RETO: A Rotary-Enhanced Transformer Operator for High-Fidelity Prediction of Automotive Aerodynamics
New rotary-enhanced transformer cuts prediction error by 16-23% on car designs
A team led by Bojun Zhang from multiple Chinese universities has developed RETO (Rotary-Enhanced Transformer Operator), a neural solver designed to predict automotive aerodynamics with significantly higher accuracy than existing models. The key innovation is a dual-stage spatial awareness mechanism: sinusoidal-cosine encodings provide global reference points, while rotary positional encodings (RoPE) use unitary rotations to capture relative displacements with translation invariance. This allows RETO to preserve local gradient details that other operators blur.
On the DrivAerML high-fidelity dataset, RETO achieves relative L2 errors of 0.089 for surface pressure and 0.097 for velocity — improvements of 23% and 19% over the Transolver baseline (0.116 and 0.121). On ShapeNet, RETO's overall error of 0.063 bests Transolver's 0.075 (16% improvement) and RegDGCNN's 0.125 (50% improvement). Information-theoretic analysis confirms RETO's attention mechanism is more focused: entropy peaks at 0.35 vs Transolver's 0.75 at 10^4 resolution. This means the model preserves localized aerodynamic gradients rather than diffusing them globally, crucial for accurate drag and lift predictions in vehicle design.
- RETO achieves 0.063 relative L2 error on ShapeNet, 16% better than Transolver (0.075)
- On DrivAerML, pressure error drops 23% (0.089 vs 0.116) and velocity error drops 19% (0.097 vs 0.121)
- Entropy peak of 0.35 (vs 0.75 for Transolver) indicates highly focused attention on local spatial features
Why It Matters
Faster, more accurate aerodynamic simulations could slash vehicle design cycles and improve EV range predictions.