IGM, a glacier evolution model accelerated by deep-learning and GPU

Abstract

<p>We give an overview of the Instructed Glacier Model (IGM) -- a new framework to model the evolution of glaciers at any scale by coupling ice dynamics, surface mass balance, and mass conservation. The key novelty of IGM is that it models the ice flow by a Convolutional Neural Network (CNN), which is trained from physical high-order ice flow mechanical models. Doing so has major advantages in both forward and inverse modelling.</p><p>In forward modelling, the most computationally demanding model component (the ice flow) is substituted by a very cheap CNN emulator. Once trained with representative data, IGM permits to model individual mountain glaciers several orders of magnitude faster than high-order ones on CPU with fidelity levels above 90 % in terms of ice flow solutions leading to nearly identical transient thickness evolution. Switching to Graphics Processing Unit (GPU) permits additional significant speed-ups, especially when modelling large-scale glacier networks and/or high spatial resolutions.</p><p>In inverse modelling, the substitution by a CNN emulator does not only speed up but facilitates dramatically the data assimilation step, i.e. the search for optimal ice thickness and ice flow parameter spatial distributions that match spatial observations at best (such as ice flow, surface topography or ice thickness profiles) while being consistent with the high-order ice flow mechanics. The reason is that inverting a CNN can take great benefit from the tools used for its training such as automatic differentiation, stochastic gradient methods, and GPU. </p><p>IGM is an open-source Python code (https://github.com/jouvetg/igm), which deals with two-dimensional gridded input and output data. Together with a companion library of trained ice flow emulators, IGM permits user-friendly, computationally highly-efficient, easy-to-customize, and mechanically state-of-the-art glacier forward and inverse modelling at any scale. We illustrate its potential by replicating a simulation of the great Aletsch Glacier, Switzerland, from 1880 to 2100, based on a Stokes model. The complete workflow (data assimilation and 220 years long forward modelling) at 100 m of resolution takes about 1-2 min on the GPU of a laptop and can be replicated and adapted easily using an online Colab notebook.</p>