Abstract

The largest uncertainty when projecting the Antarctic contribution to sea-level rise comes from the ocean-induced melt at the base of Antarctic ice shelves. Current parameterisations used to link the hydrographic properties in front of ice shelves to the melt at their base struggle to accurately simulate basal melt patterns. We suggest that deep learning can be used to tackle this issue. We train a deep feed-forward neural network to emulate basal melt rates simulated by highly-resolved ocean simulations in an idealised geometry. We explore the advantages and limitations of this new approach through sensitivity studies varying hyperparameters, input variables and training choices. We show that large neural networks perform better, that the input format of the temperature and salinity matters most, and that the neural network can be applied to conditions outside of its training range if trained appropriately. The results are promising and we make recommendations for further work with this approach.