Interpretable Neural Network Forecasting of Ocean State Transitions Using Saliency

Qi-fan Wu, Dion Häfner, Roman Nuterman, Guido Vettoretti, Markus Jochum

Track: Regular Track (Page limit: 6-8 pages)

Keywords: Time Series; Saliency Maps; Climate Dynamics

Abstract: We study sudden transitions in a key component of the climate system, the Atlantic Meridional Overturning Circulation (AMOC). Exploiting simulation results from a fully coupled climate model, we train a convolutional neural network to predict the AMOC as a result of ocean subsurface density and freshwater forcing. We find that the model can forecast transition dynamics it has never seen. Furthermore, we show how saliency maps can be used to interpret black-box neural network models in climate dynamics and enhance their performance, and we demonstrate that high saliency on excitable regions enables out-of-sample prediction of large-scale transitions. This approach opens new perspectives for interpretable, long-term AMOC forecasting.

Submission Number: 11