Abstract:A novel strategy for generating datasets is developed within the context of drag prediction for automotive geometries using neural networks. A primary challenge in this space is constructing a training databse of sufficient size and diversity. Our method relies on a small number of starting data points, and provides a recipe to interpolate systematically between them, generating an arbitrary number of samples at the desired quality. We test this strategy using a realistic automotive geometry, and demonstrate that convolutional neural networks perform exceedingly well at predicting drag coefficients and surface pressures. Promising results are obtained in testing extrapolation performance. Our method can be applied to other problems of aerodynamic shape optimization.

What problem does this paper attempt to address?

The paper aims to address the problem of predicting the drag coefficient in automotive aerodynamic design, particularly the challenges faced when using neural networks for data-driven automotive aerodynamic design. The main issues include: 1. **Constructing a sufficiently large and diverse training database**: In automotive design, a large number of Computational Fluid Dynamics (CFD) simulations are required to obtain drag data for each design point, which is both expensive and time-consuming. Additionally, due to the proprietary nature of commercial designs, there is a lack of publicly available datasets of real automotive geometries. 2. **High complexity of the parameter space**: The parameter space of modern automotive design is highly complex, leading to high sampling costs. Moreover, automotive design not only considers aerodynamic efficiency but also aesthetics, regulatory compliance, and manufacturability. To address these issues, the paper proposes a systematic approach to generate datasets for training, which relies on a small number of initial data points and provides a method to systematically interpolate between these points to generate any number of samples of the required quality. Specifically, the key contributions of the paper include: - **Systematic dataset generation technique**: A novel strategy is developed for generating datasets when using neural networks for drag prediction of automotive geometries. This method can start from a few initial data points and systematically interpolate to generate a large number of samples, thus constructing a sufficiently large training database. - **Geometry representation and interpolation**: By converting the original geometries into binary grid representations and then further into forms with Signed Distance Functions (SDF), interpolation operations can be performed. Interpolation on these SDFs using convex hulls and barycentric mapping generates new automotive geometries. - **Fluid dynamics simulation**: Large Eddy Simulation (LES) with wall models is used to obtain flow field solutions, serving as the ground truth data source for neural network training. - **Application of Convolutional Neural Networks**: Convolutional Neural Networks (CNN) are used to predict drag coefficients and surface pressure. The study also demonstrates the performance of CNNs in predicting quantities at different levels of abstraction (scalar drag coefficients, pressure vectors on axisymmetric surfaces, and full surface pressure information in three views). Through this approach, researchers can effectively generate a training dataset with sufficient diversity and demonstrate the excellent performance of neural networks in the task of drag prediction.

A systematic dataset generation technique applied to data-driven automotive aerodynamics

DrivAerNet: A Parametric Car Dataset for Data-Driven Aerodynamic Design and Graph-Based Drag Prediction

Data-Driven Car Drag Prediction With Depth and Normal Renderings

DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks

DrivAerML: High-Fidelity Computational Fluid Dynamics Dataset for Road-Car External Aerodynamics

Surrogate Modeling of Car Drag Coefficient with Depth and Normal Renderings

Drag prediction of rough-wall turbulent flow using data-driven regression

A Synergistic Framework Leveraging Autoencoders and Generative Adversarial Networks for the Synthesis of Computational Fluid Dynamics Results in Aerofoil Aerodynamics

A supervised neural network for drag prediction of arbitrary 2D shapes in low Reynolds number flows

Physics-guided Design and Learning of Neural Networks for Predicting Drag Force on Particle Suspensions in Moving Fluids

DeepJEB: 3D Deep Learning-based Synthetic Jet Engine Bracket Dataset

Active Learning over DNN: Automated Engineering Design Optimization for Fluid Dynamics Based on Self-Simulated Dataset

Data-driven modeling for unsteady aerodynamics and aeroelasticity

WindsorML: High-Fidelity Computational Fluid Dynamics Dataset For Automotive Aerodynamics

High precision aerodynamic heat prediction method based on data augmentation and transfer learning

Aero-Nef: Neural Fields for Rapid Aircraft Aerodynamics Simulations

Establishment of a Generalizable Model on a Small-Scale Dataset to Predict the Surface Pressure Distribution of Gas Turbine Blades

Airfoil aerodynamic performance prediction using machine learning and surrogate modeling

Enhanced data efficiency using deep neural networks and Gaussian processes for aerodynamic design optimization

Graph Convolutional Multi-Mesh Autoencoder for Steady Transonic Aircraft Aerodynamics

Framework Of Airfoil Max Lift-To-Drag Ratio Prediction Using Hybrid Feature Mining And Gaussian Process Regression