Multi-fidelity Data Reconstruction for Wind Pressure on Building

Accurate measurement of wind-induced pressures on buildings is crucial yet costly, especially under interference effects where high-fidelity wind tunnel experiments (EXP) remain limited in spatial resolution and sample size. In contrast, computational fluid dynamics (CFD) simulations offer scalable, low-cost data but suffer from systematic biases. Bridging this fidelity gap poses a significant challenge in wind engineering. This study proposes an AI-enhanced framework for experimental data reconstruction, leveraging multi-fidelity data fusion. The focus lies on complex aerodynamic interference scenarios between two square buildings. A comprehensive CFD–EXP dataset spanning 888 configurations is constructed, covering multiple building layouts and incident wind angles. The sparse-to-dense reconstruction is performed, where limited high-fidelity data are used to enhance predictions. Sparse experimental sensors are incorporated into multi-fidelity neural network (MFNN) and neural operator (MFNO) frameworks: MFNN provides case-specific reconstructions, while MFNO generalizes across building spacings and wind angles without retraining per case. Experimental results confirm significant improvements in spatial reconstruction under sparse supervision. The proposed framework demonstrates strong generalization and robustness. This work advances AI-enhanced experimental data reconstruction, reducing testing costs while enhancing prediction reliability in wind engineering.