Efficient Coupling of Urban Wind Fields and Drone Flight Dynamics Using Convolutional Autoencoders

Abstract

Highlights: What are the main findings? Two-dimensional convolutional autoencoders compress LES-based urban wind fields by 91% while maintaining high reconstruction fidelity. CAE-reconstructed wind fields yield flight dynamics predictions for a subscale Cessna that closely match those obtained with the original LES winds. What are the implications of the main findings? Reduced storage and memory requirements enable larger urban wind domains to be used in real-time flight dynamics simulations. Compressed wind-field representations facilitate efficient sharing and reuse of disturbance models in collaborative AAM studies. Flight safety is central to the certification process and relies on assessment methods that provide evidence acceptable to regulators. For drones operating as Advanced Air Mobility (AAM) platforms, this requires an accurate representation of the complex wind fields in urban areas. Large-eddy simulations (LES) of such environments generate datasets from hundreds of gigabytes to several terabytes, imposing heavy storage demands and limiting real-time use in simulation frameworks. To address this challenge, we apply a Convolutional Autoencoder (CAE) to compress a 40 m-deep section of an LES wind field. The dataset size was reduced from 7.5 GB to 651 MB, corresponding to a 91% compression ratio, while maintaining maximum magnitude errors within a few tenths of the spatio-temporal wind velocity. Predicted vehicle responses showed only marginal differences, with close agreement between the full LES and CAE reconstructions. These findings demonstrate that CAEs can significantly reduce the computational cost of urban wind field integration without compromising fidelity, thereby enabling the use of larger domains in real-time and supporting efficient sharing of disturbance models in collaborative studies.