A Translator That Learned the Language of the Ocean ✛
A translator at an important negotiation starts by simply repeating words, but gradually begins to anticipate what the counterpart will say. The new weather model works similarly: physics provides the framework, and machine learning learns from errors and fills in the gaps. As a result, it predicts El Niño, the warm current in the Pacific Ocean that alters weather worldwide, months in advance.
The model has two components. The atmospheric part is a simplified physics model with a horizontal resolution of 3.75°, roughly the distance from Moscow to Nizhny Novgorod. The oceanic part is purely machine-based: it predicts sea surface temperature and a new variable, the heat content of the upper 300 meters of the ocean. Previously, it was this missing heat reservoir that prevented forecasts beyond a few weeks.
The result: the model predicts the Niño 3.4 index (the average sea surface temperature anomaly in the central Pacific) better than expensive operational systems during the first 4 months of the forecast. The correlation coefficient drops to 0.5 after about 6.5 months, with the predictability horizon varying from 4 months in spring to 8 months in late summer, a manifestation of the "spring predictability barrier." The model also captures El Niño's global links to precipitation: dry conditions in the eastern Pacific and wet conditions in the southern United States.
Before this work, it was unclear whether a hybrid model with low resolution and machine learning could provide useful forecasts beyond the medium-range window of 7–10 days. Now it is clear: it can. By adding the upper-ocean heat content and tuning the trainable components to different timescales, the scientists produced a tool that competes with operational models at a fraction of the computational cost. Such an approach could benefit academic groups and small forecasting centers that lack access to supercomputers.
A translator who learns from mistakes eventually starts predicting the next phrase. The new model does the same with the language of the ocean: it is not perfect yet, but it can already hint when the Pacific is about to speak.
The ocean component of the model takes a step every 72 hours, while the atmosphere steps every 6 hours. To synchronize them, the ocean simply 'freezes' for 12 atmospheric steps until its turn to update comes around.
The idea of a hybrid model where physics and machine learning work in tandem recalls the concept of 'neuromorphic' systems from science fiction, such as in Neal Stephenson's novel 'Seveneves', where autonomous modules of a spacecraft combine rigid algorithms with learnable components for long-term forecasting.