The ocean hums like a crowded hall: in one corner, voices argue about decades; in another, about years; in a third, about the changing seasons. To pick apart these voices, climatologists used to rely on filters they tuned themselves, letting through only slow waves or only fast ones. But the Pacific plays by its own rules: a warm patch off the coast of Peru can be a brief spike or part of a long-term rearrangement. When these rhythms overlap, they baffle any linear scheme.

The new study takes a different approach: a neural network that learns to separate sounds not by preset frequencies but by the very structure of the signal. It is shown 84 years of monthly sea surface temperature maps, with the trend and seasonal cycle removed. Inside the network are just five "auditory receptors," or hidden dimensions. To keep them from eavesdropping on each other, the training includes penalties: if two receptors pick up the same rhythm, the network is punished; if their spatial patterns match, it is punished too. As a result, each receptor tunes in to its own voice.

Three voices emerged clearly. The lowest, decadal, stretches over 10–20 years and resembles the Pacific Decadal Oscillation in shape (correlation 0.84 on average across the basin, and 0.92 in the North Pacific). The middle, interannual, pulses every 3–7 years and almost exactly matches the Niño-3.4 index (correlation 0.82), which tracks El Niño. The highest, quasi-biennial, oscillates roughly every two years and most closely resembles the eastern Pacific type of El Niño (correlation 0.84), but with a shorter period than standard indices.

The key finding lies in how the voices interact. When the decadal and interannual voices sound in unison, a pattern emerges that resembles central Pacific El Niño (correlation with the C-index 0.95). When they are out of phase, a weak eastern Pacific pattern results. In other words, the slow decadal rhythm decides what the next El Niño will look like. The quasi-biennial voice, meanwhile, is tied to the development and decay of the event: it strengthens in July–November, while the interannual voice peaks in November–January.

This is not just a pretty picture. The same method was applied to two climate models, E3SM2 and CESM2, revealing that CESM2 overestimates the frequency of central Pacific El Niños and misrepresents the seasonal phase transition. Such errors were hard to spot before because they hid in the overall mix of oscillations. Now, with the voices separated, we can check whether a model hears the ocean the same way real data do. And if not, we can pinpoint which rhythm it is getting wrong.