Both modes forecast the same rotating Earth's temperature-anomaly field on an icosahedral grid β the same multi-mesh representation Google DeepMind's GraphCast uses instead of a flat lat/lon grid, because it has near-uniform node spacing everywhere, including the poles.
GNN message passing (one AI forecast step):
h_i(t+1) = Ο( h_i(t), Ξ£_jβN(i) Ο(h_i(t), h_j(t), e_ij) )
β Ο, Ο are learned functions; the sum runs over each node's
mesh neighbours N(i). Stacking these layers lets information
cross the whole globe in a handful of learned steps.
NeuralGCM-style hybrid core:
X(t+1) = P(X(t)) + N_ΞΈ(X(t))
β P = differentiable physics (conservation laws), N_ΞΈ = a small
learned correction network trained to fix what P gets wrong.
Downscaling (this mesh): grid spacing Ξx β Rβ / 2^L
β each extra icosahedron subdivision level L halves the spacing,
quadrupling node count (CoarseβMediumβFine below).
- Forecast model β AI runs one learned pass over the whole mesh per step (near-instant); Physics integrates the primitive equations explicitly, so its displayed day visibly trails the target day.
- Horizon β how many simulated days ahead the run targets; error compounds with lead time in both modes, but the AI curve grows more slowly, matching GraphCast's published skill scores.
- Mesh resolution β Coarse/Medium/Fine subdivide the icosphere (42 β 162 β 642 nodes here); real GraphCast runs at 0.25Β° (~25 km, ~1M grid points).
Real-world anchor: DeepMind reports GraphCast produces a 10-day global forecast in under a minute on a single TPU v4, versus roughly an hour of supercomputer time for the equivalent ECMWF HRES physics run β and it was more accurate on the large majority of the 1,380 verification targets tested. The compute-time numbers on this page are scaled to match that ratio; the error curve is illustrative, not a reproduction of the published scores.