GraphCast-class deterministic spherical-GNN medium-range forecasting (Phase 175.2). A real, trainable PyTorch implementation of the encoder→processor→decoder architecture on an icosahedral multi-mesh.
| Module | Phase | Contents |
|---|---|---|
mesh |
175.2.1.1 | Icosahedral multi-mesh generator (M0→M6, 10·4^L+2 nodes), per-level edge sets, multi-mesh edge union, exact spherical areas. |
grid |
175.2.1.2/.4 | Grid↔mesh geometry: radius-based grid→mesh bipartite edges and mesh-triangle containment for mesh→grid. |
layers |
— | MLP + interaction-network message passer (edge update → scatter-add aggregation → residual node update). |
encoder |
175.2.1.2 | Grid→Mesh encoder. |
processor |
175.2.1.3 | 16-round bidirectional multi-mesh processor. |
decoder |
175.2.1.4 | Mesh→Grid residual-tendency decoder (adds learned Δ to input state). |
model |
175.2.1.5 | Assembly + autoregressive rollout (up to 40 × 6 h) with gradient checkpointing for multi-step fine-tuning. |
loss |
175.2.2.4 | cos(lat) area weighting + per-variable + per-level weights; linearly-decayed multi-step rollout loss. |
diagnostics |
175.2.2.5 | Energy-conservation drift, zonal power spectrum, blur detector, and automatic early-stop on spectral collapse. |
ensemble |
175.2.2.3 | Reproducible 10-seed/leave-one-fold-out training plans and checkpoint-compatible member aggregation with mean/spread products. |
Training tasks 175.2.2.1–2 (300k-step runs on TPU-v5e/H100 pods reproducing the published 0.25° skill) require accelerator clusters and the full ERA5 archive and are not locally actionable; the architecture, loss, diagnostics, and 10-member ensemble baseline they would use are implemented and unit-tested here — including an overfit-a-batch test that proves gradients flow end-to-end.
bash
cd libs/gaia/graphcast && python3 -m pytest tests/ -q