Cross-dimensional 2d-3d car aerodynamic prediction method based on artificial intelligence algorithm
By employing a cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms, and utilizing parametric deformation, clustering, and multi-branch neural network models, the problem of high resource consumption in traditional 3D CFD simulation is solved. This method achieves efficient and low-cost aerodynamic performance prediction, adapting to different vehicle models and parameter variations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2026-05-07
- Publication Date
- 2026-07-07
AI Technical Summary
In existing technologies, traditional 3D CFD simulation consumes huge computational resources and has a long simulation cycle, while 2D CFD simulation lacks intelligent algorithm support, making it difficult to balance the efficiency and accuracy of automotive aerodynamic performance prediction, and failing to meet the needs of rapid iteration and low cost.
A cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms is adopted. Through parameterized deformation, K-Means clustering and GMM hybrid clustering, multi-branch neural network model, and cross-dimensional gain coefficient, the accurate mapping of two-dimensional features to three-dimensional aerodynamic performance is achieved.
It reduces the investment in high-computing equipment, lowers simulation time and cost, improves prediction accuracy and stability, adapts to different vehicle models and aerodynamic requirements, accommodates parameter changes, requires no additional simulation work, and enhances the engineering practicality of the prediction.
Smart Images
Figure CN122154566B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of automotive aerodynamic performance prediction technology, and particularly relates to a cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms. Background Technology
[0002] As the automotive industry moves towards higher efficiency and intelligence, rapid and accurate prediction of automotive aerodynamic performance has become crucial for shortening R&D cycles and reducing costs. Traditional automotive aerodynamic performance prediction relies on three-dimensional computational fluid dynamics (CFD) simulations, which, while ensuring accuracy, suffer from drawbacks such as huge computational resource consumption, long simulation cycles, and complex operations. It requires high-performance computing equipment and cannot meet the demands of rapid iteration and efficient verification in automotive R&D.
[0003] In existing technologies, some solutions attempt to replace 3D CFD simulation with 2D CFD simulation to save computing power, but lack targeted intelligent algorithm support, resulting in low prediction accuracy and the absence of a systematic cross-dimensional mapping method. Other solutions rely excessively on complex structural designs without focusing on algorithm optimization, making it difficult to balance prediction efficiency and accuracy. Currently, the industry lacks a prediction method that combines artificial intelligence algorithms to achieve accurate mapping of 2D to 3D aerodynamic performance while balancing efficiency and accuracy. This fails to meet the low-cost, high-precision, and high-speed aerodynamic prediction requirements in automotive R&D, thus necessitating an efficient and feasible technical solution to address these pain points. Summary of the Invention
[0004] The purpose of this invention is to provide a cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms, aiming to solve the problems mentioned in the background art.
[0005] The present invention is implemented as follows: a cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms, comprising the following steps:
[0006] Step 1: Data preparation stage, parametric deformation of the car model, the parameters are selected based on the car model. Geometric feature sensitive parameters from the normal perspective are used to form a parametric deformation table and a set of car models. Two-dimensional CFD simulation is performed on the longitudinal section of the car model set, and three-dimensional CFD simulation is performed on the undeformed car model.
[0007] Step 2: In the dataset processing stage, based on the aerodynamic coefficients obtained from the parameterized deformation table and two-dimensional CFD simulation, a hybrid clustering strategy of K-Means clustering and Gaussian mixture model (GMM) is used to perform adaptive pattern recognition and classification on the high-dimensional sample space, dividing the samples into several flow field pattern clusters with different aerodynamic trends; for each type of flow field pattern, a nonlinear regression prediction model based on artificial intelligence algorithm is constructed and trained to form a multi-branch neural network model set;
[0008] Step 3: New data prediction stage. Obtain the deformation parameters of the new 3D vehicle body model to be predicted and input them into the multi-branch neural network model set. The multi-branch neural network model set first realizes the localization of the flow field mode through cluster recognition, and performs preliminary aerodynamic prediction through the corresponding nonlinear regression prediction model. Finally, the cross-dimensional gain coefficient is introduced to perform scale correction and cross-dimensional compensation on the preliminary aerodynamic prediction results, and outputs the predicted value of the drag coefficient of the new 3D vehicle body model.
[0009] In a further technical solution, the key geometric parameters of the parametric deformation in step 1 include: approach angle, departure angle, front window tilt angle, rear window tilt angle, front fascia height, and trunk height.
[0010] In a further technical solution, in step 1, the total number of parameters for parametric deformation is set to... The number of deformed samples must reach at least There are 100 deformed samples, and all deformed samples are randomly generated using the Latin hypercube sampling method.
[0011] In a further technical solution, in step 1, the parametric deformation also includes the aerodynamic accessory geometric parameters of the automotive aerodynamic accessories, which include: tail wing angle of attack, canard angle, tail wing chord length, and canard mounting height.
[0012] In a further technical solution, in step 2, K-Means clustering uses the mapping relationship between parameterized deformation tables and aerodynamic coefficients as the core classification basis to perform aerodynamic pattern clustering on the samples. Before clustering, K-Means algorithm and GMM are used for preliminary clustering respectively. The clustering effect of the two algorithms is comprehensively evaluated by two indicators, MAE (mean absolute error) and MSE (mean squared error), and the algorithm with the smaller error is selected as the classification algorithm for the dataset.
[0013] Further technical solutions, regarding the number of clusters The optimal number of categories is found when increasing the number of categories results in a decrease of ≤5% in MAE and MSE. and the number of categories It satisfies 1≤m≤n.
[0014] A further technical solution is that, in step 3, when At that time, any class Cross-dimensional gain coefficient It conforms to the following formula:
[0015]
[0016] in, The drag coefficient of an undeformed car in a three-dimensional simulation. The drag coefficient of an undeformed car in a two-dimensional simulation.
[0017] When 2≤m≤5, any class Cross-dimensional gain coefficient It conforms to the following formula:
[0018]
[0019] in, In order to be in The drag coefficient of a randomly selected car model in a three-dimensional simulation. This represents the drag coefficient in the two-dimensional simulation of the model.
[0020] When m > 5, any class Cross-dimensional gain coefficient It conforms to the following formula:
[0021]
[0022] in, and In order to be in The drag coefficients of two randomly selected non-repeating car models in three-dimensional simulation. and Here are the two-dimensional simulation drag coefficients for these two models.
[0023] The cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms provided in this invention has the following beneficial effects:
[0024] (1) This method does not rely on excessive three-dimensional CFD simulation. It replaces full three-dimensional simulation with two-dimensional feature mapping, which reduces the investment in high-computing equipment and avoids the drawbacks of traditional three-dimensional simulation, which is time-consuming and resource-intensive, thus balancing efficiency and cost control.
[0025] (2) The method of combining cluster classification and multi-branch regression model is adopted, which effectively reduces the prediction bias compared with a single model. At the same time, the accuracy and stability of the prediction results are ensured by relying on MAE and MSE dual verification.
[0026] (3) It can be flexibly compatible with the adjustment of parameters such as the angle of attack of the tail wing and aerodynamic accessories. It can cover the deformation scenario of the vehicle body and adapt to the parameter changes of aerodynamic accessories without adding extra simulation workload.
[0027] (4) The multi-branch neural network model set can be adapted to different vehicle models and different aerodynamic requirements. It can achieve accurate prediction in different scenarios without additional adjustment of the model structure, thus avoiding repeated training.
[0028] (5) Combining the differences in aerodynamic characteristics of different sample categories, a unique cross-dimensional gain coefficient is configured for each flow field mode to avoid prediction bias caused by a single gain coefficient, so that the aerodynamic prediction accuracy of different sample categories is guaranteed and the problem of insufficient adaptability of traditional single gain coefficient is solved.
[0029] (6) Through the correction of the three-dimensional CFD results in the later stage, the cross-dimensional gain coefficient and model weight can be dynamically adjusted to further reduce the prediction error of different types of samples, adapt to the actual needs of automotive aerodynamic performance optimization, and improve the engineering practicality and operability of the overall method. Attached Figure Description
[0030] Figure 1 A flowchart illustrating the data preparation stage of the cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms provided in this embodiment of the invention;
[0031] Figure 2 A flowchart illustrating the data processing stage of the cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms provided in this embodiment of the invention;
[0032] Figure 3 A flowchart illustrating the new data prediction stage of the cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms provided in this embodiment of the invention;
[0033] Figure 4 A schematic diagram showing the trend of similarity force coefficient changes in two-dimensional and three-dimensional simulations of different clusters after clustering;
[0034] Figure 5 Comparison of two-dimensional and three-dimensional flow fields from the y-normal perspective of the same vehicle model. Detailed Implementation
[0035] 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.
[0036] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.
[0037] like Figures 1-3 As shown, an embodiment of the present invention provides a cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms, comprising the following steps:
[0038] Step 1: Data preparation stage, parametric deformation of the car model, the parameters are selected based on the car model. Geometric feature sensitive parameters from the normal perspective are used to form a parametric deformation table and a set of car models. Two-dimensional CFD simulation is performed on the longitudinal section of the car model set, and three-dimensional CFD simulation is performed on the undeformed car model.
[0039] Step 2: In the dataset processing stage, based on the aerodynamic coefficients obtained from the parameterized deformation table and two-dimensional CFD simulation, a hybrid clustering strategy of K-Means clustering and GMM is used to perform adaptive pattern recognition and classification on the high-dimensional sample space, dividing the samples into several flow field pattern clusters with different aerodynamic trends; for each type of flow field pattern, a nonlinear regression prediction model based on artificial intelligence algorithm is constructed and trained to form a multi-branch neural network model set to achieve high-precision fitting under different aerodynamic modes.
[0040] Step 3: New data prediction stage. Obtain the deformation parameters of the new 3D vehicle body model to be predicted and input them into the multi-branch neural network model set. The multi-branch neural network model set first realizes the rapid localization of the flow field mode through cluster recognition, and performs preliminary aerodynamic prediction through the corresponding nonlinear regression prediction model. Finally, the cross-dimensional gain coefficient is introduced to perform scale correction and cross-dimensional compensation on the preliminary aerodynamic prediction results, realize the mapping from two-dimensional features to three-dimensional aerodynamic performance, and output the predicted value of the drag coefficient of the new 3D vehicle body model.
[0041] In this embodiment of the invention, it should be further explained that the car model The normal viewpoint refers to the projection viewpoint with the Y-axis (horizontal, perpendicular to the vehicle's direction of travel) of the vehicle coordinate system as the normal, that is, the viewpoint observed from the side view direction of the vehicle.
[0042] In a preferred embodiment of the present invention, in step 1, the core parameters of the parametric deformation are selected based on the car model. Geometric feature sensitive parameters from the normal perspective are selected as key geometric parameters that exhibit significant changes in shape and have a high weight in influencing vehicle aerodynamic performance from this perspective. Deformation locations include, but are not limited to: approach angle, departure angle, front window sag angle, rear window sag angle, front fascia height, and trunk height. The total number of parameters for parameterized deformation is set to... The number of deformed samples must reach at least Each deformed sample is randomly generated using the Latin hypercube sampling method, and the specific values of each geometric parameter, deformation amplitude, and corresponding sample number of each deformed sample are recorded, corresponding one-to-one with the car model set.
[0043] As a preferred embodiment of the present invention, in step 1, for parametric deformation, in addition to the key geometric parameters of the automobile, the aerodynamic accessory geometric parameters can also be expanded to include the aerodynamic accessories of the automobile. The aerodynamic accessory geometric parameters include, but are not limited to: tail wing angle of attack, canard angle, and the generated parametric deformation table requires angles. Auxiliary aerodynamic parameters such as tail wing chord length and canard installation height can also be added according to actual aerodynamic prediction requirements.
[0044] In a preferred embodiment of the present invention, in step 2, K-Means clustering uses the mapping relationship between the parameterized deformation table and aerodynamic coefficients as the core classification basis to perform aerodynamic pattern clustering on the samples. Before clustering, K-Means algorithm and GMM are used for preliminary clustering respectively. The clustering effect of the two algorithms is comprehensively evaluated by two indicators: MAE (mean absolute error) and MSE (mean squared error). MAE is used to measure the fitting deviation of the mapping relationship within each subclass after clustering, and MSE is used to evaluate the overall accuracy of the fitting result. The two with smaller errors are selected as the classification algorithm for the dataset.
[0045] Number of clusters The determination of the optimal number of categories depends entirely on the combined performance of MAE and MSE. The optimal number of categories is determined when increasing the number of categories results in a decrease of ≤5% in both MAE and MSE. and the number of categories It satisfies 1≤m≤n.
[0046] In a preferred embodiment of the present invention, in step 3, when At that time, any class Cross-dimensional gain coefficient It conforms to the following formula:
[0047]
[0048] in, The drag coefficient of an undeformed car in a three-dimensional simulation. Let be the drag coefficient for a two-dimensional simulated undeformed vehicle. When 2 ≤ m ≤ 5, for any class... Cross-dimensional gain coefficient It conforms to the following formula:
[0049]
[0050] in, In order to be in The drag coefficient of a randomly selected car model in a three-dimensional simulation. Let m be the drag coefficient for the two-dimensional simulation of this model. When m > 5, for any class... Cross-dimensional gain coefficient It conforms to the following formula:
[0051]
[0052] in, and In order to be in The drag coefficients of two randomly selected non-repeating car models in three-dimensional simulation. and Here are the two-dimensional simulation drag coefficients for these two models.
[0053] Specifically, approach angle, front window tilt angle, and departure angle are selected as the core parameters for parametric deformation (i.e., Fifteen sets of deformed samples were generated through Latin hypercube sampling. Two-dimensional CFD simulations were performed on all samples, and three-dimensional CFD simulations were performed on the baseline model to construct a complete dataset. Figure 4 As shown, through hybrid clustering of K-Means and GMM, the samples were divided into two flow field modes. The second flow field mode had a higher overall two-dimensional simulation Cd value. The cross-dimensional gain coefficients corresponding to the two modes were 1.71 and 1.53, respectively. Although there was a fixed scale deviation in the simulation results of the two-dimensional and three-dimensional drag coefficients, the trend of change was highly consistent. After multi-branch model prediction and cross-dimensional gain correction, the average relative error of the predicted value of the three-dimensional drag coefficient was only 1.858%, which verified the high accuracy and reliability of the method.
[0054] Figure 5 Comparing the two-dimensional and three-dimensional flow fields of the same vehicle model, it can be seen that the flow field structures of the three-dimensional whole vehicle and the two-dimensional longitudinal section are highly similar: a strong stagnant high-pressure zone (red core area) is formed at the front of the vehicle, aerodynamic separation occurs at the junction of the hood and the front face, the roof shows high-speed flow characteristics, and the location of the tail vortex shedding area matches the color gradient pattern. At the same time, there are also subtle differences in the flow field: due to the planar assumption of the two-dimensional simulation, the low-speed diffusion range of the tail wake region in the two-dimensional model is relatively more concentrated and the shape is more compact. This difference is essentially the dimensional mapping loss of the three-dimensional flow effect in three-dimensional space on the two-dimensional plane, which is also the physical root of the introduction of a cross-dimensional gain coefficient for scale correction in this method. This coefficient can effectively correct the systematic deviation between two-dimensional and three-dimensional models, achieving high-precision aerodynamic performance prediction.
[0055] The above description is only a preferred 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 cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms, characterized in that, Includes the following steps: Step 1: Data preparation stage, parametric deformation of the car model, the parameters are selected based on the car model. Geometric feature sensitive parameters from the normal perspective are used to form a parametric deformation table and a set of car models. Two-dimensional CFD simulation is performed on the longitudinal section of the car model set, and three-dimensional CFD simulation is performed on the undeformed car model. Step 2: In the dataset processing stage, based on the aerodynamic coefficients obtained from the parameterized deformation table and two-dimensional CFD simulation, the high-dimensional sample space is adaptively identified and classified using a hybrid clustering strategy of K-Means clustering and GMM, dividing the samples into several flow field pattern clusters with different aerodynamic trends; for each type of flow field pattern, a nonlinear regression prediction model based on artificial intelligence algorithm is constructed and trained to form a multi-branch neural network model set. Step 3: New data prediction stage, obtain the deformation parameters of the new 3D body model to be predicted and input them into the multi-branch neural network model set; the multi-branch neural network model set first realizes the localization of the flow field mode through cluster recognition, and performs preliminary aerodynamic prediction through the corresponding nonlinear regression prediction model; Finally, a cross-dimensional gain coefficient is introduced to perform scale correction and cross-dimensional compensation on the preliminary aerodynamic prediction results, and output the predicted drag coefficient value of the new three-dimensional vehicle body model.
2. The cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms according to claim 1, characterized in that, In step 1, the key geometric parameters for parametric deformation include: approach angle, departure angle, front window tilt angle, rear window tilt angle, front fascia height, and trunk height.
3. The cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms according to claim 1, characterized in that, In step 1, the total number of parameters for parametric deformation is set to The number of deformed samples must reach at least There are 100 deformed samples, and all deformed samples are randomly generated using the Latin hypercube sampling method.
4. The cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms according to claim 2, characterized in that, In step 1, the parametric deformation also includes the aerodynamic accessory geometry parameters of the automotive aerodynamic accessories, which include: tail wing angle of attack, canard angle, tail wing chord length, and canard mounting height.
5. The cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms according to claim 3, characterized in that, In step 2, K-Means clustering uses the mapping relationship between parameterized deformation tables and aerodynamic coefficients as the core classification basis to perform aerodynamic pattern clustering on the samples. Before clustering, K-Means algorithm and GMM are used for preliminary clustering respectively. The clustering effect of the two algorithms is comprehensively evaluated by MAE and MSE indicators, and the two with smaller errors are selected as the classification algorithm for the dataset.
6. The cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms according to claim 5, characterized in that, For the number of clusters The optimal number of categories is found when increasing the number of categories results in a decrease of ≤5% in MAE and MSE. and the number of categories It satisfies 1≤m≤n.
7. The cross-dimensional 2D-3D automotive aerodynamic prediction method based on artificial intelligence algorithms according to claim 6, characterized in that, In step 3, when At that time, any class Cross-dimensional gain coefficient It conforms to the following formula: in, The drag coefficient of an undeformed car in a three-dimensional simulation. The drag coefficient of an undeformed car in a two-dimensional simulation. When 2≤m≤5, any class Cross-dimensional gain coefficient It conforms to the following formula: in, In order to be in The drag coefficient of a randomly selected car model in a three-dimensional simulation. This represents the drag coefficient in the two-dimensional simulation of the model. When m > 5, any class Cross-dimensional gain coefficient It conforms to the following formula: in, and In order to be in The drag coefficients of two randomly selected non-repeating car models in three-dimensional simulation. and Here are the two-dimensional simulation drag coefficients for these two models.