A mesh generation method and system

Abstract

The application relates to the field of mesh generation, and discloses a mesh generation method and system, the method comprising the following steps: acquiring an initial mesh information file; the initial mesh information comprising quadrilateral mesh and position information between the quadrilateral mesh and upper, lower, left and right nodes; constructing a mesh generation model, the mesh generation model comprising a first long short-term memory network and a deep neural network connected in sequence; inputting the initial mesh information file into the trained mesh generation model, performing semantic information embedding on the initial mesh information file by using the first long short-term memory network, and generating a mesh parameter matrix; and transmitting the mesh parameter matrix to the deep neural network to learn mesh parameter information, and generating a mesh division information file. The application improves the precision and efficiency of mesh generation.

Description

Technical Field

[0001] This invention relates to the field of mesh generation, and more specifically, to a mesh generation method and system. Background Technology

[0002] In modern CFD (Computational Fluid Dynamics) applications, automatically generating high-quality, complex mesh configurations remains a significant challenge.

[0003] Currently, the main method for automatically generating component structural meshes involves designers using general CAD modeling software to create 3D models, then dividing the component structural model into meshes, and finally analyzing the output results through iterative calculations. However, traditional mesh generation methods have limited effectiveness, require a large amount of manual operation and iterative calculations, resulting in low mesh generation efficiency and low accuracy. Summary of the Invention

[0004] To overcome the shortcomings of low efficiency and accuracy in mesh generation in existing technologies, this invention provides a mesh generation method and system.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] In the first aspect, the present invention proposes a mesh generation method, comprising the following steps:

[0007] S1: Obtain the initial mesh information file. The initial mesh information includes the quadrilateral mesh and its positional information relative to the top, bottom, left, and right nodes.

[0008] S2: Construct a grid generation model, which includes a first long short-term memory network and a deep neural network connected in sequence.

[0009] S3: Input the initial mesh information file into the trained mesh generation model, and use the first long short-term memory network to embed semantic information into the initial mesh information file to generate a mesh parameter matrix. Transmit the mesh parameter matrix to a deep neural network to learn the mesh parameter information and generate a mesh partitioning information file.

[0010] As a preferred technical solution, the first long short-term memory network includes N layers of second long short-term memory networks connected in sequence; the second long short-term memory network includes an upper long short-term memory network, a lower long short-term memory network, a left long short-term memory network, a right long short-term memory network, and a middle long short-term memory network; the middle long short-term memory network is provided with 4 input terminals, and the 4 input terminals of the middle long short-term memory network are respectively connected to the output terminals of the upper long short-term memory network, the lower long short-term memory network, the left long short-term memory network, and the right long short-term memory network.

[0011] As a preferred technical solution, in S4, the initial grid information file is semantically embedded using a first long short-term memory network to generate a grid parameter matrix Q, the specific expression of which is shown below:

[0012]

[0013]

[0014] c t =f⊙c t-1 +g⊙i

[0015] h t =o⊙tanh(c t )

[0016]

[0017] Q = [R1,...,R] k ]

[0018] Where f represents the output of the forget gate in the second long short-term memory network, g represents the new information added to the memory unit of the second long short-term memory network, and c t h is the memory unit at the current moment. t R represents the output of the hidden state in the second long short-term memory network at the current time step. k Let represent the learned parameters at grid position k, and σ(·) represent the sigmoid function. and These represent the hidden state outputs of the upper, lower, left, and right long short-term memory networks at time t-1, respectively. and Let W represent the hidden state weights of the upper, lower, left, and right long short-term memory networks at time t-1, respectively. Let X be the original input vector of the first long short-term memory network, and W be the hidden state weights of the upper, lower, left, and right long short-term memory networks. Xdenoted by , where b is the weight of the original input vector, i is the bias, and o is the input of the second long short-term memory network. This is the intermediate vector output by the intermediate layer of the Long Short-Term Memory network. and These are the intermediate vectors output by the upper, lower, left, and right long short-term memory networks, respectively.

[0019] As a preferred technical solution, after S2 and before S3, the method further includes: constructing a dataset using the initial grid information file, and dividing the dataset into a training set, a validation set, and a test set; training and tuning the grid generation model using the training set and the validation set until the loss function of the grid generation model converges, thereby obtaining a trained grid generation model;

[0020] The expression for the loss function of the mesh generation model is as follows:

[0021]

[0022]

[0023] in, For the prediction result y of the j-th grid j Compared with grid label results q j Error loss, Let J be the regularization function, Ω be the regularization term, and J be the regularization term. j H is the Jacobian ratio of the j-th grid. j Let T be the aspect ratio of the j-th grid. j Let the taper of the j-th grid be denoted as 'j'.

[0024]

[0025]

[0026] f(x i ) is the network prediction function, γ is the penalty function, λ is the regularization rate, K represents the number of grid nodes, and w represents the weight of the grid node.

[0027] As a preferred technical solution, the Jacobian ratio J of the j-th grid is... j The expression is as follows:

[0028]

[0029] Where x is the x-coordinate of the mesh after element transformation, y is the y-coordinate of the mesh after element transformation, ξ is the x-coordinate of the mesh before element transformation, and η is the x-coordinate of the mesh before element transformation.

[0030] The aspect ratio H of the j-th grid j The expression is as follows:

[0031]

[0032] Where h1 is the width of the grid and h2 is the height of the grid;

[0033] The grid taper T of the j-th grid j The expression is as follows:

[0034] T j =∑|(A j -A avg ) / A avg | max j = 1, 2, 3, 4

[0035] A avg = (A1+A2+A3+A4) / 4

[0036] Where A1 is the area of ​​the upper half of the line connecting the right diagonals in the grid, A2 is the area of ​​the lower half of the line connecting the right diagonals in the grid, A3 is the area of ​​the upper half of the line connecting the left diagonals in the grid, and A4 is the area of ​​the lower half of the line connecting the left diagonals in the grid. avg Let A1, A2, A3, and A4 be the weighted average area.

[0037] As a preferred technical solution, in S4, the mesh parameter matrix is ​​transmitted to a deep neural network to learn the mesh parameter information and generate a mesh partitioning information file, the expression of which is as follows:

[0038]

[0039] in, The output is a parameter matrix containing grid information. The parameter matrix Q is input to the hidden layer of the deep neural network. The initial grid information matrix is ​​used as input to the hidden layer of the deep neural network. This is the bias matrix of the hidden layer of the deep neural network.

[0040] As a preferred technical solution, after obtaining the initial mesh information file, the method further includes: performing a noise removal operation on the initial mesh information file, specifically: removing information other than mesh node type and mesh node storage information.

[0041] As a preferred technical solution, the semantic information includes stress characteristic layer information, strain characteristic layer information, and displacement characteristic layer information.

[0042] As a preferred technical solution, in S1, an initial mesh information file is obtained from the turotial in OpenFOAM using a filter. The initial mesh information file includes mesh information files of all case types in the turotial.

[0043] Secondly, the present invention also proposes a mesh generation system, comprising:

[0044] The acquisition module is used to acquire the initial mesh information file. The initial mesh information includes the quadrilateral mesh and its positional information relative to the top, bottom, left, and right nodes.

[0045] A building module is used to construct a mesh generation model, which includes a first long short-term memory network and a deep neural network connected in sequence.

[0046] The generation module is used to input the initial mesh information file into the trained mesh generation model, embed semantic information into the initial mesh information file using a first long short-term memory network, and generate a mesh parameter matrix. The mesh parameter matrix is ​​then transmitted to a deep neural network to learn the mesh parameter information and generate a mesh partitioning information file.

[0047] Compared with the prior art, the beneficial effects of the technical solution of the present invention are as follows: The present invention obtains the information between the quadrilateral mesh and its upper, lower, left and right nodes as input, and uses the first long short-term memory network to embed the semantic information of the initial mesh information file, thereby simulating the mutual information relationship between the simulated quadrilateral mesh and its upper, lower, left and right nodes. This makes the generated mesh more accurate and precise, greatly reduces the amount of computation, shortens the computation time, realizes the automatic generation of mesh, eliminates the operation of manual mesh division, and improves the efficiency of mesh generation. Attached Figure Description

[0048] Figure 1 This is a flowchart of the mesh generation method in the embodiments of this application.

[0049] Figure 2 This is a schematic diagram of the structure of the first long short-term memory network in the embodiments of this application.

[0050] Figure 3 This is a schematic diagram illustrating the principle of semantic information embedding in the first long short-term memory network in this embodiment of the application.

[0051] Figure 4 This is an architecture diagram of the grid generation system in the embodiments of this application. Detailed Implementation

[0052] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0054] Example 1

[0055] Please see Figure 1 This embodiment proposes a mesh generation method, including the following steps:

[0056] S1: Obtain the initial mesh information file. The initial mesh information includes the quadrilateral mesh and its positional information relative to the top, bottom, left, and right nodes.

[0057] S2: Construct a grid generation model, which includes a first long short-term memory network and a deep neural network connected in sequence.

[0058] S3: Input the initial mesh information file into the trained mesh generation model, and use the first long short-term memory network to embed semantic information into the initial mesh information file to generate a mesh parameter matrix. Transmit the mesh parameter matrix to a deep neural network to learn the mesh parameter information and generate a mesh partitioning information file.

[0059] The mesh generation method proposed in this embodiment obtains the information between the quadrilateral mesh and its upper, lower, left, and right nodes as input, and uses the first long short-term memory network to embed the semantic information of the initial mesh information file. It simulates the mutual information relationship between the simulated quadrilateral mesh and its upper, lower, left, and right nodes, making the generated mesh more accurate and precise, greatly reducing the amount of computation, shortening the computation time, realizing automated mesh generation, eliminating the operation of manual mesh division, and improving mesh generation efficiency.

[0060] Example 2

[0061] This embodiment improves upon the mesh generation method proposed in Embodiment 1, including:

[0062] S1: Obtain the initial mesh information file. The initial mesh information includes the quadrilateral mesh and its positional information relative to the top, bottom, left, and right nodes.

[0063] In this embodiment, a filter is used to read the initial mesh information file from the .cwenjian file storing mesh information in the `turotial` repository of OpenFOAM. This initial mesh information file includes mesh information files for all case types within `turotial`. Then, a noise removal operation is performed on the initial mesh information file, specifically removing information other than mesh node types and stored mesh node information.

[0064] S2: Construct a grid generation model, which includes a first long short-term memory network and a deep neural network connected in sequence.

[0065] like Figure 2 As shown, in this embodiment, the first long short-term memory network includes N layers of second long short-term memory networks connected in sequence; the second long short-term memory network includes an upper long short-term memory network, a lower long short-term memory network, a left long short-term memory network, a right long short-term memory network, and a middle long short-term memory network; the middle long short-term memory network is provided with 4 input terminals, and the 4 input terminals of the middle long short-term memory network are respectively connected to the output terminals of the upper long short-term memory network, the lower long short-term memory network, the left long short-term memory network, and the right long short-term memory network.

[0066] Since each quadrilateral grid has a neighboring grid cell on its top, bottom, left, and right sides, this embodiment also considers the influence of neighboring grid cells on the intermediate grid cells, including the nodes that make up the grid. Therefore, a second long short-term memory (LSTM) network is designed with upper, lower, left, and right LSM layers to process information from different directions for each grid cell. In the second LSM network, the upper, lower, left, and right LSM layers each input an initial grid information file for computation, and the upper, lower, left, and right LSM layers input the computation results to the intermediate LSM layer.

[0067] S3: Train the mesh generation model until the loss function of the mesh generation model converges to obtain the trained mesh generation model.

[0068] In this embodiment, the acquired initial grid information file is used as the dataset, and the dataset is divided into a training set, a validation set, and a test set in JSON format in a ratio of 6:1:3.

[0069] In this embodiment, the expression for the loss function of the mesh generation model is as follows:

[0070]

[0071]

[0072] in, For the prediction result y of the j-th grid j Compared with grid label results q j Error loss, Let J be the regularization function, Ω be the regularization term, and J be the regularization term. j H is the Jacobian ratio of the j-th grid. j Let T be the aspect ratio of the j-th grid. j Let be the taper of the j-th grid.

[0073]

[0074]

[0075] f(x i ) is the network prediction function, γ is the penalty function, λ is the regularization rate, K represents the number of grid nodes, and w represents the weight of the grid node.

[0076] In this embodiment, the Jacobian ratio J of the j-th grid is... j The expression is as follows:

[0077]

[0078] Where x is the x-coordinate of the mesh after element transformation, y is the y-coordinate of the mesh after element transformation, ξ is the x-coordinate of the mesh before element transformation, and η is the x-coordinate of the mesh before element transformation.

[0079] The Jacobian ratio is the ratio of a unit infinitesimal element after coordinate system transformation. Mathematically, it is expressed as the determinant of the Jacobian matrix calculated at each integration point of the element (the actual shape of the ideal element in free coordinates mapped to global coordinates).

[0080] In this embodiment, the aspect ratio H of the j-th grid is... j The expression is as follows:

[0081]

[0082] Where h1 is the width of the grid and h2 is the height of the grid.

[0083] The aspect ratio of a quadrilateral grid cell is the ratio of the lengths of the lines connecting the midpoints of opposite sides. The ideal aspect ratio is 1, and the acceptable value is (0, 8).

[0084] In this embodiment, the grid taper T of the j-th grid j The expression is as follows:

[0085]

[0086] A avg = (A1+A2+A3+A4) / 4

[0087] Where A1 is the area of ​​the upper half of the line connecting the right diagonals in the grid, A2 is the area of ​​the lower half of the line connecting the right diagonals in the grid, A3 is the area of ​​the upper half of the line connecting the left diagonals in the grid, and A4 is the area of ​​the lower half of the line connecting the left diagonals in the grid. avg Let A1, A2, A3, and A4 be the weighted average area.

[0088] The ideal value for the grid taper is 0, and the acceptable value is [0, 0.5].

[0089] In the specific implementation process, quadrilateral meshes are mainly used. The quality of mesh generation is checked by a loss function. It is assumed that there is a local or global optimal solution in the generation process. The value of the loss function is continuously reduced by gradient descent to find the objectively existing "optimal solution". The existing mesh division is re-divided without changing the total number of mesh nodes, so as to achieve convergence in the key areas of mechanical performance, while ensuring that the amount of computation does not increase again.

[0090] S4: Input the initial mesh information file into the trained mesh generation model, and use the first long short-term memory network to embed semantic information into the initial mesh information file to generate a mesh parameter matrix. Transmit the mesh parameter matrix to a deep neural network to learn the mesh parameter information and generate a mesh partitioning information file.

[0091] like Figure 2 and Figure 3 As shown in this embodiment, the input vector X of the initial grid information file is first defined, and its expression is as follows:

[0092] X = [v x ,v y ,h1,h2,h3,h4,x cells ,y cells ,s x ,s y ]

[0093] Among them, v x and v y h1, h2, h3, and h4 are the x and y coordinates of the four vertices of the quadrilateral grid, respectively; h1, h2, h3, and h4 are the information of the four vertices of the quadrilateral grid; x cells and y cells s represents the number of grid cells for the geometry in the X and Y directions, respectively. x and s y These represent the degree of mesh stretching or shrinking in the X and Y directions, respectively.

[0094] The initial grid information file is semantically embedded using a first long short-term memory network to generate a grid parameter matrix Q, the expression of which is shown below:

[0095] The output f of the forget gate in the second long short-term memory network is:

[0096]

[0097] The new information g added to the memory units of the second long short-term memory network is:

[0098]

[0099] Input i is:

[0100]

[0101] The output o is:

[0102]

[0103] Current memory unit c t for:

[0104] c t =f⊙c t-1 +g⊙i

[0105] The hidden state output h at the current moment in the second long short-term memory network. t for:

[0106] h t =o⊙tanh(c t )

[0107] Because the first Long Short-Term Memory network uses vectors from every layer, an N-layer network will generate 4N+1 representations (four vectors in each direction for each layer, with an initial input vector, for a total of 5). Therefore, the learned parameters R at grid position k are... k for:

[0108]

[0109] Define the mesh parameter matrix Q as:

[0110] Q = [R1,...,R] k ]

[0111] in, and These represent the hidden state outputs of the upper, lower, left, and right long short-term memory networks at time t-1, respectively. and Let W represent the hidden state weights of the upper, lower, left, and right long short-term memory networks at time t-1, respectively. Let X be the original input vector of the first long short-term memory network, and W be the hidden state weights of the upper, lower, left, and right long short-term memory networks. X denoted by , where b is the weight of the original input vector, i represents the input to the second long short-term memory network, and o represents the output of the second long short-term memory network. This is the intermediate vector output by the intermediate layer of the Long Short-Term Memory network. and These are the intermediate vectors output by the upper, lower, left, and right long short-term memory networks, respectively, and σ(·) represents the sigmoid function.

[0112] In this embodiment, the mesh parameter matrix is ​​transmitted to a deep neural network to learn the mesh parameter information and generate a mesh partitioning information file, the expression of which is as follows:

[0113]

[0114] in, The output is a parameter matrix containing grid information. The parameter matrix Q is input to the hidden layer of the deep neural network. The initial grid information matrix is ​​used as input to the hidden layer of the deep neural network. This is the bias matrix of the hidden layer of the deep neural network.

[0115] In this embodiment, the semantic information includes stress feature layer information, strain feature layer information, and displacement feature layer information. The noise-removed file is embedded with these three layers of semantic information using a first long short-term memory network. Here, the mesh parameter matrix Q incorporates the inherent semantics of the mesh, serving a pre-training function and avoiding the inability to find the optimal solution for the adaptive mesh later due to random initialization of the parameter matrix in the deep neural network.

[0116] The mesh generation method proposed in this embodiment obtains information about the quadrilateral mesh and its nodes (up, down, left, and right) as input. It uses a first long short-term memory network to embed semantic information from the initial mesh information file, simulating the mutual information relationships between the simulated quadrilateral mesh and its nodes. This results in a more accurate and precise generated mesh, significantly reducing computational load and time, achieving automated mesh generation, eliminating manual mesh division, and improving mesh generation efficiency. A loss function is used to verify the mesh generation quality. Assuming the generation process has a local or global optimum, the process is continuously optimized using the loss function. A secondary mesh division is performed on the existing mesh without changing the total number of mesh nodes, achieving convergence in key mechanical performance regions while ensuring that the computational load does not increase further, thus improving mesh generation efficiency.

[0117] Example 3

[0118] This embodiment proposes a mesh generation system, including:

[0119] The acquisition module is used to acquire the initial mesh information file. The initial mesh information includes the quadrilateral mesh and its positional information relative to the top, bottom, left, and right nodes.

[0120] A building module is used to construct a mesh generation model, which includes a first long short-term memory network and a deep neural network connected in sequence.

[0121] The generation module is used to input the initial mesh information file into the trained mesh generation model, embed semantic information into the initial mesh information file using a first long short-term memory network, and generate a mesh parameter matrix. The mesh parameter matrix is ​​then transmitted to a deep neural network to learn the mesh parameter information and generate a mesh partitioning information file.

[0122] The mesh generation system proposed in this embodiment obtains the information between the quadrilateral mesh and its nodes (up, down, left, and right) as input, and uses a first long short-term memory network to embed the semantic information of the initial mesh information file. This simulates the mutual information relationship between the quadrilateral mesh and its nodes (up, down, left, and right), making the generated mesh more accurate and precise. It also greatly reduces the amount of computation, shortens the computation time, realizes automated mesh generation, eliminates the need for manual mesh division, and improves mesh generation efficiency.

[0123] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0124] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A mesh generation method, characterized in that, include: S1: Obtain the initial mesh information file; the initial mesh information includes the quadrilateral mesh and its position information relative to the top, bottom, left, and right nodes; S2: Construct a mesh generation model, which includes a first long short-term memory network and a deep neural network connected sequentially; the first long short-term memory network includes... N The second long short-term memory network is connected in layers sequentially; the second long short-term memory network includes an upper long short-term memory network, a lower long short-term memory network, a left long short-term memory network, a right long short-term memory network, and a middle long short-term memory network; the middle long short-term memory network is provided with 4 input terminals, which are respectively connected to the output terminals of the upper long short-term memory network, the lower long short-term memory network, the left long short-term memory network, and the right long short-term memory network; A dataset is constructed using the initial grid information file, and the dataset is divided into a training set, a validation set, and a test set. The grid generation model is trained and its parameters are tuned using the training set and the validation set until the loss function of the grid generation model converges, thus obtaining a trained grid generation model. The expression for the loss function of the mesh generation model is as follows: in, For the first j Grid prediction results y j Results with grid labels q j Error loss, For regularization functions, For regularization terms, For the first j Jacobian ratio of each grid For the first j The aspect ratio of each grid cell. For the first j The grid taper of each grid; f ( y i ) is the network prediction function. For the penalty function, The regularization rate is 1. K Represents the number of grid nodes w This represents the weight corresponding to the grid node; No. j Jacobian ratio of each grid The expression is as follows: in, q The x-coordinate of the mesh after element transformation. p ξ is the ordinate of the mesh after element transformation, ξ is the abscissa of the mesh before element transformation, and η is the ordinate of the mesh before element transformation. No. j Aspect Ratio of Each Grid The expression is as follows: in, h 1 represents the width of the grid. h 2 represents the height of the grid; No. j The taper of each grid The expression is as follows: in, Let be the area of ​​the upper half of the line connecting the right diagonals in the grid. This represents the area of ​​the lower half of the line connecting the right diagonal corners of the grid. This represents the area of ​​the upper half of the line connecting the left diagonal corners of the grid. This represents the area of ​​the lower half of the line connecting the left diagonal corners of the grid. for , , and The weighted average area; S3: Input the initial grid information file into the trained grid generation model, use the first long short-term memory network to embed semantic information into the initial grid information file, and generate a grid parameter matrix; transmit the grid parameter matrix to the deep neural network to learn the grid parameter information and generate a grid partitioning information file.

2. The mesh generation method according to claim 1, characterized in that, In S4, the initial grid information file is semantically embedded using the first long short-term memory network to generate a grid parameter matrix. Q The specific expression is as follows: ⊙ ⊙ i ⊙ in, f This represents the output of the forget gate in the second long short-term memory network. g This indicates new information being added to the memory units of the second long short-term memory network. For the memory unit at the current moment, This is the output of the hidden state in the second long short-term memory network at the current moment. R k Indicates the grid position as k The parameters learned from the mesh. This represents the sigmoid function. , , and These are the upper long short-term memory network, the lower long short-term memory network, the left long short-term memory network, and the right long short-term memory network, respectively. t Output of the hidden state at time -1 , , and These are the upper long short-term memory network, the lower long short-term memory network, the left long short-term memory network, and the right long short-term memory network, respectively. t Hidden state weights at time -1 This is the original input vector of the first Long Short-Term Memory network. The weights of the original input vector. For bias, i As input to the second long short-term memory network, o This is the output of the second long short-term memory network. This is the intermediate vector output by the intermediate layer of the Long Short-Term Memory network. , , and These are the intermediate vectors output by the upper, lower, left, and right long short-term memory networks, respectively.

3. The mesh generation method according to claim 2, characterized in that, In step S4, the mesh parameter matrix is ​​transmitted to the deep neural network to learn the mesh parameter information and generate a mesh partitioning information file, the expression of which is as follows: in, The output is a parameter matrix containing grid information. The parameter matrix of the hidden layer of the input deep neural network Q , The initial grid information matrix is ​​used as input to the hidden layer of the deep neural network. This is the bias matrix of the hidden layer of the deep neural network.

4. The mesh generation method according to claim 1, characterized in that, After obtaining the initial mesh information file, the method further includes: performing a noise removal operation on the initial mesh information file, specifically: removing information other than mesh node type and mesh node storage information.

5. The mesh generation method according to claim 1, characterized in that, The semantic information includes stress characteristic layer information, strain characteristic layer information, and displacement characteristic layer information.

6. The mesh generation method according to claim 1, characterized in that, In S1, the initial mesh information file is obtained from the turotial filter in OpenFOAM using a filter. The initial mesh information file includes mesh information files for all case types in turotial.

7. A mesh generation system, applied to the mesh generation method as described in any one of claims 1 to 6, characterized in that, include: The acquisition module is used to acquire the initial grid information file; The initial mesh information includes the quadrilateral mesh and its positional information relative to the top, bottom, left, and right nodes; A building module is used to build a mesh generation model, which includes a first long short-term memory network and a deep neural network connected in sequence. The generation module is used to input the initial grid information file into the trained grid generation model, embed semantic information into the initial grid information file using a first long short-term memory network, generate a grid parameter matrix, and transmit the grid parameter matrix to a deep neural network to learn grid parameter information and generate a grid partitioning information file.

Images