Precipitation Ansatz dependent Future Sea Level Contribution by Antarctica based on CMIP5 Model Forcing

Various observational estimates indicate growing mass loss at Antarctica’s margins but also heavier precipitation across the continent. In the future, heavier precipitation fallen on Antarctica will counteract any stronger iceberg discharge and increased basal melting of floating ice shelves driven by a warming ocean. Here, we use from nine CMIP5 models future projections, ranging from strong mitigation efforts to business-as-usual, to run an ensemble of ice-sheet simulations. We test, how the precipitation boundary condition determines Antarctica’s sea-level contribution. The spatial and temporal varying 5 climate forcings drive ice-sheet simulations. Hence, our ensemble inherits all spatial and temporal climate patterns, which is in contrast to a spatial mean forcing. Regardless of the applied boundary condition and forcing, some areas will lose ice in the future, such as the glaciers from the West Antarctic Ice Sheet draining into the Amundsen Sea. In general the simulated ice-sheet thickness grows in a broad marginal strip, where incoming storms deliver topographically controlled precipitation. This strip shows the largest ice thickness differences between the applied precipitation boundary conditions too. On average 10 Antarctica’s ice mass shrinks for all future scenarios if the precipitation is scaled by the spatial temperature anomalies coming from the CMIP5 models. In this approach, we use the relative precipitation increment per degree warming as invariant scaling constant. In contrast, Antarctica gains mass in our simulations if we apply the simulated precipitation anomalies of the CMIP5 models directly. Here, the scaling factors show a distinct spatial pattern across Antarctica. Furthermore, the diagnosed mean scaling across all considered climate forcings is larger than the values deduced from ice cores. In general, the scaling is higher 15 across the East Antarctic Ice Sheet, lower across the West Antarctic Ice Sheet, and lowest around the Siple Coast. The latter is located on the east side of the Ross Ice Shelf. Plain Language Summary In the warmer future, the ice sheet of Antarctica will lose more ice at the margin, because more icebergs may calve and the warming ocean melts more floating ice shelves from below. However, the hydrological cycle is also stronger in a warmer world. As a consequence, more snowfall precipitates on Antarctica, which may balance the amplified 20 marginal ice loss. In this study, we have used future climate scenarios from various global climate models to perform numerous ice-sheet simulations. These simulations represent the Antarctic Ice Sheet. We analyze whether Antarctica will grow or shrink. In all our simulations, we find that certain areas will lose ice under all circumstances. However, depending on the method used to describe the precipitation reaching Antarctica in our simulations, parts of the Antarctic Ice Sheet may either grow or shrink in the future. How strong the precipitation grows in a warming atmosphere, may be explained by the dissimilarity between 25 1 https://doi.org/10.5194/esd-2019-78 Preprint. Discussion started: 6 January 2020 c © Author(s) 2020. CC BY 4.0 License.

Nine CMIP5 models deliver the following climate scenarios (see Table 1): control run (piControl), the historical period , as well as RCP2.6, RCP4.5, andRCP8.5 (2006-2100). These models stem from different model families (Knutti et al., 2013) and cover the range of current atmospheric (Agosta et al., 2015) and oceanographic (Sallée et al., 2013a) model 135 uncertainties although model deficiencies such as insufficient resolution can exist across all models. The transient forcing from 1850 until 2100 comprises the historical and scenario periods. Afterward, the last 30 years (2071-2100) are repeated until the year 5000 to keep the natural variability. From the control run of the climate model (piControl), we use either the first or the last 50 years. By this procedure, we could quickly identify a drift in CMIP5 models and assess its impact. Additionally, the number of scenarios is twice as large, since the mean states of the first and last 50 years differ commonly marginally. Anomaly 140 forcing is computed relative to either the first or last 50 years of the control run. In the following, the first 50 years act generally as reference.
Atmospheric and oceanic forcing is applied as annual mean forcing on top of the forcing used to spin-up the ice-sheet model (Table 2). Since CMIP5 models do not resolve ice shelves, ocean temperatures are extrapolated horizontally into the ice shelves to mimic isopycnical flow. To allow for surface melting under a warming climate, the surface mass balance (SMB) is calculated 145 following the positive degree day (PDD) approach (Braithwaite, 1995;Hock, 2005;Ohmura, 2001), where the annual 2m-air temperature standard deviation comes from daily CMIP5 model values.
The ice-sheet model PISM -based on version 0.7 -runs on a 16 km equidistant polar stereographic grid and it utilizes a hybrid system combining the Shallow Ice Approximation (SIA) and Shallow Shelf Approximation (SSA). The model utilizes a generalized version of the viscoelastic Lingle-Clark bedrock deformation model (Bueler et al., 2007;Lingle and Clark, 1985). 150 In our simulations, only the viscous part has been used because of known implementation flaws in the elastic part in our and later PISM versions. The basal resistance is described as plastic till by a Mohr-Coulomb formula to perform the yield stress computation (Bueler and Brown, 2009;Schoof, 2006). The basal melting of ice shelves is proportional to the squared thermal temperature forcing (∆T 2 force ), which is the difference between the pressure-dependent melting temperature and the actual ocean temperature above melting. Here, the parameterization considers the full depth-dependence of the ocean temperature Ice Shelves and the neighboring Antarctic Peninsula, where the ocean temperatures are lowest in the control climate ( Figure 2c).
The spatial structure of the anomalies discussed above is in general independent of the applied forcing scenario, while the scenarios determine, however, the strength of the anomalies. Regardless of the applied scenario, the discussion of the atmospheric climate anomalies indicates already that both precipitation and temperature do not necessarily grow in parallel. 205 Instead, regional differences are evident, and a simple scaling of the precipitation with temperature appears to be inadequate.

Precipitation scaling
Inspired by the Clausius-Clapeyron process, it is often assumed that with a warming atmosphere, the precipitation also raises.
ice-sheet simulations bridging several millennia often rely on climate anomalies deduced from ice cores, for instance. Based on isotopic signatures in ice cores, temperature anomalies are deduced. Inferred accumulation anomalies from these cores are 210 converted into precipitation anomalies. Together with the contemporary climate fields as a reference, the temperature scaling of precipitation is where ∆T is the temperature anomaly, ∆P the precipitation anomaly, and P 0 = P (t ref ) the precipitation reference field. The 215 scaled precipitation is The scaling deduced from ice cores varies in Antarctica between 5 %K −1 and 7 %K −1 , with a 2-sigma uncertainty of about 1 %K −1 -3 %K −1 (Figure 4, Table 3).
The region "Siple Coast" as a part of the "WAIS" region is different in many aspects. It has the smallest area compared to the other regions (Table 4), and it shows the lowest ensemble mean scaling factors for all scenarios. Also, as before, no clear trend exists between different scenarios across the entire ensemble, while the spread of trends among individual ensemble members is substantial. Furthermore, some members exhibit a negative scaling, where precipitation decreases for rising temperatures: MPI-ESM-LR under the RCP8.5 scenario and NorESM1-M under all scenarios (RCP8.5, RCP4.5, and RCP2.6). The inverted 265 sign of the scaling is in stark contrast to the ensemble mean.
In the last decades, the detected downward trend in snow accumulation in this area (Wang et al., 2017) occurs while the wider West Antarctic Ice Sheet region belongs to the most rapidly warming regions globally . It underpins that less accumulation can befall under a warming climate. Furthermore, sea ice has expanded in the Ross Sea (Haumann et al., 2016;Liu, 2004). Hence, some ensemble members seem to imitate that expanding sea ice modifies the evaporation from the 270 ocean and impacts the atmospheric circulation, which controls the flow of humid air masses, delivering precipitation to the Siple Coast. Even if NorESM1-M reproduces the overall seasonal sea ice extent cycle better than most CMIP5 models (Turner et al., 2013), it shows a unrealistic declining February sea ice trend in the Ross Sea over 1979-2005(Turner et al., 2013. MPI-ESM-LR has large negative errors in sea ice extent over the year (Turner et al., 2013). Hence the mimicry of observed features in models may occur for the wrong reason.

275
In all four large regions ("glaciered", "grounded", "EAIS Atl", "EAIS Ind", and "WAIS"), we see a trend towards lower scalings for weaker forcing scenarios in the ensemble mean, with the exception of "EAIS Ind", where the factors for RCP8.5 and RPC4.5 are indistinguishable. Also, Frieler et al. (2015) found a low dependence of the scaling factors to the RCP scenario in comparison with the dependence on the specific climate model. The region "WAIS" has on average a smaller precipitation scaling than both regions of the East Antarctic Ice Sheet ("EAIS Atl" and "EAIS Ind"), which is also reflected by the maxima 280 in these regions. Also, the Ross Ice Shelf and the adjacent Siple Coast feature on average the lowest scaling factors across the entire ice sheet. Some individual ensemble members project even negative scaling: precipitation deficit for rising temperatures.
The Siple Coast highlights definitely that, at the continental scale, it is not adequate to describe the spatial evolution of the precipitation by a fixed temperature scaling. Since the scaling exceeds mostly the commonly utilized value of 5 %K −1 , for instance, we diagnose the sea-level impact of applying the actual scaling distribution (e.g., Figure 4) versus a spatially and 285 temporally constant scaling of 2 %K −1 , 5 %K −1 , or 8 %K −1 across Antarctica.

Sea Level Impact of Precipitation Scaling by Temperature
To understand how the precipitation boundary condition impacts Antarctica's contribution to the global sea level, we inspect the precipitation fallen on Antarctica ( Figure 6). Therefore, we integrate it over the dark-blue masked region of grounded ice (map on Figure 6), perform a cumulative summation since 1850, and restrict our analysis to all ensemble members driven by 290 RCP8.5 (and anomalies relative to the first 50 years of the control run). Since accumulated precipitation over Antarctica lowers the global sea level under the assumption that ice loss (basal melting or calving) does not occur, the temporally accumulated sea-level impact curves have a negative slope (Figure 6a, b). Further on, this quantity is labeled "integrated precipitation," and the presented sea-level equivalent assumes a global ocean area of 3.61 · 10 14 m 2 (Gill, 1982).
The integrated precipitation declines more forcefully since the beginning of the 21st century, which is driven by the concurrent increase of precipitation over Antarctica (Figure 3a). The integrated precipitation shows a more pronounced temporal change, because the integral and not the mean precipitation is calculated, where the vast light precipitation regions lessen the average precipitation signal. After the year 2100, the integrated precipitation declines linearly (Figure 6b), as we adopt the forcing of the years 2071-2100 recurrently. By applying the actual precipitation anomalies (solid lines, Figure 6a, b), the sealevel drop is stronger than using a scaling of 5 %K −1 (dashed lines, Figure 6b) because the models' internal scaling exceeds 300 5 %K −1 ( Figure 5). Thus, in the year 5000, the maximal sea-level drop of 11 m (CCSM4) is nearly twice as large under the precipitation anomalies, compared to less than 6 m (MIROC-ESM) for the 5 % scaling.
The difference of the integrated precipitation between 5 %K −1 scaled and directly applied precipitation anomalies is always positive (solid lines in Figure 6 c, d). This difference ranges approximately from 1 cm (CISRO-Mk3-6-0) to 15 cm (CCSM4) in the year 2100 and from 60 cm (MPI-ESM-LR) to 550 cm (CCSM4) in the year 5000. A lower scaling of 2 %K −1 causes 305 a magnified difference (dotted line in Figure 6 c, d), which corresponds to a reduced sea-level impact if we would apply this low scaling of 2 %K −1 . It leads to differences ranging from 5 cm (MPI-ESM-LR) to 21 cm (CNRM-CM5) in 2100 and from 150 cm (MPI-ESM-LR) to 850 cm (CCSM4) in 5000.

315
Over the entire Antarctic continent, precipitation and temperature grow simultaneously in climate model simulations of the future. In concert with estimates of accumulation changes and temperature anomalies obtained from ice cores, it may (mis)lead us to scale the precipitation by the temporally evolving temperature. Therefore, fixed scaling factors are common.
However, a tendency towards higher scaling exists under more vigorous climate trends ( Figure 5), and the scaling has a clear spatial dependence (Figure 4 and 5). As a consequence, the on Antarctica accumulated snowfall for future climate projections 320 differs between the methods, which ultimately leads to biased estimates of Antarctica's contribution to the global sea level.
To assess the introduced bias, we analyze simulations of the Parallel Ice Sheet Model driven with numerous variants of the above-discussed climate conditions and a diverse set of implemented boundary conditions.

Relation between Precipitation Boundary Condition and Ice Thickness
Starting in the year 1850, we performed numerous ice-sheet simulations to analyze how the implemented precipitation bound-325 ary condition impacts the ice-sheet thickness distribution. Each climate scenario from an individual climate model (as part of the ensemble) drives an independent ice-sheet simulation. These together constitute the ensemble of ice-sheet simulations.
Hence, the average across ice-sheet simulations forms the ensemble mean. For the diagnostic, we also inspect the maximum and minimum thickness at each grid point across all ensemble members. Therefore, the field of joined extreme values could come from a diverse set of ice-sheet ensemble members and, hence, does not necessarily lead to dynamically consistent distribution.

330
Complementary ice-sheet (control) simulations are performed under the sole utilization of the reference forcing fields . In these simulations, the detected trend of about 2 mm decade −1 (sea-level equivalent) fades within the first 400 years and differs slightly between the two initial states (PISM1Eq and PISM2Eq). Hence, in the following, the subtracted trend for each single ensemble member depends on its initial state.
In the year 2100, the ice thickness for both precipitation boundary conditions (precipitation anomaly deduced from the 335 applied climate models versus scaled precipitation) increase over large parts of the Antarctic continent (Figure 7b-e). The thickness for the simulations driven by scaled precipitation grows less over substantial parts of the interior than the simulations forced by the precipitation anomalies ( Figure 7a) as the difference between scaled precipitation and applied precipitation anomaly is mostly negative. Thus, simulations driven by the precipitation anomalies accumulate more snow and grow thicker ice, which leads to a stronger sea-level drop. This result supports the analysis above ( Figure 6). A ring of a pronounced negative 340 thickness difference follows the coast, where the precipitation anomaly (Figure 2e, h, k) is enhanced. This ring emerges for a significant part of the coastal East Antarctic Ice Sheet (EAIS) and West Antarctic Ice Sheet (WAIS). For the latter ice sheet, the negative area is shifted away from the coast towards the interior ( Figure 7a). Also, a negative strip appears at the south side of the Transantarctic Mountain Range, and some grounded ice streams flowing into the Filchner-Ronne Ice Shelf are negative.
Regions of positive differences coincide with thicker ice for simulations driven by scaled precipitation. These are located 345 south of the Transantarctic Mountain Range at the northern edge of the Ross Ice Shelf, along the coastline of the WAIS, and in the coastal Terre Adélie region. There, the scaling is generally lower or falls behind the constant scaling of 5% K-1. However, this does not explain exclusively positive areas.
For both precipitation boundary conditions, the ice thicknesses of the ensemble means reveal a widespread weakening of the floating ice shelves, such as Filchner-Ronne, Ross, and Amery Ice Shelves (Figure 7b,d (Figure 8c, e).
To conclude, the ice thickness is indeed thicker for simulations driven by the precipitation anomalies ( Figure 8). Regardless of the applied precipitation boundary condition, there is widespread ice-shelf weakening, ice thinning at the margins in the ensemble mean. Ice thinning for the ensemble member of the maximal thickness highlights the most vulnerable regions, such 360 as Pine Island and Thwaites Glaciers, Amery Ice Shelf and some outlet glaciers of the EAIS.

Attribution of the driving model
All ensemble members contribute to the ensemble mean, while at a given grid location the maximum and minimum are determined by climate forcing from one particular climate model. We inspect which climate model may lead to ice thickness growth or shrinking and restrict ourselves first to the model year 2100, where the precipitation anomalies of the period 1850-365 2100 shape the ice-sheet thickness distribution of the year 2100.
Directly at margins apart from the vast ice shelves, the attributed model that drives either the maximum or minimum ice thickness shows a noisy small scale pattern, which is driven by the variety of the involved models (Figure 8d Hence, the ocean forcing drives the ice-shelf thinning here. Since the spatial pattern of the atmospheric and ocean forcing that promotes or undermines the ice thickness is not necessarily aligned, this may explain the small scale noisy pattern along the coast.

375
Beyond the direct coast strip, larger areas appear where the forcing from one climate model determines the maximum or minimum thickness, respectively. However, these extended continuous regions are often interrupted by spots controlled by the climate from other models. Also, the pattern is changing during the transient simulation starting in 1850 because the temporal evolution of the 2m-air temperature and precipitation anomalies are different for each climate model as the integrated precipitation highlights (Figure 6a, b). Furthermore, after the year 2100, where the same 30 years forcing period (2071-2100) 380 drives the ice-sheet model recurrently, the pattern evolves further ( Figure 9). Because the ice sheet has not reached the quasiequilibrium to the last 30 years forcing, the pattern alteration is ongoing.
For grounded ice, three models (CCSM4, CNRM-CM5, MIROC-ESM) determine predominantly the growing ice until the year 2100 ( Figure 8d), which is in-line with the diagnosed sea-level contribution (solid line, Figure 6a, b). CCSM4 dominates the "EAIS Atl" sector, while CNRM-CM5 dominates a band from the "EAIS Ind" sector clockwise to the Antarctica Peninsula, the year 2200, where we have applied the 30 years forcing more than three times, the emerging picture shows a consolidation of the influential spheres of the different models for both the maximum and minimum thicknesses ( Figure 9).
If we now turn towards the temperature scaled model simulations, the mean, maximum, and minimum ice thickness distributions ( Figure 10) are similar to the ones driven by the precipitation anomalies as discussed above ( Figure 7). Also, the same models determine the ice-shelf thickness of the Filchner-Ronne and Ross Ice Shelves. The latter shows that the ocean 395 controls ice-shelf thickness changes in our simulations primarily. However, we detect a stark contrast of the model determining the maximum and minimum ice thickness. For the maximum, we still have the same three models (CCSM4, CNRM-CM5, MIROC-ESM). However, the pattern has changed. CCSM4 controls a smaller area in the interior around the South Pole, and MIROC-ESM some coastal regions of the East Antarctic Continent, while the remaining majority of the grounded ice is under the control of CNRM-CM5. The most striking changes occur for the minimum. Now, NorESM1-M determines the entire WAIS 400 and some parts of "EAIS Ind", while MRI-CGCM3 dominates the remaining East Antarctic Ice Sheet.
In the latter case, temperature variations force the precipitation driven ice thickness evolution exclusively (see Equation 1).
These temperature changes do not necessarily reflect dynamical changes in the atmosphere that are accompanied by modified circulation patterns that ultimately transport and deliver the precipitation for Antarctica. Hence, the applied scaling or precipitation boundary condition impacts the temporal evolution of the Antarctic Ice Sheet geometry, which ultimately shapes 405 Antarctica's contribution to the global sea level.

Ice losses
After the spin-up, the simulations have reached a quasi-equilibrium. For the discussion of the ice losses, we concentrate on the transient period 1850-2100. The calving rate hardly changes ( Figure A12), whereas the total ice-shelf area is nearly constant until 2000 and declines afterward ( Figure A15). The ocean-driven basal melting is proportional to the squared temperature 410 difference between the pressure-dependent melting temperature and the actual ocean temperature. Since the ocean temperature increases in general (Figure 2f, i, l and Figure 3c), also the mass loss by basal melting increases, while the total shelf ice area is quasi-constant until 2000 and declines afterward ( Figure A15). For RCP8.5, the basal melting increases at the end of the 21st century quadratic. To conclude, the calving rate is nearly constant, while the basal melting increases by approximately 33 % since the year 2000.

415
The mean calving rate is about 8000 Gt year −1 and 5000 Gt year −1 for the ensemble member utilizing the parameters and the initial state of PISM1Eq and PISM2Eq, respectively ( Figure A12). The basal melting rates for PISM1Eq and PISM2Eq are similar, however, the loss rates for PISM1Eq are slightly larger than PISM2Eq ( Figure A13). The ensemble mean starts at about 550 Gt year −1 in 1850 and reaches 900 Gt year −1 in 2100.
Since floating ice shelves nourish both ice losses, these ice losses do not impact the sea-level directly. Under the assump-420 tion that the inflow of former grounded ice compensates any shelf mass loss, the reported ice losses of 8500 Gt year −1 -9000 Gt year −1 (5500-6000 Gt year −1 ) would correspond to a sea-level rise of 2.58 cm year 1 -2.74 cm year 1 (1.67 cm year 1 -1.83 cm year 1 ). The Integration over 250 years to match the period from 1850 to 2100 generates a sea-level equivalent of 6.47 m -6.85 m (4.19 m -4.57 m). However, the actual ratio between total ice mass change and the corresponding sea level response is not a 1:1 relation. Instead, on average less than 5 % of the total mass lost by both iceberg calving and floating ice-shelf 425 melting is compensated by grounded ice that raises the sea level ( Figure A8). Considering this ratio of 5 %, the sea level impact reduces to 0.32 m -0.34 m (0, 21 m -0.23 m). It is less than integrated precipitation anomalies across the Antarctic continent ( Figure 6a).
Anyhow, the integrated basal melting rates are too low and the calving rates are too high compared to observational estimates in our ensemble of ice-sheet model simulations. Besides the fact that the total loss exceeds recent observational estimates, our 430 ice sheet is in a quasi-equilibrium after the spin-up. All this may indicate that the integrated precipitation driven accumulation resulting from the RACMO precipitation reference field might be too large. However, the surface mass balance of RACMO agrees well with observational estimates (Wang et al., 2016), Beyond the year 2100 ( Figure A14), the calving rates decrease and reach a minimum in the period 3000-4000. Afterward, calving increases again slightly. Basal melting rates are subject to a slight decreasing trend (RCP2.6), nearly constant values (RCP4.5), or a negligible upward trend after the year 4000 (RCP8.5).

Precipitation Boundary condition and Sea Level
In the following, we consider all ensemble members starting from both initial states PISM1Eq and PISM2Eq. They are driven by the climate scenarios RCP2.6, RPC4.5, or RCP8.5. For each CMIP5 model, the applied anomalies have been computed either relative to the first or last 50 years of the control run simulation (piControl). The sea-level curves are shifted so that the sea-level contribution is 0 m in the year 2000. Since the spread of individual ensemble members may not follow a normal 445 distribution, we present beside the mean also the median sea-level contribution. For the RCP8.5 scenario, we highlight the spreading among models by depicting the standard deviation (1σ).
For the period 1850 until 2000, the simulated sea level contribution of Antarctica fluctuates slightly. Hence, the accumulation balances nearly the ice loss at the margin while the basal melting rated of grounded ice is steady ( Figure A9). Please note that this is not driven by any trend in the continued ice-sheet simulations under the reference climate (Table 2) since we have 450 subtracted this trend. We also detect an amplified signal for the simulations driven by the precipitation anomalies than scaled precipitation, which corresponds to the above diagnosed sea-level impact of the precipitation ( Figure 6).
After the year 2000, all our ensemble members, regardless of the forcing scenario, show a falling sea level ( Figure 12).
The basal melting of grounded ice does not impact the sea level contribution, because this basal melting rate, expressed as integrated sea-level equivalent since 1850, is the same and grows linearly for all scenarios until 2100, and only after the year 455 2500 these curves diverge ( Figure A9). Ultimately, the more vibrant growth of the accumulation in comparison to the negligible increasing combined loss of iceberg calving and basal melting of ice shelves drive the falling sea level in our simulations after the year 2000 ( Figure 12). Depending on the applied forcing and precipitation boundary condition, the global sea-level drop ranges from 2 cm to 11 cm until 2100 ( Figure 12). This result is in contrast to various publications, and we discuss it below.
If we continue our ensemble with the last 30 years of forcing until the year 5000, the sea-level contribution of those ensemble 460 members driven by the temperature scaled precipitation starts to stabilize and reaches a minimum around the year 2500.
Afterward, they begin to lose more ice at the margins than they gain in the interior. As a consequence, these will contribute after the year 3200 (RCP8.5) and 3900 (RCP2.6) to a globally rising sea level on average in our simulations, which outruns the formerly fallen sea level since 1850. In the year 5000 at the end of our simulations, these runs show a trend towards a continuously growing ice loss rate, because the curves have still an upward-directed tendency. Hence a quasi-equilibrium is not 465 established. In contrast, the simulations driven by the precipitation anomalies continue to show a falling sea level. They always contribute negatively to the global sea until the year 5000, and their ensemble mean and median sea levels tend to converge at the end of the simulations.

Sea level contribution of corrected basal melting
Since the deduced Antarctica's sea level contribution disagrees with the currently observed state showing mass loss, we apply . Under the assumption that only a fraction of the adjusted basal mass contributes to the global sea level, we apply the simulated ratio of the sea level change to the total ice mass change. For each ensemble member, this ratio is 475 the median ratio over its entire time series (For details inspect please the appendix section "Bias-corrected fluxes" A3 on page 46). Since we examine enhanced mass loss, we do not adjust the iceberg calving rates that are already higher than observed.
The determined temporal evolution of the sea level correction ( Figure A5, Equation A8) does impact the global sea level, but it does not change the sign of the contemporary sea-level evolution. Consequently, the impact on the sea level is very small as its evolution, which considers the correction, highlights ( Figure A6). If we assume instead that all of the additional mass loss This sea-level rise is larger than the observed integrated sea level rise of about 20 cm since 1850 (Church and White, 2011), which has been driven by world-wide land-water storage changes, shrinking glaciers around the globe, enhanced melting from Greenland, and thermal expansion of the ocean (Cazenave and Remy, 2011; Leclercq et al., 2011;Church and White, 2011).
To conclude: The correction exceeds observational estimates significantly under the unrealistic assumption that all additional 485 basal melting of ice shelves raises the sea level. It is unrealistic because disintegrating floating ice shelves do not impact the sea level. The correction hardly corrects the discrepancy if we apply the inferred ration of about 5 % between the simulated total ice mass loss and the simulated sea level contribution.

Conclusions
It is crucial for numerical simulations of Antarctica's sea-level contribution, how the precipitation is implemented in ice-sheet 490 simulations. The commonly used method of scaling the precipitation changes with the simulated temperature changes from ice cores or global climate models leads to a positive Antarctic sea-level contribution, i.e., a sea-level rise. However, when considering the simulated precipitation changes from the global climate models, the situation changes. In this case, numerical projections simulate a negative sea-level contribution. Major uncertainties affect these simulations, such as the partitioning of in the ice-sheet model -or the omission of important processes, such as the ocean-ice-shelf-ice-sheet interactions. While we could improve some aspects of the involved process descriptions, our simulations are state-of-the-art and suffer, thence, the same limitations as others.
In all CMIP5 models, the 2m-air temperature warms across the entire Antarctic continent without any exception (Figure 2d, g, j, and 3a), because even the minimum 2m-air temperature anomaly is positive everywhere (Appendix Figure A2d, g, j). The 500 warming enhances the hydrological cycle, which causes generally heavier precipitation ( Figure 3b) in particular along the coast of Antarctica (Figure 2e, h, k). However, the changing precipitation does not increase at the same rate with increasing temperature because it is not only thermodynamically influenced but also dynamically controlled. Given that the ensemble mean temperature scaling is different for the West and East Antarctic Ice Sheet ( Figure 5) and has a considerable spatial dependence, the dynamical component is not negligible. Instead, the region of reduced precipitation under rising air temperatures, which we 505 have identified along the Siple Coast, highlights that the dynamics could compensate or even overwhelm the thermodynamics.
The continent-wide scaling is per se problematic, even if we would adjust the scaling factor to reproduce the continental-wide average scaling. In this case, the total amount would be identical, but the spatial structure is still entirely different ( Figure 4).
Hence for a proper projection of Antarctica's sea-level contribution, the spatial pattern of the future accumulation of precipitation shall also consider the dynamical effect. Peninsula. These regions correspond to those, which have been identified across sixteen models, where ocean warming wanes marginal ice (Seroussi et al., 2019a).

515
The ocean (Etourneau et al., 2019) and atmosphere (Mulvaney et al., 2012;Thomas et al., 2009;Morris and Vaughan, 2003) is already warming along the Antarctic Peninsula. This results in a southward progressing of the annual 2m-air temperatures of -9 • C or -5 • C isotherm, which presents the range of thresholds for the stability of ice shelves suggested by Morris and Vaughan (2003, -9 • C) and Doake (2001, -5 • C), respectively. It may also enable the formation of meltwater ponds on ice shelves ) that precedes (van den Broeke, 2005) or even triggers ice-shelf disintegration (Banwell et al.,520 2013, 2019). After an ice shelf has decayed, the feeding ice streams are losing more ice, as seen for Larsen-B (Rott et al., 2011), which lowers the thickness of grounded ice. Anyhow, ice shelves along the Antarctic Peninsula have collapsed or are retreating (Cook and Vaughan, 2010;Rott et al., 1996). This observed retreat and the related ice loss will continue in our simulations under RCP8.5. Since our simulations presented here are in contrast to others that project a sea-level contribution from a shrinking Antarctic Ice Sheet, we highlight the differences before we discuss the limitations of our simulations. Some simulate Antarctica with a finer spatial resolution (Golledge et al., 2015;Pollard et al., 2015), which could improve the presentation of ice streams. These 535 streams channelize the flow of grounded ice from the interior to the margins, where they feed the attached ice shelves and discharge directly into the ocean. However, the simulated surface velocity distribution reproduces appropriate satellite-based estimates (Appendix Figure A10 and Figure A11). Others used the cliff failure parameterization supporting ice loss together with a constant ocean temperature offset of +2 • C (Pollard et al., 2015), twice as large as the amount found in our ensemble of nine CMIP5 models (Figure 3), or utilized continuously raising atmospheric and oceanographic temperature forcing (Golledge 540 et al., 2015;Mengel et al., 2015;Winkelmann et al., 2012Winkelmann et al., , 2015 beyond the year 2100. These stronger forcings alone explain a large part of the difference because we apply recurrently the forcing of the years 2071-2100 after 2100. Since the precipitation boundary condition determines if Antarctica rises or lowers the global sea level, it may be appropriate to utilize a more sophisticated surface mass balance (SMB) model. The recent publication that indicates a Greenlandification of Antarctica's margin at the end of the century (Bell et al., 2018) supports this approach, but the required atmospheric inputs 545 fields are not available at sufficient temporal resolution. Hence, this will be an option for simulations driven by the forthcoming CMIP6 model output.
Even if we apply anomalies on top of the reference background fields, we can not exclude a shock-like behavior of the simulations entirely directly following the decades after the year 1850. Since we compute the anomalies relative to the first or last 50 years, respectively, of the control run for each climate model, the anomalies are not necessarily zero at the beginning of 550 the year 1850. Hence, the ice-sheet model may experience a small jump, which may cause a wrong trend initially. Nevertheless, the long-term positive and negative sea-level contribution of Antarctica for simulations driven by temperature scaled and directly applied precipitation anomalies, respectively, are robust.
An issue could be the parameterization of the grounding line migration, where only extremely high resolution relaxes its need. However, PISM's grounding line parameterizations at medium to lower resolution is consistent with higher-order models hence, we consider our grounding line migration as reasonable. The apparent stability of ice shelves in the runs driven by the precipitation anomalies seems to comply with the safety band of ice shelves (Fürst et al., 2016), so the calving does stay outside of ice-shelf regions essential for providing buttressing for the inflowing grounded ice streams.

560
The ocean boundary condition, where ocean conditions are extrapolated into the ice-shelf cavities, drive basal ablation of ice shelves. Here, we could undoubtedly improve simulations if the ice shelves would be coupled to the driving ocean model, so that basal melting impacts the thermal structure of the ocean and, ultimately, the melt patterns. CMIP5 models neglect the oceanice-shelf interaction (Meijers, 2014), and their coarse resolution around Antarctica does not allow to represent the regional conditions (Heuzé et al., 2013;Sallée et al., 2013b). They are subject to unrealistic open-ocean convection (Heuzé et al., 2013;565 Meijers, 2014; Sallée et al., 2013a) instead of convection on or near the continental shelf (Årthun et al., 2013;Nicholls et al., 2009). All these taint the hydrographic structure along Antarctica's coasts. Hence, any improved parameterization can not rectify the existing biases in the ocean forcing. These biases are reduced if we apply ocean temperature anomalies on top of an observational-based climatological data set as performed in our study.
Since we extrapolate coastal ocean temperatures laterally into the ice-shelf cavities, the obtained ocean warming might be 570 higher if we would include the rise of the strongly warming gyre centers. If this may have been incorporated in the forcing of other groups obtaining a higher ice loss, depends on the setup details. However, it may help to bridge the gap between other studies and our simulations.
Nevertheless, the detected sea-level decrease for the used precipitation anomaly forcing is in agreement with a growing surface mass balance since 1800 AD, driven mainly by the Antarctic Peninsula region (Thomas et al., 2017). During intensive 575 El Nino years, the accumulation-driven ice height increase between Dotson Ice Shelf and Ross Ice Shelf exceeds the height reduction by basal melting processes (Paolo et al., 2018), but the ice mass is still decreasing since the low-density snowfall replaces ice with a higher density. The stability arguments of Ritz et al. (2015) confirm the apparent stability of Antarctica in our simulations. Furthermore, various recent ice-sheet model simulations, driven by selected CMIP5 climate model fields in the framework of the ISMIP6 exercise, are subject to a negative sea-level contribution under a warming climate (Seroussi et al.,580 2019b).
To evaluate the impact of the precipitation boundary condition, fully coupled simulations between a dynamic ice-sheet/shelf model and a global climate model are inevitable. The system would include the ice-shelf-ocean interaction of coupled oceanice shelves at a sufficiently high spatial resolution around Antarctica. In addition, it would contain a sophisticated surface mass balance computation. We hope these coupled atmosphere-ocean/sea-ice-ice-sheet/shelf models will overcome the discussed 585 limitations.
Code and data availability.     29 https://doi.org/10.5194/esd-2019-78 Preprint. Discussion started: 6 January 2020 c Author(s) 2020. CC BY 4.0 License. Table 4. Defined areas as part of our diagnostic. The fraction is computed relative to "glaciered." Figure 4 and Figure 5 depict these areas.   Table 2). The 2m-air temperature (a) and the total precipitation (b) are mean fields from the regional RACMO model, while the ocean temperatures come from the World Ocean Atlas 2009 (c); see Table 2 for more details. Each reference field has its colorbar above its plot. Below each reference field, the related anomalies, including their colorbar, are compiled for the period 2071-2100. Here, the second (third and fourth) row shows the anomalies for RCP8.5 (RCP4.5, RCP2.6). In these atmospheric anomaly plots, the dark-gray line follows the current coastline. All potential ocean temperatures (c, f, i, l) are a vertical mean of the depth interval from 150 m to 500 m. The white contour lines in the anomaly plots highlight the following precipitation threshold (e, h, k): 30 cm/yr. All these anomalies are the ensemble mean of the models listed in Table 1; please note that CCSM4 is not part of RCP2.6. Appendix Figure A1 and     (Table 4) named "WAIS", "EAIS Atl", and "EAIS Ind" are outlined by their green, blue, and red, respectively, boundaries (lower left legend).

Ross Sea
For further details, inspect the text, please. Appendix Figure A3 provides corresponding distributions for each climate model ensemble member. Antarctica's contour is deduced from Fretwell et al. (2013). Please note, that the RCP2.6 scenario does not include the CCSM4 model; hence, the corresponding bar is hatched.  and Amundsen Sea warm water masses that flow already into ice-shelf cavities and drive the highest basal melting rates, the potential to trigger Marine Ice Sheet Instability (MISI) exists, because WAIS has a retrograde bedrock topography. The massive ice shelves, Filchner-Ronne, Ross, and Amery, are influenced by moderate temperature increases. However, our setup misses the interaction between the ice-shelf topography and the underlying dynamically evolving ocean. Hence, the setup does not describe related circulation changes that may bring warmer water masses into the ice cavities. For instance, it has been found 860 that warmer water masses could find their way into these ice-shelf cavities and cause a strongly amplified basal mass loss under a changing climate (Hellmer et al., 2012). They have simulated an ocean warming by more then 2 • C in the Filchner Trough (eastern Filchner-Ronne Ice Shelf). At the ice shelves edge of the Filchner-Ronne Ice Shelf, our ensemble maximum ocean temperature anomaly ( Figure A1) of about 1.5 • C generates a much weaker forcing.

865
If one calculates temperature scaling factors out of the CMIP5 model simulated temperature and precipitation changes, it turns out that the temperature scaling factor of the precipitation is different for each model and therefore shows an inhomogeneous spatial pattern (Appendix Figure A3). Furthermore, the details of the scaling factors depend on if we determine the anomalies, which drive our ice-sheet simulations, relative to the first or last 50 years of the corresponding piControl runs. If we alternatively compute the anomalies relative to the first 30 years of the transient historical period , we obtain also slightly 870 different results. However, all these differences do not changes the spatial structure significantly, and they have a neglectable impact compared to the choice of the driving model.
The scaling across all model tends to be highest for the EAIS, where the part facing the Atlantic Ocean exhibits highest scalings ( Figure 5). The WAIS has a lower scaling and the embedded region "Siple Coast" has on average the lowest scaling.
There is a tendency for a higher scaling under a more vigorously changing climate across all regions, except for the smallest 875 region "Ross." This tendency exists for the ensemble mean and across models characterized by a larger than average scaling.
The detected precipitation deficit in reanalysis data and shallow ice cores in the "Siple Coast" region (Wang et al., 2017) is represented by most models by a reduced scaling, while only NorESM1-M shows a precipitation deficit under a raising air temperature for all future climate scenarios. When considering the whole Antarctica, the difference between the grounded ice sheet only and all glaciated regions (including ice shelves) is small.

880
The highest scaling spread between the first and last 50 years piControl reference period has MIROC-ESM across all inspected regions and scenarios, which is probably related to the pronounced trend of the global 2m-air temperature (0.67 • C) between these two reference periods in our ensemble. Otherwise, the spread is related to enhanced/amplified long-term re-gional climate variability expressed by differing values in the reference period. For example, CCSM4 or MPI-ESM-LR are subject to a larger spread in the Atlantic sector of the EAIS, while in the neighboring Indian sector the variability is negligible.

885
The higher spread of the smaller subregion Ross within the WAIS sector supports this interpretation (at least for the models CCSM4, CanESM2, HadGEM2-ESM).
There exists a tendency towards a higher scaling of coastal areas that are subject to incoming storm tracks, which potentially deliver heavier precipitation events that are also controlled by the rising topography height towards the interior of Antarctica.
In the majority of the simulations, we identify a lower scaling in WAIS and also a low to negative scaling in the area of the 890 Ross Ice Shelf and the adjacent parts of the WAIS.

A2 Marginal ice loss by ocean-driven basal melting and iceberg calving
We turn our analysis to the individual mass balance terms: Iceberg calving, basal melting in the ice-shelf cavities, and surface mass balance. To recap: the surface mass balance is obtained by applying the individual spatial atmospheric model forcing on top of the reference fields obtained from RACMO, while the basal melting is calculated by adding ocean anomalies on top of 895 the World Ocean Atlas climatology ( Table 2). The calving is composed of three processes (thickness calving, Eigen-calving, kill mask calving) as part of the Parallel Ice Sheet Model (PISM) simulations. Here, the analysis focuses predominantly on the period from 1850 to 2100, because after 2100, we reapply the forcing from 2071-2100 recurrently.
Until 2100, the temporal evolution of the iceberg calving rates of individual ensemble members is subject to some variability, which is typical for such event-based mass losses. For some models, we could identify some reduced calving of 20 % around  Rignot et al., 2013). Also the combined observed mass loss by calving and basal melting of ice shelves, which is about 915 calving rate from our ensemble members starting from PISM2Eq. Therefore, our ensemble mean ice loss rate exceeds current estimates, which could lead to an overestimation of the total sea-level rise in our simulations.
The basal melting rate of floating ice shelves, in the following just termed basal melting rates, is the second ice mass loss process. The basal melting rate anomaly is computed relative to the 50 years between 1951 and 2000. We could identify 920 immediately that the basal melting rates have risen between 10 % and 100 % since the 1850s ( Figure A13). The simulated historical trend is nearly independent of the initial state selection (PISM1Eq and PISM2Eq) and reference to compute the ocean temperature anomaly. For each climate model scenario, the anomalies are computed relative to the first or last 50 years of the pre-industrial climate (piControl) simulations. However, only for MIROC-ESM the reference state (first vs. last 50 years piControl) matters, because this model is subject to not negligible trend during the piControl phase.

925
In the future, the basal melting rate will further increase between 10 % and more than 100 %. The latter increase is consistent with results from specialized ocean simulations. These simulations resolve ice shelves, include the ocean-ice-sheet interaction explicity, are driven by future projection from various climate models (Naughten et al., 2018;Hellmer et al., 2012).
The temporal evolution of the actual basal melting rate ( Figure A13) increases until 2100 and falls back afterward onto the value of the year 2071 because we apply the last 30-years-forcing recurrently after 2100. Also, for the basal melting the 930 separation of ensemble members starting from PISM1Eq and PISM2Eq is apparent. However, both groups are close to the ensemble mean, which is in contrast to the calving rate. The basal melting rates of all ensemble members underestimate the observational basal melting rates.
Since, in general, the observed calving rate is lower than the basal melting rate, our model ensemble swaps the importance of basal melting and iceberg calving. Also the sum of the calving rate and basal melting rate exceeds the observed estimates.

935
Hence, our simulations could tend to overestimate ice loss and, ultimately, sea-level rise.
The ensemble mean calving and basal melting rates stay nearly constant or reach a maximum of around 2100 and scenarios with a higher forcing (RCP8.5 vs. RCP4.5, for instance) cause more ice loss by both calving and basal melting. Beyond 2100, ice loss rates decrease in general ( Figure A14). Since the temporal variability remains high also after 2100, our approach works where the last 30 years of the forcing until 2100 is recurrently applied afterward. To highlight the primary trend in the temporal 940 evolution after 2100, a 250-year running mean is applied after 2100.
The basal melting rates of the stronger forcing scenario RCP8.5 show a minimum of around 3500 and increase afterward slightly, while the other scenarios (RCP4.5 and RCP2.6) indicate a tendency for stabilization at the end of our simulation in the year 5000 ( Figure A14). Over the entire period, the basal melting rate is higher for the stronger forcing scenarios. This result reflects the dependence of the basal melting on the ocean temperature because the ocean temperature anomalies are warmer 945 for the scenarios with a higher forcing.
The calving rates before 2100 tend to be slightly higher for the RCP8.5 scenario. However, after 2100, we detect the sharpest fall of the ice loss rates for the scenario RCP8.5 and an average decrement for RCP4.5 and a moderate reduction for RCP2.6 ( Figure A14). Around 3000, RCP8.5 calving reaches its minimum, followed by an enhanced increase for 500 years and a moderate increase afterward. Forcing scenarios with a lower strength reach the minimum later, so that RCP4.5 has its minimum 950 around 3200, while RCP2.6 shows the minimum around 3700. At this time, the ensemble mean calving rates of the RCP4.5 and RCP2.6 are similar (please note that RCP2.6 does not include simulations driven by CCSM4). The trends of all scenarios converge around 4000.
In the long term, the most active basal melting occurs for the stronger forcing scenarios, while the highest calving occurs under scenarios with a lower forcing. The calving rate controls the evolution of the total ice mass loss in our simulations. Before 955 the year 2100, RCP8.5 has the highest calving rates, while these are lowest shortly afterward. The ordering of the scenarios with the highest calving rates (RCP2.6) is those with the lowest forcing (RCP2.6) and vise versa (RCP8.5). The ensemble mean of the basal melting increases by 60 %-70 %, 70 %-85 %, and 90 %-115 %, for RCP2.6, RCP4.5, and RCP8.5, respectively.
The fractional calving change of the ensemble mean is between +2 %-−4 %, +2 %-−10 %, and +2 %-−19 % for RCP2.6, RCP4.5, and RCP8.5, respectively. Across these, we detect that the most substantial ice-shelf area reduction occurs for RCP8.5 960 and the lowest for RCP2.6. Our simulations suggest that the warmer climate causes a stronger ice-shelf retreat and a stronger drop in the calving rate in the period, where the ice shelf could adjust to the quasi-equilibrium forcing. Based on these results we conclude: warmer climate drives more basal melting and enhances calving so that we obtain smaller ice shelves. Under the assumption that the smaller total integral of the ice shelves area is proportional to the calving probability (under else nearly unaltered conditions), we ultimately reduce the calving rates.
970 and the fraction of the original simulated flux to the reference flux (F ref ) The corrected flux F cor using Equation A1 is defined as With Equation A2 we obtain where z l is the sea level and m ice the total ice mass, which includes grounded and floating ice. We use here p = median(p(t)) so that each ensemble member is characterized by one value for its entire time series. If p = 1 ρAoce , 100 % of flux difference (Equation A4) contributes immediately to the sea level of the global ocean with an area of A oce .
The total ice mass (m ice ) changes are driven by four terms where F SM B (t) is the surface mass balance flux and F G (t) the basal mass flux of grounded ice ( Figure A9). We assume that these two terms in the brackets do not change regardless of the applied corrections to the last two terms F B and F D . Hence the difference in the ice mass change is Now we relate the temporal evolution of the sea level to the total ice mass changes by utilizing the Equation A5 so that we obtain: where the sea-level difference ∆z l (t) is The Figures A5 and A7 depict the sea-level difference for two cases. If the additional mass loss contributes immediately to a rising sea level ( Figure A7), the sea level rise of 30 cm is larger than the actual sea level rise since 1850 of about 20 cm (Church and White, 2011).This case is not realistic because a melting floating ice shelf does not impact the sea level. Only the flow of grounded ice across the grounding line to feed an ice shelf or the direct loss of grounded ice rise the sea level.  Figure A8 shows the proportion of the deduced ratio to the 100 % ratio. Only very few ensemble member losses about 15% of the maximum value of p = 1/(ρ · A oce ). In contrast, the mean and median value of this proportion is generally less then 5 %.
For all ensemble members driven by the precipitation anomaly, this proportion is on average 4.7 % with a median of 3.9 %.
It is even lower for ensemble members driven by the temperature scaled precipitation. The median amounts 0.7 % and the corresponding mean is 0.9 %. Please note that some ensemble members under the temperature scaled precipitation are subject 1010 to a negative scaling. This result confirms the above presented low positive and negative scaling seen for restricted regions ( Figure 5). It highlights also that simulations driven by temperature scaled precipitation could show unexpected results.
48 https://doi.org/10.5194/esd-2019-78 Preprint. Discussion started: 6 January 2020 c Author(s) 2020. CC BY 4.0 License.  Table 2). The 2m-air temperature (a) and the total precipitation (b) are mean fields from the regional RACMO model, while the ocean temperatures come from the World Ocean Atlas 2009 (c). Below each reference field, the related maximum anomalies are compiled for the period 2071-2100. Here, the second (third and fourth) row shows the anomalies for RCP8.5 (RCP4.5, RCP2.6). The dark-gray line follows the current coastline. All potential ocean temperatures (c, f, i, l) are a vertical mean of the depth interval from 150 m to 500 m. The white contour lines in the anomaly plots highlight the following thresholds. 2m-air temperature (d, g, j): 8 • C; total precipitation (e, h, k) 50 cm year −1 . All these anomalies are the ensemble maximum of the models listed in Table 1; please note that CCSM4 is not part of RCP2.6. Figure 2 shows the corresponding ensemble mean fields.  Table 2). The 2m-air temperature (a) and the total precipitation (b) are mean fields from the regional RACMO model, while the ocean temperatures come from the World Ocean Atlas 2009 (c). Below each reference field, the related minimum anomalies are compiled for the period 2071-2100. Here, the second (third and fourth) row shows the anomalies for RCP8.5 (RCP4.5, RCP2.6). The dark-gray line follows the current coastline. All potential ocean temperatures (c, f, i, l) are a vertical mean of the depth interval from 150 m to 500 m. The white-grey lines in the anomaly plots highlight the following thresholds. 2m-air temperature (d, g, j): 0 • C; total precipitation (e, h, k) 0 cm year −1 ; potential ocean temperature 0 • C. All these anomalies are the ensemble minimum of the models listed in Table 1; please note that CCSM4 is not part of RCP2.6. Figure 2 shows the corresponding ensemble mean fields.  Figure A3. Temperature scaled precipitation under the RCP8.5 scenario for nine CMIP5 models (Table 1): Period 2051-2100. The ice-sheet simulations are driven by anomalies relative to the first 50 years of the related piControl climate scenario. In the dotted regions enclosed by black contours, the combined simulated scaling and the standard deviation contains the value of 5 %K −1 . Gray dashed lines follow this 5 %K −1 contour. The scaling values deduced from ice cores are shown at their location (mean and the 2-sigma uncertainty). The regions named "WAIS," "EAIS Atl," and "EAIS Ind" are outlined by their green, blue, and red, respectively, boundaries (lower left legend). For further details, inspect the text, please. Figure 4 shows the corresponding ensemble mean. The contours of the Antarctic continent are deduced from Fretwell et al. (2013). rcp85 -anomaly rcp45 -anomaly rcp26 -anomaly rcp85 -scaled rcp45 -scaled rcp26 -scaled Figure A5. The diagnosed sea level correction as defined by Equation A8 covers the period from 1850 until 2100. Here, the ratio p(t) (Equation A5) is the temporal median for each ensemble member (please see Figure A7 for the corresponding figure, where all additional mass loss rises immediately the sea level). The correction is computed relative to the year 2000 as in Figure 12. The resulting sea level for the entire period from 1850 until 5000 depicts the Figure A6 Figure A8. The ratio between the actual sea level contribution due to mass loss and the sea level equivalent of corresponding mass. Individual ensemble members are shown as crosses. A red "x" represents a member that is driven by the precipitation anomaly, while a blue-green "+" indicates those driven by the temperature scaled precipitation. In the latter case, light blue circles highlight members with negative ratios.

A4 Appendix Figures
Vertical lines mark median values for these two groups; see also legend. The corresponding mean values are listed on the left side. The term p(t) is defined by the Equation A5 and p100 = 1 ρ·Aoce , where ρ is the density and Aoce represents the global ocean area. as a measure of the variability among the ice-sheet ensemble members driven by various climate models (Table 1). Depoorter (2013) Rignot (2013) Liu (2015) Depoorter (2013) Rignot (2013) Liu (2015) Basal Melting Calving condition. For the RCP8.5 scenario, the shading highlights the standard deviation (1-sigma) as a measure of the variability among the icesheet ensemble members driven by various climate models (Table 1). Please note the different axes for both subfigures. The dashed frame in the right subfigure depicts the value range of the left subfigure.