Deep learning-based T-shaped joint guidance grid self-adaption method

Abstract

The invention provides a deep learning-based T-shaped joint guidance grid self-adaption method, which comprises the following steps of S1, generating a T-shaped joint model of random geometric parameters, and defining a geometric shape through parametric design; s2, predicting the size distribution of grid units by using a deep learning model, and guiding the generation of a non-uniform grid; s3, integrating a finite element solver to carry out linear elastic mechanical calculation, and automatically setting material parameters and boundary conditions; and S4, outputting strain energy, stress distribution and an adaptive grid result through a visual tool. According to the method, the optimal non-uniform grid unit size distribution is obtained through the deep learning model, then the more accurate non-uniform grid is obtained, the deep learning model predicts the optimal grid distribution, the grid adjustment process based on posterior error estimation in a traditional method is avoided, and the accuracy and efficiency can be remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computational mechanics research, and in particular to a method for T-joint guided grid adaptation based on deep learning. Background Art

[0002] T-joints are numerous in ship structures and are often the weakest link in the overall structure. Workmanship defects and stress concentration can lead to failure and damage in T-joints. Traditional experimental methods can verify the strength of T-joints, but these tests are destructive, costly, and time-consuming. With the rapid development of scientific computing technology, finite element simulation has become an effective alternative to physical experiments. Finite element simulation is low-cost, highly efficient, and can intuitively demonstrate the stress distribution in a structure, providing guidance for design.

[0003] Currently, finite element simulation of T-joints is a tedious process. First, engineers must create a geometric model in 3D drawing software. Then, they must import the geometric model into the finite element software, create a mesh, set material parameters and boundary conditions, and perform calculations. Finally, the results must be visualized using the finite element software's own or independent post-processing software. This entire simulation process is not only time-consuming but also requires significant skill and experience from the engineer.

[0004] One of the most critical technologies in finite element simulation is mesh generation. Mesh generation divides a continuous area into a finite number of units to achieve spatial discretization. Unstructured meshes are widely used because they can be automatically generated and encrypted. Generally speaking, the smaller the size of the mesh unit, the higher the accuracy of the calculation. However, the time cost of the calculation increases superlinearly with the increase in the number of mesh units. In order to break the constraint between calculation accuracy and speed, non-uniform meshes are widely used. Adaptive non-uniform meshes are encrypted at key locations and sparse at other locations. The difficulty lies in determining where to encrypt to achieve the best computational efficiency. The usual method is to encrypt at locations with high computational errors to achieve an equal distribution of errors. In order to obtain the error, it is necessary to perform error analysis based on the finite element solution, which is also called a posteriori error estimation. In order to generate a non-uniform mesh, a finite element solution must be performed first, which is very costly.

[0005] Publication number CN 114239328 A discloses a method for predicting residual stress and deformation in T-joint welding, comprising the following steps: S1: creating a three-dimensional model of a steel plate T-joint weld in ABAQUS, assigning material properties, meshing the three-dimensional model, defining the welding finite element simulation direction, and setting the welding step, cooling step, and boundary conditions; S2: programming a double-wire MAG welding program language using the Dlux program in Visual Studio software; S3: submitting the coordinate information of the meshed three-dimensional T-joint weld model in ABAQUS, as well as the set material properties, welding step, cooling step, and boundary conditions, to the Visual Studio software, activating the Visual Studio software and running the Dflux language controlled by the Inter Fortran program to perform welding finite element analysis. Prior art methods are not only time-consuming and difficult, but also lack accuracy in prediction. Summary of the Invention

[0006] In view of this, the present invention aims to propose a method for guiding grid adaptation of T-joints based on deep learning, so as to solve the problems that the existing methods are not only time-consuming and difficult, but also have inaccurate prediction methods.

[0007] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0008] A method for guiding mesh adaptation of a T-joint based on deep learning, comprising the following steps:

[0009] S1, generate a T-joint model with random geometric parameters and define the geometric shape through parametric design;

[0010] S2. Use deep learning models to predict grid cell size distribution and guide non-uniform grid generation;

[0011] S3, calling the finite element solver to perform simulation calculation;

[0012] S4. Output strain energy, stress distribution and adaptive mesh results through visualization tools.

[0013] In this setup, the random geometric model generated through parametric design covers a variety of sizes and shape combinations of T-joints. The deep learning model further improves the accuracy of finite element calculations by training these diverse samples. Furthermore, by learning the geometric characteristics and stress distribution patterns of a large number of T-joint samples, the deep learning model obtains the optimal non-uniform mesh unit size distribution, thereby obtaining a more accurate non-uniform mesh. By predicting the optimal mesh distribution, the deep learning model avoids the mesh adjustment process based on a posteriori error estimation in traditional methods, achieving greater efficiency. The deep learning-driven mesh adaptation strategy achieves significant improvements in prediction accuracy, simplicity, and computational efficiency, while providing an efficient and versatile solution for the simulation analysis of complex welded structures.

[0014] Furthermore, the parameterized design includes the upper plate thickness l1, the lower plate thickness l2, the weld width 0.5l1 and the weld lower angle tangent value tan, and the parameter range is: 0.4 m <l1<2.0 m,0.2 m<l2<1.0 m,0.5<tan<2.0。

[0015] Furthermore, both ends of the lower plate of the T-shaped joint are fixedly constrained, and a three-dimensional load is applied to the end surface of the upper plate, and each component of the three-dimensional load is less than 10,000 Pa.

[0016] Furthermore, the deep learning model is a neural network built using Python libraries TensorFlow and Keras, the input of which includes geometric parameters, position information and load parameters, and the output is the grid cell size.

[0017] Furthermore, the neural network is a fully connected neural network with 7 layers, with 32, 64, 128, 64, 32, 8, and 1 neurons in each layer, respectively. This setting enables better training and fitting effects and more accurate predictions.

[0018] Furthermore, the training data of the deep learning model is generated by:

[0019] Generate T-joint models with random geometric parameters;

[0020] Finite element solution is performed through uniform background grid to obtain strain energy distribution;

[0021] A non-uniform grid is generated according to the strain energy distribution, and the node position information and the corresponding grid size are extracted as training data.

[0022] This setting improves the accuracy and efficiency of finite element calculations.

[0023] Furthermore, the position information is encoded using a median coordinate system and input into a neural network to predict the grid size.

[0024] Furthermore, the open source tool Tetgen is integrated for mesh generation.

[0025] Furthermore, the neural network training configuration includes: a loss function using mean square error, an optimizer using Adam, and a training cycle of 30 rounds, with each batch containing 128 data points. This configuration ensures good training convergence, better training fit, and more accurate predictions.

[0026] Furthermore, the grid adaptation method includes:

[0027] Defining the mesh size based on the strain energy distribution, where the mesh size is inversely proportional to the strain energy;

[0028] The number of grid cells is adjusted to a preset range through a loop process, and the preset range is determined by the user.

[0029] This setting makes the finite element calculation more efficient.

[0030] Furthermore, the finite element solver is FreeFem++, which is used to solve the linear elastic Lamé equation, and the user inputs material and boundary condition parameters in the interface.

[0031] Furthermore, the visualization tool is Paraview, and the output includes Mises stress, X-direction normal stress and strain energy density distribution diagram.

[0032] Compared with the prior art, the method for guiding grid adaptation of T-joints based on deep learning described in the present invention has the following advantages:

[0033] 1) This invention uses a deep learning model to obtain the optimal non-uniform grid cell size distribution, thereby obtaining a more accurate non-uniform grid. By predicting the optimal grid distribution, the deep learning model avoids the grid adjustment process based on posterior error estimation in traditional methods, significantly improving accuracy and efficiency. It also reduces the computation time of posterior error analysis and improves grid generation efficiency.

[0034] 2) The random geometric models generated by this invention through parametric design cover a variety of sizes and shapes of T-joints. By training these diverse samples, the deep learning model further improves the accuracy of finite element calculations. Moreover, the parametric design enables one-click simulation, lowering the operational threshold.

[0035] 3) The adaptive grid obtained by the present invention can balance calculation accuracy and efficiency and is suitable for complex geometries. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a schematic diagram of the simulated geometric shape of the present invention, showing the geometric parameters of the upper plate, lower plate and weld of the T-joint;

[0037] Figure 2 This is a screenshot of the program interface of the present invention, including a parameter input box, a calculation button, and a four-view result display area. DETAILED DESCRIPTION

[0038] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0039] The present invention proposes a method for guiding grid adaptation of T-joints based on deep learning, comprising the following steps:

[0040] S1, generate a T-joint model with random geometric parameters and define the geometric shape through parametric design;

[0041] Specifically, if Figure 1 The T-joint simulation diagram shown in the figure, the parameterized design includes the upper plate thickness l1, the lower plate thickness l2, the weld width 0.5l1 and the weld angle tangent value tan, the overall width of the T-joint is 4m. Its parameter range is: 0.4 m <l1<2.0 m,0.2 m<l2<1.0 m,0.5<tan<2.0。

[0042] Specifically, fixed displacement constraints are applied to both ends of the lower plate of the T-joint, and loads are applied to the upper plate end face along the positive x, y, and z directions, with the components in each direction less than 10,000 Pa. The finite element calculations in this method use the French open-source software FreeFem++. The code defines the linear elastic Lame algorithm, material parameters, and boundary conditions.

[0043] S2. Use deep learning models to predict grid cell size distribution and guide non-uniform grid generation;

[0044] Specifically, the training data of the deep learning model is generated in the following way:

[0045] Generate T-joint models with random geometric parameters;

[0046] Finite element solution is performed through uniform background grid to obtain strain energy distribution;

[0047] A non-uniform grid is generated according to the strain energy distribution, and the node position information and the corresponding grid size are extracted as training data.

[0048] More specifically, deep learning models require a large amount of data for training. Data generation involves uniformly distributed random numbers within a range of parameters. A dataset of 500 random finite element problems was selected. Within a single simulation problem domain, there are numerous nodes, each corresponding to a piece of training data. The input data consists of three geometric parameters, ten positional information components, and three load components. The output data is the mesh element size at the corresponding location. Positional information is expressed using mean value coordinates. To obtain the mesh element size for the training data, a high-quality non-uniform mesh must be generated. First, a uniform, relatively sparse background mesh is generated. A finite element solution is then performed on this mesh, and the strain energy distribution is calculated. The mesh element size is defined based on the inverse of the strain energy: small elements at locations with high strain energy and large elements at locations with low strain energy. Finally, a mesh generator is used to generate the non-uniform mesh based on the geometric boundaries and the element size distribution. A loop is used to identify the number of cells in the non-uniform mesh and compare it with the desired number of cells. The mesh is then repeatedly refined or thinned until the number of cells falls within the required range. Then the grid cell size around each node of the background grid is calculated.

[0049] After obtaining the dataset, a deep learning model was developed. The neural network was built using the Python libraries TensorFlow and Keras. We used a 7-layer fully connected neural network, with 32, 64, 128, 64, 32, 8, and 1 neurons in each layer, respectively. Training used mean square error as the loss function and Adam as the optimization method. Training was conducted for 30 rounds, with each batch containing 128 data points. After training, the neural network was used to generate a non-uniform mesh. First, a geometric shape was generated based on the size parameters entered by the user. A sparser uniform mesh was then generated on this geometry as the background mesh. The mean coordinates and other parameters of each node on the background mesh were read in as the input tensor to the neural network. The neural network reads in this tensor and outputs the mesh size at each node position. The mesh generator Tetgen reads in the background mesh and the cell size distribution to generate the non-uniform mesh.

[0050] S3, calling the finite element solver to perform simulation calculation;

[0051] Specifically, after the user enters key parameters and clicks Start, the program automatically generates an executable file readable by FreeFem++, containing boundary conditions and the solution method. FreeFem++ then reads this file and the mesh generated by the neural network to quickly perform a finite element solution.

[0052] S4. Output strain energy, stress distribution and adaptive mesh results through visualization tools;

[0053] Specifically, after the calculation is completed in step S3, the program calls the open source post-processing software Paraview to display the strain energy, mesh, Mises stress and x-direction normal stress, as shown in Figure 2. Figure 2 The program interface screenshot of the present invention is shown.

[0054] As shown in Table 1 below, the maximum Mises stress values obtained using the above method under typical input parameters.

[0055] Table 1 Maximum Mises stress values under typical input parameters

[0056]

[0057] It can be seen from Table 1 that the result of calculating the grid generated by the neural network using the method of the present invention is close to that of the traditional adaptive grid, that is, the method of the present invention is not only fast but also highly accurate.

[0058] Although the present invention is disclosed as above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope defined by the claims.

Claims

1. A method for T-joint guided grid adaptation based on deep learning, characterized in that: The steps include: S1, generate a T-joint model with random geometric parameters and define the geometric shape through parametric design; S2. Use deep learning models to predict grid cell size distribution and guide non-uniform grid generation; S3, calling the finite element solver to perform simulation calculation; S4. Output strain energy, stress distribution and adaptive mesh results through visualization tools.

2. The method according to claim 1, characterized in that The parameterized design includes the upper plate thickness l1, the lower plate thickness l2, the weld width 0.5l1 and the weld lower angle tangent value tan, and the parameter range is: 0.4 m <l1<2.0 m,0.2 m<l2<1.0 m,0.5<tan<2.0。 3. The method according to claim 1, characterized in that The deep learning model is a neural network built using the Python libraries TensorFlow and Keras. The input includes geometric parameters, position information and load parameters, and the output is the grid cell size.

4. The method according to claim 3, characterized in that The training data for the deep learning model is generated in the following way: Generate T-joint models with random geometric parameters; Finite element solution is performed through uniform background grid to obtain strain energy distribution; A non-uniform grid is generated according to the strain energy distribution, and the node position information and the corresponding grid size are extracted as training data.

5. The method according to claim 4, characterized in that The position information is encoded using a median coordinate system and input into a neural network to predict the grid size.

6. The method according to claim 1, characterized in that The grid adaptation method comprises: Defining the mesh size based on the strain energy distribution, where the mesh size is inversely proportional to the strain energy; The number of grid cells is adjusted to a preset range through a loop procedure.

7. The method according to claim 1, characterized in that The finite element solver is FreeFem++, which is used to solve the linear elastic Lamé equation. The user inputs material and boundary condition parameters in the interface.

8. The method according to claim 3, characterized in that The neural network is a fully connected neural network, which has 7 layers, and each layer has 32, 64, 128, 64, 32, 8, and 1 neurons respectively.

9. The method according to claim 8, characterized in that The training configuration of the fully connected neural network includes: the loss function is mean square error; the optimizer is Adam; the training cycle is 30 rounds, and each batch contains 128 data.

10. The method according to claim 1, characterized in that The visualization tool is Paraview, and the output includes Mises stress, X-direction normal stress and strain energy density distribution diagram.

Images