The bed topography, zb, is derived from profiles of airborne and ground-based common-offset ice-penetrating radar surveys distributed over Kronebreen from 2009, 2010 and 2014 . The initial surface topography includes different available surface digital elevation models (DEMs) and is described in .

Ice surface velocities were derived from feature tracking of TerraSAR-X image pairs in slant range using correlation windows of 200×200 pixels at every 20 pixels, and subsequently ortho-rectified to a pixel size of 40 m using a DEM . Images were acquired roughly every 11 days for the period May–October 2013. Uncertainties in surface velocity are estimated to be ∼ 0.4 md-1 and comprise a co-registration error (±0.2 pixels) and errors arising from unavoidable smoothing of the velocity field over the feature-tracking window. Ice-front positions were manually digitised from the same images used for feature tracking after they had been orthorectified to a pixel size of 2 m using a surface DEM .

We use surface velocity and frontal position data described above to test the effects of sliding and undercutting on calving using different models in a global approach. This one-way offline coupling approach is divided into three parts using six models (see Fig. ): inversion for sliding and computation of geometry evolution (with Elmer/Ice), determining undercutting (with the energy balance model, subglacial hydrology model, plume model and undercutting model) and computing calving (with HiDEM). In this paper, we use the output of five different models as input for the discrete particle model, HiDEM, in order to compare the modelled calving front to observations for different configurations of sliding, geometry and undercutting.

We set t0 at the velocity acquisition just before the first melt and the following observational times are set at each observation of surface velocity. The exact dates are summarised in Table .

First, we infer the sliding velocity at each observational time from surface velocities using an adjoint inverse method implemented in Elmer/Ice with an updated geometry from observations. At each iteration, i, corresponding to an observed front position, Fiobs, the front and the surface are dynamically evolved during the observation time interval (roughly 11 days) with Elmer/Ice with a time step of 1 day. By the end of the observation interval, the front has advanced to a new position, Fi+1elmer. Here we use i+1 because this is the position the front would have at ti+1 in the absence of calving. Second, given subglacial drainage inferred from modelled surface runoff, a plume model calculates melt rates based on the subglacial discharge for each iteration, which are subsequently applied to the front geometry at subglacial discharge locations. At each iteration, the front geometry takes into account the undercutting modelled at the former iteration. Finally, the sliding velocity, geometry and undercutting (when applicable) are taken as input to the calving particle model HiDEM for each iteration and a new front, Fi+1hidem, is computed for four iterations, i={0,4,6,11}, which represent interesting cases (see comments in Table ). More details about each aspect of the model process are given in the following sections.

We call this approach an offline coupling because inputs to the HiDEM are output results from Elmer/Ice and undercutting model but not vice versa. In Elmer/Ice, we use the observed frontal positions. A completely coupled physical model would use the output of HiDEM, the modelled front position, as input to the ice flow model Elmer/Ice and the undercutting model. It would also calculate the basal friction from a sliding law rather than an inversion. In principle, such an implementation is possible using the same model components as this study.

At the base of the glacier, we use a linear relation for sliding of the form τb+βvb=0, with τb the basal shear stress and vb the basal velocity. The basal friction coefficient, β, is optimised at each observational time to best reproduce observed velocity distribution at the surface of the glacier as described in . This is done by using a self-adjoint algorithm of the Stokes equations for an inversion e.g. and implemented in Elmer/Ice . The inversion is performed using the method of Lagrange multipliers to minimise a cost function including the observed horizontal surface velocities and a Tikhonov regularisation. We use an unstructured mesh, with spatial repartition of elements based on the mean observed surface velocities in the horizontal plane (roughly 30 m resolution close to the front). Vertically, the 2-D mesh is extruded with 10 levels (roughly 10 m resolution close to the front). More details on the Elmer/Ice modelling (viscosity, ice temperature, iterations, etc.) are given in .

After each inversion, the temporal evolution of the glacier is mathematically described by the kinematic boundary condition defined at the surface, ∂zs∂t+vx(zs)∂zs∂x+vy(zs)∂zs∂y-vz(zs)=a˙s(ti), which describes the evolution of the free surface elevation, z=zs, for a given net accumulation, a˙s(ti), calculated using a coupled modelling approach after , described in the next section. We use a time step of 1 day during the interval of time between two acquisitions. Equation () is solved alongside the Stokes equation, coupled to the latter by the velocities. The basal sliding velocity is not evolved and stays equal to the result of the inversion. When the front is advanced, the mesh is stretched to match the new front position. No new element or node is created and the basal sliding coefficients are extrapolated towards the new front. The new surface is in fact only used as an input for the next iteration. There is no interpolation of the basal sliding coefficients between two observational dates.

We assume that the front is vertical above the water line so that the observed front position (at the surface of the glacier) is the same at sea level. We call Fiobs(z=0), the front position observed at time ti with z=0 at the sea level and Fi+1elmer(z=0), the advanced modelled front position after Δt=ti+1-ti (see Fig. ). The front is advanced by imposing a Lagrangian scheme over a distance equal to the ice velocity multiplied by the time step. We do not account for the submarine melting during the advance because we only have observations at the beginning and the end of each time span. Instead, we lump frontal melting by applying an undercutting after the advance, as explained hereafter.

The surface mass balance, a˙s, and runoff are simulated with a coupled energy balance–snow modelling approach . The coupled model solves the surface energy balance to estimate the surface temperature and melt rates. The subsurface routine simulates density, temperature and water content changes in snow and firn while accounting for meltwater percolation, refreezing and storage. The model is forced with 3-hourly meteorological time series of temperature, precipitation, cloud cover and relative humidity from the Ny-Ålesund weather station (eKlima.no; Norwegian Meteorological Institute). Elevation lapse rates for temperature are calculated using output from the Weather Research and Forecast (WRF) model , while the precipitation lapse rate is taken from ; zero lapse rates are assumed for cloud cover and relative humidity. Surface runoff is modelled on a 100 m × 100 m grid.

The temporal subglacial discharge at the calving front is estimated from integration of daily surface runoff assumed to be directly transferred down to the glacier bed. Assuming the basal water pressure at over burden, the flow path of the meltwater towards the glacier front is determined from the hydraulic potential surface defined as ϕ=ρig(zs-zb)+ρwgzb, with g the gravitational acceleration. The grid is the same as that used for surface runoff. The flow path along the hydraulic potential surface is determined by D-infinity flow method where the flow direction from a grid cell is defined as the steepest triangular facets created from the eight neighbouring grid cells . The flow from the centre grid cell is distributed proportionally to the two cells that define the steepest facet. The flow is accumulated as the meltwater is routed along the calculated hydraulic potential surface towards the front and outlet points at the front are determined by identifying flow rates higher than 1 m3s-1. The hydraulic potential surface is filled before flow accumulation is calculated to avoid sinks.

A high-resolution plume model is used here to simulate the behaviour of subglacial discharge at the terminus of Kronebreen. The model is based upon the fluid dynamics code Fluidity , which solves the Navier–Stokes equations on a fully unstructured three-dimensional finite element mesh. The model formulation builds upon the work of , with the addition of a large eddy simulation (LES) turbulence model and the use of the synthetic eddy method (SEM) at the inlet .

The geometry of the model is adapted to Kronebreen by setting the water depth to 100 m and initialising the model with ambient temperature and salinity profiles collected from ringed seals instrumented with GPS-equipped conductivity, temperature and depth satellite relay data loggers (GPS-CTD-SRDLs) . These data were collected between 14 August and 20 September 2012 from a region between 1 and 5 km away from the glacier terminus and are taken as representative of the ambient conditions in the fjord during summer. Melt rates are calculated on the terminus using a three-equation melt parameterisation described by and and implemented in Fluidity by . Velocities driven by ocean circulation are typically around 2–3 orders of magnitude smaller than plume velocities and therefore neglected.

The model is spun up for 1000 model seconds until the turbulent kinetic energy in the region of the plume reaches a steady state and thereafter run for 10 min of steady-state model time. Melt rates are extracted from the duration of the steady-state period and then time-averaged and interpolated onto a uniform 1 m × 1 m grid covering a 400 m wide section of the glacier terminus.

The high computational cost of the model means that it cannot be run continuously over the study period, nor can the full range of discharges and oceanographic properties be tested. Instead, representative cases Md using the ambient ocean properties described above and discharges d of 1, 10, 50 and 100 m3s-1 were tested and the melt rate profiles for intermediate discharges were linearly interpolated from these cases.

We assume a vertically aligned surface front at the beginning of the melt season. We know the position of the front, F0obs(z=0), for the time span of each satellite image. The front is spatially digitised with 10 m spacing in the horizontal space and 1 m spacing in the vertical space. We use the combination of observed front, advanced front from Elmer/Ice and melt rates from the plume model to estimate the daily amount of undercutting. At each iteration, i, the sum of the daily undercutting during the observation interval is subtracted from the front.

When the first discharge occurs, the melt rate calculated with the plume model in 2-D is summed for the period of time between t0 and t1 and projected to the advanced front F1elmer(z=0) (advanced from F0obs(z=0)) at the location of the subglacial outlets and ice is removed normal to the front. This yields a new position of the front at depth z below sea level called F1elmer(z). At the second iteration, t2, we know where the front would be if there had not been any calving between t1 and t2: F2elmer(z=0), which is the advanced front from the observed position at t1, F1obs(z=0). Therefore we can transfer the whole undercutting from previous iteration to F2elmer(z) if F1obs(z=0) is situated in front of F1elmer(z) (see Fig. b–c). Otherwise, the undercutting would have been fully or partly calved away (see Fig. b–c). We then apply the new undercutting on this new geometry given the melt rates between t1 and t2.

At time ti, the modelled front position at depth z (advanced by Elmer/Ice from the observed front position at ti-1) is Fielmer(z) and the observed front position is Fiobs(z=0). We advance this observed front with Elmer/Ice during Δt=ti+1-ti to obtain the front position Fi+1elmer(z=0) at ti+1. We want to determine Fi+1elmer(z) and depth z given the melt rate calculated between ti and ti+1 and the state of the undercutting from the previous front Fielmer(z) updated by the observed front Fiobs(z=0). Three different cases, depending on the relative position of the observed and modelled fronts at depth z, are then possible as shown in Fig. :

if the new observed position Fiobs(z=0) is behind Fielmer(z=0) and in front of Fielmer(z), the melted undercutting is kept and advances in the flow direction the same distance as the surface modelled front Fi+1elmer(z=0) (see Fig. a);

The melt summed up between ti and ti+1 is then applied to Fielmer(z) to obtain Fi+1elmer(z) and so on. Frontal melt above the surface has not been taken into account so that the effect of submerged ice feet is not described. The bed topography, the new geometry (surface elevation, front position with or without undercutting) and the basal friction are then interpolated onto a 10 m × 10 m grid to feed the HiDEM and a new front, Fi+1hidem, is modelled after calving for the four selected iterations (i={0,4,6,11}).

The fracture dynamics model is described in detail in . This first-principles model is constructed by stacking blocks connected by elastic and breakable beams representing discrete volumes of ice. For computational efficiency, we use a block size of 10 m.

At the beginning of a fracture simulation, the ice has no internal stresses and contains a few randomly distributed broken beams, representing small pre-existing cracks in the ice. The dynamics of the ice are computed using a discrete version of Newton's equation of motion, iteration of time steps, and inelastic potentials for the interactions of individual blocks and beams. As the ice deforms under its own weight, stresses on the beams increase, and if stress reaches a failure threshold the beam breaks and the ice blocks become disconnected but continue to interact as long as they are in contact. In this way cracks in the ice are formed. For computational reasons, we initialise the glacier using a dense packed face-centred cubic (fcc) lattice of spherical blocks of equal size. This introduces a weak directional bias in the elastic and fracture properties of the ice. The symmetry of the underlying fcc lattice is, however, easily broken by the propagating cracks. The ground under the ice or at the seafloor is assumed to be elastic with a linear friction law that varies spatially (Eq. ).

The time step is limited by the travel time of sound waves through a single block and is thereby set to 10-4 s. If the stress in the ice exceeds a fracture threshold, crevasses will form and ice may calve off the glacier. The duration of a typical calving event at Kronebreen is a few tens of seconds followed by a new semi-equilibrium when the ice comes to a rest. The model runs for ∼100 s, which takes 2 days of computing time. As HiDEM cannot be triggered too often because of computational limitations, we simulate ice flow with Elmer/Ice and compute calving with HiDEM thereafter for the selected iterations. Calving events will then appear as fewer but bigger events compared to observations. If the time step is changed, the overall rate of change stays roughly within ±50 % .

The basal friction coefficients, β, at the front of Kronebreen are in the order of 108–1012 kgm-2s-1 , and in order to avoid instabilities building up, a cut-off value, above which particles are assumed to be stuck to the bed substrate, is fixed at β=1012 kgm-2s-1.

HiDEM reads a file with surface and bed coordinates on a grid and a file with surface and basal ice (to take into account the undercutting) coordinates. For simulations with an undercutting at a discharge location and in order to avoid complication in the HiDEM (position of the basal ice), we remove particles below the maximum melt (no ice foot as shown by the dashed line in Fig. ). In the ocean, the basal friction coefficient is extrapolated downstream of the front and taken equal to the mean of the values further up from the terminus in case the ice advances. An ice block is calved when all bonds are broken from the glacier even though it does not separate from the front.

There is a clear separation of timescales between the velocities of sliding (∼mday-1) and calving ice (∼ms-1). This gives us the opportunity to rescale friction so that we can more effectively simulate calving: even if we scale down friction by, for example, 2 orders of magnitude and increase sliding accordingly to ∼100 md-1, there is still negligible sliding during calving events which last tens of seconds or perhaps a minute. However, a rescaling speeds up the frequency of calving, and we can thus “speed up”, within reason, the few minutes of HiDEM simulation to effectively model calving which would otherwise take tens of hours or days and thus be practically impossible to simulate with HiDEM. By applying scaling, the calving events modelled during the simulation of HiDEM (a few minutes) correspond to the sum of calving events that would happen during the timescale of sliding. The scaling factor that we use is the same for the whole domain and for all simulations. We use a friction scaling factor for β equal to 10-2 (or sliding velocity scaled up by 102), and simulations run until calving stops and a new quasi-static equilibrium is reached.

In a fully coupled model, the altered ice geometry after calving could then be re-implemented in the flow model, acting as the initial state for a continued prognostic simulation with the continuum model. Here, this back-coupling is replaced by prescribing the next observed configuration.

The mean volumetric frontal ablation rate (or mean volumetric frontal calving rate) at the ice front at time ti, a˙c(ti), is the difference between the ice velocity at the front, vw(ti), and the rate of change of the frontal position, ∂L/∂t, integrated over the terminus domain Σw as defined in . This yields a˙c(ti)=∫Σwvw(ti)-∂L∂tdΣw, with ∫Σw∂L∂tdA=ΔA(ti)ti-ti-1∫z∈Σwdz and ΔA(ti), the area change at the terminus over the interval of time between two observations ti-ti-1. We want to compare the ablation rates from Fielmer for observed and modelled cases. For the observed case, the mean volumetric ablation rate is calculated between the advanced front Fielmer and the observed front Fiobs. For the modelled case, during one simulation with HiDEM, several calving events are triggered. Volumetric calving rate is then inferred from the difference between the initial, Fielmer, and final position, Fihidem, of the front, after calving has stopped. The total subaqueous melt rate, a˙m, at the front of the glacier is omitted in this balance.

We investigate the effect of three different parameters on calving activity: the geometry, gi, corresponding to the frontal position and topography; the sliding velocity, mainly influenced by the basal friction parameter, βi; and the undercutting, ui, at the subglacial discharge locations for four distinct times ti=t0,t4,t6,t11 (see Table ). The different configurations are referred as C(gi,βj,ui). If ui=0, there is no undercutting and hence a vertical ice front at the subglacial discharge location. At t=0, the melt season has not started yet, so there is no modelled undercutting. At t=11, the melt season is finished and there is no modelled undercutting. If j≠i, the geometry, gi, is taken at ti and the basal friction, βj, at tj to assess the roles of geometry and basal sliding velocity. We investigate basal friction at t0 and t6 since the former has maximum friction and the latter minimum friction of the studied cases. The configurations studied in this paper are summarised in Table .

The basal friction coefficient, β, for the four runs presented above, is shown in Fig. . At t0, before the melt season, the basal friction is high and roughly homogeneous over the first kilometre. At t4, when the surface runoff is the highest, the pattern is similar but with a large offset. The lowest friction is reached at t6, particularly at the front and in the southern part of the glacier. The highest friction is reached at t11 a kilometre from the front. Close to the front position, however, the friction is still high.

The hydrological model predicts that there are two main subglacial channels with discharge exceeding 1 m3s-1 of water (see Fig. a). This is in accordance with satellite and time-lapse camera images showing upwelling at these locations . Modelled surface melt and discharge at the northern outlet – in short northern discharge (ND) – starts 6 June and ends 1 October, while the discharge at the southern outlet (SD) starts 21 June and ends 22 August. Fluxes at ND clearly exceed those at SD as shown in Fig. b and Table .

The melt rate profiles calculated by the plume model for four different volumes of subglacial discharge are shown in Fig. .

At a discharge of 1 m3s-1, melt rates are low (< 2.5 md-1), with the maximum melt rate occurring at depth and negligible melt rates close to the water line. At 10 m3s-1, the melt profile reaches the surface and has highest melt rates (∼ 3.5 md-1) along the plume column. With 50 and 100 m3s-1 discharge, the highest melt rates are attained at the ocean surface on the sides of the plume column (∼ 5 and ∼ 6 md-1 respectively). In general, low discharges drive maximum melt within the plume and at depth, while higher discharges drive stronger surface gravity currents and therefore give higher melt rates at the surface.

The modelled frontal position is summarised in Fig.  in plan view and vertical view at the discharge locations. In most cases for the ND location, where the discharge is the highest, the melt profile (Fig. d) creates an undercut profile concentrated right near the waterline. found similar results when modelling melt rates at shallow grounding lines (100–250 m) given 250 m3s-1 discharge. It is interesting to see that the observed front after calving, Fiobs (dashed line in Fig. a–b), generally falls behind the undercut front before calving, Fielmer(z) (thick line in Fig. b).

The frontal submerged undercutting driven by the plume differs in shape from one location to another. In the first 50 m below the surface, the undercutting at the SD is not as abrupt as at the ND and is also smaller (Fig. c–e). Where the discharge is the highest, the melt rate peaks just below the waterline and stretches laterally from the vertical centreline of the plume. The lateral extent of melting is much lower at depth. At the SD, melting is strongest at depth due to lower discharge rates and less vigorous buoyant ascent of the plume. One should keep in mind that our modelling approach neglects the change of the front during the period of interest between two observations of frontal positions (11 days for most cases). In reality, calving would occur more often during that period, causing such large undercuttings, like the modelled ones, to not be possible. This simplification has consequences for the next step when the particle model handles the calving of icebergs due to front imbalance.

The observed mean volumetric calving rate averaged over the entire calving front volume of ice, a˙cobs, is the difference between the frontal velocity, vwobs(ti), and the rate of position change, ∂Lobs/∂t, integrated over the terminus domain. These quantities and the total modelled ice mass melted by the plume normalised per day (when an undercutting is prescribed) are given in Table .

To assess the performance of the offline coupling, we evaluate the mean volumetric calving rate averaged over the entire calving front volume of ice (see Eq. ), and the mean absolute distance between the modelled and the observed front, |L|‾. These are presented in Fig.  for each configuration as well as the observed mean volumetric calving rate. Figure  shows the different front positions after the HiDEM simulation for each configuration of the studied time. Figure  shows strain rates modelled by HiDEM that resemble an observed crevasse patterns (yellow lines representing crevasses).

At t0, before the melt started, the front has retreated at a rate of 7.93 ×105 m3 d-1 with a frontal ice flux of 2.63 ×105 m3 d-1, mostly in the middle part with a calved area of 5.1×104 m2. The HiDEM produces a slightly higher mean volumetric calving rate, 9.76 ×105 m3 d-1, with a vertical ice front configuration (red line C(g0,β0,0) in Fig. a) at a mean distance of 32 m from the observed front. However, calving is concentrated south of SD in a zone of high ice velocity and high strain rates as modelled by HiDEM (see Fig. ).

With peak surface runoff, at t4, the observed mean volumetric calving rate equals 7.97 ×105 m3 d-1, similar to t0 but with higher ice velocities (3.68 ×105 m3 d-1). Observed retreat at and north of ND is significant but is not reproduced by the configuration with a vertical ice front (red line C(g4,β4,0) in Fig. b). Instead the front is retreating south of SD in the same fashion as for t0. The mean volumetric calving rate (6.82 ×105 m3 d-1) is therefore close to the observed value, but the mean distance between the observed and the modelled front is close to 60 m (see Fig. ). For the undercutting configuration (blue line C(g4,β4,u4) in Fig. b), the mean volumetric calving rate is also overestimated at the same location but the observed retreat around ND is matched by the HiDEM. The mass removed by undercutting represents 11.8 % of the total observed mean volumetric calving rate (see Table ) and is therefore non-negligible. At the SD, the observed front is advancing (see Fig. b) and regardless of the applied modelled front configuration (with or without undercutting), a similar slight retreat is modelled. In this case, the undercutting has no influence on the calving.

Vertical front configuration at t6 (red line C(g6,β6,0) in Fig. c), during a period of accelerated glacier flow, results in slower modelled mean volumetric calving rate (16.26 ×105 m3 d-1) than observed (26.94 ×105 m3 d-1) and no front position change at both SD and ND, leading to a mean distance to the observed front close to 60 m. With the undercut configuration (blue line C(g6,β6,u6) in Fig. b), modelled mean volumetric calving rate (23.60 ×105 m3 d-1) is similar to observation and the front positions at discharge locations are reproduced even though the undercutting only represents 5.2 % of the observed mean volumetric calving rate. The modelled front is still intensively breaking up south of SD, but, at that date, it matches the observed retreat.

At the end of the melt season at t11, when subglacial discharge has ended, the observed front retreats at a rate of 24.99 ×105 m3 d-1 despite a frontal basal friction higher than at the last studied iteration resulting in an averaged frontal velocity of 2.56 ×105 m3 d-1. But, as shown in Fig. , the sliding velocity is higher (lower basal friction, β11) close to the front than further upglacier. Large calving events occur at both former discharge locations where the bed elevation is lower than anywhere else. The calving front modelled by HiDEM (red line C(g11,β11,0) in Fig. d) manages to reproduce this behaviour but overestimates the retreat for the region in between, where the pattern of high strain rate is also denser (see Fig. ).

Two configurations vary the friction coefficient, β, to assess the role of sliding in the calving process. If the basal friction is set according to t6 and the geometry to t0 (orange line C(g0,β6,0) in Fig. a), the mean volumetric calving rate exceeds observations by more than a factor of 2 (16.40 ×105 m3 d-1), similar to C(g6,β6,0), yet with matching spatial frontal patterns as C(g0,β0,0) as well as strain rate distribution with elevated rates close to the calved zones. If the geometry of t6 is simulated with the basal friction of t0 (orange line C(g6,β0,0) in Fig. c), it is striking to notice again that the calved zones are similar to the vertical front configuration at t6 but the mean volumetric calving rate is similar to the observed one at t0. High strain rates are less pronounced than with the basal friction of t6 but concentrated at the same locations.

Our plume model uses a fixed, planar ice front to calculate submarine melt rates rather than a time-evolving geometry. This assumption is supported by , who showed that the shape of the submerged ice front does not have a significant feedback effect on plume dynamics or submarine melt rates. However, the same study suggests that the total ablation driven by submarine melting will increase due to the greater surface area available for melting. To take this effect into account in our undercutting model, submarine melt rates are horizontally projected onto the undercut front modelled at the previous iteration.

By using ambient temperature and salinity profiles that do not vary in time, we neglect the inter- and intra-annual variability in Kongsfjorden. This variability can affect the calculated melt rate in two ways: (i) the three-equation melt parameterisation explicitly includes the temperature and salinity at the ice face, and (ii) the ambient stratification affects the vertical velocity and neutral buoyancy height of the plume. The direct effect of changes in temperature and salinity on the melt equations are well tested. Past studies using uniform ambient temperature and salinity conditions have found a linear relationship between increases in ambient fjord temperatures and melt rates, with the slope of the relationship dependent upon the discharge volume . Salinity, on the other hand, has been shown to have a negligible effect on melt rates . However, with a non-uniform ambient temperature and salinity, the effects of changes in the stratification on the plume vertical velocity and neutral buoyancy are much more complex. The stratification in Kongsfjorden is a multi-layer system, with little or no direct relationship between changes in different layers . Therefore, testing cases by uniformly increasing or decreasing the salinity would not be informative for understanding the true effects of inter- and intra-annual variability. The high computational expense of the plume model used here means that it is not yet feasible to run the model on the timescales necessary to understand this variability, nor to run sufficient representative profiles to provide a useful understanding of the response. Previous work has suggested that intra-annual changes in the ambient stratification are small enough that plumes are relatively insensitive to these changes and that plume models forced with variations in runoff and a constant ambient stratification can qualitatively reproduce observations . For these reasons, we highlight this as a limitation of the current implementation and suggest that this should be addressed in future investigations of plume behaviour. A model based upon one-dimensional plume theory e.g. would be less computationally expensive and may allow some of these limitations to be addressed. However, such a model would not capture the strong surface currents driven by the plume which are important for the terminus morphology studied here.

For ND (Fig. b and d), the undercutting is in line with the observed front to a certain extent, particularly for t4. However, for SD, apart from t3, no apparent correlation between modelled undercutting and observed front location seems to exist. However, Fig.  shows that modelling calving with undercutting at SD and ND for t4 and t6 gives a good fit to observation. The difference in agreement with the observed front position and the modelled calving could possibly be explained by the uncertainty in discharge or the different character of the plume at high and low discharge. The low dependence of calving front position on modelled undercutting in situations of low discharge seems to have no major influence on the performance of the calving model. At Kronebreen, the high discharge relative to the shallow depth of the terminus drives strong gravity currents at the surface as water is rapidly exported horizontally away from the plume. The melt rates driven by these gravity currents are significant, as shown in Fig. , and in some cases dominate over the melt rates driven by the plume at depth. The difference between low and high discharges is therefore slightly counterintuitive. At low discharges, when maximum melt rates occur at depth, the terminus is more undercut but in a narrower area; meanwhile, at higher discharges, strong undercutting occurs but over a much wider area of the terminus. This suggests that calving behaviour may be very different in these two situations.

Because the imposed undercuttings are the product of melt during the whole interval between observations, the model results should be treated with caution. compared HiDEM calving for specified undercuttings of different sizes and showed that calving magnitude increases with undercutting size. For small undercuttings, calving simply removes part of the overhang, but for large undercuttings calving removes all of the overhang plus additional ice. The mechanisms are different in each case: low-magnitude calving for small undercuttings occurs through collapse of part of the unsupported overhang, whereas high-magnitude calving for large undercuttings involves forward rotation of the whole front around a pivot point located at the base of the undercut cliff. The long time-step intervals (11 or 18 days) between the starting geometry and the HiDEM simulation in the present study might therefore bias the results towards higher calving events. Testing this possibility is beyond the scope of the present paper, but it remains an important goal for future research. Despite this caveat, our results compare well with observations and yield valuable insights into the calving process.

Firstly, the HiDEM results show that undercutting associated with meltwater plumes is an essential factor for calving during the melt season (t4 and t6). Surface melt leads to the formation of a subglacial drainage system that ultimately releases the water into the ocean from discharge points at the front of the glacier. Simulations without frontal undercutting at these subglacial discharge locations do not agree well with observed frontal positions and mean volumetric calving rates. In contrast, simulations with frontal undercutting reproduce the retreat reasonably well at these locations, particularly where the discharge is high such as at ND. The largest discrepancy between modelled and observed calving is in the region south of SD at t4. Here, the model predicts calving of a large block, whereas the observed front underwent little change. This largely reflects the rules used for calving in HiDEM: any block that is completely detached from the main ice body is considered as calved, even if only separated by a narrow crack from the rest of the glacier and still sitting at its original position. This is the case for the large “calved” region south of SD at t4, where the block may have been completely detached but remained grounded and in situ. If this were to occur in nature, it would not register as a calving event on satellite images. The discrepancy between model results and observations at this locality therefore may be more apparent than real.

Secondly, the model results replicate the observed high calving rates at t11, after the end of the melt season, when there is no undercutting. At this time, the observed mean volumetric calving rate is 24.99×105 m3d-1, which compares well with the HiDEM rate of 28.50×105 m3d-1. These values are much higher than those at the start of the melt season, when there is also zero undercutting. This contrast can be attributed to the high strain rates in the vicinity of the ice front at t11, which would encourage opening of tensile fractures (Fig. ). In turn, the high strain rates result from low basal friction (Fig. d), likely reflecting stored water at the glacier bed after the end of the melt season. It is possible that geometric factors also play a role in the high calving rates at t11, because the mean ice front height is greater at that time than at t0, reflecting sustained calving retreat during the summer months, which would have increased longitudinal stress gradients at the front . This interpretation is supported by experiments C(g0,β6,0) and C(g6,β0,0), in which the basal friction values are transposed for non-undercut ice geometries at t0 and t6. Imposing low friction (β6) at t0 produces mean volumetric calving rates similar to (but smaller than) those observed at t6, whereas imposing high basal friction (β0) at t6 produces low volumetric calving rates similar to those observed at t0. The influence of basal friction on calving rates is consistent with the results of , who found that a strong correlation exists between frontal ablation rates and ice velocity at Kronebreen when velocity is high. Low basal friction is associated with both high near-terminus strain rates and high velocities, facilitating fracturing and high rates of ice delivery to the front. Our experiments do not include varying fjord water temperature, so we cannot corroborate the strong correlation between frontal ablation and fjord temperature observed by . However, our results are consistent with their finding that melt undercutting is a primary control on calving rates, with an additional role played by ice dynamics at times of high velocity.