GenCast-class probabilistic diffusion forecasting (Phase 175.3). A conditional EDM diffusion model over +12 h residual fields, using the Phase 175.2 spherical GNN as its score network.
| Module | Phase | Contents |
|---|---|---|
edm |
175.3.1.1 | EDM preconditioning (c_skip/out/in/noise), loss weight, log-normal training-noise, ρ=7 sampling schedule (σ_min 0.02, σ_max 88). |
denoiser |
175.3.1.1 | GNN score network (encoder→multi-mesh processor→decoder) conditioned on the previous two states + noise level, wrapped by EDM preconditioning. |
samplers |
175.3.1.2 | Deterministic DPM-Solver++(2M) (20-step) and stochastic Heun with Langevin churn. |
ensemble |
175.3.1.4 | Batched multi-seed ensemble generation with 12 h autoregressive rollout (15-day, 50-member). |
calibration |
175.3.2.1 | Fair-estimator CRPS, spread-skill ratio, Talagrand rank histogram, reliability diagram. |
extremes |
175.3.2.2 | P99.9 percentile bias, L-moments GEV fit + return levels, empirical (Gringorten) return levels. |
clustering |
175.3.2.3 | PCA (SVD) + k-means++ scenario clustering with representative members and scenario probabilities. |
Training 175.3.1.3 (1979–2018 ERA5, CRPS vs IFS-ENS on 1320 targets) needs the full archive and accelerator clusters and is not locally actionable; the model, samplers, and every verification metric it depends on are implemented and unit-tested here — including a DPM-Solver++ test that recovers the exact Gaussian marginal from the analytic optimal denoiser, and CRPS values checked by hand.
bash
cd libs/gaia/gencast && python3 -m pytest tests/ -q