Data Scaling Boosts High-Resolution Weather Forecasting Accuracy

The development of high-resolution global weather forecasting models, particularly those based on machine learning, has been significantly hampered by the scarcity of fine-grained data. While reanalysis data exists, it's typically at a coarser 0.25° resolution, and attempts to fine-tune models on limited 0.1° samples have proven insufficient due to irreversible information loss. This research introduces BaguanHR, a novel framework that shifts focus from model transfer to data transfer, addressing this bottleneck directly. BaguanHR leverages variable-wise super-resolution (SR) to synthesize vast amounts of 0.1° data from the widely available ERA5 reanalysis. The core insight is that SR is a more robust mechanism for resolution transfer than direct forecasting, exhibiting lower conditional entropy and input amplification. By generating this extensive synthetic-plus-real dataset, BaguanHR enables ML-based models to achieve unprecedented accuracy. The framework's performance is remarkable, surpassing both current ML-based methods and the traditional IFS-HRES model for over 85% of lead times within 72 hours. Furthermore, the study identified a clear power-law scaling effect, where doubling the data reduces RMSE by approximately 4.6-4.9% for 72-hour and 120-hour forecasts, respectively. These findings strongly suggest that data availability, rather than model architecture, is the primary constraint for high-resolution ML-based weather forecasting, and super-resolution offers a general solution.