Physics-constrained generative machine learning-based high-resolution downscaling of Greenland's surface mass balance and surface temperature

Bochow, Nils; Hess, Philipp; Robinson, Alexander

doi:10.5194/tc-20-1841-2026

Articles | Volume 20, issue 3

https://doi.org/10.5194/tc-20-1841-2026

© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/tc-20-1841-2026

© Author(s) 2026. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 20, issue 3

Research article

|

30 Mar 2026

Research article |

| 30 Mar 2026

Physics-constrained generative machine learning-based high-resolution downscaling of Greenland's surface mass balance and surface temperature

Nils Bochow, Philipp Hess, and Alexander Robinson

Data sets

Supplementary Data (Training & test data, checkpoints) N. Bochow et al. https://doi.org/10.5281/zenodo.18241574

Model code and software

Fast, Scale-Adaptive, and Uncertainty-Aware Downscaling of Earth System Model Fields with Generative Machine Learning, Code Ocean capsule 9094997, version 1 P. Hess et al. https://doi.org/10.24433/CO.2150269.v1

Short summary

This study presents a fast, physics-guided machine-learning method that downscales coarse climate fields to fine resolution while enforcing conservation of large-scale totals. Trained on regional climate simulations and driven by Earth system model output, it handles extremes and outperforms linear interpolation, providing realistic, high-resolution forcing for ice-sheet models and improving projections of Greenland’s sea-level contribution.