All projects
Research / Climate Sciencein-progress

Greenland Ice Sheet Calving Model

Finite-element glacier model of Greenland's calving margin on Firedrake and icepack: meshes the IMBIE2 basin geometry in polar stereographic, drives calving from the effective strain rate, and is built to calibrate on 1995–2007 satellite observations and hindcast the 2007–2015 retreat.

PythonFiredrakeicepackGeoPandasxarrayNetCDF

Skills involved

PythonData ScienceResearch

What This Is

A computational glaciology project: modelling where and how fast the Greenland Ice Sheet loses ice at its marine margins. Calving the mechanical breakup of ice into icebergs at the terminus is one of the largest sources of uncertainty in sea-level projections, because it is a fracture process being represented inside continuum flow models that were never designed for fracture.

The model is built on Firedrake (finite-element PDE solving by automatic code generation from a variational form) and icepack, the glacier flow library that sits on top of it. The intended evaluation target is NASA's Cryosphere Model Comparison Tool, which scores ice-sheet models against remote-sensing observations rather than against each other.

The Setup

Geometry. Greenland's basins come from the IMBIE2 v1.3 shapefile, reprojected to EPSG:3413 (NSIDC Sea Ice Polar Stereographic North) working in metres on a polar projection rather than degrees is non-negotiable once you are computing strain rates. The individual drainage basins are unioned into a single Greenland-wide domain and meshed at 5 km resolution, with a CG2 vector space for velocity and CG1 for thickness the standard Taylor-Hood-style pairing that keeps the momentum solve stable.

Observations. Annual ice-mask and thickness fields (ISMIP-style sftgif, the CF standard name for land-ice area fraction) are interpolated from their native grid onto the mesh nodes. The experiment design is a genuine hindcast: calibrate on 1995–2007, then predict 2007–2015 and compare against what actually happened, rather than tuning on the period you report.

Calving law. The terminus retreat rate is driven by the effective strain rate:

ε     = sym(∇u)
ε_eff = √(ε : ε)
c     = coefficient · D · ε_eff

masked to the terminus and capped.

This is the strain-rate family of calving laws stretching drives crevasse propagation drives breakup rather than a fixed rate or a pure water-depth criterion. The parameter set alongside it (Glen's law A = 3×10⁻²⁵, n = 3, friction C, σ_max = 0.12 MPa) is the standard viscous-flow parameterisation the full icepack solve needs.

Output. Time series of calving volume flux, ice area and mean thickness, plus the final ice mask, written to CF-attributed NetCDF so it drops straight into an observation-comparison pipeline.

Where It Actually Stands

Honest status: the scaffolding is real, the physics solve is not finished.

The current time loop uses a prescribed velocity parameterisation u ~ 50 + 2000·exp(−h/800) as a stand-in rather than solving the momentum balance through icepack. Runs so far have fallen back to the synthetic-data path because the observation NetCDF wasn't wired in, which means the output plots in the repo show a flat, essentially null result. They are plumbing tests, not findings, and are treated as such.

The calibration step is likewise a placeholder: it bins the observed area-change rate into three brackets and picks one of three calving coefficients. That is a sanity harness, not an inversion.

What's Next

The gap between here and a defensible model is well-defined:

  • Replace the velocity stand-in with a real solve icepack's IceStream/IceShelf models, so the strain rate driving the calving law comes from the actual momentum balance rather than a thickness-dependent guess
  • Wire in the real observation fields and drop the synthetic fallback entirely; a model that silently substitutes fake data is worse than one that fails loudly
  • Proper calibration: gradient-based inversion of the calving coefficient against observed terminus positions, which is exactly what Firedrake's adjoint capability exists for
  • Move from a rectangle mesh to a boundary-conforming mesh of the actual coastline the current mesh is a bounding box with the geometry interpolated onto it
  • Per-basin rather than ice-sheet-wide: the northwest basin is the natural first target, since it's where the fastest-retreating outlet glaciers are

The vendored icepack tutorials (synthetic ice shelf, Larsen ice shelf inverse problem, hybrid ice stream) are the reference path for each of these steps.

Last updated Jul 27, 2026

← Back to all projects