Turbine gas film cooling variable fidelity modeling method and system based on field adaptive migration

Abstract

The invention aims to solve the problem that the performance of a deep learning-based air film cooling performance prediction model is seriously limited by the quality and quantity of training data, so that a traditional single-fidelity modeling method is always difficult to achieve balance between data cost and model performance. The invention provides a field adaptive migration-based turbine gas film cooling variable fidelity modeling method and system, and belongs to the technical field of rapid prediction of gas turbine gas film cooling performance. The method comprises the following steps: establishing a parameterization expression system taking a single gas film hole as a modeling unit, and realizing accurate geometric description of the complex gas film hole; constructing a multi-scale multi-channel self-attention network structure combined with gradient loss so as to improve the reconstruction precision of the model on a multi-local high-gradient complex flow field; and a transfer learning mechanism is introduced to form a progressive modeling strategy for data with different fidelity, so that efficient generalization of the model in a multi-fidelity scene is realized. The method can be applied to cooling design of air-cooled turbine blades of aero-engines and gas turbines.

Description

Technical Field

[0001] This invention belongs to the field of rapid prediction technology of gas turbine film cooling performance, specifically involving a variable fidelity modeling method for turbine film cooling based on domain adaptive migration. Background Technology

[0002] Gas turbines play a crucial role in aviation power, ship propulsion, ground power generation, and other industrial fields due to their compact structure, high power-to-weight ratio, and rapid start-up. The compressor, combustion chamber, and gas turbine are collectively known as the three main components of a gas turbine. The turbine, as the primary power-generating component, converts the internal energy of the high-temperature, high-pressure gas exiting the combustion chamber into kinetic energy, outputting mechanical work. However, turbines often operate in harsh environments with high temperature and pressure. To ensure the long-term stable and safe operation of the turbine, film cooling is one of the important thermal protection measures for the turbine's hot-end components.

[0003] Currently, the mainstream approach in engineering remains to first screen and optimize the air film pore arrangement scheme through CFD before conducting physical experiments. However, engineering design systems that use numerical simulation as a performance evaluation method require a large number of iterative calculations. To balance the accuracy of the results with the computational cost, RANS methods remain the most widely used evaluation method in engineering. On the one hand, continuing to develop high-performance computing capabilities is an effective approach; on the other hand, reducing the order of the original equations and using approximate models to replace part of the numerical iteration process will greatly improve the efficiency of parameter search and help to carry out large-scale optimization to find the global optimum.

[0004] The widespread adoption of advanced, refined measurement methods and high-precision CFD methods has provided massive data support for the introduction of machine learning. As a research method based on big data, machine learning excels at establishing highly nonlinear mapping relationships, an advantage favored by researchers who apply it to fluid mechanics problems. In the field of film cooling, the introduction of machine learning can be broadly divided into two categories. One is using machine learning methods to correct turbulence models. This type of work focuses on using machine learning methods to predict the distribution of dimensionless parameters in the flow field, providing anisotropic corrections for turbulence models and improving the computational accuracy of RANS solutions for film cooling problems. Currently, most two-equation turbulence models rely on linear eddy viscosity models to close the Reynolds stress. Existing research shows that LEVM cannot provide satisfactory accuracy when solving flow problems with curvature, collisions, and separation effects. Using high-fidelity data to correct the closing terms and empirical constants of turbulence models can effectively improve the applicability and computational accuracy of RANS methods, which has significant practical value in engineering applications. However, this type of research is still based on solving high-order partial differential equations and has not fundamentally solved the problem of long iteration times in the optimization process. Furthermore, although modifying the closure terms of the RANS turbulence model for anisotropic correction helps improve the model's accuracy in capturing complex flow phenomena, as Duraisamy et al. stated, even if the quantities calculated using DNS completely replace specific terms in the RANS closure, the overall prediction results are still unsatisfactory due to the assumptions and approximations at various levels of the model, coupled with the potential ill-conditioned nature of the RANS equations. This makes it difficult to meet the actual needs of a fast evaluation method in efficient optimization systems.

[0005] Therefore, researchers proposed a second technical approach: establishing a direct mapping from geometric information and boundary conditions to the adiabatic efficiency field of film cooling, and constructing a surrogate model for performance evaluation. Based on the types of film cooling aperture layouts addressed by the researchers and their understanding of the film cooling efficiency distribution prediction problem, these approaches can be broadly categorized into three types: direct prediction methods, time-series advancement methods, and superposition principle methods. Direct prediction methods, as the name suggests, directly establish a mapping model from the cooling layout to the efficiency distribution without introducing any prior assumptions. Wang et al., based on an image mapping strategy, used the projected contour image of the cooling layout on the end face as input data. They achieved rapid prediction of the cooling performance of complex film cooling layouts using a convolutional neural network and developed a reinforcement learning-based layout optimization method around the surrogate model, proposing design criteria for different heat load distributions.

[0006] The temporal progression method incorporates the concept of spatiotemporal iterative progression from numerical solutions into the design of neural network architectures and training strategies. This method posits that downstream flow is the result of the cumulative effect of upstream flow and local disturbances. Based on this idea, Yang et al. proposed an iterative operator neural network. The convolution kernel and sliding window convolution analogous to the spatial progression of numerical methods, while the iterative computation characteristics of the recurrent neural network analogous to temporal progression, reduce the prediction errors for temperature and pressure by approximately 20% compared to the directly mapped cGAN model. Yu et al. established a prediction model for the distribution of seepage cooling efficiency in a randomly perforated flat plate based on the iterative operator network. Furthermore, they established extended databases for different operating conditions and at different resolutions to verify the transferability of the proposed method. In subsequent work, the flat plate seepage cooling model was transferred to the blade seepage cooling problem, accelerating model convergence while avoiding overfitting caused by small samples.

[0007] The superposition principle method, developed from Seller's research, aims to decompose the cooling efficiency distribution of complex multi-row film cooling arrays into a weighted superposition of the cooling efficiency of small units, either a single row of holes or combinations of several holes. Due to the simplicity and intuitiveness of the superposition formula, many studies have focused on improving and modifying this method to broaden its applicability. Wang et al. proposed a physically enhanced neural network operator framework, SDNO, based on the superposition principle for predicting end-wall film cooling efficiency. This framework consists of a computational network and a superposition network. They trained the computational network using a series of randomly arranged basic hole units, including combinations of single, double, triple, and five holes, while the superposition network calculated the spatial weight distribution of different basic unit combinations.

[0008] In summary, existing deep learning-based prediction models for turbine film cooling performance rely on massive amounts of data, which still primarily come from simulations and experimental measurements. This means that in engineering practice, high-fidelity, high-value data is often sparsely distributed in terms of spatiotemporal resolution due to high acquisition costs, while low-fidelity data, although abundant, cannot accurately capture the true physical characteristics of film cooling flow. Therefore, traditional single-fidelity modeling methods still face a trade-off between data cost and overall model performance in practical applications. When training a model solely with low-fidelity data, the reliability of the prediction results is questionable; while training solely with high-fidelity data often leads to overfitting due to the small data volume, rendering the model unable to extrapolate to new data. Summary of the Invention This invention addresses the challenge that the performance of deep learning-based film cooling performance prediction models is severely limited by the quality and quantity of training data. This leads to a dilemma: high-fidelity data is scarce due to its high cost, and using it solely for training can easily result in overfitting and poor generalization ability; while the vast amount of low-fidelity data lacks accuracy, making the prediction results of models trained on it unreliable. Therefore, this invention proposes a variable-fidelity modeling method and system for turbine film cooling based on domain adaptive transfer.

[0009] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a variable fidelity modeling method for turbine film cooling based on domain adaptive migration, the method comprising the following steps: Input data representation steps: For the input data, accurately represent the complex array geometry of turbine film cooling based on lexical units; The steps for constructing a parameter mapping network are as follows: Construct a parameter list mapping multi-path attention network and construct a gradient loss function for the network; The progressive modeling steps based on transfer learning are as follows: The parameter mapping network in step 2 is trained with low-fidelity data to learn the basic physical characteristics of film cooling flow. The learnable parameters of the low-fidelity parameter mapping network are used as prior information and as the initial parameters of the high-fidelity prediction model. The prediction model with initial parameters is fine-tuned and trained with high-fidelity data to obtain a high-fidelity film cooling efficiency rapid prediction model.

[0010] Furthermore, the specific steps for expressing the above input data are as follows: Based on the lexical units in natural language processing, each hole is treated as an abstract character. ; Treating the set of all holes as a short sentence forms... ,in N The number of air film pores for each array layout, d The number of input features includes geometric parameters, dimensionless flow direction, and position. Dimensionless expansion position Inclination angle θ Composite angle φ aperture D Boundary condition parameter: air blowing ratio BR Compared to temperature TR ; The output data is a dimensionless temperature field of the blade surface. The surface temperature field data is extracted from the CFD results using a uniform grid. The turbine blade surface is uniformly divided into 256 sections along the spanwise direction. 256 sampling points are uniformly placed along the blade arc curve on the suction and pressure sides of each section. Finally, the blade surface temperature field, which is used as a label, is organized into 256×512 image data.

[0011] Furthermore, the above parameter list maps the multipath attention network to include an upstream boundary and geometry processing module and a downstream multipath flow field prediction module; The upstream module is used to receive the parameter list encoding and obtain the global features of the current air film aperture layout through latent vector calculation and spatial weight allocation of each aperture. The downstream module is used to receive global features of size 128×16×32 through three channels, and allocates different flow data sizes to the three channels through the PixelShuffle upsampling module.

[0012] Furthermore, the core computational unit of the above three pathways is two cascaded residual sliding window attention network modules RSTB. Each RSTB contains six sliding window attention computational units SwinT. Multi-scale features are fused through an element-wise summation aggregation function, and finally the cooling efficiency distribution field prediction value is obtained again through a PixelShuffle upsampling module and a convolutional output module.

[0013] Furthermore, the working data sizes of the three paths are 16×32, 32×64 and 64×128, respectively, and the number of RSTB modules used in each branch is 6, 4 and 2, respectively.

[0014] Furthermore, the gradient loss function described above adopts a composite loss of mean squared error and pixel error, and in this loss function, the weight allocation adopts an adaptive calculation method, which is autonomously allocated as the loss propagates back during the iteration process.

[0015] Secondly, the variable fidelity modeling method for turbine film cooling based on domain adaptive migration described in this invention can be entirely implemented using computer software. Therefore, correspondingly, this invention also provides a variable fidelity modeling system for turbine film cooling based on domain adaptive migration.

[0016] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the turbine film cooling variable fidelity modeling method based on domain adaptive migration as described above.

[0017] Fourthly, the present invention also provides a computer device, the device including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the turbine film cooling variable fidelity modeling method based on domain adaptive migration described in any one of the above.

[0018] The beneficial effects of this invention are as follows: This invention proposes a variable fidelity modeling method for turbine film cooling based on domain adaptive transfer. This method is a variable fidelity progressive modeling method for turbine film cooling, which includes three core parts: (1) establishing a parameterized expression system with a single film pore as the modeling unit to achieve accurate description of complex film pore geometry; (2) constructing a multi-scale, multi-path self-attention network structure combined with gradient loss to improve the model's reconstruction accuracy for complex flow fields with multiple local high gradients; (3) introducing a transfer learning mechanism to form a progressive modeling strategy for different fidelity data, so as to achieve efficient generalization of the model in multi-fidelity scenarios.

[0019] This invention can be applied to the cooling design of air-cooled turbine blades for aero-engines and gas turbines. Using cooling efficiency data of varying quantities and fidelity, a rapid prediction model for wall cooling efficiency is established, with its fidelity gradually improved. This method provides a new approach for low-cost construction of turbine aerodynamic heat transfer databases and for achieving self-evolution of the model. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the variable fidelity modeling method for turbine film cooling based on domain adaptive migration described in this embodiment of the invention. Figure 2 This refers to the network architecture of the ModelPara parameter mapping network described in this embodiment of the invention. Figure 3 This refers to the transfer learning framework for the turbine film cooling air-thermal coupling temperature field prediction model described in this embodiment of the invention. Figure 4 The two models described in this embodiment of the invention are the reconstruction results of some samples on the training set, wherein... Figure 4 (a) Reconstruction results of a portion of the samples on the training set for the ModelPara model built using the proposed method. Figure 4(b) Reconstruction results of a portion of the training set for the ModelImg model built using existing mature methods; Figure 5 The two models described in this embodiment of the invention are the reconstruction results of some samples on the test set, wherein... Figure 5 (a) Reconstruction results of a portion of the ModelPara model built using the proposed method on the test set. Figure 5 (b) Reconstruction results of the ModelImg model built using existing mature methods on a portion of the test set; Figure 6 This is a comparison of the lateral average results of two sets of models on the training set in this embodiment of the invention. Figure 6 (a) The lateral averaging results of some samples during training of the ModelPara model built using the proposed method. Figure 6 (b) The lateral averaging results of some samples during training of the ModelImg model built using existing mature methods; Figure 7 This is a comparison of the lateral average results of two sets of models on a partial sample of the test set, as described in this embodiment of the invention. Figure 7 (a) The partial sample lateral mean results of the ModelPara model built using the proposed method on the test. Figure 7 (b) Lateral averaging results of partial samples in the test of the ModelImg model established by existing mature methods; Figure 8 The convergence process of the ModelPara model described in this embodiment of the invention varies with the proportion of the training set, wherein... Figure 8 (a) shows the change in residual convergence of the model on the training set as a function of the proportion of the training set. Figure 8 (b) shows the change in residual convergence of the model on the test with the proportion of the training set; Figure 9 This refers to the average metric of the model described in this embodiment of the invention on the entire dataset as a percentage of the training set, wherein... Figure 9 (a) represents the PSNR response as a function of the training set proportion. Figure 9 (b) is the response of SSIM as a function of the training set proportion; Figure 10 These are the convergence residual curves of the migration model and the non-migration model described in this embodiment of the invention; Figure 11 The two sets of models described in this embodiment of the invention are the reconstruction results on a portion of the samples, wherein... Figure 11 (a) shows the reconstruction results of the transfer model on a subset of samples. Figure 11 (b) shows the reconstruction results of the non-transfer model on a subset of samples; Figure 12This is a comparison of the lateral average results of the two models described in this embodiment of the invention on a portion of the samples. Figure 12 (a) Comparison of lateral average results of the transfer model on a subset of samples. Figure 12 (b) Comparison of lateral averaging results for a subset of samples using the no-transfer model; Figure 13 The PSNR and SSIM distributions of the two models described in this embodiment of the invention are shown on the entire dataset, wherein... Figure 13 (a) shows the PSNR distributions of the transfer learning model and the non-transfer learning model on the dataset. Figure 13 (b) shows the SSIM distribution of the transfer model and the non-transfer model on the dataset. Detailed Implementation

[0022] The specific implementation details (such as experimental setup, operation procedures, data processing steps, and example parameters) of the "Variable Fidelity Modeling Method for Turbine Film Cooling Based on Domain Adaptive Migration" provided in this specification are primarily intended for illustrative purposes rather than limiting definitions, aiming to help those skilled in the art thoroughly understand the principles and implementation of the invention. However, those skilled in the art should understand that these details represent only one feasible embodiment, and the core concept of the invention can be fully realized through other technical means or alternative solutions not described in detail, without departing from its spirit and essence. Furthermore, the omission of details of conventional experimental methods and apparatus known in the art in the specification is to avoid redundant information interfering with the understanding of the innovation points. This does not mean that these known technologies are not required during implementation, and those skilled in the art should be able to supplement and apply them based on their professional knowledge.

[0023] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

[0024] Example 1, Combination Figure 1 This embodiment addresses the challenge that the performance of deep learning-based film cooling performance prediction models is severely limited by the quality and quantity of training data. This leads to a dilemma: high-fidelity data is scarce due to its high cost, and using it solely for training can easily result in overfitting and poor generalization ability; while massive amounts of low-fidelity data lack accuracy, making the prediction results of models trained with it unreliable. Therefore, traditional single-fidelity modeling methods consistently struggle to achieve a balance between data cost and model performance. To address this, this embodiment proposes a variable-fidelity modeling method for turbine film cooling based on domain adaptive transfer, comprising the following steps: Input data representation steps: For the input data, accurately represent the complex array geometry of turbine film cooling based on lexical units; The steps for constructing a parameter mapping network are as follows: Construct a parameter list mapping multi-path attention network and construct a gradient loss function for the network; The progressive modeling steps based on transfer learning are as follows: The parameter mapping network in step 2 is trained with low-fidelity data to learn the basic physical characteristics of film cooling flow. The learnable parameters of the low-fidelity parameter mapping network are used as prior information and as the initial parameters of the high-fidelity prediction model. The prediction model with initial parameters is fine-tuned and trained with high-fidelity data to obtain a high-fidelity film cooling efficiency rapid prediction model.

[0025] The flowchart of the variable fidelity modeling method for turbine film cooling based on domain adaptive migration proposed in this embodiment is as follows: Figure 1 As shown, it includes three core parts: (1) establishing a parameterized expression system with a single air film pore as the modeling unit to achieve accurate description of complex air film pore geometry; (2) constructing a multi-scale, multi-path self-attention network structure combined with gradient loss to improve the model's reconstruction accuracy for complex flow fields with multiple local high gradients; (3) introducing a transfer learning mechanism to form a progressive modeling strategy for data of different fidelity levels, so as to achieve efficient generalization of the model in multi-fidelity scenarios.

[0026] Example 2, Combination Figure 2 and Figure 3 This embodiment is a detailed explanation of the variable fidelity modeling method for turbine film cooling based on domain adaptive migration described in Embodiment 1 above. Input data representation steps: For the input data, accurately represent the complex array geometry of turbine film cooling based on lexical units; Specifically: For the input data, we choose the concept of lexical units from natural language processing problems, treating each 'hole' as an abstract character. Treating the set of all holes as a short sentence, forming ,in N The number of air film pores for each array layout, d The number of input features includes geometric parameters, dimensionless flow direction, and position. Dimensionless expansion position Inclination angle θ Composite angle φ aperture D Boundary condition parameter: air blowing ratio BR Compared to temperature TRThe output data is a dimensionless temperature field of the blade surface. Surface temperature field data is extracted from the CFD results using a uniform grid. The turbine blade surface is uniformly divided into 256 sections along the spanwise direction, and 256 sampling points are uniformly placed along the blade's arc curve on both the suction and pressure sides of each section. The final labeled blade surface temperature field is organized into 256×512 image data.

[0027] Furthermore, to eliminate the influence of different input feature dimensions and accelerate convergence, the input parameters were normalized, with all parameters scaled to [0, 1]. The output value is a dimensionless temperature field, which, since it is in the range of [0, 1], does not require normalization.

[0028] The steps for constructing a parameter mapping network are as follows: Construct a parameter list mapping multi-path attention network and construct a gradient loss function for the network; Specifically: The parameter mapping network comprises an upstream boundary and geometry processing module and a downstream multi-path flow field prediction module. The upstream module receives the parameter list encoding and obtains the global features of the current film vent layout through latent vector calculation and spatial weight allocation for each vent. The three subsequent paths receive global features of size 128×16×32 and assign different flow data sizes to each path via a PixelShuffle upsampling module. The core computational unit for each of the three paths is two cascaded Residual Sliding Window Attention Network (RSTB) modules, each RSTB containing six SwinT sliding window attention computation units. Multi-scale features are fused using an element-wise summation aggregation function, and finally, the predicted cooling efficiency distribution field value is obtained through another PixelShuffle upsampling module and a convolutional output module. The working data sizes for different paths are 16×32, 32×64, and 64×128, respectively, and the number of RSTB modules used for each branch is 6, 4, and 2, respectively. The neural network architecture is as follows: Figure 2 As shown.

[0029] Furthermore, the loss function adopts a composite loss of mean square error and pixel error, as shown in Equation (1). In this loss function, the weight allocation adopts an adaptive calculation method, which is autonomously allocated as the loss propagates back during the iteration process.

[0030] (1) (2) (3) In equation (1), Different loss terms can be represented by MSE and gradient in this embodiment; These are learnable parameters used to calculate the loss weights for each iteration; Represents the temperature field of the reference wall. Represents the predicted temperature field. H and W These represent the number of sampling points in the spanwise and chordwise directions of the target field, respectively.

[0031] Furthermore, the network and necessary environment are implemented based on the PaddlePaddle framework. During training, data is input in mini-batches with a batch size of 16. The Adam algorithm is used to train the model. An exponential learning rate scheduler is employed, with an initial learning rate of 1×10⁻⁶. -4 The decay factor is 0.9, meaning that the learning rate steps to 0.9 times the previous value every 250 epochs.

[0032] The progressive modeling steps based on transfer learning are as follows: The parameter mapping network in step 2 is trained with low-fidelity data to learn the basic physical characteristics of film cooling flow. The learnable parameters of the low-fidelity parameter mapping network are used as prior information and as the initial parameters of the high-fidelity prediction model. The prediction model with initial parameters is fine-tuned and trained with high-fidelity data to obtain a high-fidelity film cooling efficiency rapid prediction model.

[0033] Specifically: Numerical simulations and experimental data at different fidelity levels for the same film cooling layout scheme describe the same physical process at different degrees of approximation, exhibiting similar overall evolutionary trends and patterns. High-fidelity data is often scarce, insufficient in scale to train a high-fidelity rapid prediction model for cooling efficiency with a wide range of applicable parameters; while low-fidelity data can be obtained at low cost. Therefore, this embodiment trains the model with a large amount of low-fidelity data, enabling it to learn the basic characteristics of turbine film cooling flow heat transfer problems; these learnable parameters of the model are then used as prior information and transferred to a new high-fidelity dataset for fine-tuning, reducing the dependence of the air-heat coupling prediction model on data and the number of iterations required for training to convergence. Based on the transfer learning method, this embodiment proposes a feasible strategy, such as... Figure 3 As shown. Traditional neural network training methods use random number initialization to assign initial values ​​to the network's learnable parameters, while transfer learning methods use the parameter set of the trained air-film cooled neural network ModelPara as the initial value, providing a small amount of label data from air-thermal coupling calculations to fine-tune the model, achieving efficient generalization of the model in multi-fidelity scenarios. The initial learning rate is set to 1×10. -5 The loss function is a combination of MSE and pixel gradient loss. The weight allocation uses an adaptive function, which is optimized along with the loss term in each iteration. All other configurations remain completely consistent with the above.

[0034] Example 3: This example provides a specific implementation step to illustrate the actual operation of the variable fidelity modeling method for turbine film cooling based on domain adaptive migration described in Example 1 or 2 above.

[0035] 1. Data Preparation In practical applications, it is generally advisable to prepare two datasets with different fidelity levels. Low-fidelity datasets are typically larger in size, but due to model simplifications and ideal assumptions made during acquisition, their description of real physical processes often deviates significantly. High-fidelity data, on the other hand, is usually acquired from high-precision experimental measurements and numerical calculations, providing an accurate description of real physical processes; however, due to higher acquisition costs, its size is usually smaller. Empirically, the size of high-fidelity data should be no less than 10% of that of low-fidelity data.

[0036] The air film pore layout is parameterized and organized into a parameter list. Input parameters include dimensionless flow direction position. Dimensionless expansion position Inclination angle θ Composite angle φ aperture D blower BR Compared to temperature TR The output data is a dimensionless temperature field of the blade surface. Surface temperature field data is extracted from the CFD results using a uniform grid. The turbine blade surface is uniformly divided into 256 sections along the spanwise direction, and 256 sampling points are uniformly placed along the blade's arc curve on both the suction and pressure sides of each section. The final blade surface temperature field, used as a label, is organized into 256×512 image data. The input and output data are normalized, scaling the range to [0, 1].

[0037] 2. Low-fidelity model training The network uses low-fidelity data, and all associated code is written in Python, built on the open-source deep learning framework PaddlePaddle. Training the network on a computer platform equipped with specific hardware is recommended, typically using the Adam optimizer with default parameters.

[0038] 3. Performance verification of the low-fidelity model After the model is completed, its performance on the dataset is evaluated using mean squared error, relative error, peak signal-to-noise ratio (PSNR), and structured similarity criterion. Statistical indicators are compiled into a results table. Contour plots of the dimensionless temperature field for predicted and actual values, as well as violin plots of the PNR and structured similarity criterion distributions, are then plotted. Finally, lateral mean dimensionless temperature curves are drawn.

[0039] 4. High-fidelity model training High-fidelity data was used, with low-fidelity model parameters serving as prior information and initial values ​​for optimization and training. The network and its associated code were all written in Python and built on the open-source deep learning framework PaddlePaddle. The network was trained on a computer platform equipped with specific hardware; the Adam optimizer was typically recommended, using default parameters.

[0040] 5. Model Prediction Once the model is trained, it can predict the wall temperature field for the new cooling layout scheme. Only normalized experimental data needs to be provided; importing this data into the model will yield the corresponding experimental boundary conditions. The normalization index should utilize statistical characteristics of the data from both low-fidelity and high-fidelity datasets.

[0041] Example 4, Combination Figures 4 to 13 This embodiment further illustrates the beneficial effects of the variable fidelity modeling method for turbine film cooling based on domain adaptive migration described in this invention through simulation. First, a comparison was made with existing image mapping models. Both models were trained for 10,000 epochs. The proposed parameter list mapping model (ModelPara) and the image mapping model used as a baseline (ModelImg) converged to 6.0 × 10⁻⁶ on the training set. -5 and 1.2×10 -4 It converged to 4.9 × 10 on the test set. -4 and 5.3×10 -4 Table 1 (Statistical metrics of ModelPara and ModelImg on the training and test sets) summarizes the different statistical metrics of the two sets of models.

[0042] Table 1

[0043] Figure 4 and Figure 5 The paper presents the reconstructed dimensionless temperature field contour maps of two mapping models on partial samples of the training and test sets. The samples shown are those with the mean squared error closest to the median of their respective datasets, as these best represent the general performance of the models on those datasets. The first row of each subplot displays CFD reference values, the middle row displays the model prediction results, and the third row displays the relative error distribution. The relative error contour maps show that the errors in the reconstructed temperature field mainly exist near the outlets of each pore and near the boundaries of the film gas core region. The temperature gradient in this region is relatively large, which is unfavorable for accurate fitting by the convolutional network. However, the introduction of the attention mechanism and pixel gradient loss in the proposed method alleviates this problem.

[0044] It can be seen that ModelPara exhibits a more significant advantage over ModelImg on the training set. For the samples shown, considering the maximum relative error, the ModelPara reconstructed field is 50% of the ModelImg reconstructed field, and the former also shows fewer error concentration areas overall. On the test set, the performance difference between the two models is no longer significant. Although, for the samples shown, the maximum relative error of ModelPara is 1.5 times that of ModelImg, the former generally has lower values ​​in most areas and fewer high error concentration areas.

[0045] Secondly Figure 6 and Figure 7 The curves showing the variation of the lateral mean dimensionless temperature along the chord direction for the corresponding samples are further illustrated, where the region [-1, 0] represents the area from the pressure-side trailing edge to the leading edge, and the region [0, 1] represents the area from the leading edge to the suction-side trailing edge. It can be seen that although most film cooling vents are located on the pressure-side surface, resulting in a more complex temperature field distribution, both models show a high degree of fit with the CFD reference values. Overall, the image mapping results show more significant errors. The reason for this can still be traced back to the accuracy of the two mapping methods in reconstructing the geometry and location distribution of the film cooling vents. Compared to flat-plate film cooling, turbine blade film cooling has a more pronounced impact in two aspects: firstly, the wall curvature stretches the projected profile of the cold air outlet; secondly, the cross-sectional area of ​​the film cooling vent outlet differs significantly from the blade surface scale, hindering the accurate representation of the blade surface geometry.

[0046] Furthermore, the input symbolic distance function field of the image mapping is based on simplified geometry, which to some extent ignores the influence of wall curvature. This is one of the reasons for the inaccurate geometric characterization of the air film pores in ModelImg. Figure 4 The relative error contour maps of (b) and 5(b) also reflect this, showing that the predicted field exhibits extremely high errors at some air film aperture outlet locations, significantly different from the surrounding pixels. In contrast, parameter mapping seems more suitable for this scenario, allowing the neural network to autonomously learn the mapping relationship between geometric parameters and the actual shape and spatial location of the air film aperture through abstract expression. Although limited by the model's own learning ability, the geometric recognition of some apertures is still inaccurate, such as... Figure 4 (a) and 5(a), but their number and proportion are lower than those of image mapping. On the other hand, due to the limitation of computational cost, the sampling point density on the blade surface is limited, and the number of pixels occupied by the air film pore region is very small. Some pores can only be allocated 2 to 3 pixels. The characterization of the flow near the pore outlet is actually distorted, resulting in a high gradient between the pixel values ​​in these regions and the surrounding areas. This is not conducive to fitting by neural networks based on convolutional computation.

[0047] Finally, the training time of each model in the turbine cooling temperature field reconstruction task was compared, as summarized in Table 2 (CFD solution and computation and prediction time of the four models). The results show that increasing the number of RSTB modules in each branch increases model complexity. The time cost of network optimization is determined by the three branch paths. At this point, the time consumption of the two mapping strategies is similar, while parameter mapping is slightly longer. This is because the number of film cooling pores is much greater in turbine cooling than in plate cooling, increasing the computational cost in the geometric feature extraction stage. However, once training is complete, the prediction time of the model is almost negligible compared to CFD computation, giving it a significant advantage in the early stages of rapid optimization design for turbine cooling.

[0048] Table 2

[0049] After the low-fidelity model is trained, the effectiveness of the proposed transfer learning-based fidelity-decreasing modeling method in reducing retraining costs is demonstrated. Keeping the model configuration and training hyperparameters completely consistent, five training sets were set up according to different allocation criteria, with quantities of 10, 40, 80, 160, and 320 sets, representing 2.8%, 11.1%, 22.2%, 44.4%, and 88.9% of the model, respectively. The remainder was used as the test set. First, the impact of the training set ratio on model convergence is shown, such as... Figure 8 As shown, each model was trained for 10,000 epochs. As the proportion of the training set increased, the accuracy of the model improved on both the training and test sets, with a more significant improvement on the test set.

[0050] When the model hyperparameters remain consistent, the model's learning ability remains unchanged. Therefore, most of the training samples converge to similar accuracy without significant differences. The model's accuracy on the test set reflects its generalization level. Except for the scheme with only 10 training sets, all other schemes maintain an accuracy of 10. -4 The convergence residuals below the order of magnitude indicate that the transfer learning model has high generalization ability.

[0051] Furthermore, the number of iterations required for the model to fully converge on the test set appears to increase with the increase in the proportion of the training set. When the number of training samples is small, the model reaches the minimum residual on the test set after only about 200 iterations, followed by a slight increase in loss, exhibiting obvious overfitting. However, when the training data is increased to 40 sets, the model no longer overfits and converges after about 500 iterations. This indicates that transfer learning models can reduce the dependence of new models on training data, achieving comparable accuracy with fewer iterations.

[0052] Figure 9This demonstrates the response of the model's average PSNR and SSIM across the entire dataset to the proportion of the training set. As the proportion of samples included in the training set increases, the model's prediction accuracy across the entire dataset gradually improves, exhibiting a monotonic increase in both metrics. Simultaneously, it can be observed that in the early stages of increasing the training set proportion, the gradient of the model's accuracy response curve is high; however, as the training set gradually covers the entire dataset, the growth rate gradually levels off, indicating that increasing the training set proportion has a diminishing marginal effect on improving model accuracy. In fact, from... Figure 9 (b) shows that when the proportion of training data reaches more than 22%, further increasing the amount of training data has little value for the final performance of the model.

[0053] To further demonstrate this conclusion, Table 3 (Accuracy and Training Costs with Different Input Data Ratios) summarizes the training time cost, model convergence steps, and model accuracy. When the training data ratio increases to 22%, the accuracy gain achieved when the model reaches convergence is essentially balanced with the time cost. Further increasing the number of training samples increases both the computational cost of data acquisition and the cost-effectiveness of model training. Considering both the adjusted training time and the accuracy achieved by the model, a model with a training set ratio of 22.2% will be used in subsequent analyses.

[0054] Table 3

[0055] Further analysis of the multi-dimensional performance of the transfer learning model was conducted by training a new model from scratch using the same dataset, termed the non-transfer learning model. Table 4 summarizes the statistical metrics of the transfer learning model and the non-transfer learning model. Overall, both models achieved accurate predictions of the air-thermal coupling temperature field. However, the transfer learning model demonstrated a significant advantage on both the training and test sets, outperforming the non-transfer learning model across all evaluation metrics.

[0056] Table 4

[0057] Figure 10 This demonstrates the impact of introducing transfer learning on model convergence. The transfer learning model converged on the test set after only about 700 iterations, while the model without transfer learning required approximately 4000 iterations to achieve basic convergence. Furthermore, it was observed that the residual of the transfer learning model on the training set was consistently lower than that of the model without transfer learning after the same number of iterations. This indicates that introducing transfer learning can effectively reduce the training cost of the model and achieve comparable or even better reconstruction results with less training data. Observing the convergence curve on the test set, the transfer learning model effectively reduced the degree of overfitting on the same dataset size.

[0058] Figure 11 The visualization results of the temperature field reconstructions of partial samples by two models are presented, showing the samples whose mean squared errors are closest to the median of their respective datasets. Since transfer learning itself requires less training data in the target domain and focuses more on the model's adaptability to the target domain sample space, the training and test sets are no longer distinguished; instead, the performance level across the entire dataset is considered comprehensively. The results show that both models achieve high accuracy in reconstructing the temperature field of the air-thermal coupled wall. In terms of relative error distribution, the maximum relative error of the transfer model is approximately 14%, with an average of approximately 0.7%. In contrast, the maximum relative error of the non-transfer model reaches 32%, with an average of approximately 1.5%, exhibiting a more significant error, mainly distributed at the outlets of the cold air vents.

[0059] Figure 12 The lateral average cooling efficiency curves for the corresponding samples are shown. Overall, the predicted values ​​still maintain a high degree of fit with the reference values, but there is a more significant deviation compared to the film cooling prediction model. This is mainly due to the difference in training set size; the film cooling model used 89% of the data, while the air-thermal coupling model of this invention only used 22% of the data.

[0060] Figure 13 The PSNR and SSIM distributions of the two models across all data are presented. Corroborating previous analyses, the transfer learning model demonstrates higher overall reconstruction accuracy on the dataset than the non-transfer learning model. Figure 13 (b) shows a more pronounced SSIM distribution, with the transfer model results exhibiting a tighter distribution and a higher mean. In contrast, Figure 13 The PSNR distribution shown in (a) is relatively discrete, regardless of whether it is a transfer model or a non-transfer model, mainly due to the small size of the training set.

[0061] Example 5: The turbine film cooling variable fidelity modeling method based on domain adaptive migration described in any of the above embodiments can be fully implemented using computer software. Therefore, correspondingly, this embodiment provides a turbine film cooling variable fidelity modeling system based on domain adaptive migration, the system comprising: A storage device for the accurate representation of complex array geometry for lexical-based turbine film cooling; A storage device for constructing a parameter list mapping multipath attention network and for constructing a gradient loss function for the network; A storage device is used to train the parameter mapping network with low-fidelity data so that it learns the basic physical characteristics of film cooling flow, and uses the learnable parameters of the low-fidelity parameter mapping network as prior information as the initial parameters of the high-fidelity prediction model; and uses high-fidelity data to fine-tune the prediction model with the initial parameters to obtain a high-fidelity film cooling efficiency rapid prediction model.

[0062] Example 6: This example provides a computer-readable storage medium storing a computer program. When the computer program is run by a processor, it executes the variable fidelity modeling method for turbine film cooling based on domain adaptive migration described in any of the above examples.

[0063] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be accomplished by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0064] Example 7: This example provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the variable fidelity modeling method for turbine film cooling based on domain adaptive migration described in any of the above examples.

[0065] This embodiment provides a computer device. The hardware device in this part is a general model and is not shown in the figure. The system includes a processor and a memory. The processor and the memory can be connected by a bus or other means. The memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs and modules, as well as corresponding program instructions / modules. The processor executes various functional applications and data processing by running the non-transitory software programs, instructions and modules stored in the memory, so as to realize the turbine film cooling variable fidelity modeling method and steps based on domain adaptive migration in the above method embodiment.

[0066] The memory may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, mobile communication networks, and combinations thereof.

[0067] One or more modules are stored in the memory. When the processor executes, it performs the method steps in the embodiments. In this way, the invention objective can be achieved through the method, apparatus and process of the present invention. The specific details of the computer device described above can be understood by referring to the relevant descriptions and effects in the embodiments, and will not be repeated here.

[0068] The above description of the technical solution provided by the present invention through several specific embodiments is intended to highlight the advantages and benefits of the technical solution provided by the present invention. However, the above-described specific embodiments are not intended to limit the present invention. Any reasonable modifications and improvements to the present invention, reasonable combinations of implementation methods and equivalent substitutions based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A variable-fidelity modeling method for turbine film cooling based on domain adaptive migration, characterized in that, The methods include: Input data representation steps: For the input data, accurately represent the complex array geometry of turbine film cooling based on lexical units; The steps for constructing a parameter mapping network are as follows: Construct a parameter list mapping multi-path attention network and construct a gradient loss function for the network; The progressive modeling steps based on transfer learning are as follows: The parameter mapping network in step 2 is trained with low-fidelity data to learn the basic physical characteristics of film cooling flow. The learnable parameters of the low-fidelity parameter mapping network are used as prior information and as the initial parameters of the high-fidelity prediction model. The prediction model with initial parameters is fine-tuned and trained with high-fidelity data to obtain a high-fidelity film cooling efficiency rapid prediction model.

2. The variable fidelity modeling method for turbine film cooling based on domain adaptive migration according to claim 1, characterized in that, The input data is specifically expressed as follows: Based on the lexical units in natural language processing, each hole is treated as an abstract character. ; Treating the set of all holes as a short sentence forms... ,in N The number of air film pores for each array layout, d The number of input features includes geometric parameters, dimensionless flow direction, and position. Dimensionless expansion position Inclination angle θ Composite angle φ aperture D Boundary condition parameter: air blowing ratio BR Compared to temperature TR ; The output data is a dimensionless temperature field of the blade surface. The surface temperature field data is extracted from the CFD results using a uniform grid. The turbine blade surface is uniformly divided into 256 sections along the spanwise direction. 256 sampling points are uniformly placed along the blade arc curve on the suction and pressure sides of each section. Finally, the blade surface temperature field, which is used as a label, is organized into 256×512 image data.

3. The variable fidelity modeling method for turbine film cooling based on domain adaptive migration according to claim 2, characterized in that, The input data is normalized, and all data is scaled to [0, 1].

4. The variable fidelity modeling method for turbine film cooling based on domain adaptive migration according to claim 1, characterized in that, The parameter list maps the multipath attention network to include an upstream boundary and geometry processing module and a downstream multipath flow field prediction module. The upstream module is used to receive the parameter list encoding and obtain the global features of the current air film aperture layout through latent vector calculation and spatial weight allocation of each aperture. The downstream module is used to receive global features of size 128×16×32 through three channels, and allocates different flow data sizes to the three channels through the PixelShuffle upsampling module.

5. The variable fidelity modeling method for turbine film cooling based on domain adaptive migration according to claim 4, characterized in that, The core computational unit of the three pathways is two cascaded residual sliding window attention network modules (RSTB). Each RSTB contains six sliding window attention computation units (SwinT). Multi-scale features are fused through an element-wise summation aggregation function. Finally, the predicted value of the cooling efficiency distribution field is obtained through a PixelShuffle upsampling module and a convolutional output module.

6. The variable fidelity modeling method for turbine film cooling based on domain adaptive migration according to claim 4, characterized in that, The working data sizes for the three paths are 16×32, 32×64 and 64×128, respectively, and the number of RSTB modules used in each branch is 6, 4 and 2, respectively.

7. The variable fidelity modeling method for turbine film cooling based on domain adaptive migration according to claim 1, characterized in that, The gradient loss function uses a composite loss of mean squared error and pixel error. In this loss function, the weight allocation adopts an adaptive calculation method and is autonomously allocated as the loss propagates back during the iteration process.

8. A variable-fidelity modeling system for turbine film cooling based on domain adaptive migration, characterized in that, The system is implemented based on the variable fidelity modeling method for turbine film cooling based on domain adaptive migration as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the variable fidelity modeling method for turbine film cooling based on domain adaptive migration as described in any one of claims 1-7.

10. A computer device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the variable fidelity modeling method for turbine film cooling based on domain adaptive migration as described in any one of claims 1-7.

Images