Limitation of super-resolution machine learning approach to precipitation downscaling.

The present study explores the potential of super-resolution machine learning (ML) models for precipitation downscaling from 100 to 12.5 km at hourly timescale using the Conformal Cubic Atmospheric Model (CCAM) data over the Australian domain. Two approaches were examined: the perfect approach, which trains the ML model using coarsened high-resolution data as input (i.e., CCAM 12.5 km data coarsened to 100 km), and the imperfect approach, which uses original coarse-resolution data as input (i.e., CCAM model simulation at 100 km resolution) and in both the cases high-resolution data (i.e., CCAM 12.5 km simulation) is used as target. In the perfect case, the ML model (ML) accurately reproduces high-resolution climatology and extremes. However, the ML model with CCAM 100 km simulation data as input (i.e., in the imperfect setting) underestimates the magnitude of the output and introduces spatial inconsistencies, while the ML model captures high-resolution structures but underestimates extremes. This suggests that the super-resolution ML model approach is inappropriate for precipitation downscaling because of the spatial inconsistencies between the coarse and high-resolution simulations. Additionally, we introduced sensitivity-based diagnostics beyond standard evaluation methods to understand model behaviour and identify structural issues. These diagnostics reveal that both models increase precipitation inputs non-linearly without creating spurious spatial relationships. However, the ML model outputs precipitation in high-altitude regions regardless of input, highlighting the structural issue of the ML model. Our study highlights the challenges in using super-resolution ML models for precipitation downscaling, introduces several useful diagnostics for assessing the super-resolution ML models and their physical realism, and provides ideas to explore to improve ML-based precipitation downscaling.