The ocean resembles a giant web. Touch one thread, and a distant corner trembles. The neural network from this study learned to weave such a web to predict El Niño.

El Niño is a warm current that, every few years, brings floods or droughts to different countries. Previously, forecasts were made with physical models, but they often erred several months ahead. Other neural networks treated the map as a picture and missed that distant ocean points influence each other.

The new model represented the ocean as a graph: 1345 points on a map, connected by invisible lines. Each point stored water temperature for three months. The connections were not predetermined; the model itself decided which points should exchange information, keeping only the eight strongest for each point. Signals then propagated along these connections, like rumors through a social network.

The result outperformed previous approaches. For a six-month forecast, the correlation with observations was 0.780, whereas a convolutional neural network achieved 0.762. For the strong El Niño of 2015, the model yielded a correlation of 0.948 over the year. Most importantly, the learned web matched what physicists already knew about the ocean. On short timescales, key points are located where temperature is measured. By six months, the influence shifts to the western Pacific, and by 23 months, to the Indian and Atlantic Oceans.

Previously, neural networks were black boxes: they provided forecasts but did not explain why. Now, the learned connections can be verified. It is as if the web woven by the machine turned out to be the very one the ocean weaves itself.