The Greenland Ice Sheet (GrIS) is a major cryospheric contributor to global sea level rise , with numerical models predicting accelerating, although uncertain, GrIS mass loss throughout the 21st century and beyond . However, numerous aspects of the GrIS's thermodynamics and hence motion remain enigmatic, including the widespread presence of large (greater than 1/3 of total ice thickness) englacial plumes found by tracing reflections of equal age in radargrams (i.e., isochrones; Figs. A, , ; ). Mapping and visual inspection of automatically-tracked disrupted radiostratigraphy shows that these large plume-like features (hereafter plumes) are mostly found in the northern part of the GrIS (Figs. A, ) but a consensus formation mechanism has not yet been identified. Although the plumes themselves are unlikely to be critical in interpretation of ongoing mass loss processes, clarifying their formation mechanism may reveal important information about the rheology, basal thermal state, and stability of the locations where they are or are not found – ultimately improving representation of ice flow in model projections.

These plumes have previously been hypothesized to result from basal freeze on , or travelling basal slippery spots , which both require that the bed be at least locally or temporarily thawed. Separately, and show that convergent flow, rheological anisotropy, and a rough bed are sufficient to form small-scale (<100 m) folds, but that, when basal slip and freeze on processes are excluded, large-scale folds require density gradients induced by thermal expansion and significantly lower viscosity. Another way to describe such temperature- and buoyancy-driven fold formation – and which lies along the same process continuum – is convection. In thermal convection, layers of ice heated geothermally from below (or cooled from above) thermally expand at the bottom (or thermally contract at the top), creating an unstable density gradient and forcing a vertical flow of material.

Here, we use “local convection” to refer to a temperature- and density-controlled process generating self-sustaining upwards motion and disrupted plume-like structures emanating from the bed, superimposed upon primary thermodynamic processes driving interior ice towards the ice sheet's margins, that are relatively isolated spatially. We explore whether ice convection – a process with a contentious history in theoretical glaciology – can explain observations of these large plumes, also known as disrupted basal units. previously proposed convection within ice sheets, but only for full-thickness convection (rather than stagnant-lid) and attracting strong objection . Both Hughes and Fowler approach convection analytically only, by estimating a Rayleigh number Ra, the dimensionless ratio of heat transfer via upwards mass transport (i.e. convection) vs. thermal conduction (Appendix A1, ). In these analytical models, convection initiates when a critical Rayleigh number is exceeded (∼650–1700 in and ), reached already if a 2500 m ice column has a uniform effective viscosity below 4×1014 Pa s (see below and Appendix A1 for further information on effective viscosity). While Hughes and Fowler found Ra values close enough to the critical value to warrant consideration of convection, applying a purely analytic approach to the GrIS is not ideal. The formulation of thermal diffusion in Ra does not capture dynamical effects important in ice sheet flow, such as horizontal shearing or downwards motion from snowfall; and the critical Ra value is itself tied to the particular boundary conditions and the initial perturbation geometry (e.g., ) making an analytical approach challenging for a dynamically and materially complicated system perched close to the onset of convection behaviour. We therefore consider a direct numerical modelling approach more appropriate for investigating the question of convection in terrestrial ice sheets.

We use the geodynamics software package ASPECT 2.5.0 to test our convection hypothesis. The setup is adjusted to simulate a 25 km along-flow two-dimensional slice through an ice sheet (Fig. ), or a 22 km along-flow by 18 km across-flow three-dimensional cuboid (Fig. ). ASPECT is used in place of a conventional ice sheet model due to its extensive benchmarking in convection problems and built-in functionality to model buoyancy forces, which are lacking in modern ice-sheet models. Similar geodynamics models have been previously used to study convection in the shells of icy moons . To facilitate a broad parameter sweep at low computational expense, and to isolate the influence of the parameters in question, we simplify the domain to have a uniform ice thickness (2.5 km as a reference value).

For all simulations except those that explicitly consider snowfall surface mass balance is set to zero, i.e., no snowfall or surface melting (vz,s=0 at the surface boundary condition where v=(vx,vy,vz) is the velocity field, the subscript x represents the along-flow distance, the subscript z is depth, and the subscript y is across-flow distance in 3-D simulations). When applied, surface shearing velocity vx,s is uniform across the domain's top surface with the rigid vx,b=0 condition maintained at the base. This Dirichlet velocity condition on a fixed surface differs from a “standard” ice sheet model, where surface velocity is an emergent result of ice geometry and flow parameters, but is suitable for our purposes as we treat surface velocity as an independent variable in simulations. Regardless, the net effect on background (i.e., not convection controlled) stress and strain fields is similar. Keeping vx,b=0 is likely a firmer control on basal velocity than the possible décollement observed in radargrams (Fig. C). Therefore, while increasing vx,s in the model can simulate plume behaviour as actual surface velocity increases (Fig. C), the comparison is not one-to-one. In reality, surface displacement could also be accommodated by basal slip or a thin shear layer beneath the plumes (perhaps visible in Fig. ) meaning our modelled plume behaviour represents the lower limit of stratigraphic disruption under a given surface velocity. Similarly, the placement of the initial perturbation in snowfall runs will influence the balance between horizontal and vertical velocity components (Fig. ). Further, snowfall, surface velocity, and ice thickness all exhibit moderate variation over the millennial timescales important for convection . Resolution is determined by the ASPECT requirement to set grid spacings as a given number of even divisions. In the case of a 2500 m thickness and 6 divisions this gives horizontal and vertical resolutions of 390 and 39 m, respectively.

In most ice-sheet models, stress is related to strain rate with the Nye–Glen isotropic flow law 1ϵ˙=Aτen-1τ, where ϵ˙ is the strain-rate tensor, τ is the deviatoric stress tensor, τe2=12tr(τ2) defines the effective stress τe (Pa), n is the flow exponent generally assumed as 3 or 4 , and A=EA0exp-QR1T-1T0 (Pa⁻ⁿ a⁻¹) is the creep parameter, where E is the enhancement factor, A0 is the creep prefactor, Q (J mol⁻¹) is the activation energy, R is the ideal gas constant (J mol⁻¹ K⁻¹), T0=263.2 K, and T (K) is ice temperature. E is defined as ϵ˙m/ϵ˙o where ϵ˙m is the measured strain rate and ϵ˙o is the strain rate predicted by Eq. (). The value and influence of E therefore varies depending on the choice of A0 and n; we use the default values in as a widely used reference (Table ), which makes comparison with existing ice-sheet models more straightforward. E varies based on deformation type but is often assumed to be ∼4–6 for the GrIS when using n=3, though it has been inferred to be up to 12 in Antarctic shear margins and even 120 in mountain glaciers . In most ice-sheet models, separate Q values are used for high (T>T0≈263.1 K) and low temperatures, but for simplicity in ASPECT here we simplify this to one mid-range value, which has limited effect (Fig. ). Moreover, the pressure dependence of A is neglected.

We also simplify to a Newtonian rheology for ASPECT by setting τen-1 constant. We use n=3 with τe=50 kPa as a reasonable starting value (Figs. D, ) based on an Ice-sheet and Sea-level System Model simulation with further information, and justification for a constant τe value provided in Appendix A2. A Newtonian rheology is appropriate here as strain rates due to convection are small compared to those from background ice flow (Fig. ), and convection can be considered a secondary phenomenon in this sense, though the implications of the rheological setup and choice of τe are covered further in the Discussion and Appendix A2. Prescribing τe is also necessary as a full ice-sheet stress state can not be accurately replicated in a simplified along-flow slice. Effective viscosity η can then be calculated as 2η=12Aτen-1-1. Rearranging Eq. () yields the functional dependence of E: 3E=cη-1τe1-n, where c=(2A0)-1exp⁡QRT-1-T0-1. The temperature-dependent viscosity is then controlled by varying E (Fig. ) as a tool to test rheological variation. We furthermore describe some results using E but note that the actual effective viscosity remains a complex function of ongoing and historic stress and deformation states.

ASPECT solves the governing equations of convection, 4∇⋅v=0(conservation of mass)5-∇⋅2ηϵ˙+∇p′=-βρ¯T′g(conservation of momentum)6ρ¯Cp∂T∂t+v⋅∇T-∇⋅κ∇T=Fint(conservation of energy) where p (Pa) is pressure, Cp (J kg⁻¹ K⁻¹) is heat capacity, κ (W m⁻¹ K⁻¹) is thermal conductivity, β (K⁻¹) is thermal expansion coefficient, ρ¯ (kg m⁻³) is the reference density, g (m s⁻²) is acceleration due to gravity, and Fint (W m⁻²) is the sum of all other heating terms. We set Fint=0, thereby ignoring adiabatic heating and neglecting strain heating, to prevent simulations with greater vx,s and hence greater strain heating from evolving a different rheology along flow, though we note that such heating will soften ice and may further facilitate convection. These equations follow the Boussinesq approximation – that density variations are small enough to be neglected everywhere except for in the buoyancy term βρ¯T′g – which is valid for very slow-flowing materials without abrupt density changes. This solution method simplifies the temperature field to T=T¯+T′, where T¯ is a constant reference temperature and T′ is the temperature perturbation; analogous perturbations are formed for the pressure and density fields.

Two baseline temperature profiles are used, representing the colder NEEM ice-core site in northern Greenland and the warmer DYE-3 in southern Greenland respectively (Figs. A, ). We apply a transformation, T2=T1+TaTb(Tb+Ta) where T1 is the original temperature profile, Tb is the basal temperature and Ta is an adjustment term used to raise the basal temperature to -2 °C. The temperature profiles are stretched and compressed when adapted to the range of ice thicknesses. This is not a perfect representation of ice-sheet ice temperature, but allows more direct comparison between simulations of different thicknesses, and other temperature estimation methods are subject to their own uncertainties. As we keep the basal temperature uniform across the domain (Dirichlet condition, therefore indirectly neglecting geothermal heat flux which nonetheless modulates through a realistic range over our simulations; see Results, Fig. ) and also want to consider ice some distance from the ice core site, a slightly higher fixed basal value is appropriate as a midpoint between the ice-sheet interior and margins. Above ∼3000 m, the basal temperature is then technically above the pressure-melting-point, even though a non-slip condition is imposed at the base at all times. Simulations where H>3000 m comprise a small proportion of our overall ensemble and we do not change the basal temperature for this subset. An initial temperature perturbation replicating a fold is created 5 km in from the inflow side (3.5 km in the case of snowfall simulations) using two Gaussian functions of opposing signs. We refer to two temperature perturbation sizes for 2D runs: large, used for most simulations (Fig. ), and medium (Fig. ). In the 3D runs a simpler approach is taken, with a cube of uniform 273 K ice measuring 3000 m × 5000 m × 750 m as the initial perturbation. As the initial temperature gradient is not linear (Fig. ), using a larger initial perturbation allows convection to occur in a more realistic temperature field without a delay for initial plume development, though results are relatively insensitive to the initial perturbation (Figs. B, C, ). ASPECT input files and scripts to recreate the temperature perturbations and perform other post-processing operations are provided in the Open Research Section. Values for set parameters are given in Table  and Figs.  and illustrate boundary conditions.

For each temperature profile we focus on the influence of four variables on the maximum upwards-directed vertical velocity, max(vz): the enhancement factor (E), shear velocity (vx,s), ice thickness (H), and snow accumulation rate (vz,s) (labelled in Table , Fig. ). Additional 3D simulations are included for runs B and F, which consider the parameter space covering observed plumes (NEEM temperature profile). Defining a threshold for convection under a given parameter space is not straightforward, but we focus on max(vz) over time as a reasonable indicator. Nonetheless, even if max(vz) trends towards zero over time, the englacial stratigraphy will still be slightly disrupted during this transition period. The total buoyancy forces in the 3D simulation will also be greater than in 2D as we are able to model an isolated plume rather than a laterally extensive fold. We divide behaviour into three zones, focusing on the 2D simulations that cover a broader parameter space. Suppressed convection is defined where max(vz) at 20 kyr is below 0.01 m yr⁻¹ or where max(vz) at 20 kyr has dropped by 0.03 m yr⁻¹ or more relative to its value at 4 kyr. Amplifying convection is defined where max(vz) at 20 ka exceeds 0.4 m yr⁻¹ or where max(vz) has increased by 0.1 m yr⁻¹ or more between 4–20 kyr ka. Sustained convection then occupies the space between these two zones. This approach allows us to isolate which parameter combinations may produce sufficient upwards flow to account for the distribution of large englacial plumes (Fig. A).

While our modelling is substantially more sophisticated than calculation of a single Ra value, the general behaviour in our simulations can still be understood in terms of the Ra number (Appendix A1), with greater values of E and H prompting convection. Given the additional complications of varying surface velocity, snowfall, initial perturbation, and viscosity profile, we find that there is no single critical value of E that describes this transition, but Figs. , suggest that, for τe=50 kPa, 45≤E≤75 encapsulates a range of behaviour for the NEEM temperature profile sufficient to form features similar to those observed in radiostratigraphy (Fig. ). Considering shear over the domain of 1 m yr⁻¹ with no snowfall (Fig. B) in 3D, max(vz) begins to increase from 4–14 kyr between 40<E<50, around the same point at which 2D convection is considered to be sustained under our definition. In Fig. B at E=75, max(vz) is consistently increasing over time and when E=60 convection is still classified as sustained for snowfall rates exceeding 0.15 m yr⁻¹. Figure  shows suppressed, amplifying, and sustained behaviour as a time series under different scenarios.

DYE-3 requires much lower E values to transition between suppressed, sustained, and amplifying zones compared to NEEM when other parameters are equivalent (Fig. a; cf. Figs. A, B and A, C). The steep temperature gradient at the base of the DYE-3 profile is sustained over a shorter height, resulting in lower buoyancy forces overall, but this is compensated by lower viscosity in the upper portion of the domain (Fig. ). Ice thickness is an important factor for both profiles, with sustained convection becoming infeasible below thicknesses of around 2000 m (hence its use as a boundary in Fig. B), though note that ice thickness is also intertwined to an extent with its influence on the basal temperature gradient. Vertical transport rates begin to level off with increasing ice thickness for the DYE-3 profile, pointing towards a maximum rate of upwards motion. However, surface velocity and snowfall exert perhaps the most important overall constraints on max(vz), with convection becoming infeasible as surface velocity in our simulations increases from 1 to 3 m yr⁻¹ or as snowfall increases beyond 0.1 to 0.3 m yr⁻¹, with the precise cut-off value depending on E and the temperature profile (Fig. D–F, J–L).

The modelled plume geometry varies substantially across the parameter space. In the suppressed convection zone, a perturbation is still produced but is not sustained (Figs. D, D). Amplifying convection can prompt a self-sustaining plume chain with significant temperature variation (Figs. A, C). Sustained convection (Figs. B, F, L, A, B, E) produces plumes more similar to the folds observed in radargrams, though plume length is slightly shorter. 3D simulations produce roll-over in the down-flow direction, but characteristic symmetric overturning typical of convection in an otherwise static medium in the across-flow direction (Figs. D, E). Using a smaller perturbation in 2D simulations does not appreciably alter this pattern, but does slightly and consistently shift down the maximum vz over the simulation time frame (Fig. a). The heat flux required to sustain the fixed basal temperature (∼40–70 mW m⁻¹, Fig. ) is compatible with proposed rates of geothermal heat flux beneath the GrIS .

Our results suggest five main thresholds for convection to occur within the Greenland Ice Sheet: (1) an initial temperature and therefore density perturbation; (2) ice thickness must be greater than around 2200 m; (3) snowfall must be less than around 0.15 m yr⁻¹; (4) total horizontal shear through the column convection is occurring in must be less than around 1 m yr⁻¹; and (5) the effective viscosity profile range should fall within ∼2×1012–3×1014 (equivalent to an enhancement factor of ∼45–75 if τe=50 kPa) – roughly an order of magnitude lower than may typically be anticipated (Fig. ). Condition (1) is easily satisfied by bedrock perturbations (e.g. Fig. B, C) or basal folding induced by processes such as convergence and rheological anisotropy ; condition (2) is satisfied for a large region of the interior of the central ice sheet (Fig. B); and condition (3) is primarily satisfied in north Greenland (Fig. E). Condition (4) is satisfied by low surface velocities throughout large parts of northern central Greenland (Fig. C), with longer residence time being unique to the northern central ice divide (Fig. B); and ). However, observed plumes are always located in regions where surface velocity is >1 m a⁻¹ and sometimes found where surface velocity is >10 m a⁻¹. In part, this may arise from surface velocity being mitigated by basal slip and shear beneath the plumes (see also Methods) but this issue is addressed further as we progress through the discussion.

We consider the possibility that effective viscosities for Greenland may be lower than generally assumed for north Greenland, satisfying condition (5), after first discussing other aspects of plume morphology and distribution. First, modelled plume widths (Fig. ) are comparable but slightly narrower than observations. This may occur due to the continued disruption of plumes as the velocity field evolves after they have attained their maximum amplitude in a thicker, colder, slower, and larger palaeo-GrIS . However, this places plumes in a delicate balance between transport downstream and thinning to below the 1/3H observed threshold. Relatedly, from Eq. () effective viscosity is proportionally related to effective strain as 7η∝ϵ˙e(1-n)/n or, if n=3, as η∝ϵ˙e-2/3, where ϵ˙e2=12tr(ϵ˙2) is the effective strain rate. This creates another balance whereby increasing the effective strain (or effective stress through τe∝ϵ˙e1/n) reduces the effective viscosity, encouraging convection, but also increases ice-column disturbance, discouraging convection. While our experiments focus on along-flow slices, it is possible that this may assist explanations regarding the presence of tall plumes with off-axis orientations just outside ice stream margins where surface velocity is higher. Here effective strain is greater, and hence effective viscosity reduced, but through-column horizontal shear is not excessive (rather, rotation can increase effective strain while not significantly disrupting plumes). Modelled upwards velocity rates may push through pRES measurement and location uncertainty making analysis of repeat radar surveys a feasible way to test if these plumes are actively expanding. Notably, the convection plumes generated in our 3D simulations (Fig. E) more closely resemble the geometry of observed folds (Fig. C, D), and we are not aware of another mechanism that is hypothesized to produce this unique type of geometry.

Such settings are rarer or absent in southern Greenland, where escape times from the central ice divide to the 2000 m principal contour line are only a little over 10 kyr. The significant snowfall (≥0.35 m yr⁻¹) and hence downwards motion in south Greenland (Fig. E) also likely limits the possibility of convection in this region (Fig. J) and will have done for at least the past 9 kyr .

Considering existing hypotheses for the formation of the observed folds, basal freeze-on may be limited as a general explanation capable of explaining plume ubiquity in north Greenland given the requirements for large volumes of basal water in a region that is not likely to be pervasively thawed . Further, freeze-on may not explain the fairly consistent sizing and spacing (at ∼10 km, Fig. ) of some north Greenland plumes, which contrast the much more spatially extensive freeze-on layers in East Antarctica . Travelling slippery spots develop clearly in an controlled setup, but also require thawed bed areas in the same region to facilitate at least a degree of slip and further do not appear to align with the observation of a highly deformed basal layer beneath the plumes (Fig. C), which may be more consistent with high rates of basal ice deformation than basal sliding . Travelling slippery spots may also not be compatible with the 3D geometry of observed plumes (Fig. B, D) or with ice motion over a rough bed. Additionally, neither mechanism accounts for an apparent absence of H>1/3 plumes in south Greenland. However, the basal thermal state may have been different ∼10 kyr ago and we do not rule out these two processes contributing to the onset of an initial perturbation or playing a role in their continued development. Our imposition of a no-slip, constant-temperature basal boundary also means that possible feedbacks between convective heat dispersal and basal sliding are not recognised. These may complicate plume geometry in a similar manner to that explored in . Rheological contrasts and convergence , as covered in may also interact with convection in ways not explored here. Last, geothermal heat flux decreases throughout our simulations towards ∼4 mW m⁻¹. This will have some bearing upon model outcome but as outlined in the Methods a linear temperature profile with constant geothermal heat flux presents its own misrepresentations. In any case, we still obtain stable max(vz) time series for the sustained regime, while the amplifying regime may or may not reach a steady state (Fig. ). We highlight these possibilities to motivate further work on englacial plumes and more clearly determine to what degree it is necessary for convection to operate in concert with additional processes.

Is it possible that effective viscosity values lower than commonly assumed are feasible for the northern GrIS? Independent of convection being possibly the only feasible mechanism for large plume formation, we suggest that an affirmative answer may be appropriate. Large plumes are mostly found in areas with a relatively larger proportion of pre-Holocene ice (Fig. A). Beyond this observation fulfilling the requirement for relatively stable ice (condition 3), older ice from the Last Glacial Period is consistently measured or inferred to be significantly less viscous than Holocene ice , as a result of fabric development and a higher impurity content from a drier and dustier ice age. Despite the importance of this softer ice for interpretation of the overall motion of the GrIS, very few direct measurements exist. Borehole closure rates from ice divides likely reflect stresses inconsistent with basal shearing, and such locations are often explicitly selected for their lack of a history of extensive horizontal shear . To our knowledge, no laboratory measurements have been conducted on ice resembling that found within what we hypothesize to be englacial convection plumes. It may therefore be possible that basal ice in north Greenland is sufficiently soft as to permit convective plume formation. Alternatively, additional plume-forming processes operating in parallel may expand the effective viscosity range required for their formation. Similarly, parallelly-operating processes may also account for the presence of plumes in regions of slightly greater velocity than the threshold indicated here. More complex modelling featuring additional processes and tests on field specimens presents the clearest opportunity to directly asses our hypothesis.

In situ rheology is also modulated by anisotropy, which is not included in our simulations. suggest an important role for anisotropy in the formation of large plume-like folds (their Fig. 4) as a result of direction-dependent softening due to directional alignment of the c axis. Plume-forming motion will rotate initially bed-planar fabric such that it also broadly aligns with the dominant shear direction in plume formation. However, anisotropy itself describes a stress-orientation dependent rheology, rather than a specific softening. The role of anisotropy in then comes in part from their implementation which allows the viscosity acting along the plane perpendicular to the c-axis maximum (denoted η2) to decrease by a factor of three and fall below the 1×1013 Pa lower limit set for the isotropic run (their Table S2). Such a decrease is sufficient to reach the effective basal viscosity values (∼3×1012 Pa s) in our E=40 and E=60 simulations, where local convection becomes increasingly viable. This is not to say that progressive anisotropic softening is not an important process here. In glacier settings, bulk viscosity will generally decrease as ice develops stronger crystallographic anisotropy, though the effect is stress-state dependent , meaning such softening may also be a contributing factor for plume locations occurring at distance from ice divides. In our application, as in many others , the enhancement factor operates as a simple parameterisation of anisotropic effects without recourse to a stress-orientation dependent tensorial flow law. Consequently, our results provide a plausible estimate of how progressive anisotropic softening may affect η2, which likely exerts the strongest rheological control on convection onset.

Ice is more accurately represented as a non-linear shear-thinning fluid in most situations, and may also be more non-linear than the linearised approximation of n=3 implemented in this study, with growing evidence for n=4 in some regions . Our use of Newtonian rheology allows us to test a broad parameter space but may miss non-linear interactions caused by the plumes themselves which both increase and decrease effective strain rates and thereby influence the effective viscosity (Eq. ). Increasing values of n away from unity may increase the importance of these non-linear stress responses within the plumes; intuitively one may anticipate in a direction that more readily facilitates plume formation, though this will depend upon the appropriate values for A0 and Z and the resultant effective strain field. We emphasise, however, that rate-weakening in plumes is still expected to be small compared to the main coastward movement of the ice sheet, which exerts a first order control over effective stress (Figs. , ). In any case, we hope that our results closely isolate the effective rheological thresholds for ice-sheet convection, which permits a narrower starting point for future numerical models featuring more complex and computationally costly thermodynamics.

A lower effective viscosity of basal ice will significantly influence ice dynamics, similar to the influence of an increased flow exponent, n . If ice-sheet models are initiated under fixed assumptions of higher ice viscosity, then inversions for basal traction will overcompensate by producing unrealistically low basal traction values and bias the resulting projections . Convection-driven plumes also present a mechanism that draws warmer and lower-viscosity basal ice upwards – counteracted by the transport of higher-viscosity (colder) interior ice downwards. Exploring the possible implications of lower viscosity ice and convection-driven mixing – and their influence upon inferred basal traction – is therefore warranted to better quantify the errors that may be introduced into predictive ice-sheet models. Finally, the relative lack of large plume observations in the Antarctic Ice Sheet, outside of the Gamburtsev Mountains, may simply result from colder temperatures there and hence higher viscosities in the upper ice column, which limit convection , or from a sampling bias given the comparative paucity of radar-sounding observations in Antarctica .

Our modelling indicates that, following an initial perturbation, local convection is possible within the Greenland Ice Sheet under conditions that are not unrealistically far from the existing consensus on ice rheology. This hypothesis could explain the observed spatial distribution of large plumes in Greenland, with surface velocity, accumulation rates and ice rheology exerting the strongest controls on convection viability and hence plume formation. A corollary of this result is that ice in northern Greenland may be softer than commonly assumed. Appropriately probing and then implementing these constraints into ice-sheet models may help reduce compensatory errors and improve the accuracy of their future projections.