| Intended Task/Domain: |
Kilometer-scale global atmosphere emulation (weather and climate research) |
| Model Type: |
Transformer with neighborhood attention (two-stage "PixelDiT": semantic backbone + pixel refinement) |
| Intended Users: |
Weather and climate ML researchers / developers building or evaluating km-scale global atmosphere emulators |
| Output: |
148 channels on the SCREAM ne1024pg2 cubed-sphere grid (6 × 2048 × 2048 columns, ~4.9 km): potential temperature, zonal/meridional wind, geopotential height, vertical pressure velocity and specific humidity on 24 model levels, 2-m temperature, surface pressure and liquid/ice surface precipitation fluxes |
| Describe how the model works: |
STRATA autoregressively predicts the SCREAM atmospheric state 10 minutes ahead. The current state (145 channels) plus four forcing channels (solar zenith angle, surface geopotential, sub-grid orography, land fraction) are normalized and processed tile by tile on the cubed-sphere; a semantic transformer on 4×4-pixel patches captures mesoscale structure and a pixel-resolution transformer refines the output at native resolution. The network predicts a mixed state/tendency target which is de-normalized and added to the input state; longer forecasts are produced by feeding predictions back as inputs. |
| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: |
Not Applicable |
| Technical Limitations & Mitigation: |
The model emulates one specific simulation (SCREAM v1) rather than observations, so it inherits SCREAM's biases and is not calibrated against real-world data. The model was trained on only 17 simulated days from three SCREAM runs, so it has seen a limited range of seasons and climate states. Forecast skill degrades with lead time (multi-day) and small-scale features such as individual convective cells decorrelate within hours; tile-based inference can introduce faint tile-boundary artifacts that are mitigated with overlapping halos and sub-patch jitter. Users should validate on their own use case. |
| Verified to have met prescribed NVIDIA quality standards: |
Yes |
| Performance Metrics: |
Area-weighted Root Mean Square Error (RMSE) and bias per variable and level against SCREAM, radially averaged power spectra, precipitation statistics |
| Potential Known Risks: |
The model may mispredict the location, intensity or timing of severe weather (e.g. convective systems, tropical cyclones, heavy precipitation). It is a research model and should not be used as the sole basis for safety-critical decisions. |
| Licensing: |
Linux Foundation OpenMDW License Agreement, version 1.1 |
