El Niño: a new method of explainability reveals what the neural networks used to predict it ‘see’

The AGOP method outperforms conventional explanatory techniques on three out of four criteria and can even trigger or halt a simulated El Niño ten months later.

Explaining the predictions of a neural network remains an open problem in physics. Three researchers compared several explainable AI methods (gradients, Integrated Gradients, GradientSHAP) applied to networks that predict the El Niño Southern Oscillation from data from the Zebiak-Cane model. They scored them on four criteria: sensitivity, attribution, robustness and spatial coherence of the patterns highlighted.

Their new method, AGOP (average of the external product of gradients), exploits global information on gradients to identify an important direction for each situation. It obtains the best scores of attribution, robustness and coherence in the configurations tested; only the simple gradient method exceeds it in sensitivity, by construction. AGOP brings out a depth signal of the equatorial thermocline in accordance with the theory of the recharge oscillator, as well as an area of the Southeast Pacific that may be specific to the model used.

The authors go further by disrupting the physical model along the identified directions: these disturbances are enough to suppress extreme episodes, or, from a quasi-neutral state, to cause a strong El Niño or a strong La Niña ten months later. A way to generate physical hypotheses from AI.