A multi-agent based CFD simulation method and cloud platform system

Abstract

The application provides a kind of multi-agent-based CFD simulation method and cloud platform system, by fusing multimodal big model, computer vision, intelligent pre and post processing module and cloud computing technology, a set of low threshold, high efficiency, strong intelligent CFD simulation automation system is constructed, so that users describe simulation requirements in natural language on the web side, the system completes parameter extraction and model preparation through pre-processing agent, and automatically executes simulation tasks and post-processing analysis, realizes the whole process closed loop from "language input" to "result output", realizes the simulation and wind load assessment function of building. The application reduces the operation threshold, improves the work efficiency and process consistency by improving the automation degree, provides more accurate and objective analysis support by accurately and intelligently identifying risks, and improves the delivery quality and readability by integrating multimodal results and generating reports.

Description

Technical Field

[0001] This invention belongs to the field of building simulation technology, specifically relating to a CFD simulation method and cloud platform system based on multi-agent systems. Background Technology

[0002] With accelerated urbanization and innovative architectural design concepts, a large number of complex-shaped high-rise buildings, bridges, and stadiums have emerged. These buildings are highly sensitive to wind loads, easily leading to structural instability and high-frequency vibrations, especially in areas with high wind speeds or near coastlines. Therefore, accurate assessment of wind loads is crucial for guiding wind-resistant design.

[0003] In the early stages of building design, wind load assessment primarily relies on physical wind tunnel tests and numerical wind tunnel simulations. Physical wind tunnel tests are costly and time-consuming, making them unsuitable for frequent modifications and testing during the design process. In numerical wind tunnel simulations, CFD technologies are relatively mature both domestically and internationally, with commercial CFD software such as ANSYS Fluent demonstrating powerful capabilities and accuracy in complex flow simulations. While these technologies have been applied in engineering practice in my country, engineering examples of rapid wind load assessments based on current standards remain limited, and users require substantial experience in mesh generation and mature preprocessing capabilities. Traditional numerical wind tunnel analysis reports often face challenges such as high operational difficulty for architects and structural engineers, lengthy design phases, and difficulty in quickly adapting to design changes. This, to some extent, limits their application in the wind resistance assessment and design of building envelopes.

[0004] On the other hand, the rapid development of large language models has brought the application of artificial intelligence into a new stage. Various intelligent agents based on NLP (Natural Language Processing) and multimodal large models have demonstrated good versatility and application capabilities in many fields such as mathematics and computer science.

[0005] To address the challenges of applying CFD in architectural design, some domestic teams have actively explored and implemented solutions. Some teams have developed simulation cloud platforms to provide users with lighter-weight post-processing services, eliminating the tedious process of manually analyzing fluid results. These platforms offer engineers intuitive data visualization and in-depth analysis, thus better assisting in design optimization. Other teams have chosen to embed lightweight simulation modules into structural design software to directly generate simulation results relevant to specifications, thereby facilitating engineers during the design process. However, most existing simulation cloud platforms focus on fixed-mode task uploading and result display, failing to deeply understand the issues designers truly care about. Designers still need to rely on their own experience to summarize and analyze the visualized results, significantly diminishing the platform's application value. Furthermore, the accuracy of calculation results from lightweight simulation modules in structural design software still needs improvement. Compared to mature mainstream CFD software, their calculation precision lags behind, making it difficult to meet the demands of high-precision wind load assessment. Furthermore, these existing methods generally lack in-depth analysis and feedback of the results from large language models, and still require users to have basic fluid analysis and CFD skills, thus failing to fully meet the analysis needs of architects and structural engineers in their practical work.

[0006] The specific drawbacks and problems of existing technologies include: 1. The complexity of operating traditional CFD software. Traditional CFD software, such as ANSYS Fluent and COMSOL, uses graphical interfaces and parameter settings, requiring users to master complex menu operations and a large number of professional parameters.

[0007] Specific issues include: setting boundary conditions requires users to be familiar with fluid mechanics theory and understand professional concepts such as turbulence models and wall functions; mesh generation quality directly affects calculation accuracy, but mesh parameter adjustment requires extensive CFD experience; solver selection and numerical format setting involve dozens of professional parameters, and incorrect settings can lead to calculation divergence or result distortion; convergence judgment and iterative control require expert-level theoretical knowledge.

[0008] 2. The problem of low automation in the simulation process. The existing CFD analysis process relies heavily on manual operation. From geometric modeling and mesh generation to result post-processing, each step requires professional personnel to complete step by step.

[0009] Specific problems include: the geometric preprocessing stage requires manual cleaning of CAD models, simplification of detailed components, and repair of geometric defects, which takes 2-8 hours; mesh generation requires repeated debugging of mesh density and quality, and only expert-level operation can ensure mesh convergence; boundary condition setting depends on the engineer's deep understanding of the physical problem, and batch automated configuration cannot be achieved; the post-processing stage requires manual extraction of key data, drawing of cloud maps, and writing of analysis reports, which takes 1-2 days per project.

[0010] 3. The subjectivity and experience-dependent nature of risk identification. Traditional CFD result analysis mainly relies on experts manually reviewing visualizations such as velocity contour maps and pressure distribution maps, and identifying risk areas based on experience.

[0011] Specific problems include: inconsistent standards for identifying risk areas, leading to different conclusions from different experts regarding the same simulation results; manual analysis is prone to overlooking local high-risk areas, especially hidden locations in complex geometric structures; risk level classification is highly subjective, lacking quantitative risk assessment indicators and objective judgment standards; and the efficiency of batch analysis of a large number of simulation cases is extremely low, with expert resources becoming a bottleneck.

[0012] 4. Issues with the efficiency and standardization of results report generation. The writing of CFD analysis reports relies entirely on manual labor, and the entire process, from data extraction and chart creation to textual expression, lacks standardization and automation.

[0013] Specific problems include: report writing is time-consuming, with a single project report taking 4-16 hours, which seriously affects the project delivery cycle; the report format and content standards are not consistent, and the quality and depth of reports from different engineers vary significantly; key data extraction and chart generation require switching between multiple software programs, which can easily lead to data errors; the expression of professional terminology and the writing of analytical conclusions require engineers to have high writing skills. Summary of the Invention

[0014] The technical problem to be solved by this invention is to provide a CFD simulation method and cloud platform system based on multi-agent systems for simulating buildings and assessing wind loads.

[0015] The technical solution adopted by this invention to solve the above-mentioned technical problems is as follows: a CFD simulation method based on multiple agents, comprising the following steps: S1: Input simulation requirements in the form of natural language, and parse the natural language through a large language model and convert it into CFD simulation parameters; S2: Perform geometric preprocessing on the user-uploaded building model, output the building simulation model, and import it into the CFD simulation database along with the CFD simulation parameters to generate a CFD simulation case. S3: Cloud-based scheduling and parallel solving of CFD simulation cases, horizontal slicing of calculation results and multi-level flow field analysis, outputting CFD simulation result graphs; S4: Perform image analysis on the CFD simulation results and extract high-risk areas by combining the HSV color space and building component recognition algorithm; S5: Synchronizes cloud simulation results to local storage for post-processing, including intelligent analysis based on multimodal large models, integration of fixed templates and parameter information, generation of engineering content, and output of standardized deliverables.

[0016] According to the above scheme, in step S1, based on the parameter combination of the user's structured input, the corresponding CFD simulation configuration is retrieved through keyword indexing and condition matching algorithms, the input parameters are mapped to simulation settings, and a parameter structure that meets the requirements of CFD solution is output; the specific steps are as follows: S11: Establish a CFD professional parameter database, and create a multi-dimensional mapping relationship table including wind speed-boundary conditions, wind direction-inlet settings, and terrain type-roughness parameters; the database adopts a hierarchical storage structure, including standardized data such as standard working condition templates, boundary condition configurations, and grid density settings; S12: Integrates a large language model to build an intelligent parameter tuning engine, which identifies and processes special input requirements through professional prompt templates; establishes a requirement-parameter adjustment rule library, which automatically triggers corresponding parameter optimization when specific keywords are identified; S13: Design a standardized CFD simulation parameter data structure, using JSON format to encapsulate complete simulation information including boundary conditions, solver configuration, and mesh parameters; establish a parameter verification mechanism to verify the physical rationality and numerical stability of the fused parameter combination.

[0017] According to the above scheme, in step S2, the preprocessing operations include model coordinate normalization, component simplification, mesh generation, and boundary condition region identification; the specific steps are as follows: In the model preprocessing stage, the user-uploaded building model is geometrically analyzed to extract the triangular mesh of the building's outer surface; the axial bounding box algorithm is used to calculate the geometric center of the model and translate it to the origin of the numerical wind tunnel coordinate system; the length, width, and height of the model are calculated and stored in the database; a mature solution control file scheme is generated, and the mesh generation file is generated through the key building parameters and the parameter structure obtained in step S1. In the boundary condition setting stage, the computational domain is rotated according to the wind direction angle input in step S1 to make the inlet boundary perpendicular to the incoming flow direction; the inlet wind speed profile is consistent with the atmospheric boundary layer wind profile determined by the specification and written into the 0 / U file to match the target turbulence intensity and wind profile index; the outlet boundary is set as a pressure outlet and the side boundary is set as symmetry to reduce the computational domain size, the building surface is marked and associated with the turbulent wall function; During the scheme generation phase, a steady-state or transient solver is selected based on the flow type, the time step and maximum Courant number are configured, the MPI parallel computing task allocation is optimized to ensure computational efficiency, and a complete CFD simulation scheme is generated.

[0018] According to the above scheme, the specific steps in step S3 are as follows: S31: Allocate computing nodes in the cloud, call command-line tools to perform parallel computing and simulation solutions for CFD simulation cases; monitor residual curves in real time and save simulation output results; S32: Read the calculation results by parsing the internal and boundary grids; prioritize reading the data from the last time step and accurately correspond it to the BIM model using regular expressions; S33: Create a horizontal slice at a specified height, preserving the flow field data of the plane completely; use the building's geometric center as a reference and perform spatial clipping using a four-step clipping method according to a preset multiple; S34: Performs multi-level flow field analysis and outputs CFD simulation results; calculates three-dimensional wind speed based on the vector synthesis principle, uses the exponential wind profile formula for height correction, and automatically calculates the wind speed amplification factor; simultaneously outputs two types of cloud maps, wind speed and wind speed ratio, to meet the requirements of wind environment analysis.

[0019] According to the above scheme, the specific steps in step S4 are as follows: S41: A multi-scale feature extraction algorithm based on computer vision is used to perform geometric analysis on the CFD post-processed cloud map; based on the bounding box size of the building, an adaptive kernel function is used to perform morphological dilation operation, expanding pixels outward proportionally to dynamically construct a region of interest mask; S42: Convert the RGB three-channel image to the HSV cylindrical coordinate system, and dynamically adjust the color threshold parameters according to the analysis type output by the preprocessing operation in step S2 to complete the HSV color space risk identification. S43: Locate the risk blocks corresponding to high wind speeds within the building area, identify independent risk blocks through the connected component labeling algorithm, and mark the risk areas using an adaptive bounding box selection algorithm.

[0020] Furthermore, in step S41, the building outlines are identified by the Sobel gradient operator and the Canny edge detector, straight line features are extracted by combining the Hough transform, and the minimum convex hull of the building complex is constructed using the Graham scan algorithm.

[0021] Furthermore, in step S42, the specific steps are as follows: For wind environment analysis scenarios, when users focus on "pedestrian comfort", the color angle is set to H∈[0°,30°]∪[330°,360°] to identify red-orange areas, corresponding to uncomfortable areas with wind speeds >3m / s; when focusing on "ventilation effect", the angle is expanded to H∈[45°,75°] to identify yellow areas, corresponding to windless areas with wind speeds <1m / s. For wind load analysis scenarios, when users perform "structural safety assessment", the color angle is set to H∈[0°,15°]∪[345°,360°] to identify the dark red area, which corresponds to the extremely high wind pressure area; A risk mask is generated by combining a preset color threshold, and salt-and-pepper noise is eliminated by opening operation of a 3×3 structuring element, while closing operation optimizes edges and eliminates noise.

[0022] According to the above scheme, the specific steps in step S5 are as follows: S51: Input the labeled images generated by risk identification and the embedded expert experience prompts into the multimodal large language model to automatically generate targeted analysis suggestions and optimization measures; including intelligently generating professional content including flow field feature analysis, risk cause explanation and improvement measure suggestions based on parameters such as location, area and wind speed characteristics of the risk area and combined with professional knowledge of building wind engineering. S52: Integrate the analysis content generated in step S51 with the fixed structured report template and user-inputted project parameter information to form a complete standardized CFD evaluation report, including an overview, simulation overview, main results, result analysis, model establishment and mesh generation, basic principles of numerical simulation, and data analysis methods. This report includes the k-ε turbulence model governing equations, OpenFOAM boundary condition settings, SimpleFoam solver iterative methods and wind pressure coefficient calculation formulas, as well as numerical solution methods based on the finite volume method and convergence analysis.

[0023] A cloud platform system includes, The requirement transformation submodule is used to input simulation requirements in the form of natural language, and then parses the natural language through a large language model and converts it into CFD simulation parameters. The preprocessing submodule is used to perform geometric preprocessing operations on the user-uploaded building model, output the building simulation model, and import it into the CFD simulation database together with the CFD simulation parameters to generate CFD simulation cases. The simulation submodule is used to schedule and solve CFD simulation cases in parallel in the cloud, perform horizontal slicing and multi-level flow field analysis on the calculation results, and output CFD simulation result graphs. The risk labeling submodule is used to perform image analysis on the CFD simulation results and extract high-risk areas by combining the HSV color space and building component recognition algorithm. The report submodule is used to synchronize cloud simulation results to the local machine and perform post-processing workflows, including intelligent analysis based on multimodal large models, integration of fixed templates and parameter information, generation of engineering content, and output of standardized deliverables.

[0024] A computer memory storing a computer program executable by a computer processor, the computer program executing a multi-agent CFD simulation method.

[0025] The beneficial effects of this invention are as follows: 1. The present invention provides a CFD simulation method and cloud platform system based on multi-agent technology. By integrating multimodal large language model (LLM), computer vision, intelligent pre- and post-processing modules and cloud computing platform, a low-threshold, high-efficiency and highly intelligent automated CFD (Computational Fluid Dynamics) simulation system is constructed. It completes the entire process of automation from language input, model processing, simulation calculation to result output, and realizes the functions of simulating buildings and assessing wind loads.

[0026] 2. This invention proposes an automated CFD (Computational Fluid Dynamics) simulation system driven by natural language interaction. It allows non-CFD professionals to input their analysis requirements in natural language via a webpage and automatically completes steps such as simulation parameter configuration, building model preprocessing, CFD case generation and deployment, cloud computing, result visualization and intelligent evaluation. Finally, it generates a richly illustrated intelligent analysis report and supports interactive optimization suggestions based on a large language model, aiming to achieve the goals of "low threshold, high efficiency and intelligence".

[0027] 3. This invention addresses the high technical barriers to entry in traditional CFD simulation analysis, which requires users to have a strong foundation in fluid mechanics theory and software operation experience. It relies on professional engineers to manually configure model parameters, boundary conditions, and mesh generation, resulting in a cumbersome, time-consuming process that demands advanced expertise, making it difficult for ordinary engineers to independently complete simulation analysis. This invention significantly lowers the operational threshold by introducing natural language parsing and intelligent parameter matching mechanisms, allowing non-professionals to drive the complete simulation process using simple expressions, thus greatly reducing the barrier to entry.

[0028] 4. This invention addresses the technical problem of complex operation in traditional CFD simulation analysis. Existing technologies mostly adopt semi-automatic processes, requiring manual intervention in multiple stages, resulting in low automation of the entire simulation process. From pre-processing to post-processing, a large amount of manual intervention is needed, and there is a lack of intelligent end-to-end solutions. This invention significantly improves the level of automation, opening up the entire process of automation from simulation requirement input to result output, including parameter extraction, model processing, simulation calculation, image analysis and report generation, which greatly improves work efficiency and process consistency.

[0029] 5. This invention addresses the inefficiency of traditional CFD simulation analysis. Traditional risk assessment relies heavily on manual interpretation of simulation results, risk identification and result analysis depend on human experience, and the identification of key risk points in simulation results mainly relies on expert experience. This approach is highly subjective, inefficient, and lacks objective and standardized intelligent identification methods. This invention provides more accurate and intelligent risk identification by combining HSV image processing, building area recognition, and risk masking technology. It can automatically identify high-wind-speed areas, calculate key indicators, and mark potential risk points, providing more accurate and objective analytical support.

[0030] 6. This invention addresses the technical problem of insufficient intelligence in traditional CFD simulation analysis. Traditional methods require manual compilation of charts and parameter data, resulting in low efficiency in expressing simulation results and generating reports. Professional report writing is time-consuming, has inconsistent formats, and is difficult to achieve batch delivery in engineering. This invention integrates multimodal results and generates reports, automatically integrates simulation images, structured data, and recognition results, and calls language models to generate standardized reports, achieving unified expression from "figure-data-text", thus improving delivery quality and readability.

[0031] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of an embodiment of the present invention.

[0034] Figure 2 This is a flowchart illustrating an embodiment of the present invention.

[0035] Figure 3 This is a schematic diagram of natural language parsing according to an embodiment of the present invention.

[0036] Figure 4 This is a schematic diagram of the overall grid and the grid surrounding the target building in an embodiment of the present invention.

[0037] Figure 5 This is a wind speed cloud map for identifying dangerous areas, as described in an embodiment of the present invention.

[0038] Figure 6 This is a schematic diagram of the wind speed amplification factor cloud map for identifying dangerous areas according to an embodiment of the present invention. Detailed Implementation

[0039] 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 embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0040] Example 1 See Figure 1 The specific steps of a multi-agent-based CFD simulation method are as follows: The overall process is as follows Figure 1 and Figure 2 As shown, it can be divided into five stages: natural language preprocessing agent stage, model preprocessing stage, CFD simulation solution stage, risk point identification stage, and postprocessing agent stage.

[0041] I. Natural Language Parsing Stage (Pre-processing Agent) Users input simulation requirements via natural language on the web interface (e.g., "Analyze the ventilation performance of a building under the prevailing summer wind direction"). The system, based on a large language model (using DeepSeek R1), uses a semantic parsing engine to automatically convert user natural language input into CFD simulation parameters. The core technology implementation path is as follows: CFD professional parameter mapping; An intelligent parameter tuning engine based on a large language model; Parameter structure generation.

[0042] This phase focuses on the intelligent conversion of "language to parameters," providing a semantic foundation for subsequent simulation configuration. The user input interface is shown in the attached image. Figure 3 As shown.

[0043] First, a CFD professional parameter database is established, creating multi-dimensional mapping tables for wind speed-boundary conditions, wind direction-inlet settings, and terrain type-roughness parameters. The database employs a hierarchical storage structure, containing standardized data such as standard working condition templates, boundary condition configurations, and mesh density settings. Based on user-structured input parameter combinations, the system quickly retrieves corresponding CFD simulation configurations using keyword indexing and condition matching algorithms, achieving direct mapping from parameter input to simulation settings. The output parameter structure conforming to CFD solution requirements serves as the input basis for subsequent modeling and simulation processes.

[0044] Secondly, an intelligent parameter tuning engine is built by integrating a large language model to handle special requirements described in the user's notes box. Professional prompt templates enable the model to accurately recognize expressions such as "focus on high wind speed areas" and "consider the interference effect of building clusters." A requirement-parameter adjustment rule library is established, automatically triggering parameter optimization when specific keywords are identified: for example, increasing mesh density for "refined analysis" and adjusting turbulence model configuration for "turbulence effect," achieving intelligent adjustment of simulation parameters based on natural language understanding.

[0045] Finally, a standardized CFD simulation parameter data structure was designed, encapsulating complete simulation information such as boundary conditions, solver configuration, and mesh parameters in JSON format. A parameter verification mechanism was established to check the physical rationality and numerical stability of the fused parameter combinations. A standardized data interface was used to facilitate data exchange with subsequent model preprocessing and numerical solution modules, ensuring automated integration of the simulation process.

[0046] II. Model Preprocessing Stage (Architectural Simulation Model Processing) After natural language parsing is completed, the user uploads the architectural model, and the system automatically performs geometric preprocessing operations, including: Model coordinate normalization; Component simplification (e.g., ignoring small-sized components); Grid generation; Identify boundary condition areas (such as air inlets, air outlets, and the ground).

[0047] The output of this stage is a building simulation model that meets the requirements of CFD simulation. This model, along with the parameter information output from the natural language parsing stage, is imported into the CFD simulation database to ultimately generate a CFD simulation case. Related mesh renderings are shown below. Figure 4 As shown.

[0048] In the model preprocessing stage, the system first performs geometric analysis on the user-uploaded building model. Using the C# language and geometry3Sharp, it extracts triangular meshes (STL format) from the building's outer surface and calculates the model's geometric center using the AABB (Axial Bounding Box) algorithm, translating it to the origin (0,0,0) of the numerical wind tunnel coordinate system. Simultaneously, the system automatically calculates the model's bounding box (length, width, height) and stores this data in a database, providing parameter basis for subsequent mesh generation.

[0049] Taking OpenFOAM as an example, an adaptive numerical wind tunnel mesh setting scheme is generated using BlockMesh and combined with length, width, and height information from the database. The snappyHexMesh scheme is used to automatically set the mesh refinement setting. Surface features (such as eaves and balcony edges) are extracted using surfaceFeatureExtractDict, and curvature and spacing refinement criteria (refinementSurfaces) are set. In the near-wall region (from the building surface to 1.5 times the building height), the system automatically applies boundary layer refinement (layers parameter) to ensure that the y+ value meets the requirements of the turbulence model.

[0050] It should be noted that in conventional calculations, each example file of the open-source OpenFOAM solver must contain three folders: 0, system, and constant. These include the initial and boundary conditions of the computational domain (folder 0), mesh generation files (blockMeshDict, snappyHexMeshDict, and surfaceFeatureExtractDict), and solution control files (controlDict, fvSchemes, and fvSolution). This system can automatically generate mature solution control file schemes. It can intelligently generate mesh generation files based on key building parameters and the parameter structure obtained in the first stage. Specifically, as shown below... Figure 4 As shown.

[0051] During the boundary condition setting phase, the system rotates the computational domain based on the wind direction angle input in the first phase, ensuring the inlet boundary is perpendicular to the incoming flow direction. The inlet wind velocity profile is consistent with the atmospheric boundary layer wind profile defined in the specifications and is written into the 0 / U file, matching the target turbulence intensity and wind profile index. The outlet boundary is set as a pressure outlet (pressureInletOutletVelocity), the side boundaries are set as symmetry to reduce the computational domain size, and the building surface is marked as no-slip (wall type) and associated with the turbulent wall function.

[0052] After mesh generation and boundary condition setting, the system automatically generates a complete CFD simulation scheme. Based on the flow type (steady-state / transient), a steady-state or transient solver is selected, and the time step and maximum Courant number are configured in system / controlDict. Furthermore, the allocation of MPI parallel computing tasks is optimized using decomposeParDict to ensure computational efficiency.

[0053] III. CFD Simulation Solution Stage The system automatically uploads simulation cases to the cloud server (which deploys the open-source CFD software OpenFOAM) and completes the following operations: Task scheduling and parallel solving; Automatically invoke command-line tools to execute simulations; Save the simulation output results (such as velocity field and pressure field) to a database or file system.

[0054] This phase focuses on completing simulation calculation tasks efficiently, in batches, and automatically.

[0055] Specifically, during the simulation and solution phase, the system submits the task to the cloud server, uses the Slurm cluster management system to allocate computing nodes, and calls mpirun to start parallel computing. During the solution process, the system monitors the residual curves in real time (e.g., the residual for the continuity equation is <1e-5) for engineers to test and adjust.

[0056] After calculation, the system performs post-processing analysis using Python's PyVista library. First, it uses POpenFOAMReader from the PyVista library to read the calculation results. This module is specifically optimized for the OpenFOAM data structure, automatically identifying and parsing key data such as the internal mesh and boundary mesh. The system prioritizes reading data from the last time step to ensure the analysis is based on fully converged flow field results. Simultaneously, it automatically extracts model naming information from the Settings.txt file using regular expressions, achieving precise correspondence with the BIM model.

[0057] After extracting the calculation results, a horizontal slice is created at a specified height (default pedestrian height 1.5m) to fully preserve the flow field data of that plane. Then, using the building's geometric center as a reference, spatial clipping is performed according to a preset multiple (usually 3 times the building size). The system employs a four-step clipping method (x / +x / y / +y directions) to ensure the accuracy of the clipping boundaries and automatically calculates the quality index of the clipped mesh, triggering adaptive optimization when necessary. This process can reduce the amount of data by more than 70% while fully preserving the key flow field characteristics around the building.

[0058] The system provides multi-level flow field analysis capabilities. It calculates three-dimensional wind speed based on the vector synthesis principle, performs height correction using the exponential wind profile formula, and automatically calculates the wind speed amplification factor. Under wind environment analysis requirements, it simultaneously outputs both wind speed and wind speed ratio cloud maps. A specially designed fault-tolerant mechanism ensures that reasonable results are still provided even in cases of missing or abnormal parameters.

[0059] It should be noted that numerical wind tunnel calculations often only focus on fluid variables around the target building and load-related variables on the target building's surface. Taking wind environment analysis as an example, the first step is to select the hourly average wind speed U, and then determine the range of automatic flow field clipping based on parameters such as length, width, and height in the database, thereby enabling the automatic generation of wind speed cloud maps and wind speed amplification factor cloud maps at pedestrian height.

[0060] IV. Risk Point Identification Stage (Intelligent Analysis of Simulation Images) The system performs image-based analysis on CFD simulation results (such as velocity contour maps and pressure distribution maps), and intelligently extracts high-risk areas by combining the HSV color space and building component recognition algorithms, as detailed in the attached document. Figure 5 and appendix Figure 6 As shown. Mainly includes: Building area identification and masking construction; HSV color space risk identification; Risk area marking This stage focuses on the intelligent identification and labeling analysis of high-risk areas, providing a basis for subsequent analysis and document generation.

[0061] In the building area identification and processing stage, the system employs a multi-scale feature extraction algorithm based on computer vision to perform geometric analysis on the CFD post-processing cloud image. Building outlines are identified using the Sobel gradient operator and the Canny edge detector, while straight-line features are extracted using Hough transform. The Graham scan algorithm is then used to construct the minimum convex hull of the building complex. Based on the building's bounding box size, an adaptive kernel function is used for morphological dilation, proportionally expanding outwards by a certain number of pixels to dynamically construct a region of interest (ROI) mask. This mask is used to limit the identification range of high-risk areas, thereby improving the accuracy and targeting of the identification.

[0062] In the HSV color space risk identification stage, the system converts the RGB three-channel image to the HSV cylindrical coordinate system and dynamically adjusts the color threshold parameters according to the analysis type parsed by the intelligent agent in the first stage of preprocessing. For wind environment analysis scenarios, when the user focuses on "pedestrian comfort," the hue angle is set to H∈[0°,30°]∪[330°,360°] to identify red-orange areas, corresponding to uncomfortable areas with wind speeds >3m / s; when focusing on "ventilation effect," the focus expands to H∈[45°,75°] to identify yellow areas, corresponding to windless areas with wind speeds <1m / s. For wind load analysis scenarios, when the user performs "structural safety assessment," the hue angle is set to H∈[0°,15°]∪[345°,360°] to identify deep red areas, corresponding to areas with extremely high wind pressure. The system generates a risk mask by combining a preset color threshold, and eliminates salt-and-pepper noise through opening operations of 3×3 structuring elements and closing operations to optimize edges and eliminate noise.

[0063] During the risk area labeling phase, the system locates the risk blocks corresponding to high wind speeds within the building area, identifies independent risk blocks through the Connected Component Labeling algorithm, and draws yellow rectangular label boxes around the risk areas using an adaptive bounding box selection algorithm, adding serial numbers such as "Risk 1" and "Risk 2" to provide a basis for subsequent analysis and document generation.

[0064] V. Post-processing agent stage After the CFD simulation is completed, the system synchronizes the simulation results (such as velocity field, pressure field, vorticity field, etc.) from the cloud to the local machine via the API interface, and starts the post-processing analysis workflow, which mainly includes the following four functions: Intelligent analysis based on multimodal large models; Integration of fixed templates and parameter information; Engineering content is automatically generated; Standardized delivery results output.

[0065] This stage automatically generates professional evaluation reports that meet engineering standards by integrating professional templates, intelligent analysis algorithms, and large language model capabilities.

[0066] In the intelligent analysis phase, the system inputs the labeled images generated from risk identification and embedded expert experience prompts into a multimodal large language model (the system uses a QVQ-MAX model) to automatically generate targeted analysis suggestions and optimization measures. Based on parameters such as the location, area, and wind speed characteristics of the risk area, combined with professional knowledge of building wind engineering, the system intelligently generates professional content such as flow field characteristic analysis, explanation of risk causes, and suggestions for improvement measures.

[0067] During the report integration phase, the system integrates the intelligently generated analysis content with the fixed structured report template and the project parameter information entered by the user to form a complete standardized evaluation report containing seven parts: overview, simulation overview, main results, result analysis, model establishment and mesh generation, basic principles of numerical simulation, and data analysis methods.

[0068] The report uses standard CFD terminology and mathematical expressions, and includes technical details such as the governing equations of the k-ε turbulence model, the setting of OpenFOAM boundary conditions, the iterative method of the SimpleFoam solver, and the formula for calculating the wind pressure coefficient, as well as professional content such as numerical solution methods based on the finite volume method and convergence analysis.

[0069] It should be noted that the final output is a 10-15 page Word technical document, which includes high-resolution color wind speed cloud map and wind speed magnification cloud map, detailed numerical analysis tables, targeted optimization suggestions (such as L-shaped deflector design, building layout adjustment, etc.), and provides A / B / C level wind environment assessment conclusions, realizing the automated and engineering delivery of assessment results, which is convenient for project archiving and decision-making.

[0070] In summary, this embodiment constructs an automated CFD simulation method integrating natural language interaction, automatic model processing, simulation solving, intelligent identification, and result generation. Through the organic collaboration of five stages—from parsing parameter intent using natural language, to model preprocessing and automatic simulation solving, and then to intelligent risk point identification and automatic result generation—a closed-loop process of "language-driven, data-linked, automated computation, and intelligent results" is achieved. The system significantly lowers the professional threshold for CFD simulation, improves the efficiency and accuracy of simulation configuration, enhances the intelligent level of risk identification, and completes engineering delivery in the form of standardized reports, providing an efficient, intelligent, and scalable technical solution for airflow analysis and risk assessment in complex building environments.

[0071] Example 2 The steps in this embodiment are the same as in Embodiment 1, except that each step is applied to a specific instance. Specifically, it includes the following steps: This invention has undergone thorough system function testing, performance verification, and practical engineering applications, proving that the technical solution is entirely feasible. The test environment was based on a local 80-core Linux server, integrating OpenFOAM and DeepSeek + Alibaba Cloud multimodal large model architecture to process user input and initial models. Twenty typical building models were selected, and ten architectural engineers and CFD experts were invited to participate in a one-month system verification process.

[0072] The system's functional test results at each stage demonstrate excellent technical performance. In the natural language parsing stage, the semantic parsing accuracy reached 94.5% for 50 different complexity requirements, with a parameter extraction completeness of 91.2% and an average processing time of 5.3 seconds. In the model preprocessing stage, it supports multiple architectural model formats, achieving a 100% success rate in automatic processing, improving efficiency by 85% compared to traditional manual modeling. In the risk point identification stage, through expert annotation and comparison, the accuracy rate for building area identification reached 96.8%, and the detection accuracy for high-risk areas reached 89.4%, exceeding traditional methods by 15%. In the post-processing intelligent agent stage, the generated analysis report achieved 98.5% completeness and 91.7% accuracy in using professional terminology.

[0073] The system has been successfully applied in 10 real-world engineering projects, including typical cases such as wind environment analysis for a 200-meter twin-tower commercial complex, pollutant diffusion assessment in a chemical industrial park, and ventilation optimization in a high-density residential community. The application has yielded significant results, reducing simulation setup time from the traditional 2 days to 30-60 minutes, saving 30-40% in analysis costs, and shortening project design cycles by 10-20%.

[0074] Taking a certain super high-rise project as an example, the project is located in Danzhou City, Hainan Province. It is necessary to conduct a summer monsoon environmental assessment on the proposed super high-rise building complex to ensure the wind comfort and safety around the buildings.

[0075] Phase 1: Natural Language Preprocessing Agent Project engineers input relevant information through the system's web interface, including geographical location (Danzhou City, Hainan Province), topographic information (Category A), wind direction (southeast wind, S), wind speed (2.2 m / s), and remarks (analyzing the wind environment performance of a super high-rise project under summer southerly monsoon conditions, with a focus on high-wind-speed areas between building clusters), as detailed in the attached document. Figure 3 As shown.

[0076] The system automatically performs semantic parsing based on the large language model (DeepSeek R1), including the following parts: 1. Core simulation objective: Wind environment analysis of building complex; 2. Extract and match key parameters to the CFD parameter library: geographic location information, topographic information, wind direction, and wind speed; 3. Specific needs identification: High wind speed area identification.

[0077] The output parameter structure includes: the purpose is wind environment analysis, the incoming wind speed is 2.2m / s, the wind direction is 90°, the nominal turbulence intensity is 0.12, the gradient wind height is 300m, the cutoff height is 5m, the analysis focuses on the identification of high wind speed areas, and the output is in JSON format.

[0078] Phase Two: Model Preprocessing Users upload BIM model files containing high-rise building complexes. The system automatically executes: 1. Model coordinate normalization: Align the origin of the building model to the origin of the numerical wind tunnel coordinate system, calculate the length, width, and height of the model's bounding box, and store it in the database. 2. Numerical wind tunnel mesh generation: Taking OpenFOAM as an example, BlockMesh is used in conjunction with the length, width and height information in the database to generate an adaptive numerical wind tunnel mesh setting scheme, and snappyHexMesh is used to automatically set the encrypted mesh setting scheme.

[0079] It should be noted that in conventional solver calculations, each example file of the open-source OpenFOAM solver must contain three folders: 0, system, and constant. These include the initial and boundary conditions of the computational domain (folder 0), mesh generation files (blockMeshDict, snappyHexMeshDict, and surfaceFeatureExtractDict), and solver control files (controlDict, fvSchemes, and fvSolution). In the model preprocessing stage of this system, mature solver control file schemes can be automatically generated. Mesh generation files can be intelligently generated using key building parameters and the parameter structure obtained in the first stage, as detailed below. Figure 4 As shown.

[0080] 3. Boundary condition identification: Based on the parameter structure data of the first stage, the direction of the numerical wind tunnel is adjusted to be consistent with the wind direction. Parameters such as incoming wind speed and nominal turbulence intensity are input into the boundary conditions at the inlet of the numerical wind tunnel. The function of identifying boundary conditions such as the flow field outlet and symmetrical boundaries is automatically completed, thereby completing the initial marking of the computational domain and the generation of boundary condition (folder 0) files.

[0081] 4. Generate CFD simulation scheme: After mesh generation and boundary condition setting, based on the mature solver and turbulence model setting scheme in the system, the user-uploaded model and CFD calculation parameter configuration are integrated to automatically generate a computable CFD simulation file for CFD calculation.

[0082] Phase 3: CFD Simulation Solution 1. Simulation solution process: The CFD simulation files generated in the second stage are automatically synchronized to the cloud server, and CFD simulation calculation and analysis are carried out. Based on the native OpenFOAM, the automatic CFD mesh generation and convergence calculation of the building model are completed.

[0083] 2. Python-based post-processing: After the calculation is completed, the CFD results are post-processed using PyVista and VTK. In numerical wind tunnel calculations, only the fluid variables around the target building and the load-related variables on the target building surface are of concern. In this embodiment, the hourly average wind speed U is first selected, and the range of automatic flow field clipping is determined based on the length, width, and height parameters in the database, thereby realizing the automatic generation of wind speed cloud maps and wind speed amplification factor cloud maps at the pedestrian height of the building.

[0084] Phase 4: Risk Identification After the CFD simulation calculation is completed, the system performs intelligent analysis on the generated wind speed cloud map and wind speed magnification factor cloud map at the pedestrian height of the building. The main steps include: Building area recognition: Automatically identify the outline of building groups and generate the convex hull of the building groups; Masking range expansion: Centered on the building complex, expand outwards to a 20-pixel area as the key analysis region; High wind speed area identification and marking: Identify red areas in the HSV color space.

[0085] Four risk points were identified using wind speed cloud maps at pedestrian height for this project, and three more risk points were identified using wind speed amplification factor cloud maps at pedestrian height. (See attached image) Figure 5 and Figure 6 .

[0086] Fifth stage: Post-processing agent After risk point identification is completed, the system synchronizes the simulation results to the cloud and initiates the post-processing analysis workflow, which mainly includes the following functions: Intelligent analysis based on a multimodal large model: The system inputs the labeled images generated from risk identification and embedded expert experience prompts into a multimodal large language model (QVQ-MAX) to automatically generate targeted analysis suggestions and optimization measures. Based on parameters such as the location, area, and wind speed characteristics of the risk area, combined with professional knowledge of building wind engineering, the system intelligently generates professional content such as flow field characteristic analysis, explanations of risk causes, and suggestions for improvement measures.

[0087] Intelligent risk quantification analysis: Combines in-depth analysis of the risk areas identified in the fourth stage, automatically extracts key parameters such as the Reynolds number of the vortex shedding area, the local acceleration coefficient of the funnel effect, the area of ​​the windless zone, and the turbulent kinetic energy density, and conducts compliance verification with GB / T50378-2019 "Evaluation Standard for Green Buildings" and "Standard for Wind Tunnel Test Methods for Building Engineering".

[0088] Fixed template and parameter information integration: The system integrates the intelligently generated analysis content with fixed structured report templates and user-inputted project parameter information to form a complete standardized evaluation report containing seven parts: overview, simulation overview, main results, result analysis, model establishment and mesh generation, basic principles of numerical simulation, and data analysis methods.

[0089] Standardized deliverables: The final output is a 12-page Word technical document, which includes high-resolution color wind speed cloud map and wind speed magnification cloud map, targeted optimization suggestions, and A / B / C level wind environment assessment conclusions. This achieves automated and engineered delivery of assessment results, facilitating project archiving and decision-making.

[0090] In summary, this invention constructs an automated CFD simulation system integrating natural language interaction, automatic model processing, simulation solving, intelligent identification, and result generation. Through the organic collaboration of five stages—from parsing parameter intent using natural language, to model preprocessing and automatic simulation solving, and then to intelligent risk point identification and automatic result generation—a closed-loop process of "language-driven, data-linked, automated calculation, and intelligent results" is achieved. The system significantly lowers the professional threshold for CFD simulation, improves the efficiency and accuracy of simulation configuration, enhances the intelligent level of risk identification, and completes engineering delivery in the form of standardized reports. It provides an efficient, intelligent, and scalable technical solution for airflow analysis and risk assessment in complex building environments. Taking a "super high-rise project" as an example, the entire process from requirement input to report generation takes approximately one hour, improving efficiency by more than 90% compared to traditional manual operation.

[0091] 1. The automatic generation process for CFD simulation cases proposed in this invention primarily utilizes OpenFOAM. OpenFOAM computational files are clearly organized, with data stored in easily parsable key-value pairs and adhering to a specific directory structure and naming conventions. This allows for automatic code generation and replacement through operations such as reading, locating, and modifying text content, enabling efficient automated processing of parameters and model settings, thereby improving the efficiency of computational model construction and modification. Therefore, other CFD software with similar characteristics can also serve as solvers for this system, such as the open-source solver OpenLB, or the LBM solver XFlow, which primarily uses the lattice Boltzmann method. Modifications to the technical solution can be achieved simply by adapting the code to the naming rules of the corresponding CFD software computational files.

[0092] 2. The intelligent analysis process proposed in this invention currently focuses on outdoor wind environment and outdoor wind load, and users can complete the corresponding analysis by uploading a single building model. However, this process can be adapted to various application scenarios in traditional CFD analysis, such as indoor airflow organization and thermal environment analysis, pollutant diffusion analysis, etc. As long as users use more detailed indoor models and describe their analysis needs in natural language, the system can realize the application of technical solutions for various application scenarios after completing the relevant code adaptation.

[0093] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0094] Example 3 This embodiment is used to build a cloud platform system based on the principles of the above method embodiments, including... The requirement transformation submodule is used to input simulation requirements in the form of natural language, and then parses the natural language through a large language model and converts it into CFD simulation parameters. A novel intelligent conversion mechanism from natural language to CFD parameters: This system pioneers the introduction of a large language model into the CFD simulation workflow, constructing a method for extracting operating parameters based on semantic understanding. The system can accurately identify key CFD parameters such as wind speed, wind direction, terrain level, and boundary type from user-input natural language, achieving automatic mapping from "human language" to "simulation language," significantly reducing the professional threshold for simulation setup. Users input analysis requirements in natural language through a web interface, and the system automatically parses and extracts key simulation parameters, such as wind speed, wind direction, terrain level, simulation area, and boundary type, using deep semantic understanding technology, achieving intelligent semantic conversion from "natural language description" to "standardized simulation configuration." The system can identify simulation targets, environmental scenarios, and physical model settings, automatically generating a complete operating condition configuration file containing solver parameters, initial conditions, and boundary condition types. This provides a standardized input interface for mesh generation and numerical calculation, significantly reducing the technical threshold for CFD simulation and enabling non-professional users to quickly start high-precision fluid simulation tasks.

[0095] The preprocessing submodule is used to perform geometric preprocessing operations on the user-uploaded building model, output the building simulation model, and import it into the CFD simulation database together with the CFD simulation parameters to generate CFD simulation cases. Automated CFD Preprocessing Workflow for Building Models: A framework for automated CFD preprocessing technology that integrates geometric processing, boundary recognition, and mesh generation to meet the needs of building simulation. The system's preprocessing agent automatically matches the CFD parameter database, sets boundary conditions and analysis objectives; simultaneously, it performs geometric preprocessing on the user-uploaded building model (coordinate normalization, component simplification, boundary recognition, mesh generation, etc.) to generate a standardized CFD model suitable for simulation.

[0096] The simulation submodule is used to schedule and solve CFD simulation cases in parallel in the cloud, perform horizontal slicing and multi-level flow field analysis on the calculation results, and output CFD simulation result graphs. CFD Simulation Automated Case Generation and Cloud Scheduling Mechanism: This mechanism utilizes templates and scripts to automatically assemble, upload, and remotely schedule simulation cases, forming a one-click simulation workflow that requires no manual intervention. Based on set parameters and the processed model, the system automatically generates complete CFD cases and uploads them to a remote server for remote parallel simulation calculations using CFD solvers such as OpenFOAM.

[0097] The risk labeling submodule is used to perform image analysis on the CFD simulation results and extract high-risk areas by combining the HSV color space and building component recognition algorithm. An image recognition algorithm for CFD simulation results based on HSV color thresholding was developed. This method uses HSV color space thresholding to segment and identify wind speed levels, and combines morphological algorithms to automatically extract high-risk wind speed areas. The method is clearly structured, highly adaptable, and can be widely used for the automatic analysis and annotation of images such as wind speed maps and pressure maps, improving the interpretability of the results. The system automatically visualizes CFD simulation results (such as velocity maps and pressure distribution maps) and uses an image processing method based on the HSV color space for automatic identification of risk areas. Through learning and standardizing the color mapping rules in the simulation images, the system presets multiple HSV color threshold ranges corresponding to different wind speed levels or flow field characteristics; for example, high-wind-speed areas are usually represented by dark red or orange-red. During the recognition process, the system converts the image from RGB space to HSV space, and combines masking operations and morphological processing techniques to accurately extract high-wind-speed areas exceeding the set safety threshold.

[0098] The report submodule is used to synchronize cloud simulation results to the local machine and perform post-processing workflows, including intelligent analysis based on multimodal large models, integration of fixed templates and parameter information, generation of engineering content, and output of standardized deliverables.

[0099] A multimodal-driven intelligent method for generating structured reports: The system integrates multi-source data such as simulation images, structured parameters, and semantic tags, combined with fixed report templates and expert-provided prompts, to guide a large language model to automatically generate Word-format CFD evaluation reports. This mechanism is characterized by high professionalism, clear expression, and a high degree of automation, significantly reducing the cost of manual report writing and improving the efficiency and practical value of results communication. Based on the analysis results, the system integrates multimodal information such as simulation images, structured parameters, and semantic tags, combined with fixed structured report templates and embedded expert-provided prompts, to guide a large language model to generate professional CFD evaluation reports. The template includes pre-set modules such as project overview, working condition configuration, wind field diagrams, risk analysis, and optimization suggestions. The prompts incorporate commonly used terminology, logical expressions, and judgment rules in the field of building simulation, ensuring consistency and authority in the report content in terms of logic, professionalism, and expression style. The final output Word document has a clear structure and combines text and graphics, improving both report generation efficiency and enhancing the engineering application value of simulation results.

[0100] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0101] This embodiment also includes a processor, a communication interface, a memory, and a communication bus; wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the program is executed by the processor, the processor performs the steps of a multi-agent CFD simulation method.

[0102] This embodiment also provides a computer-readable storage medium storing executable instructions that, when executed by a processor, enable the processor to implement a multi-agent-based CFD simulation method.

[0103] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0104] Furthermore, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0105] This application is described with reference to the flowchart of the method and computer program product according to Embodiment 1 and the block diagram of the device (system) according to Embodiment 3. It should be understood that each step or block in the flowchart or block diagram, as well as combinations of steps or blocks in the flowchart or block diagram, can be implemented by computer program instructions.

[0106] These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which are executable by the processor of the computer or other programmable data processing device, produce instructions for implementing the process. Figure 1 One or more processes or boxes Figure 1 A cloud platform system that specifies the functions in one or more boxes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes or boxes Figure 1 The steps of a multi-agent-based CFD simulation method are specified in one or more boxes.

[0109] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A CFD simulation method based on multiple agents, characterized in that: Includes the following steps: S1: Input simulation requirements in the form of natural language, and parse the natural language through a large language model and convert it into CFD simulation parameters; S2: Perform geometric preprocessing on the user-uploaded building model, output the building simulation model, and import it into the CFD simulation database along with the CFD simulation parameters to generate a CFD simulation case. S3: Cloud-based scheduling and parallel solving of CFD simulation cases, horizontal slicing of calculation results and multi-level flow field analysis, outputting CFD simulation result graphs; S4: Perform image analysis on the CFD simulation results and extract high-risk areas by combining the HSV color space and building component recognition algorithm; S5: Synchronizes cloud simulation results to local storage for post-processing, including intelligent analysis based on multimodal large models, integration of fixed templates and parameter information, generation of engineering content, and output of standardized deliverables.

2. The CFD simulation method based on multiple agents according to claim 1, characterized in that: In step S1, based on the parameter combination of the user's structured input, the corresponding CFD simulation configuration is retrieved through keyword indexing and condition matching algorithms, mapping the input parameters to simulation settings, and outputting a parameter structure that meets the requirements of CFD solution; the specific steps are as follows: S11: Establish a CFD professional parameter database, and create a multi-dimensional mapping relationship table including wind speed-boundary conditions, wind direction-inlet settings, and terrain type-roughness parameters; the database adopts a hierarchical storage structure, including standardized data such as standard working condition templates, boundary condition configurations, and grid density settings; S12: Integrates a large language model to build an intelligent parameter tuning engine, which identifies and processes special input requirements through professional prompt templates; establishes a requirement-parameter adjustment rule library, which automatically triggers corresponding parameter optimization when specific keywords are identified; S13: Design a standardized CFD simulation parameter data structure, using JSON format to encapsulate complete simulation information including boundary conditions, solver configuration, and mesh parameters; Establish a parameter verification mechanism to verify the physical rationality and numerical stability of the fused parameter combination.

3. The CFD simulation method based on multiple agents according to claim 1, characterized in that: In step S2, the preprocessing operations include model coordinate normalization, component simplification, mesh generation, and boundary condition region identification; the specific steps are as follows: In the model preprocessing stage, the user-uploaded building model is geometrically analyzed to extract the triangular mesh of the building's outer surface; the axial bounding box algorithm is used to calculate the geometric center of the model and translate it to the origin of the numerical wind tunnel coordinate system; the length, width, and height of the model are calculated and stored in the database; a mature solution control file scheme is generated, and the mesh generation file is generated through the key building parameters and the parameter structure obtained in step S1. In the boundary condition setting stage, the computational domain is rotated according to the wind direction angle input in step S1 to make the inlet boundary perpendicular to the incoming flow direction; the inlet wind speed profile is consistent with the atmospheric boundary layer wind profile determined by the specification and written into the 0 / U file to match the target turbulence intensity and wind profile index; the outlet boundary is set as a pressure outlet and the side boundary is set as symmetry to reduce the computational domain size, the building surface is marked and associated with the turbulent wall function; During the scheme generation phase, a steady-state or transient solver is selected based on the flow type, the time step and maximum Courant number are configured, the MPI parallel computing task allocation is optimized to ensure computational efficiency, and a complete CFD simulation scheme is generated.

4. The CFD simulation method based on multiple agents according to claim 1, characterized in that: The specific steps in step S3 are as follows: S31: Allocate computing nodes in the cloud, call command-line tools to perform parallel computing and simulation solutions for CFD simulation cases; monitor residual curves in real time and save simulation output results; S32: Read the calculation results by parsing the internal and boundary grids; prioritize reading the data from the last time step and accurately correspond it to the BIM model using regular expressions; S33: Creates a horizontal slice at a specified height, preserving the flow field data of that plane completely; Based on the geometric center of the building, spatial cutting is performed using a four-step cutting method according to a preset multiple. S34: Performs multi-level flow field analysis and outputs CFD simulation results; calculates three-dimensional wind speed based on the vector synthesis principle, uses the exponential wind profile formula for height correction, and automatically calculates the wind speed amplification factor; simultaneously outputs two types of cloud maps, wind speed and wind speed ratio, to meet the requirements of wind environment analysis.

5. The CFD simulation method based on multiple agents according to claim 1, characterized in that: The specific steps in step S4 are as follows: S41: A multi-scale feature extraction algorithm based on computer vision is used to perform geometric analysis on the CFD post-processed cloud map; based on the bounding box size of the building, an adaptive kernel function is used to perform morphological dilation operation, expanding pixels outward proportionally to dynamically construct a region of interest mask; S42: Convert the RGB three-channel image to the HSV cylindrical coordinate system, and dynamically adjust the color threshold parameters according to the analysis type output by the preprocessing operation in step S2 to complete the HSV color space risk identification. S43: Locate the risk blocks corresponding to high wind speeds within the building area, identify independent risk blocks through the connected component labeling algorithm, and mark the risk areas using an adaptive bounding box selection algorithm.

6. The CFD simulation method based on multiple agents according to claim 5, characterized in that: In step S41, the outlines of buildings are identified by the Sobel gradient operator and the Canny edge detector, straight line features are extracted by combining the Hough transform, and the minimum convex hull of the building complex is constructed by the Graham scan algorithm.

7. The CFD simulation method based on multiple agents according to claim 5, characterized in that: The specific steps in step S42 are as follows: For wind environment analysis scenarios, when users focus on "pedestrian comfort", the color angle is set to H∈[0°,30°]∪[330°,360°] to identify red-orange areas, corresponding to uncomfortable areas with wind speeds >3m / s; when focusing on "ventilation effect", the angle is expanded to H∈[45°,75°] to identify yellow areas, corresponding to windless areas with wind speeds <1m / s. For wind load analysis scenarios, when users perform "structural safety assessment", the color angle is set to H∈[0°,15°]∪[345°,360°] to identify the dark red area, which corresponds to the extremely high wind pressure area; A risk mask is generated by combining a preset color threshold, and salt-and-pepper noise is eliminated by opening operation of a 3×3 structuring element, while closing operation optimizes edges and eliminates noise.

8. The CFD simulation method based on multiple agents according to claim 1, characterized in that: The specific steps in step S5 are as follows: S51: Input the labeled images generated by risk identification and the embedded expert experience prompts into the multimodal large language model to automatically generate targeted analysis suggestions and optimization measures; including intelligently generating professional content including flow field feature analysis, risk cause explanation and improvement measure suggestions based on parameters such as location, area and wind speed characteristics of the risk area and combined with professional knowledge of building wind engineering. S52: Integrate the analysis content generated in step S51 with the fixed structured report template and user-inputted project parameter information to form a complete standardized CFD evaluation report, including an overview, simulation overview, main results, result analysis, model establishment and mesh generation, basic principles of numerical simulation, and data analysis methods. This report includes the k-ε turbulence model governing equations, OpenFOAM boundary condition settings, SimpleFoam solver iterative methods and wind pressure coefficient calculation formulas, as well as numerical solution methods based on the finite volume method and convergence analysis.

9. A cloud platform system for implementing the multi-agent-based CFD simulation method of any one of claims 1 to 8, characterized in that: The requirement transformation submodule is used to input simulation requirements in the form of natural language, and then parses the natural language through a large language model and converts it into CFD simulation parameters. The preprocessing submodule is used to perform geometric preprocessing operations on the user-uploaded building model, output the building simulation model, and import it into the CFD simulation database together with the CFD simulation parameters to generate CFD simulation cases. The simulation submodule is used to schedule and solve CFD simulation cases in parallel in the cloud, perform horizontal slicing and multi-level flow field analysis on the calculation results, and output CFD simulation result graphs. The risk labeling submodule is used to perform image analysis on the CFD simulation results and extract high-risk areas by combining the HSV color space and building component recognition algorithm. The report submodule is used to synchronize cloud simulation results to the local machine and perform post-processing workflows, including intelligent analysis based on multimodal large models, integration of fixed templates and parameter information, generation of engineering content, and output of standardized deliverables.

10. A computer memory, characterized in that: It contains a computer program that can be executed by a computer processor, which performs a multi-agent-based CFD simulation method as described in any one of claims 1 to 8.

Images