SamudrACE speeds up climate simulation with AI

SamudrACE can simulate 1,500 years of global climate in a single day using one NVIDIA H100 GPU. Ai2's system combines AI, ocean and atmosphere models to explore climate scenarios at a speed traditional models cannot match.

The tool was developed with researchers from NYU, Princeton, M2LInES and NOAA's GFDL laboratory. Its goal is not to replace physics-based climate models, but to emulate them much faster.

The problem with current climate models

Global climate models are computer simulations that represent how the atmosphere, oceans, ice and land surface evolve. They are essential for studying climate change, but they require enormous computing power.

A 100-year projection can take weeks on a supercomputer. And one simulation is not enough: scientists need to run hundreds of versions to compare scenarios and measure uncertainty.

SamudrACE addresses this bottleneck with an AI emulator. These systems learn to reproduce the behavior of complex physics-based models, but at a much lower computational cost.

Ocean and atmosphere in the same system

SamudrACE's main innovation is that it connects three-dimensional atmosphere and ocean models. This matters because the two systems constantly influence each other: they exchange heat and moisture, alter currents and affect phenomena such as El Niño.

Earlier emulators could focus on the atmosphere or the ocean separately. The problem is that some climate patterns only appear when both interact realistically.

SamudrACE uses two specialized components:

• ACE2, which simulates the atmosphere and land surface in six-hour steps.
• Samudra, which predicts how the ocean and sea ice evolve in five-day steps.

Every five days, the information accumulated by ACE2, such as heat and moisture exchange, is passed to Samudra. The resulting ocean state, including surface temperature and ice coverage, then returns to the atmosphere. This creates a continuous cycle between the two systems.

What results has it achieved

On an NVIDIA H100, SamudrACE reaches a speed of 1,500 years of climate simulation per day. The GFDL climate model it was trained on, CM4, advances approximately 16 years per day and requires thousands of CPU cores.

According to Ai2, the comparison represents a 3,750-fold reduction in energy use compared with that traditional simulation. The speed did not eliminate the essential climate patterns the system was expected to reproduce.

SamudrACE managed to:

• Represent El Niño variability more realistically than an earlier AI model.
• Reproduce its effects on global precipitation.
• Maintain a stable and accurate seasonal cycle for sea ice in the Arctic and Antarctic.
• Run century-scale simulations with average climate errors comparable to those of its components and the original physics-based model.

What this changes for research

With a tool like this, scientists can run many more simulations and better study the range of possible futures. For example, they could analyze how a major volcanic eruption would affect temperatures over the following decade or calculate the probability of several extreme El Niño events over a short period.

That lets you compare more scenarios without reserving a supercomputer for weeks. It also makes it easier to measure uncertainty, which is crucial when making decisions about agriculture, water, infrastructure or disaster prevention.

But SamudrACE still has an important limitation. The version presented was trained only on preindustrial climate conditions. Ai2 warns that it does not expect the system to generalize well to future climates, especially those with higher carbon dioxide concentrations.

The next step will be to train it on CM4 simulations that include CO₂ levels of up to four times preindustrial values. Until that happens, you should understand SamudrACE as evidence that AI can accelerate coupled climate models, not as a tool ready to answer every question about the future climate.