Climate science has a speed problem. The complex, physics-based models that predict our planet’s future are foundational to understanding climate change, but they are agonizingly slow. Running a single century-long projection can tie up a supercomputer for weeks. This computational bottleneck limits how many future scenarios scientists can explore.
Researchers at the Allen Institute for AI (Ai2) and its partners just dropped a potential solution: an AI climate emulator called SamudrACE that can simulate 1,500 years of global climate in a single day on one NVIDIA H100 GPU. According to a paper published on arXiv, this represents a staggering 3,750-fold reduction in energy usage compared to the traditional model it was trained to mimic.
The real breakthrough isn’t just speed, but complexity. SamudrACE is the first AI model to successfully couple full 3D simulations of both the ocean and the atmosphere. This is a critical step because the constant exchange of energy between these two systems drives major climate phenomena like the El Niño-Southern Oscillation (ENSO), the Pacific warming and cooling pattern that influences everything from California droughts to Australian bushfires.
Previous AI emulators could model the atmosphere or the ocean separately, but they struggled to capture the complex behaviors that emerge when the two interact. By linking two state-of-the-art emulators—ACE2 for the atmosphere and Samudra for the ocean—SamudrACE creates a stable feedback loop. The result is a model that can realistically simulate the powerful, globe-spanning influence of El Niño, a feat that uncoupled AI models simply can't match.
A new era for climate 'what-ifs'
This leap in efficiency could fundamentally change climate research. Instead of painstakingly running a handful of simulations, scientists using an AI climate emulator like SamudrACE can run massive ensembles of thousands of projections. This allows for a much deeper understanding of uncertainty and the full range of possible climate outcomes.
Researchers can now rapidly test “what-if” scenarios that were previously too computationally expensive. Questions like, “How would a major volcanic eruption affect global temperatures over the next decade?” or “What are the odds of a string of extreme El Niño events?” can be explored in days, not years.
SamudrACE is still a proof-of-concept. The current version was trained only on pre-industrial climate data, so it isn't ready to project our high-carbon future just yet. The team’s next step is to retrain the model on simulations with much higher CO2 levels. If successful, it could provide science with a powerful, high-speed crystal ball for navigating the complexities of a warming planet.
