CN-121980876-A - User-free intelligent generation method for cascade simulation grid

CN121980876ACN 121980876 ACN121980876 ACN 121980876ACN-121980876-A

Abstract

The invention relates to the technical field of aviation industry, and provides a method, a system, equipment and a medium for intelligently generating a blade grid simulation grid without operation of a user, wherein the method comprises the following steps of S1, constructing a large language model; S2, obtaining a demand target, S3, decomposing the demand to obtain grid demand parameters and simulation demand parameters, S4, generating and verifying a script, including self-programming and verifying, S5, grid generation and grid quality assessment, driving a grid generation tool to generate grids, after initial grid generation, giving a fine tuning instruction according to a fed-back grid quality report and a large language model, iterating until the fine tuning instruction reaches a preset quality standard, S6, performing simulation execution and feedback modification, starting simulation calculation to obtain a simulation result file and a monitoring process file, and obtaining a final grid simulation result. According to the invention, grids are generated according to requirements through a large language model, the grid quality is optimized, the simulation calculation is performed, the grid quality optimization is fed back, and the problems that grid division depends on engineer experience and the consumed time is too long are solved.

Inventors

ZHANG WEIHAO
ZHAO QINGCHAO
LI LELE

Assignees

北京航空航天大学

Dates

Publication Date: 20260505
Application Date: 20260206

Claims (7)

1. A user-friendly intelligent generation method of a cascade simulation grid is characterized by comprising the following steps: s1, constructing a model, and constructing a large language model oriented to cascade simulation, wherein the large language model carries out fine adjustment training through corpus in the field of aeroengines to obtain a trained large language model adapting to professional terms and task requirements in the field of aeroengine cascade simulation; S2, obtaining a demand target, and obtaining a simulation demand of a user through interaction of the large language model and the user, wherein the simulation demand comprises demand parameters and physical targets of grids and leaf grids; S3, decomposing the requirement, namely decomposing the requirement through the large language model, and firstly identifying key feature words in the simulation requirement, wherein the large language model matches the key feature words with knowledge learned in the large language model training process in the step S1, and the large language model analyzes the requirement to obtain grid requirement parameters and simulation requirement parameters which need to be executed on grids; s4, script generation and verification, which specifically comprises the following steps: s4.1, self-programming, namely compiling a complete grid generation script according to the blade grid geometric parameters and the flow field characteristics among the grid demand parameters obtained in the step S3 and combining the control grammar of each grid generation program learned in the fine-tuning training process, and giving out a complete configuration file of the large language model in simulation software by matching with the grid according to the simulation demand; S4.2, verifying whether a conflict link exists in the running process of the file through the large language model diagnosis, and modifying the grid generation script through the large language model to ensure that no conflict error exists; S5, intelligent grid generation and grid quality evaluation are carried out, a grid generation tool is driven, geometric drawing and calculation domain blocking are carried out according to a grid generation script generated by the large language model, then grid coarse generation is carried out, after initial grid generation, a first fine tuning instruction is given out by the large language model according to a fed-back grid quality report, then the large language model is accessed to a specific requirement in the generation script to generate a second fine tuning instruction, the grid is modified again according to the fed-back grid quality report, a correction method is found out from expert experience to generate a third fine tuning instruction, then a possible modification scheme is proposed through experience in a corpus, the grid is updated continuously, more fine tuning instructions are given out after the process is iterated for a plurality of times until the grid visually and data reach a preset quality standard; s6, simulation execution and feedback modification After the grid is ready, submitting a simulation task to finite element computing software, and depending on the simulation demand parameters obtained in the step S3, generating a simulation script by the large language model to provide computing settings of the simulation software, starting simulation computation to obtain a simulation result file, and monitoring output information of a solver in real time to obtain a monitoring process file; And (3) carrying out evaluation analysis on the simulation result file and the monitoring process file, generating a specific grid modification suggestion by the large language model, returning to the step S4.1, returning to the generation stage of the grid, modifying the part needing modification in the grid generation program, regenerating the grid, carrying out simulation again, and repeating until the simulation result is stable and is judged to meet the initial requirement of the user, thereby obtaining the final grid simulation result.
2. The intelligent generation method of the cascade simulation grid without operation of a user according to claim 1, wherein the intelligent generation method comprises the following steps: The large language model constructing step in the step S1 comprises the following steps: s1.1, selecting a base model, namely selecting DeepSeek-V3.2-Exp as the base model; s1.2, collecting and preprocessing a field corpus, namely collecting a multi-source corpus comprising a cascade pneumatic design document, a CFD simulation report, a grid division guide, a turbulence model theory description, a professional software operation manual and a field expert experience summary text, and cleaning, labeling and structuring; S1.3, supervising and trimming training, namely, adopting an instruction trimming technology to construct a training sample pair, wherein the input of the sample pair is natural language, and the output of the sample pair is a structural characteristic parameter; S1.4, evaluating and iterating, namely optimizing the performance of the model by combining an automatic test and a manual evaluation; And S1.5, system integration, namely deploying the trained large language model, and integrating with a grid generating tool, simulation software and a quality evaluation module.
3. The method for intelligently generating a user-friendly cascade simulation grid according to claim 1, wherein in step S3, the grid demand parameters include: the geometric requirement parameter is that a user provides geometric characteristics of the part; grid base parameters, namely required grid types, grid types commonly used for solving the problem, global maximum and minimum size requirements and grid quantity targets; grid encryption parameters, namely, aiming at experience encryption instructions of a specific area, grid encryption required by a user; Grid quality parameters, namely specifying quality indexes of skewness, length-width ratio and orthogonality; The simulation physical parameters are working medium attribute, boundary condition, turbulence model selection and steady state or transient solver setting.
4. The intelligent generation method of the cascade simulation grid without operation of the user according to claim 1, wherein the evaluation and analysis process of the step S6 comprises the following steps: S6.1, inputting the simulation result file and the file of the monitoring process into the large language model for evaluation, and evaluating whether the physical target proposed in the step S2 is met or not; if the requirement is met, the large language model judges that the grid and the simulation result meet the requirement, and an optimal grid file, a simulation setting file, a result file, a process generated by the large language model and result analysis are obtained; S5.2 judging that the condition is not met or the user judges that the requirement is not met through the large language model, diagnosing the problem of the whole process according to the existing expert experience, and back-pushing to a grid problem which possibly causes the problem.
5. The intelligent generation system of the cascade simulation grid without operation of a user is characterized by comprising the following components: The model building module is used for building a large language model oriented to the cascade simulation, and the large language model carries out fine adjustment training through corpus in the field of aeroengines to obtain a trained large language model adapting to professional terms and task requirements in the field of aeroengine cascade simulation; The demand target module is used for obtaining the simulation demand of the user through the interaction between the large language model and the user, wherein the simulation demand comprises demand parameters and physical targets for grids and leaf grids; The large language model matches the key feature words with knowledge learned in the training process of the large language model, and the large language model analyzes the requirements to obtain grid requirement parameters and simulation requirement parameters which need to be executed on grids; The system comprises a script generation and verification module, a large language model, a simulation module and a simulation module, wherein the script generation and verification module is specifically used for self-programming and verification, wherein the self-programming is that the large language model is used for compiling a complete grid generation script according to the blade grid geometric parameters and flow field characteristics among the obtained grid demand parameters and combining the control grammar of each grid generation program learned in the fine tuning training process, and the large language model is matched with a grid to give a complete configuration file in simulation software so that the complete configuration file can be matched with the grid to enter the grid; The intelligent grid generation and grid quality evaluation module is used for driving a grid generation tool, carrying out geometric drawing and calculation domain blocking according to a grid generation script generated by the large language model, then carrying out grid coarse generation, giving out a first fine tuning instruction according to a fed back grid quality report after initial grid generation, enabling the large language model to be connected with a specific requirement in the generation script to generate a second fine tuning instruction, modifying the grid again according to the fed back grid quality report, finding a correction method from expert experience to generate a third fine tuning instruction, and then providing a possible modification scheme through experience in a corpus, continuously updating the grid, iterating for a plurality of times, and further giving out more fine tuning instructions until the grid visually and data reach a preset quality standard; The simulation execution and feedback modification module is used for submitting a simulation task to finite element calculation software after the grid is ready, generating a simulation script by virtue of the simulation demand parameters, providing calculation setting of the simulation software by virtue of the large language model, starting simulation calculation to obtain a simulation result file, monitoring output information of a solver in real time to obtain a monitoring process file, carrying out evaluation analysis on the simulation result file and the monitoring process file, generating a specific grid modification suggestion by virtue of the large language model, returning to a grid generation stage by virtue of the intelligent grid generation and grid quality evaluation module, modifying a part needing modification in a grid generation program, then regenerating the grid, carrying out simulation again, and repeating until a simulation result is stable and judged to meet the initial demand of a user, and obtaining a final simulation result.
6. Computer device, characterized in that it comprises an input interface and an output interface, and further comprises a processor adapted to implement one or more instructions, a computer storage medium storing one or more instructions adapted to be loaded by the processor and to perform the generation method according to any one of claims 1-4.
7. A computer storage medium storing one or more instructions adapted to be loaded by a processor and to perform the method of generating of any one of claims 1 to 4.

Description

User-free intelligent generation method for cascade simulation grid Technical Field The invention relates to the field of aviation industry, in particular to an intelligent generation method of a cascade simulation grid without operation of a user. Background Meshing of turbine cascades has been the focus of research in the field of computational fluid mechanics. The accuracy and manner of division of the grid largely determines the accuracy and reliability of computational fluid dynamics problems. To capture a boundary layer or wake effect, a region grid needs to be encrypted and geometry fitted. In the conventional design operation, the grid generation takes up to about 60% of the manpower time in the whole calculation period, and if grid optimization and other processes are performed, the grid generation takes up to 90% of the manpower time. For the quality of the meshing, the meshing experience takes decisive role. The high grid density may improve accuracy and geometry, but consumes a lot of computation memory and time, and is costly. Grid sparsity, although computation is fast, there is a possibility of large errors occurring, and grid quality may not be too tight resulting in non-convergence of the computation. A high quality grid is particularly important for the convergence of the simulation calculation, the convergence of the residual curve of the high quality grid is generally better, and calculation with a slightly lower grid quality may not converge the residual curve. For the problems of negative volume grids, excessive length-width ratio, abrupt grid size change and the like in the dividing process, the problems are solved completely by relying on the experience of researchers. In order to improve the traditional method, the development of an automatic meshing technology can obviously reduce the threshold of computational fluid mechanics, so that artificial intelligence technology is proposed to realize meshing. The existing intellectualization is to use artificial intelligence to assist in grid generation, for example, a Graph Neural Network (GNN) is used for grid smoothing or a genetic algorithm is used for grid optimization, but the methods are all used for modifying grids under the condition that human beings finish preliminary grid division, and cannot realize starting from a research target, completely understand the requirements of users, completely get rid of experience dependence and finish grid division by themselves. The most important technical scheme at the grid division of the leaf grating is a completely manual division method. In the field of specialized computational fluid mechanics, after manually creating a grid and completing the division, an operator needs to perform grid quality inspection and manual optimization. If the calculation problem occurs later, the optimization is still performed manually according to the feedback of the calculation result. In particular, when problems such as viscosity, turbulence, and complex flow are involved, the number of iterations required for operation increases significantly. For some commercial software, the automatic partitioning technique is hard-programmed. For a Computer AIDED DESIGN (CAD) model, the commercial software will generate the grid according to a fixed program, following the geometric shape. Such grids have not yet reached the standards for direct use in today's engineering and academic research. There are also experimental techniques for optimizing grid node locations after manually initially completing grid partitioning using machine learning techniques such as Graph Neural Networks (GNNs). A disadvantage of these prior art techniques is that the manual method relies heavily on the CFD knowledge and meshing experience of the engineer. The continuous adjustment and trial-and-error of the ring joint can greatly prolong the grid division working time. Different quality of grids divided by different engineers may cause different calculation accuracy and time of the same problem, and different errors, so that reliability of simulation results is challenged. The automatic grid dividing module of the existing commercial software saves time for generating the preliminary grid, but the grid still needs human intervention and adjustment to participate in the actual calculation process, and meanwhile, the generating program is relatively dead, and can not better meet the diversified research requirements. The grid optimization method adopting machine learning can only be well applied to 2D grids, is limited by the complexity of 3D grids, and cannot optimize three-dimensional grids at present. Disclosure of Invention The invention aims to provide a method, a system, equipment and a medium for intelligently generating a blade grid simulation grid without operation of a user, which are used for generating the grid after understanding the user demands through a large language model, automatically optimizing the grid quality, automat