HiResCastNet: A 3D Deep Learning Model for Probabilistic Regional High‐Resolution Subseasonal Forecasting Through Downscaling

Abstract

Abstract High‐resolution subseasonal forecasts, spanning two to 6 weeks, play a vital role in bridging the gap between short‐range forecasts and seasonal predictions. They provide crucial information for decision‐making in disaster and hazard preparedness, early warning systems, water resource management, and renewable energy. However, generating subseasonal forecasts at high spatial and temporal resolutions remains a significant challenge. Global subseasonal forecasting systems such as European Center for Medium‐Range Weather Forecasts (ECMWF) typically provide outputs at a coarse spatial resolution of about 31 km and a temporal resolution of 12 hr, which limits their usefulness for regional‐scale planning and impact assessments. Our three‐dimensional (3D) high‐resolution regional subseasonal forecasts are generated by dynamically downscaling the ECMWF subseasonal forecasts using the Weather Research and Forecasting (WRF) model, producing outputs at a spatial resolution of 4 km and a temporal resolution of 3 hr. Yet, the dynamical downscaling model is computationally very demanding, especially when using multiple ensemble members. To address this limitation, we introduce HiResCastNet, a data‐driven, probabilistic downscaling framework combining Swin‐transformer and a diffusion model, which serves as an efficient surrogate for the traditional downscaling system. HiResCastNet takes spatio‐temporal coarse resolution forecasts from ECMWF subseasonal system as inputs, and generates 3D high‐resolution forecasts at 4 km spatial resolution and 3‐hourly intervals. Our results suggest that HiResCastNet significantly reduces computational cost while maintaining strong agreement with WRF physical model. It reproduces the dynamical behavior across variables such as wind and temperature and handles the downscaling of multiple ensemble members efficiently, which highlights its potential for operational use.

The authors' abstract, as published at the source. Journal of Geophysical Research Machine Learning and Computation, 2026 · DOI ↗