Capabilities of IGM, a thermo-mechanical glacier evolution model accelerated by deep-learning and GPU

Abstract

We present the concepts and capabilities of IGM (https://github.com/jouvetg/igm), a fast and accessible Python model that simulates the evolution of glaciers at any scale by coupling ice thermomechanics, surface mass balance, and mass conservation. IGM models the ice flow by physics-informed deep learning. Specifically, we use a convolutional neural network, which is trained to minimise the energy associated with high-order ice flow physics. Based on the Tensorflow library, IGM performs a suite of fast, vectorised, and differentiable operations, which can be accelerated with a graphics processing unit (GPU). In turn, this allows fully-parallelised implementations of key model components in glacier modelling applications such as the positive degree day surface mass balance scheme, the enthalpy scheme for modelling the thermal regime of ice, or the integration of a large amount of particle trajectories for modelling debris transportation. As a result, IGM combines coding simplicity and modularity, high computational efficiency, state-of-the-art thermomechanical modelling, and efficient data assimilation thanks to underlying automatic differentiation tools. We demonstrate the capability of IGM for two different applications. First, we present a complete workflow (including OGGM-based data preprocessing, inverse and forward modelling, rendering of results) that allows us to model any mountain glacier in the world within a few minutes requiring only its Randolph Glacier Inventory (RGI) ID. Second, we present an application in paleo glacier modelling by simulating the entire European Alpine ice field (about 800 km long) in high resolution (200 m) during the Last Glacial Maximum.