Mixture-of-Experts AI cuts vehicle drag by 10% in shape optimization
94.3% trend accuracy achieved on heterogeneous aerodynamic datasets
Engineering shape optimization has long been bottlenecked by manual expert setup and unreliable surrogate models, especially for heterogeneous geometry databases. A new framework from researchers at multiple institutions addresses this by integrating knowledge-based constraints directly into a DFFD-based deformation operator. The core innovation is a Mixture-of-Experts Neural Operator (MoE-NO) that learns to handle diverse aerodynamic shapes with high precision. To ensure reliability, the system uses Mahalanobis distance on the MoE-NO encoder to detect out-of-distribution designs and trigger physics-solver feedback for local sample enrichment, blending data-driven speed with trust in physics.
On in-house datasets for MPV, SUV, and Sedan models, MoE-NO achieved a test-set MAPE of 1.16% and a trend-prediction accuracy of 94.34%, beating best baselines of 1.52% and 90.34% respectively. More importantly, vehicle shape-optimization experiments yielded CFD-validated drag coefficient reductions of approximately 4% to 10%. This represents a significant step toward automated, high-confidence aerodynamic design that can be deployed in real engineering workflows without constant human oversight.
- MoE-NO outperforms baselines: 1.16% MAPE vs 1.52% on aerodynamic drag prediction
- Uncertainty estimation (Mahalanobis distance) flags out-of-distribution designs for physics-solver validation
- CFD-validated drag reductions of 4% to 10% across MPV, SUV, and Sedan geometries
Why It Matters
Automates aerodynamic shape optimization with high confidence, reducing reliance on expert engineers and manual tuning.