamic characteristics of this airfoil under a 100m / s flow."

[0130] The system execution flow is as follows:

[0131] Step A: The multimodal large language model (GPT-4V or Qwen-VL) identifies the airfoil contour in the sketch and uses an edge detection algorithm (Canny operator) to extract feature points;

[0132] Step B: The visual model uses a contour fitting algorithm (such as B-spline curve fitting) to measure the chord length, maximum thickness, leading edge radius, and other geometric parameters of the airfoil;

[0133] Step C: Based on the textual description, determine the incoming flow velocity as U = 100 m / s and the Reynolds number as Re = 6.7 × 10⁻⁶. 5 (Assume air and chord length 1m);

[0134] Step D: The system determines that the flow is turbulent, and it is recommended to use the simpleFoam solver in conjunction with the k-omega SST turbulence model;

[0135] Step E: Generate the configuration file and execute the simulation;

[0136] This embodiment demonstrates the system's multimodal understanding capability. Users do not need to accurately describe geometric parameters; they only need to provide a sketch, which lowers the professional threshold.

[0137] Case 3: Predictive Optimization Case;

[0138] User input: "Simulate supersonic airfoil flow, Ma=2.0";

[0139] The system execution flow is as follows:

[0140] Step A: Architect generates the initial configuration and selects the rhoCentralFoam compressible solver;

[0141] Step B: Before execution, the predictive optimization module performs a configuration pre-check;

[0142] Step C: Inspection revealed that the time step Δt = 0.001s. According to the CFL condition, the CFL number is approximately 2.5 > 1, indicating a risk of numerical instability.

[0143] Step D: The system proactively proposes optimization suggestions: "The number of CFLs is too large. It is recommended to reduce the time step to 0.0004s or increase the mesh density."

[0144] Step E: The user adopts the suggestion, and the system automatically adjusts deltaT in controlDict to 0.0004;

[0145] Step F: Perform the simulation. The simulation converged successfully, avoiding failure caused by an excessively large number of CFLs.

[0146] This embodiment demonstrates the system's predictive optimization capabilities, proactively identifying potential problems and providing suggestions before execution, thus avoiding ineffective calculations and saving time costs.

[0147] Case 4: Acceleration Case Study Using Dedicated Small Models;

[0148] User input: "Simulate the aerodynamic characteristics of the standard NACA0012 airfoil under the condition Re=1e6";

[0149] The system execution flow is as follows:

[0150] Step A: The Architect agent extracts task features: geometry = standard airfoil NACA0012, Re = 1e6, turbulent flow;

[0151] Step B: Calculate task complexity: G=2 (standard geometry), P=5 (turbulence), N=1 (common task), C_score = 0.3×2 + 0.5×5 + 0.2×1 = 3.3 < 5;

[0152] Step C: The system determines that it is a routine task and calls the dedicated small model (Qwen2.5-7B-Instruct, with 7-8B parameters).

[0153] Step D: The "Boundary Condition Classification Model" (a text classifier finely tuned based on the BERT-base architecture) of the preprocessing intelligent model group automatically identifies the following: the inlet is the velocity inlet, the outlet is the pressure outlet, the airfoil surface is the wall, and the upper and lower boundaries are the far field. The classification accuracy is 98.5%, and the time taken is 0.5 seconds.

[0154] Step E: Solve the "Initial Field Intelligent Generation Model" of the intelligent model group (based on the U-Net convolutional neural network architecture). Learn the flow field characteristics from similar historical cases to generate reasonable initial velocity and pressure fields. Use physical information neural network (PINN) constraints to ensure that the Navier-Stokes equations are satisfied, avoid starting from zero field, and speed up the convergence speed.

[0155] Step F: InputWriter generates the configuration file, Runner executes the simulation, and convergence occurs in 150 iterations (300 iterations are required if starting from zero field).

[0156] This example demonstrates the application value of a dedicated small model: compared to traditional methods that do not use intelligent models at all, the number of convergence iterations is reduced by 50%, and the overall efficiency is improved by about 2 times.

[0157] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0158] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A CFD simulation intelligent agent system based on multimodal and knowledge enhancement, characterized in that, It adopts a six-layer decoupled architecture: the first layer is the user interaction layer; the second layer is the CFD agent large model collaboration layer; the third layer is the special scenario small model layer; the fourth layer is the knowledge management and reasoning layer; the fifth layer is the CFD tool adaptation layer; and the sixth layer is the basic support and computing layer. The user interaction layer includes a natural language interface, a sketch input interface, a file upload interface, and a real-time log interface; The CFD agent large model collaboration layer includes Architect agent, InputWriter agent, Runner agent, and Reviewer agent, and supports message passing and context memory; The specialized scenario mini-model layer includes a pre-processing intelligent model group, a solution intelligent model group, and a post-processing intelligent model group; the knowledge management and reasoning layer includes a knowledge document library, a historical case library, an execution script library, and a command document library. The CFD tool adapter layer supports one or more of the following CFD software: OpenFOAM, Fluent, STAR-CCM+, and CFX. The basic support and computing layer includes a data storage management module, a computing hardware resource module, and a computing scheduling and management module.

2. The CFD simulation intelligent agent system based on multimodal and knowledge enhancement according to claim 1, characterized in that, The preprocessing intelligent model group includes a geometry recognition and cleaning model, a mesh adaptive generation model, and a boundary condition classification model; the solution intelligent model group includes an initial field intelligent generation model, a flow field accelerated convergence model, a convergence recognition and diagnosis model, a physical information neural network solution model, and an intelligent turbulence modeling model; the postprocessing intelligent model group includes an aerodynamic characteristic prediction model, a multi-source data fusion model, a flow field intelligent enhancement model, and a flow field super-resolution reconstruction model.

3. The CFD simulation intelligent agent system based on multimodal and knowledge enhancement according to claim 2, characterized in that, The Architect agent is responsible for task decomposition and subtask planning, breaking down the CFD simulation task into multiple subtasks, each subtask corresponding to the generation of a configuration file; the InputWriter agent is responsible for automatically generating CFD configuration files. For each subtask, it searches for similar configuration file templates and generates configuration file content that meets user needs based on a large language model. The Runner agent is responsible for running the CFD simulation and monitoring the execution status, executing the generated execution script, monitoring the simulation process in real time, and determining whether the simulation has been successfully completed or failed. The Reviewer agent is responsible for error detection, diagnosis, and correction. When the Runner agent detects a simulation failure, it initiates a hierarchical error correction process.

4. An autonomous optimization method for a CFD simulation intelligent agent system based on multimodal and knowledge enhancement, implemented using the CFD simulation intelligent agent system based on multimodal and knowledge enhancement as described in any one of claims 1-3, characterized in that, The process includes the following steps: S1. Multimodal task understanding and knowledge retrieval: Receive multimodal input from the user, understand and parse it through a multimodal large language model, extract key information of the CFD simulation task, and enhance retrieval through a knowledge management and reasoning layer to obtain multi-source retrieval results; S2. Multi-agent collaborative configuration generation: The CFD agent large model collaboration layer and the special scenario small model layer use a message passing mechanism to decompose tasks, generate configuration files, execute simulations and verify results for multi-source retrieval results. If the result verification is successful, proceed to step S5; if the verification fails, proceed to step S3. S3. Hierarchical Error Detection and Adaptive Correction: When the result verification fails, the Reviewer agent initiates a hierarchical error correction process, selects different correction strategies according to the error complexity, and obtains the corrected configuration file; S4. Predictive error prevention and proactive optimization: Before the simulation is executed, the knowledge management and reasoning layer predicts potential problems and proactively optimizes the configuration by static analysis and historical pattern matching of the corrected configuration file, and then returns to step S2 to generate multi-agent collaborative configuration. S5. Intelligent post-processing and result extraction: For successfully verified results, a post-processing workflow is automatically generated to extract the target physical quantity and generate visualized results; S6. Continuous Knowledge Base Updates and Adaptive Learning: The configuration, results, and correction strategies of each simulation are recorded in the knowledge management and reasoning layer. The knowledge management and reasoning layer is trained through an adaptive improvement method until the target evaluation conditions are met.

5. The autonomous optimization method for a CFD simulation intelligent agent system based on multimodal and knowledge enhancement according to claim 4, characterized in that, The specific implementation method of step S1 includes the following steps: S1.

1. Receive multimodal input from the user, including natural language text input, sketch input, image input, and file input; S1.

2. Extract physical type, geometric parameters, state parameters, boundary conditions, and solver requirements through a multimodal large language model to obtain key information for the CFD simulation task; S1.

3. Key information for CFD simulation tasks is retrieved and enhanced through a knowledge management and reasoning layer. A multi-stage, multi-source, and cross-validation retrieval strategy is adopted. The multi-source retrieval results are weighted and fused to obtain the comprehensive similarity score, which is calculated using the following formula: ； in, To calculate the overall similarity score, For the first The weight coefficient of each knowledge source For the first The similarity scores of each knowledge source; the weight coefficients for knowledge management and reasoning layer retrieval are set as follows: knowledge document base weight 0.4, historical case base weight 0.3, command document base weight 0.2, and execution script base weight 0.

1.

6. The autonomous optimization method for a CFD simulation intelligent agent system based on multimodal and knowledge enhancement according to claim 5, characterized in that, Step S2: Each agent listens to specific outputs and publishes messages to the environment message bus. The environment maintains a message queue to ensure that agents execute in the order of dependencies, supporting iterative loops. The collaboration between the CFD agent large model collaboration layer and the special scenario small model layer is achieved by automatically selecting the model through a task complexity scoring mechanism. The task complexity scoring formula is: ； in, Score for complexity. For geometric complexity, For the complexity of the physical model, For novelty, , , These are the weighting coefficients for geometric complexity, physical model complexity, and novelty, respectively; when Use a dedicated scene small model layer. When using a large model collaboration layer of CFD agents.

7. The autonomous optimization method for a CFD simulation intelligent agent system based on multimodal and knowledge enhancement according to claim 6, characterized in that, The specific implementation method of step S3 includes the following steps: S3.

1. Error Classification and Complexity Scoring: The Reviewer agent extracts error output information from the CFD solver and classifies it based on error feature keywords; S3.

2. Establish an adaptive hierarchical upgrade, including a Level 1 fast correction layer, which is suitable for syntax errors and formatting errors, and uses a small-scale language model in a dedicated scenario small model layer for correction; Level 2 intelligent correction layer is suitable for unreasonable parameters and conflicting boundary conditions. It uses the multimodal large language model in the CFD agent large model collaboration layer to correct the problem by combining physical constraints. Level 3 reconstruction correction layer is suitable for solver selection errors and task comprehension deviations. It re-plans the task to correct the problem. S3.

3. Set the adaptive hierarchical upgrade method to automatically upgrade to the next level and repeat the correction process if the current level correction fails.

8. The autonomous optimization method for a CFD simulation intelligent agent system based on multimodal and knowledge enhancement according to claim 7, characterized in that, The specific implementation method of step S4 includes the following steps: S4.

1. Configuration Pre-check: The knowledge management and reasoning layer performs physical parameter rationality checks, mesh quality estimation, file integrity checks, and parameter dependency verification on the generated configuration to obtain the pre-check results; S4.

2. Historical Failure Pattern Matching: Extract historical failure cases from the historical case library, calculate the feature similarity between the current configuration and the historical failure cases, and if the similarity is greater than a threshold, issue a warning and provide historical experience; the feature similarity calculation uses cosine similarity, and the calculation formula is: ； Where, and are the feature vectors of the two configurations respectively, is the vector dot product, and is the Euclidean norm of the vector; S4.

3. Proactive Optimization Suggestions: The knowledge management and reasoning layer proposes optimization suggestions based on pre-check results and historical experience; S4.

4. User Confirmation and Automatic Application: Display warning information and optimization suggestions to the user, and automatically apply the optimization suggestions after the user makes a selection.

9. The autonomous optimization method for a CFD simulation intelligent agent system based on multimodal and knowledge enhancement according to claim 8, characterized in that, The specific implementation method of step S5 includes the following steps: S5.

1. Post-processing task identification: Extracting target physical quantities from user requirements; S5.

2. Generate post-processing script: Based on the target physical quantity, generate the execution script and VTK format conversion command; S5.

3. Generate Data Extraction Script: Automatically generate script, read post-processed data files, extract target physical quantities, and save as JSON format; S5.

4. Visualization and Report Generation: Use visualization tools to generate flow field cloud maps and comparison charts.

10. The autonomous optimization method for a CFD simulation intelligent agent system based on multimodal and knowledge enhancement according to claim 9, characterized in that, The specific implementation method of step S6 includes the following steps: S6.

1. Success Case Recording: Record user requirements, generated configuration files, simulation results, extract feature vectors of successful cases, and add them to the knowledge literature base; S6.

2. Failure Case and Correction Strategy Record: Record the error type, error message, correction method, and correction result of the failure, and add it to the historical case library; S6.

3. Knowledge Base Vector Update: Regularly vectorize newly added cases and update the FAISS vector index; S6.

4. Dynamic adjustment of recommendation weights: The recommendation weights are dynamically adjusted based on the usage frequency and success rate of the cases.

Images