FGN-class direct tropical-cyclone prediction (Phase 175.4).
| Module | Phase | Contents |
|---|---|---|
functional_head |
175.4.1.1 | Joint cyclone functional: prob-of-cyclone, centre, Vmax, MSLP, RMW, R34/R50/R64 quadrants, eye flag — with softplus/sigmoid range constraints and R34≥R50≥R64 monotonicity. |
trajectory |
175.4.1.2 | Transformer decoder over the past 24 h track + backbone latent → Gaussian position-increment distribution, with stochastic trajectory sampling (track cone). |
losses |
175.4.1.3 | Combined loss: Haversine position, Huber Vmax/MSLP, soft-binned Saffir-Simpson cross-entropy, Brier formation; Gaussian trajectory NLL. |
atcf |
175.4.1.4 | NHC ATCF a-deck writer (BASIN/CY/time/TAU/Lat/Lon/Vmax/MSLP/dev-level + 34/50/64 kt wind-radii lines). |
integration |
175.4.1.5 | Adapters: @oya/drones no-fly cone + hunter tasking, @cybele/construction Rankine-vortex asset-hardening alerts, @veritas/news bulletin generator. |
geo |
— | NumPy + differentiable Torch Haversine distance. |
Training 175.4.1.3 to "replicate the published FGN NHC landfall-skill win over HAFS/HWRF" needs the operational archive and is not locally actionable; the combined objective it optimises is implemented and unit-tested (overfit-a-track tests prove the head and trajectory model train end-to-end).
bash
cd libs/gaia/cyclone-forecast && python3 -m pytest tests/ -q