How a Neural Network Catches El Niño by Its Long-Distance Connections ✎

Original: "Graph Neural Networks for Improved El Niño Forecasting"
arXiv:2012.01598 · 2020-12-02 · arXiv non-exclusive · independent analysis · 2 min
The ocean talks to itself, and a graph neural network learns to hear that conversation and predicts El Niño.

Every few years, a warm current called El Niño appears in the tropical Pacific. It is like an uninvited guest that turns a house upside down: causing droughts in some places and floods in others. Scientists want to warn of its arrival in advance, but the ocean does not give up its secrets. The cause of El Niño is not in a single location but in how different parts of the ocean communicate, even when thousands of kilometers apart. Conventional computer models only see neighbors, like a person who hears only the neighbors next door but does not know that someone across town opened a faucet and changed the pressure in the pipes.

A graph neural network is a program that works not with a grid but with a network: each ocean point becomes a node, and the program tunes the connections between them as it learns.

Scientists gave such a program a map of the ocean and let it decide which points are strongly connected and which are weakly connected. It turned out that for a one-month forecast, it predicts El Niño more accurately than previous models: a correlation of 0.9867 versus 0.94. For three months ahead, it is also better: 0.8936 versus 0.8742. This is like a resident who finally learns about the faucet across town not from a wet ceiling but from a slight drop in water pressure. However, for six months ahead, the new model still lags because it lacks data on how much heat is stored in the upper ocean layer. But the scientists already plan to add that data and hope the model will then warn of the ocean's whims even earlier.

The name 'El Niño' (Spanish for 'the boy') was given by Peruvian fishermen who noticed that the warm current arrives around Christmas. And the graph neural networks used in the study were originally developed for predicting road traffic.

In Kim Stanley Robinson's novel 'Forty Signs of Rain', climatologists battle the consequences of abrupt changes in ocean currents. Improved El Niño forecasts could be the very tool that helps humanity prepare for the ocean's whims.

Concepts
graph neural network neural network regression spring predictability barrier forecast ENSO El Niño index equatorial wave propagation Jin-Timmermann model El Niño-Southern Oscillation La Niña ERSST Large Language Models Multivariate ENSO Index NOAA CPC NOAA PMEL / TAO Generative Adversarial Networks deep learning in medical imaging Graphs Machine Learning TAO/TRITON moorings