Reproduce “Distance-Aware Attention and Wall-Distance Expert Routing for Transformer-Based 3D Flow Prediction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sanghyeon Kim, Sunwoong Yang, Namwoo Kang
Wall-distance conditioning is incorporated into both information retrieval and feature transformation in transformer surrogates for three-dimensional automotive aerodynamics. On DrivAerML, the combined method lowers relative errors for surface pressure, wall shear, volume pressure, and velocity by up to 12.5% in the 50-case setting, with larger gains at 300 training cases and consistent improvements across alternative backbones and vehicle geometries.
Key insights
- DA-CA changes the content of a query rather than simply weighting nearby anchors, enabling different attention heads to learn distinct near-wall and outer-flow retrieval patterns.
- The two components address different stages of prediction: DA-CA mainly improves information retrieval near the wall, while SVMoE prevents a shared transformation from favoring near-wall points at the expense of the wake and far field.
- Unsupervised routing organizes volume experts into wall-distance bands and surface experts into spatially coherent geometric regions, providing an interpretable physical partition without manually assigned labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.