Data-driven method and apparatus for predicting three-dimensional combustion flow field of ramjet engines

By using a data-driven approach, fully connected neural networks and multi-path 3D neural networks are employed to decompose and reconstruct the 3D combustion flow field of a ramjet engine. This solves the problems of huge computational resource consumption and excessively long time cycles in existing technologies, enabling fast and accurate flow field prediction and improving design and optimization efficiency.

CN123118975APending Publication Date: 2026-11-17NORTHWESTERN POLYTECHNICAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611613467.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

Smart Images

  • Figure CN123118975A_ABST
    Figure CN123118975A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of three-dimensional combustion flow field prediction technology for ramjet engines. It discloses a data-driven method and device for predicting the three-dimensional combustion flow field of ramjet engines, which obtains combustion test parameters of ramjet engines; converts the combustion test parameters into one-dimensional test parameters corresponding to a three-dimensional uniform grid generated by the ramjet engine; inputs the one-dimensional test parameters into a fully connected neural network to obtain a one-dimensional flow path parameter distribution; extends and maps the one-dimensional flow path parameter distribution onto a three-dimensional uniform grid to obtain a three-dimensional flow path parameter distribution; inputs the three-dimensional flow path parameter distribution into a multi-path three-dimensional neural network to obtain the predicted three-dimensional combustion flow field of the ramjet engine. This invention effectively avoids the huge computational load and convergence difficulties caused by directly fitting high-dimensional flow fields end-to-end through a processing mechanism of dimensionality reduction encoding, one-dimensional evolution guidance, and three-dimensional reconstruction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional combustion flow field prediction technology for ramjet engines, and particularly relates to a data-driven method and apparatus for predicting three-dimensional combustion flow field of ramjet engines. Background Technology

[0002] The three-dimensional combustion flow field of a ramjet engine describes the flow and combustion state of the airflow, fuel, and flame inside the combustion chamber, which differs from the simplified two-dimensional planar assumption. The three-dimensional combustion flow field contains supersonic incoming flow and vortex structures formed by fuel injection. At the same time, shock waves and chemical reactions are coupled with each other. The flame and component distribution exhibit significant spanwise inhomogeneity, which affects mixing efficiency, flame stability, and wall thermal load. It is the core object of ramjet engine research.

[0003] Three-dimensional combustion flow field prediction mainly employs two methods: high-precision computational fluid dynamics numerical simulation (CFD) and ground-based experimental verification. Numerically, three-dimensional compressible CFD with chemical reactions is the primary method, commonly using RANS, DES / hybrid RANS-LES methods to solve multi-component Navier-Stokes equations and capture the coupling effects of shock waves, vortex structures, and combustion heat release. This is the core tool for predicting flow field distribution in the early stages.

[0004] However, CFD has a large mesh size and the calculation of chemical reaction source terms is time-consuming. A single working condition may take several days or even weeks, making it difficult to quickly and batch evaluate a large number of schemes, and the efficiency of parameter optimization is very low. Summary of the Invention

[0005] The purpose of this invention is to provide a data-driven method and apparatus for predicting the three-dimensional combustion flow field of a ramjet engine, so as to improve the prediction efficiency of the three-dimensional combustion flow field of a ramjet engine.

[0006] This invention employs the following technical solution: a data-driven method for predicting the three-dimensional combustion flow field of a ramjet engine, comprising the following steps: Obtain combustion test parameters for the ramjet engine; the combustion test parameters are the historical combustion flow field parameters or test condition parameters for the previous N time moments. The historical combustion flow field parameters include temperature, pressure and Mach number, and the test condition parameters include the geometric parameters and operating conditions of the ramjet engine. The combustion test parameters were converted into one-dimensional test parameters corresponding to the three-dimensional uniform grid generated by the ramjet engine. One-dimensional experimental parameters are input into a fully connected neural network to obtain a one-dimensional parameter distribution along the path; The one-dimensional friction parameter distribution is extended and mapped onto a three-dimensional uniform grid to obtain the three-dimensional friction parameter distribution; The three-dimensional flow parameter distribution along the path is input into a multi-path three-dimensional neural network to obtain the predicted three-dimensional combustion flow field of the ramjet engine.

[0007] Another technical solution of the present invention: a data-driven three-dimensional combustion flow field prediction device for ramjet engines, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor implements the above-described method when executing the computer program.

[0008] The beneficial effects of this invention are as follows: Through the processing mechanism of dimensionality reduction encoding, one-dimensional evolution guidance, and three-dimensional reconstruction, this invention decomposes the high-dimensional nonlinear flow field prediction problem into two stages: axial dominant trend learning and spatial detail reconstruction. Based on the rapid solution of one-dimensional flow parameter distribution by a fully connected network, it effectively avoids the huge computational load and convergence difficulties brought about by directly fitting the high-dimensional flow field end-to-end. With the help of multi-path three-dimensional neural networks to extract and fuse the differentiated features of different location regions, the prediction time can be significantly shortened while maintaining physical consistency. Thus, it effectively solves the technical problems of existing prediction methods, such as huge consumption of computational resources and excessively long time cycles, which make it difficult to support large-scale parameter optimization and real-time performance evaluation. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of a data-driven three-dimensional combustion flow field prediction method for ramjet engines according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the basic configuration of the ramjet engine in an embodiment of the present invention; Figure 3 This is a visual comparison diagram of the unstructured mesh and original flow field data with the interpolated uniform mesh and interpolated flow field data in an embodiment of the present invention; Figure 4 This is a schematic diagram of a multi-path three-dimensional neural network architecture in an embodiment of the present invention; Figure 5 This is a CFD result diagram of the three-dimensional pressure field of the ramjet engine in an embodiment of the present invention; Figure 6 The image shows the predicted three-dimensional pressure field of a ramjet engine obtained using the method of this invention. Figure 7 This is a CFD result diagram of the three-dimensional temperature field of the ramjet engine in an embodiment of the present invention; Figure 8 The image shows the predicted results of the three-dimensional temperature field of a ramjet engine obtained using the method of this invention. Figure 9 This is a normalized temperature distribution result obtained based on CFD at 10ms in an embodiment of the present invention; Figure 10 This is a normalized temperature distribution result obtained using the method of the present invention at 10ms in an embodiment of the present invention; Figure 11 for Figure 9 and Figure 10A schematic diagram of the relative error of the normalized temperature distribution; Figure 12 This is a normalized temperature distribution result obtained based on CFD at 30ms in an embodiment of the present invention; Figure 13 This is a normalized temperature distribution result obtained by using a neural network prediction at 30ms in an embodiment of the present invention; Figure 14 for Figure 12 and Figure 13 A schematic diagram of the relative error of the normalized temperature distribution. Detailed Implementation

[0010] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0011] This invention discloses a data-driven method for predicting the three-dimensional combustion flow field of a ramjet engine, such as... Figure 1 As shown, the process includes the following steps: obtaining combustion test parameters of the ramjet engine; the combustion test parameters are historical combustion flow field parameters or test condition parameters from the previous N time points, where the historical combustion flow field parameters include temperature, pressure, and Mach number, and the test condition parameters include the geometric parameters and operating parameters of the ramjet engine; converting the combustion test parameters into one-dimensional test parameters corresponding to a three-dimensional uniform grid generated by the ramjet engine; inputting the one-dimensional test parameters into a fully connected neural network to obtain a one-dimensional flow path parameter distribution; extending and mapping the one-dimensional flow path parameter distribution onto a three-dimensional uniform grid to obtain a three-dimensional flow path parameter distribution; inputting the three-dimensional flow path parameter distribution into a multi-path three-dimensional neural network to obtain the predicted three-dimensional combustion flow field of the ramjet engine.

[0012] This invention employs a processing mechanism of dimensionality reduction encoding, one-dimensional evolution guidance, and three-dimensional reconstruction to decompose the high-dimensional nonlinear flow field prediction problem into two stages: axial dominant trend learning and spatial detail reconstruction. Based on the rapid solution of one-dimensional flow parameter distribution using a fully connected network, it effectively avoids the huge computational load and convergence difficulties caused by directly fitting high-dimensional flow fields end-to-end. By leveraging the differentiated extraction and fusion of features from different locations using a multi-path three-dimensional neural network, the prediction time can be significantly shortened while maintaining physical consistency. This effectively solves the technical problems of existing prediction methods, such as huge computational resource consumption and excessively long time cycles, which make it difficult to support large-scale parameter optimization and real-time performance evaluation.

[0013] Based on the differences in the predicted scenarios, combustion test parameters are categorized into two sources. The first category consists of historical combustion flow field parameters from the previous N time points, suitable for time-series predictions under known initial flow field evolution conditions. This data primarily originates from high-precision numerical simulation results or ground-based test measurements, including key physical quantities such as temperature, pressure, and Mach number corresponding to specific time points. The second category comprises test condition parameters, applicable to scenarios with only engine configuration and operating condition settings but no initial flow field data. These parameters cover the geometric parameters of the ramjet engine (such as the size of the isolator section, combustion chamber expansion angle, support plate position, and cavity size) and operating condition parameters (such as flight Mach number, flight pressure, and fuel injection equivalence ratio). During the data acquisition phase, historical flow field data can be extracted by reading pre-stored CFD simulation database files, or operating condition parameters on the test bench can be acquired in real time through a sensor acquisition system. The acquired raw data must undergo cleaning and format standardization processing to eliminate noise interference and dimensional differences, ensuring the input quality for subsequent data conversion.

[0014] In this invention, one-dimensional experimental parameters are used to standardize input data of different dimensions and formats to adapt to the subsequent neural network processing architecture. A three-dimensional uniform grid refers to a pre-constructed structured grid with a fixed number of grid cells along the X, Y, and Z axes (e.g., 600×100×100). Converting combustion experimental parameters into one-dimensional experimental parameters essentially maps three-dimensional spatially distributed physical quantities or high-dimensional geometric descriptions into a one-dimensional vector sequence characterizing the flow field's variation trend.

[0015] If the combustion test parameters are historical combustion flow field parameters, the conversion process includes: first, obtaining the unstructured mesh of the ramjet engine; mapping the coordinates of the unstructured mesh and the corresponding flow field data to the predetermined three-dimensional uniform mesh nodes using a three-dimensional linear interpolation algorithm; then, performing surface averaging calculation on the flow field parameters of each flow section in the three-dimensional uniform mesh to obtain a one-dimensional parameter sequence along the flow direction at N time points, i.e., one-dimensional test parameters.

[0016] If the combustion test parameters are test condition parameters, the conversion process focuses on encoding geometric parameters into one-dimensional geometric parameters and mapping operating condition parameters into one-dimensional operating condition parameters before merging them. In the efficient calculation process of ramjet engine flow field, this operation of converting complex physical space data into one-dimensional parameters can effectively reduce the dimensionality of input data and extract the core features that determine the evolution of the flow field, unifying test parameters from different sources into the same data space, thus laying the foundation for the input of fully connected neural networks.

[0017] Next, a fully connected neural network is used to learn the nonlinear mapping relationship from one-dimensional experimental parameters to the flow field distribution along the flow path. The fully connected neural network consists of a multilayer sensing mechanism, with fully connected neurons in adjacent layers, and nonlinear factors are introduced through activation functions to fit complex functional relationships. The one-dimensional flow field parameter distribution refers to the distribution of physical quantities along the engine flow direction, typically including the curves of key parameters such as pressure, temperature, and Mach number as a function of location.

[0018] During the model execution phase, one-dimensional experimental parameters serve as input features. These parameters are weighted and summed across layers of a fully connected neural network, undergoing nonlinear transformation to output a predicted one-dimensional friction parameter vector. In one embodiment, the fully connected neural network may contain four fully connected layers. The input layer receives the one-dimensional experimental parameters, which are a combination of one-dimensional geometric parameters and one-dimensional operating condition parameters. The output layer generates a 3×600 one-dimensional vector, corresponding to pressure, temperature, and Mach number at 600 flow directions. The network structure can also be adjusted to a deeper residual fully connected network to improve feature extraction capabilities based on data complexity. The one-dimensional friction parameter distribution output in this step reflects the macroscopic evolution trend of the flow field in the mainstream direction, providing physical constraints and a baseline framework for subsequent reconstruction of three-dimensional spatial details.

[0019] Then, the low-dimensional one-dimensional friction parameter distribution is restored to three-dimensional space to achieve data dimensionality matching. The extended mapping refers to broadcasting or copying the parameter values ​​(such as pressure and temperature) corresponding to each flow direction location in the one-dimensional friction parameter distribution to all transverse and spanwise grid points corresponding to that flow direction location, according to the topology of a three-dimensional uniform grid. For example, the i-th element of the one-dimensional vector is assigned to all nodes in the three-dimensional uniform grid whose x-coordinate is equal to the i-th grid coordinate; or, based on pre-calculated three-dimensional distribution weights, the one-dimensional friction parameter distribution is weighted and extended to reflect the radial distribution gradient.

[0020] Through this process, the original one-dimensional flow parameter distribution, which only contained flow direction changes, was expanded into tensor data containing complete three-dimensional coordinate information, i.e., a three-dimensional flow parameter distribution. This data volume not only contains the distribution of the core physical quantities of the flow field but also implicitly includes the geometric spatial structure information of the engine. Its data dimension is completely consistent with the spatial dimension of the three-dimensional uniform grid, thus forming the direct input tensor of the multi-path three-dimensional neural network. This approach of first reducing the dimension for trend prediction and then expanding the dimension for spatial expansion utilizes the physical characteristics of the ramjet engine flow field being autonomously induced along the flow direction, effectively reducing the difficulty of directly predicting high-dimensional flow fields.

[0021] Finally, a deep learning model is used to reconstruct the detailed three-dimensional structure of the flow field. A multi-path 3D neural network is a convolutional neural network architecture specifically designed for processing three-dimensional volumetric data. It contains multiple parallel feature extraction paths to capture the feature differences of the flow field in different spatial regions or at different scales. The three-dimensional flow path parameter distribution is used as input, and after forward propagation through the multi-path 3D neural network, the output is a predicted three-dimensional combustion flow field, namely, the pressure field, temperature field, and Mach number field containing complete spatial distribution information.

[0022] As a specific implementation, a multi-path 3D neural network can be designed with a structure containing multiple encoders and one decoder. The encoders extract deep features through 3D convolution and pooling operations, while the decoder restores the flow field resolution through deconvolution or upsampling operations. The network can also incorporate an attention mechanism to dynamically adjust the weights at different spatial locations to focus on key flow regions such as the recirculation zone or shock surface. During inference, the network fills in the lateral and spanwise details of the flow field based on the input 3D flow path parameter distribution and the geometric and flow field mapping rules learned during training, ultimately generating high-precision 3D combustion flow field data. The output of this step can be directly used for engine performance evaluation, combustion stability analysis, and optimization design, thus replacing time-consuming traditional CFD simulations.

[0023] The aforementioned technical processes work together to achieve rapid and accurate prediction of the three-dimensional combustion flow field of a ramjet engine. First, a fully connected neural network extracts the macroscopic physical evolution trend along the flow direction from one-dimensional experimental parameters. This low-dimensional prediction provides a basic framework conforming to physical laws for subsequent three-dimensional reconstruction, avoiding non-physical oscillations that may occur in end-to-end prediction. Subsequently, by extending and mapping the one-dimensional flow parameter distribution to a three-dimensional uniform grid, a three-dimensional flow parameter distribution containing spatial geometric information is generated, providing the multi-path three-dimensional neural network with rich inputs that include both flow field trends and spatial constraints. Based on this, the multi-path three-dimensional neural network utilizes its powerful three-dimensional feature extraction capabilities to reconstruct the fine spatial structure of the flow field, compensating for the lack of lateral and spanwise details in one-dimensional prediction. Figure 9-14 As shown in the comparison of prediction results at 10ms and 30ms, the predicted 3D combustion flow field generated by this method maintains a high degree of consistency with the high-fidelity simulation results in shock wave position, flame morphology, and parameter distribution, with a significant reduction in average error. This collaborative mechanism of one-dimensional trend guidance and three-dimensional detail reconstruction, while ensuring computational efficiency, fully utilizes the prior physical knowledge of the flow field, effectively solving the problems of long computation time and difficulty in quickly evaluating a large number of schemes using traditional CFD methods, and significantly improving the iterative efficiency of ramjet engine design and optimization.

[0024] In one embodiment, inputting the three-dimensional friction parameter distribution into a multi-path three-dimensional neural network includes: determining the location region of the three-dimensional friction parameter distribution in the ramjet engine; and feeding the three-dimensional friction parameter distribution of different location regions into different encoders of the multi-path three-dimensional neural network.

[0025] Based on the acquired three-dimensional flow parameter distribution and considering the physical structural characteristics of the ramjet engine, the flow field data is divided into different regions along the flow direction, such as the isolation section, the first-stage combustion chamber, and the second-stage combustion chamber. The division of these regions is defined by the axial coordinate distribution of the support plates, cavities, and expansion sections in the engine's geometric model. Specifically, by analyzing the coordinate range of a three-dimensional uniform grid in the X-axis direction, the entire flow field data is sliced ​​and mapped to the corresponding structural partitions, thereby separating the flow field data of the isolation section, the first-stage combustion chamber, and the second-stage combustion chamber. This physics-driven region division strategy decouples the complex global three-dimensional flow field into several local subfields with dominant physical mechanisms, reducing the complexity of processing global data using a single model.

[0026] For each of the divided location regions, its corresponding 3D path parameter distribution is input into its respective independent encoder branch in the multi-path 3D neural network. For example... Figure 4 As shown, the multi-path 3D neural network is configured with multiple encoders operating in parallel (encoder 1, encoder 2, and encoder 3 as shown in the figure), each encoder specifically designed to process the flow field features of a particular region. Simultaneously, the multi-path 3D neural network also includes a decoder, which restores the data size to reconstruct the 3D flow field. Features output from the convolutional blocks in the encoders are passed to the decoder via skip connections and fused with the upsampled features from the corresponding level of the decoder, achieving the fusion of shallow and deep features. For example, the 3D friction parameter distribution of the isolation section is fed into encoder 1 to extract shock wave evolution and boundary layer development features; data from the first-stage combustion chamber is fed into encoder 2 to capture the mixing characteristics of the strut wake and fuel injection; and data from the second-stage combustion chamber is fed into encoder 3 to analyze the flame propagation law and pressure recovery characteristics within the expansion section. This multi-path parallel processing mechanism allows each encoder to specifically learn nonlinear mapping relationships under different spatial scales and flow patterns, effectively avoiding the feature smoothing problem that may occur when a single network model processes strong gradient flow fields (such as concave cavity recirculation zones), thereby significantly improving the prediction accuracy and robustness of the flow field structure in the key combustion region of the ramjet engine.

[0027] Through structured region division and differential coding, this method can effectively address the complex spatial heterogeneity of the combustion chamber of a ramjet engine, laying a precise feature extraction foundation for subsequent reconstruction of a high-fidelity three-dimensional combustion flow field using a decoder.

[0028] When the combustion test parameters are test condition parameters, converting the combustion test parameters into one-dimensional test parameters corresponding to the three-dimensional uniform grid generated by the ramjet engine includes: encoding the geometric parameters into one-dimensional geometric parameters; merging the one-dimensional geometric parameters and one-dimensional operating condition parameters to obtain one-dimensional test parameters; wherein, the one-dimensional operating condition parameters are obtained through operating condition parameter mapping.

[0029] Geometric parameters characterize the physical structure of a ramjet engine and are key static features determining the spatial distribution of the combustion flow field. Given that original geometric parameters typically exist in the form of computer-aided design models or discrete coordinate point sets and have high dimensionality, this invention encodes them into fixed-length, low-dimensional dense vectors—one-dimensional geometric parameters. Specifically, based on a three-dimensional uniform grid generated by the ramjet engine, the geometric boundary information of the engine flow channel is extracted, and the three-dimensional structure is projected into a two-dimensional contour image along the flow direction and transverse direction. This two-dimensional contour image can intuitively reflect key configuration features such as the flow channel expansion ratio, the position of the support plate, and the depth of the cavity. Subsequently, a visual attention neural network is used to extract and encode features from the two-dimensional contour image, mapping the image pixel information into a vector representation in a high-dimensional feature space. The visual attention neural network can include a multi-head self-attention mechanism, which automatically focuses on local geometric features that significantly affect the combustion flow (such as the leading edge of the support plate and the trailing edge of the cavity) by weighted aggregation of information from different regions of the image, thereby compressing the geometric parameters containing complex topological structures into a fixed-size one-dimensional geometric parameter vector (e.g., a 2048-dimensional vector). Through the above encoding process, unstructured geometric shapes are converted into numerical features with translation invariance and local perception capabilities, which facilitates efficient processing by subsequent neural networks.

[0030] One-dimensional operating condition parameters are quantitative indicators characterizing the engine's operating environment and working state, derived from operating condition parameters. To eliminate the adverse effects of the order-of-magnitude differences between different physical dimensions (such as pressure, Mach number, and equivalence ratio) on model convergence, the operating condition parameters need to be converted into one-dimensional operating condition parameters through mapping operations. This mapping process can include: using normalization or standard deviation standardization to linearly or nonlinearly map the original operating condition parameters (such as flight Mach number, flight pressure, and injection equivalence ratio) to a preset interval (e.g., [0,1] or [-1,1]), thereby generating a one-dimensional operating condition parameter vector with a unified numerical range and eliminated dimensional influence; or, using a fully connected neural network as a feature mapper to map discrete operating condition parameter values ​​into high-dimensional feature vectors. This fully connected neural network can contain a multilayer perceptron and introduce nonlinear transformation capabilities through activation functions to capture potential coupling relationships that may exist between different operating condition parameters. After obtaining the one-dimensional geometric parameters and one-dimensional operating condition parameters, a tensor concatenation operation is used to merge them along a specific dimension, forming a combined vector containing complete input information, i.e., the one-dimensional test parameters. This combined vector simultaneously contains the engine's structural characteristics (provided by the one-dimensional geometric parameters) and operational state characteristics (provided by the one-dimensional operating condition parameters), serving as input data for the subsequent fully connected neural network, thus achieving a unified representation of the physical configuration and test conditions. By merging the one-dimensional geometric parameters and the mapped one-dimensional operating condition parameters, the neural network can jointly learn the synergistic influence mechanism of geometric shape and operating conditions on the combustion flow field in a high-dimensional feature space.

[0031] In one specific implementation, encoding the geometric parameters into one-dimensional geometric parameters includes: generating a geometric model of the ramjet engine based on the geometric parameters, and generating a three-dimensional uniform mesh based on the geometric model; selecting a cross-section perpendicular to the axis along the flow direction of the ramjet engine at the center position of the flame stabilizer; selecting a longitudinal section passing through the engine axis; projecting the mesh coordinates of the three-dimensional uniform mesh onto the cross-section and the longitudinal section to obtain the flame stabilizer profile and the flow channel profile; and extracting the flame stabilizer parameters and flow channel parameters of the ramjet engine based on a visual attention neural network, respectively, using the flame stabilizer profile and the flow channel profile as inputs, and encoding them into one-dimensional geometric parameters.

[0032] To transform abstract design parameters into a computer-processable digital geometric representation, a parametric geometric model of the ramjet engine is first constructed. This model, serving as a mathematical description of the physical structure, can automatically generate a solid model of the engine by calling a computer-aided design software interface, setting numerical geometric parameters such as the flow channel expansion angle, total combustion chamber length, support plate position coordinates, cavity depth, and aspect ratio as driving variables. Based on this, to meet the input requirements of subsequent deep learning models for regular tensor data, the generated geometric model undergoes mesh discretization. Specifically, a structured or regularized three-dimensional uniform mesh is constructed within the fluid domain, ensuring that mesh nodes are spatially uniformly distributed and that their topology maintains a one-to-one correspondence with physical spatial coordinates. Each mesh node contains its absolute coordinate information in the ramjet engine coordinate system (including X-axis, Y-axis, and Z-axis coordinates). The establishment of the three-dimensional uniform mesh not only clarifies the physical domain boundary for flow field prediction but also provides a definite spatial sampling benchmark for subsequent dimensionality reduction projection of the three-dimensional spatial information.

[0033] Two-dimensional features of the engine's geometry are extracted from a three-dimensional uniform mesh through spatial transformation operations. Along the flow direction of the ramjet engine, a cross-section perpendicular to the axis is selected at the center of the flame stabilizer. The grid coordinates at this cross-section are extracted from the three-dimensional uniform mesh and projected onto the cross-section to generate a flame stabilizer profile. Specifically, the extracted grid point coordinates are mapped onto the cross-section to form a binary or grayscale image reflecting the geometry of the flame stabilizer cross-section. The resolution of this image matches the spanwise number of grid points, and the grayscale values ​​of the image pixels are used to characterize whether the location belongs to the fluid domain or its distance from the axis. The flow channel profile is generated by projecting the three-dimensional uniform mesh coordinates onto a longitudinal section passing through the engine axis. It is mainly used to capture the overall profile features of the engine along the flow direction, such as the expansion trend of the flow channels and the distribution of the contraction and expansion sections of the combustion chamber. This projection operation transforms the complex three-dimensional geometric topology into two-dimensional image data, enabling efficient extraction of geometric features using mature computer vision algorithms. This effectively avoids the high computational complexity and difficulty in capturing local structural features associated with directly processing three-dimensional point cloud data.

[0034] By utilizing a neural network architecture with an attention mechanism, the automatic perception and encoding of key geometric features is achieved. The visual attention neural network used here refers to a combination of architectures such as convolutional neural networks and Transformers that incorporate self-attention or channel attention mechanisms. The network takes the generated flame stabilizer contour map and flow channel contour map as input. During feature extraction, the attention mechanism automatically assigns higher weights to key regions that significantly affect the flow field (such as the trailing edge of the cavity and the leading edge of the support plate, which are areas of strong shear), thereby suppressing background noise interference.

[0035] For the flame stabilizer profile, the network focuses on identifying the arrangement and thickness distribution of the support plates, as well as the geometry of the cavity. These features directly determine flame stability and fuel mixing, and are therefore extracted and encoded as flame stabilizer parameters. For the flow channel profile, the network focuses on extracting macroscopic geometric features such as the flow channel cross-sectional area variation and expansion angle. These features affect the compression and expansion processes of the main flow path, and are therefore extracted and encoded as flow channel parameters.

[0036] The encoding process aims to map the extracted high-dimensional semantic features into low-dimensional dense vectors. Specifically, a fully connected layer can be placed at the end of the network to flatten the feature map and map it into a one-dimensional vector of a preset length. This one-dimensional geometric parameter is represented as a continuous sequence of real numbers, which not only preserves the information of the original geometric parameters numerically but also implies the spatial topological relationships and potential physical connections between geometric structures. Through this encoding method, complex engine geometric information is transformed into a numerical form that can be directly processed by the fully connected neural network, providing accurate geometric boundary condition inputs for subsequent one-dimensional along-the-path parameter distribution prediction.

[0037] It should be noted that when projecting the grid coordinates of the three-dimensional uniform grid along the flow direction, the number of flame stabilizer profile diagrams is determined according to the number of stages of the flame stabilizer.

[0038] The cross-sectional projection operation is adaptively adjusted based on the actual arrangement of the physical flame stabilization structure within the ramjet engine. In the combustion chamber design of a ramjet engine, the flame stabilizer (such as a support plate or cavity) is a key component for controlling airflow mixing and flame stability. The number of flame stabilizer stages directly determines the number of core disturbance regions within the flow field. If a fixed number of flame stabilizer profiles are used, when the engine configuration changes from single-stage to multi-stage flame stabilization, the fixed projection slices may not cover all critical flame stabilization areas, resulting in the loss of geometric feature information for some flame stabilizers. Therefore, during the projection operation, the number of flame stabilizer stages is first identified based on the current engine configuration (or directly read from preset information). Then, based on the number of flame stabilizer stages, the position and number of cross-sections selected along the flow direction are determined to ensure that a corresponding flame stabilizer profile exists at the physical location of each flame stabilizer.

[0039] During system initialization or geometric parameter analysis, information about the flame stabilizer configuration is read from the geometric parameters, and the number of flame stabilizer stages is counted. Subsequently, along the combustion chamber flow direction, cross-sections are selected at the center positions of each flame stabilizer stage, and flame stabilizer profile diagrams are generated. For example, for a two-stage flame stabilizer configuration containing a primary support plate and a secondary support plate, the number of flame stabilizer profile diagrams is determined to be 2, corresponding to the cross-sections at the center positions of the primary and secondary support plates, respectively.

[0040] The coordinate information of all grid centers in the 3D uniform grid is extracted, and grid points whose coordinate values ​​match the cross-sectional coordinates at the center of the target flame stabilizer are selected. The coordinates or physical properties of these points are mapped to a 2D plane to form a flame stabilizer profile representing the flow channel cross-section at that location. Thus, the number of flame stabilizer profiles corresponds one-to-one with the number of flame stabilizer stages, enabling the subsequent visual attention neural network to extract specific geometric features for each independent flame stabilizer unit.

[0041] This adaptive, quantity-determining projection method ensures that the generated two-dimensional image completely covers all critical flame stabilizer structures inside the engine. This physically-driven dynamic projection mechanism avoids the omission of geometric features due to insufficient or misaligned projections, ensuring that the visual attention neural network in the above embodiment receives a complete and aligned input image regardless of the number of flame stabilizer stages in the engine configuration. This allows for accurate extraction of flow path parameters and flame stabilizer parameters, improving the model's generalization ability and encoding accuracy for engine configurations of varying complexity.

[0042] In one embodiment, generating a three-dimensional uniform mesh based on a geometric model includes: generating an unstructured mesh based on the geometric model, and interpolating the unstructured mesh into a predetermined three-dimensional uniform mesh.

[0043] When constructing the three-dimensional computational domain of a ramjet engine, unstructured meshes are used to accurately adapt to the engine's complex geometric boundaries, such as the structures of support plates, cavities, and flow channel walls. These meshes are typically composed of tetrahedral, pyramidal, or prism elements, and their flexible node distribution allows them to closely fit curved surfaces and meet the requirements for refinement in local high-gradient regions. The process of generating unstructured meshes based on the geometric model defines the physical spatial topology of the flow field, ensuring the geometric fidelity of subsequent flow field numerical simulations or data acquisition. Since the data arrangement of unstructured meshes lacks a regular three-dimensional array structure, it is difficult to directly adapt to neural network models that rely on regular data formats; therefore, it is necessary to map them onto a predetermined three-dimensional uniform mesh.

[0044] Interpolating unstructured meshes into a predetermined 3D uniform mesh aims to convert irregular data in physical space into a regular tensor format that can be processed by deep learning algorithms. This process calculates the flow field values ​​corresponding to each grid point in the predetermined 3D uniform mesh based on known flow field parameters (such as pressure, temperature, and Mach number) and their spatial coordinates at the nodes of the unstructured mesh using an interpolation algorithm. For example, a 3D linear interpolation method is used, utilizing multiple known nodes in the unstructured mesh closest to the target uniform mesh point, and determining the parameter values ​​of the target point based on inverse distance weights or a spatial shape function, thereby preserving the physical continuity of the original data. The predetermined 3D uniform mesh has a fixed and uniformly distributed number of grids in the X, Y, and Z axes, for example, a 600×100×100 size structure. Through this transformation, discrete and topologically complex unstructured mesh data is reconstructed into a standard 3D matrix form, satisfying the input requirements of deep learning models such as convolutional neural networks, and providing a unified coordinate reference for subsequent extraction of flame stabilizer and flow channel profile maps, thus achieving strict consistency between geometric feature encoding and flow field prediction in terms of data dimensions.

[0045] More specifically, when the combustion test parameters are the same as the test condition parameters, the neural network training in this embodiment includes the following steps: Step 1): Establish engine sample datasets with different geometric parameters and operating conditions; the specific steps are as follows: Step 1-1): Parametrically configure the ramjet engine geometry and extract key parameters (such as flow channel expansion ratio, overall length, and the position of the support plate / cavity). Determine the engine geometry design space and the design space for calculation conditions, and obtain a series of engine samples using orthogonal sampling.

[0046] In this step, the parametric modeling method uses a ramjet engine with a multi-stage support plate-cavity combustion chamber combination as the basic configuration; for example... Figure 2 As shown, the geometric parameters include flow channel shape, combustion chamber expansion angle, total length and combustion chamber outlet expansion ratio related to the flow channel, support plate position, number of support plates, support plate thickness, cavity length, cavity depth, cavity spacing and cavity depth-to-diameter ratio related to the first-stage support plate, first-stage cavity, second-stage support plate and second-stage cavity, etc. The operating parameters include flight Mach number, flight pressure and injection equivalence ratio of each stage, etc.

[0047] A neural network dataset was created by orthogonally sampling the values ​​of various geometric parameters. The dataset contains 180 samples in total, with 80% used for training and 20% for testing. For ease of understanding, three representative sets of samples from the dataset were selected, as shown in Table 1.

[0048] Table 1 Representative Samples Steps 1-2): Generate an engine geometric model based on the structural parameters of the ramjet engine, and generate an engine numerical simulation mesh based on the engine geometric model.

[0049] The numerical simulation mesh for the engine is an unstructured mesh with an average mesh size of 3 mm, a first boundary layer height of 0.1 mm, and a total mesh size of approximately 8 million to 10 million.

[0050] Steps 1-3): Based on the engine numerical simulation mesh, numerical simulations were performed on the configurations with different geometric parameters in Step 1-1) under different flight Mach numbers and injection equivalence ratios to obtain the original data of the engine's three-dimensional combustion flow field.

[0051] The numerical simulation used the SST k-omega turbulence model, and the fuel used was a substitute with the molecular formula C. 10 H 22 The kerosene fuel and chemical kinetic model used the eleven-component, ten-step overall reaction kinetic mechanism of kerosene, and the turbulent combustion interaction model used the PaSR model to obtain the raw data of the three-dimensional combustion flow field of the ramjet engine.

[0052] Steps 1-4): The original flow field data in Steps 1-3) has the same grid size as the grid in Step 1-2), which is an unstructured grid with an average size of 3 mm. The grid coordinate information and the original flow field data are interpolated into a new uniform grid to facilitate deep learning computation. The grid coordinates serve as the input values ​​in the deep learning dataset, and the flow field data serves as the labels in the deep learning dataset.

[0053] The interpolation method is three-dimensional linear interpolation, using nearest neighbor points weighted by distance. The grid information includes the absolute coordinates of each grid center point in the ramjet engine coordinate system, and the raw flow field data includes the pressure, temperature, and Mach number (Mach number refers to the flow Mach number at each grid point) for the corresponding grid. The uniform grid has 600, 100, and 100 grid cells along the X, Y, and Z axes, respectively. Therefore, the interpolated grid coordinate information and flow field data are both vectors of size 3×600×100×100. Figure 3 As shown, this is a visualization comparison diagram of unstructured mesh and original flow field data with interpolated uniform mesh and interpolated flow field data. The left figure shows unstructured mesh and original flow field data, and the right figure shows interpolated uniform mesh and interpolated flow field data.

[0054] Step 2): Encode and merge the engine's geometric information and operating conditions, and calculate the one-dimensional friction parameter distribution using a fully connected neural network. The specific steps are as follows: Step 2-1): Select a cross section perpendicular to the axis at the center of each support plate along the flow direction, and select a longitudinal section passing through the engine axis; project the grid coordinates of the three-dimensional uniform grid obtained in Step 1-4) onto the cross section and the longitudinal section respectively to obtain the flame stabilizer profile and the flow channel profile; use a visual attention neural network to extract the engine's flame stabilizer parameters and flow channel parameters respectively, and encode them into one-dimensional geometric parameters.

[0055] The ramjet engine was projected to produce three two-dimensional images, including two flame stabilizer profiles and one flow channel profile. The flow channel profile characterizes the engine's support plate position, expansion angle, and other flow channel geometric parameters; the two-dimensional image size is 600×100. The flame stabilizer profile characterizes the number and thickness of the support plates, and the two-dimensional image size is 100×100. Table 2 shows partial data from the flame stabilizer profile at the interface between the second-stage flame stabilizer support plate and the wall, where 0 represents the support plate (non-fluid domain), and the remaining data represent the absolute Y-axis coordinates of the grid center point.

[0056] Table 2. Examples of partial data from the flame stabilizer outline diagram. The visual attention neural network first crops and divides the three contour images into eight blocks, then uses fully connected layers to linearly map each of the eight blocks into eight one-dimensional vectors of size 1×128. A sinusoidal positional encoding is added to each one-dimensional vector to mark the positional relationships between the vectors representing the eight blocks. A multi-head self-attention mechanism is then used to map the input one-dimensional vectors into geometric feature vectors. This multi-head self-attention mechanism consists of four attention heads and four self-attention layers. The final output is a one-dimensional vector of size 2048.

[0057] Step 2-2): Use a fully connected neural network to map the operating parameters into one-dimensional operating parameters in vector form, and merge them with the one-dimensional geometric parameters in vector form from Step 2-1) to obtain one-dimensional test parameters.

[0058] Operating parameters include flight Mach number, flight pressure, first-stage injection equivalence ratio, and second-stage injection equivalence ratio.

[0059] The fully connected neural network consists of four fully connected layers. The inputs are the flight Mach number, flight pressure, first-stage jet equivalence ratio, and second-stage jet equivalence ratio. The output is a one-dimensional vector of size 256. Finally, it is merged with the vector from step 2-1). The first 10 values ​​of the outputs from steps 2-1) and 2-2) are shown in Table 3 below.

[0060] Table 3 Examples of output data for steps 2-1) and 2-2) Step 2-3): Using a fully connected neural network, input the merged one-dimensional test parameters obtained in step 2-2) and output the one-dimensional friction parameter distribution of the engine along the flow direction, including pressure, temperature and Mach number.

[0061] The fully connected neural network contains four fully connected layers. It takes the one-dimensional experimental parameters obtained by merging steps 2-1) and 2-2) as input and outputs a one-dimensional path parameter distribution containing pressure, temperature, and Mach number, with a size of 3×600. The first 20 values ​​of the output after partial pressure, temperature, and Mach number normalization are shown in Table 4 below.

[0062] Table 4. Examples of partial output data from a fully connected neural network Step 3): Broadcast the one-dimensional flow parameter distribution to a three-dimensional uniform grid to obtain the three-dimensional flow parameter distribution, thereby achieving dimensional matching between the one-dimensional average data and the three-dimensional flow field data, which can then be used as input to the three-dimensional neural network.

[0063] Step 4): Use a multi-path 3D neural network to predict the 3D pressure, temperature, and Mach number fields, output the predicted 3D combustion flow field, compare the calculation error with the true value, and carry out neural network training. Step 4-1): Combine the results of Step 3) with the grid coordinates obtained in Step 1-4) to obtain a three-dimensional flow field + grid information-based three-dimensional flow parameter distribution, which serves as the input value for the multi-path three-dimensional neural network.

[0064] The vector representing the grid coordinate information has three dimensions, each with a size of 600×100×100. The vector obtained in step 3) has three dimensions: pressure, temperature, and Mach number, each with the same size as the vector representing the grid coordinate information. The input value size of the multipath 3D neural network is 6×600×100×100.

[0065] Step 4-2): Divide the input values ​​into three location regions according to the isolation section of the engine combustion chamber, the first-stage combustion chamber, and the second-stage combustion chamber. Use a multi-path 3D neural network (composed of an encoder and a decoder) to extract the flow field features of the three regions respectively, and calculate the 3D combustion flow field; The region is divided along the flow direction according to the engine's structural characteristics. The multi-path 3D neural network consists of three encoders and one decoder. The encoders reduce the data size through convolutional blocks (composed of convolutional layers and pooling layers) and extract one-dimensional parametric features that integrate the engine's geometric features. The decoder restores the data size through deconvolutional layers and upsampling to reconstruct the 3D flow field. Each encoder and decoder contains four convolutional blocks and incorporates skip connections.

[0066] Step 4-3): Using the flow field data from Steps 1-4) as the ground truth for the neural network, train the neural network. Use the PyTorch neural network function library to build the training framework, training for a total of 500 epochs with a learning rate of 0.02. Save the optimal model parameters after training.

[0067] After the neural network training is completed, the model parameters are loaded. The new engine configuration is preprocessed according to steps 1-2) and 1-4) to obtain the grid coordinate information of a uniform grid. The combustion test parameters and grid coordinate information are used as input data to the neural network for forward inference to generate a three-dimensional flow field.

[0068] In one embodiment, the flight Mach number is Ma6, the flight pressure is 50 kPa, and the first-stage and second-stage equivalence ratios are 0.6 and 0.4, respectively. The engine expansion angle is set to 2°, the support plate is located at an expansion ratio of 1.6, and the remaining parameters are consistent with the baseline configuration in the dataset. Combustion test parameters and grid coordinate information are used as input data to pass to a neural network for forward inference to generate a three-dimensional flow field. Specifically, according to... Figure 5 (true_p represents the three-dimensional pressure field of the CFD) and Figure 6 A comparison of the CFD and predicted three-dimensional pressure fields (where pred_p represents the predicted three-dimensional pressure field) shows that the location and peak value of the pressure rise are basically consistent between the CFD and predicted three-dimensional pressure fields. After the first-stage cavity, the pressure in both results drops sharply, demonstrating consistency. Figure 7 (true_T represents the three-dimensional temperature field of the CFD) and Figure 8 A comparison of the three-dimensional temperature field predicted by the method of this invention (pred_T represents the three-dimensional temperature field predicted by the method of this invention) shows that both the CFD and the predicted three-dimensional temperature field exhibit a certain flame rise phenomenon behind the first-stage support plate, indicating that the method of this invention can capture the ignition delay behind the support plate. Furthermore, calculations show that the average error of the three-dimensional flow field is less than 5%, proving the effectiveness of the method of this invention.

[0069] This invention extracts the flame stabilizer profile and flow channel profile of the engine separately and encodes them using a visual attention neural network. This enables the perception of different geometric parameters of the engine and the learning of the differences in the three-dimensional combustion flow field generated by these different geometric parameters. This allows for rapid and accurate prediction of ramjet engine performance, reducing the time required for flow field calculations.

[0070] The multi-path three-dimensional neural network proposed in this invention divides the engine into three regions, uses three encoders to learn the flow characteristics of the three regions that have significant differences, and uses the same decoder to fuse the features and reconstruct the flow field.

[0071] This invention is applicable to combustion flow field prediction when engine geometry parameters and operating conditions change simultaneously, and has a high model generalization ability. In the combustion flow field data processing, the irregular computational grid is first mapped to the regular learning domain, and visual attention neural network, fully connected neural network, and multi-path three-dimensional neural network are used for step-by-step prediction, which can effectively improve the modeling accuracy and reduce the prediction error of the flow field.

[0072] The prediction method used in this invention employs a "one-dimensional parameter-three-dimensional flow field" prediction strategy, which introduces strong constraints into the neural network learning process that conform to the physical evolution law of the internal flow field of the ramjet engine. That is, it first captures the trend of the axial core physical parameters of the dominant flow field structure, and then reconstructs the three-dimensional spatial details under its guidance. This effectively reduces the complexity and uncertainty of directly fitting the high-dimensional, nonlinear three-dimensional flow field to the model, avoids the non-physical oscillations and error propagation commonly found in direct end-to-end prediction, and significantly improves the overall accuracy and reliability of three-dimensional flow field parameter prediction.

[0073] In another embodiment, when the combustion test parameters are the historical combustion flow field parameters of the previous N time moments, converting the combustion test parameters into one-dimensional test parameters corresponding to the three-dimensional uniform grid generated by the ramjet engine includes: obtaining the unstructured grid of the ramjet engine; interpolating the unstructured grid and the historical combustion flow field parameters into the predetermined three-dimensional uniform grid; and performing surface averaging calculation on the historical combustion flow field parameters in the three-dimensional uniform grid to obtain the one-dimensional test parameters of the N time moments.

[0074] Obtaining unstructured meshes is fundamental for numerical simulation analysis of complex combustion chamber geometries. Compared to structured meshes, unstructured meshes are more adaptable to complex boundaries such as internal supports, cavities, and injectors in ramjet engines, and can more accurately capture complex flow details such as the interaction between shock waves and boundary layers, and the shear layer formed by fuel injection. In this embodiment, the raw data for the unstructured mesh comes from a pre-performed CFD simulation. For high-quality unstructured meshes generated for a specific engine configuration, the average size can be set to 3 mm, the height of the first boundary layer mesh can be set to 0.1 mm, and the total number of meshes is typically between 4 million and 10 million to ensure the convergence accuracy of the numerical solution.

[0075] To eliminate the spatial irregularities of the original CFD data and transform it into a regular tensor form that is easily processed by deep learning models, the unstructured mesh and the historical combustion flow field parameters attached to its nodes need to be mapped to a predetermined three-dimensional uniform mesh. Since the spatial distribution of nodes in the unstructured mesh is disordered, it cannot be directly used as input to convolutional neural networks or fully connected neural networks. Therefore, a regular three-dimensional uniform mesh needs to be constructed as an intermediate carrier.

[0076] The predetermined three-dimensional uniform grid is a regular spatial discrete domain pre-divided according to the overall geometric dimensions and computational accuracy requirements of the ramjet engine, with a fixed number of grids along the flow direction, transverse direction, and spanwise direction. For example, the size of the three-dimensional uniform grid can be set to 600 nodes along the flow direction, 100 nodes along the transverse direction, and 100 nodes along the spanwise direction, thus forming a 600×100×100 data matrix.

[0077] The interpolation process can employ a three-dimensional linear interpolation algorithm. For non-fluid domain grid points located on the engine wall or inside the support plate, their flow field parameters are filled with specific values ​​(such as 0) to distinguish between the fluid domain and the solid domain. After this interpolation process, the historical combustion flow field parameters, originally scattered across millions of unstructured nodes, are reconstructed into structured three-dimensional tensor data, and a strict one-to-one correspondence is established between the flow field parameters and the grid coordinates, providing a standardized data foundation for subsequent surface averaging calculations.

[0078] Performing surface averaging calculations on historical combustion flow field parameters in a three-dimensional uniform grid is a core step in reducing the dimensionality of high-dimensional spatiotemporal flow field data and extracting one-dimensional flow evolution features. This process aims to extract the dominant physical laws along the combustion chamber flow direction from the complex three-dimensional spatial distribution, simplifying the three-dimensional flow field at each moment into a one-dimensional flow distribution, thereby adapting to the input requirements of fully connected neural networks for time series prediction.

[0079] For each moment (in The following averaging calculation logic is executed: For all transverse and spanwise grid points with the same flow direction coordinates in the three-dimensional uniform grid, their corresponding flow field parameter values ​​are extracted. By calculating the arithmetic mean of the parameters of all fluid domain grid points within the cross section, the one-dimensional average parameter at that flow direction location is obtained.

[0080] Through the above surface averaging calculation, the original three-dimensional flow field data (including pressure, temperature, and Mach number) is compressed into a one-dimensional vector of size 3×600. For the historical data of the first N time steps, this one-dimensional vector can be further downsampled or directly concatenated to finally form an input tensor of size N×250 or N×600, which represents the one-dimensional experimental parameters for the N time steps. This one-dimensional processing not only greatly reduces the computational complexity of the subsequent fully connected neural network, but also effectively filters out non-uniform disturbances in the lateral and spanwise directions, retaining one-dimensional macroscopic features that can reflect the overall performance of the combustion chamber (such as total pressure recovery and flame propagation front position), thereby significantly improving the prediction accuracy and robustness of transient processes such as forced ignition.

[0081] Through the above process, the geometric topological information of the unstructured grid and the transient physical information of the historical flow field are mapped to a unified regular grid space, and the one-dimensional features along the flow path are extracted by surface averaging calculation, which constitutes a solid data input basis for predicting the evolution of unsteady flow fields.

[0082] It should be noted that when faced with different types of combustion test parameters, fully connected neural networks and multi-path three-dimensional neural networks are trained separately.

[0083] For the two types of combustion test parameters with significantly different physical properties and data distributions, independent training processes are configured to adapt to different mapping mechanisms. Specifically, the different types of combustion test parameters cover two main input modes: one is a combination of static parameters defined based on the ramjet engine's geometry and actual operating conditions, corresponding to the test condition parameter scenarios described in the previous embodiments; the other is a sequence of historical combustion flow field parameters characterizing the dynamic evolution of the combustion flow field over time, corresponding to the historical combustion flow field parameter scenarios for the first N time moments described in the previous embodiments. Given that the former focuses on the spatial mapping of geometric boundary conditions and steady-state operating parameters to the final flow field, while the latter focuses on the evolution of the flow field itself in the time dimension, using a unified neural network model to process mixed data can easily lead to model convergence difficulties or decreased prediction accuracy due to differences in physical mechanisms. Therefore, this embodiment adopts a type-specific independent training strategy: for the first type of test condition parameters, a fully connected neural network and a multi-path 3D neural network are constructed and trained specifically for learning the geometry-operating condition-to-flow field mapping relationship; for the second type of historical combustion flow field parameters, a fully connected neural network and a multi-path 3D neural network are constructed and trained specifically for learning the relationship between time series and future flow field prediction.

[0084] The specific implementation of this training strategy first involves dividing the original training dataset according to the type of combustion test parameters, forming a geometric condition dataset and a time-series historical dataset.

[0085] For the geometric condition dataset, experimental condition parameters are used as network inputs, and corresponding 3D combustion flow field simulation data or experimental measurement data are used as ground truth labels. Supervised learning training is performed on the fully connected neural network and the multi-path 3D neural network. In this process, the fully connected neural network is optimized to accurately extract geometric and condition features and output a one-dimensional flow parameter distribution; the multi-path 3D neural network reconstructs a high-precision 3D spatial flow field based on the one-dimensional distribution, thereby establishing an accurate mapping from static conditions to the spatial flow field.

[0086] For the time-series historical dataset, the sequence of historical combustion flow field parameters from the first N time steps is used as the network input, and the three-dimensional combustion flow field data at time step N+1 is used as the ground truth label. This is used to train a separate set of fully connected neural networks and a multi-path three-dimensional neural network. Through this process, the fully connected neural network captures and grasps the temporal evolution trend of the flow field, while the multi-path three-dimensional neural network, based on this, achieves spatial prediction of the transient flow field, thereby establishing a dynamic prediction model of the flow field from historical time series to future time steps.

[0087] By training independently for different input paradigms, the above method effectively isolates two learning tasks with significantly different physical mechanisms, avoiding the negative transfer effect that a single model might produce when processing heterogeneous data. This task separation mechanism allows each neural network to focus on feature extraction and pattern learning within its own domain, significantly reducing the complexity of model optimization. Furthermore, this method improves generalization performance and prediction accuracy in configuration optimization design scenarios and transient ignition prediction scenarios, ensuring the effectiveness and robustness of the ramjet engine three-dimensional combustion flow field prediction method under different engineering application requirements.

[0088] More specifically, the encoder consists of several concatenated convolutional blocks; the multipath 3D neural network also includes a decoder, which consists of several concatenated convolutional blocks; the convolutional blocks in the encoder and the convolutional blocks in the decoder have skip connections.

[0089] As the core feature extraction unit of a multi-path 3D neural network, the encoder is responsible for extracting hierarchically expressed flow field features from the input 3D flow parameter distribution. This encoder employs a deep architecture design, consisting of several convolutional blocks connected in sequence. This stacked structure allows the network to progressively capture flow field patterns from local details to global distributions. Each convolutional block typically contains basic components such as convolutional layers, activation functions, and normalization layers. Convolutional operations slide across 3D space to aggregate neighborhood information, and activation functions introduce nonlinear transformations to enhance the model's expressive power. As the number of convolutional layers increases, the encoder's receptive field gradually expands, and the identified flow field structure evolves from basic edge features such as shock surfaces and shear layers to more complex semantic features such as vortex structures, recirculation zones, and combustion reaction zones. This gradual feature extraction process provides a rich and complete feature representation foundation for subsequent flow field reconstruction.

[0090] The decoder is responsible for mapping the high-dimensional abstract features extracted by the encoder back to the reconstruction of the three-dimensional physical flow field space. Its architecture corresponds to that of the encoder, also consisting of several cascaded convolutional blocks. In the complete topology of the multi-path 3D neural network, the decoder is located after the encoder. It gradually restores the spatial resolution of the flow field data through deconvolution (transposed convolution) and upsampling operations, ultimately outputting a 3D combustion flow field prediction result with the same input size. The cascaded sequence of convolutional blocks in the decoder is responsible for decoding the compressed feature map level by level. Each level of convolutional block not only uses the features of the current layer for reconstruction but also integrates contextual information from different levels of the encoder. Through this multi-level cascaded structure, the decoder can transform the abstract physical laws learned in the encoding stage into specific flow field parameter distributions, such as the precise three-dimensional spatial arrangement of pressure, temperature, and Mach number, thereby achieving a reliable mapping from the high-dimensional feature space to the low-dimensional physical space.

[0091] Skip connections constitute a key link in the multi-path 3D neural network architecture, aiming to solve the gradient vanishing problem in deep network training and promote the effective fusion of cross-layer features. This connection mechanism directly skips the output features of a specified convolutional block in the encoder to the corresponding convolutional block in the decoder, and then performs element-wise concatenation with the upsampled features from the previous level at the input of the decoder convolutional block. This cross-layer connection allows the decoder to directly utilize the high-resolution geometric and texture information retained in the shallow network of the encoder when reconstructing flow field details, such as the boundary layer gradient of the combustion chamber wall, the initial jet shape of the nozzle, and the sharp edges of the flame front. Compared with the traditional autoencoder structure, skip connections avoid information loss during multiple downsampling and encoding processes, ensuring the integrity of high-frequency detail information in the flow field. In actual operation, the data stream flows from the encoder to the decoder, and the residual channel acts as a bypass to inject the original low-level physical features into the reconstruction path without loss, so that the final output 3D combustion flow field maintains the rationality of global physical laws while possessing extremely high local spatial accuracy and boundary fidelity.

[0092] The encoder extracts high-dimensional abstract features of the flow field layer by layer through concatenated convolutional blocks, providing a semantic foundation for flow field reconstruction. The decoder decodes the features level by level through concatenated convolutional blocks and restores the spatial resolution, realizing the mapping from features to the flow field. Skip connections serve as cross-layer information channels between the encoder and decoder, directly injecting high-resolution detailed features from the encoder's shallow layers into the corresponding layers of the decoder. The three work together to form a closed-loop feature transfer system of extraction-transfer-reconstruction, effectively solving the problems of information loss and degradation in deep networks, and ensuring the dual accuracy of the 3D combustion flow field prediction results of the ramjet engine in terms of both macroscopic laws and microscopic details.

[0093] In one specific embodiment, the specific execution process is given when the combustion test parameters are the historical combustion flow field parameters of the previous N time moments.

[0094] Step 1): Obtain a dataset of forced ignition processes (e.g., tens of milliseconds) under different operating conditions; the specific steps are as follows: Step 1-1): For a certain type of engine, confirm the inflow conditions (flight Mach number and flight pressure at the combustion chamber inlet) and the injection equivalence ratio (first-stage support plate and second-stage support plate) calculation conditions.

[0095] The engine is a multi-stage strut-cavity combustor combination engine; operating parameters include flight Mach number, flight pressure, and injection equivalence ratio of each stage. The total number of datasets is 180, of which 80% are used for training and 20% for testing.

[0096] Steps 1-2): Generate the numerical simulation mesh for the engine.

[0097] The numerical simulation mesh for the engine is an unstructured mesh with an average mesh size of 3 mm, a first boundary layer height of 0.1 mm, a total mesh size of approximately 4 million, and an output time interval of 0.2 ms.

[0098] Steps 1-3): Conduct unsteady-state numerical simulations, using the high enthalpy jet of the methane rocket in the concave cavity (second stage concave cavity) for forced ignition of the engine, and accumulate raw data of the engine's three-dimensional combustion flow field ignition process (Mach number, pressure, and temperature of the flow field). The numerical simulation used the SST k-omega turbulence model, and the fuel used was a substitute with the molecular formula C. 10 H 22 The chemical kinetic model for kerosene fuel uses the eleven-component, ten-step overall reaction kinetic mechanism of kerosene, and the turbulent combustion interaction model uses the PaSR model.

[0099] Step 2): Interpolate the original mesh and flow field data obtained in Steps 1-2) and 1-3) to a uniform engine mesh that can adapt to deep learning computation. Use the flow field information from the previous N time steps to predict the flow field at time N+1. Predict the one-dimensional friction parameter distribution through a fully connected neural network. The specific steps are as follows: Step 2-1): The raw flow field data obtained in Steps 1-3) is based on an unstructured mesh with an average size of 3 mm. To adapt to the computational requirements of the deep learning model, an interpolation algorithm (such as linear interpolation, nearest neighbor interpolation, or cubic interpolation) is used to uniformly interpolate and map the unstructured mesh and its corresponding raw flow field data to a set of preset three-dimensional uniform meshes. The number and size of these three-dimensional uniform meshes are not limited by the original unstructured mesh.

[0100] The interpolation method is three-dimensional linear interpolation, and the non-fluid domain is filled with 0; the original flow field data includes the grid point coordinates, Mach number, pressure and temperature of the three-dimensional flow field.

[0101] Step 2-2): For a set of interpolated unsteady flow field data, perform surface averaging calculations on the flow field information of the first N time steps to obtain the one-dimensional parameter distribution along the flow path of the first N time steps.

[0102] Construct a fully connected neural network, taking the average one-dimensional parameter distribution along the surface at the first N time steps as input and the average one-dimensional parameter distribution along the surface at time N+1 as output, and train the fully connected neural network model.

[0103] The value of N can be adjusted according to the task, with a minimum of 3; the fully connected neural network contains four fully connected layers, with the number of neurons being 750, 1024, 900 and 600.

[0104] First, the input includes information on the Mach number, pressure, and temperature in the three-dimensional flow field for the first N time steps, with a grid size of 45 (horizontal) × 60 (vertical) × 500 (flow direction).

[0105] A surface averaging calculation is performed on the three-dimensional parameter field to obtain a surface-averaged one-dimensional parameter distribution with a size of 500. Then, it is downsampled to a vector of size 250, and the downsampled surface-averaged one-dimensional parameter distributions at N time points are concatenated to obtain a vector of size N×250, which is used as the input of the fully connected network.

[0106] Step 3): The one-dimensional parameter distribution is uniformly extended and mapped onto a three-dimensional uniform grid to achieve dimensional matching between the one-dimensional average data and the three-dimensional flow field data, which serves as the input to the three-dimensional neural network.

[0107] The one-dimensional output is broadcast to three-dimensional space, and the one-dimensional profile information is extended along the horizontal spatial dimension. Spatial correction is performed according to the division of fluid domain and non-fluid domain. This information is used as the input of the multi-path three-dimensional neural network. In this embodiment, the multi-path three-dimensional neural network has three input and output channels. Mach number, pressure, and temperature are used as inputs. The Mach number, pressure, and temperature are predicted simultaneously, and high-precision and high-efficiency prediction of complex flow phenomena is achieved. The multipath 3D neural network consists of three encoders, each with five convolutional layers, and one decoder. The number of feature channels in the encoders is 8→16→32→64→128, and in the decoder it is 128→64→32→16→8. It uses 3×3×3 convolutional kernels for learning, with each kernel automatically learning to recognize specific flow field structure patterns. A feature representation system from microscopic to macroscopic is constructed through multi-layer stacking. A bottleneck layer connects the encoder and decoder to compress information and extract the essential physical laws controlling the evolution of the flow field. This prevents overfitting and ensures the model can capture common flow characteristics under different operating conditions. Skip connections are added to achieve the fusion of shallow and deep features.

[0108] Step 4): After the neural network training is complete, input the surface-averaged one-dimensional parameters for the first N time steps, and output the three-dimensional flow field at time step N+1. Through M iterations, obtain the unsteady three-dimensional flow field information for all time steps from N+1 to N+M.

[0109] This invention can obtain the corresponding result at time N+1 based on the surface-averaged one-dimensional parameter distribution of the first N time steps, reducing the difficulty of modeling unsteady information and enabling rapid and accurate prediction of unsteady processes such as pressure rise and flame propagation during the forced ignition process of a ramjet engine, thereby reducing the time required for flow field calculation.

[0110] like Figure 9-14 The figure shown is a comparison of the predicted results of the forced ignition process of the ramjet engine at 10ms and 30ms in this invention. In this embodiment, the flight Mach number is Ma4, and the equivalence ratios of the first and second stages are 0.5 and 0.5, respectively. Figure 9 The normalized temperature distribution result obtained from CFD at 10ms. Figure 10 The normalized temperature distribution obtained using the method of this invention at 10 ms is the result. Figure 11 The relative error between the method of this invention and CFD at 10ms; Figure 12 The normalized temperature distribution obtained from CFD at 30ms is the result. Figure 13 The normalized temperature distribution obtained using the method of this invention at 30ms is the result. Figure 14 The relative error between the method of this invention and the CFD results at 30ms.

[0111] As shown in the figure, the temperature field prediction results obtained by the flow field prediction method of this invention can reproduce the spatial distribution characteristics of the temperature field obtained by numerical simulation quite well. At 10ms and 30ms, the high-temperature region, low-temperature region, and temperature gradient change region in the prediction results of this invention are basically consistent with the corresponding regions in the CFD results in terms of position and shape, indicating that this invention can capture the main characteristics of the flow field evolution over time. At the same time, the error results show that the prediction error is mainly distributed in the local strong gradient region and the geometric structure change region, while the prediction error of the main flow field region is relatively small. Therefore, this invention can effectively predict the flow field temperature distribution at different times without having to perform a complete numerical iteration solution again.

[0112] The present invention also discloses a data-driven three-dimensional combustion flow field prediction device for ramjet engines, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement any of the above methods.

[0113] The present invention also discloses an embodiment that provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0114] The present invention also provides a computer program product that, when run on a data storage device, enables the data storage device to implement the steps in the above-described method embodiments.

[0115] If the integrated unit module is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a storage device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0116] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0117] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0118] It should be noted that the data used in the implementation of this invention were all collected or gathered through legal and compliant channels, and the collection and gathering activities fully comply with the requirements of relevant laws, regulations and industry standards; the existing technical methods involved in this invention were also obtained and used through legal and compliant means.

Claims

1. A data-driven method for predicting the three-dimensional combustion flow field of a ramjet engine, characterized in that, Includes the following steps: Obtain combustion test parameters of the ramjet engine; the combustion test parameters are historical combustion flow field parameters or test condition parameters for the previous N time moments, the historical combustion flow field parameters include temperature, pressure and Mach number, and the test condition parameters include the geometric parameters and operating condition parameters of the ramjet engine; The combustion test parameters are converted into one-dimensional test parameters corresponding to the three-dimensional uniform grid generated by the ramjet engine. The one-dimensional test parameters are input into a fully connected neural network to obtain a one-dimensional parameter distribution along the process. The one-dimensional friction parameter distribution is extended and mapped onto the three-dimensional uniform grid to obtain the three-dimensional friction parameter distribution. The three-dimensional flow parameter distribution is input into a multi-path three-dimensional neural network to obtain the predicted three-dimensional combustion flow field of the ramjet engine.

2. The data-driven three-dimensional combustion flow field prediction method for ramjet engines as described in claim 1, characterized in that, Inputting the three-dimensional path parameter distribution into a multi-path three-dimensional neural network includes: Determine the location region of the three-dimensional friction parameters distributed in the ramjet engine; The three-dimensional path parameter distributions of different location regions are fed into different encoders of the multi-path three-dimensional neural network.

3. The data-driven three-dimensional combustion flow field prediction method for ramjet engines as described in claim 2, characterized in that, When the combustion test parameters are test condition parameters, converting the combustion test parameters into one-dimensional test parameters corresponding to the three-dimensional uniform grid generated by the ramjet engine includes: The geometric parameters are encoded as one-dimensional geometric parameters; The one-dimensional geometric parameters and one-dimensional operating parameters are combined to obtain the one-dimensional test parameters; wherein, the one-dimensional operating parameters are obtained by mapping the operating parameters.

4. The data-driven three-dimensional combustion flow field prediction method for ramjet engines as described in claim 3, characterized in that, Encoding the geometric parameters into one-dimensional geometric parameters includes: A geometric model of the ramjet engine is generated based on the geometric parameters, and a three-dimensional uniform mesh is generated based on the geometric model. A cross-section perpendicular to the axis is selected at the center of the flame stabilizer along the flow direction of the ramjet engine; Select a longitudinal section passing through the engine axis; The grid coordinates of the three-dimensional uniform grid are projected onto the cross section and longitudinal section to obtain the flame stabilizer outline and the flow channel outline, respectively. Using the flame stabilizer profile and flow channel profile as inputs, the flame stabilizer parameters and flow channel parameters of the ramjet engine are extracted based on a visual attention neural network and encoded as one-dimensional geometric parameters.

5. The data-driven three-dimensional combustion flow field prediction method for ramjet engines as described in claim 4, characterized in that, Generating a three-dimensional uniform mesh based on the geometric model includes: An unstructured mesh is generated based on the geometric model, and the unstructured mesh is interpolated into a predetermined three-dimensional uniform mesh.

6. The data-driven three-dimensional combustion flow field prediction method for ramjet engines as described in claim 4, characterized in that, When projecting the grid coordinates of a three-dimensional uniform grid along the flow direction, the number of flame stabilizer profile diagrams is determined according to the number of flame stabilizer stages.

7. The data-driven three-dimensional combustion flow field prediction method for ramjet engines as described in claim 2, characterized in that, When the combustion test parameters are historical combustion flow field parameters from the previous N time moments, converting the combustion test parameters into one-dimensional test parameters corresponding to the three-dimensional uniform grid generated by the ramjet engine includes: Obtain the unstructured mesh of the ramjet engine; The unstructured mesh and historical combustion flow field parameters are interpolated into a predetermined three-dimensional uniform mesh; The historical combustion flow field parameters in the three-dimensional uniform grid are calculated by surface averaging to obtain one-dimensional experimental parameters at N time points.

8. The data-driven three-dimensional combustion flow field prediction method for ramjet engines as described in claim 2, characterized in that, For different types of combustion test parameters, fully connected neural networks and multi-path three-dimensional neural networks were trained respectively.

9. The data-driven three-dimensional combustion flow field prediction method for ramjet engines as described in claim 2, characterized in that, The encoder consists of several concatenated convolutional blocks; The multipath 3D neural network also includes a decoder, which consists of several concatenated convolutional blocks; The convolutional blocks in the encoder and the convolutional blocks in the decoder have skip connections.

10. A data-driven three-dimensional combustion flow field prediction device for ramjet engines, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1-9.