Exploration of computational formulations for wind-induced interference effects on high-rise buildings via Kolmogorov–Arnold networks

In dense urban environments, wind-induced interference effects introduce significant uncertainties in predicting aerodynamic forces on high-rise buildings. Conventional methods such as wind tunnel tests and computational fluid dynamics (CFD) suffer from high cost and long runtime, while multivariate regression analysis (MRA) lacks the ability to capture nonlinear couplings, and black-box models (e.g., CatBoost) fail to ensure physical consistency. To overcome these limitations, this study proposes a KM-KAN-SR framework that integrates Kolmogorov–Arnold Networks (KAN) with K-means clustering (KM) and symbolic regression (SR) to derive explicit aerodynamic force formulas. Benchmarking results highlight the superior performance of KM-KAN-SR. Specifically, KM-KAN-SR achieved R 2 values of 0.931 and 0.961 for C Fx_mean and C Fy_mean , respectively, significantly higher than those of CFD (0.830 and 0.795) and MRA (0.849 and 0.532). Moreover, the expressions derived by KM-KAN-SR are on average 50 % less complex than those of conventional KAN-SR and remain concise and interpretable. In terms of efficiency, KM-KAN-SR generates predictions within milliseconds, whereas CFD requires millions of grid cells and hours of computation under large-eddy simulation. Sensitivity analyses further reveal that KM-KAN-SR preserves smooth, physically consistent aerodynamic trends, unlike CatBoost which exhibits step-like discontinuities. Overall, the KM-KAN-SR framework demonstrates high predictive accuracy, low formula complexity, strong physical consistency, and orders-of-magnitude faster computation, providing a robust and interpretable tool for wind-resistant design of high-rise buildings. • KAN with K-means are integrated to derive explicit aerodynamic force formulas for high-rise building interference. • Wind incidence angle as key was identified via KAN pruning. • KM-KAN-SR achieves outperforming CFD and MRA. • Reduces formula complexity by 50 % vs KAN-SR, cutting computation time vs CFD.