Learning Kilometer-Scale Weather Prediction with Global-Regional Alignment
ScaleCast reuses pretrained global weather models to drive high-resolution regional forecasts without retraining for each driver.
The paper says its global-regional alignment modules map global and regional representations onto regional locations, then mix that guidance with local neighborhood dynamics. In tests on ERA5 and CERRA, the model improved surface and upper-air regional forecasts. A single trained ScaleCast model worked with Pangu-Weather, GraphCast, and HRES as global inputs. Fine-tuning on HRRR at 3 km showed transfer to another regional setup, with case studies reporting better cyclone position, core pressure, temperature, and humidity agreement. ArXiv · AI/CL/LG's note
The paper says its global-regional alignment modules map global and regional representations onto regional locations, then mix that guidance with local neighborhood dynamics. In tests on ERA5 and CERRA, the model improved surface and upper-air regional forecasts. A single trained ScaleCast model worked with Pangu-Weather, GraphCast, and HRES as global inputs. Fine-tuning on HRRR at 3 km showed transfer to another regional setup, with case studies reporting better cyclone position, core pressure, temperature, and humidity agreement. ArXiv · AI/CL/LG's note
score 4