the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 15 Jul 2026
An accurate present-day model of the Antarctic ice sheet is necessary but insufficient for an accurate projection of its future evolution
Abstract. Accelerated retreat of the Antarctic ice sheet (AIS) dominates both the magnitude of, and the uncertainty in, high-end sea-level projections. One frequently cited caveat in the computational models used to create these projections is that many of them poorly represent the observed present-day state of the AIS. Following the thought that a model should accurately represent the present before it can be expected to accurately represent the future, there have been discussions in the ice-sheet modelling community to weight or even exclude individual models from ensemble studies based on how well they reproduce present-day observations.
In this study, we create a wide ensemble of realisations of the AIS with a numerical ice-sheet model under different parametric assumptions, utilising inverse methods to ensure every realisation reproduces snapshot observations of the present-day geometry, velocity, and thinning rates of the AIS to a degree that is comparable with the most recent ISMIP6 ensemble. These model realisations project anything between −0.3 m and +1.5 m of sea-level contribution (SLC) by 2300, with the sea-level rate at that time varying from −0.7 to +9.7 mm/yr. There is no clear relation between the degree to which these model realisations reproduce snapshot observations of the present-day AIS, and the amount of SLC they project. The main factors contributing to the wide variance in projected SLC are the sliding law and the distribution of slipperiness of the subglacial bed. Since these quantities are typically derived by inversion from observations of the ice flow, they can only be known as accurately as other processes affecting the flow, such as subglacial hydrology, ice damage, and calving. The resulting compensating errors are known to affect the modelled present-day state of the ice sheet differently than its response to future climate change, explaining why models with a nearly identical initial state can produce such different SLC projections. This implies that weighting or excluding models from ensemble studies based only on their agreement with snapshot observations of the present-day ice sheet might not be a wise choice. Instead, comparing to temporal observations, and to novel observations of the ice sheet’s sub-surface, might be more helpful to reduce the uncertainty in Antarctica’s future contribution to sea-level rise.
- Preprint
(3182 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-4041', Anonymous Referee #1, 17 Aug 2026
-
AC1: 'Reply on RC1', Tijn Berends, 23 Oct 2026
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-4041/egusphere-2026-4041-AC1-supplement.pdf
-
AC1: 'Reply on RC1', Tijn Berends, 23 Oct 2026
-
RC2: 'Comment on egusphere-2026-4041', Anonymous Referee #2, 08 Sep 2026
This is a manuscript with a very important message, well stated in the title, which I wholeheartedly agree with. Using a large ensemble of model simulations with a variety of parameter and forcing choices, it presents evidence that calibrating ice sheet models to reproduce snapshots of a present-day state is not sufficient for accurate predictions of future states. It will be a valuable paper, and certainly should be published.
The manuscript is thorough, with results well presented and the argument clearly made. I found myself agreeing with many well-made points throughout, and I don’t have any major recommendations for changes to the current content. I would however be interested to see a little more discussion on a couple of points:
1. The assessment of model quality by matching a snapshot of data is the idea being argued against. However, with reliable observations over Antarctica being available for the past two to three decades, what are the authors’ thoughts on using a match to the observational trend over a longer time period as a metric? I think this may have been one idea under consideration for the ISMIP7 analysis, using the historical runs required from models in that protocol.
2. While most inversion processes are still based on single snapshots of data, this is quite widely recognised as an issue (although never directly addressed in this way before, to my knowledge). In some cases (e.g. Goldberg & Holland, 2022), efforts are being made to introduce transient calibration, in which observational data from multiple timestamps is considered in the inversion process. This is seen to produce larger contributions to sea level in century-scale simulations. Related to the above point, do the authors think that matching multiple snapshots over time in this way could lead to an improved ability to judge the ‘quality’ of model setups from their initialisation, if the method were more widespread?I have a few specific comments below.
Line 74-6: A very good point; I’ve always thought initialising to a steady state is bound to introduce errors, since the real system is not steady.
Line 105: I wouldn’t class 5km as being a particularly high resolution for accurately representing grounding line movement, although this perhaps depends on the scale of the problem. Specifying that this is for whole-of-Antarctica simulations would clarify.
Line 203: A brief explanation of ABUMIP is needed, as not all readers will immediately know what it is. My suggestion would be to omit the reference to it here and simply say “The experiments presented in Appendix A…”, then give more detail in the appendix.
Line 227-30: What is the justification for the choice of 40,000 years etc. for the initialisations? Do you have a specific criteria by which steady state is defined, that is met within that timeframe?
Line 348: I think it’s important to note here that there is a significant increase in SLC at the upper end of the range covered by the high-resolution grounding lines, especially considering this high resolution is only applied to one region. It doesn’t surprise me that there’s little difference between 20km and 5km, but I expect if the 2km and 1km resolution were applied to the Siple coast and Ronne ice shelf where your model sees most retreat, the grounding line resolution would have a much larger effect. Convergence tests in another model have shown that mesh resolution should be under 2km to robustly simulate grounding line migration (Sun & Gudmundsson, 2026).
Line 385-6: I would argue that inversions do nothing to compensate for the lack of calving in models, but that’s a minor point of wording. The overall argument being made here is entirely right!
Line 406: This raises an interesting difference between the ‘nudging’ method and the other common approach of adjoint inversion. In the latter, observation errors are taken into account in the cost function. Many of the same points apply to both methods however, regarding the non-uniqueness of solutions and potentially reaching a ‘correct’ solution for the wrong reasons. While tangential to the main point of this paper (and I’m certainly not suggesting any further experiments!), it would be interesting to see some thoughts on whether the choice of inversion method would make any difference to your results.
References:Goldberg, D.N. and Holland, P.R., 2022. The relative impacts of initialization and climate forcing in coupled ice sheet‐ocean modeling: Application to Pope, Smith, and Kohler glaciers. Journal of Geophysical Research: Earth Surface, 127(5), p.e2021JF006570.
Sun, S. and Gudmundsson, G.H., 2026. The potential of marine ice cliff instability for Amundsen Sea glaciers revisited. Annals of Glaciology, 67, p.e15.
Citation: https://doi.org/10.5194/egusphere-2026-4041-RC2 -
AC2: 'Reply on RC2', Tijn Berends, 23 Oct 2026
The comment was uploaded in the form of a supplement: https://egusphere.copernicus.org/preprints/2026/egusphere-2026-4041/egusphere-2026-4041-AC2-supplement.pdf
-
AC2: 'Reply on RC2', Tijn Berends, 23 Oct 2026
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 278 | 249 | 30 | 557 | 29 | 34 |
- HTML: 278
- PDF: 249
- XML: 30
- Total: 557
- BibTeX: 29
- EndNote: 34
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Review of: An accurate present-day model of the Antarctic ice sheet is necessary but insufficient for an accurate projection of its future evolution
Berends et al. present a thorough and well-presented study considering the relationship between the 'quality' of an ice sheet model's initial state and its projection over centennial timescales. This is an important question as ISMIP exercises have begun to use the capability of a model to reproduce present conditions as an indicator of its ability to produce tractable projections. Using an ensemble intended to mimic the diversity of ISMIP submissions, but retaining a more controlled overview of the differences by keeping the underlying model the same, the authors find that there is in fact no clear relationship between the quality of the present day snapshot and future projections. This is an important, well-founded, and largely well-described finding, and it is certainly well-deserving of publication following minor revisions. I was impressed by the comprehensive nature of the study, and I have no further suggestions for additional simulations or experiments.
General comments:
Section 2 is a good and comprehensive overview of how the model is forced. What is missing however is a good description of how errors and instance quality are calculated *before* we get to the results. I would maybe put this as an additional section before the present section 3. This should include the full RMSE calculation (only introduced now at l244) and the Q_init terms (so basin, pan). I think the future projection methodology should also be introduced before, rather than alongside, the results pertaining to it (at the moment, l315 onwards is a methods paragraph in a results section). It would also be useful to introduce the idea that Q_init terms can be compared to sea level contribution within the methodology, rather than in the results, so that the reader understands why the methods are structured as they are.
In addition, all of your simulations show a very small sea level contribution for the first 100 years or even more, which I believe falls below many other projections (and implies a deceleration from currently observed melt rates). This is not central to your study as one would assume that the disconnect between initialisation skill and projection could be proven even for an idealised ice sheet case, but I think it does warrant a bit more discussion than it currently receives.
Sliding terms are a bit mixed up, going between 'roughness', 'till friction angle', 'slipperiness'. Suggest choosing one and sticking (pun intended) to it.
I'm not totally clear how your 'nudging' schemes relate to what are more generally called inversion schemes (such as Morlighem et al., 2013, https://agupubs.onlinelibrary.wiley.com/doi/full/10.1002/jgrf.20125). I would suggest making your definition clearer, and simply using inversion if your methods are suitably comparable to the likes of Morlighem et al. 2013. Usually when I hear nudging I think of the approach taken by paleao simulations where a fully automated inversion procedure is not possible. Speaking of which, it would be nice to get some description of how your findings might relate to the palaeo simulations that also feature in ISMIP ensembles.
Specific comments:
Abstract:
This is a cool paper and it does a lot of things, but I don't believe it warrants a long double-paragraph abstract
I think acronym usage should generally be minimised and I do not think sea level contribution needs an acronym (especially not in the abstract)
I note that there are no studies available (to my knowledge) that actually link friction to 'novel observations of the ice sheet's sub-surface', and that this is only cursorily addressed in the discussion. I would be happy to see further affirmative coverage in the discussion, or saying that developing such a method is important, but otherwise I think this suggestion operates a bit like a research trope as it appears to perennially be just out of reach.
l17 These to Our if they are yours
l21 subglacial bed to glacier bed
l24 differently than is a bit imprecise, 'force the system away from...' or similar?
Introduction:
l33 I think worth very briefly mentioning the nature of these instabilities if you are going to introduce instabilities at all
l60 (bulk) to bulk or [].
l64 non-existence to not existing at all
l64 I would introduce that these can be derived from inversions or 'nudging/adjustment' procedures in the case of simulations running from a palaeo state
l83 You may want to consider starting one of these sentences with a word other than however. However, this is just a stylistic choice
l99 the surficial to surface
l105 I don't know if 5 km can be correctly described as high resolution, and it would be good to give a maximum resolution. I see this is in Table 6, but good to put the info here, too.
l117 i.e. to as, or consider writing out in fully vectorised form
l127 See also the general comment, but a bit more description about this procedure is needed. How is the nudging achieved, is an inversion algorithm used?
l130 See an earlier comment, but I think good to make this introduction a bit clearer, and perhaps to include a bit more information about how other ISMIP models handle this (and why you use the methods you do)
l137 Suggest calling this height-velocity nudging scheme as H is not defined for a while (or to introduce the nomenclature in the first sentence of the subsection)
l143 Can we get a bit more information on the f_u and f_d functions?
l144 As above, I think replacing i.e. with a word is appropriate
l150 I'm not sure personal comms falls under EGU citing guidelines. Suggest just removing the parenthetical statement
l169 As with section 2.3, suggest putting in a preface to the sub-shelf melt options. WOA as an acronym is not introduced until later in this subsection. I suggest just calling it World Ocean Atlas throughout (or at the very least in the subsection heading)
l171 Suggest stating which equation from Favier, and perhaps also why that choice
l188 Further clarification about nudging vs. inversion schemes
191 I don't see a definition of BMB (basal mass balance?) and it's only mentioned four times -- remove the acronym?
l216 I think sections 3 and 4 could be folded into a larger results section that features a brief introduction to guide the reader
l223 This information could go higher up when resolution is introduced
l226 This is a bit unclear. Multiple resolutions to get to the start point of one simulation, or separate simulations. If so, why the difference in years? Needs a bit more explanation.
l231 resolved that were already resolved, consider rewording
l234 also/instead to also or instead of. Also, specify the place for most currently observed mass loss (Antarctica, west Antarctica)
l279 Worth giving the simulation moniker, or referencing best rather than this?
l333 How is it that it stabilises for such a long time when we would really expect observed melt rates to continue (see also general comment)?
l329 Here and in at least one other instance, I think better to start the paragraph with something more descriptive than introducing a figure
l403 This sentence is not super clear, and could probably just be removed if not tightened
l405 This paragraph is not really necessary to the central thesis and reads more like a commentary, but I appreciate the points being made. I just think that as it isn't mentioned in the abstract only the real heads will get around to reading it, probably still worth including
l416 Specify what both refers to
l420 I think it is complicated, but not necessarily impossible. See for example Maier et al. (2021, https://tc.copernicus.org/articles/15/1435/2021/) and Jager et al. 2026 (a and b, https://essopenarchive.org/doi/full/10.22541/essoar.177099457.70593031/v1).
l440 I think it would be great to foreshadow this conclusion in the abstract
Figures:
Fig. 2 a minor point, but UFE on its own has not been defined. As with Fig. 5, I think the radial grid needs explaining in the caption and should probably co-exist with a normal perpendicular grid.
Fig. 5 I think the radial grid is so that Q_init is visible, but I think this is confusing at first glance. I would suggest using normal perpendicular grid lines and overlying the radial grid if you feel strongly about it, and also mentioning it in the caption. Also handy to put sliding law, nudging method, and so on alongside A, B, C, ...
Fig. 7 I would be personally interested to see how/if each group of model differences alters the final model geometry (appreciate that Fig. 7 is grounded cells not geometry), but feel free to ignore this
Fig. 8 Could you make a figure with all simulations where they are coloured by there Q_int values? Might be a neat illustration, but just a suggestion
Fig. 9 This figure could be improved by colour-coding the dots to specific groups (for example, Weertman-type and otherwise)
Appendix A, nowhere is the acronym for ABUMIP (or ABUK) expanded