Abstract
Future Greenland ice sheet (GrIS) melt projections are limited by the lack of explicit melt calculations within most global climate models and the high computational cost of dynamical downscaling with regional climate models (RCMs). Here, we train artificial neural networks (ANNs) to obtain relationships between quantities consistently available from global climate model simulations and annually integrated GrIS surface melt. To this end, we train the ANNs with model output from the Community Earth System Model 2.1 (CESM2), which features an interactive surface melt calculation based on a downscaled surface energy balance. We find that ANNs compare well with an independent CESM2 simulation and RCM simulations forced by a CMIP6 subset. The ANNs estimate a melt increase for 2,081–2,100 ranging from 414 ± 275 Gt yr−1 (SSP1‐2.6) to 1,378 ± 555 Gt yr−1 (SSP5‐8.5) for the full CMIP6 suite. The primary source of uncertainty throughout the 21st century is the spread of climate model sensitivity.
Highlights
Greenland is losing mass at an accelerating rate since the 1990s (Bamber et al, 2018; Mouginot et al, 2019; Shepherd et al, 2020) in response to global warming
We find that artificial neural networks (ANNs) compare well with an independent Community Earth System Model 2.1 (CESM2) simulation and regional climate models (RCMs) simulations forced by a Climate Model Intercomparison Project Phase 6 (CMIP6) subset
We use the ANNs to predict melt from five atmospheric variables obtained from an independent CESM2 historical/SSP5-8.5 simulation not used for training or cross-validation and compare to the explicit melt calculation in this simulation
Summary
Greenland is losing mass at an accelerating rate since the 1990s (Bamber et al, 2018; Mouginot et al, 2019; Shepherd et al, 2020) in response to global warming. The classical approach to use climate simulations for estimates of ice sheet surface melt is through positive-degree-day schemes (Braithwaite, 1995; Wake & Marshall, 2015) While these schemes are often used for computational efficiency, their performance is poor when surface melt/ablation is high (Bauer & Ganopolski, 2017), which we can expect for most global warming scenarios. State-of-the-art 21st century projections of GrIS surface melt come from RCMs (Fettweis et al, 2013; Mottram et al, 2017; van Angelen et al, 2013) These models are run at high resolution and with a surface energy balance-based calculation of melt. They are computationally expensive and require external forcing from a global climate model (Noël et al, 2020)
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
