Abstract
Traditional aerodynamic coefficient prediction relies on computationally intensive CFD simulations or costly wind tunnel tests. This paper presents a modular deep-learning framework that predicts these coefficients by fusing 3D point cloud geometry with flow conditions through a cross-attention mechanism. By incorporating physical-informed training, the model ensures consistency with aerodynamic laws and achieves superior generalization. Evaluations on the DrivAerNet++ and a custom AircraftData dataset demonstrate state-of-the-art performance, particularly in data-scarce scenarios.
Introduction
In aerospace and automotive engineering, accurately predicting aerodynamic coefficients such as lift ($C_L$) and drag ($C_D$) is vital for design optimization. While data-driven AI models offer significant speed advantages over traditional Computational Fluid Dynamics (CFD), they often struggle with the complex interplay between 3D geometry and varying flow conditions (Mach number and Angle of Attack). Furthermore, pure data-driven approaches can produce physically inconsistent results and require vast amounts of high-fidelity data, which is often scarce.
Methodology
The proposed framework addresses these challenges through a modular architecture consisting of four main components:
- Shape Encoder: Based on the PointGPT architecture, it processes 3D point clouds using Farthest Point Sampling (FPS) and K-Nearest Neighbors (KNN) to extract local geometric features.
- Flow Condition Encoder: A Multi-Layer Perceptron (MLP) encodes environmental parameters—Mach number (Ma) and Angle of Attack (AoA)—into high-dimensional vectors.
- Cross-Attention Fusion: This core component utilizes a Transformer-based decoder with cross-attention. It allows the model to dynamically weight geometric features based on the specific flow conditions, effectively "attending" to critical surfaces like wings or fuselages as the AoA changes.
- Physical-Informed Training: To ensure physical consistency, fundamental aerodynamic equations (e.g., relating axial force, side force, and pitch moment) are integrated into the loss function. This "Physics-Informed Neural Network" (PINN) approach guides the model toward physically plausible solutions.
Experimental Setup
The model was evaluated using two datasets:
- DrivAerNet++: A public dataset of 8,000 diverse car designs with high-fidelity CFD labels.
- AircraftData: A custom dataset of 1,000 aircraft geometries across 96 flow conditions, totaling approximately 100,000 samples.
Results and Analysis
The framework achieves state-of-the-art (SOTA) performance on the DrivAerNet++ benchmark, outperforming established models like PointNet and RegDGCNN across multiple metrics, including Mean Squared Error (MSE) and $R^2$ score.
Key insights from the study include:
- Dynamic Attention: Visualization of attention maps confirms that the model shifts its focus across the geometry in response to changing flow conditions, mimicking physical airflow interactions.
- Transfer Learning Benefits: Utilizing a pre-trained shape encoder allows the model to leverage general 3D geometric knowledge, significantly improving performance when aerodynamic data is limited.
- Physics-Informed Fine-tuning: Combining pre-trained weights with physical constraints yielded the lowest Mean Relative Error (MRE) of $4.610 \times 10^{-2}$, demonstrating that physics acts as a powerful supervisory signal for extrapolation.
Conclusion
This research introduces a robust, physically grounded framework for aerodynamic coefficient prediction. By combining modular encoders with cross-attention fusion and physics-informed training, the model achieves high accuracy and generalization. Future work will explore integrating more complex partial differential equations, such as Navier-Stokes, and applying the model to real-time design optimization loops.
Summary updated