Regional Down-Scaling with Generative Diffusion Models

Data-driven time steppers—trained to emulate the time steps of a global analysis like ERA5—have achieved state-of-the-art performance for global weather forecasting. However, higher resolution is needed to predict weather and climate changes on local scales relevant to us. A common setup used for these down-scaling tasks is to nest a high-resolution regional model within a coarse-resolution global forecast or analysis. However, such regional models are still expensive, so we have trained large generative diffusion models instead with 3 years of regional analysis data provided by the Central Weather Bureau of Taiwan. Specifically, the diffusion models generate 2km atmospheric fields with only snapshots of 0.25 deg ERA5 data as input. We demonstrate the fidelity of the generated fields through systematic probabilistic scoring as well as case studies with strong mesoscale (e.g. tropical cyclones) and synoptic-scale (e.g. frontal systems) organization. The diffusion models are computationally cheaper than physics-based models and require less information about the atmospheric state than data-driven time steppers. High-resolution forecasts can be made by downscaling forecast outputs from traditional or data-driven global forecasts. This proves the concept that a machine learnt statistical down-scaling approach can produce atmospheric states with realistic variability.