A warm patch off the coast of Peru does not appear every year, but when it does, deserts are drenched with rain and Australia is plagued by wildfires. This is El Niño, a phenomenon that climate models struggle to reproduce. Simple models do well at estimating how often it occurs but lose spatial details. Complex models capture the three-dimensional picture but are systematically biased and demand enormous computational resources.
The authors of the new study acted like a conductor who brings together a soloist and an orchestra. The soloist here is a simple stochastic model that sets the tempo: it generates a sequence of warm and cold events in a narrow band along the equator. The orchestra is an intermediate model with a neural network that adapts to the soloist and fills in the rest of the ocean picture. The neural network, with 40 million parameters, was pretrained on a 1200-year CESM2 run and then fine-tuned on real GODAS data.
The model was run for 1000 years, with the last 960 analyzed. The results show spectral peaks at periods of 3–4 years for sea surface temperature in the Niño-3 and Niño-4 regions, and a seasonal cycle with a minimum in June. The distribution of longitudes of maximum temperature anomalies shows two peaks, as observed, whereas CESM2 produces a single peak in the central Pacific. For eastern Pacific El Niño events, the model correctly placed anomalies in temperature, wind stress, and sea surface height, correcting CESM2's biases.
Heat budget analysis confirmed that the main positive mechanisms are the along-equator heat transport by currents and the influence of subsurface temperature on the surface. The residual term acts as a negative feedback, preventing the model from running away.
This approach paves the way for multi-module climate models, where each component handles a part of the task. Long time series from the two-core model are suitable for pretraining neural network forecasting models and for analyzing rare events. In the future, versions with more detailed ocean or atmosphere components may emerge. For now, the orchestra plays more harmoniously than a large symphonic ensemble, and at a lower cost.
The model, with 40 million neural-network parameters, works in tandem with a simple stochastic model that requires almost no computational resources; together they are faster than CESM2 and more accurate statistically.