Emulating a Subglacial Hydrology Model with a Neural Operator

Subglacial hydrology exerts an important control on ice flow and influences the evolution of downstream hydrology, as well as the occurrence of glacial lake outburst floods. However, large-scale modelling of subglacial hydrology remains computationally expensive due to the presence of nonlinear processes.

Within our DeLIGHT (Deep-Learning-Informed Glacio-Hydrological Threat) framework, we aim at enabling coupled simulations of ice flow, subglacial hydrology, and downstream hydrology, with the goal of improving predictions of ice flow evolution and the timing of peak runoff. For this purpose, we will leverage recent advances in deep learning. As a first step, this research focuses on the development of a subglacial hydrology emulator trained using output from the Glacier Drainage System (GlaDS) model implemented within Elmer/Ice, with the aim of applicability to mountain glacier catchments worldwide. The emulator is based on a class of deep learning architecture called neural operators, which allow for better generalisation compared to classical neural networks.

To generate the training set, GlaDS is forced using meltwater inputs derived from a calibrated degree-day model, which is driven by daily climate data spanning the 2000–2010 period. We select 70 glaciers spanning a wide range of physiographic characteristics across Svalbard, Scandinavia, the Alps, and Central and Southeast Asia to provide a representative range of mountain-glacier subglacial hydrological scenarios within the training set. We present results from the training simulations and initial directions for the development of the emulator.