NeuralGCM

Published in Nature on July 22, 2024 (DOI 10.1038/s41586-024-07744-y), NeuralGCM is Google Research's hybrid general circulation model that fuses a traditional physics-based dynamical core with machine-learning parameterizations, trained end-to-end and fully differentiable. Developed with ECMWF, it produces 1- to 15-day forecasts competitive with the best conventional models and realistic multi-decadal climate simulations, while running 70,000 simulated days in 24 hours on a single TPU versus 19 days on nearly 14,000 CPU cores.

NeuralGCM is a research model without public weights or a hosted API, and its climate projections inherit the usual caveats about training on historical data in a changing climate. Its significance is architectural: it demonstrated that end-to-end differentiable hybrid modeling can match both weather and climate benchmarks, pointing toward cheaper, more accessible Earth-system simulation rather than replacing operational forecast centers outright.