Domain-Decomposed Graph Neural Network Surrogate Modeling for Ice Sheets

Adrienne M. Propp, Mauro Perego, Eric C. Cyr, Anthony Gruber, Amanda A. Howard, Alexander Heinlein, Panos Stinis, Daniel M. TartakovskyarXiv 2025 · DOI · arXiv

Abstract

Predicting ice sheet dynamics is essential for climate forecasting but is computationally expensive due to the complex nonlinear partial differential equations (PDEs) involved. This work introduces a physics-inspired Graph Neural Network (GNN) surrogate that leverages a Hamiltonian bracket-based architecture, domain decomposition, and transfer learning. Applied to Greenland's Humboldt Glacier, the model provides accurate velocity predictions at a fraction of the cost of traditional solvers, enabling robust uncertainty quantification (UQ) that was previously computationally prohibitive.

Introduction

The movement of ice sheets is a critical factor in sea-level rise projections. However, modeling these dynamics requires solving complex PDEs on large, unstructured meshes. Traditional numerical solvers, while accurate, are computationally intensive, often taking hours or days for a single simulation. This latency creates a bottleneck for Uncertainty Quantification, which requires thousands of model runs to explore parameter spaces such as basal friction. To address this, researchers have developed a GNN-based surrogate capable of high-fidelity predictions with extreme efficiency.

Methodology

The surrogate model incorporates three core innovations to maintain physical accuracy on unstructured grids:

  1. Bracket-based GNNs: Utilizing a Hamiltonian formulation, the network treats its depth as continuous "pseudo-time." This prevents the "oversmoothing" common in standard GNNs, preserving the sharp spatial gradients necessary for modeling fast-moving ice streams.
  2. Domain Decomposition (DD): The large glacier mesh is partitioned into smaller subdomains using spectral clustering. This allows for parallelized training and enables the model to specialize in regional physics, such as the distinct dynamics of the glacier's interior versus its terminus.
  3. Transfer Learning: By pre-training on high-complexity regions (like the terminus) and fine-tuning on the rest of the domain, the model achieves high accuracy even when high-fidelity training data is scarce.
The unstructured mesh of the Humboldt Glacier used for the GNN surrogate.

Data and Features

The study focuses on the Humboldt Glacier in Greenland, using data generated by the MPAS-Albany Land Ice (MALI) model. The mesh consists of 18,544 nodes and 54,962 edges. The model processes several key physical inputs to predict ice velocity, including ice thickness, bed topography, and basal friction—the latter being the primary source of uncertainty in ice sheet modeling.

Spatial distribution of input features: Bed Topography, Basal Friction, and Ice Thickness.

Key Results

The surrogate model demonstrates remarkable performance across several metrics:

  • Inference Speed: Once trained, the model generates predictions in approximately 15 milliseconds per sample, compared to the hours required for traditional PDE solvers.
  • Accuracy: The combination of DD and transfer learning effectively eliminated errors near the glacier terminus, a region notorious for high velocity and steep gradients.
  • UQ Capability: The fine-tuned model successfully recovered the full distribution of ice mass flux, capturing the variance necessary for meaningful uncertainty quantification, whereas baseline models often only predicted the mean correctly.

Conclusion and Future Work

This research demonstrates that physics-inspired GNNs, combined with domain decomposition, can serve as powerful surrogates for complex geophysical systems. While currently a "snapshot" surrogate, future work aims to extend this to autoregressive time-stepping and the development of "foundation models" capable of generalizing across different glaciers globally with minimal fine-tuning.

Summary updated

BibTeX
@misc{propp2025domain,
  author = {Adrienne M. Propp and Mauro Perego and Eric C. Cyr and Anthony Gruber and Amanda A. Howard and Alexander Heinlein and Panos Stinis and Daniel M. Tartakovsky},
  title = {{Domain-Decomposed Graph Neural Network Surrogate Modeling for Ice Sheets}},
  year = {2025},
  eprint = {2512.01888},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.01888},
}
Figure