-
- A43R-2373: Can a Single Sounding Predict the Weather? An AI Downscaling Framework for Local-Circulation-Dominated Flow over Complex Topography
-
Board 2373‚ Hall EFG (Poster Hall)NOLA CC
Author(s):Generic 'disconnected' Message
Fucent Hsuan Wei Hsu, National Taiwan University (First Author, Presenting Author)
Min-Ken Hsieh, National Taiwan University
Chien-ming Wu, National Taiwan University
High-resolution forecasts typically rely on supercomputers that run for hours to resolve winds around mountains and coastlines, whereas fast AI models often ingest so many meteorological inputs that the truly impactful factors become obscured. We show that a handful of “gold-nugget” atmospheric clues are enough. By feeding an AI model a single weather-balloon sounding—32 elements describing the large-scale vertical wind profile—we can rebuild 2 km wind maps over Taiwan's rugged terrain in seconds, much like prompting ChatGPT with a sentence and receiving an image. The lean input lets us trace each variable's influence and reveals that low-level wind direction, wind speed, and the diurnal cycle are critical to reconstructing the wind field; it also enables a modern transformer architecture that sharpens accuracy in mountainous terrain and under unfamiliar weather patterns. Because the model runs on widely available hardware, it can deliver high-resolution wind products to students, forecasters, and emergency managers without supercomputers or huge data downloads, offering a transparent, resource-efficient path to AI downscaling. More broadly, the study shows that AI models need not remain inscrutable black boxes—they can be engineered to illuminate, rather than obscure, the physical processes that govern the atmosphere.
Scientific DisciplineSuggested ItinerariesNeighborhoodType
Enter Note
Go to previous page in this tab
Session
