HRRRCast: a data-driven emulator for regional weather forecasting at convection allowing scales

The High-Resolution Rapid Refresh (HRRR) model is a convection-allowing model used in operational weather forecasting across the contiguous United States (CONUS). To provide a computationally efficient alternative, we introduce HRRRCast, a data-driven emulator built with advanced machine learning techniques. HRRRCast includes two architectures: a ResNet-based model (ResHRRR) and a Graph Neural Network-based model (GraphHRRR). ResHRRR utilizes convolutional neural networks enhanced with squeeze-and-excitation blocks and Feature-wise Linear Modulation, and supports probabilistic forecasting via the Denois-ing Diffusion Implicit Model (DDIM). To better handle longer lead times, we train a single model to predict multiple lead times (1h, 3h, and 6h), and then use a greedy rollout strategy during inference. When evaluated on composite reflectivity over the full CONUS domain using ensembles of 3 to 10 members, ResHRRR outperforms HRRR forecast at light rainfall threshold (20 dBZ) and achieves competitive performance at moderate thresholds (30 dBZ). Our work advances the pioneering StormCast model described in Pathak et al. [21] by: a) training on the full CONUS domain, b) training on multiple lead times to improve long-range performance, c) using analysis data for training instead of the +1h post-analysis data inadvertently used in StormCast, and d) incorporating future Global Forecast System (GFS) weather states as inputs, adding a downscaling component that significantly improves long-lead forecast accuracy. Grid-based, neighborhood-based, and object-based verification metrics confirm improved storm placement, lower frequency bias, and enhanced success ratios compared to HRRR. Additionally, HRRRCast's ensemble forecasts maintain sharper spatial detail and reduced blurriness than deterministic baselines, with power spectra more closely matching HRRR analysis. While GraphHRRR underperforms in its current form, it lays the groundwork for future probabilistic graph-based forecasting. Overall, HRRRCast represents a step toward efficient, data-driven regional weather prediction with competitive accuracy and ensemble capability. Introduction Recent advances in machine learning weather prediction (MLWP) have shown great promise in complementing or even replacing traditional numerical weather prediction (NWP) systems, particularly at global scales. Several studies have demonstrated that data-driven models can rival the skill of physics-based models at a fraction of the computational cost, enabling applications such as ensemble forecasting and climate downscaling with greater efficiency [2, 12, 13, 23, 18, 17]. However, while progress in global MLWP is substantial, the transition to high-resolution regional forecasting-especially at convection-allowing scales (km-scale) - remains an active area of research. These authors have made equal contributions.