Major research outcome
- Prof. Sang Lee’s Research Group Develops Real-Time Urban Turbulence Machine Learning Model for Safe Urban Air Mobility (UAM) Operations
- 관리자 |
- 2026-05-04 20:54:28|
- 50
- 2026-05-04 20:54:28|
Accurate prediction of wind flow (turbulence) between complex urban buildings is essential for the safe operation of Urban Air Mobility (UAM), as well as for analyzing pedestrian comfort and pollutant dispersion. While traditional high-fidelity Large-Eddy Simulations (LES) provide high accuracy, their prohibitive computational costs have limited their application in real-time forecasting or large-scale scenario analysis. This research overcomes these temporal constraints and presents an innovative machine-learning-based numerical model that accelerates the practical real-time application of urban flow prediction.
Under the supervision of Professor Sang Lee, lead author Jun Park (Postdoctoral Researcher), Haechan Kim (Master's Student), and Professor Panagiotis Tsiotras (Georgia Institute of Technology) developed ‘3DSwinUrbanNet,’ a model based on the 3D Swin-Transformer architecture. The core technology of this model is 3D building geometry embeddings, which directly integrate building information into the encoder, backbone, and decoder stages of the machine learning model. This allows the model to accurately recognize wind patterns even in complex terrain environments with randomized building heights or 45-degree rotations. By designing the model to perform high-precision forecasting at 1-minute time steps—a significant increase from the 0.05-second steps used in traditional LES—computational efficiency has been maximized.
The results demonstrate that 3DSwinUrbanNet completed a 20-minute urban flow forecast in just 2.58 seconds, a task that typically requires approximately 1 hour and 25 minutes (5,120 seconds) using traditional LES. This represents a 1,984-fold speedup, proving that real-time forecasting is possible on a consumer-grade graphic card (RTX 4090) without the need for expensive supercomputers. Despite this rapid acceleration, the model faithfully reproduces critical physical phenomena, such as leading-edge updrafts and inter-building recirculation, with LES-level fidelity. This achievement provides essential real-time wind information for UAM trajectory planning and safety guidelines, while also offering a technical roadmap for pollutant dispersion prevention and smart city weather forecasting systems. By significantly lowering the barrier to high-precision urban turbulence forecasting through ML, this technology is expected to enhance the feasibility of next-generation future mobility industries.
This study was published in April 2026 in 'Physics of Fluids', a top-tier international journal in the field of fluid mechanics (ranked within the Top 5% and Q1 in the JCR Physics, Fluids & Plasma category).
DOI: https://doi.org/10.1063/5.0324405
Under the supervision of Professor Sang Lee, lead author Jun Park (Postdoctoral Researcher), Haechan Kim (Master's Student), and Professor Panagiotis Tsiotras (Georgia Institute of Technology) developed ‘3DSwinUrbanNet,’ a model based on the 3D Swin-Transformer architecture. The core technology of this model is 3D building geometry embeddings, which directly integrate building information into the encoder, backbone, and decoder stages of the machine learning model. This allows the model to accurately recognize wind patterns even in complex terrain environments with randomized building heights or 45-degree rotations. By designing the model to perform high-precision forecasting at 1-minute time steps—a significant increase from the 0.05-second steps used in traditional LES—computational efficiency has been maximized.
The results demonstrate that 3DSwinUrbanNet completed a 20-minute urban flow forecast in just 2.58 seconds, a task that typically requires approximately 1 hour and 25 minutes (5,120 seconds) using traditional LES. This represents a 1,984-fold speedup, proving that real-time forecasting is possible on a consumer-grade graphic card (RTX 4090) without the need for expensive supercomputers. Despite this rapid acceleration, the model faithfully reproduces critical physical phenomena, such as leading-edge updrafts and inter-building recirculation, with LES-level fidelity. This achievement provides essential real-time wind information for UAM trajectory planning and safety guidelines, while also offering a technical roadmap for pollutant dispersion prevention and smart city weather forecasting systems. By significantly lowering the barrier to high-precision urban turbulence forecasting through ML, this technology is expected to enhance the feasibility of next-generation future mobility industries.
This study was published in April 2026 in 'Physics of Fluids', a top-tier international journal in the field of fluid mechanics (ranked within the Top 5% and Q1 in the JCR Physics, Fluids & Plasma category).
DOI: https://doi.org/10.1063/5.0324405
| Attach File |
|---|