Ai2 and Google team up on NeuralGCM+: one open climate model for atmosphere and ocean, aimed at regional rainfall decades out

Ai2 is launching a climate modeling initiative to develop NeuralGCM+, an open-source model that extends simulation from the atmosphere to a coupled atmosphere, ocean and sea-ice system, aimed at climate trends over years to decades, such as changes in rainfall and water availability in the southwestern United States through mid-century. Google Research's original NeuralGCM simulated the atmosphere alone, using physical equations for large-scale dynamics and AI for small-scale processes such as cloud formation and rainfall; Google Research has since developed it for nearer-term forecasts over weeks to months, while Ai2 turns to decadal climate trends. In a public benchmark, the atmosphere-only version of Google Research's NeuralGCM simulated a year of atmospheric conditions for the entire globe in eight minutes on a single TPU, indicating the speed such hybrid models can reach.

In NeuralGCM+, the AI component handles only local processes within vertical slices of the atmosphere, rather than learning how the atmospheric state evolves as a whole as Ai2's ACE models do. In one study, ACE2 completed experiments on how ocean-temperature patterns affect climate roughly 100 times faster than the conventional atmospheric model it was trained to emulate, but ACE models trained on conventional climate-model simulations inherit those models' limitations, and versions trained only on historical data struggle to project future climates not yet experienced. NeuralGCM's hybrid design gives the AI a simpler learning problem and may help it generalize as the climate warms; the plan proceeds on three fronts: coupling the ocean and sea ice, adding more established physics and testing how well the model reproduces temperature and rainfall changes in warmer climates it did not encounter during training, and releasing models and frameworks with documented workflows while improving performance on widely used GPUs.

This news describes an initiative Ai2 is launching with three-year funding from Google.org; NeuralGCM+ is still in development, and no simulation results for that model are given. The demonstrated capability comes from the atmosphere-only version of Google Research's NeuralGCM on a public benchmark and from Ai2's ACE2 running roughly 100 times faster than the conventional atmospheric model it emulates in experiments on how ocean-temperature patterns affect climate.

The text states that demonstrating reliable future projections remains a central research task; models trained only on historical data struggle to make reliable projections of future climates not yet experienced, and ACE models trained on conventional climate-model simulations inherit those models' limitations. Whether coupling the ocean and sea ice yields stable simulations spanning decades, and how well the model reproduces temperature and rainfall changes in warmer climates it did not encounter during training, remain goals to be tested.