Abstract:

Urban Air Mobility (UAM) experiences difficulties from uncertain surface level conditions, sometimes referred to as "microweather," such as varying levels of urban wind velocity. Standard computational methods for weather simulation can be inefficient or fail to represent turbulence as stochastic. This paper introduced probabilistic generative modeling framework to establish the connection between macro- to micro- weather to develop a framework to predict wind realistically from existing, already established regional forecasts. Moreover, state-of-the-art Deep Generative Models (DGMs) are used for development, both Denoising Diffusion Probabilistic Models (DDPM) and Flow Matching (FM) Models, to produce distributions of wind velocities at high resolution. The DGMs achieve better performance relative to Gaussian Mixture Models (GMM) to capture the conditional dependencies between regional forecasts and the behavior of local fluctuations. The approach is validated against Sonic Detection and Ranging (SoDAR) data, finding that DGMs produce statistically consistent wind profiles, while improving accuracy. Overall, the framework represents a significant advancement for UAM flight planning, safety, and operational efficiency while also representing scalable, AI-based prediction of microweather information that will facilitate autonomous and eVTOL aviation ecosystems