Simulating the ice flow of marine ice sheets and outlet glaciers with IGM, a physics-informed deep-learning model

Marine sectors of ice sheets and marine-terminating glaciers are pivotal to cryospheric mass loss. Despite their limited areal extent, marine regions strongly regulate ice discharge across the grounding line. Moreover, although most glaciers are not marine-terminating, those that are represent a large share of total glacier ice volume and, among glaciers, dominate the potential glacier contribution to future sea-level rise. Accurately representing marine regions in ice-flow models is thus essential.

Here, we present current progress towards extending IGM to account for marine regions. IGM is a model that uses physics-informed machine learning to simulate ice-flow dynamics (Jouvet and Cordonnier, 2023). In IGM, the mapping between glacier configuration (e.g., geometry) and ice velocity can be obtained either with classical numerical approaches or by learning a neural-network surrogate through the optimization of its weights. The model is implemented in Python with a modular design, which facilitates the implementation and modification of individual physical components, and enables the use of high-performance libraries such as TensorFlow for GPU computing. IGM has demonstrated orders-of-magnitude speedups over classical solvers and has enabled continental-scale, long-term simulations of mountain glaciers (e.g., Leger et al., 2025).

However, IGM was not originally developed for marine settings. Extending it to such regions is challenging because (i) the stress balance differs from that of grounded ice, with negligible basal friction, (ii) the resulting dynamics are markedly more nonlocal due to the stronger influence of membrane stresses (the elliptic terms in the stress balance), and (iii) the transition from grounded to floating ice occurs over short spatial scales, on the order of a few hundred meters. We describe the progress made to address these challenges, including a multiscale strategy that locally decouples distinct flow regimes. We will present results on idealized test cases, with particular attention to numerical accuracy and computational efficiency.