SamudrACE: Highly efficient coupled global climate modeling with the Ai2 climate emulator

What if we could explore countless climate futures in the time it currently takes to simulate one?

Understanding and predicting climate change is a critical scientific challenge of our time. For decades, scientists have relied on complex, physics-based global climate models (GCMs)— sophisticated computer simulations that serve as our crystal ball for Earth’s ever-changing conditions. These models are foundational to climate science, but they have a significant limitation: they’re incredibly slow. Running a single 100-year projection can take weeks on a supercomputer, and researchers need to run hundreds to fully explore the range of possible outcomes.

At Ai2, we're working to overcome this barrier by building AI-powered emulators—systems that learn to mimic GCMs but run orders of magnitude faster.

Our latest breakthrough is SamudrACE, a fast and accurate AI emulator that – for the first time – couples 3D models of both the ocean and the atmosphere, giving it a deep understanding of global Earth system interactions. Developed in partnership with our colleagues at NYU, Princeton, M2LInES, and NOAA’s Geophysical Fluid Dynamics Laboratory (GFDL), SamudrACE can simulate 1,500 years of global climate in a single day running on an NVIDIA H100 GPU.

The coupling challenge

The Earth's climate is an interconnected system, dominated by the exchange of energy between the atmosphere and the ocean. This interplay gives rise to critical climate phenomena like the El Niño-Southern Oscillation (ENSO), a natural pattern of warming and cooling in the Pacific Ocean that influences weather around the world. ENSO affects everything from California droughts to Australian bushfires, making it a key factor in seasonal prediction.

While previous AI emulators have shown promise in speeding up atmospheric or oceanic simulations individually, they’ve struggled to capture the emergent behavior that arises only when the two systems are realistically coupled. Think of it like a weather app that only shows air temperature—useful, but missing crucial context. Without this coupling, an emulator can't fully represent the dynamics of our climate. 

SamudrACE’s breakthrough is its success in creating a stable, realistic, and fully coupled atmosphere-ocean system.

Multiple components, one system

SamudrACE is not a single monolithic model. Instead, it follows the same successful paradigm as traditional GCMs by using separate, specialized components that communicate through a "coupler." We built it by linking two existing state-of-the-art emulators:

• ACE2, an emulator for the atmosphere and land surface, developed by our Climate Modeling team
• Samudra, an emulator for the ocean developed by collaborators at M2LInES and extended by the Climate Modeling team to prognose sea ice

In our framework, ACE2 r simulates the atmosphere in 6-hour increments. Relevant information, like the surface fluxes of heat and moisture, is averaged over a 5-day period. This averaged data is then passed to Samudra, which uses it – along with the current state of the ocean – as input to predict a single corresponding 5-day step. 

The system feeds the resulting ocean state, including sea surface temperature and sea ice cover, back into ACE2—-creating a continuous and stable feedback loop. This physics-informed coupling strategy allows SamudrACE to function as a fully integrated climate model emulator.

The results: speed, stability, and realism

Running on a single NVIDIA H100 GPU, SamudrACE can simulate 1,500 years of global climate in a single day. This represents a 3,750-fold reduction in energy usage compared to the traditional GFDL Climate Model v4 (CM4) simulation it was trained on, which requires thousands of CPU cores to achieve a much slower simulation rate of about 16 years per day.

This speed doesn’t come at the cost of accuracy or stability. SamudrACE can generate centuries-long simulations with low time-average climate biases comparable to its uncoupled components and the original GCM.

Most importantly, SamudrACE successfully simulates the complex, emergent patterns of a coupled climate system. Our research shows that it generates more realistic ENSO variability than a prior AI model (Figure 1). It also accurately captures ENSO’s influence on global precipitation patterns. Finally, it produces a stable – and highly accurate – seasonal cycle of sea ice extent in both the Arctic and Antarctic.

Why this matters for climate science—and what’s next

The ability to run fast, accurate, and coupled climate models unlocks new possibilities for researchers. 

With models like SamudrACE, scientists can run large ensembles of simulations to better quantify uncertainty and understand the full range of potential climate outcomes. It allows for the rapid exploration of different climate scenarios and "what-if" questions (e.g., “How would a major volcanic eruption affect global temperatures over the next decade?,” “What are the chances of multiple extreme El Niño events occurring within a short timeframe?”) that would be computationally prohibitive with traditional models. 

SamudrACE serves as a powerful proof-of-concept, demonstrating that data-driven models can emulate the most complex aspects of our climate system.

To date, we have only trained SamudrACE on pre-industrial climate conditions. We do not expect this version of SamudrACE to generalize well to future climate states. To get a model that works well across a wider range of climates, we plan to train SamudrACE on CM4 simulations with carbon dioxide values up to quadruple the pre-industrial value. This new dataset will no doubt provide more exciting challenges and opportunities for further advancement of AI-driven coupled climate simulations.

To learn more, see the SamudrACE paper here, and explore the repo on Hugging Face.