ACE-class climate emulation & long rollouts (Phase 175.6).
| Module | Phase | Contents |
|---|---|---|
fno |
175.6.1.1 | Fourier Neural Operator (spectral conv with mode truncation) backbone. |
forcing |
175.6.1.4 | IPCC/Myhre radiative-forcing expressions (CO2 log, CH4 with N2O overlap, aerosol). |
emulator |
175.6.1.1/.4 | Forcing-conditioned one-year climate step + autoregressive rollout; ScenarioEmulator for tunable-forcing what-if paths. |
feedback |
175.6.1.2 | Coupled sea-ice-albedo & soil-moisture feedbacks as FNO-in-FNO (physical prior + learned correction). |
budgets |
175.6.1.1 | Area-weighted global mean, GMST series/anomaly, per-decade drift/trend, TOA energy imbalance. |
hindcast |
175.6.1.3 | Decadal-hindcast driver + validation (anomaly correlation, RMSE, warming-trend error) vs a reference observational series. |
Task 175.6.1.1 as written ("train on CMIP6 for 100-year rollouts with drift ≤ 0.05 K/decade") needs the CMIP6 archive and long accelerator runs and is not locally actionable; its FNO architecture and the drift diagnostics that gate it are implemented and tested here. The hindcast (175.6.1.3) validation harness runs offline against a supplied reference anomaly series — the HadCRUT5/GISTEMP/AMIP records load through a data seam.
bash
cd libs/gaia/climate-emulator && python3 -m pytest tests/ -q