Automobile aerodynamic optimization method based on vortex system characteristics and ternary coupling ai

Abstract

This invention relates to the field of automotive aerodynamics and intelligent design technology, providing a method for automotive aerodynamic optimization based on vortex characteristics and ternary coupled AI. The method includes: dividing the wake into multiple rectangular wake zones based on vortex characteristics according to vehicle type; parametrically deforming the rear region of the vehicle; extracting the regional average value of vortex characteristic quantities within each wake zone through flow field numerical simulation; constructing a ternary coupled dataset of geometric features, wake zone features, and drag coefficients; constructing and training a ternary coupled AI model; using the trained ternary coupled AI model to predict the wake zone features and drag coefficients of new samples; combining contribution analysis and global optimization to obtain optimization priorities; and providing feedback on the features most deserving optimization. This invention achieves accurate prediction and targeted optimization of the aerodynamic characteristics of the vehicle rear, improving aerodynamic optimization efficiency.

Description

Technical Field

[0001] This invention belongs to the field of automotive aerodynamics and intelligent design technology, and particularly relates to an automotive aerodynamic optimization method based on vortex characteristics and three-dimensional coupled AI. Background Technology

[0002] With increasingly stringent global requirements for energy conservation and emission reduction in automobiles, aerodynamic drag has become a key factor affecting fuel consumption of traditional gasoline vehicles and driving range of electric vehicles. Effectively controlling the vehicle's wake structure is one of the core technological directions for reducing aerodynamic drag. However, the wake flow fields of stepped-back (sedan) and straight-back (SUV / MPV) vehicles exhibit complex and significantly different three-dimensional vortex structures.

[0003] The existing research and practice have the following technical shortcomings:

[0004] (1) Lack of fine-grained partitioning: Traditional methods often use overall flow field analysis or geometric equal division, which fail to make a clear physical distinction between different regions (such as A-pillar vortex, C-pillar vortex, bottom vortex, etc.) based on the actual vortex core distribution, making it difficult to quantify the contribution of each vortex system to the total drag.

[0005] (2) Optimization relies on experience and trial and error: Tail shape optimization relies heavily on engineers' experience and "trial and error" simulation iteration, which has a long cycle, high cost, and makes it difficult to locate the key vortex system area that causes high drag, thus failing to achieve targeted drag reduction.

[0006] (3) AI models are black boxes and have poor generalization ability: Pure data-driven aerodynamic prediction models lack physical mechanism constraints, are prone to learning false correlations, and have poor interpretability. At the same time, the range of changes in the drag coefficient of automobiles is narrow, and conventional AI models are prone to falling into the "fixed value fitting" defect, and have insufficient ability to predict small nonlinear changes.

[0007] (4) Poor vehicle adaptability: Existing methods do not fully consider the essential differences in the topology of the vortex system at the rear of the stepped-back and straight-back models, making it difficult for a single method to be universally applicable.

[0008] Therefore, there is an urgent need for a vehicle aerodynamic optimization method that can integrate physical mechanisms, has high-precision prediction capabilities, and can provide clear optimization targets. Summary of the Invention

[0009] The purpose of this invention is to provide a method for optimizing automotive aerodynamics based on vortex characteristics and ternary coupled AI, aiming to solve the problems mentioned in the background art.

[0010] The embodiments of the present invention are implemented as follows: a vehicle aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI, comprising the following steps:

[0011] Step 1: Based on the vehicle type and vortex characteristics, divide the wake into multiple rectangular wake zones; the vehicle types include stepped-back vehicles and straight-back vehicles;

[0012] Step 2: Perform parametric geometric deformation on the rear region of the car to obtain multiple sets of car models with different shapes; perform flow field numerical simulation on each car model, extract the regional average value of vortex characteristic quantities in each wake partition, and construct a three-dimensional coupled dataset that couples geometric features, wake partition features, and drag coefficient.

[0013] Step 3: Construct a ternary coupled AI model, which includes a mapping network between geometric features and wake partition features, a mapping network between wake partition features and drag coefficient, feature contribution analysis, and global model optimization. The mapping network between wake partition features and drag coefficient is trained using a dual verification mechanism. The ternary coupled dataset from Step 2 is input into the ternary coupled AI model to complete training, establishing the mapping relationship between geometric features, wake partition features, and drag coefficient.

[0014] Step 4: Input the geometric parameters of the new sample into the trained ternary coupled AI model to predict the characteristics of each wake zone and the drag coefficient value of the new sample; obtain the optimization priority through feature contribution analysis and global model optimization, and feed back the geometric features and wake zone features that should be optimized most for the new sample based on the optimization priority.

[0015] A further technical solution is that, in step 1, for a stepped-back car, the wake is divided into the A-pillar vortex region, C-pillar vortex region, tail-end rising vortex region, bottom longitudinal vortex region, hairpin vortex region, and far-flow region based on the flow characteristics; for a straight-back car, the wake is divided into the upper recirculation bubble region, C-pillar vortex region, tail-end rising vortex region, bottom longitudinal vortex region, and far-flow region based on the flow characteristics.

[0016] In a further technical solution, the C-pillar vortex area of ​​a stepped-back car is divided into two parts, C1 and C2, located above and behind the trunk lid, respectively; while the C-pillar vortex area of ​​a straight-back car is only a part, located on the side and rear of the trunk lid.

[0017] In a further technical solution, in step 1, the point where the longitudinal section of the car, the ground, and the rear section of the car intersect is taken as the origin of the partition; the partition coordinate axes are respectively defined by the positive directions of x*, y*, and z* from the front to the rear of the car, the normal direction of the longitudinal section of the car from the passenger side to the driver side, and the bottom to the top of the car; the units of the partition coordinate axes are x*=x / L, y*=2y / W, and z*=z / H, where L, W, and H are the length, width, and height of the car, respectively, and x, y, and z are the actual dimensions in the direction of the coordinate axis.

[0018] A further technical solution is that, in step 1, the rectangular wake zones of the stepped-back car are categorized in the z* direction as follows: hairpin vortex zone (1.00, 1.12), A-pillar vortex zone (0.82, 1.00), C1 sub-zone (0.64, 0.82), C2 sub-zone (0.39, 0.82), rear upturn vortex zone (0.15, 0.39), bottom longitudinal vortex zone (0, 0.15), and far-flow zone (0, 0.82).

[0019] The rectangular wake regions of a hatchback car are defined in the z* direction as follows: upper recirculation bubble region (0.39, 1.00), C-pillar vortex region (0.39, 1.00), rear updraft vortex region (0.15, 0.39), bottom longitudinal vortex region (0, 0.15), and far-flow region (0, 1.00).

[0020] In a further technical solution, in step 1, the division range of each zone of the stepped-back car in the y* direction is as follows: the range of the A-pillar vortex zone, the hairpin vortex zone, and the C1 sub-zone is (0, 1.00); the range of the C2 sub-zone, the rear upward vortex zone, the bottom longitudinal vortex zone, and the far-flow zone is (0, 1.10).

[0021] The range of each zone in the y* direction for a straight-back car is as follows: the range of the rear upward vortex zone, the bottom longitudinal vortex zone, and the far-flow zone is (0, 1.10); the range of the upper recirculation bubble zone is (0, 1.00); and the range of the C-pillar vortex zone is (1.00, 1.10).

[0022] A further technical solution is that, in step 1, the rectangular wake zones of the stepped-back car are categorized in the x* direction as follows: hairpin vortex zone (-0.40, -0.18), A-pillar vortex zone (-0.40, 0), C1 sub-zone (-0.26, 0), C2 sub-zone (0, 0.25), rear upturn vortex zone (0, 0.25), bottom longitudinal vortex zone (0, 0.25), and far-flow zone (0.25, 0.50).

[0023] The x* direction division range of each zone of the straight-back car is as follows: upper recirculation bubble zone (0, 0.25), C-pillar vortex zone (0, 0.25), rear upward vortex zone (0, 0.25), bottom longitudinal vortex zone (0, 0.25), and far-flow zone (0.25, 0.50).

[0024] In a further technical solution, in step 2, the parametric geometric deformation includes adjusting at least one of the following: rear upturn angle, trunk lid height, rear window tilt angle, and bottom diffuser angle.

[0025] The eddy current characteristic quantities include at least one of the following: vorticity, Q criterion, and λ2 criterion.

[0026] In a further technical solution, in step 3, the two network components, namely the mapping network of geometric features and wake partition features and the mapping network of wake partition features and drag coefficient, are implemented using a general machine learning or deep learning architecture, including but not limited to BP neural networks, time series networks and other intelligent fitting models.

[0027] In a further technical solution, the dual verification mechanism in step 3 is as follows:

[0028] First verification: If the difference between the predicted values ​​of the drag coefficient of any two test samples is less than 0.001, the model is determined to be stuck in fixed value fitting and will be automatically retrained.

[0029] Second verification: If the predicted value of the resistance coefficient of any test sample exceeds 80% to 120% of the average value of the training set, the network training process will be automatically restarted until the prediction accuracy requirement is met.

[0030] A further technical solution, in step 3, specifically involves: performing weight sensitivity analysis based on the trained mapping network of geometric features and wake partition features, and the mapping network of wake partition features and drag coefficient, calculating the contribution weight of each geometric feature and wake partition feature to the drag coefficient. The contribution weight is obtained by extracting the weights from the input layer to the hidden layer of each neural network in the ternary coupling model, calculating the average absolute value of the weights, and normalizing them. The magnitude of the absolute value of the weights is positively correlated with the strength of the influence of the corresponding variable on the drag coefficient.

[0031] In a further technical solution, in step 3, the optimization algorithm used for global model optimization includes at least one of random global search, genetic algorithm, multi-island genetic algorithm, particle swarm optimization, simulated annealing algorithm, and differential evolution algorithm, in order to obtain the theoretically optimal global geometric eigenvalue and the theoretically optimal wake partition eigenvalue with the minimum drag coefficient.

[0032] In a further technical solution, in step 4, the optimization priority calculation method is: optimization priority = (standardized value of theoretical global optimal feature value - standardized value of current feature) × contribution weight;

[0033] Among them, the theoretical global optimal eigenvalues ​​include the theoretical global optimal geometric features and the theoretical global optimal wake partition features. One or more features with the largest absolute value of optimization priority are determined as the features that should be optimized first.

[0034] The standardized value of the current feature is calculated as follows:

[0035]

[0036]

[0037]

[0038] in, , and These are the standardized values ​​for geometric characteristics, wake zoning characteristics, and drag coefficient, respectively. , and These are the original values ​​for geometric characteristics, wake zoning characteristics, and drag coefficient, respectively. , and These are the training set mean values ​​for geometric features, wake zoning features, and drag coefficient, respectively. , and is the standard deviation of the training set for the corresponding feature.

[0039] The automotive aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI provided in this invention has the following beneficial effects:

[0040] (1) It can clearly define the contribution of each geometric parameter and vortex system to drag. The code can provide feedback on the geometric position and vortex region that should be optimized first, guiding engineers to perform precise drag reduction and avoiding blind iteration and waste of computing resources.

[0041] (2) This application uses artificial vortex physical partitioning based on the real characteristics and structure of the wake, so that the AI ​​model fits the real flow, avoids the black box problem of pure data-driven, and has high physical interpretability.

[0042] (3) By using non-uniform physical partitioning, the modeling variables are reduced; at the same time, the dual verification mechanism effectively overcomes the problems caused by the narrow range and weak changes of automotive aerodynamic data, ensuring high-precision prediction.

[0043] (4) Different zoning schemes were designed to address the differences between the rear vortex systems of stepped and straight-back vehicles, covering mainstream passenger car models and with a wide range of applications.

[0044] (5) After the model training is completed, the aerodynamic performance prediction and optimization suggestions of the new model can be output in seconds, which greatly shortens the iteration cycle in the early development stage.

[0045] (6) The characteristics of each vortex zone are very different, and the vorticity of different zones can differ by a factor of 100. The optimization priority is based on standardized characteristic values, making the priority more convincing. Attached Figure Description

[0046] Figure 1 A schematic diagram of the partitions of a hatchback car from the y-normal perspective;

[0047] Figure 2 A schematic diagram of the partitions of a hatchback car from the x-normal viewpoint;

[0048] Figure 3 This is a schematic diagram of the partitions of a hatchback car from the y-normal perspective.

[0049] Figure 4 This is a schematic diagram of the partitions of a hatchback car from the x-normal viewpoint.

[0050] Figure 5 This is a ternary coupled AI model architecture;

[0051] Figure 6 This is a schematic diagram of the deformation parameters;

[0052] Figure 7 The image shows the test set results of the 3D coupled AI model (where a is the prediction result from M to O, and b is the prediction result from O to N).

[0053] Figure 8 The geometric and rectangular wake partition numbers that should be modified first for the test set samples. Detailed Implementation

[0054] 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.

[0055] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0056] like Figure 5 As shown, an embodiment of the present invention provides a vehicle aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI, which includes the following steps:

[0057] Step 1: Divide the wake into regions based on vortex system characteristics;

[0058] H, L, and W represent the length, width, and height of a hatchback or direct-hatch car, respectively; the origin of the partition is the point where the longitudinal section of the car, the ground, and the rear cross-section of the car intersect; the partition coordinate axes (x*, y*, and z*) are defined with the positive directions of x*, y*, and z* respectively: from the front to the rear of the car, the normal direction of the longitudinal section of the car (from the passenger side to the driver side), and from the bottom to the top of the car; the units of the partition coordinate axes are x*=x / L, y*=2y / W, and z*=z / H, where x, y, and z are the actual dimensions in that coordinate axis direction. Figure 1 and Figure 2As shown, for a stepped-back car, the wake is first divided into rectangular wake zones based on flow characteristics, including the A-pillar vortex zone (corresponding to A in the figure), the C-pillar vortex zone, the upward-rising vortex zone at the tail (corresponding to D in the figure), the bottom longitudinal vortex zone (corresponding to G in the figure), the hairpin vortex zone (corresponding to R in the figure), and the far-flow zone (corresponding to P in the figure). Figure 3 and Figure 4 As shown, for a hatchback vehicle, the wake is first divided into rectangular regions based on flow characteristics: the upper recirculation bubble region (corresponding to U in the figure), the C-pillar vortex region (corresponding to C in the figure), the tail-end upward vortex region (corresponding to D in the figure), the bottom longitudinal vortex region (corresponding to G in the figure), and the far-flow region (corresponding to P in the figure). It should be noted that in the actual division, all regions are located on the positive y-axis. Figure 2 and Figure 4 The division on both sides of the y* axis is simply for ease of display.

[0059] Step 2: Establish a ternary coupled dataset;

[0060] Parametric geometric deformation of the car, mainly focusing on the rear region, was performed to obtain multiple sets of car models with different shapes. Numerical flow field simulation was performed on each car model to extract the regional average value of vortex characteristic quantities in each wake region, and a three-dimensional coupled dataset was constructed that couples geometric features, wake region features, and drag coefficient.

[0061] Step 3: Construction of the ternary coupled AI model;

[0062] The ternary coupling AI model consists of four parts: a mapping network between geometric features and wake partition features, a mapping network between wake partition features and drag coefficient, feature contribution analysis, and global model optimization. The mapping network between wake partition features and drag coefficient employs a dual-validation mechanism during training. The ternary coupling dataset from step 2 is input into the ternary coupling AI model to complete training, establishing the mapping relationship between geometric features, wake partition features, and drag coefficient.

[0063] Step 4: Prediction and feature optimization of new samples;

[0064] The geometric parameters of the new sample are input into the ternary coupled AI model trained in step 3. The ternary AI model can predict the characteristics of each wake region and the drag coefficient of the sample through the mapping network of geometric features and wake region features, as well as the mapping network of wake region features and drag coefficient. The optimization priority is obtained through feature contribution analysis and global model optimization. The ternary coupled AI model feeds back the geometric features and wake region features that should be optimized most for the new sample based on the optimization priority.

[0065] In a preferred embodiment of the present invention, in step 1, the C-pillar vortex region of the stepped-back car is divided into two parts: a C1 sub-region and a C2 sub-region (corresponding to...). Figure 1 and Figure 2 The C1 and C2 pillars of a hatchback car are located above and behind the trunk lid, respectively; the C-pillar vortex area of ​​a straight-back car is only a part, located on the side and rear of the trunk lid.

[0066] In a preferred embodiment of the present invention, in step 1, the rectangular wake regions of the stepped-back car are as follows in the z* direction: hairpin vortex region (1.00, 1.12), A-pillar vortex region (0.82, 1.00), C1 sub-region (0.64, 0.82), C2 sub-region (0.39, 0.82), rear updraft vortex region (0.15, 0.39), bottom longitudinal vortex region (0, 0.15), and far-flow region (0, 0.82); the rectangular wake regions of the straight-back car are as follows in the z* direction: upper recirculation bubble region (0.39, 1.00), C-pillar vortex region (0.39, 1.00), rear updraft vortex region (0.15, 0.39), bottom longitudinal vortex region (0, 0.15), and far-flow region (0, 1.00).

[0067] In a preferred embodiment of the present invention, in step 1, the division range of each zone of the stepped-back car in the y* direction is as follows: the range of the A-pillar vortex zone, the hairpin vortex zone, and the C1 sub-zone is (0, 1.00); the range of the C2 sub-zone, the rear upward vortex zone, the bottom longitudinal vortex zone, and the far-flow zone is (0, 1.10).

[0068] The range of each zone in the y* direction for a straight-back car is as follows: the range of the rear upward vortex zone, the bottom longitudinal vortex zone, and the far-flow zone is (0, 1.10); the range of the upper recirculation bubble zone is (0, 1.00); and the range of the C-pillar vortex zone is (1.00, 1.10).

[0069] In a preferred embodiment of the present invention, in step 1, the rectangular wake zones of the stepped-back car are categorized in the x* direction as follows: hairpin vortex zone (-0.40, -0.18), A-pillar vortex zone (-0.40, 0), C1 sub-zone (-0.26, 0), C2 sub-zone (0, 0.25), rear upturn vortex zone (0, 0.25), bottom longitudinal vortex zone (0, 0.25), and far-flow zone (0.25, 0.50).

[0070] The x* direction division range of each zone of the straight-back car is as follows: upper recirculation bubble zone (0, 0.25), C-pillar vortex zone (0, 0.25), rear upward vortex zone (0, 0.25), bottom longitudinal vortex zone (0, 0.25), and far-flow zone (0.25, 0.50).

[0071] In a preferred embodiment of the present invention, in step 2, the parametric geometric deformation includes adjusting at least one of the following: rear upturn angle, trunk lid height, rear window tilt angle, and bottom diffuser angle.

[0072] The eddy current characteristic quantities include at least one of the following: vorticity, Q criterion, and λ2 criterion.

[0073] In a preferred embodiment of the present invention, in step 3, the two networks, namely the mapping network of geometric features and wake partition features and the mapping network of wake partition features and drag coefficient, are implemented using a general machine learning or deep learning architecture, including but not limited to BP neural networks, time series networks and other intelligent fitting models.

[0074] In a preferred embodiment of the present invention, the dual verification mechanism in step 3 is as follows:

[0075] First verification: If the difference between the predicted values ​​of the drag coefficient of any two test samples is less than 0.001, the model is determined to be stuck in fixed value fitting and will be automatically retrained.

[0076] Second verification: If the predicted value of the resistance coefficient of any test sample exceeds 80% to 120% of the average value of the training set, the network training process will be automatically restarted until the prediction accuracy requirement is met.

[0077] In a preferred embodiment of the present invention, in step 3, the feature contribution analysis specifically involves: performing weight sensitivity analysis based on the trained mapping network of geometric features and wake partition features, and the mapping network of wake partition features and drag coefficient, calculating the contribution weight of each geometric feature and wake partition feature to the drag coefficient. The contribution weight is obtained by extracting the weights from the input layer to the hidden layer of each neural network in the ternary coupling model, calculating the average absolute value of the weights, and normalizing them. The magnitude of the absolute value of the weights is positively correlated with the influence intensity of the corresponding variable on the drag coefficient.

[0078] In a preferred embodiment of the present invention, in step 3, the optimization algorithm used for global model optimization includes at least one of random global search, genetic algorithm, multi-island genetic algorithm, particle swarm optimization, simulated annealing algorithm, and differential evolution algorithm, in order to obtain the theoretically optimal global geometric eigenvalue and the theoretically optimal wake partition eigenvalue with the minimum drag coefficient.

[0079] In a preferred embodiment of the present invention, in step 4, the optimization priority calculation method is: optimization priority = (standardized value of theoretical global optimal feature value - standardized value of current feature) × contribution weight;

[0080] The theoretically optimal eigenvalues ​​include the theoretically optimal geometric features and the theoretically optimal wake partition features. The larger the absolute value of the optimization priority, the higher the priority of that feature. One or more features with the largest absolute value of optimization priority are identified as the features that should be optimized first.

[0081] The standardized value of the current feature is calculated as follows:

[0082]

[0083]

[0084]

[0085] in, , and These are the standardized values ​​for geometric characteristics, wake zoning characteristics, and drag coefficient, respectively. , and These are the original values ​​for geometric characteristics, wake zoning characteristics, and drag coefficient, respectively. , and These are the training set mean values ​​for geometric features, wake zoning features, and drag coefficient, respectively. , and is the standard deviation of the training set for the corresponding feature.

[0086] like Figure 6 As shown, a three-dimensional coupled AI model was constructed using a stepped-back car as the object to achieve wake order reduction and feedback optimization. First, the car model was deformed, and three geometric features were selected: the rear window tilt angle M1, the ducktail M2, and the departure angle M3, with variation ranges of [-100, 100], [-50, 100], and [-50, 100] (in millimeters), respectively. The numerical simulation software used was STAR CCM+, with a simulated wind speed of 28 m / s. Seven wake zones were divided according to the aforementioned method and numbered (O1-O7 correspond to the A-pillar vortex zone, C1 sub-zone, C2 sub-zone, tail-end uplift vortex zone, bottom longitudinal vortex zone, far-flow zone, and hairpin vortex zone, respectively). Probes were evenly distributed according to the size of each zone, totaling 3790 probes. The zone with the fewest probes had 270, and the zone with the most had 900. The average probe vortex value of each zone was used as the wake zone characteristic.

[0087] A ternary coupled dataset is constructed, consisting of geometric features (feature number M), wake scrambling features (feature number O), and drag coefficient (feature number N). The training and test sets are randomly divided in an 8:2 ratio. To ensure simplicity in the implementation, both the mapping networks for geometric features and wake scrambling features, and the mapping network for wake scrambling features and drag coefficient, employ a two-layer cascaded backpropagation (BP) neural network structure. The first layer (M→O) has 12 hidden neurons, and the second layer (O→N) has 10 hidden neurons. The training algorithm uses Bayesian regularization, with a maximum of 250 iterations, a training objective error of 0.0005, and an initial learning rate of 0.008. A dual verification mechanism is introduced: the predicted drag coefficient values ​​are constrained to have a spacing of no less than 0.001, and the predicted drag coefficient values ​​are within 80%–120% of the training set mean, ensuring the model's generalization and stability.

[0088] like Figure 7 a and Figure 7 As shown in b, after training, the prediction of wake zoning features using geometric features showed good performance across all seven test set samples. Regarding the prediction of drag coefficients using wake zoning features, the predicted values ​​of all eight test samples closely matched the actual values. The mean absolute error (MAE) of drag prediction on the test set was 0.008242, and the mean squared error (MSE) was 8.759 × 10⁻⁶. -5 The actual values ​​and predicted values ​​are in high agreement, which verifies the accuracy and reliability of the model's predictions.

[0089] The code performs feature contribution analysis and global model optimization based on the trained ternary coupled AI model. The global optimization algorithm in this embodiment is a stochastic global search algorithm with progress feedback. According to the optimization priority calculation method, the geometric parameters to be adjusted first and the corresponding wake vortex region are automatically determined for each test sample. For example... Figure 8 As shown, test samples 1-8 respectively recommend prioritizing the modification of geometric parameters M1 / M2 and wake characteristics O1 / O4, thus achieving a closed loop from accurate wake field prediction to shape optimization.

[0090] 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 method for automotive aerodynamic optimization based on vortex system characteristics and three-dimensional coupled AI, characterized in that, Includes the following steps: Step 1: Based on the vehicle type and vortex characteristics, divide the wake into multiple rectangular wake zones; the vehicle types include stepped-back vehicles and straight-back vehicles; Step 2: Perform parametric geometric deformation on the rear region of the car to obtain multiple sets of car models with different shapes; perform flow field numerical simulation on each car model, extract the regional average value of vortex characteristic quantities in each wake partition, and construct a three-dimensional coupled dataset that couples geometric features, wake partition features, and drag coefficient. Step 3: Construct a ternary coupled AI model, which includes a mapping network between geometric features and wake partition features, a mapping network between wake partition features and drag coefficient, feature contribution analysis, and global model optimization. The mapping network between wake partition features and drag coefficient is trained using a dual verification mechanism. The ternary coupled dataset from Step 2 is input into the ternary coupled AI model to complete training, establishing the mapping relationship between geometric features, wake partition features, and drag coefficient. Step 4: Input the geometric parameters of the new sample into the trained ternary coupled AI model to predict the characteristics of each wake zone and the drag coefficient value of the new sample; obtain the optimization priority through feature contribution analysis and global model optimization, and feed back the geometric features and wake zone features that should be optimized most for the new sample based on the optimization priority.

2. The automotive aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI according to claim 1, characterized in that, In step 1, for a stepped-back car, the wake is divided into the A-pillar vortex region, C-pillar vortex region, tail-end rising vortex region, bottom longitudinal vortex region, hairpin vortex region, and far-flow region based on the flow characteristics; for a straight-back car, the wake is divided into the upper recirculation bubble region, C-pillar vortex region, tail-end rising vortex region, bottom longitudinal vortex region, and far-flow region based on the flow characteristics.

3. The automotive aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI according to claim 2, characterized in that, The C-pillar vortex area of ​​a stepped-back car is divided into two parts, C1 and C2, located above and behind the trunk lid, respectively; the C-pillar vortex area of ​​a straight-back car is only a part, located on the side and rear of the trunk lid.

4. The automotive aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI according to claim 3, characterized in that, In step 1, the origin of the partition is the point where the longitudinal section of the car, the ground, and the rear section of the car intersect. The positive directions of the partition coordinate axes are x*, y*, and z*, respectively, from the front to the rear of the car, the normal direction of the longitudinal section of the car from the passenger side to the driver side, and from the bottom to the top of the car. The units of the partition coordinate axes are x*=x / L, y*=2y / W, and z*=z / H, where L, W, and H are the length, width, and height of the car, respectively, and x, y, and z are the actual dimensions in the direction of the coordinate axis.

5. The automotive aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI according to claim 1, characterized in that, In step 2, the parametric geometry deformation includes adjusting at least one of the following: rear upturn angle, trunk lid height, rear window tilt angle, and bottom diffuser angle; The eddy current characteristic quantities include at least one of the following: eddy current, Q criterion, and λ2 criterion.

6. The automotive aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI according to claim 1, characterized in that, In step 3, the two network components, namely the mapping network of geometric features and wake partition features, and the mapping network of wake partition features and drag coefficient, are implemented using a general machine learning or deep learning architecture.

7. The automotive aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI according to claim 1, characterized in that, In step 3, the dual verification mechanism is as follows: First verification: If the difference between the predicted values ​​of the drag coefficient of any two test samples is less than 0.001, the model is determined to be stuck in fixed-value fitting and will be automatically retrained. Second verification: If the predicted value of the resistance coefficient of any test sample exceeds 80% to 120% of the average value of the training set, the network training process will be automatically restarted until the prediction accuracy requirement is met.

8. The automotive aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI according to claim 1, characterized in that, In step 3, the feature contribution analysis specifically involves: performing weight sensitivity analysis based on the trained mapping network of geometric features and wake partition features, as well as the mapping network of wake partition features and drag coefficient, to calculate the contribution weight of each geometric feature and wake partition feature to the drag coefficient. The contribution weight is obtained by extracting the weights from the input layer to the hidden layer of each neural network in the ternary coupling model, calculating the average absolute value of the weights, and normalizing them. The magnitude of the absolute value of the weight is positively correlated with the strength of the influence of the corresponding variable on the drag coefficient.

9. The automotive aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI according to claim 1, characterized in that, In step 3, the optimization algorithm used for global model optimization includes at least one of random global search, genetic algorithm, multi-island genetic algorithm, particle swarm optimization, simulated annealing algorithm, and differential evolution algorithm, in order to obtain the theoretically optimal global geometric eigenvalue and the theoretically optimal wake partition eigenvalue with the minimum drag coefficient.

10. The automotive aerodynamic optimization method based on vortex system characteristics and three-dimensional coupled AI according to claim 9, characterized in that, In step 4, the optimization priority calculation method is as follows: optimization priority = (standardized value of theoretical global optimal feature value - standardized value of current feature) × contribution weight; Among them, the theoretical global optimal eigenvalues ​​include the theoretical global optimal geometric features and the theoretical global optimal wake partition features. One or more features with the largest absolute value of optimization priority are determined as the features that should be optimized first. The standardized value of the current feature is calculated as follows: in, , and These are the standardized values ​​for geometric characteristics, wake zoning characteristics, and drag coefficient, respectively. , and These are the original values ​​for geometric characteristics, wake zoning characteristics, and drag coefficient, respectively. , and These are the training set mean values ​​for geometric features, wake zoning features, and drag coefficient, respectively. , and is the standard deviation of the training set for the corresponding feature.

Images