Climatologists manually define phenomena like El Niño using thresholds of temperature anomalies in specific regions. Alongside this work, the archive hosts studies where neural networks search for new galaxies, separate climate signals from weather noise, or infer properties of neutron stars. All use machine learning to extract structure from complex data, but this work stands apart: it is the first to apply a causal approach to real climate data rather than synthetic examples.
The authors used NCEP-DOE Reanalysis 2 data from 1979 to 2014 and extracted weekly averages of zonal wind and sea surface temperature in the equatorial Pacific band (140°–280°E, 10°S–10°N). They adapted their previous causal feature learning algorithm by replacing probability density estimation with regression of the conditional mean. First, a neural network with three hidden layers predicted temperature from wind; then the regression results were clustered to obtain wind macro-states. For each wind state, k-nearest neighbors among temperature maps were computed, and these representations were clustered into temperature macro-states.
The algorithm identified four wind states and four temperature states. Two temperature states matched El Niño and La Niña precisely: more than 75% of weeks in these states exceeded the weak anomaly threshold (0.5°C), and more than 50% exceeded the strong anomaly threshold (1.5°C). In comparison, k-means clustering even with 16 clusters did not achieve such accuracy. On shuffled data, the algorithm found no significant states, confirming its sensitivity to real causal connections. When the number of states increased from 4 to 16, the accuracy of identifying El Niño and La Niña grew slowly, and the states themselves split into subclasses.
This work shows that causal feature learning can automatically extract climate macro-variables without manual tuning. Unlike neighboring studies that use conventional machine learning for prediction or clustering, this is the first to apply a causal approach to real wind and temperature data. This opens the way to discovering new climate phenomena and to more rigorous definitions of known ones. The method also allows formulating hypotheses for expensive climate simulations.
The method can be extended to other variable pairs in the climate system, such as pressure and precipitation. A theoretical extension to systems with feedback loops would allow testing causal hypotheses with climate models. It may also uncover new types of El Niño, as suggested by the fragmentation of states when their number is increased.
The results will affect climatology, oceanography, and data science.
Test the method on other climate variables and develop a theoretical foundation for systems with feedback loops.
In climatology, the precise definition of El Niño and its causes remain open questions. This work offers a way to automatically identify such phenomena from data, which could help resolve debates about El Niño types.
The name 'El Niño' comes from the Spanish for 'the Christ child,' because the warm current often appears around Christmas.