IGM: an accessible, modular, differentiable, and GPU-accelerated high-order ice flow model

Guillaume Jouvet, Brandon Finley, Thomas Gregov, Sebastian Rosier

We present the Instructed Glacier Model (IGM, v3.2), an open-source framework for simulating glacier evolution from single-glacier to mountain-range scales. IGM is built on a single design principle: all physical processes, including ice flow, surface mass balance, thermodynamics, and mass conservation, are expressed as short sequences of operations on raster grids. This workflow runs natively on GPUs and scales efficiently to large domains. Within this framework, higher-order ice flow is handled by a differentiable mapping trained by automatic differentiation to satisfy the Blatter--Pattyn equations, reducing the computational cost dramatically compared to traditional CPU-based approaches. The framework is structured around five guiding principles: (i) accessibility, through a Python package and a simple YAML-based interface; (ii) modularity, enabling modellers to extend or replace individual components without modifying the core code; (iii) reproducibility, through automated configuration logging and continuous integration; (iv) scalability, enabling large-domain simulations on GPU; and (v) explorability, facilitating ensemble runs, parameter sweeps, and model calibration. We describe the physical and numerical foundations of IGM, its software architecture, and illustrate its capabilities through benchmark tests and real-world glacier simulations.