• Hybrid AI–CFD frameworks for curved-roof wind prediction. • Ten surrogate models compared using a validated CFD dataset. • ETR showed the best balance of accuracy and efficiency. • GS-SVR performed best in low-speed wind prediction. • Practical guidance is provided for model selection in design. Large-span buildings with curvilinear roofs generate complex separation and wake flows that affect pedestrian wind comfort and design performance, yet conventional Computational Fluid Dynamics (CFD) remains too computationally expensive for rapid iteration. This study aims to develop and comparatively evaluate hybrid Artificial Intelligence–CFD surrogate frameworks for fast and reliable prediction of wind velocities around curved-roof buildings, with emphasis on robustness across different velocity regimes. A parametric workflow generated 105 curved-roof geometries and 1,050 steady Reynolds-Averaged Navier–Stokes simulations to form a consistent training–testing dataset, and the CFD outputs were validated using wind-tunnel hot-wire measurements. Ten surrogate frameworks were implemented, covering linear regression, support vector regression, neural networks, gradient-boosting methods, and ensemble forests. Beyond conventional pointwise indicators (accuracy and computational cost), a distribution-aware assessment was introduced using cumulative distribution consistency to check whether each surrogate reproduces the full velocity distribution, especially in low-speed ranges relevant to comfort. Results show that Extra Trees achieves the best overall accuracy–efficiency trade-off, grid-search-optimized support vector regression provides the highest precision for low-velocity prediction but at markedly higher training cost, and Light Gradient Boosting Machine offers rapid screening with stable, acceptable accuracy. The comparative findings provide actionable guidance for selecting surrogate frameworks to accelerate wind-environment evaluation and support sustainable curved-roof building design.
Gan et al. (Fri,) studied this question.