Fine-scale downscaling & bias correction (Phase 175.7).
| Module | Phase | Contents |
|---|---|---|
corrdiff |
175.7.1.1 | CorrDiff-class regression UNet + EDM residual-diffusion super-resolution (25 km → 2 km). |
station_bias |
175.7.1.2 | Per-station bias-correction MLP consuming the 25 km ensemble (CPU-trainable per station); fail-loud registry. |
quantile_mapping |
175.7.1.3 | Empirical + multivariate (Schaake-shuffle) quantile mapping and derived variables — wind-power density (½ρV³), HDD/CDD, GDD. |
urban |
175.7.1.4 | Diurnal UHI-aware urban downscaling (~50 m) over Urban-Atlas/OSM imperviousness + vegetation. |
Task 175.7.1.1 ("train the CorrDiff model, validated on HRRR/CERRA") needs paired regional reanalyses and is not locally actionable; the architecture is implemented and its regression stage is overfit-tested. The station-bias task (175.7.1.2) is local — the per-station MLP trains on CPU and its overfit test recovers a synthetic warm-bias + diurnal signal.
bash
cd libs/gaia/downscale && python3 -m pytest tests/ -q