The glacier is now being asked to file its own weight
the glacier is now being asked to file its own weight: an lstm trained on 30,000 observations from 30 swiss glaciers predicts mass balance on eight glaciers it never saw, a generative model downscales greenland’s surface mass balance 32 fold from 160 to five kilometers, and iceboost v2.0 learned from over 7 million thickness measurements to size every glacier on earth.
the ice now files its own mass report.
Context
The Mass Balance Machine (van der Meer and co-authors, 2026) is an LSTM trained on 30,000 observations from 30 glaciers in the Swiss Alps and tested on eight glaciers left out of training. It predicts winter and annual point surface mass balance with a root mean squared error of 0.35 and 0.78 meters water equivalent on those unseen glaciers, and its glacier-wide results agree with geodetic estimates.
The Cryosphere paper (published March 30, 2026) uses a consistency model, a generative network, to downscale Greenland surface mass balance and surface temperature by a factor of up to 32, from 160 to 5 kilometer grid spacing. It is trained on monthly MARv3.12 regional climate model output and enforces a conservation constraint at inference. On the test set it reports a continuous ranked probability score of 5.37 mm water equivalent per month for mass balance, better than interpolation.
IceBoost v2.0 is a gradient-boosted decision tree scheme (XGBoost and CatBoost) trained on about 7 million ice thickness measurements, 7,069,690 points from 1,661 glaciers after cleaning. It models every glacier in Randolph Glacier Inventory v6.0 and v7.0 and gives a global volume of 149 plus or minus 38 thousand cubic kilometers, consistent with an earlier ensemble estimate of 147 plus or minus 28.
The note's description of the LSTM matches the abstract: 30,000 observations, 30 glaciers, eight held-out glaciers. The target is point surface mass balance over the Swiss Alps, so the result is for Alpine glaciers and not a global claim. The paper's accuracy is given as RMSE, not as a percent.
The Greenland model downscales by up to 32 times, as the note says, and the paper trains it on one regional climate model (MARv3.12), so it learns that model's behaviour and not direct observations. It also applies bias-corrected Earth system model output as input to show the idea can work on climate model fields.
The note says IceBoost 'sized every glacier on earth' from over 7 million thickness measurements. That matches the abstract, which also says confidence is highest at high latitudes and weaker on steep mountain terrain, and that its error is 20 to 45% lower than other solutions in the high Arctic only. The model is boosted trees, not a neural network. The three studies measure different things: point mass balance, ice sheet surface mass balance and ice thickness.
Related work
- Glacier Mass Balance Modeling Using a Long Short-Term Memory Network (Journal of Geophysical Research: Machine Learning and Computation, via Jordi Bolibar's page) ↗Source for the Swiss LSTM results.
- Physics-constrained generative machine learning-based high-resolution downscaling of Greenland's surface mass balance and surface temperature (The Cryosphere) ↗Source for the 32 times downscaling and the scores.
- Machine-learned global glacier ice volumes (arXiv, IceBoost v2.0) ↗Source for the IceBoost training data, method and global volume.
- Welcome to the ICEBOOST v2.0 viewer ↗Web viewer for the IceBoost v2.0 outputs.
Watch next
- Read the Swiss LSTM paper for its comparison with earlier models. Check whether the Greenland model has been tested against observations rather than MAR output.
Sources
- Glacier Mass Balance Modeling Using a Long Short-Term Memory Network (Journal of Geophysical Research: Machine Learning and Computation, via Jordi Bolibar's page)jordibolibar.org
- Physics-constrained generative machine learning-based high-resolution downscaling of Greenland's surface mass balance and surface temperature (The Cryosphere)tc.copernicus.org
- Machine-learned global glacier ice volumes (arXiv, IceBoost v2.0)arxiv.org
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 3 October 2026 at 23:02 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →