Ai2 与 Google 合作开发新一代气候模型 NeuralGCM+

Ai2 is launching a new climate modeling initiative with multimillion-dollar, three-year funding from Google.org and technical support from Google Research. The funding will expand our Climate Modeling team as we develop NeuralGCM+, an open-source model designed to make climate projections faster and more reliable.

The project builds on Google Research’s NeuralGCM, which uses physical equations to simulate large scale dynamics and AI to represent smaller-scale processes such as cloud formation and rainfall. We will extend it to simulate the atmosphere and ocean together, capturing how oceans absorb and redistribute heat—a major influence on long-term climate.

Since its original release, Google Research has been developing NeuralGCM for nearer-term forecasts over weeks to months, such as how warm or wet an upcoming summer might be. Ai2’s focus will be climate trends over years to decades, including how rainfall and water availability could change through mid-century in regions such as the southwestern United States. 

Our ambition is to put these tools within reach of more scientists and help communities make better-informed decisions about water planning and climate adaptation.

“Driven by deep climate science expertise and a shared commitment to open-access research, Ai2 is uniquely positioned to extend Google’s NeuralGCM model into the next generation of long-term climate projections. We look forward to supporting this critical work, which is vital for helping global industries and communities prepare for the future.” — Alex Diaz, Head of Applied Science, Google.org

“Much of the progress in Earth system AI modeling has focused on weather forecasting, in the range of days to months. But communities also need to understand what could happen over decades as they plan for a changing climate. That longer view needs sustained research, and it’s where we see Ai2 making a contribution. We’re grateful for Google.org’s support that will allow us to push even further in this direction, adding to the work we’re already doing like the ongoing development of the ACE climate model.” — Oliver Watt-Meyer, Climate Modeling Lead, Ai2

Conventional climate models have shaped our understanding of global warming for decades, but their computational demands limit the experiments scientists can run. Faster, less expensive simulations could let researchers explore more possible futures and investigate persistent uncertainties in regional temperature and rainfall trends. Google’s existing NeuralGCM illustrates that potential—in a public benchmark, an atmosphere-only version simulated a year of atmospheric conditions for the entire globe in eight minutes on a single TPU.

We’ve already made climate simulations faster and more computationally efficient with our Ai2 Climate Emulator (ACE), a family of models that learn how atmospheric conditions change over time. 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.

Speed, however, is only part of the challenge. ACE models trained on conventional climate-model simulations inherit those models’ limitations. Other versions learn from reconstructions of past atmospheric conditions and can accurately simulate historical variability and warming trends. But models trained exclusively on historical data struggle to make reliable projections of future climates we have not yet experienced.

NeuralGCM’s hybrid design may help address this challenge. Unlike ACE, which learns how the atmospheric state evolves as a whole, NeuralGCM’s AI component focuses on local processes within vertical slices of the atmosphere. This gives the AI a simpler learning problem and may help it generalize as the climate warms. Demonstrating reliable future projections remains a central research task.

We will develop NeuralGCM+ along three connected fronts:

• Connect the atmosphere and ocean. We aim to link NeuralGCM with an ocean and sea-ice model to produce stable simulations spanning decades. A key goal is to reduce regional temperature and rainfall errors when simulating the past 50 years, compared with conventional climate models.
• Improve the response to a warming world. We will add more established physics and test how well the model reproduces temperature and rainfall changes in warmer climates it did not encounter during training, using conventional climate-model simulations as a reference.
• Make the tools easier to use and extend. Building on NeuralGCM’s existing open-source code and available model weights, we plan to release our models and modeling frameworks with documented workflows that enable outside researchers to train and adapt models and contribute code. We will also improve performance on widely used GPUs.

Our long-term ambition is a shared research platform that climate modeling labs can use and improve together, expanding the science that informs how communities prepare for a changing climate.

Learn more about Ai2’s climate modeling research and explore the existing NeuralGCM code and documentation.

Subscribe to receive monthly updates about the latest Ai2 news.