| Intended Task/Domain: |
High-resolution regional weather forecasting over the Continental United States (CONUS) |
| Model Type: |
Diffusion Transformer (DiT) |
| Intended Users: |
Weather and climate ML-based researchers / developers building or evaluating regional km-scale weather forecasting pipelines. |
| Output: |
99 surface and model-level variables on the HRRR Lambert Conformal Conic grid (3 km, 1024×1792): near-surface winds (u10m, v10m), 2 m temperature (t2m), mean sea-level pressure (mslp), hybrid model-level winds (u_hl*, v_hl*), temperature (t_hl*), specific humidity (q_hl*), geopotential (z_hl*), pressure (p_hl*) at levels 1–11, 13, 15, 20, 25, 30, and composite reflectivity (refc). |
| Describe how the model works: |
StormCast-CONUS autoregressively predicts 99 atmospheric state variables at 3 km resolution and 1-hour time steps. At each step, the DiT backbone takes 99 state variables, 26 ERA5 large-scale conditioning variables (pressure-level winds, temperature, geopotential, and specific humidity), static inputs (surface elevation statistics and land/water mask), and the cosine of the solar zenith angle. Neighborhood attention (NATTEN) replaces global self-attention to make computation tractable at 1024×1792 resolution. The model generates a denoised prediction of the next atmospheric state, which is fed back as input for the following step. |
| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: |
Not Applicable |
| Technical Limitations & Mitigation: |
Forecasts are limited to the CONUS domain and may degrade for weather phenomena outside the 2018–2025 training period or for rare, extreme events underrepresented in HRRR analyses. Autoregressive rollout accumulates errors over longer forecast horizons. The model does not enforce physical conservation laws. Ensemble generation (multiple stochastic draws) is recommended to characterize forecast uncertainty and partially mitigate single-member error. |
| Verified to have met prescribed NVIDIA quality standards: |
Yes |
| Performance Metrics: |
Root Mean Square Error (RMSE), Continuous Ranked Probability Score (CRPS), Anomaly Correlation Coefficient (ACC), Equitable Threat Score (ETS) for composite reflectivity |
| Potential Known Risks: |
The model may incorrectly predict the location, intensity, or timing of severe weather events (e.g., convective systems, heavy precipitation). Forecast quality degrades at longer lead times. Outputs should not be used as the sole basis for operational safety-critical decisions without integration into a validated forecasting system. |
| Licensing: |
Linux Foundation OpenMDW License Agreement, version 1.1 |
