A Graph-Based GAN-Based Approach to Optimize Cooling Channel Design for Gas Turbine Blades

Abstract

This invention relates to a graph-based GAN-based method for optimizing cooling channels in gas turbine blades, belonging to the field of blade cooling technology. The method includes: first, establishing a parametric model to determine the design variables to be optimized and the constraint equations for each parameter, and generating different cooling channels using a parametric modeling approach. Second, sampling and generating structures that satisfy the model constraints within the design parameter space, and calculating the model's physical fields. Next, preprocessing the data, including data filtering and information preprocessing. Then, building multiple physical field prediction models to achieve a direct mapping from design variables and grid node information to physical field information. Following this, building a graph-based GAN-based generative adversarial model to achieve a continuous mapping from sampling to the cooling channel model. Finally, combining the generative adversarial model and multiple physical field prediction models, obtaining the gradient information between the optimization objective and the cooling channel shape, and using gradient descent to optimize and obtain design parameter values ​​that meet the design requirements.

Description

Technical Field

[0001] This invention belongs to the field of blade cooling technology, specifically relating to a method for designing and optimizing cooling channels for gas turbine blades based on Graph-Based GAN. Background Technology

[0002] Gas turbines, characterized by rapid start-up and shutdown and flexible operation, play a vital supporting role in power systems. To improve the efficiency and power of gas turbines, increasing the turbine inlet temperature is the most direct and effective method. This subjects the turbine blades to extremely high heat loads over extended periods, necessitating a series of cooling measures for high-temperature protection. Among these, internal blade cooling is one of the most effective methods. Therefore, researching and optimizing the design of internal cooling channels for turbine blades is of paramount importance for enhancing blade safety, ensuring the safe and stable operation of gas turbine units, reducing significant economic losses, and preventing catastrophic accidents.

[0003] Current research on gas turbine cooling channels mainly includes numerical and experimental studies. Numerical studies have advantages such as low cost and convenient calculation, and can obtain data that is difficult to obtain through experimental measurements. However, computational fluid dynamics-based methods require extensive and repeated analyses of the aero-thermal characteristics of turbine blades under different models and operating conditions, resulting in drawbacks such as large computational load, high memory requirements, and long computation time. This significantly increases the design cycle of turbine blade cooling channels and makes real-time control of the unit impossible. Experimental studies have higher reliability and can obtain data that most closely approximates the actual operating conditions, but they also have drawbacks such as difficulties in setting up experimental environments, high complexity in designing experimental components, and high costs.

[0004] On the one hand, both traditional numerical and experimental research methods have certain drawbacks. For numerical methods, the research object is usually a cooling channel with a defined profile, and each update iteration requires multiple steps such as design, modeling, mesh generation, calculation, and post-processing, resulting in a long optimization cycle. For experimental methods, due to the high design and manufacturing costs of test pieces, it is almost impossible to optimize cooling channel design parameters using experimental methods. On the other hand, rapidly developing artificial intelligence algorithms have powerful data processing capabilities and have been widely and successfully applied in many fields. Using artificial intelligence algorithms to design and optimize gas turbine cooling channels not only has higher optimization efficiency, saving significant time costs in optimizing cooling channel design parameters for gas turbine blades, but also achieves better optimization results, generating cooling channel models with better overall performance. Models obtained based on artificial intelligence algorithms are small in size, fast in speed, and highly transferable, making them suitable for the design and optimization of internal cooling channels in various blades. Summary of the Invention

[0005] The purpose of this invention is to quickly obtain high-performance turbine blade cooling channels to ensure the safe operation of gas turbine units. It provides a graph-based GAN-based method for designing and optimizing gas turbine blade cooling channels, which solves the problems of low optimization efficiency, poor optimization effect, long calculation cycle, high cost and unfavorable industrial application in traditional methods.

[0006] This invention is achieved using the following technical solution:

[0007] The method for optimizing the design of cooling channels for gas turbine blades based on graph-based GAN includes the following steps:

[0008] 1) Parametric model:

[0009] For the internal cooling channels of turbine blades, the cooling channel profile and cooling structure layout are selected as design variables to be optimized, and the design space and cooling channel profile parameters S are determined. i and cooling structure layout parameters R j The constraint equations between them are used to generate cooling channels of different shapes using a parametric modeling method, where i = 1, 2, 3, ..., n, n is the number of cooling channel profile parameters, and j = 1, 2, 3, ..., m, m is the number of cooling structure layout parameters;

[0010] 2) Sample generation:

[0011] Based on the sampling principle, in the design parameter space {S i R j The design parameters are randomly sampled within the matrix to form a design parameter matrix A. Then, a model structure set {C} satisfying the constraints is generated based on the sampled design parameter matrix A. p The finite element model set {M} is automatically generated using a scripting language. p}, according to the specified boundary conditions, the finite element model set {M p The model is used to perform calculations to obtain the physical field information dataset of the model. Where p is the number of sampling points, q is the number of recorded field data, and CO0 is the information matrix of all network nodes. This includes temperature field data, pressure field data, and velocity field data;

[0012] 3) Data preprocessing:

[0013] Based on the calculated physical field information dataset The calculated data was filtered, removing datasets with unacceptable heat transfer levels and data points with calculation anomalies. The filtered design variable matrix and physical field information datasets are A1 and A2, respectively. Meanwhile, for different deep learning networks, the physical field information of finite element calculation is preprocessed to make the output of the physical field conform to the input format of the deep learning model.

[0014] 4) Construct a physical field prediction model based on deep learning:

[0015] Multiple physics prediction models were built and trained to generate a dataset of physics information from the design variable matrix A and the network node information matrix CO1. Mapping between;

[0016] 5) Build a Graph-Based GAN generative adversarial network model:

[0017] Build a graph-based GAN generative adversarial network model to learn how to generate cooling channel models, and achieve a continuous mapping from sampling to cooling channel profiles and then to cooling channel models;

[0018] 6) Optimization based on gradient descent:

[0019] A gradient optimization model is formed by combining a generative network and a physics prediction model, and this model is used to obtain the optimization objective f. opt Based on the gradient information of the cooling channel profile, construct the optimization objective f. opt With cooling channel profile parameter S i and cooling structure layout parameters R j The differential relationship is used to optimize the design parameters with the best overall cooling performance by employing the gradient descent method.

[0020] A further improvement of the present invention is that it further includes the following steps:

[0021] 7) Algorithm maintenance:

[0022] In practice, when adding new design parameters, the constraint equations for the new design parameters are formulated according to the model parameterization method in step 1). Then, a new design parameter matrix is ​​formed in the new design parameter space and a new finite element model set and physical field information dataset are constructed according to the method in step 2). Next, the new physical field dataset is screened and preprocessed according to the method in step 3). Then, a new physical field prediction model based on deep learning is constructed according to the method in step 4). After that, a new cooling channel generation model based on GAN is built according to the method in step 5). Finally, the differential relationship between the optimization objective and the new design variables is constructed according to the method in step 6), and the gradient descent method is used for optimization to obtain the new design parameter values ​​with the best comprehensive cooling performance.

[0023] A further improvement of the present invention is that step 1) is implemented as follows:

[0024] 101) Determine the cooling channel profile parameter S i and cooling structure layout parameters R j ;

[0025] 102) Determine the cooling channel profile parameter S i and cooling structure layout parameters R j The constraint equations are expressed as follows:

[0026] D i,min ≤S i ≤D i,max

[0027] D j,min ≤R j ≤D j,max

[0028] Among them, D i，min and D i，max The cooling channel profile parameters S are respectively i The lower and upper bounds of D j，min and D j，max The cooling structure layout parameters R are respectively j The lower and upper bounds;

[0029] In the actual design process, due to the existence of geometric constraints, the variables are not absolutely independent. In order to ensure the closure of the geometric model, the preceding variables will affect the effective value range of the subsequent variables. Therefore, the value range of the subsequent variables should be dynamically adjusted according to the constraints determined by the preceding variables to form a new effective value range. Then, the variables are randomly sampled within this range.

[0030] 103) Based on the constraint equations, different shapes of cooling channels are generated using parametric modeling.

[0031] A further improvement of the present invention is that step 2) is implemented as follows:

[0032] 201) In the design parameter space {S i R j Based on the optimization objective, a sampling method is established, and design parameters are randomly sampled to form a design parameter matrix A;

[0033] 202) Generate a model structure set {C} that satisfies the constraints. p The finite element model set {M} is automatically generated using a scripting language. p};

[0034] 203) Determine the boundary conditions of the cooling channel model based on the actual working conditions, and apply the same boundary conditions to the finite element model set {M}. pThe physical field information dataset of all models on the dataset is obtained by performing calculations on all models.

[0035] Wherein, the design parameter matrix A is a p×m or p×n matrix, and the physical field information dataset... The specific format is as follows:

[0036]

[0037] Where 'a' represents the number of nodes in each physical field data, 'x', 'y', and 'z' represent the coordinates of the node in the three directions, 'f' represents the physical field data of the node, and the subscript represents the node number.

[0038] A further improvement of the present invention is that, in step 2), 201), the sampling method is selected from the principles of hierarchical sampling, Latin hypercube sampling, Holton sampling, and Poisson disk sampling.

[0039] A further improvement of the present invention is that step 3) is implemented as follows:

[0040] 301) Filter the calculation data and remove the datasets that do not meet the heat transfer requirements. The design variable matrix after filtering is A1.

[0041] 302) Continue filtering the calculation data, removing data points with calculation anomalies. The resulting physics information dataset is:

[0042] 303) Preprocess the physical field information calculated by finite element method to make the output of the physical field conform to the input format of the deep learning model.

[0043] A further improvement of this invention is that, in step 4) 303), if a convolutional neural network is selected as the deep learning model, a pixel-based processing method is adopted. The interpolation function within the cell is calculated using the coordinates of the grid cell and the information of the cell nodes to obtain an approximate mapping relationship between the coordinates and the physical field information, thus obtaining the result for any point. Specifically, D×H grid nodes are uniformly distributed within a region Ω containing the model, with some nodes inside the cell and others outside the model. For the cell formed by node eifg and node c of the interpolation grid, and after transforming the cell to the natural coordinate system, the interpolation function for point c is:

[0044] f ζ (x,y)=f ζ (0,0)(1-η)(1-ζ)+f ζ (1,0)η(1-ζ)+f ζ (0,1)(1-η)ζ+f ζ (1,1)ηζ

[0045] By using pixel-based interpolation, the dataset of the physical field distribution of the cooling channels is equivalent to N×D×H×C. f The image format is defined as follows: N represents the number of samples, D and H represent the width and height of the converted image, respectively, and C... f This represents the number of channels in the image, corresponding to the number of predicted physical field types;

[0046] If a graph convolutional neural network is chosen as the deep learning model, the adjacency relationships between grid nodes are described in mathematical graph form. Specifically, the grid's connectivity is described as G = {V, E}, where V represents the set of all nodes in the finite element grid, and E represents all edges of the grid. Nodes are numbered sequentially starting from 1, without repetition or gaps. The adjacency relationships of all nodes in G are represented by a symmetric adjacency matrix AD.

[0047]

[0048] A further improvement of this invention is that the specific implementation method of step 4) is as follows:

[0049] 401) Select a suitable network structure and build various physical field prediction models;

[0050] 402) Define the network loss function, whose expression is:

[0051]

[0052] Where Φ represents the learned parameters of the network model, and ψ represents the real field parameters. Here, E is the prediction field parameter, D is the mapping function, and the loss function is selected from MSELoss and MAELoss based on the actual situation.

[0053] 403) Train the network model to convert the design variable matrix A and the network node information matrix CO1 into a physical field information dataset. The mapping between them is expressed as:

[0054]

[0055] in, This represents the mapping relationship between inputs and outputs.

[0056] A further improvement of the present invention is that the specific implementation method of step 5) is as follows:

[0057] 501) Determine the network structure parameters and build a Graph-Based GAN generative adversarial network model;

[0058] 502) Train a GAN (Generative Adversarial Network) model to achieve a continuous mapping from sampling to cooling channel profiles, and then to the cooling channel model. The network model training process can be expressed as follows:

[0059] min G max D V(D,G)=E x [log(D(X))]+E G(Z) [log(1-D(G(Z)))]

[0060] Where G represents the generative model, D represents the discriminative model, and V represents the adversarial loss function.

[0061] A further improvement of this invention is that the specific implementation method of step 6) is as follows:

[0062] 601) Select the optimization objective f opt Its expression is:

[0063]

[0064] Where w k As weight, The expression is x is the design variable, i.e., {S} i R j}, where Θ is the input parameter;

[0065] 602) Based on a deep learning model, establish the cooling channel profile parameters S i and cooling structure layout parameters R j To optimize objective f opt The direct mapping relationship;

[0066] 603) Construct a gradient-based automatic differentiation mechanism, the expression of which is:

[0067]

[0068] Where τ is the step size factor controlling the update. To optimize objective f opt The gradient;

[0069] 604) The gradient descent method is used for optimization to obtain the design parameter values ​​corresponding to the cooling channel with the best overall cooling performance.

[0070] The present invention has at least the following beneficial technical effects:

[0071] Compared with traditional methods, this invention provides a graph-based GAN-based method for designing and optimizing cooling channels for gas turbine blades. It can construct a direct mapping relationship between design variables and optimization objectives, and quickly obtain the optimal cooling channel for the objective variables based on the gradient descent method.

[0072] Furthermore, this invention avoids the drawback of traditional methods where manual optimization is prone to getting trapped in local optima by randomly sampling within a multidimensional design space and obtaining the global optimal solution for the target variable based on the gradient descent method.

[0073] Furthermore, this invention obtains the cooling channel performance of gas turbine blades under different design parameters by sampling and calculating the multi-dimensional design space. It does not require exploring the impact of a single design parameter on the cooling channel performance, and therefore does not require designers to have extensive engineering experience and analytical capabilities.

[0074] Furthermore, by establishing a deep learning-based physical field prediction model, this invention can quickly predict the physical field information of the cooling channel of a gas turbine blade based on design variables.

[0075] Furthermore, by constructing an expression of design variables and physical field information with optimization objectives, this invention can quickly predict the performance of cooling channels for gas turbine blades based on design parameters and physical field information.

[0076] In summary, this invention provides a simple, convenient, fast, and effective method for designing and optimizing cooling channels for gas turbine blades, which is suitable for designing and optimizing internal cooling channels for various turbine blades. Attached Figure Description

[0077] Figure 1 This is a flowchart illustrating the design optimization method for cooling channels of gas turbine blades based on Graph-Based GAN, as described in this invention.

[0078] Figure 2 This is a schematic diagram showing the profile parameters and settings of the cooling channel for the gas turbine blades in the first embodiment of the present invention.

[0079] Figure 3 This is a schematic diagram of the Graph-Based GAN model structure in this invention.

[0080] Figure 4 This is a schematic diagram of the cooling structure layout and parameters of the gas turbine blade cooling channel in the second embodiment of the present invention.

[0081] In the diagram: L1-L 10 R1-R8 is the length of the straight segment, R1-R8 is the length of the transition arc, α1-α8 is the angle of the transition arc, θ1 is the angle of the inlet segment, θ2 is the angle of the outlet segment, L1-L10 RR1-RR is the length of the straight segment of the rib column. 10 For the diameter of the rib column, HH1-HH 10 This is the distance between the rib and the wall. Detailed Implementation

[0082] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Without departing from the above-described spirit of the invention, the method of the present invention is not only applicable to the design of gas turbine cooling channels, but can also be extended to the design and optimization of various channels according to actual problems.

[0083] First implementation method:

[0084] like Figure 1 As shown, the present invention provides a method for optimizing the design of cooling channels for gas turbine blades based on Graph-Based GAN, comprising the following 7 steps:

[0085] 1) Parametric model:

[0086] For the internal cooling channels of turbine blades, the cooling channel profile is selected as the design variable to be optimized, and the design space and cooling channel profile parameter S are determined. i The constraint equations between them are used to generate different cooling channels using a parametric modeling method, where i = 1, 2, 3, ..., n, and n is the number of cooling channel profile parameters.

[0087] The specific implementation method of step 1) is as follows:

[0088] 101) Determine the cooling channel profile parameter S i In this embodiment, the lengths of the straight line segments L1-L are included. 10 The radius of the transition section arc is R1-R8, the angle of the transition section arc is α1-α8, the included angle of the inlet section is θ1, and the included angle of the outlet section is θ2.

[0089] 102) Determine the cooling channel profile parameter S i The constraint equations are expressed as follows:

[0090] D i,min ≤S i ≤D i,max

[0091] Among them, D i，min and D i，max The cooling channel profile parameters S are respectively i The lower and upper bounds.

[0092] Due to the existence of geometric constraints, the variables are not absolutely independent. In order to ensure the closure of the geometric model, the preceding variables may affect the effective value range of the subsequent variables. Therefore, the value range of the subsequent variables should be dynamically adjusted according to the constraints determined by the preceding variables to form a new effective value range. Then, the variables are randomly sampled within this range.

[0093] 103) Based on the constraint equations, different shapes of cooling channels are generated using parametric modeling.

[0094] 2) Sample generation:

[0095] Based on specific sampling principles, in the design parameter space {S i The design parameters are randomly sampled within the matrix to form a design parameter matrix A. Then, a model structure set {C} satisfying the constraints is generated based on the sampled design parameter matrix A. p The finite element model set {M} is automatically generated using a scripting language. p}, according to the specified boundary conditions, the finite element model set {M p The model is used to perform calculations to obtain the physical field information dataset of the model. Where p is the number of sampling points, q is the number of recorded field data, and CO0 is the information matrix of all network nodes. This includes temperature field data, pressure field data, and velocity field data.

[0096] The specific implementation method for step 2) is as follows:

[0097] 201) In the design parameter space {S i Based on the optimization objective, a suitable sampling method is established. Sampling principles such as hierarchical sampling, Latin hypercube sampling, Holton sampling, and Poisson disk sampling can be selected to randomly sample design parameters and form a design parameter matrix A.

[0098] 202) Generate a model structure set {C} that satisfies the constraints. p The finite element model set {M} is automatically generated using a scripting language. p};

[0099] 203) Determine the boundary conditions of the cooling channel model based on the actual working conditions, and apply the same boundary conditions to the finite element model set {M}. p The physical field information dataset of all models on the dataset is obtained by performing calculations on all models.

[0100] Wherein, the design parameter matrix A is a p×m matrix, and the physical field information dataset... The specific format is as follows:

[0101]

[0102] Where 'a' represents the number of nodes in each physical field data, 'x', 'y', and 'z' represent the coordinates of the node in the three directions, 'f' represents the physical field data of the node, and the subscript represents the node number.

[0103] 3) Data preprocessing:

[0104] Based on the calculated physical field information dataset The calculated data was filtered, removing datasets with unacceptable heat transfer levels and data points with calculation anomalies. The filtered design variable matrix and physical field information datasets are A1 and A2, respectively. Simultaneously, for different deep learning networks, the physical field information calculated by finite element method is preprocessed to make the output of the physical field conform to the input format of the deep learning model.

[0105] The specific implementation method for step 3) is as follows:

[0106] 301) Filter the calculation data and remove the datasets that do not meet the heat transfer requirements. The design variable matrix after filtering is A1.

[0107] 302) Continue filtering the calculation data, removing data points with calculation anomalies. The resulting physics information dataset is:

[0108] 303) Preprocess the physical field information calculated by finite element method to make the output of the physical field conform to the input format of the deep learning model;

[0109] like Figure 3 As shown, in this embodiment, a graph convolutional neural network is chosen as the deep learning model. The adjacency relationships between grid nodes are then described mathematically in graph form. Specifically, the grid's connectivity is described as G = {V, E}, where V represents the set of all nodes in the finite element grid, and E represents all edges of the grid. Nodes are numbered sequentially starting from 1, without repetition or gaps. The adjacency relationships of all nodes in G can be represented as a symmetric adjacency matrix AD.

[0110]

[0111] 4) Construct a physical field prediction model based on deep learning:

[0112] Multiple physics prediction models were built and trained to generate a dataset of physics information from the design variable matrix A and the network node information matrix CO1. The mapping between them.

[0113] The specific implementation method for step 4) is as follows:

[0114] 401) Select a suitable network structure and build various physical field prediction models;

[0115] 402) Define the network loss function, whose expression is:

[0116]

[0117] Where Φ represents the learned parameters of the network model, and ψ represents the real field parameters. Here, E is the prediction field parameter, D is the mapping function, and the loss function chosen is MSELoss.

[0118] 403) Train the network model to convert the design variable matrix A and the network node information matrix CO1 into a physical field information dataset. The mapping between them is expressed as:

[0119]

[0120] in, This represents the mapping relationship between inputs and outputs.

[0121] 5) Build a GAN-based cooling channel generation model:

[0122] like Figure 3 As shown, a Graph-Based GAN generative adversarial network model is built to learn the generation method of the cooling channel model, realizing a continuous mapping from sampling to cooling channel profile and then to the cooling channel model;

[0123] The specific implementation method for step 5) is as follows:

[0124] 501) Determine appropriate network structure parameters and build a Graph-Based GAN generative adversarial network model;

[0125] 502) Train a GAN (Generative Adversarial Network) model to achieve a continuous mapping from sampling to cooling channel profiles, and then to the cooling channel model. The network model training process can be expressed as:

[0126] min G max D V(D,G)=E x [log(D(X))]+E G(Z) [log(1-D(G(Z)))]

[0127] Where G represents the generative model, D represents the discriminative model, and V represents the adversarial loss function.

[0128] 6) Optimization based on gradient descent:

[0129] A gradient optimization model is formed by combining a generative network and a physics prediction model, and this model is used to obtain the optimization objective f. opt Based on the gradient information of the cooling channel profile, construct the optimization objective f. opt With cooling channel profile parameter S i The differential relationship is used to optimize the design parameters with the best overall cooling performance by employing the gradient descent method.

[0130] The specific implementation method for step 6) is as follows:

[0131] 601) Select the optimization objective f opt Its expression is:

[0132]

[0133] Where w k As weight, The expression is x is the design variable, i.e., {S} i}, where Θ is the input parameter.

[0134] 602) Based on a deep learning model, establish the cooling channel profile parameters S i To optimize objective f opt The direct mapping relationship.

[0135] 603) Construct a gradient-based automatic differentiation mechanism, the expression of which is:

[0136]

[0137] Where τ is the step size factor controlling the update. To optimize objective f opt The gradient.

[0138] 604) The gradient descent method is used for optimization to obtain the design parameter values ​​corresponding to the cooling channel with the best overall cooling performance.

[0139] 7) Algorithm maintenance:

[0140] In practice, when adding new design parameters, the constraint equations for the new design parameters are formulated according to the model parameterization method in step 1). Then, a new design parameter matrix is ​​formed in the new design parameter space and a new finite element model set and physical field information dataset are constructed according to the method in step 2). Next, the new physical field dataset is screened and preprocessed according to the method in step 3). Then, a new physical field prediction model based on deep learning is constructed according to the method in step 4). After that, a new cooling channel generation model based on GAN is built according to the method in step 5). Finally, the differential relationship between the optimization objective and the new design variables is constructed according to the method in step 6), and the gradient descent method is used for optimization to obtain the new design parameter values ​​with the best comprehensive cooling performance.

[0141] Second implementation method:

[0142] The specific difference between this implementation method and the first implementation method is that, for example... Figure 4 As shown, step 101) is: determining the cooling structure layout parameters R. j In this embodiment, the length of the straight segment L1-L of the rib column is included. 10 , Rib diameter RR1-RR 10 , distance from rib to wall HH1-HH 10 .

[0143] Step 102) is: Determine the cooling structure layout parameters R. j The constraint equations are expressed as follows:

[0144] D j,min ≤R j ≤D j,max

[0145] Among them, D j，min and D j，max The cooling structure layout parameters R are respectively j The lower and upper bounds.

[0146] Based on the cooling structure layout parameter R j The resulting design parameter space is {R} j}

[0147] The above implementation methods are only used to illustrate the present invention and are not intended to limit the present invention. Although the present invention has been described in detail, those skilled in the art should understand that various combinations, modifications or substitutions of the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and should be covered within the scope of the claims of the present invention.

Claims

1. A method for optimizing the design of cooling channels for gas turbine blades based on Graph-Based GAN, characterized in that, Includes the following steps: 1) Parametric model: For the internal cooling channels of turbine blades, the cooling channel profile and cooling structure layout are selected as design variables to be optimized, and the design space and cooling channel profile parameters are determined. S i and cooling structure layout parameters R j The constraint equations between them are used to generate cooling channels of different shapes using a parametric modeling approach. i =1, 2, 3..., n , n The number of cooling channel profile parameters. j =1, 2, 3..., m , m Number of parameters for the cooling structure layout; 2) Sample generation: Based on the sampling principle, in the design parameter space { S i , R j Randomly sampled design parameters within the matrix to form a design parameter matrix. A Then, based on the design parameter matrix obtained from sampling A Generate a set of model structures that satisfy the constraints. C p }, using a scripting language to automatically generate finite element model sets { M p }, apply the specified boundary conditions to the finite element model set { M p The model is used to perform calculations to obtain the physical field information dataset of the model. CO 0, },in, p The number of sampling points. q The number of field data points recorded. CO 0 represents the information matrix of all network nodes. This includes temperature field data, pressure field data, and velocity field data; 3) Data preprocessing: Based on the calculated physical field information dataset { The calculated data was filtered, removing datasets with unacceptable heat transfer levels and abnormal data points. The filtered design variable matrix and physical field information dataset were as follows: A 1 and { CO 1, Meanwhile, for different deep learning networks, the physical field information of finite element calculation is preprocessed to make the output of the physical field conform to the input format of the deep learning model. 4) Construct a physical field prediction model based on deep learning: Multiple physical field prediction models were built and trained to achieve the goal of designing variable matrices. A 1 and network node information matrix CO 1 to physical field information dataset { Mapping between}; 5) Build a Graph-Based GAN generative adversarial network model: Build a graph-based GAN generative adversarial network model to learn how to generate cooling channel models, and achieve a continuous mapping from sampling to cooling channel profiles and then to cooling channel models; 6) Optimization based on gradient descent: A gradient optimization model is formed by combining a generative network and a physics prediction model, and the optimization objective is obtained using this model. f opt Based on the gradient information of the cooling channel profile, construct the optimization objective. f opt With cooling channel profile parameters S i and cooling structure layout parameters R j The differential relationship is used to optimize the design parameter values ​​with the best overall cooling performance by employing the gradient descent method.

2. The method for designing and optimizing cooling channels for gas turbine blades based on Graph-Based GAN according to claim 1, characterized in that, It also includes the following steps: 7) Algorithm maintenance: In practice, when adding new design parameters, the constraint equations for the new design parameters are formulated according to the model parameterization method in step 1). Then, a new design parameter matrix is ​​formed in the new design parameter space and a new finite element model set and physical field information dataset are constructed according to the method in step 2). Next, the new physical field dataset is screened and preprocessed according to the method in step 3). Then, a new physical field prediction model based on deep learning is constructed according to the method in step 4). After that, a new cooling channel generation model based on GAN is built according to the method in step 5). Finally, the differential relationship between the optimization objective and the new design variables is constructed according to the method in step 6), and the gradient descent method is used for optimization to obtain the new design parameter values ​​with the best comprehensive cooling performance.

3. The method for designing and optimizing cooling channels for gas turbine blades based on Graph-Based GAN as described in claim 1, characterized in that, The specific implementation method of step 1) is as follows: 101) Determine the cooling channel profile parameters S i and cooling structure layout parameters R j ; 102) Determine the cooling channel profile parameters S i and cooling structure layout parameters R j The constraint equations are expressed as follows: in, D i，min and D i，max These are the cooling channel profile parameters. S i The lower and upper bounds, D j，min and D j，max Cooling structure layout parameters R j The lower and upper bounds; In the actual design process, due to the existence of geometric constraints, the variables are not absolutely independent. In order to ensure the closure of the geometric model, the preceding variables will affect the effective value range of the subsequent variables. Therefore, the value range of the subsequent variables should be dynamically adjusted according to the constraints determined by the preceding variables to form a new effective value range. Then, the variables are randomly sampled within this range. 103) Based on the constraint equations, different shapes of cooling channels are generated using parametric modeling.

4. The method for designing and optimizing cooling channels for gas turbine blades based on Graph-Based GAN according to claim 3, characterized in that, The specific implementation method of step 2) is as follows: 201) In the design parameter space { S i , R j Based on the optimization objective, a sampling method is established, and design parameters are randomly sampled to form a design parameter matrix. A ; 202) Generate a set of model structures that satisfy the constraints { C p }, using a scripting language to automatically generate finite element model sets { M p }; 203) Determine the boundary conditions of the cooling channel model based on the actual working conditions, and apply the same boundary conditions to the finite element model set { M p The dataset { is used to perform calculations on all models on the dataset to obtain the physical field information of the models.} CO 0, }; Among them, the design parameter matrix A for p × m or p × n The matrix, physical field information dataset { The specific format of} is: in, a This indicates the number of nodes in each physical field data point. x , y , z These represent the coordinates of the node in three directions. f This represents the physical field data of a node, with the subscript indicating the node number.

5. The method for designing and optimizing cooling channels for gas turbine blades based on Graph-Based GAN according to claim 4, characterized in that, In step 2), 201), the sampling method is selected from stratified sampling, Latin hypercube sampling, Holton sampling, and Poisson disk sampling principles.

6. The method for designing and optimizing cooling channels for gas turbine blades based on Graph-Based GAN according to claim 4, characterized in that, The specific implementation method of step 3) is as follows: 301) Filter the calculation data, remove datasets whose heat transfer levels do not meet the requirements, and the resulting design variable matrix is: A 1; 302) Continue filtering the calculation data, removing data points with calculation anomalies. The resulting physics information dataset is { CO 1, }; 303) Preprocess the physical field information calculated by finite element method to make the output of the physical field conform to the input format of the deep learning model.

7. The method for designing and optimizing cooling channels for gas turbine blades based on Graph-Based GAN according to claim 6, characterized in that, In step 3), if a convolutional neural network is chosen as the deep learning model, a pixel-based processing method is used. The interpolation function within the grid cell is calculated using the coordinates and node information of the grid cell to obtain an approximate mapping relationship between the coordinates and the physical field information, thus obtaining the result for any point. Specifically, D×H grid nodes are uniformly distributed within a region Ω containing the model, with some nodes inside the cell and others outside the model. For each node… eifg The nodes of the formed cells and interpolation mesh c And by transforming the unit to the natural coordinate system, the point c The interpolation function is: By using pixel-based interpolation, the dataset of the physical field distribution of the cooling channels is equivalent to... N × D × H × C f The image format, among which, N Indicates the number of samples. D and H These represent the width and height of the converted image, respectively. C f This represents the number of channels in the image, corresponding to the number of predicted physical field types; If a graph convolutional neural network is chosen as the deep learning model, the adjacency relationships between grid nodes are described in mathematical graph form. Specifically, the grid's connectivity is described as G={V,E}, where V represents the set of all nodes in the finite element mesh, and E represents all edges of the mesh. Nodes are numbered sequentially starting from 1, without repetition or gaps. The adjacency relationships of all nodes in G are represented by a symmetric adjacency matrix AD. 。 8. The method for designing and optimizing cooling channels for gas turbine blades based on Graph-Based GAN according to claim 6, characterized in that, The specific implementation method for step 4) is as follows: 401) Select a suitable network structure and build various physical field prediction models; 402) Define the network loss function, whose expression is: in, These are the learning parameters for the network model. For real field parameters, Here, E represents the predicted field parameters, and E is the mapping function. D For the computational domain, loss Choose between MSELoss and MAELoss loss functions based on the actual situation; 403) Train the network model to achieve the design of the variable matrix. A 1 and network node information matrix CO 1 to physical field information dataset { The mapping between} is expressed as: in, This represents the mapping relationship between inputs and outputs.

9. The method for designing and optimizing cooling channels for gas turbine blades based on Graph-Based GAN according to claim 8, characterized in that, The specific implementation method for step 5) is as follows: 501) Determine the network structure parameters and build a Graph-Based GAN generative adversarial network model; 502) Train a GAN (Generative Adversarial Network) model to achieve a continuous mapping from sampling to cooling channel profiles, and then to the cooling channel model. The network model training process can be expressed as follows: in G To generate models, D To discriminate the model, V To counteract the loss function.

10. The method for designing and optimizing cooling channels for gas turbine blades based on Graph-Based GAN according to claim 9, characterized in that, The specific implementation method for step 6) is as follows: 601) Select the optimization objective f opt Its expression is: in w k As weight, The expression is , For design variables, i.e. { S i , R j }, For input parameters; 602) Based on a deep learning model, establish cooling channel profile parameters. S i and cooling structure layout parameters R j To optimize the target f opt The direct mapping relationship; 603) Construct a gradient-based automatic differentiation mechanism, the expression of which is: in, To control the update step size factor, To optimize the target f opt The gradient; 604) The gradient descent method is used for optimization to obtain the design parameter values ​​corresponding to the cooling channel with the best overall cooling performance.

Images