Geometry-Aware Anisotropic Boundary Correction for Aerodynamic Simulation

2026-06-08

AI Summary

This work addresses the limitation of conventional neural operators in aerodynamic simulations, which typically treat boundary regions isotropically and thus fail to accurately capture the distinct physical behaviors along tangential and normal directions near walls. To overcome this, the authors propose the GeoABC framework, which uniquely transforms static boundary geometry from an input feature into a structural prior embedded within the intermediate representations of the neural operator, thereby explicitly modeling near-wall anisotropy. The approach incorporates a geometry-conditioned anisotropic boundary correction module that can be flexibly integrated into various backbone architectures. Evaluated on both 2D airfoil and 3D automotive aerodynamic tasks, GeoABC reduces the average relative L2 error in near-boundary regions by 38%, substantially narrowing the performance gap of mainstream neural operators in wall-proximal predictions.

Abstract

Aerodynamic simulation is a key component of engineering shape design, where core quantities such as the surface pressure coefficient strongly depend on flow dynamics near solid boundaries. Neural operators provide an efficient alternative to expensive Computational Fluid Dynamics (CFD) solvers. However, conventional methods treat the boundary region isotropically, failing to account for the distinct physical behaviors along the boundaries. In reality, the aerodynamic process exhibits anisotropy: along the tangential direction, flow propagates along the wall; along the normal direction, physical quantities are constrained by the wall. To explicitly model the distinct physical behaviors, we propose GeoABC, a geometry-conditioned anisotropic boundary correction framework. GeoABC leverages the boundary geometries to introduce direction-aware boundary correction into the intermediate representations of neural operators, transforming boundary geometry from static input features into a structural prior that modulates physical prediction. On 2D airfoil and 3D car tasks, GeoABC consistently adapts to multiple neural operator backbones, reducing near-boundary relative $L_2$ error by $\sim$38\% on average, narrowing the structural near-wall gap shared by mainstream neural operators, and advancing neural operators toward high-fidelity aerodynamic simulation.

Problem

• aerodynamic simulation
• anisotropic boundary
• neural operators
• boundary correction
• near-wall flow

Innovation

• anisotropic boundary correction
• geometry-aware neural operators
• aerodynamic simulation
• boundary geometry prior
• near-wall flow modeling