The Pacific Ocean is not uniformly warm: near the coasts of South America, the water alternately warms and cools. These changes occur at different rates. Some last a couple of years, others several years, and still others decades. They overlap like three conversations in one room. Previously, scientists tried to separate them using filters they devised themselves, keeping only slow waves or only fast ones. But the ocean does not obey such rules.

Now a different approach has been devised. A neural network examines ocean temperature maps over 84 years and learns on its own to extract three distinct rhythms. Inside it, there are only five "listeners." To keep them from interfering with each other, penalties are added during training: if two listeners catch the same rhythm, the network is penalized. As a result, each listener tunes into its own voice.

The first voice is the slowest, decadal one. The second is intermediate, repeating every few years and closely resembling El Niño. The third is the fastest, roughly every two years. It turns out that the slow voice determines what the upcoming El Niño will be like: a warm patch near the coast or in the middle of the ocean. The fast voice is linked to the onset and end of El Niño: it strengthens in summer and autumn, before the event develops.

This helps validate computer climate models. One of them, CESM2, was tested, and it was found to make El Niño too often "central" and to confuse seasons. Now, by separating the voices, we can understand which rhythm the model gets wrong.