AI and the World's Hard Problems · Physics · arXiv · 2026-10-26
CoPiT: A conditional Position-induced Transformer for aerodynamic field prediction on unstructured meshes
High-fidelity RANS simulations on unstructured meshes are accurate but costly, which makes many-query tasks such as design exploration expensive. Existing neural operators often simply concatenate global parameters with pointwise inputs, leaving their effect implicit. The proposed CoPiT builds on a position-induced Transformer and injects the global condition through multi-stage feature-wise linear modulation, reducing surface and volume losses by about 54% and 24% relative to Transolver on AirfRANS and achieving the lowest normalized test MSE among compared baselines on BlendedNet++.