Optimizing CorrDiff for High-Resolution Downscaling: Sensitivity to Training Design and Multivariate Conditioning over Complex Terrain in Taiwan
High-resolution atmospheric downscaling is essential for regional climate applications, yet the systematic optimization of generative AI/ML-based frameworks for operational deployment remains insufficiently characterized. This study evaluates the sensitivity and robustness of Corrective Diffusion (CorrDiff) using a 33-year paired dataset comprising ERA5 reanalysis and the 2-km Taiwan Re-Analysis Downscaling (TReAD) dataset over Taiwan’s complex terrain. Controlled experiments examine sensitivity to training data length, atmospheric predictor configuration and multivariate output strategy, and variable-dependent downscaling performance. Precipitation downscaling skill generally improves with training length, with diminishing marginal gains beyond approximately 20 years within the 5–25-year range examined. A pronounced land–sea asymmetry in improvement is identified, consistent with a terrain-locking interpretation in which complex orography produces persistent, learnable precipitation structures over land. Joint downscaling of multiple surface variables generally improves precipitation spatial skill relative to single-variable approaches. When ERA5 precipitation is available as a predictor, a surface-only multivariate configuration approaches the performance of the full three-dimensional configuration; when precipitation is removed, three-dimensional atmospheric information becomes more valuable, while adding 850-hPa specific humidity recovers most of the resulting RMSE degradation, particularly during convectively dominated seasons. CorrDiff also achieves substantially larger RMSE reductions for 2-m temperature (∼80%) and 10-m wind speed (∼60–70%) than for precipitation (∼25%), consistent with the greater spatial coherence generally attributed to these variables. These findings provide practical guidance for configuring generative downscaling systems over complex terrain under present-day climate conditions, while emphasizing that optimal predictor design depends on predictor availability and application requirements.
Authors
- Wan‐Ling Tseng (ORCID: https://orcid.org/0000-0002-6644-9965)
- Tzu‐Ting Lo
- Yunjing Chen (ORCID: https://orcid.org/0009-0003-6321-9536)
- Yun-Ting Jhu
- Jing-Shan Hong
- Ching-Teng Lee
Institutions
- National Tsing Hua University (TW)
- Central Weather Administration (TW)
Publication Details
- Journal
- Artificial Intelligence for the Earth Systems
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1175/aies-d-26-0056.1
- Primary Topic
- Meteorological Phenomena and Simulations
- Type
- article
- Field-Weighted Citation Impact
- 0.00