An orchestra plays, but you only hear one violin. That's how river flooding used to be predicted: take the water temperature in a tiny square of the Pacific Ocean and use it to judge whether the Amazon would flood. But the ocean is vast, and heat flows from one place to another, like music moving between instruments.

Scientists let the neural network listen to the whole orchestra: they showed it ocean surface temperatures over 56 years across a huge area from the Indian Ocean to the mid-Pacific. The network learned to predict how much water would flow through the Amazon and Congo. It turned out to make smaller errors than old methods that only looked at one violin. The network also revealed that not only waters near South America matter, but also the Indian Ocean, and even connections between distant regions.

But there's a catch: when the network trained on computer models of the ocean, it found many long-distance connections, but in real observations there are fewer. This suggests that models may exaggerate how strongly different ocean regions influence each other. Now that we know where the network looks, we can improve both the forecasts and the climate models. The ocean conducts rivers with its full orchestra, and we are learning to hear every musician.