Accessing ice effective viscosity using physics-informed deep learning

Ice flows in response to stresses according to the flow law that involves ice viscosity. An accurate description of effective ice viscosity is essential for predicting the mass loss of the Antarctic Ice Sheet, yet measurement of ice viscosity is challenging at a continental scale. Lab experiments of polycrystalline ice shows that the effective viscosity of ice obeys a power-law relation with the strain rate, known as Glen’s flow law. However, it remains unclear how processes at ice-shelf scales impact the effective viscosity of glacial ice. Here, we leverage the availability of remote-sensing data and physics-informed deep learning to infer the effective ice viscosity and examine the rheology, i.e. flow law,  of glacial ice in Antarctic Ice Shelves. We find that the rheology of ice shelves differs substantially between the compression and extension zones. In the compression zone near the grounding line the rheology of ice closely obeys power laws with exponents in the range 1<n<3, consistent with prior laboratory experiments. In the extension zone, which comprises most of the total ice-shelf area, ice performs complex rheological behavior, deviating from laboratory findings. We also discover the areas where ice viscosity appears non-isotropic.