A translator who learned a language from thousands of books can interpret a conversation without being distracted by the noise in the hall. Similarly, a neural network trained on a vast archive of model climate predicts winter weather from ocean temperatures in July.

The ocean stores heat and releases it slowly, like a masonry stove after firing. If the water tens of meters deep was warmer than usual in July, winter in some regions may be milder. Conventional forecasting models try to compute every detail of the physics, but they often stumble over inaccurate data and simplifications. The neural network takes a different route: instead of solving equations, it looks for patterns in 127,137 years of ready-made simulations. That's like learning a language from thousands of books rather than from a grammar textbook.

First, the model learned to predict winter from the ocean, then it mastered information compression, and finally all its parts were fine-tuned together. The neural network was then run on real data from 1982 to 2018 and compared with four leading dynamical systems. It turned out that the neural network is more accurate on nearly all continents, and in the eastern tropical Pacific its forecasts are almost perfect. Moreover, the neural network can act as a mirror for older models: if the neural network's forecast, trained on one model's data, beats that model's own forecast, the problem lies not in the data but in the model's design. This approach helps identify weak spots and makes forecasts more reliable for agriculture, water resources, and energy.