Aircraft surface physical field prediction method based on language model

Abstract

The invention discloses an aircraft surface physical field prediction method based on a language model, and belongs to the field of aircraft design and deep learning. Comprising the following steps: S1, generating an initial shape of an aircraft, and parameterizing the shape through a parameterization method with physical meanings to determine a post-parameterization parameter deformation range; s2, generating a new appearance, a surface grid and a space grid according to the determined parameter deformation range; s3, performing CFD solution on the new appearance; and S4, taking the shape and the flight state of the aircraft as features, taking the surface physical field distribution as labels, inputting codes into a Bert model for training, and obtaining an available model after logistic regression. And S5, obtaining the surface physical field distribution of the aircraft. The method has the advantages that parameterized geometry is not limited by grid generation; constructing an aircraft surface physical field prediction model based on the language model; the constructed model directly takes aircraft coordinate points and flight states as inputs, does not depend on grid types, and solves the contradiction between the efficiency and the precision of a traditional CFD (Computational Fluid Dynamics) method.

Description

Technical Field

[0001] The present invention relates to the field of aircraft design and deep learning technology, and in particular to a method for predicting the physical field of an aircraft surface based on a language model. Background Art

[0002] For aircraft design, the lift and drag coefficients come from the integral of surface pressure and friction resistance; the radar cross-sectional area is obtained by integrating the surface current; and the surface friction lines can intuitively reflect the shock waves, flow separation and other phenomena on the aircraft surface. Accurate prediction of surface physical fields is a key link in the design of high-performance aircraft. At present, in the field of aircraft design, aircraft design methods based on computational fluid dynamics (CFD) are widely used. With the continuous improvement of aircraft performance, the aircraft envelope is getting larger and larger. The prediction accuracy of CFD methods based on solving the Reynolds-averaged NS (RANS) equation at the envelope boundary has dropped significantly, but the high-precision detached eddy simulation (DES) and large eddy simulation (LES) are difficult to be widely used due to their huge computational workload. In terms of aircraft optimization design, the optimization method based on the proxy model needs to build a sample library, and the cost of repeatedly calling CFD calculations is unbearable in high-dimensional design problems; the optimization method based on the adjoint equation mainly relies on the solution of the main control equation of the flow field and the adjoint equation. It can be seen that the contradiction between the efficiency and accuracy of the traditional CFD method has become a key factor restricting the design of high-performance aircraft.

[0003] In recent years, the development of artificial intelligence methods is in the ascendant, and the "universal approximation" ability of neural networks provides a new idea for solving flow fields. At present, researchers have conducted a lot of research on the use of convolutional neural networks (CNN) for flow field prediction, and have achieved good results in the prediction of flow fields of airfoils and simple wings. However, the CNN-based method requires the projection of geometric shapes into Cartesian grids, or the conversion of physical grids into elliptical computational grids, which makes this type of method mainly applicable to structured grids, which restricts its development to complex shapes. On the other hand, current research mainly focuses on two-dimensional airfoils or simple wings, and there are still few physical field prediction methods for complex three-dimensional configurations. The emergence of the Transformer architecture has brought new developments to natural language processing and computer vision. Based on this architecture, the number of neural network parameters has increased significantly, and large language models represented by Bert, GPT, Llama, etc. have developed rapidly. At the same time, various downstream tasks based on large language models have also emerged in an endless stream. The present invention utilizes the powerful feature extraction capability of the large language model, based on the Bert model, and combines the Cartesian solver to propose a language model-based aircraft surface physical field prediction method that is applicable to any grid type, replacing the time-consuming CFD solution and providing a new means for aircraft design. Summary of the invention

[0004] In order to solve the contradiction between the efficiency and accuracy of the CFD method used in the above-mentioned prior art, which depends on the grid type, the present invention provides a method for predicting the physical field of the aircraft surface based on a language model.

[0005] To implement the above technical solution, the present invention provides a method for predicting the physical field of an aircraft surface based on a language model, comprising the steps of:

[0006] S1: Generate the initial shape of the aircraft, and parameterize the shape through a parameterization method with physical meaning to determine the deformation range of the parameters after the parameters;

[0007] S2: Generate a new shape based on the parameter deformation range determined in step S1, and generate a surface mesh of the aircraft shape

[0008] Grids and spatial grids;

[0009] S3: Perform CFD solution on the new shape generated in step S2;

[0010] S4: The aircraft shape and flight state are used as features, and the surface physical field distribution is used as labels. They are encoded and input into the Bert model for training. After logistic regression, a qualified model is obtained.

[0011] S5: Based on the trained model, the aircraft shape coordinate points and flight status are used as input to obtain the physical field distribution on the aircraft surface.

[0012] Furthermore, in step S4, a parameterized method with physical meaning is used to construct a three-dimensional sample data set, and a Bert language model and a regression model are used to obtain a prediction model of the physical field of the aircraft surface.

[0013] Furthermore, in step S3, the deformed shape is solved using a Cartesian grid; the data set includes inviscid and viscous results.

[0014] Furthermore, in step S4, the grid point coordinates and flight status are used as features, and the corresponding surface physical fields are used as labels. After being tokenized, they are input into the Bert model training, and then the prediction model training is completed through the logistic regression model.

[0015] In summary, the present invention has the following beneficial effects compared with the prior art:

[0016] The present invention utilizes powerful feature extraction capabilities and relies on language models to construct a prediction model for the physical field of the aircraft surface. The constructed model directly uses the aircraft coordinate points and flight status as input, does not rely on the grid type, and achieves second-level prediction of the surface physical field of any aircraft configuration at sub-, trans- and supersonic speeds, solving the contradiction between efficiency and accuracy of traditional CFD methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0018] Figure 1 It is a flow chart for realizing the present invention;

[0019] Figure 2 The initial shape defined in the present invention;

[0020] Figure 3 The present invention adopts parameterized and deformed shapes;

[0021] Figure 4 It is the training and inference flow chart of the present invention;

[0022] Figure 5 This is a schematic diagram of the model used in this paper;

[0023] Figure 6 Comparison between the surface pressure cloud map predicted by the present invention and the CFD solution;

[0024] Figure 7 Comparison of cross-sectional pressure distribution predicted by the present invention and obtained by CFD calculation. DETAILED DESCRIPTION

[0025] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0026] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form can also include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0027] See also Figures 1 to 7 As shown, the present invention provides a method for predicting the physical field of an aircraft surface based on a language model, comprising the steps of:

[0028] S1: Generate the initial shape of the aircraft, and parameterize the shape through a parameterization method with physical meaning to determine the deformation range of the parameters after the parameters;

[0029] S2: Generate a new shape based on the parameter deformation range determined in step S1, and generate a surface mesh of the aircraft shape

[0030] Grids and spatial grids;

[0031] S3: Perform CFD solution on the new shape generated in step S2;

[0032] S4: The aircraft shape and flight state are used as features, and the surface physical field distribution is used as labels. They are encoded and input into the Bert model for training. After logistic regression, a qualified model is obtained.

[0033] S5: Based on the trained model, the aircraft shape coordinate points and flight status are used as input to obtain the physical field distribution on the aircraft surface.

[0034] like Figure 1 As shown, firstly, an initial shape is generated according to the requirements, and the variables to be parameterized and the range of parameter variation are determined through a parametric method. Figure 2 The initial configuration with parameterization that has been deformed is given. In this example, the present invention selects the root chord length (Root_chord), the wing sweep angle (Sweep angle), the Kink angle (Angle 1) and the wing tip angle (Span2) as parameterized variables. Then, the deformed shape and the mesh for CFD evaluation are generated according to the variation range, such as Figure 3 As shown. Then, CFD calculations are performed on the shape to obtain surface pressure distribution, velocity distribution, etc. In this example, taking the flying wing layout as an example, 10,000 different shapes are generated to evaluate the flow field distribution under different Mach numbers, angles of attack, and sideslip angles, totaling 120,000 samples. Then, the sample shape coordinate points and flight status are used as features, and the surface physical field is used as a label. These quantities are encoded and input into the Bert model for training, as shown in Figure 4 shown. Figure 5 The Bert model is given, which uses the Transformer encoding structure. After the model is trained, the aircraft shape and flight status are input to obtain the physical field distribution on the aircraft surface.

[0035] The comparison between the model prediction and CFD evaluation is shown in the figure below. Figure 6 and Figure 7 shown. Figure 6 The pressure cloud diagram of a certain shape surface is compared with the CFD solution when the Mach number is 0.85, the angle of attack is 3°, and the sideslip angle is 0°. Figure 7 A comparison between the predicted pressure distribution of a certain section and that obtained by CFD solution is given, and it can be seen that the two are in good agreement.

[0036] As a preferred embodiment, in step S4, a three-dimensional sample data set is constructed by a parameterized method with physical meaning, and a Bert language model and a regression model are used to obtain a prediction model for the physical field of the aircraft surface, which can achieve second-level prediction of the physical field of the surface of any configuration.

[0037] As a preferred method, in step S3, the parameterized physical shape can realize arbitrary deformation of global parameters (sweep angle, root-to-tip ratio) and component parameters (rack position, rack shape, wing angle) without being restricted by the grid. The deformed shape is solved by an efficient and high-precision Cartesian grid, and the data set contains inviscid and viscous results.

[0038] As a preferred embodiment, in step S4, the grid point coordinates and flight status are used as features, and the corresponding surface physical fields are used as labels. After being tokenized, they are input into the Bert model training, and then the prediction model training is completed through the logistic regression model.

[0039] After inputting the shape coordinate points and flight status, the physical field distribution on the aircraft surface can be obtained. It is not restricted by the grid type and the topological structure properties, and has cross-topology, variable state, and second-level prediction capabilities.

[0040] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting the physical field of an aircraft surface based on a language model, characterized in that: Includes steps: S1: Generate the initial shape of the aircraft and parameterize the shape using a parameterization method with physical meaning. To determine the range of parameter deformation after the parameters; S2: Generate a new shape according to the parameter deformation range determined in step S1, and generate a surface mesh and a space mesh of the aircraft shape; S3: Perform CFD solution on the new shape generated in step S2; S4: Encode the aircraft shape and flight state as features and the surface physical field distribution as labels. Input the Bert model for training, and obtain a qualified model after logistic regression; S5: Based on the trained model, the aircraft shape coordinate points and flight status are used as input to obtain the physical field distribution on the aircraft surface.

2. The method for predicting the physical field of an aircraft surface based on a language model according to claim 1, characterized in that: In step S4, a three-dimensional sample data set is constructed using a parameterized method with physical meaning, and a prediction model of the aircraft surface physical field is obtained using a Bert language model and a regression model.

3. The method for predicting the physical field of an aircraft surface based on a language model according to claim 1, characterized in that: In step S3, the deformed shape is solved using a Cartesian grid; the data set contains inviscid and viscous results.

4. The method for predicting the physical field of an aircraft surface based on a language model according to claim 1, characterized in that: In step S4, the grid point coordinates and flight status are used as features, and the corresponding surface physical fields are used as labels. After being tokenized, they are input into the Bert model training, and then the prediction model training is completed through the logistic regression model.

Images