Identifying Arctic states vulnerable to rapid ice loss with a convolutional neural network

Abstract

This work investigates the incidence of rapid ice-loss events in the Arctic, in which a substantial fraction of September sea-ice is lost in only a few years. Such events are seen to occur in climate model simulations forced with increasing greenhouse gas concentrations, but not as yet in the relatively short observational record. We train a convolutional neural network (CNN) using six large climate model ensembles to diagnose vulnerability to rapid ice-loss events. The trained model is shown to separate higher and lower risk states in out-of-sample climate models via a leave-one-out cross-validation. The model emphasises the role of persistently thin sea-ice, remaining sea-ice extent and sustained high upper ocean heat content as predictors of rapid ice-loss. The CNN is applied to observational data over the satelite-era showing a signal for rapid ice-loss prior to a drop around 2007 and an increasing risk from around 2015 to present. Such rapid-ice loss events would likely have dramatic consequences for Arctic ecosystems, coastal communities and global shipping.