Abstract
High-fidelity Computational Fluid Dynamics (CFD) is the gold standard for automotive design but remains computationally prohibitive for rapid iteration. While Machine Learning (ML) surrogate models offer millisecond-scale predictions, individual architectures often struggle to generalize across the complex, multi-scale geometries of modern vehicles. This paper introduces a Mixture of Experts (MoE) framework that dynamically integrates three distinct AI architectures—Graph Neural Networks, Neural Operators, and Convolutional Neural Networks. By utilizing a gating network with entropy regularization, the proposed model achieves a 20% reduction in pressure prediction error compared to the best standalone expert on the DrivAerML dataset.
Introduction
In automotive aerodynamics, engineers face a trade-off between the accuracy of high-fidelity simulations and the speed of surrogate models. Current ML approaches are fragmented by architectural "inductive biases": Graph Neural Networks (GNNs) excel at irregular shapes, Neural Operators capture global flow patterns, and Convolutional Neural Networks (CNNs) efficiently process smooth surfaces.
Addressing the "no free lunch" theorem in aerodynamics, this research asks whether these disparate strengths can be combined. Instead of selecting a single "best" model, the authors propose an intelligent meta-learning framework that routes specific geometric features to the most capable architecture.
Methodology
The proposed Mixture of Experts (MoE) framework consists of three primary components:
- Heterogeneous Experts: The system employs three state-of-the-art models: DoMINO (a point-cloud-based neural operator), X-MeshGraphNet (a multi-scale GNN), and FigConvNet (a CNN-based global convolution network).
- Gating Network: A Multi-Layer Perceptron (MLP) acts as a router, analyzing local geometry and surface normals to assign weights to each expert at every point on the vehicle's surface.
- Entropy Regularization: To prevent "expert collapse"—where the system over-relies on a single model—the authors introduced entropy regularization into the loss function. This forces the gating network to maintain a diverse strategy and leverage the unique strengths of all three experts.
Experimental Setup
The framework was evaluated using the DrivAerML dataset, which includes 500 unique geometric variants of the DrivAer generic vehicle model. Unlike simplified steady-state datasets, DrivAerML utilizes high-fidelity, scale-resolving RANS-LES simulations, providing a rigorous benchmark for surface pressure, wall shear stress, and volumetric flow fields.
Results and Discussion
The MoE model demonstrated superior performance across all metrics. It achieved an L-2 prediction error of 0.08 for pressure, representing a 20% improvement over the best individual expert (DoMINO) and a 61.9% improvement over the weakest. Similar gains of 13-17% were observed in wall shear stress predictions.
As shown in the figure above, the gating network learns a spatially-variant strategy that aligns with physical principles. It naturally assigns higher weights to the point-based operator (DoMINO) in regions of high pressure and stagnation, while favoring the grid-based CNN (FigConvNet) for large, flat surfaces like the roof and doors. This synergy allows the composite model to outperform its individual components by identifying and utilizing "pockets of expertise."
Conclusion and Future Work
This research demonstrates that a Mixture of Experts approach effectively mitigates the limitations of individual ML architectures in complex aerodynamic tasks. Future work will focus on extending this framework to 3D volumetric flow fields and exploring graph-based gating networks to better capture vehicle topology. Additionally, the entropy of the gating network may serve as a valuable proxy for model uncertainty, identifying regions where traditional CFD validation remains necessary.
Summary updated