A deep learning-based method for predicting flow field in a nozzle

Abstract

This invention relates to a deep learning-based method for predicting the flow field inside a nozzle, comprising: constructing a variable-configuration nozzle geometry library based on parametric curves and performing CFD calculations to generate a flow field dataset; performing dual-channel encoding processing of the geometric boundaries using the signed distance field (SDF) and the identifier matrix (IM); constructing a deep learning model based on a U-Net backbone network with an embedded dual-ended convolutional compression excitation (DE-ConvSE) adapter; pre-training the model using a basic dataset; employing an incremental learning strategy for new nozzle configurations, freezing the backbone network parameters and fine-tuning only the adapter module; and finally, using the trained model to achieve real-time prediction of the velocity, temperature, and pressure fields inside the nozzle. This invention can significantly reduce CFD simulation computation time, achieve rapid and high-precision prediction of the nozzle flow field, and provide effective technical support for nozzle design optimization.

Description

Technical Field

[0001] This invention belongs to the fields of computational fluid dynamics and artificial intelligence, and particularly relates to a method for predicting the flow field inside a nozzle based on deep learning. Background Technology

[0002] As a core component of an engine, the nozzle plays a crucial role in energy conversion and thrust generation. To obtain flow field data inside the nozzle and determine the optimal parameters for engine operation, engineers typically employ methods such as computational fluid dynamics (CFD) simulations to evaluate the nozzle's performance under different inlet conditions. Furthermore, accurate prediction of the nozzle flow field is of great significance for nozzle structural optimization design and other related aspects.

[0003] While CFD simulation is widely used due to its accuracy, its high computational cost and complex computational process remain challenges in acquiring flow field data. In recent years, with the rise of deep learning, artificial intelligence has offered new approaches to solving traditional computational fluid dynamics problems. Compared to traditional CFD methods, deep learning offers advantages such as shorter processing time, batch processing capability, and no need for mesh generation, providing a more efficient and economical solution for nozzle optimization design problems.

[0004] In recent years, deep learning (DL) has provided new insights into solving fluid mechanics problems. Existing flow field prediction methods are mostly based on convolutional neural networks (CNNs), physical information neural networks (PINNs), or neural operators (FNOs). However, existing technologies still have significant shortcomings:

[0005] (1) Limited generalization ability: Traditional image-based deep learning models are usually trained under fixed geometric parameters, which makes it difficult to cope with variable configuration scenarios where geometric parameters such as nozzle throat diameter and expansion ratio change significantly;

[0006] (2) Difficulty in capturing shock waves: The flow field in the nozzle often contains complex discontinuous flow structures such as shock waves and expansion waves. Conventional networks are difficult to accurately capture the location and intensity of local shock waves while ensuring the correct global trend.

[0007] (3) Catastrophic forgetting problem: When the model needs to adapt to a new nozzle configuration, if full-fine-tuning is used, the model will often "forget" the flow field characteristics of the old configuration, resulting in a significant decrease in the prediction accuracy of the original task, which cannot meet the needs of continuous learning and design iteration.

[0008] Therefore, there is an urgent need for a real-time flow field prediction method that can adapt to variable nozzle configurations, balance computational efficiency and accuracy, and effectively overcome catastrophic forgetting. Summary of the Invention

[0009] To address the issues of long computation time and difficulty in meeting the requirements of large-scale iteration in existing computational fluid dynamics (CFD) simulations for nozzle design, as well as the shortcomings of existing deep learning methods in terms of poor generalization ability and catastrophic forgetting of fine-tuned models when faced with nozzles with variable configurations, this invention provides a deep learning-based method for predicting the flow field inside the nozzle. This method can not only significantly reduce the consumption of computational resources, but also quickly adapt to new geometric configurations through incremental learning, achieving high-precision and real-time prediction of the flow field inside the nozzle, providing efficient technical support for the structural optimization design of engine nozzles.

[0010] To achieve one of the above-mentioned objectives, one embodiment of the present invention provides a method for predicting the flow field inside a nozzle based on deep learning, the method comprising the following steps:

[0011] S1. Construct a dataset and perform preprocessing. The dataset consists of several sets of nozzle geometric features and corresponding temperature field, velocity field and pressure field data.

[0012] S2. Construct a multi-physics prediction network based on the U-Net architecture. The multi-physics prediction network includes a shared encoder and three independent decoder branches, and a double-ended convolutional compression excitation adapter module is embedded after each convolutional block of the shared encoder and decoder branches.

[0013] S3. Train the multiphysics prediction network based on the dataset preprocessed in step S1.

[0014] S4. Using the multiphysics prediction network trained in step S3, perform flow field prediction for the nozzle under test.

[0015] As a further improvement to one embodiment of the present invention, the dataset construction method in step S1 includes:

[0016] S11. Several nozzle geometries with different expansion profiles are generated using parametric modeling methods.

[0017] S12. For each nozzle geometry generated in step S11, use computational fluid dynamics (CFD) to perform flow field simulation and obtain the corresponding temperature field, velocity field and pressure field data.

[0018] S13. Calculate the minimum directed distance from each grid point in the computational domain to the nozzle wall, and construct the symbolic distance field;

[0019] S14. Construct an identification matrix with the same resolution as the computational domain grid, where each element in the identification matrix represents the physical meaning of the corresponding region;

[0020] S15. The symbolic distance field (SDF) and the identification matrix are spliced ​​together in the channel dimension to form the nozzle geometry.

[0021] S16. The geometric features of each nozzle and the corresponding temperature field, velocity field and pressure field data form a set of "geometry-flow field" samples, which in turn form a dataset.

[0022] As a further improvement to one embodiment of the present invention, step S2, the method for building a multiphysics prediction network based on the U-Net architecture, includes:

[0023] S21. A multi-layer convolutional neural network is used as a shared encoder, with nozzle geometric features as input.

[0024] S22. For the temperature field, velocity field and pressure field, three independent decoder branches are constructed, and each decoder branch is connected to the encoder in a skip connection.

[0025] As a further improvement to one embodiment of the present invention, the loss function of the multiphysics prediction network is:

[0026] ,

[0027] ,

[0028] ,

[0029] ,

[0030] in, , and Let represent the loss functions of the velocity field, temperature field, and pressure field channels, respectively. , and These represent the positions in the nth sample. The predicted values ​​for the velocity field, temperature field, and pressure field are given. , and respective positions The model represents the CFD simulation results of the velocity field, temperature field, and pressure field. These represent the weights of the velocity field, temperature field, and pressure field, respectively.

[0031] As a further improvement to one embodiment of the present invention, The adaptive acquisition method is as follows:

[0032] ,

[0033] Where c = 1, 2, 3, This represents the mean of channel c. These represent the positions in the nth sample. The true values ​​of the velocity field, temperature field, and pressure field at that location. These represent the central locations in the dataset. The mean of the true values ​​of the velocity field, temperature field, and pressure field; N, H, and W represent the batch size of the dataset, the height of the nozzle, and the width of the nozzle, respectively.

[0034] As a further improvement to one embodiment of the present invention, step S2, the method for building a multiphysics prediction network based on the U-Net architecture, further includes:

[0035] S23. Embed a dual-ended convolutional compression excitation adapter module after each convolutional block in the shared encoder and decoder branches, and perform residual connections; the dual-ended convolutional compression excitation adapter module includes a convolutional stack layer and a compression-excitation SE module, and introduces learnable gating coefficients.

[0036] As a further improvement to one embodiment of the present invention, the method further includes:

[0037] If the geometry of the nozzle under test is not included in the database, an incremental learning strategy is adopted to construct an incremental dataset, and only the bi-ended convolutional compressed excitation adapter module in the multiphysics prediction network is learned.

[0038] As a further improvement to one embodiment of the present invention, the method further includes:

[0039] S5. Based on the flow field prediction results of step S4, obtain the specific impulse of the nozzle under test:

[0040] ,

[0041] ,

[0042] In the formula, I represents the specific impulse of the nozzle under test, and F represents the thrust of the nozzle under test. This represents the mass flow rate inside the nozzle being tested. For gas density, The axial component of the nozzle exit velocity to be measured is denoted as . , These represent the pressure at the nozzle exit and the atmospheric pressure, respectively; S represents the nozzle exit cross-sectional area.

[0043] To achieve one of the above-mentioned objectives, an embodiment of the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and when the program is executed on the processor, it implements the steps in the deep learning-based nozzle in-flow field prediction method described above.

[0044] To achieve one of the above-mentioned objectives, an embodiment of the present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps in the deep learning-based nozzle flow field prediction method described above.

[0045] Compared with existing technologies, the deep learning-based nozzle flow field prediction method of the present invention has the following significant advantages:

[0046] By constructing a U-Net network with an embedded dual-ended convolutional compressed excitation adapter and adopting an incremental learning strategy of freezing the backbone and fine-tuning only the adapter, the catastrophic forgetting problem in the prediction of variable-configuration nozzles is effectively solved. Only a small number of samples are needed to achieve high-precision adaptation of new configurations. At the same time, the dynamic loss scaling mechanism balances the numerical differences of multi-physics fields. While ensuring the accuracy of detailed features such as shock waves, it achieves real-time prediction that is two orders of magnitude faster than traditional CFD simulation, providing low-cost and high-efficiency technical support for the rapid optimization design of nozzles. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the implementation of the present invention.

[0048] Figure 2 This is a schematic diagram of an engine nozzle according to an embodiment of the present invention.

[0049] Figure 3 This is a schematic diagram of the CFD mesh and key regions implemented in this invention.

[0050] Figure 4 This is a schematic diagram of the dataset sampling points in the implementation of this invention.

[0051] Figure 5 This is a schematic diagram of the symbolic distance field (SDF) of this invention.

[0052] Figure 6 This is a schematic diagram of the identifier matrix (IM) implemented in this invention.

[0053] Figure 7 This is a schematic diagram of the DE-ConvSE U-Net neural network architecture of the present invention.

[0054] Figure 8 This is a schematic diagram of a single branch structure of the DE-ConvSE U-Net neural network of the present invention.

[0055] Figure 9This is a schematic diagram of the flow field inside the nozzle under under-expansion using a deep learning model implemented in this invention. (a) shows the predicted temperature field inside the nozzle under under-expansion conditions, (b) shows the CFD simulation results of the temperature field inside the nozzle under the corresponding boundary conditions, (c) shows the predicted velocity field inside the nozzle under under-expansion conditions, (d) shows the CFD simulation results of the velocity field inside the nozzle under the corresponding boundary conditions, (e) shows the predicted pressure field inside the nozzle under under-expansion conditions, and (f) shows the CFD simulation results of the pressure field inside the nozzle under the corresponding boundary conditions.

[0056] Figure 10 This is a schematic diagram of the flow field inside the nozzle under over-expansion using a deep learning model implemented in this invention. (a) shows the predicted temperature field inside the nozzle under over-expansion, (b) shows the CFD simulation result of the temperature field inside the nozzle under the corresponding boundary condition, (c) shows the predicted velocity field inside the nozzle under over-expansion, (d) shows the CFD simulation result of the velocity field inside the nozzle under the corresponding boundary condition, (e) shows the predicted pressure field inside the nozzle under over-expansion, and (f) shows the CFD simulation result of the pressure field inside the nozzle under the corresponding boundary condition. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present invention, and any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the scope of protection of the present invention.

[0058] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0059] This embodiment discloses a deep learning-based method for predicting the flow field inside a nozzle, such as... Figure 1 As shown, it includes the following steps:

[0060] S1. Dataset construction and preprocessing based on CFD simulation calculations: The specific steps are as follows:

[0061] S11. Determine the pressure and temperature at the nozzle outlet, change the gas temperature and pressure at the nozzle inlet within a certain range, combine the boundary conditions at the nozzle inlet, and form a sample library based on this combination.

[0062] In this embodiment, a Laval nozzle is selected as the basic configuration, such as... Figure 2 As shown, in order to construct a dataset covering flow fields with various geometric configurations, a parametric modeling method is used to design the variable configuration of the nozzle's expansion section.

[0063] S11. Nozzle Geometry Parameterization and Sampling: To achieve diverse nozzle geometries, this embodiment uses a fourth-order Bézier curve to parameterize the nozzle's expansion section profile. The Bézier curve formula is as follows:

[0064] ,

[0065] Among them, P i Indicates the first One control point, These are curve parameters. For example... Figure 3 As shown, this embodiment selects five control points (P0, P1, P2, P3, P4) to control the curve shape. The positions of the starting point P0 (throat) and the ending point P4 (exit) are fixed or constrained within a small range according to design requirements, while the three middle control points (P1, P2, P3) are the main design variables. Using the Latin hypercube sampling method, these control points are uniformly sampled within a preset design space, generating a large number of nozzle geometries with different expansion profiles, thereby constructing a basic geometric library with sufficient geometric diversity.

[0066] S12. Mesh Generation and CFD Simulation: For each nozzle geometry generated in S11, flow field simulation is performed using computational fluid dynamics (CFD) software, specifically including the following steps:

[0067] S121. Mesh Generation: The computational domain is divided using a structured quadrilateral mesh, such as... Figure 4 As shown. To accurately capture boundary layer flow and shock wave characteristics, mesh refinement was performed in the throat and near-wall regions. In this embodiment, the total number of meshes in a single case is approximately 22379, and the height of the first layer of meshes on the wall is set to 0.0001 mm to ensure the dimensionless wall distance y + It is approximately 0.12 to meet the accuracy requirements for viscous flow calculations.

[0068] S122. Boundary Conditions and Solution: Set the nozzle inlet as a pressure inlet boundary, and give the total temperature (e.g., 833K) and total pressure (e.g., varying from 0.41MPa to 4.1MPa, covering under-expansion and over-expansion conditions); set the outlet as a pressure outlet boundary, and set the ambient pressure (e.g., 101.325kPa) and ambient temperature (300K); use adiabatic no-slip boundary conditions for the wall.

[0069] S123. Solver Settings: A steady-state solver based on the Reynolds-averaged Navier-Stokes (RANS) equations is used in conjunction with the shear stress transport (SST) k-ω turbulence model for calculation. The working fluid is set to an ideal gas.

[0070] S124. Obtain the temperature field, velocity field, and pressure field data corresponding to each geometric configuration through CFD calculation, and use them as the output dataset of the deep learning model.

[0071] S13. Construction of Dual-Channel Geometric Feature Input (SDF and IM): To enable the neural network to understand complex geometrical boundaries, this invention constructs a dual-channel feature tensor containing a signed distance field (SDF) and an identity matrix (IM) as model input, specifically including the following steps:

[0072] S131. Construct the Symbolic Distance Field (SDF): Calculate the minimum directed distance from each grid point in the computational domain to the nozzle wall. The definition rule is: positive values ​​for the interior region of the nozzle, negative values ​​for the exterior region, and 0 for the nozzle wall. Figure 5 As shown, SDF can provide smoothly varying geometric gradient information, which helps convolutional neural networks extract shape features.

[0073] S132. Constructing the Identification Matrix (IM): To give the model a clear physical boundary semantics, an integer identification matrix consistent with the computational grid resolution is constructed. Regions are encoded according to their physical meaning: for example, 0 represents the outside of the nozzle, 1 represents the inside of the nozzle, 2 represents the inlet, 3 represents the outlet, 4 represents the axis of symmetry, and 5 represents the nozzle wall, etc. Figure 6 As shown.

[0074] S133. Concatenate the generated SDF matrix and IM matrix along the channel dimension to form an input tensor of size H×W×2.

[0075] S14. Data Normalization and Dataset Partitioning: Extract the temperature, velocity, and pressure field data obtained from CFD calculations and map them to the same grid nodes as the input tensor using Cartesian grid interpolation. To accelerate model convergence, perform max-min normalization on each physics field data, scaling its numerical range to the interval [0, 1] or [-1, 1].

[0076] The final basic dataset contains 1,000 pairs of "geometry-flow field" samples, which are randomly divided into training and validation sets in a 7:3 ratio for pre-training of the subsequent basic model.

[0077] S2. Construct the DE-ConvSE U-Net model, such as Figure 7 and Figure 8 As shown, the specific steps are as follows:

[0078] S21. Construct a feature extraction encoder. Use a multi-layer convolutional neural network as the encoder. Through continuous convolution operations and downsampling pooling, extract the local geometric and physical features of the nozzle flow field step by step.

[0079] S22. Construct a multi-branch independent decoder. The decoder is divided into three independent branches: pressure, temperature and speed. The feature map resolution is gradually restored through upsampling operations, and the shallow spatial features of the encoder are fused with the deep global semantic features of the output through skip connections at the corresponding levels.

[0080] S23. Calculate the loss functions for each channel of velocity, temperature, and pressure. , and :

[0081] ;

[0082] ;

[0083] ;

[0084] In the formula: , , These represent the nth sample point in... The predicted values ​​for velocity, temperature, and pressure are given. , , The model is represented at the point. The CFD simulation results.

[0085] S24. Weight the losses from pressure, temperature, and velocity channels to obtain the loss function of the flow field prediction model:

[0086] ,

[0087] In the formula, These represent the weights of the velocity field, temperature field, and pressure field, respectively.

[0088] Furthermore, a variance-based dynamic loss scaling mechanism is employed to adaptively weight the losses in pressure, temperature, and velocity channels to balance the magnitude of losses across different physical quantities; the adaptive weighting expression is as follows:

[0089] ;

[0090] In the formula This represents the mean of channel c. Represents the position of the nth sample. The actual value of the c-th channel; N, H, and W represent the batch size, height, and width, respectively.

[0091] S25. Construct the DE-ConvSE adapter module. The DE-ConvSE adapter module contains a convolutional stack layer and a squeeze-and-excitation (SE) module. It adjusts the relationship between the input data through the attention mechanism, and finally introduces learnable gating coefficients and performs residual connections.

[0092] Specifically, the output calculation formula for the adapter is as follows:

[0093] ,

[0094] in As input features, For convolution operations, To compress the excitation operation, The learnable gating coefficients are set to a small value during initial training to preserve the backbone network characteristics in the early stages of training.

[0095] S26. Maintain the backbone U-Net model structure and embed DE-ConvSE adapter modules into the convolutional blocks at both ends of the encoder and decoder.

[0096] S3. Basic model pre-training, specific implementation methods include:

[0097] S31. Input the symbolic distance field (SDF) and the label matrix (IM) from the basic dataset into the network, generate the predicted flow field through forward propagation, calculate the gradient of the total objective loss function using the backpropagation algorithm, and update the parameters of the encoder, decoder and adapter modules until the model converges.

[0098] S32. Save the basic model parameters for incremental model learning.

[0099] S4. Incremental learning and adaptation, specifically implemented through the following methods:

[0100] S41. Construct a small sample incremental dataset: Obtain a small amount of nozzle flow field data under new geometric configurations as an incremental learning sample set;

[0101] S42. Sample data augmentation: Perform geometric transformation processing on the incremental learning sample set. The geometric transformation includes random flipping, rotation, scaling and translation operations to expand the spatial diversity of the sample data.

[0102] S43. Freeze the backbone network: Keep the weight parameters of all convolutional and pooling layers of the encoder and decoder in the U-Net model obtained by pre-training in step S3 unchanged, so that it retains the general flow field feature extraction capability of the basic configuration.

[0103] S44. Adapter parameter fine-tuning: The model is iteratively trained using a data-enhanced sample set. The internal parameters and gating coefficients of the DE-ConvSE adapter modules embedded in each layer of the network are updated only through the backpropagation algorithm, so that the model can adapt to the flow field characteristics of the new geometric configuration.

[0104] Specifically, the DE-ConvSE U-Net neural network processes the input data as follows:

[0105] (1) Dual-channel geometric feature input

[0106] The input layer receives the preprocessed nozzle flow field geometric feature tensor (including the symbolic distance field SDF and the identifier matrix IM channel) and passes it to the shared feature extraction encoder.

[0107] (2) Encoder feature extraction and adaptation

[0108] The encoder processes the input data through four cascaded convolutional downsampling modules. In each module, the data first undergoes a nonlinear feature transformation through convolutional layers, batch normalization layers, and ReLU activation functions. Subsequently, the feature map enters an embedded DE-ConvSE adapter module. The adapter recalibrates the feature channel weights based on the current geometric configuration information through a "compression-excitation" mechanism, enhancing the response of key geometric features. Finally, the adapted feature map is downsampled through a max-pooling layer (with a stride of 2) to achieve spatial compression and feature extraction. The deep, high-dimensional feature map output by the encoder is then fed into the bottleneck layer.

[0109] (3) Feature fusion of bottleneck layer

[0110] The bottleneck layer acts as a bridge connecting the encoder and decoder. It performs further convolution processing on the highest-dimensional abstract features of the encoder output, extracts the implicit flow-geometry coupling features, and copies and distributes them to the three independent decoder branches at the back end.

[0111] (4) Multi-branch decoding and physical field reconstruction

[0112] The decoder is divided into three parallel branches: temperature, speed, and pressure. Each branch performs the following processing independently:

[0113] (4.1) Upsampling and Skip Connection: The spatial size of the feature map is enlarged by 2 times by deconvolution operation, and the enlarged feature map is spliced ​​and fused with the high-resolution feature map of the corresponding layer of the encoder (Skip Connection) in the channel dimension to recover spatial details such as shock wave position lost during downsampling.

[0114] (4.2) Decoder adaptation: The fused feature map is processed again by a convolutional layer and then enters the DE-ConvSE adapter module embedded in the decoder. The adapter here uses incrementally learned updated parameters to finely adjust the features for specific physical fields (such as pressure changes at shock waves).

[0115] (4.4) Final physical field output: After four levels of upsampling and adaptation, the feature map is restored to the same original resolution as the input grid. Finally, each branch maps the high-dimensional features to a single-channel output through a 1×1 convolutional layer, generating the predicted temperature field distribution map, velocity field distribution map, and pressure field distribution map, respectively.

[0116] S5. Use the trained model to quickly predict the nozzle flow field. Specific implementation methods include:

[0117] S51. Prepare the pressure, temperature and velocity at the engine nozzle inlet and fill them into a Cartesian grid as input to the model. The model will extract the corresponding flow field spatial features based on the flow field image features, perform high-dimensional mapping of the flow field spatial features through a deconvolution layer, and generate a predicted flow field image with corresponding boundary conditions.

[0118] Depend on Figure 9 As can be seen, under the over-expansion state, the model's values ​​for pressure, velocity, and temperature are 1.61%, 4.32%, and 2.77%, respectively. This demonstrates that the model can accurately capture the flow field characteristics and predict the internal flow field of the engine nozzle under the over-expansion state with high precision.

[0119] Depend on Figure 10 As can be seen, under the over-expansion state, the model's values ​​for pressure, velocity, and temperature are 2.12%, 5.38%, and 3.97%, respectively. This demonstrates that the model can accurately capture the flow field characteristics inside the engine nozzle under the over-expansion state, such as shock waves and flow separation, and can predict the internal flow field of the engine nozzle under the over-expansion state with high accuracy.

[0120] S6. Predict the nozzle specific impulse parameters using the trained model. Specific implementation methods include:

[0121] S61. Based on the model prediction results, the nozzle wall velocity and the pressure at the inlet and outlet are extracted. Based on the gas density inside the nozzle and the diameters of the nozzle inlet and outlet, the specific impulse of the nozzle is obtained through the following calculation steps:

[0122] ;

[0123] In the formula, I represents the specific impulse of the nozzle, and F represents the nozzle thrust. The mass flow rate inside the nozzle is represented by the following formula:

[0124] ;

[0125] In the formula: For gas density, This represents the axial component of the exit velocity. , These refer to the pressure at the nozzle exit and atmospheric pressure, respectively.

[0126] Furthermore, the parameters obtained from the nozzle simulation are compared with the parameters predicted by the neural network, and the error calculation formula is shown below;

[0127] ;

[0128] In the formula, The specific impulse is the one predicted by the neural network. The specific impulse is obtained based on CFD simulation. The error statistics for both methods in this embodiment are shown in Table 1:

[0129] Table 1 Comparison of Specific Impulse Prediction Errors

[0130] Pressure MPa Temperature K Simulated specific impulse N s / kg Model predicted specific impulse N s / kg Error % 7.8 3500 2360.4 2337.6 0.9 9.0 3400 2284.5 2298.6 0.6 10.0 3500 2450.8 2340.9 1.2

[0131] Another advantage of deep learning models is their significant reduction in the time consumption of numerical simulations. Table 2 shows the time consumption for acquiring flow field information using CFD simulation and deep learning in batches of 1, 10, and 100. To ensure fairness, both calculations were performed using an Intel i5-12400F processor with 2 cores, and the convergence residual was uniformly set to 1×10⁻⁶. -6 The time consumption of the two methods in this embodiment is compared in Table 2:

[0132] Table 2 Comparison of time consumption for CFD simulation and deep learning batch simulation

[0133] Simulation number CFD simulation time t / s Deep learning simulation time t / s 1 119 1.7 10 1190 20.3 100 11900 182.6

[0134] As shown in Tables 1 and 2, compared with CFD simulation, the deep learning model achieves a two-order-of-magnitude acceleration in acquiring large-scale flow field data within the engine nozzle, and can achieve stable and efficient rapid prediction, making it more suitable for applications with multiple batches and high real-time requirements. In the early design process of the engine nozzle, it can quickly iterate different design schemes, reducing the need for extensive CFD calculations and thus greatly improving design efficiency.

[0135] In another embodiment of the present invention, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and when the program is executed on the processor, the steps in the deep learning-based nozzle in-flow field prediction method described above are implemented.

[0136] In another embodiment of the present invention, a storage medium is also provided, the storage medium storing a computer program, which, when executed by a processor, implements the steps in the deep learning-based nozzle in-flow field prediction method as described above.

[0137] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the modules described above can be referred to the corresponding process in the aforementioned method implementation, and will not be repeated here.

[0138] The modules described as separate components may or may not be physically separate. Similarly, the components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0139] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in a combination of hardware and software functional modules.

[0140] The integrated modules implemented as software functional modules described above can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer system (which may be a personal computer, server, or network system, etc.) or processor to execute some steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for predicting the flow field inside a nozzle based on deep learning, characterized in that: The method includes the following steps: S1. Construct a dataset and perform preprocessing. The dataset consists of several sets of nozzle geometric features and corresponding temperature field, velocity field and pressure field data. S2. Construct a multi-physics prediction network based on the U-Net architecture. The multi-physics prediction network includes a shared encoder and three independent decoder branches, and a double-ended convolutional compression excitation adapter module is embedded after each convolutional block of the shared encoder and decoder branches. S3. Train the multiphysics prediction network based on the dataset preprocessed in step S1. S4. Using the multiphysics prediction network trained in step S3, perform flow field prediction for the nozzle under test.

2. The method according to claim 1, characterized in that, The dataset construction method in step S1 includes: S11. Several nozzle geometries with different expansion profiles are generated using parametric modeling methods. S12. For each nozzle geometry generated in step S11, use computational fluid dynamics (CFD) to perform flow field simulation and obtain the corresponding temperature field, velocity field and pressure field data. S13. Calculate the minimum directed distance from each grid point in the computational domain to the nozzle wall, and construct the symbolic distance field; S14. Construct an identification matrix with the same resolution as the computational domain grid, where each element in the identification matrix represents the physical meaning of the corresponding region; S15. The symbolic distance field (SDF) and the identification matrix are spliced ​​together in the channel dimension to form the nozzle geometry. S16. The geometric features of each nozzle and the corresponding temperature field, velocity field and pressure field data form a set of "geometry-flow field" samples, which in turn form a dataset.

3. The method according to claim 1, characterized in that, The method for building a multiphysics prediction network based on the U-Net architecture in step S2 includes: S21. A multi-layer convolutional neural network is used as a shared encoder, with nozzle geometric features as input. S22. For the temperature field, velocity field and pressure field, three independent decoder branches are constructed, and each decoder branch is connected to the encoder in a skip connection.

4. The method according to claim 3, characterized in that, The loss function of the multiphysics prediction network is: ， ， ， ， in, , and Let represent the loss functions of the velocity field, temperature field, and pressure field channels, respectively. , and These represent the positions in the nth sample. The predicted values ​​for the velocity field, temperature field, and pressure field are given. , and respective positions The model represents the CFD simulation results of the velocity field, temperature field, and pressure field. These represent the weights of the velocity field, temperature field, and pressure field, respectively.

5. The method according to claim 4, characterized in that, The adaptive acquisition method is as follows: ， Where c = 1, 2, 3, This represents the mean of channel c. These represent the positions in the nth sample. The true values ​​of the velocity field, temperature field, and pressure field at that location. These represent the central locations in the dataset. The mean of the true values ​​of the velocity field, temperature field, and pressure field; N, H, and W represent the batch size of the dataset, the height of the nozzle, and the width of the nozzle, respectively.

6. The method according to claim 3, characterized in that, The method for building a multiphysics prediction network based on the U-Net architecture in step S2 also includes: S23. Embed a dual-ended convolutional compression excitation adapter module after each convolutional block in the shared encoder and decoder branches, and perform residual connections; the dual-ended convolutional compression excitation adapter module includes a convolutional stack layer and a compression-excitation SE module, and introduces learnable gating coefficients.

7. The method according to claim 6, characterized in that, The method further includes: If the geometry of the nozzle under test is not included in the database, an incremental learning strategy is adopted to construct an incremental dataset, and only the bi-ended convolutional compressed excitation adapter module in the multiphysics prediction network is learned.

8. The method according to claim 6, characterized in that, The method further includes: S5. Based on the flow field prediction results of step S4, obtain the specific impulse of the nozzle under test: ， ， In the formula, I represents the specific impulse of the nozzle under test, and F represents the thrust of the nozzle under test. This represents the mass flow rate inside the nozzle being tested. For gas density, The axial component of the nozzle exit velocity to be measured is denoted as . , These represent the pressure at the nozzle exit and the atmospheric pressure, respectively; S represents the nozzle exit cross-sectional area.

9. An electronic device, comprising a memory and a processor, characterized in that: The memory stores a computer program that can run on the processor, and when the program is executed on the processor, it implements the steps of the method as described in any one of claims 1-7.

10. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images