| 17 Aug 2026
Predicting meltwater ponding on Antarctic ice shelves using a firn model emulator
Abstract. Depletion of firn air can lead to meltwater ponding. If conditions are unfavourable, it can cause hydrofracturing and disintegration of Antarctic ice shelves. Projections of firn air content can therefore be used in ice-sheet model simulations to prescribe which ice-shelf sections become vulnerable to collapse, and when. However, using a full firn model for simulating firn air content for a wide range of climate forcings is computationally expensive. We therefore developed an XGBoost emulator for firn air content, a fast machine-learning model trained to reproduce the firn air content from a full firn densification model. The emulator needs annual means of snowfall, surface melt, rain, air temperature, and wind speed as input. Using scenario and spatial blocking cross-evaluation, we find that the emulator predicts > 94 % of the variance in the firn air content of the full firn densification model. When a simulated firn air content of less than 2 m is assumed to represent the onset of melt ponding, the emulator shows 6–7 % false positives and 14–17 % false negatives. We demonstrate that surface melt is an important input variable, by constructing an alternative emulator that does not require surface melt. This emulator explains > 82 % of the variance in firn air content, and has poorer agreement with firn-model simulated melt pond locations. Including surface melt results in a better emulator, because the non-linear behaviour of snow surface albedo is hard to predict from atmospheric variables only.
Competing interests: At least one of the (co-)authors is a member of the editorial board of The Cryosphere.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.
Review for „Predicting meltwater ponding on Antarctic ice shelves using a firn model emulator“ by Veldhuijsen et al. submitted to The Cryosphere.
General comments:
This well-written and clearly structured study presents a newly trained emulator for the semi-empirical 1D firn model IMAU-FDM to simulate firn air depletion across all Antarctic ice shelves in a computationally efficient way. The authors chose a XGBoost approach, and describe the training and validation steps of the emulator in detail, which gives the reader a clear description of the fidelity and limitations of this firn emulator for forecasting firn air depletion. It is identified and highlighted that having access to “surface melting rates” as an input/training variable increases the emulator performance considerably over pure temperature-based input variables, which is explained by the complex nonlinearities and feedbacks of the melt process. By building two emulators, one with, and one without meltwater production as input, the authors test a more commonly available input dataset but highlight that without explicitly inputting “surface melt” in the training dataset, the emulator will train on and reproduce the relationship between temperature and melt rates from the input-generating atmospheric model, here a unidirectional coupling to RACMO2 output, which will then be transferred to any emulator forcing dataset. Thus, the no-melt emulator seems to be mimicking the specific RACMO2-FDM co-performance, much more than the stand-alone FDM performance.
Once trained, the authors briefly present two application cases for the emulators, namely applying them to known cases of ice-shelf collapses in the past, and forcing them with data from other regional climate models to explain the sensitivity to the forcing dataset.
The study does a good job of presenting the intentions, application areas, and limitations of the firn air emulator, and it highlights that emulator performance relies crucially on the appropriate input dataset. I suggest swift publication after only minor reviews and my only major criticism is the misleading title.
The manuscript presents all emulator performance and validation in terms of firn air content (FAC) and, specifically, firn air depletion, a precursor to surface meltwater ponding. However, the authors highlight themselves in paragraph l. 449-455, there are additional complications between firn air content/depletion and actual meltwater ponding, which I expected the study to address in detail based on the title. Instead, what the authors intended and did was predicting “firn air depletion”, which can only be interpreted as indicative of meltwater ponding, especially since the emulated FDM model uses a simple bucket scheme for meltwater percolation, which does not allow for blocking of meltwater flow and standing water on top of ice lenses, which seems to be important when the aim would have been specifically surface meltwater ponding. Thus, I suggest renaming the title to “Predicting firn air depletion on Antarctic ice shelves using a firn model emulator” or similar.
Specific comments:
L. 143: “This value replaces the true average firn air content”. It seems that assigning negative values for the low firn air days is difficult to revert/interpret. How do you interpret emulator predictions of negative values? Why are no negative values shown in Fig 5.? How do you get to “m” of negative firn air content in Fig 8 a and b if the emulator can only predict THAT the firn has experienced 5 consecutive days of less than 0.1m? The choices for these “artificial negative values” also seem to be arbitrary. Could you elaborate on the decision to score “-1 for every 5 days of less than 0.1m”? How does the emulator performance change if you summed -0.1 for each 3 days of less than 5cm?
L. 253: related to the comment above: the highest biases are found in the low firn air content values, which might have to do with how you treat negative emulator predictions?
Fig. 3: From a), it looks like region 7 (and 8) should possibly be split if the idea is to divide the ice shelves by MoA ratio. Or are the “inland” grid points not used anyways? (L. 171) Then I would recommend to only show the grid points that are used in the emulator evaluation and training.
L. 217: It seems relevant to mention briefly that ERA5 can reproduce the AWS observations. Please include a short quantitative analysis on the comparison of AWS and ERA5 at these sites.
L. 386-387: This seems to be a critical piece of information that should be conveyed to any emulator user. I therefore would suggest mentioning it again in the Conclusions and/or the Abstract.
L. 386 and Fig. 4: MoA and Accumulation seem to be very important features for the emulators, yet the manuscript only stresses the importance of realistic melt values in future scenarios. Given the uncertainty of model predictions of accumulation in Antarctica, it seems necessary to me to also include a short paragraph on the influence of uncertain accumulation predictions when using the emulator for future predictions. (Coulon et al. 2025)
L. 400 and L. 468: If the FDM simulations are run as offline model, how is the “snow melt-albedo” feedback included in the “melt emulator” in contrast to the “no melt emulator”?
Technical corrections:
L. 101: Add the simulation period for the “future projections” in this sentence.
Fig. 1 and 3: use the same colors for the same regions in Fig 1 and 3 since they refer to each other.
L. 181: “FAC” is never introduced and not routinely used in the text. It introduced, it should be used throughout the text.
Fig. 6, Fig. 7, Fig. 13, Fig. 14: I would appreciate a “tabular labeling” or rows and columns of the subplots in these figures in addition to the explanation given in the caption.
Fig. 7: The choice of displaying the results on a map seems not very suitable. Maybe consider a scatterplot with colorscale for the different scenarios?
And Fig7 caption: replace “shorter periods” with “less years” to be clearer.
L. 374: Delete “3.3”
L. 463: exchange “to be prescribed” to “as input”
L. 463: include a description of the period and region in which the emulator can explain >94% of the variance. Standalone this value is not very informative.
Bibliography:
Coulon, V., Klose, A. K., Edwards, T., Turner, F., Pattyn, F., and Winkelmann, R.: From short-term uncertainties to long-term certainties in the future evolution of the Antarctic Ice Sheet, Nat Commun, 16, 10385, https://doi.org/10.1038/s41467-025-66178-w, 2025.