El Niño is a warm current in the tropical Pacific that every few years alters weather worldwide, causing droughts, floods, and hurricanes. Predicting its onset is challenging because it involves teleconnections, that is, long-distance interactions between regions thousands of kilometers apart. Classical models based on convolutional neural networks capture local patterns well but struggle with such global connections. A whole research direction has formed around this problem: some researchers predict El Niño from time series of the index, others from temperature maps, and still others build climate networks where nodes are grid points and links are statistical dependencies. Our work sits alongside these attempts but takes a step further: instead of fixed local filters, we give the model the freedom to discover important connections between any points in the ocean on its own.
The authors used sea surface temperature data from the ERSSTv5 dataset, training on 1871 to 1973 and validating on 1984 to 2020. They represented the ocean map as a graph: each grid point became a node, and the model could adjust connections between points during training. Unlike convolutional neural networks, which assume important patterns are nearby, the graph network allows information to flow along any edge, for example, from the western Pacific to the eastern Pacific. The architecture was adapted from a recent work on traffic forecasting and tailored to climate data. In the first stage, they used a simple two-layer network without predefined edges to test the idea.
Initial results are encouraging. For one-month-ahead forecasts, the graph network achieved a correlation of 0.9867, whereas the convolutional model from a previous study reached about 0.94. For three months ahead, the correlation was 0.8936 versus 0.8742. For six months ahead, the graph model still lags: 0.6776 against 0.7616. The authors attribute this to the small dataset (only 1233 points in the training sample) and to the fact that for longer forecasts, sea surface temperature alone is insufficient; data on upper-ocean heat content are needed.
This work shows for the first time that graph neural networks are applicable to seasonal climate forecasting. That sets it apart from neighboring studies that relied on convolutional networks or classical statistical methods. The main advantage of the graph approach is interpretability: after training, one can examine which edges the model deemed important and compare them with known teleconnections. If these connections align with physical patterns, it will strengthen confidence in the model and help climatologists better understand the mechanisms of El Niño.
In the future, the authors plan to add other variables to the model, such as ocean heat content, sea level pressure, and wind. This should improve forecasts for six months and beyond. They are also considering using CMIP5 climate model data for pretraining to compensate for the lack of observations. Another direction is to forecast not only the index but also the spatial distribution of temperature anomalies, which would allow classifying El Niño types.
Improved El Niño forecasts will help prepare for droughts and floods in agriculture, energy, and public health. Moreover, the success of graph networks in this task could extend to other climate phenomena, such as other oceanic oscillations or weather forecasting.
Immediate next steps are to incorporate ocean heat content data into the model and apply transfer learning from CMIP5 models. It is also worth trying deeper architectures and comparing them with ensembles of dynamical models.
For climatology, a key unresolved problem is the predictability of El Niño beyond six months, the so-called 'spring predictability barrier'. This work directly addresses it by proposing a new class of models that may better capture slow global processes.
The name 'El Niño' (Spanish for 'the boy') was coined by Peruvian fishermen who noticed the warm current arriving around Christmas. The graph neural networks used in this study were originally developed for traffic forecasting.
In Kim Stanley Robinson's novel 'Forty Signs of Rain', climatologists struggle with the consequences of abrupt changes in ocean currents. Improved El Niño forecasts could be exactly the tool that helps humanity prepare for the ocean's whims.