An intelligent agent model modeling method based on flow field characteristics
By using an intelligent aerodynamic surrogate model based on flow field characteristics and leveraging the DeepOnet and BranchNet/TrunkNet architectures, the problems of high aerodynamic modeling costs and insufficient interpretability are solved, achieving efficient and low-cost aerodynamic prediction and optimization.
Patent Information
- Application Number
- CN202611252777.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-03
AI Technical Summary
Existing technologies have high computational costs in aerodynamic modeling, which are difficult to reduce effectively. Furthermore, the interpretability and robustness of the models are insufficient, especially when studying complex flow fields and multi-parameter variants.
An aerodynamic intelligent proxy model based on flow field characteristics is adopted. By combining the DeepOnet model with BranchNet and TrunkNet, the core aerodynamic features of pressure distribution data are extracted through orthogonal decomposition. A deep operator neural network model is established to directly predict the aerodynamic parameters of untested shapes.
It significantly reduces aerodynamic prediction errors, saves more than 50% of CFD computation costs, improves model transparency and interpretability, and enhances evaluation and optimization capabilities in complex aerodynamic environments.
Smart Images
Figure CN123031622A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of aircraft aerodynamic characteristic prediction technology, specifically relating to a modeling method for intelligent aerodynamic proxy models based on flow field characteristics. Background Technology
[0002] In the field of aerodynamic computation, aerodynamic modeling is a core component. Traditional methods directly model integral forces (normal force, axial force, and moment), relying on a large amount of high-precision CFD calculation data. However, acquiring this data is costly and computationally expensive, especially when dealing with complex flow fields and multi-parameter variants, where the demand for computational resources increases dramatically, limiting research efficiency. Complex phenomena such as flow separation, shock waves, and transonic speeds make direct aerodynamic modeling particularly difficult. Feature extraction techniques such as Proper Orthogonal Decomposition (POD) and autoencoders (AE) have been widely used in flow feature analysis. They can extract effective physical features from limited flow field data, and the difficulty is lower than direct aerodynamic modeling. There is a strong correlation between flow field data such as pressure distribution and aerodynamic forces. Existing techniques typically first obtain the distributed forces through CFD calculations and then integrate them to obtain the aerodynamic forces, but this method is computationally expensive. There is an urgent need in this field for an aerodynamic modeling method that can significantly improve prediction accuracy and reduce computational costs, especially under conditions of small sample training.
[0003] Traditional single-branch networks perform unconstrained mixing of two types of heterogeneous parameters at the input layer or shallow layers, resulting in stable geometric information and noisy state information becoming entangled in the feature space. This mechanism has significant drawbacks: First, noise easily propagates and amplifies in deeper layers of the network, leading to overfitting of the upstream flow field model and producing spurious correlation fitting; second, the lack of physical decoupling between geometric and flow representations, and the black-box splicing mapping, severely weakens the model's interpretability; third, when the upstream model's output distribution shifts or noise increases significantly, the generalization ability and robustness of the single-branch architecture drop sharply, making it difficult to support highly reliable engineering applications. Therefore, there is an urgent need for a new aerodynamic prediction architecture that can explicitly decouple multi-source heterogeneous inputs and balance physical consistency with noise robustness. Summary of the Invention
[0004] The purpose of this invention is to overcome one of the problems existing in the prior art and provide a modeling method for aerodynamic intelligent proxy model based on flow field characteristics. This method can significantly improve the prediction accuracy of integral forces such as normal force, axial force, and moment coefficient, greatly reduce the dependence on high-precision CFD calculation samples, save computational costs, and enhance the physical interpretability of the model. It is suitable for rapid evaluation and optimization in complex aerodynamic environments.
[0005] This invention provides a method for modeling an aerodynamic intelligent proxy model based on flow field characteristics, comprising the following steps: Obtain pressure distribution data for aerodynamic models with different shapes; Extract the core aerodynamic features from the pressure distribution data; Based on the obtained aerodynamic model shape and core aerodynamic features, a deep operator neural network model is established to predict the pressure distribution of untested shapes. Based on the geometry and pressure distribution of the untested shape, a machine learning surrogate model is used to predict the key aerodynamic parameters of the untested shape.
[0006] Furthermore, pressure distribution data for aerodynamic models with different shapes were obtained through computational fluid dynamics simulation analysis.
[0007] Furthermore, the extraction of the core aerodynamic features of the pressure distribution data includes: using orthogonal decomposition to reduce the dimensionality of the pressure distribution data and extract features, filtering redundant information, and retaining the core aerodynamic features.
[0008] Furthermore, the machine learning proxy model adopts the DeepOnet model. The original geometric parameters of the aircraft are input into BranchNet, which learns the basic aerodynamic response patterns determined by the geometry through multilayer perceptron mapping, outputting a set of geometric response basis coefficients. Flow field feature parameters from the upstream prediction model are input into TrunkNet, which uses multilayer perceptron mapping to characterize the modulation information of the aerodynamic response under specific flow conditions, outputting the corresponding flow activation patterns. The geometric response basis coefficients output by BranchNet and the flow activation patterns output by TrunkNet are coupled by an inner product operation to obtain the final aerodynamic prediction result.
[0009] Furthermore, the key aerodynamic parameters of the untested shape include: normal force coefficient, axial force coefficient, and moment coefficient.
[0010] The advantages of this invention compared to the prior art are as follows: Compared to traditional direct modeling methods using integral forces, this invention employs an aerodynamic prediction model that embeds low-dimensional physical features using TCA (or optimized feature extraction methods). This significantly reduces prediction errors, achieving a reduction of 40%-50% or more compared to direct modeling, with the effect becoming even more pronounced with increased sample size. This framework can save over 50% of the computational cost of CFD samples while maintaining the same accuracy. By performing aerodynamic visualization analysis in the low-dimensional feature space, the correlation between aerodynamic forces and flow field characteristics (such as shock waves and pressure distribution) can be clearly demonstrated, improving the model's transparency and reliability, and providing intuitive guidance for optimizing aerodynamic characteristics (such as drag reduction). Through feature importance and coupling analysis, the optimal feature combination can be selected to improve model performance.
[0011] The following comparison examines the feature enhancement modeling performance of the original Kriging-based model and the DeepOnet-based model. The comparison reveals that DeepOnet achieves effective utilization of flow field feature parameters, particularly in CA, CN, and Cm, where it reduces performance by approximately half compared to the existing Kriging surrogate model (e.g., Figure 4 (As shown).
[0012] When comparing the MAE (mean absolute error) of different models on the test set (e.g.) Figure 3 As shown in the diagram, we can observe that the DeepOnet model performs particularly well under the feature augmentation modeling approach. Specifically, for methods that directly model the initial features, the ResNet model generally has a higher MAE value than the Kriging model on various sample sets, indicating that its ability to capture the original features is not as good as Kriging's. However, when DeepOnet is incorporated into the feature augmentation modeling approach, its performance is significantly improved. Compared to other augmentation modeling methods such as Kriging and ResNet, DeepOnet not only optimizes the relational representation of data features through network layer learning, but also captures deeper patterns between different data, bringing higher fitting accuracy to the model and achieving lower prediction errors. Its performance on the CA, CN, and Cm datasets also shows that DeepOnet has better robustness and generalization than Kriging and ResNet, with its MAE values on these sample sets concentrated in the lowest or second-lowest range.
[0013] Comparison of Aerodynamic Modeling Errors Attached Figure Description
[0014] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of a method for modeling an aerodynamic intelligent proxy model based on flow field characteristics according to an embodiment of the present invention.
[0015] Figure 2 This is a diagram of an aerodynamic intelligent agent model architecture based on flow field characteristics according to an embodiment of the present invention.
[0016] Figure 3 This is a comparison chart of the six-component prediction accuracy of the model according to an embodiment of the present invention.
[0017] Figure 4 This is a comparison chart of the prediction accuracy of axial force, normal force, and moment of the model according to an embodiment of the present invention. Detailed Implementation
[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0019] Example 1 This invention provides a modeling method for aerodynamic intelligent proxy models based on flow field characteristics, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps: S01: Obtain pressure distribution data for aerodynamic models with different shapes; S02: Extract the core aerodynamic features of the pressure distribution data; S03: Based on the obtained aerodynamic model shape and core aerodynamic features, establish a deep operator neural network model to predict the pressure distribution of untested shapes; S04: Based on the geometry and pressure distribution of the untested shape, a machine learning surrogate model is used to predict the key aerodynamic parameters of the untested shape.
[0020] Furthermore, pressure distribution data for aerodynamic models with different shapes were obtained through computational fluid dynamics simulation analysis.
[0021] Furthermore, the extraction of the core aerodynamic features of the pressure distribution data includes: using orthogonal decomposition to reduce the dimensionality of the pressure distribution data and extract features, filtering redundant information, and retaining the core aerodynamic features.
[0022] Furthermore, the machine learning agent model adopts the DeepOnet model, such as... Figure 2 As shown, the original geometric parameters of the aircraft are input into BranchNet, and the basic aerodynamic response patterns determined by the geometry are learned through multilayer perceptron mapping, outputting a set of geometric response basis coefficients. The flow field feature parameters from the upstream prediction model are input into TrunkNet, and the modulation information of the aerodynamic response under specific flow conditions is characterized through multilayer perceptron mapping, outputting the corresponding flow activation patterns. The geometric response basis coefficients output by BranchNet and the flow activation patterns output by TrunkNet are coupled by inner product to obtain the final aerodynamic prediction results.
[0023] Furthermore, the key aerodynamic parameters of the untested shape include: normal force coefficient, axial force coefficient, and moment coefficient.
[0024] Example 2 This invention provides a method for modeling an aerodynamic intelligent proxy model based on flow field characteristics, comprising the following steps: 1) Acquisition of characteristic data: First, computational fluid dynamics (CFD) simulations are used to calculate and analyze aerodynamic models with different geometries to obtain numerical data on pressure distribution. This is the data foundation for subsequent analysis.
[0025] 2) Dynamic feature extraction: Advanced machine learning techniques, such as orthogonal decomposition (POD, TCA, etc.) experimentally used in this scheme, are used to effectively reduce the dimensionality of pressure data and extract features, filter redundant information, and retain core aerodynamic features.
[0026] 3) Intelligent Mapping and Prediction Construction: A deep neural network model (DeepOnet) is used to model the pressure distribution data with high precision, predicting the pressure distribution of new, untested shapes a priori. The next step inputs the predicted pressure data into a machine learning surrogate model (e.g., the Kriging model). The input parameters include basic geometric input variables and key physical features extracted from the pressure prediction results. Ultimately, the model predicts three key aerodynamic parameters: the normal force coefficient C_N, the axial force coefficient C_A, and the moment coefficient C_m. Theoretically, the prediction results do not require complex integration steps and directly achieve the required aerodynamic accuracy.
[0027] 4) Intelligent Optimization and Validation: Extensive experiments were conducted by controlling the number of training samples (e.g., 200, 300, 400). The results compared traditional methods using only geometric parameters with methods employing different feature extraction techniques such as POD, autoencoders, and TCA. In a lower-dimensional physical feature space, the differences in the effectiveness of different feature extraction methods in improving aerodynamic prediction accuracy were analyzed, and the number of CFD computation samples saved (over 50%) was quantitatively evaluated. Simultaneously, based on an improved visualization technique, the patterns of aerodynamic changes with features were intuitively displayed, which is beneficial for the interpretability of the algorithm and the formulation of optimization strategies.
[0028] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A modeling method for aerodynamic intelligent surrogate models based on flow field characteristics, characterized in that, Includes the following steps: Obtain pressure distribution data for aerodynamic models with different shapes; Extract the core aerodynamic features from the pressure distribution data; Based on the obtained aerodynamic model shape and core aerodynamic features, a deep operator neural network model is established to predict the pressure distribution of untested shapes. Based on the geometry and pressure distribution of the untested shape, a machine learning surrogate model is used to predict the key aerodynamic parameters of the untested shape.
2. The aerodynamic intelligent proxy modeling method based on flow field characteristics according to claim 1, characterized in that, Pressure distribution data of aerodynamic models with different shapes were obtained by computational fluid dynamics simulation analysis.
3. The aerodynamic intelligent proxy modeling method based on flow field characteristics according to claim 1 or 2, characterized in that, The extraction of the core aerodynamic features from the pressure distribution data includes: using orthogonal decomposition to reduce the dimensionality of the pressure distribution data and extract features, filtering redundant information, and retaining the core aerodynamic features.
4. The aerodynamic intelligent proxy modeling method based on flow field characteristics according to claim 1, characterized in that, The machine learning agent model uses the DeepOnet model.
5. The aerodynamic intelligent proxy modeling method based on flow field characteristics according to claim 1, characterized in that, The key aerodynamic parameters of the untested shape include: normal force coefficient, axial force coefficient, and moment coefficient.