Samudra: The AI Ocean Emulator That Runs 150 Times Faster Than Traditional Models

Running climate simulations at resolutions fine enough to capture crucial ocean dynamics has traditionally been impossible without supercomputers and months of processing time. Now, researchers at CDS have developed an AI model that can perform these simulations in hours rather than days.

A team including CDS Joseph and Herbert Keller Professor of Applied Math & Professor of Mathematics and Data Science Laure Zanna and CDS Associate Professor of Mathematics and Data Science Carlos Fernandez-Granda has created Samudra, an AI-powered ocean emulator that can simulate global ocean behavior for centuries while remaining stable. Their work, detailed in the paper “Samudra: An AI Global Ocean Emulator for Climate,” represents a significant advancement in climate modeling technology.

“Training this network allows us to predict short changes in the ocean using existing simulation data. Once trained, you can unleash it and run it for years and years, giving you a reasonable long-term simulation,” Fernandez-Granda said. Zanna added: “Samudra demonstrates the promise of deep learning for computationally cheap ocean simulations and the potential to reduce energy usage.”

Why Ocean Modeling Matters

The oceans play a critical role in our climate system, absorbing more than 90% of excess heat and 25% of carbon dioxide emissions. Understanding ocean processes is essential for predicting climate change impacts and developing mitigation strategies. However, traditional ocean modeling requires enormous computational and energy resources.

Samudra addresses this challenge by using a neural network architecture to create a digital twin of the ocean that runs 150 times faster than traditional ocean models. This breakthrough allows researchers to perform in just hours what would typically take days on thousands of CPU cores.

How Samudra Works

The team trained their AI model on data from OM4, a state-of-the-art ocean general circulation model. Samudra successfully reproduces key ocean variables, including sea surface height, ocean currents, temperature, and salinity throughout the ocean’s depth.

What sets Samudra apart from previous emulation attempts is its stability over extended periods. The model can run for centuries without diverging from realistic behavior — a challenge that has plagued previous neural network approaches to climate modeling.

“A good property of this model is that it’s stable and can run for a long time,” noted Fernandez-Granda. “A lot of models explode after a while since they’re trained to learn very short-term transitions.”

Practical Applications

While still in development, Samudra’s potential applications are vast. The ability to run large ensembles of climate simulations could help scientists better understand climate uncertainty and extreme events. It could also accelerate data assimilation for more accurate ocean state estimation and improve operational forecasting.

These capabilities are particularly valuable for studying different climate scenarios, as they allow researchers to explore how different parameters might influence future outcomes without the prohibitive computational costs of traditional modeling.

The model is part of a larger initiative called M²LInES (Multiscale Machine Learning In coupled Earth System Models), a five-year project led by Zanna and funded by Schmidt Sciences. This project aims to improve climate modeling with AI across various scales and components of the Earth system.

Challenges and Future Work

Despite its impressive capabilities, Samudra still faces challenges. The team notes that while the model reproduces variability like El Niño-Southern Oscillation events well, it struggles to capture the correct magnitude of forcing trends while simultaneously remaining stable.

“We’re really not there yet in terms of making predictions,” Fernandez-Granda explained. “At the moment, we’re just trying to match what the actual climate simulations are doing. These are baby steps, but that’s where we ultimately would like to go.”

The research team continues to refine the model, working toward a fully coupled ocean-atmosphere emulator that could further revolutionize physics discovery and climate science in particular.

Samudra represents more than just a proof of concept — it’s a functional tool that climate scientists can use immediately. The model and its source code are publicly available, allowing researchers worldwide to leverage this technology for their own studies. “Samudra opens the door to democratizing ocean and climate modeling,” said Zanna. “Anyone anywhere can use it to explore how the physics of the oceans work and think of exciting new scientific discoveries.”

As climate modeling continues to evolve, AI emulators like Samudra may become essential tools in our quest to understand and address the challenges of climate change.

See also Zanna’s comprehensive blog post on the model.

By Stephen Thomas