Enhancing Regional Machine Learning Weather Prediction Using Tailored Loss Functions and High-Resolution RRJ-ClimCORE Reanalysis

Recent advances in machine learning-based weather prediction (MLWP) have achieved accuracy comparable to operational numerical weather prediction systems, while offering much faster and more energy-efficient inference. However, global MLWP models struggle to forecast localized extremes, including typhoons and heavy rainfall. This is mainly due to coarse resolution of training data and the reliance on mean squared error (MSE) loss, which inclines toward spatial smoothing. This study develops a regional MLWP model for Japan trained on high-resolution regional reanalysis data with the employment of variable-specific loss functions for wind, pressure, geopotential height, and precipitation. Ablation experiments demonstrate that these loss functions contribute in complementary ways, substantially improving the forecast skill relative to conventional MSE loss. Case studies of an extratropical cyclone and Typhoon Nanmadol show improved forecasts of cyclone intensity, strong winds, and orographic rainfall. Leveraging regional reanalysis with the tailored loss design is therefore an effective strategy for regional MLWP.