Fifty-eight hours. That is how long the Parallel Ice Sheet Model, or PISM, needed to march the Antarctic ice sheet from 2025 to 2100 on the setup used here, chewing through roughly 46 minutes of processor time for every single model year. A team at Victoria University of Wellington got a machine learning stand-in to reproduce the same 76 years of output in about an hour and a half.
That gap is the whole point. Physics-based ice sheet models encode what glaciologists know about how ice deforms, slides, floats and melts, and they are trusted for exactly that reason. But they are slow, and slowness has a scientific cost. Researchers want ensembles: hundreds or thousands of runs under different emissions futures, different melt assumptions, different starting conditions, because that spread is what tells you how uncertain a sea level projection really is. At 58 hours a run, you get a handful.
Paula Maddigan and colleagues, working across the university's data science centre and its Antarctic Research Centre, tried a deliberately unfashionable approach. No neural networks. They took a previously published PISM simulation, downsampled to a coarse 51 by 51 grid, and asked whether ordinary tree-based algorithms could learn to mimic it. Each grid cell, each year from 2015 to 2100, carried three climate inputs (air temperature, ocean temperature, precipitation) and three things to predict: the ice mask (is this cell grounded ice, floating ice, or open ocean?), the ice thickness in metres, and the ice velocity in metres per year.
The trick was in what they fed the models. Ice has memory. What a patch of Antarctica does this year depends heavily on what it was doing last year and the decade before, and on what its neighbours are doing. So the team built features that encode both: lagged values going back four years, ten-year rolling averages, and the average conditions across the 3 by 3 block of cells surrounding each point. They tested wider neighbourhoods, 5 by 5 and 7 by 7, and found them slightly worse. At this resolution, only the immediate neighbours carry useful signal.
Random Forest, an algorithm from 2001 that averages the votes of many decision trees, beat both of the fancier gradient boosting methods the team tried. Against 2,257 grid points per year, it got the ice mask wrong on an average of three cells, with a typical error of 3.36 metres on thickness and 65.6 metres per year on velocity. The gradient boosting models had a specific failure the forest avoided: they sometimes predicted negative ice thickness or negative velocity, which is not a thing.
Where the emulator breaks
The honest part of the paper is section 5.8. The authors went looking for the grid points where their model failed worst, and found them clustered around the Ross and Filchner ice shelves, the Amery Ice Shelf, and Jutulstraumen Glacier, places where ice floats and where the boundary between grounded and floating ice moves.
One point on the Ross Ice Shelf tells the story. In the PISM simulation, something changes around 2075: ocean temperature climbs, velocity picks up, thickness drops, and the cell flips from grounded to floating. When the emulator was fed real simulated data each year, it tracked this fine. When it had to run recursively, feeding its own predictions back in as inputs the way it would have to for a genuine forecast, it missed the transition entirely and then compounded that error for the remaining 25 years. The authors are direct about why: the climate signals in their inputs were too weak to trigger the shift, and the model has no features representing ice shelf breakup, grounding line migration, or delayed ice-ocean interactions. The physics that drives the change was never given to it.
Recursive forecasting degraded performance across the board. Mask errors rose from about 3 cells per year to around 30; thickness error went from 3.36 to 29.6.
Why it matters
Antarctic melt is projected to be the single largest contributor to sea level rise this century, potentially 34 centimetres by 2100 under a high-emissions pathway, and metres by 2300. Governments planning seawalls and managed retreat need to know not just the central estimate but the range around it, and that range comes from running models many times over.
What this paper offers is not a replacement for PISM. It is a fast approximation of one specific PISM run, and the authors call it proof-of-concept throughout. The emulator learned from PISM's output, so it inherits whatever PISM assumes and cannot exceed it.
The explainability work may prove more durable than the speed. Using SHAP, a method for attributing a prediction to its inputs, the team could show what the model was leaning on: last year's thickness above all, with ocean temperature mattering more for thickness and precipitation more for velocity, and ocean temperature's influence growing over the century in a way consistent with basal melting being a slow, persistent process. That an emulator can explain itself in terms glaciologists recognise is what makes it worth checking rather than merely fast. Next, the team plans to test it against other ice sheet models and to build features that capture fracture and grounding line retreat, the exact mechanisms it currently cannot see.