pith. sign in

arxiv: 2507.22485 · v1 · submitted 2025-07-30 · ⚛️ physics.geo-ph · cs.AI

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

Nils Bochow , Philipp Hess , Alexander Robinson This is my paper

Pith reviewed 2026-05-19 03:10 UTC · model grok-4.3

classification ⚛️ physics.geo-ph cs.AI
keywords Greenland ice sheetsurface mass balancedownscalingconsistency modelphysics-constrainedgenerative modelingsea level riseregional climate
0
0 comments X

The pith

A physics-constrained consistency model downscales Greenland surface mass balance and temperature from 160 km to 5 km while preserving coarse-scale conservation laws.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a generative machine learning method to produce high-resolution maps of Greenland ice sheet surface mass balance and temperature. It trains a consistency model on regional climate simulations and conditions it on topography and solar input. The key step is enforcing a hard conservation rule at inference time so that totals of mass and temperature match the original coarse fields. This matters for sea-level projections because current downscaling is either too slow for large ensembles or loses fine-scale detail. The resulting model runs quickly, reproduces observed spatial variability, and works on inputs from global Earth system models.

Core claim

The authors introduce a consistency model for downscaling surface mass balance and surface temperature fields by a factor of up to 32. The model is trained on monthly outputs of the MAR regional climate model and conditioned on ice-sheet topography and insolation. Enforcing a hard conservation constraint during inference approximately preserves the coarse-scale sums of surface mass balance and temperature. This property also supports generalization to extreme climate states without retraining. On held-out test data the constrained model records a continuous ranked probability score of 6.31 mm water equivalent for surface mass balance and 0.1 K for temperature, outperforming interpolation, as

What carries the argument

Consistency model with a hard conservation constraint enforced only at inference time that maintains approximate preservation of integrated surface mass balance and temperature on the coarse grid.

If this is right

  • The downscaled fields supply realistic high-resolution forcing for ice-sheet simulations.
  • Fast inference makes the method practical to embed inside Earth-system and ice-sheet model workflows.
  • The same framework can directly downscale bias-corrected fields from global Earth system models such as NorESM2.
  • Spatial power spectra of the downscaled output match the training data across scales.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same conservation-enforced generative approach could be tested on other ice sheets or climate variables where integrated budgets must be respected.
  • The speed of the method opens the possibility of running large ensembles of high-resolution Greenland projections that were previously too expensive.
  • Because retraining is not required for extreme states, the model may prove useful for paleoclimate reconstructions or very high-emission scenarios.

Load-bearing premise

That enforcing the conservation constraint only during inference is enough to make the model generalize reliably to climate conditions far outside the training distribution.

What would settle it

Run the downscaled fields through a high-resolution regional climate model or compare them against independent observations for periods with climate states more extreme than the training data and check whether the coarse-scale totals remain conserved and the fine-scale patterns stay realistic.

Figures

Figures reproduced from arXiv: 2507.22485 by Alexander Robinson, Nils Bochow, Philipp Hess.

Figure 1
Figure 1. Figure 1: Exemplary downscaling of surface mass balance and temperature at surface for random month from test set. (a) Random month from held-out test set pooled with a factor of 16 (80 km) resolution. This serves as input to our model. (b) Downscaled SMB field with the CM model using hard constraints to 5 km resolution. The small scale variability is recovered, especially at the ice-sheet margins. (c) High-resoluti… view at source ↗
Figure 2
Figure 2. Figure 2: Schematic depiction of the consistency model (CM) workflow. We train the CM on the high-resolution RCM MARv3.12 with topography and insolation as auxiliary input. During inference, the low resolution input fields are noised with an adaptive noising strength and subsequently denoised by the trained CM with conservative constraints. The small-scale variability is recovered by the CM. 10 [PITH_FULL_IMAGE:fig… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of downscaled SMB and Ts fields. (a) Downscaled SMB field of test set with hard constraints during inference for warm month with negative SMB at the margins (blue) with a 5 km resolution. The mean absolute error (in mmWE) and Pearson correlation with the GT field are denoted. (b) Same as (a) but unconstrained CM. (c) Coarsened MAR field by a factor of 16 (80 km resolution), which is … view at source ↗
Figure 4
Figure 4. Figure 4: Normalized power spectral density (PSD) for different noise scales. The mean radially averaged normalized PSD for different noising scales is shown. We downscale the test set that was coarsened by a factor of 16 and linearly interpolated using the CM (unconstrained). Additionally, the PSD of an unconditional ensemble is shown (red). The coarsened field (orange) generally underestimates the small scale vari… view at source ↗
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Continuous ranked probability scores (CRPS) for different noise scales over the test set for SMB and Ts. (a) Spatial mean CRPS for the noising time t = 1. For each month of the test set, an ensemble of 50 realizations is generated and compared to the corresponding ground truth. The spatiotemporal mean CRPS is 8.67 mmWE/month. The CRPS is largest at the margins of the ice sheet, where the variability is lar… view at source ↗
read the original abstract

Accurate, high-resolution projections of the Greenland ice sheet's surface mass balance (SMB) and surface temperature are essential for understanding future sea-level rise, yet current approaches are either computationally demanding or limited to coarse spatial scales. Here, we introduce a novel physics-constrained generative modeling framework based on a consistency model (CM) to downscale low-resolution SMB and surface temperature fields by a factor of up to 32 (from 160 km to 5 km grid spacing) in a few sampling steps. The CM is trained on monthly outputs of the regional climate model MARv3.12 and conditioned on ice-sheet topography and insolation. By enforcing a hard conservation constraint during inference, we ensure approximate preservation of SMB and temperature sums on the coarse spatial scale as well as robust generalization to extreme climate states without retraining. On the test set, our constrained CM achieves a continued ranked probability score of 6.31 mmWE for the SMB and 0.1 K for the surface temperature, outperforming interpolation-based downscaling. Together with spatial power-spectral analysis, we demonstrate that the CM faithfully reproduces variability across spatial scales. We further apply bias-corrected outputs of the NorESM2 Earth System Model as inputs to our CM, to demonstrate the potential of our model to directly downscale ESM fields. Our approach delivers realistic, high-resolution climate forcing for ice-sheet simulations with fast inference and can be readily integrated into Earth-system and ice-sheet model workflows to improve projections of the future contribution to sea-level rise from Greenland and potentially other ice sheets and glaciers too.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces a physics-constrained consistency model (CM) for downscaling Greenland surface mass balance (SMB) and surface temperature fields from 160 km to 5 km resolution (factor of up to 32). Trained on monthly MARv3.12 regional climate model outputs and conditioned on ice-sheet topography and insolation, the framework enforces a hard conservation constraint at inference to preserve coarse-scale sums and enable generalization to extreme states without retraining. On the MARv3.12 test set it reports CRPS of 6.31 mmWE for SMB and 0.1 K for temperature, outperforming interpolation; spatial power-spectral analysis is used to show faithful reproduction of variability across scales. The model is further applied to bias-corrected NorESM2 Earth-system model fields to illustrate direct ESM downscaling.

Significance. If the central performance and generalization claims hold, the work would provide a fast, integrable method for generating high-resolution climate forcing suitable for ice-sheet simulations, with potential to improve sea-level rise projections from Greenland and other ice masses. The use of a consistency model for few-step sampling and the explicit hard constraint are clear technical strengths that distinguish the approach from standard generative or interpolation baselines.

major comments (2)
  1. [Abstract] Abstract: The claim that the hard conservation constraint 'ensure[s] ... robust generalization to extreme climate states without retraining' is load-bearing for the paper's applicability to ESM inputs such as NorESM2, yet no quantitative out-of-distribution tests (e.g., amplified warming or extreme SMB anomaly regimes) are reported. All numerical metrics (CRPS, spectral analysis) are confined to the in-distribution MARv3.12 test set, leaving the robustness assertion unsupported by direct evidence.
  2. [Methods] Methods / Results: The exact mathematical form of the hard conservation constraint, its enforcement mechanism during the few-step sampling, and any ablation studies isolating its contribution are not provided. Without these details it is impossible to verify whether the constraint merely preserves coarse sums on in-distribution data or genuinely improves extrapolation, which directly affects the soundness of the generalization claim.
minor comments (2)
  1. The abstract and results sections would benefit from explicit reporting of uncertainty (error bars or standard deviations) on the CRPS values and from a quantitative metric (e.g., integrated spectral error) to accompany the power-spectral plots.
  2. Training hyperparameters, data-split details, and the precise conditioning procedure on topography/insolation should be expanded for reproducibility, ideally with a supplementary table or pseudocode.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough review and constructive feedback on our manuscript. We address each of the major comments in detail below and outline the revisions we plan to make.

read point-by-point responses

  1. Referee: [Abstract] The claim that the hard conservation constraint 'ensure[s] ... robust generalization to extreme climate states without retraining' is load-bearing for the paper's applicability to ESM inputs such as NorESM2, yet no quantitative out-of-distribution tests (e.g., amplified warming or extreme SMB anomaly regimes) are reported. All numerical metrics (CRPS, spectral analysis) are confined to the in-distribution MARv3.12 test set, leaving the robustness assertion unsupported by direct evidence.

    Authors: We appreciate the referee highlighting this point regarding the generalization claim. The NorESM2 application provides an initial demonstration on inputs from a different distribution, and the hard constraint is designed to support physical consistency for such cases. However, we agree that dedicated quantitative out-of-distribution tests would strengthen the evidence. In the revised manuscript we will add a new subsection with tests on amplified warming scenarios and extreme SMB anomaly regimes to directly evaluate robustness. revision: yes

  2. Referee: [Methods] The exact mathematical form of the hard conservation constraint, its enforcement mechanism during the few-step sampling, and any ablation studies isolating its contribution are not provided. Without these details it is impossible to verify whether the constraint merely preserves coarse sums on in-distribution data or genuinely improves extrapolation, which directly affects the soundness of the generalization claim.

    Authors: We agree that these technical details are necessary for full reproducibility and to substantiate the claims. In the revised manuscript we will add the exact mathematical formulation of the hard conservation constraint to the Methods section, describe its enforcement mechanism within the few-step sampling procedure of the consistency model, and include ablation studies that compare constrained versus unconstrained variants on both in-distribution and out-of-distribution data. revision: yes

Circularity Check

0 steps flagged

No significant circularity; framework relies on empirical training and held-out evaluation

full rationale

The paper describes training a consistency model on monthly MARv3.12 regional climate outputs, conditioning on topography and insolation, then enforcing a hard conservation constraint only at inference time to preserve coarse-scale sums. Performance is reported via CRPS scores and power-spectral analysis on a test split from the same data source, with a separate qualitative demonstration on bias-corrected NorESM2 fields. These steps constitute standard supervised learning and post-hoc constraint application rather than any self-definitional loop, fitted parameter renamed as prediction, or load-bearing self-citation chain. The claimed generalization to extreme states is an assertion supported by the constraint mechanism and external ESM input, not a reduction to the training inputs by construction. The overall derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on training data from MARv3.12 outputs and the assumption that a generative model can learn scale-consistent mappings when conditioned on topography and insolation while obeying conservation.

free parameters (1)
  • consistency model hyperparameters
    The generative model contains numerous parameters learned during training on MAR data; exact count and selection process not specified in abstract.
axioms (2)
  • domain assumption Hard conservation of SMB and temperature sums must hold on the coarse scale after downscaling
    Invoked during inference to preserve coarse-scale totals as described in the abstract.
  • domain assumption Training on MARv3.12 monthly outputs is representative for generalization to other climate states
    Underlying the claim of robust generalization without retraining.

pith-pipeline@v0.9.0 · 5818 in / 1447 out tokens · 52250 ms · 2026-05-19T03:10:27.444876+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages · 3 internal anchors

  1. [1]

    Morlighem, M. et al. BedMachine v3: Complete Bed Topography and Ocean Bathymetry Mapping of Greenland From Multibeam Echo Sounding Combined With Mass Conservation. en. Geophysical Research Letters 44, 11, 051–11, 061. ISSN : 1944-8007. doi:10.1002/2017GL074954 (2017)

    work page doi:10.1002/2017gl074954 1944

  2. [2]

    Trusel, L. D. et al. Nonlinear rise in Greenland runoff in response to post-industrial Arctic warming. en. Nature 564, 104–108. ISSN : 1476-4687. doi:10.1038/s41586-018-0752-4 (Dec. 2018)

    work page doi:10.1038/s41586-018-0752-4 2018

  3. [3]

    Van den Broeke, M. R.et al. On the recent contribution of the Greenland ice sheet to sea level change. English. The Cryosphere 10, 1933–1946. ISSN : 1994-0416. doi:10.5194/tc-10-1933-2016 (Sept. 2016)

    work page doi:10.5194/tc-10-1933-2016 1933

  4. [4]

    & Wood, M

    Choi, Y ., Morlighem, M., Rignot, E. & Wood, M. Ice dynamics will remain a primary driver of Greenland ice sheet mass loss over the next century. en. Communications Earth & Environment 2, 1–9. ISSN : 2662-4435. doi:10.1038/s43247-021-00092-z (Feb. 2021)

    work page doi:10.1038/s43247-021-00092-z 2021

  5. [5]

    Goelzer, H. et al. The future sea-level contribution of the Greenland ice sheet: a multi-model ensemble study of ISMIP6. English. The Cryosphere 14, 3071–3096. ISSN : 1994-0416. doi:10.5194/tc-14-3071-2020 (Sept. 2020)

    work page doi:10.5194/tc-14-3071-2020 1994

  6. [6]

    Payne, A. J. et al. Future Sea Level Change Under Coupled Model Intercomparison Project Phase 5 and Phase 6 Scenarios From the Greenland and Antarctic Ice Sheets. en. Geophysical Research Letters 48, e2020GL091741. ISSN : 1944-8007. doi:10.1029/2020GL091741 (2021)

    work page doi:10.1029/2020gl091741 1944

  7. [7]

    Very high resolution regional climate model simulations over Greenland: Identifying added value

    Lucas-Picher, P.et al. Very high resolution regional climate model simulations over Greenland: Identifying added value. en. Journal of Geophysical Research: Atmospheres 117. ISSN : 2156-2202. doi:10.1029/2011JD016267 (2012)

    work page doi:10.1029/2011jd016267 2012

  8. [8]

    Mottram, R. et al. Surface mass balance of the Greenland ice sheet in the regional climate model HIRHAM5: Present state and future prospects. Low Temperature Science75, 105–115 (Mar. 2017)

    work page 2017

  9. [9]

    Fettweis, X. et al. Estimating the Greenland ice sheet surface mass balance contribution to future sea level rise using the regional atmospheric climate model MAR. English. The Cryosphere 7, 469–489. ISSN : 1994-0416. doi:10.5194/tc-7-469-2013 (Mar. 2013)

    work page doi:10.5194/tc-7-469-2013 1994

  10. [10]

    Fettweis, X. et al. Reconstructions of the 1900–2015 Greenland ice sheet surface mass balance using the regional climate MAR model. English. The Cryosphere 11, 1015–1033. ISSN : 1994-0416. doi:10.5194/tc-11-1015- 2017 (Apr. 2017)

    work page doi:10.5194/tc-11-1015- 1900

  11. [11]

    Van Dalum, C. T.et al. First results of the polar regional climate model RACMO2.4. English. The Cryosphere 18, 4065–4088. ISSN : 1994-0416. doi:10.5194/tc-18-4065-2024 (Sept. 2024)

    work page doi:10.5194/tc-18-4065-2024 1994

  12. [12]

    Aschwanden, A. et al. Contribution of the Greenland Ice Sheet to sea level over the next millennium. Science Advances 5, eaav9396. doi:10.1126/sciadv.aav9396 (June 2019)

    work page doi:10.1126/sciadv.aav9396 2019

  13. [13]

    Garbe, J., Albrecht, T., Levermann, A., Donges, J. F. & Winkelmann, R. The hysteresis of the Antarctic Ice Sheet. en. Nature 585, 538–544. ISSN : 1476-4687. doi:10.1038/s41586-020-2727-5 (Sept. 2020)

    work page doi:10.1038/s41586-020-2727-5 2020

  14. [14]

    & Winkelmann, R

    Beckmann, J. & Winkelmann, R. Effects of extreme melt events on ice flow and sea level rise of the Greenland Ice Sheet. English. The Cryosphere 17, 3083–3099. ISSN : 1994-0416. doi:10.5194/tc-17-3083-2023 (July 2023)

    work page doi:10.5194/tc-17-3083-2023 1994

  15. [15]

    & Winkelmann, R

    Zeitz, M., Reese, R., Beckmann, J., Krebs-Kanzow, U. & Winkelmann, R. Impact of the melt–albedo feedback on the future evolution of the Greenland Ice Sheet with PISM-dEBM-simple. English.The Cryosphere 15, 5739–5764. ISSN : 1994-0416. doi:10.5194/tc-15-5739-2021 (Dec. 2021)

    work page doi:10.5194/tc-15-5739-2021 1994

  16. [16]

    Bochow, N. et al. Overshooting the critical threshold for the Greenland ice sheet. en. Nature 622, 528–536. ISSN : 1476-4687. doi:10.1038/s41586-023-06503-9 (Oct. 2023)

    work page doi:10.1038/s41586-023-06503-9 2023

  17. [17]

    & Boers, N

    Bochow, N., Poltronieri, A. & Boers, N. Projections of precipitation and temperatures in Greenland and the impact of spatially uniform anomalies on the evolution of the ice sheet. English. The Cryosphere 18, 5825–5863. ISSN : 1994-0416. doi:10.5194/tc-18-5825-2024 (Dec. 2024)

    work page doi:10.5194/tc-18-5825-2024 1994

  18. [18]

    Seroussi, H. et al. ISMIP6 Antarctica: a multi-model ensemble of the Antarctic ice sheet evolution over the 21st century. English. The Cryosphere 14, 3033–3070. ISSN : 1994-0416. doi:10.5194/tc-14-3033-2020 (Sept. 2020)

    work page doi:10.5194/tc-14-3033-2020 1994

  19. [19]

    Nowicki, S. et al. Experimental protocol for sea level projections from ISMIP6 stand-alone ice sheet models. English. The Cryosphere 14, 2331–2368. ISSN : 1994-0416. doi:10.5194/tc-14-2331-2020 (July 2020)

    work page doi:10.5194/tc-14-2331-2020 1994

  20. [20]

    Sellevold, R. et al. Surface mass balance downscaling through elevation classes in an Earth system model: application to the Greenland ice sheet. English. The Cryosphere 13, 3193–3208. ISSN : 1994-0416. doi:10.5194/ tc-13-3193-2019 (Dec. 2019). xi Physics-Constrained Generative Downscaling for Greenland PREPRINT

    work page 1994

  21. [21]

    Noël, B. et al. A daily, 1 km resolution data set of downscaled Greenland ice sheet surface mass balance (1958–2015). English. The Cryosphere 10, 2361–2377. ISSN : 1994-0416. doi:10.5194/tc-10-2361-2016 (Oct. 2016)

    work page doi:10.5194/tc-10-2361-2016 1958

  22. [22]

    & Cervone, G

    Tedesco, M., Colosio, P., Fettweis, X. & Cervone, G. A computationally efficient statistically downscaled 100 m resolution Greenland product from the regional climate model MAR. English. The Cryosphere 17, 5061–5074. ISSN : 1994-0416. doi:10.5194/tc-17-5061-2023 (Nov. 2023)

    work page doi:10.5194/tc-17-5061-2023 1994

  23. [23]

    & Boers, N

    Bochow, N., Poltronieri, A., Rypdal, M. & Boers, N. Reconstructing historical climate fields with deep learning. Science Advances 11, eadp0558. doi:10.1126/sciadv.adp0558 (Apr. 2025)

    work page doi:10.1126/sciadv.adp0558 2025

  24. [24]

    Lam, R. et al. Learning skillful medium-range global weather forecasting.Science 382, 1416–1421. doi:10.1126/ science.adi2336 (Dec. 2023)

    work page 2023

  25. [25]

    P., and Hoyer, S.: Neural general circulation models for weather and climate, Nature, 632, 1060–1066, https://doi.org/10.1038/s41586-024-07744-y,

    Kochkov, D.et al. Neural general circulation models for weather and climate. en. Nature 632, 1060–1066. ISSN : 1476-4687. doi:10.1038/s41586-024-07744-y (Aug. 2024)

    work page doi:10.1038/s41586-024-07744-y 2024

  26. [26]

    & Boers, N

    Hess, P., Aich, M., Pan, B. & Boers, N. Fast, scale-adaptive and uncertainty-aware downscaling of Earth system model fields with generative machine learning. en. Nature Machine Intelligence 7, 363–373. ISSN : 2522-5839. doi:10.1038/s42256-025-00980-5 (Mar. 2025)

    work page doi:10.1038/s42256-025-00980-5 2025

  27. [27]

    Bi, K. et al. Accurate medium-range global weather forecasting with 3D neural networks. en.Nature 619, 533–538. ISSN : 1476-4687. doi:10.1038/s41586-023-06185-3 (July 2023)

    work page doi:10.1038/s41586-023-06185-3 2023

  28. [28]

    Price, I. et al. Probabilistic weather forecasting with machine learning. en. Nature 637, 84–90. ISSN : 1476-4687. doi:10.1038/s41586-024-08252-9 (Jan. 2025)

    work page doi:10.1038/s41586-024-08252-9 2025

  29. [29]

    Watt-Meyer, O. et al. ACE2: accurately learning subseasonal to decadal atmospheric variability and forced responses. en. npj Climate and Atmospheric Science 8, 205. ISSN : 2397-3722. doi: 10.1038/s41612- 025- 01090-0 (May 2025)

    work page doi:10.1038/s41612- 2025

  30. [30]

    Jouvet, G. et al. Deep learning speeds up ice flow modelling by several orders of magnitude. en. Journal of Glaciology 68, 651–664. ISSN : 0022-1430, 1727-5652. doi:10.1017/jog.2021.120 (Aug. 2022)

    work page doi:10.1017/jog.2021.120 2021

  31. [31]

    & Cordonnier, G

    Jouvet, G. & Cordonnier, G. Ice-flow model emulator based on physics-informed deep learning. en.Journal of Glaciology 69, 1941–1955. ISSN : 0022-1430, 1727-5652. doi:10.1017/jog.2023.73 (Dec. 2023)

    work page doi:10.1017/jog.2023.73 1941

  32. [32]

    Wang, Y ., Lai, C.-Y ., Prior, D. J. & Cowen-Breen, C. Deep learning the flow law of Antarctic ice shelves.Science 387, 1219–1224. doi:10.1126/science.adp3300 (Mar. 2025)

    work page doi:10.1126/science.adp3300 2025

  33. [33]

    Rosier, S. H. R., Bull, C. Y . S., Woo, W. L. & Gudmundsson, G. H. Predicting ocean-induced ice-shelf melt rates using deep learning. English. The Cryosphere 17, 499–518. ISSN : 1994-0416. doi:10.5194/tc-17-499-2023 (Feb. 2023)

    work page doi:10.5194/tc-17-499-2023 1994

  34. [34]

    Bolibar, J. et al. Deep learning applied to glacier evolution modelling. English. The Cryosphere 14, 565–584. ISSN : 1994-0416. doi:10.5194/tc-14-565-2020 (Feb. 2020)

    work page doi:10.5194/tc-14-565-2020 1994

  35. [35]

    A minimal machine-learning glacier mass balance model

    Van der Meer, M.et al. A minimal machine-learning glacier mass balance model. English. The Cryosphere 19, 805–826. ISSN : 1994-0416. doi:10.5194/tc-19-805-2025 (Feb. 2025)

    work page doi:10.5194/tc-19-805-2025 1994

  36. [36]

    Steidl, V ., Bamber, J. L. & Zhu, X. X. Physics-aware machine learning for glacier ice thickness estimation: a case study for Svalbard. English. The Cryosphere 19, 645–661. ISSN : 1994-0416. doi:10.5194/tc-19-645-2025 (Feb. 2025)

    work page doi:10.5194/tc-19-645-2025 1994

  37. [37]

    & Robel, A

    Verjans, V . & Robel, A. Accelerating Subglacial Hydrology for Ice Sheet Models With Deep Learning Methods. en. Geophysical Research Letters 51, e2023GL105281. ISSN : 1944-8007. doi:10.1029/2023GL105281 (2024)

    work page doi:10.1029/2023gl105281 1944

  38. [38]

    Harris, L., McRae, A. T. T., Chantry, M., Dueben, P. D. & Palmer, T. N. A Generative Deep Learning Approach to Stochastic Downscaling of Precipitation Forecasts. en. Journal of Advances in Modeling Earth Systems 14, e2022MS003120. ISSN : 1942-2466. doi:10.1029/2022MS003120 (2022)

    work page doi:10.1029/2022ms003120 1942

  39. [39]

    & Boers, N

    Hess, P., Lange, S., Schötz, C. & Boers, N. Deep Learning for Bias-Correcting CMIP6-Class Earth System Models. en. Earth’s Future11, e2023EF004002. ISSN : 2328-4277. doi:10.1029/2023EF004002 (2023)

    work page doi:10.1029/2023ef004002 2023

  40. [40]

    Aich, M. et al. Conditional diffusion models for downscaling & bias correction of Earth system model precipitation Apr. 2024. doi:10.48550/arXiv.2404.14416

    work page doi:10.48550/arxiv.2404.14416 2024

  41. [41]

    Towards Principled Methods for Training Generative Adversarial Networks

    Arjovsky, M. & Bottou, L. Towards Principled Methods for Training Generative Adversarial NetworksJan. 2017. doi:10.48550/arXiv.1701.04862

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1701.04862 2017

  42. [42]

    & Deck, K

    Bischoff, T. & Deck, K. Unpaired Downscaling of Fluid Flows with Diffusion Bridges. EN.Artificial Intelligence for the Earth Systems 3. ISSN : 2769-7525. doi:10.1175/AIES-D-23-0039.1 (May 2024)

    work page doi:10.1175/aies-d-23-0039.1 2024

  43. [43]

    & Sutskever, I.Consistency Models May 2023

    Song, Y ., Dhariwal, P., Chen, M. & Sutskever, I.Consistency Models May 2023. doi:10.48550/arXiv.2303. 01469

    work page doi:10.48550/arxiv.2303 2023

  44. [44]

    & Boers, N.Diffusion models for probabilistic precipitation generation from atmospheric variables Apr

    Aich, M., Bathiany, S., Hess, P., Huang, Y . & Boers, N.Diffusion models for probabilistic precipitation generation from atmospheric variables Apr. 2025. doi:10.48550/arXiv.2504.00307. xii Physics-Constrained Generative Downscaling for Greenland PREPRINT

    work page doi:10.48550/arxiv.2504.00307 2025

  45. [45]

    Glaude, Q. et al. A Factor Two Difference in 21st-Century Greenland Ice Sheet Surface Mass Balance Projections From Three Regional Climate Models Under a Strong Warming Scenario (SSP5-8.5). en. Geophysical Research Letters 51, e2024GL111902. ISSN : 1944-8007. doi:10.1029/2024GL111902 (2024)

    work page doi:10.1029/2024gl111902 1944

  46. [46]

    Hard-constrained deep learning for climate downscaling

    Harder, P.et al. Hard-constrained deep learning for climate downscaling. J. Mach. Learn. Res. 24, 365:17534– 365:17573. ISSN : 1532-4435 (July 2023)

    work page 2023

  47. [47]

    & Fettweis, X

    Delhasse, A., Beckmann, J., Kittel, C. & Fettweis, X. Coupling MAR (Modèle Atmosphérique Régional) with PISM (Parallel Ice Sheet Model) mitigates the positive melt–elevation feedback. English. The Cryosphere 18, 633–651. ISSN : 1994-0416. doi:10.5194/tc-18-633-2024 (Feb. 2024)

    work page doi:10.5194/tc-18-633-2024 1994

  48. [48]

    & Abbeel, P

    Ho, J., Jain, A. & Abbeel, P. Denoising Diffusion Probabilistic Models Dec. 2020. doi:10.48550/arXiv.2006. 11239

    work page doi:10.48550/arxiv.2006 2020

  49. [49]

    Score-Based Generative Modeling through Stochastic Differential Equations

    Song, Y .et al. Score-Based Generative Modeling through Stochastic Differential EquationsFeb. 2021. doi:10. 48550/arXiv.2011.13456

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  50. [50]

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric

    Zhang, R., Isola, P., Efros, A. A., Shechtman, E. & Wang, O.The Unreasonable Effectiveness of Deep Features as a Perceptual Metric Apr. 2018. doi:10.48550/arXiv.1801.03924

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1801.03924 2018

  51. [51]

    xclim: xarray-based climate data analytics

    Bourgault, P.et al. xclim: xarray-based climate data analytics. en. Journal of Open Source Software 8, 5415. ISSN : 2475-9066. doi:10.21105/joss.05415 (May 2023)

    work page doi:10.21105/joss.05415 2023

  52. [52]

    J., Sobie, S

    Cannon, A. J., Sobie, S. R. & Murdock, T. Q. Bias Correction of GCM Precipitation by Quantile Mapping: How Well Do Methods Preserve Changes in Quantiles and Extremes? EN.Journal of Climate 28, 6938–6959. ISSN : 0894-8755, 1520-0442. doi:10.1175/JCLI-D-14-00754.1 (Sept. 2015)

    work page doi:10.1175/jcli-d-14-00754.1 2015

  53. [53]

    PyPDD: a positive degree day model for glacier surface mass balance Oct

    Seguinot, J. PyPDD: a positive degree day model for glacier surface mass balance Oct. 2019. doi:10.5281/ zenodo.3467639

    work page 2019

  54. [54]

    Spatial and seasonal effects of temperature variability in a positive degree-day glacier surface mass-balance model

    Seguinot, J. Spatial and seasonal effects of temperature variability in a positive degree-day glacier surface mass-balance model. en. Journal of Glaciology 59, 1202–1204. ISSN : 0022-1430, 1727-5652. doi: 10.3189/ 2013JoG13J081 (Jan. 2013)

    work page 2013

  55. [55]

    & Greve, R

    Calov, R. & Greve, R. A semi-analytical solution for the positive degree-day model with stochastic tem- perature variations. en. Journal of Glaciology 51, 173–175. ISSN : 0022-1430, 1727-5652. doi: 10 . 3189 / 172756505781829601 (Jan. 2005)

    work page 2005

  56. [56]

    Crameri, F., Shephard, G. E. & Heron, P. J. The misuse of colour in science communication. en.Nature Communi- cations 11, 5444. ISSN : 2041-1723. doi:10.1038/s41467-020-19160-7 (Oct. 2020). xiii

    work page doi:10.1038/s41467-020-19160-7 2041