Recent warming trends of the Greenland ice sheet documented by historical firn and ice temperature observations and machine learning.

Surface melt on the Greenland ice sheet has been increasing in intensity and extent over the last decades due to Arctic atmospheric warming. Surface melt depends on the surface energy balance, which includes the atmospheric forcing but also the thermal budget of the snow, firn and ice near the ice sheet surface. The temperature of the ice sheet subsurface has been used as an indicator of the thermal state of the ice sheet's surface. Here, we present a compilation of 4612 measurements of firn and ice temperature at 10 m below the surface (T10m) across the ice sheet, spanning from 1912 to 2022. The measurements are either instantaneous or monthly averages. We train an artificial neural network model (ANN) on 4597 of these point observations, weighted by their relative representativity, and use it to reconstruct T10m over the entire Greenland ice sheet for the period 1950–2022 at a monthly timescale. We use 10-year averages and mean annual values of air temperature and snowfall from the ERA5 reanalysis dataset as model input. The ANN indicates a Greenland-wide positive trend of T10m at 0.2 ∘ C per decade during the 1950–2022 period, with a cooling during 1950–1985 (-0.4 ∘ C per decade) followed by a warming during 1985–2022 (+0.7 ∘ per decade). Regional climate models HIRHAM5, RACMO2.3p2 and MARv3.12 show mixed results compared to the observational T10m dataset, with mean differences ranging from -0.4 ∘ C (HIRHAM) to 1.2 ∘ C (MAR) and root mean squared differences ranging from 2.8 ∘ C (HIRHAM) to 4.7 ∘ C (MAR). The observation-based ANN also reveals an underestimation of the subsurface warming trends in climate models for the bare-ice and dry-snow areas. The subsurface warming brings the Greenland ice sheet surface closer to the melting point, reducing the amount of energy input required for melting. Our compilation documents the response of the ice sheet subsurface to atmospheric warming and will enable further improvements of models used for ice sheet mass loss assessment and reduce the uncertainty in projections. [ABSTRACT FROM AUTHOR]