A hypersonic vehicle aerodynamic heat prediction method, device and medium

Abstract

The application discloses a hypersonic vehicle aerodynamic heat prediction method and device and a medium, and relates to the technical field of aerospace heat protection. In view of the problem that the traditional aerodynamic heating prediction of a hypersonic vehicle under complex working conditions is time-consuming, a neural operator prediction model is constructed by taking the three-dimensional geometric shape of the surface of the vehicle and macro flight state parameters as physical inputs. Through multi-scale physical feature extraction and adaptive fusion, the multi-scale evolution process of shock compression and viscous dissipation in the flow field is represented, and a mapping relationship from the macro flight condition to the full-field heat flux density distribution of the surface of the vehicle is established. The application reduces the dependence of the traditional method on artificial grid division, realizes high-precision prediction on complex configurations such as blunt cones and lifting bodies, has small errors in the stagnation point and leading edge high-gradient areas, and improves the evaluation efficiency and physical consistency of the heat protection system design.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of thermal protection design for aerospace vehicles and artificial intelligence, and specifically relates to a method, equipment and medium for predicting the aerodynamic heat of hypersonic vehicles. Background Technology

[0002] In the development of hypersonic vehicles, accurately predicting their aerodynamic heating during actual flight trajectories is a crucial prerequisite for the refined design of thermal protection systems (TPS). During hypersonic flight, shock wave compression and viscous dissipation are significant near the vehicle's nose and leading edge, resulting in high thermal loads. The fluid's kinetic energy is converted into internal energy, generating a high-temperature boundary layer near the surface. Heat is primarily transferred to the vehicle surface through convection and viscous diffusion.

[0003] Currently, major aerodynamic heating prediction methods suffer from the following limitations: First, traditional computational fluid dynamics (CFD) methods can provide detailed flow and temperature field information, but when applied to complex aircraft configurations spanning the entire mission profile or lifecycle, they face high computational costs and long processing times. This limits their use in iterative design optimization, real-time evaluation, or uncertainty quantification involving multiple flight conditions. Second, while traditional wind tunnel experiments and semi-empirical formulas (such as Fay-Riddell correlations and enthalpy-based approximations) provide engineering methods, their experimental costs are high, measurement coverage is sparse, and prediction accuracy decreases in complex flow field regions. Third, although the pure data-driven surrogate models introduced in recent years have high computational efficiency, they often struggle to maintain physical consistency and are often difficult to generalize to unknown configurations outside their predefined training domain. Fourth, networks that partially incorporate physical constraints or graph networks specifically designed for irregular data still require significant time for manual feature engineering and prior feature selection (such as estimating boundary layer edge features) when handling aerodynamic heating prediction. This reliance on explicit geometric feature extraction reduces modeling efficiency and limits the model's ability to accelerate in practical applications. Therefore, it is necessary to develop an efficient modeling framework that can both preserve the physical laws represented by the underlying governing equations and achieve generalization across multiple aerodynamic configurations, thereby improving prediction accuracy and physical interpretability while maintaining computational efficiency. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, and addressing the difficulty in simultaneously achieving computational efficiency, generalization ability, and physical consistency in existing aerodynamic thermal prediction methods, the present invention aims to provide an aerodynamic thermal prediction method, computer equipment, and storage medium for hypersonic vehicles. The invention aims to construct an end-to-end neural operator prediction model to predict the surface heat flux density distribution of complex hypersonic vehicles under actual flight trajectories. This model is suitable for the rapid prediction of the surface aerodynamic heat flux density distribution of complex-shaped hypersonic vehicles and helps improve the physical consistency of the prediction results.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Firstly, a method for predicting the aerodynamic heat of a hypersonic vehicle is provided, comprising the following steps: Step 1: Obtain multimodal input data of the surface of the aircraft to be predicted. The multimodal input data includes three-dimensional spatial coordinates r = ( x , y , z And the flight status parameters associated with that spatial coordinates. f (r), wherein the flight state parameters include at least Mach number, flight altitude and angle of attack; Step 2 involves preprocessing and feature encoding the multimodal input data. Specifically, the three-dimensional spatial coordinates are Fourier encoded using multiple log-interval frequency bands to map low-dimensional coordinate features into high-dimensional position embedding information, thereby enhancing the model's ability to perceive high-frequency spatial details. For example, the Fourier encoding process of the three-dimensional spatial coordinates uses six log-interval frequency bands for high-dimensional position feature extraction. Step 3: Integrate the high-dimensional position embedding information with the flight state parameters. f (r) After concatenation, the input nonlinear multilayer perceptron preprocessor is projected onto the hidden dimension and a learnable global bias vector is broadcast to inject global contextual information. Step 4: Use the adaptive fusion module to enhance and fuse local and global information, and input the enhanced multimodal features into an encoder containing multiple physical scale blocks. Perform multi-scale clustering of surface spatial nodes at different resolutions to capture complex flow field features. Step 5: Input the multi-scale flow field features output by the encoder into the decoder, and perform progressive refinement processing through layer normalization, physical perception attention mechanism and multilayer perceptron operation in sequence; Step 6: Through the output projection module, the decoded hidden representation is mapped to the full-field heat flux density distribution on the aircraft surface to obtain the aerodynamic thermal prediction results.

[0006] In a second aspect, a computer device is provided, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the hypersonic vehicle aerothermal prediction method described in the first aspect.

[0007] Thirdly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the hypersonic vehicle aerothermal prediction method described in the first aspect.

[0008] Compared with the prior art, the beneficial effects of the present invention are: 1) Improved end-to-end prediction and computation efficiency.

[0009] This invention constructs an end-to-end neural operator architecture, reducing the reliance on manual mesh generation and manual prior feature extraction in traditional methods. The model learns a nonlinear mapping from flight parameters (Mach number, altitude, angle of attack) and geometric features to the overall heat flux density, which can reduce the computation time compared to high-fidelity CFD simulations, support faster flow field evaluation, and is suitable for rapid aerodynamic and thermal design optimization and trajectory-based risk assessment.

[0010] 2) Generalization ability and prediction accuracy.

[0011] By employing multi-scale geometric encoding and an adaptive fusion mechanism, this invention can balance local large gradient flow characteristics with gradient spatial distribution. On four representative configurations—blunt cone, bicone, biellipsoid, and lifting body—the average relative errors are 1.13%, 1.00%, 0.97%, and 0.47%, respectively. In stagnation points and leading-edge regions where traditional neural networks are prone to severe oscillations and phase shifts, the relative error of this invention can be maintained within 1%, approaching the accuracy of computational fluid dynamics benchmarks.

[0012] 3) Physical consistency and interpretability.

[0013] Physical systems inherently exhibit multi-scale characteristics. This invention introduces a multi-scale physical awareness attention mechanism to simultaneously capture interactions across multiple scales. By performing multi-scale clustering on the latent representation and adaptively fusing local and global attention, the model aligns the network attention distribution with the aerodynamic feature distribution of CFD samples. Regions with significant shock compression, strong viscous dissipation, and high physical correlation are assigned relatively high representation weights, which helps ensure that the prediction results conform to the physical mechanisms characterized by the three-dimensional unsteady compressible Navier-Stokes equations and Fourier's law of heat conduction. Attached Figure Description

[0014] Figure 1This is a schematic diagram of the end-to-end network architecture and information flow of the SynoPhys multi-scale neural operator in an embodiment of the present invention.

[0015] Figure 2 This is a comparison cloud map showing the aerodynamic heat flux density distribution prediction results of the collaborative multi-scale neural operator on blunt cone, bicone, biellipsoid and lifting body configurations in the embodiments of the present invention, and the CFD calculation reference values.

[0016] Figure 3 This is a cloud map showing the relative error distribution of the model prediction in an embodiment of the present invention.

[0017] Figure 4 The principal component analysis clustering results of the model's internal latent features in this embodiment of the invention are used to characterize the model's physical interpretability.

[0018] Figure 5 This is a spatial distribution diagram of the attention feature norm of the latent features inside the model in this embodiment of the invention, used to characterize the physical interpretability of the model. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] To address the high cost of traditional computational fluid dynamics in calculating full-mission profiles for complex configurations, and the limitations of existing pure data-driven models such as insufficient physical consistency and poor generalization ability, this invention proposes an aerodynamic-thermal prediction method for hypersonic vehicles, building an end-to-end neural operator prediction architecture based on SynoPhys. This method acquires the vehicle's spatial coordinates and flight state parameters, and captures both global and local flow field characteristics by combining multi-scale geometric encoding with an adaptive fusion mechanism.

[0021] like Figure 1 As shown, the method of this invention mainly includes steps such as data acquisition, feature encoding, adaptive fusion, multi-scale extraction, decoding mapping, and model training. The SynoPhys architecture is based on the Transformer mechanism, allowing the model to capture long-range spatial dependencies within the functional data space. By mapping geometric features, flow conditions, and trajectory parameters to the target field, the need for intermediate transformations or manual feature extraction is reduced. Detailed descriptions follow: Step 1: Obtain multimodal input data of the surface of the aircraft to be predicted.

[0022] The multimodal input data of this invention includes three-dimensional spatial coordinates r = ( x , y ,z And the flight status parameters associated with that spatial coordinates. f (r) The flight status parameters reflect the current flight conditions and include at least Mach number, flight altitude, and angle of attack. The three-dimensional spatial coordinates can be obtained by discretizing the aircraft's geometric model, while the flight status parameters can be obtained from airborne sensors or mission profile data.

[0023] Step 2: Preprocess and feature encode the obtained multimodal input data.

[0024] The preprocessing here mainly refers to converting the original three-dimensional coordinates into a data tensor format suitable for neural networks to read and compute; feature encoding is to solve the problem that low-dimensional coordinates are difficult to represent complex geometric changes. By using six log-interval frequency bands to perform Fourier encoding on the three-dimensional spatial coordinates, the low-dimensional coordinates are mapped to high-dimensional position embedding information to obtain their high-dimensional position features, thereby extracting high-frequency detail features in space.

[0025] Step 3: The obtained high-dimensional position embedding information is concatenated with the corresponding flight state parameters and used as input to the nonlinear multilayer perceptron preprocessor. This is then projected onto the hidden dimension, and a learnable global bias vector is broadcast to inject global contextual information.

[0026] In this step, the nonlinear multilayer perceptron preprocessor consists of several fully connected layers and nonlinear activation functions. The specific projection method involves multiplying the weight matrices and adding a bias to uniformly map the concatenated multimodal feature dimensions to the hidden layer dimensions (e.g., 128 dimensions) set by the network.

[0027] The broadcasting process of the learnable global bias vector involves defining a global feature vector in the network, copying it in the spatial dimension (Broadcast), and adding it to the features of each local spatial coordinate point, so that each local node can perceive the global context of the flow field.

[0028] Step 4: Use the adaptive fusion module to combine local and global information for feature enhancement, and input the enhanced features into an encoder containing multiple physical scale blocks. Perform multi-scale clustering at different resolutions to capture flow field features.

[0029] In this step, the adaptive fusion module can be a feature-weighted network containing learnable weights (such as a gating mechanism).

[0030] The aforementioned "local information" comes from the spatial features after Fourier encoding and the current flight parameters, while the "global information" comes from the broadcast global bias vector. This module fuses the two, so that the features contain both local details and global information.

[0031] The fused features are fed into the encoder, which consists of a series of cascaded physical scale blocks stacked together. Each physical scale block includes pre-normalization processing, a physical awareness attention module, and a feedforward multilayer perceptron, and residual connections are applied after each operation to mitigate gradient vanishing.

[0032] Specifically, a single physical scale block first performs pre-normalization processing on the input features, then calculates the correlation weights between different surface nodes according to the multi-scale clustering results through the physical perception attention module, and adds the weighted aggregated features with the original input features as residuals; the residual output is then input into the feedforward multilayer perceptron for nonlinear transformation, and is output to the next physical scale block again through residual connection.

[0033] Preferably, the resolution of the multi-scale clustering in this step is preset to 16, 32, and 64.

[0034] The execution mechanism of the physical perception attention module includes: performing layer normalization on the input features to stabilize the gradient distribution during training; and limiting the query range of each node to physically adjacent and motionally similar regions based on the surface node neighborhood obtained by multi-scale clustering. Within this neighborhood, query, key, and value features are generated through linear mapping, and the correlation weights between query, key, and value features are calculated. These weights quantitatively characterize the physical influence of neighboring nodes on the current node. The neighborhood features are weighted and aggregated according to the correlation weights to achieve selective fusion of neighborhood information. The enhanced multi-scale flow field features are output through residual connections, preserving both large-scale trends and incorporating local correlations. The physical perception attention module in step 4 is used to capture flow field correlations at different resolutions within the physical scale block of the encoder, while the physical perception attention mechanism in step 5 is used to progressively refine the multi-scale flow field features output by the encoder in the decoder. Both use the same attention calculation unit but are located in different network stages.

[0035] Step 5: Input the multi-scale flow field features output by the encoder into the decoder, and perform progressive refinement processing through layer normalization, physical perception attention mechanism and multilayer perceptron operation.

[0036] Specifically, after receiving the multi-scale flow field features output by the encoder, the decoder first performs layer normalization to stabilize the feature distribution, and then uses a physical perception attention mechanism to perform weighted fusion of the multi-scale features based on the physical correlation of surface nodes. Subsequently, the channel dimension mapping and nonlinear refinement are completed by a multilayer perceptron, and the decoded features are output layer by layer through residual connections.

[0037] Step 6: Through the output projection module, the decoded hidden representation is mapped to the full-field heat flux density distribution on the aircraft surface to obtain the aerodynamic thermal prediction results.

[0038] In this step, the output projection module includes layer normalization and linear transformation units; during the training process, the model applies physical consistency constraints based on Fourier's law of heat conduction to the full-field heat flux density distribution.

[0039] Specifically, the output projection module first performs layer normalization on the decoded hidden representation, and then maps the hidden dimension to the heat flux density output dimension through linear transformation to obtain the full-field heat flux density distribution corresponding to each surface space node; the physical consistency constraint constrains the output heat flux density distribution during the pre-training process.

[0040] In this invention, the model training process refers to the model pre-training process completed before executing step 1, which is used to constrain the full-field heat flux density distribution generated by the output projection module under the supervision of computational fluid dynamics sample data; after the model pre-training is completed, the obtained model is used to execute the aerodynamic thermal prediction process from step 1 to step 6.

[0041] More preferably, before performing step 1, a model training process based on computational fluid dynamics is also included, specifically: Parametric sampling simulations were performed on various aircraft geometries using a computational fluid dynamics solver to generate a sample dataset containing surface spatial nodes and corresponding wall heat flux density distributions as a benchmark reference. The physical parameters of height, angle of attack and Mach number in the sample dataset are Z-score standardized, and the target reference value of heat flux density is logarithmically transformed and Z-score scaled. Constructing a collaborative multi-scale neural operator network, the architecture of which includes N t There are 1 Transformer layer, each configured with 1 Transformer layer. N h Hidden units and N a We employ an attention head, using the Smooth L1 loss function as the optimization objective, and iteratively train the AdamW optimizer, combining a cosine annealing scheduling strategy to update the network weights until the model converges.

[0042] Based on the above steps, this invention reduces the need for re-subdividing the CFD mesh or manually extracting prior features for each prediction, and can learn the nonlinear mapping from flight parameters and geometric coordinates to the overall heat flux density. This invention can predict the aerodynamic heating of complex aircraft configurations such as blunt cones, bicones, biellipsoids, and lifting bodies quickly and accurately under varying operating conditions, with relative errors controlled within 1% in critical high-gradient stagnation point regions. This improves the computational efficiency and physical consistency of aerodynamic thermal prediction and can be used for aerodynamic design optimization and mission-level thermal analysis.

[0043] The following is a specific embodiment of the present invention.

[0044] Step 1: Generation and preprocessing of training data Step 101: Generate the basic dataset using a CFD solver. For representative high Mach number configurations (blunt cone, double cone, double ellipsoid, lifting body), a parametric sampling design is adopted, with parameters covering flight altitudes from 20 km to 60 km, Mach numbers from 5 to 20, and angles of attack from 0° to 20°.

[0045] Step 102: Solve the three-dimensional unsteady compressible Navier-Stokes equations to obtain the spatially resolved wall heat flux density distribution as a reference database for supervised learning and verification, generating a total of 1008 CFD samples.

[0046] Step 103, Dataset Splitting and Standardization. The dataset is divided into a 60% training set, a 20% validation set, and a 20% test set. Physical parameters such as altitude, angle of attack, and Mach number are then applied. Z -Score normalization; due to the highly uneven distribution of heat flux density (high at stagnation points, low on the leeward side), a logarithmic transformation is performed on the heat flux target, followed by... Z -score scaling is used to address the inherent imbalance in the training distribution.

[0047] Step 2, Construction of the Cooperative Multi-Scale Neural Operator Model Step 201, construct the encoder input layer. Receive multimodal input, including spatial coordinates r = ( x , y , z The coordinates and their corresponding flight state parameters are used. High-dimensional position embeddings are obtained by Fourier encoding the coordinates using frequency bands with six log-intervals. These embeddings are then concatenated with the flight state parameters and projected onto the hidden dimensions of the model using a multilayer perceptron (MLP) preprocessor.

[0048] Step 202: Inject global information. A learnable global bias vector is broadcast at each spatial location, and the local and global information are combined through an adaptive fusion module.

[0049] Step 203: Construct the encoder backbone. The enhanced features are fed into a series of physical scale blocks, each containing pre-normalization, a physical awareness attention module, and a feedforward MLP, with residual connections applied after each operation. The attention mechanism performs multi-scale clustering at resolutions of 16, 32, and 64.

[0050] Step 204: Construct the neural operator decoder. Features enter the decoder and are progressively refined through modules including layer normalization, physical awareness attention, and MLP operations.

[0051] Step 205: Construct the output mapping. Using an output projection module that includes layer normalization and linear transformation, the hidden representation is mapped to the specified output dimension, resulting in a continuous physical field of surface heat flux density after physical consistency constraints. q w The entire architecture consists of eight Transformer layers, each with 128 hidden units and eight attention heads.

[0052] Step 3, Model Training and Validation Testing Step 301: Train the model using the PyTorch framework on an NVIDIA RTX 2080 Ti GPU. The optimization objective is the Smooth L1 loss function.

[0053] Step 302: Update parameters using the AdamW optimizer, setting the initial learning rate to 10. -3 The weight decays to 10. -5 And it is supplemented by a cosine annealing learning rate scheduler within a specified period.

[0054] Step 303: After training, test the model. Select the sample with the largest relative error in the test set for evaluation (e.g., a blunt cone condition at an altitude of 20 km, Mach number 5, and angle of attack of 20°). Figure 2 As shown in the figure. The results show that the aerothermal field profile and spatial distribution predicted by the model are largely consistent with the CFD reference data. Furthermore, for geometric configurations of varying complexity, such as blunt cones, bicones, biellipsoids, and lifting bodies, the model can reconstruct the high heat flux region near the stagnation point and capture the physical process of heat flux gradually decaying downstream, demonstrating the model's ability to learn the physical laws of aerothermal distribution.

[0055] Figure 3 The relative error distribution cloud map shown indicates that the relative error of the model remains within approximately 1% in most areas of the aircraft surface. In regions with high gradient characteristics and complex flows, such as stagnation points and leading edges, the deviation is relatively small. However, a few regions with slightly larger relative errors (approximately 10%) are mainly concentrated in areas with lower heat flux values, such as the leeward side. This phenomenon is due to the relatively small denominator amplifying the relative error, which is related to the small denominator in the relative error calculation. This result demonstrates that the prediction results have good accuracy and spatial consistency.

[0056] The aerothermal field profile output by the model is largely consistent with the CFD reference data in terms of amplitude and spatial distribution, reproducing the high heat flux region near the stagnation point and its gradual decay downstream. In key flow field regions, the predicted principal component analysis latent features form clear separated clusters based on local thermodynamic states, such as... Figure 4 As shown, the information entropy of the stagnation region is low and the attention weight is high, indicating that its feature distribution is related to the high-speed aero-thermodynamic control mechanism.

[0057] Figure 5 The spatial distribution of attention feature norms of latent features within the model is presented. The results show that regions with higher feature norms are mainly concentrated in physically relevant locations with complex flow fields and strong gradients, such as stagnation points and sharp leading edges. This indicates that the model can allocate higher representation weights to regions highly correlated with aerodynamic heating processes, revealing a feature distribution related to the high-speed aerodynamic thermodynamic control mechanism.

[0058] This invention provides a computer device, including a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the steps of the hypersonic vehicle aerodynamic thermal prediction method described in this invention. The memory includes, but is not limited to, volatile memory and / or non-volatile memory. The non-volatile memory may be a read-only memory (Read-Only Memory). Only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DR RAM).

[0059] This invention also provides a computer storage medium storing a computer program thereon, which, when executed by a computer, enables the computer to perform the methods described in the above-described method embodiments. Alternatively, this invention also provides a computer program product containing instructions that, when executed by a computer, cause the computer to perform the methods described in the above-described method embodiments.

[0060] In some embodiments of the present invention, the computer program may be divided into one or more modules, which are stored in the memory and executed by a processor to perform the method provided in the embodiments of the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the encoding device.

[0061] When implemented using software, it can be implemented entirely or partially as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0062] The above description is an embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the aerodynamic heat of a hypersonic vehicle, characterized in that, Includes the following steps: Step 1: Obtain multimodal input data of the surface of the aircraft to be predicted. The multimodal input data includes three-dimensional spatial coordinates r = ( x , y , z And the flight status parameters associated with that spatial coordinates. f (r), wherein the flight state parameters include at least Mach number, flight altitude and angle of attack; Step 2: Preprocess and feature encode the multimodal input data. Specifically, the three-dimensional spatial coordinates are Fourier encoded through multiple logarithmically spaced frequency bands, thereby mapping low-dimensional coordinate features into high-dimensional position embedding information to enhance the model's ability to perceive high-frequency spatial details. Step 3: Integrate the high-dimensional position embedding information with the flight state parameters. f (r) After concatenation, the input nonlinear multilayer perceptron preprocessor is projected onto the hidden dimension and a learnable global bias vector is broadcast to inject global contextual information. Step 4: Use the adaptive fusion module to enhance and fuse local and global information, and input the enhanced multimodal features into an encoder containing multiple physical scale blocks. Perform multi-scale clustering of surface spatial nodes at different resolutions to capture complex flow field features. Step 5: Input the multi-scale flow field features output by the encoder into the decoder, and perform progressive refinement processing through layer normalization, physical perception attention mechanism and multilayer perceptron operation in sequence; Step 6: Through the output projection module, the decoded hidden representation is mapped to the full-field heat flux density distribution on the aircraft surface to obtain the aerodynamic thermal prediction results.

2. The aerodynamic thermal prediction method for hypersonic vehicles according to claim 1, characterized in that, In step 2, the Fourier encoding process of the three-dimensional spatial coordinates uses six logarithmically spaced frequency bands to extract high-dimensional position features.

3. The aerodynamic thermal prediction method for hypersonic vehicles according to claim 1, characterized in that, In step 3, the learnable global bias vector is broadcast to inject global contextual information, as implemented as follows: Define a global feature vector, copy it in the spatial dimension, and add it to the feature vector of each local spatial coordinate point, so that each local node can perceive the global context of the flow field.

4. The aerodynamic thermal prediction method for hypersonic vehicles according to claim 1, characterized in that, In step 4, the adaptive fusion module is a feature-weighted network containing learnable weights. The local information comes from the spatial features after Fourier encoding and the current flight parameters, and the global information comes from the broadcast global bias vector.

5. The aerodynamic thermal prediction method for hypersonic vehicles according to claim 1, characterized in that, In step 4, the physical scale block includes pre-normalization processing, a physical perception attention module, and a feedforward multilayer perceptron, and residual connections are applied after each operation; the resolution of the multi-scale clustering is preset to 16, 32, and 64. The execution mechanism of the physical perception attention module includes: performing layer normalization on the input features, calculating the correlation weights between query, key, and value features based on the surface node neighborhood obtained by multi-scale clustering, weighting and aggregating the neighborhood features according to the correlation weights, and outputting the enhanced multi-scale flow field features through residual connections; the physical perception attention module in step 4 is used to capture flow field correlations at different resolutions within the physical scale block of the encoder, and the physical perception attention mechanism in step 5 is used to progressively refine the multi-scale flow field features output by the encoder in the decoder. Both use the same attention calculation unit but are located in different network stages.

6. The aerodynamic thermal prediction method for hypersonic vehicles according to claim 1, characterized in that, In step 6, the output projection module includes layer normalization and linear transformation units; during model training, the physical consistency constraint of Fourier's heat conduction law is applied to the full-field heat flux density distribution; The model training process refers to the model pre-training process completed before executing step 1, which is used to constrain the full-field heat flux density distribution generated by the output projection module under the supervision of computational fluid dynamics sample data; after the model pre-training is completed, the obtained model is used to execute the aerodynamic thermal prediction process from step 1 to step 6.

7. The aerodynamic thermal prediction method for hypersonic vehicles according to claim 1, characterized in that, Before performing step 1, a model pre-training process based on computational fluid dynamics is also included, specifically: Parametric sampling simulations were performed on various aircraft geometries using a computational fluid dynamics solver to generate a sample dataset containing surface spatial nodes and corresponding wall heat flux density distributions as a benchmark reference. The physical parameters of height, angle of attack, and Mach number in the sample dataset were analyzed. Z -Score normalization processing, which performs logarithmic transformation and Z-score scaling on the target reference value of heat flux density; A collaborative multi-scale neural operator network was constructed, and the Smooth L1 loss function was used as the optimization objective. The AdamW optimizer is used for iterative training, and the network weights are updated using a cosine annealing scheduling strategy until the model converges. The resulting model is then used to execute steps 2 through 6.

8. The aerodynamic thermal prediction method for hypersonic vehicles according to claim 7, characterized in that, The architecture of the collaborative multi-scale neural operator network includes N t There are 1 Transformer layer, each configured with 1 Transformer layer. N h Hidden units and N a One point of attention.

9. A computer device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the hypersonic vehicle aerothermal prediction method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the hypersonic vehicle aerodynamic thermal prediction method according to any one of claims 1 to 8.

Images