Ai2 grabs $5M from Google to stretch weather AI decades into the future

Google is handing Seattle's Allen Institute for AI roughly $5 million, plus another $1.8 million in cloud credits, to take a weather model built by Google Research and stretch it decades into the future. The three-year project, announced Wednesday morning, pairs cash from Google.org — Google's philanthropic arm — with technical muscle from Google Research.

The payoff Ai2 is chasing: an open-source model that gives scientists, water agencies and planners a faster, sharper read on how rainfall, temperatures and water supplies could shift over the next five to 50 years. Four new hires will join Ai2's existing eight-person climate modeling team to build it.

The new model, NeuralGCM+, builds on Google's NeuralGCM, which blends traditional physics equations with AI. Oliver Watt-Meyer, Ai2's director of research for AI climate modeling, reviewed NeuralGCM for Nature when Google published the research in 2024 and flagged two weak spots: some long simulation runs broke down, and when researchers simulated oceans about 7 degrees Fahrenheit warmer than today, the atmosphere predictions stopped making physical sense.

Ai2's plan attacks both. It will wire NeuralGCM+ into an ocean and sea ice model to chase stable, decades-long simulations, and layer in more established physics — then test its warmer-world predictions against conventional climate models. "We don't know if it'll work necessarily," Watt-Meyer said, calling the warmer-world reliability the hardest piece, one that could outlast the three-year timeline.

The work circles back to a project started around 2016 by late Microsoft co-founder Paul Allen at Vulcan, his Seattle holding company. The group landed at Ai2 in 2021, about two and a half years after Allen's death; he had founded the nonprofit institute in 2014. The original pitch was physics plus AI, shelved because the AI couldn't be trained well inside climate models written in the decades-old Fortran language — a problem Google solved by writing NeuralGCM in modern code.

Ai2 already runs a fully AI model family, the Ai2 Climate Emulator (ACE), which is much faster but struggles to project conditions the planet hasn't experienced; insurers and reinsurers use it in research. First up is releasing NeuralGCM's training code, currently tied to Google's internal systems, within a year so outside researchers can train their own versions on ordinary GPUs. Watt-Meyer also credited AI coding agents with reshaping what his team thinks it can finish, and pointed to Southwest rainfall and water planning around Lake Powell on the Colorado River as prime use cases.