A Generative Machine Learning Surrogate Model for High-Fidelity Urban Wind Field Prediction

Accurate prediction of urban wind fields is critical for applications such as Unmanned Aerial System (UAS) trajectory planning, pedestrian comfort analysis, and urban heat island mitigation. However, traditional high-fidelity Computational Fluid Dynamics (CFD) simulations are computationally prohibitive for extensive parametric analyses, while low-fidelity empirical models lack the necessary localized accuracy within complex urban morphologies.

This work presents the development of a Machine Learning Surrogate Model designed to bridge the gap between low-fidelity approximations and high-fidelity simulations. Using generative diffusion architectures, the proposed data-driven surrogate serves as a lightweight, data-fusion-capable tool for rapid wind-flow prediction. The model is trained on a comprehensive dataset generated through parametric simulations of a cityscape, combining high-fidelity solvers with low-fidelity kinematic approximations. The training space systematically covers the range of wind magnitudes and directions derived from annual observed wind roses.

To ensure broad applicability, the surrogate model is designed to generalize across varying building densities, enabling accurate interpolation of flow fields transitioning from dense, high-rise downtown centers to sparse suburban layouts. Furthermore, the model aims to incorporate a robust data-fusion methodology to assimilate sparse, real-world meteorological measurements, dynamically enhancing predictive accuracy.

Ultimately, this generative framework demonstrates substantial downstream utility, offering near-high-fidelity accuracy at a fraction of the computational cost, thereby enabling rapid, reliable environmental modeling for advanced urban infrastructure and autonomous aerial navigation.