[A-AS05] Weather, Climate, and Environmental Science Studies using High-Performance Computing

[AAS05-P08] Preliminary Study on Constructing Grid-Based Meteorological Data for Taiwan Using Machine Learning Models

*Yu-Chi Wang¹, Mei-Ling Tang¹, Yi-Sheng Wang¹, Jay Chen², Ming-Long Lee¹ (1.National Center for High-performance Computing, National Applied Research Laboratories. , 2.NVIDIA)

Keywords:Machine Learning, Gridded Observational Data, CorrDiff Model, Downscaling

Taiwan's complex topography presents challenges for establishing meteorological observation stations in mountainous regions, resulting in discontinuities in spatial coverage. The " Taiwan Climate Change Projection Information Platform" project provides long-term, high-resolution gridded daily meteorological data at a 1-km resolution, bridging the gaps in station-based observations. However, since these datasets rely on in-situ measurements, they cannot be generated in real-time.
In this study, we employ NVIDIA’s recently released generative AI model, CorrDiff, to downscale the 25-km global grid data from the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5 reanalysis to a 1-km resolution for Taiwan. This machine learning approach enables the generation of high-resolution gridded meteorological data even in the absence of complete station coverage. Given the substantial computational resources required for model training, we utilize eight NVIDIA V100 Tensor Core GPUs to achieve the necessary performance. Our model is trained on 15 years of global 25-km grid data, and its outputs are validated against gridded observational datasets to assess accuracy and reliability.