UniAero: A Unified Framework for Global Drag and Local Flow Field Prediction

ICLR 2026 Conference Submission15216 Authors

Keywords: Automotive aerodynamic, computational fluid dynamics, drag coefficient, transformer

TL;DR: UniAero is a unified framework that jointly predicts global drag and local surface fields (pressure, WSS) from 3D vehicle geometry ~1 s per car, higher accuracy.

Abstract:

Automotive aerodynamic design relies simultaneously on global metrics (e.g., drag coefficient $C_d$) and local flow field information (e.g., surface pressure and wall shear stress). While the former dictates overall fuel efficiency, the latter informs detailed design and performance optimization. Existing deep-learning surrogates typically focus on either global or local predictions individually, failing to optimize both tasks jointly, thus limiting their effectiveness in iterative design processes. To address this, we propose UniAero, a unified framework that jointly predicts $C_d$ and dense surface fields from 3D automotive geometry. UniAero combines (i) a Physically Stratified Mixture-of-Experts (Phys$^{2}$MoE) with scale-sensitive experts and physics-aware gating for multi-task, multi-scale learning; (ii) Serialized Patch Attention to process large meshes efficiently while preserving long-range interactions; and (iii) a hierarchical encoder with Geometry-aware Position Encoding (GPE) to capture subtle shape cues.
Experiments on three industrial datasets demonstrate that, by explicitly leveraging the inherent coupling between global and local aerodynamic phenomena through joint modeling. UniAero reduces drag error by 12% and improves local-field accuracy by 16% over strong single-task baselines, with ~1s inference per vehicle on a single GPU, far faster than CFD simulations. With its superior accuracy, speed, and coherence, UniAero holds significant promise for automotive aerodynamic design. The code is available at https://anonymous.4open.science/r/ICLR2026UniAero.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 15216