Bochow et al. (2026) Physics-constrained generative machine learning-based high-resolution downscaling of Greenland's surface mass balance and surface temperature

Identification

• Journal: ˜The œcryosphere
• Year: 2026
• Date: 2026-03-30
• Authors: Nils Bochow, Philipp Hess, Alexander Robinson
• DOI: 10.5194/tc-20-1841-2026

Research Groups

• Alfred Wegener Institute, Helmholtz Centre for Polar and Marine Research, Potsdam, Germany
• Potsdam Institute for Climate Impact Research, Potsdam, Germany
• Department of Mathematics and Statistics, Faculty of Science and Technology, UiT The Arctic University of Norway, Tromsø, Norway
• Munich Climate Center and Earth System Modelling Group, Department of Aerospace and Geodesy, Technical University of Munich, Munich, Germany

Short Summary

This study introduces a novel physics-constrained generative machine learning framework, based on consistency models, to downscale Greenland's surface mass balance (SMB) and surface temperature (T_s) fields by a factor of up to 32 (from 160 km to 5 km grid spacing). The method ensures physical conservation during inference, enabling robust generalization to extreme climate states and providing realistic, high-resolution climate forcing for ice-sheet simulations with fast computational efficiency.

Objective

• To develop and evaluate a physics-constrained generative machine learning framework for high-resolution downscaling of Greenland's surface mass balance (SMB) and surface temperature (T_s) fields.
• To enable efficient generation of realistic, high-resolution climate forcing for ice-sheet simulations, capable of generalizing to extreme climate conditions and integrating into Earth System Model (ESM) workflows.

Study Configuration

• Spatial Scale: Downscaling from low-resolution fields (e.g., 160 km or 80 km grid spacing) to a high-resolution 5 km grid spacing for the Greenland Ice Sheet.
• Temporal Scale: Monthly outputs; training data from 1950–2100; projections and validation extending to 2100 (NorESM2) and 2300 (CESM2-WACCM).

Methodology and Data

• Models used:
  • Consistency Model (CM): A generative machine learning architecture based on a U-Net backbone, trained to learn a direct mapping from noisy inputs to clean data.
  • Regional Climate Model (RCM): MARv3.12 (Modèle Atmosphérique Régional) for generating high-resolution training data and comparison.
  • Earth System Models (ESMs): NorESM2 (Norwegian Earth System Model) and CESM2-WACCM (Community Earth System Model version 2 with Whole Atmosphere Community Climate Model) for downscaling applications.
  • Bias Correction: Quantile Data Mapping (QDM) using the xclim Python library.
  • Physics-based Model: Simple Positive Degree Day (PDD) model (PyPDD implementation) for a hybrid downscaling approach.
• Data sources:
  • Monthly outputs of SMB and Ts from MARv3.12 simulations (forced by CMIP5 and CMIP6 models) over 1950–2100 (21,432 monthly fields).
  • Bias-corrected SMB (approximated as sum of precipitation, evaporation, runoff) and Ts fields from NorESM2 (SSP-5.85 scenario) and CESM2-WACCM (extended high-emission SSP-5.85 scenario).
  • Auxiliary conditioning inputs: Ice sheet height (topography) and spatially-varying monthly mean insolation.

Main Results

• The physics-constrained Consistency Model (CM) successfully downscales SMB and T_s fields for the Greenland Ice Sheet by a factor of up to 32 (from 160 km to 5 km resolution).
• On the test set, the constrained CM achieves a continuous ranked probability score (CRPS) of 5.37 mm w.e. per month for SMB and 0.1 K for surface temperature, outperforming interpolation-based downscaling.
• Spatial power-spectral analysis demonstrates that the CM faithfully reproduces variability across spatial scales, with the hard-constrained field's power spectral density (PSD) being indistinguishable from the ground truth.
• The hard-constrained CM shows a mean absolute error (MAE) of 9.9 mm w.e. per month and a Pearson correlation (r) of 0.98 for SMB, and MAE of 0.19 K and r of 1.00 for T_s (for 16x downscaling).
• The physics-constrained approach ensures approximate preservation of SMB and temperature sums on the coarse spatial scale, enabling robust generalization to extreme climate states (e.g., high-emission scenarios beyond the training period) without retraining.
• Downscaling 85 years of monthly SMB and surface temperature fields with the constrained CM takes approximately 30 minutes on a single NVIDIA H100 Tensor Core GPU, demonstrating high computational efficiency.
• The model can directly downscale bias-corrected ESM fields (NorESM2, CESM2-WACCM) and can be used in a hybrid manner with physics-based models like the PDD model.

Contributions

• Introduction of a novel physics-constrained generative modeling framework (Consistency Model) for high-resolution downscaling of Greenland's SMB and T_s, significantly advancing beyond previous statistical or unconstrained ML methods.
• Achieves an unprecedented downscaling factor of up to 32 (from 160 km to 5 km), quadrupling the resolution gain of prior generative downscaling work.
• Implements hard conservation constraints during inference, ensuring physical consistency (mass/energy conservation) and robust generalization to out-of-sample extreme climate conditions without the need for retraining.
• Provides a computationally fast inference method for generating realistic, high-resolution climate forcing, making it suitable for integration into Earth-system and ice-sheet model workflows.
• Demonstrates the model's ability to directly downscale outputs from various sources, including regional climate models (MAR) and Earth System Models (NorESM2, CESM2-WACCM), and in a hybrid setup with a PDD model.
• The approach does not explicitly prescribe elevation-SMB relationships, offering greater flexibility and potential applicability to other ice sheets or glaciers.

Funding

• UiT Aurora Centre Program, UiT – The Arctic University of Norway (2020)
• Research Council of Norway (project no. 314570)
• European Union's Horizon Europe research and innovation programme (ClimTip project, grant agreement no. 101137601)
• European Union (ERC, FORCLIMA; grant no. 101044247)
• Ministry of Research, Science and Culture (MWFK) of Land Brandenburg (high performance computer system at Potsdam Institute for Climate Impact Research)
• Alfred-Wegener-Institut Helmholtz-Zentrum für Polar- und Meeresforschung (article processing charges)

Citation

@article{Bochow2026Physicsconstrained,
  author = {Bochow, Nils and Hess, Philipp and Robinson, Alexander},
  title = {Physics-constrained generative machine learning-based high-resolution downscaling of Greenland's surface mass balance and surface temperature},
  journal = {˜The œcryosphere},
  year = {2026},
  doi = {10.5194/tc-20-1841-2026},
  url = {https://doi.org/10.5194/tc-20-1841-2026}
}