A translator in a large hall hears both a whisper right next to their ear and a conversation at the far end. The same mind picks up local nuances while keeping track of the overall discussion. That is how ResoNet works: a neural network that predicts El Niño.

Until now, forecasts beyond a year have been unreliable: the ocean is too complex, and models failed to connect distant regions. ResoNet does it differently. Convolutional layers, like an attentive ear, catch local temperature anomalies, for instance, a warm patch off the coast of Peru. The transformer, like hearing across the whole hall, links those patches to processes in the Indian Ocean and the South Pacific. Together, they produce forecasts 19–26 months ahead.

What matters is not only what the network predicts, but also why. The integrated gradients method reveals which pixels of the input map the network focused on most strongly. It turns out that for long-range forecasts, ResoNet pays attention to the South Pacific and Indian Ocean, while for shorter leads, it focuses on the equatorial region. This aligns with known mechanisms: warm water accumulates off the equator before El Niño, and the Indian Ocean acts as a capacitor, releasing heat with a delay.

For the first time, a neural network has captured an asymmetry: for El Niño at 18 months lead, the South Pacific is important, whereas for La Niña, the eastern equatorial region matters. Previously, such differences were only seen in dynamical models. Now the translator not only repeats what it hears, but also explains which voices in the hall were decisive. This is what reduces skepticism: ResoNet is not a black box, but a tool that can be checked against the physics of the ocean.