AI carries glacier physics to new sites

GlaUnTI adds a neural correction to a temperature-based glacier model, adjusting parameters governing snowfall and melt. Konstantin Maslov’s team used 52 European glaciers for training, 13 for validation and 13 for testing, across records spanning 1995–2024.[1]

The observations begin with fieldwork. The World Glacier Monitoring Service describes measuring ice loss at stakes on the glacier tongue and snow gain higher up, including pits dug to the previous summer’s surface. Snow and ice have different densities, so their thickness changes are converted into water equivalents before comparison. Measurements at individual points must then be extrapolated across the glacier. That makes a glacier-wide balance an estimate built from observations, with an additional spatial calculation.[2]

Temperature-based modelling compresses some of that complexity into a melt factor: how much surface loss accompanies warmth above a chosen threshold. An independent 2024 instrument study by Andrew Wickert and colleagues shows why calibrating that factor demands care. Their automated stake at Argentina’s Perito Moreno glacier recorded surface distance, temperature and humidity every five minutes in February–March 2020. Fitted melt factors changed through the record; the authors identified changing weather and the evolving surface beneath the sensor among possible explanations. A fitted coefficient carries the conditions under which it was measured.[3]

The new study gave conventional physics a local advantage: earlier records from each test glacier. For annual point measurements in 2019–2024, that locally tuned model’s root mean square error was 0.905 metres water equivalent; GlaUnTI, without local retuning, reached 0.883. Its neural correction used geometry, climate and available surface-class maps.[1]

The experiment tests spatial transfer. Geometry came from observations; long-term evolution also requires ice dynamics.[1]

Report an error or suggest a correction