StormCast¶
NANWC202440 GBNVIDIAPyTorch
StormCast is NVIDIA's generative convection-allowing model. It forecasts 99 HRRR variables on a 3 km window of the HRRR grid over the central United States with a 1-hour step, conditioned on coarse global fields. A regression network makes a first guess and a diffusion network corrects it, so an ensemble comes from the sampler's noise rather than from perturbed initial conditions.
Skill¶
Pick a metric and variable; hover for exact values at each lead time. Use the Month selector for seasonal (DJF/MAM/JJA/SON) and per-month skill against the all-month curve, the Init hour selector for skill by initialization time, the Event selector for skill during named weather events, the View selector for the skill of every initial condition, and the Baseline selector to overlay persistence and climatology reference forecasts.
Evaluation¶
8-member ensemble · 20 initial conditions · 12-hour horizon · 5 variables
Scores are uniformly weighted on the model grid and aggregated over the initial conditions. Evaluation is done against HRRR. Initial conditions rotate through the 00Z/03Z/06Z/09Z/12Z/15Z/18Z/21Z hours.
| Type | 8-member ensemble |
| Initial conditions | 20 (2025) |
| Initial condition source | HRRR |
| Verification (ground truth) | HRRR |
| Lead times | 1 h to 12 hours |
| Variables scored | 5 |
| Metrics | RMSE, MAE, RMSE (ensemble mean), CRPS, Spread, Spread / Skill |
| Events | 14 March outbreak (Missouri, Arkansas) (2025-03-14 to 2025-03-15), 2 April outbreak (Mid-South) (2025-04-02 to 2025-04-03) |
Variables¶
Scored output variables (5)
| Name | Description | Unit | Group |
|---|---|---|---|
mslp |
Mean sea level pressure | Pa | Surface |
refc |
Maximum/Composite radar reflectivity | dBZ | Surface |
t2m |
Temperature at 2m | K | Surface |
u10m |
U-component (eastward, zonal) of wind at 10 m | m s⁻¹ | Surface |
v10m |
V-component (northward, meridional) of wind at 10 m | m s⁻¹ | Surface |
All of the model's output variables that have ERA5 verification are scored.
Data¶
The numbers behind the plot are in eval_scores_stormcast.json, exported by
the eval recipe scorecard
(scorecard/export_scores.py --docs) -- one value per metric, variable
and lead time, in the variable's own units.
Reproducibility¶
Run and environment details
| Date scored | 2026-09-14 |
| Scores written | 2026-09-14 |
| GPUs | 8 x NVIDIA H100 80GB HBM3 (single node) |
| PyTorch | 2.12.0a0+0291f960b6.nv26.04.48445190 |
| CUDA | 13.2 |
| Python | 3.12.3 |
| Repo commit | 6a641cf5184b |
| Provenance source | run |
| Exported | 2026-09-14 |
| Locked dependencies | uv.lock @ 6a641cf5184b |
Event-based scoring¶
Unlike the global scorecards, this page comes from an event campaign. Each event is a space-time window over the evaluation data showcasing a physical event. Initial conditions run every 3 hours from 12 hours before the window to its end. The headline curves pool every event's initial conditions, and the Event selector shows one episode at a time. Truth is the HRRR analysis, and scores are uniformly weighted on the model grid. The campaign definition is stormcast_2025_events.yaml and the event scoring lives in the evaluation recipe.
Reference¶
Pathak, J., Cohen, Y., Garg, P., Harrington, P., Brenowitz, N., Durran, D., Mardani, M., Vahdat, A., Xu, S., Kashinath, K., and Pritchard, M. (2026). Kilometer-scale convection-allowing model emulation using generative diffusion modeling. Science Advances, 12(5), eadv0423.