Modeling of anisotropic ice flow in Elmer/Ice using machine learning (M/F)

id="offer-description" class="ecl-u-type-heading-2 ecl-u-mt-none ecl-u-mb-l">Offer Description -------Ice flow in polar ice caps and mountain glaciers is largely due to plastic deformation, which can be extremely slow. The rheology of polar ice exhibits strong viscoplastic anisotropy, linked to its microstructure (preferred crystal orientation), which evolves with plastic deformation. Such strongly anisotropic behavior should have a marked influence on the flow dynamics of ice caps, particularly their response to climate transitions, but it has never been fully accounted for in models. The relationship between rheology and microstructure is very well captured by mean-field homogenization methods, which account for ice deformation mechanisms at the fine scale (micrometers) to estimate the effective behavior at the decimeter scale, depending on the material's microstructure. These methods, although very powerful, remain far too computationally expensive to be implemented as a behavior law in a finite element code for polar ice cap flow (scale: 1,000 km). We therefore propose to train a neural network on a nonlinear homogenization model in order to reproduce the nonlinear and evolving anisotropic rheology of ice, and then to incorporate this neural network into a large-scale flow model. Our initial work in this field has demonstrated the accuracy and extraordinary computational efficiency of AI trained on the homogenization model. We will thus be able to study the impact of ice