A translator who learned a language from thousands of books can interpret a conversation without being distracted by noise in the hall. Similarly, a neural network trained on 127,137 years of model climate produces a winter forecast in a fraction of a second, faster and more accurately than bulky dynamical systems.

Seasonal forecasting months ahead is an attempt to peer into the future by relying on slow processes in the ocean. The ocean stores heat and releases it gradually to the atmosphere, like a stone oven after firing. If in July the water at depths of 5–300 meters is warmer than usual, winter in some regions may be milder. But to capture this link, you need to account for thousands of details: currents, evaporation, clouds, topography. Dynamical models that solve physics equations stumble over three barriers: imprecise initial data, simplified schemes for small-scale processes, and too few ensemble members. A neural network sidesteps these barriers differently: instead of trying to reproduce physics, it seeks patterns in a giant archive of ready-made simulations.

The model's architecture resembles a translator with two blocks. An encoder compresses the pair "July ocean–winter weather" into a short code, just as a translator grasps the essence of a phrase. A decoder expands that code into a forecast map, just as a translator speaks the answer in another language. The code retains only what matters for the ocean–atmosphere connection. Noise, that is, random fluctuations, does not interfere: the model learns to separate signal from interference, just as an experienced translator ignores a cough in the hall.

Training proceeded in three steps. First, the model learned to predict winter from the ocean without any code; then it mastered compression; and finally all parameters were fine-tuned together. After that, the neural network was run on real ocean data from 1982 to 2018 and compared with four leading dynamical systems. The result was unexpected: the correlation of the forecast with observations for temperature in the eastern tropical Pacific exceeded 0.9, and for precipitation and temperature over nearly all continents, the neural network was more accurate than the benchmarks.

But accuracy is not the only point. The neural network can serve as a mirror for dynamical models. If the forecast from a neural network trained only on data from a single model turns out better than that model's own forecast, then the problem lies not in the data but in the model's formulation. In this way, the authors showed that initialization errors and simplified schemes contribute differently across regions. Sensitivity maps confirmed that the neural network relies on physically meaningful patterns: warmth in the eastern tropical Pacific is linked to drought over the Maritime Continent, as predicted by the Walker circulation.

A translator who learned a language from books not only translates but also explains why a phrase sounds the way it does in the original. A neural network trained on model climate not only predicts weather but also shows where dynamical systems go wrong and why. This means that seasonal forecasts for agriculture, water resources, and energy could become more reliable, and developers of climate models gain a tool for identifying weak points.