The Hybrid Future

The response to GraphCast wasn't rejection — it was adoption. The ECMWF built AIFS, its own learned model in the GraphCast mold, and in 2025 began running it operationally alongside its physics model — the first open ML forecast from a major center.

Deeper still is NeuralGCM (2024, Nature): keep a classical dynamical core on a real grid, but let neural networks learn the small-scale physics corrections, then train the whole hybrid end-to-end. It matched the ECMWF ensemble on medium-range benchmarks while running multi-decade, climate-timescale simulations orders of magnitude faster. Physics provides the skeleton; learning fills in what the grid can't see.