An orchestra of two musicians plays from the same score, but each hears only their own line. The atmosphere and ocean in a climate model resemble such a duet: they exchange heat and moisture every few hours, yet until recently their neural network counterparts could not keep a common rhythm. SamudrACE has made them play together for centuries without missing a beat.

Two ready-made emulators, ACE2 for air and Samudra for water, were trained separately on 155 years of simulation from the GFDL CM4 climate model. They were then linked through a physical interface: every 6 hours the atmosphere passes heat, moisture, and wind stress fluxes to the ocean, while the ocean averages them over 5 days and returns surface temperature and ice concentration. This exchange happens in physical space, not in the hidden representations of the neural networks, keeping the model understandable and stable.

The key result: SamudrACE reproduces El Niño with realistic amplitude and a three-year peak in its spectrum. Previously, this phenomenon, which arises only from atmosphere-ocean interaction, had not been achieved in any fully trainable coupled emulator. However, low-frequency variability (periods longer than 4 years) is underestimated: La Niña is weaker than in the reference, and decadal Pacific oscillations are smoothed out.

Before this work, no one had managed to couple three-dimensional neural network emulators of the atmosphere and ocean so that they ran stably for centuries and reproduced key climate phenomena. Now the recipe is known: train components independently, couple them through physical fluxes, and fine-tune jointly. Already today, hundreds of climate scenarios can be run in days instead of months, accelerating tests of hypotheses about internal climate variability and responses to external forcing.

The orchestra now plays in harmony, but the score is still pre-industrial. The next step is to add emulators for biogeochemistry, the carbon cycle, and land, to move toward scenarios with rising CO2. If the musicians master new notes, climate experiments will become as fast as weather forecasts.