Experimental Data Acquisition and Reconstruction based on CFD calibration

Accurate wind pressure prediction around urban buildings is essential for structural safety and environmental comfort. Wind tunnel (EXP) tests offer high accuracy but are time-consuming and costly. Computational Fluid Dynamics (CFD) simulations are efficient and low-cost but often have physical approximations. To bridge this gap, we propose a CFD-calibrated experimental data acquisition and reconstruction framework using Multi-Fidelity Neural Networks (MFNN), which efficiently fuses low-fidelity CFD and high-fidelity EXP data for improved prediction and generalization.

The framework includes:

(1) CFD→EXP(one single layer) mean pressure mapping using MLP;

(2) Spatial reconstruction from CFD to EXP using DeepONet for 3D fields;

(3) Main→Interfering Column Prediction, inferring unmeasured experiment values via spatial interaction mappings learned from CFD;

(4) MFNN (single case), fusing low-fidelity CFD and sparse high-fidelity EXP data;

(5) MFNO (multiple cases), extending MFNN for multiple layouts and wind angles. Validated on 888 CFD-EXP cases (37 layouts × 24 wind directions), MFNN outperforms interpolation baselines, especially in regions of complex flow. It enhances accuracy and robustness in unseen layouts and wind angles.

This study provides a cost-effective and high-fidelity approach for wind field estimation in urban environments, contributing to digital twin development and aerodynamic design optimization.