Fast-flowing (>5 km yr−1) Greenland tidewater glaciers1,2,3 contribute to sea-level rise through increased ice discharge4, sensitively coupled to terminus position5 and flow resistance at the glacier bed6,7. Glacier basal resistance depends on water pressure and distribution within the subglacial drainage system8,9 and is controlled by where, when, and how much surface melt reaches the bed10,11. The flow of the inland Greenland Ice Sheet is sensitive to the variability of surface-melt inputs on hourly to seasonal timescales5,12,13. Sea-level-rise contributions from the ice sheet, however, are dominated by the calving of fast-flowing tidewater glaciers at the marine margin4 —a region where the effects of surface-melt forcing on ice flow are poorly understood14,15.

Our limited understanding of coupled tidewater-glacier hydrology and ice flow—and the potential future response of tidewater glaciers to expected changes in surface-melt production16— is due, in part, to the difficulty of obtaining contemporaneous observations of subglacial water flow and high-temporal-resolution ice-flow velocities for this important class of glaciers. Recent predictions based on inferences from low-time-resolution remote-sensing observations of tidewater-glacier seasonal velocity patterns suggest that changes in flow speeds caused by changes in melt could be very important for ice discharge; or, perhaps, negligible6,17,18,19. These divergent predictions hinge on a widely used binary framework for classifying the subglacial drainage system as efficient, if there is a late-season velocity minimum, or inefficient, if there is not3,7,12,13,17,18,19,20. This framework is observationally justified for inland ice-sheet regions12,13, but whether it can be reliably applied to determine the subglacial hydrology of fast-flowing tidewater glaciers is unknown15. Determining the nature of the drainage system and its response to meltwater input is critical for sea-level rise predictions, as increased meltwater could have a tendency to both increase or decrease ice discharge8,9,10.

Helheim Glacier (Fig. 1a) is one of the fastest-flowing tidewater glaciers in Greenland2,21, and has the largest ice discharge at present4. Helheim Glacier velocity responds to calving events21, tides22, surface melt23,24,25, and ice-mélange coherence2 on sub-seasonal timescales. Unlike many alpine glaciers11 and inland regions of the Greenland Ice Sheet12,13, however, Helheim does not exhibit a reduction in velocity over the latter half of the melt season2,17,18,21, leading some authors to infer that high subglacial water pressures are sustained throughout the melt season17,18. Theoretical work8,9,10 that matches observations at alpine glaciers11 would classify such a high-pressure drainage system as inefficient, with surface meltwater input leading to increased basal sliding.

Fig. 1: Helheim Glacier, East Greenland.
figure 1

a Landsat image from 2001/182 showing supraglacial lake L1, (triangles) GPS array, and (circle) Automatic Weather Station (AWS) deployed in 2007. (thick black line) Glacier terminus position on 2007/236. July 2007 surface velocities64 shown in black contours at 1000 m yr−1 intervals. Inset shows (star) location of Helheim Glacier in Greenland. b Landsat images of L1 (red box in a) from 2007/206–231. Velocity response to L1 drainage begins at 2007/229.5 (Fig. 2). Source data are provided as a source data file. Landsat images courtesy of the U.S. Geological Survey.

Supraglacial lake drainages26 offer rare opportunities to observe the ice-flow response to isolated injections of surface melt. Such events have provided key constraints on ice-sheet behavior at lower flow speeds (0.1–0.6 km yr−1)26,27,28,29. The flow response to lake drainage along the length of a high-discharge, fast-flowing tidewater glacier has never been measured. Here, we present observations from a lake drainage at Helheim Glacier showing that the hydrological system is dominated by an efficient drainage network that is capable of rapidly exporting the additional meltwater input from the lake, with an increase, and subsequent decrease, in drainage-system water pressure. The water-pressure changes are small and short-lived, such that the lake drainage has minimal impact on the glacier’s longer-term average flow speed.

Results and discussion

Like other major Greenland tidewater glaciers2,3,6,7, Helheim Glacier hosts multiple supraglacial lakes and water-filled crevasses30 (Fig. 1a). The lakes fill and drain during the melt season30,31, exposing optically bright surfaces of smooth, previously submerged ice in the days immediately following drainage26,30 (Fig. 1b, Supplementary Figs. S1 and S2). In 2007, a network of geodetic-quality, dual-frequency GPS receivers was deployed on the main tributary of Helheim Glacier21,22,23,24,25 (Fig. 1a; “Methods”). Satellite images on days 222 and 231 of 2007 (2007/222 and 2007/231) (Fig. 1b) and GPS estimates of glacier-surface velocity (Fig. 2a and Supplementary Fig. S3) capture the drainage of supraglacial lake L1 located 18 km from the terminus on 2007/229. Across all eight GPS stations recording high-quality data (Supplementary Information), the glacier velocity response to lake drainage beginning on 2007/229 is characterized by a 1-day increase in along-flow speed \(v\) of ~4%, followed by a reduction in speed to 2% below pre-drainage velocities that is sustained for as much as 2 days (Fig. 2b). We observe ~0.06 m of vertical uplift occurring over 12 h at lake-proximal stations IS27–29 (Supplementary Fig. S4).

Fig. 2: Glacier velocities during supraglacial lake drainage.
figure 2

a Along-flow velocities \(v\) plotted as \((v-{v}_{228})/{v}_{228}\), where \({v}_{228}\) are individual average station velocities on 2007/228. Gray shading shows ±1σ errors on the velocity ratios. Black bar gives the velocity-ratio scale. Along-flow velocities are plotted by station distance from the glacier terminus; the y-intercept of each horizontal gray line is the station distance from the terminus. Time of (gray) \({t}_{0}\), (green) \({t}_{{peak}}\), and (purple) \({t}_{{node}}\) shown for each station with ±1σ error bars in time. b Averaged velocity pulse ±1 standard deviation of the velocity distribution for stations located near the lake (blue; IS27–29) and terminus (red; IS35–39). Velocity records are temporally aligned at \({t}_{0}\). c Additional displacement, where a value of 0 m indicates no change in displacement relative to where the station would have been if the station maintained a velocity of \({v}_{228}\) over the timeseries. Source data are provided as a source data file.

The along-flow velocity response to lake drainage begins at lake-proximal station IS27 and reaches near-terminus stations IS35–39 ~3.5 h later (Fig. 2a). We characterize the velocity response with three time points: (1) \({t}_{0}\), the time of an initial increase in velocity; (2) \({t}_{{peak}}\), the time of maximum velocity response; and, (3) \({t}_{{node}}\), the end of a positive velocity response. Each time point occurs first at the stations near the lake, and later at stations farther down-glacier. Stations within <2 km of each other exhibit temporal scatter; time points are within ±0.83 h of each other at lake-proximal stations IS27–29 and ±1.08 h of each other at terminus-proximal stations IS35–39. We interpret \({t}_{0}\) to indicate the arrival of high basal water pressures \({P}_{w}\) of the subglacial flood11, and \({t}_{{peak}}\) to indicate the time at which the subglacial system becomes maximally over-pressurized29,32. At this time, flood waters have likely overwhelmed existing subglacial channels and escaped laterally into neighboring regions of the bed8,9,13,32,33. As a network-wide measure of the temporal impact of the lake-drainage event on glacier flow, we use the duration over which the positive velocity response arrives and subsides across the GPS network. Calculated as the time between \({t}_{0}\) at IS27 and \({t}_{{node}}\) averaged across stations IS35–39, we find a duration of 28.16 h.

Interpreting the glacier velocity pulse and uplift as a response to lake drainage is consistent with additional independent sets of observations from on- and off-ice Automatic Weather Stations34 and a pressure transducer record of drainage of the same lake in 2009 that was seen to take 9 h (Supplementary Figs. S2 and S5). Other hypotheses, such as a stationary reduction in basal traction or a terminus-specific forcing, are not compatible with the observations. A stationary perturbation in basal traction cannot explain the velocity pulse because the pulse has a temporal lag as it propagates down-glacier over a distance (>16 km) that is 4–8 times larger than the longitudinal stress-coupling length scale35 of the main tributary of Helheim Glacier36. The pulse initiation location and down-glacier propagation direction disqualify a terminus-specific forcing, such as a calving event21, tidal modulation of flow22, or a change in mélange coherence37.

The down-glacier propagation speed of the peak in the velocity pulse we observe at Helheim Glacier represents the fast end-member along a continuum of previously observed pulse-propagation speeds from draining supraglacial lakes38, ice-dammed lakes11, stored basal water39, and precipitation events40, and is within the range observed for jökulhlaups41 (Table 1). In the upper terminus region between IS27 and IS26, the peak propagated at 0.96 ± 0.15 m s−1, slowing to 0.52 ± 0.09 m s−1 in the lower-terminus region between IS26 and the four terminus-proximal stations. Here, near the marine terminus, effective pressures \(N\) (\({N=P}_{i}-{P}_{w},\, {{{{{\rm{where}}}}}}\,{P}_{i}\) is the ice overburden pressure) at the glacier bed have been inferred to be low24,25. In general, low effective pressure near the grounding line of marine-terminating outlets promotes subglacial conduit widening, resulting in lower water flow speeds closer to the grounding line42.

Table 1 Previous observations of down-glacier flood propagation speeds following melt events, rain events, jökulhlaups, and lake drainages

Our observations show that flood events can temporarily modify the speed of even very fast outlet glaciers, but that the net effect on ice advection is small compared to background flow speeds—partly because ice velocity is suppressed below background speeds for ~2 days following the drainage-related peak (Fig. 2a, b). Integrating along-flow velocities from 2 days prior to 5 days following \({t}_{0}\) yields an additional, flood-related, ice displacement of −0.03 to +0.15 m for stations IS26–29 and +0.17 to +0.65 m for stations IS35–39 (Fig. 2c). This additional ice displacement during the week of the lake drainage is small relative to background flow speeds of 77–168 m week−1 (Supplementary Fig. S3). The timescale for the drainage system to return to a pre-flood state is longer than the duration of increased velocities, and roughly equivalent to the durations of suppressed velocities observed following mid-melt-season lake drainages on the western margin of the ice sheet28. The post-drainage slowdown we observe is consistent with previous observations at other glaciers11,39 of flow deceleration that follows meltwater supply exceeding a critical rate of water flow10. This glacier velocity response, combined with theoretical support from previous numerical modeling9,10,32,33, indicates that an efficient drainage system, with interacting channelized and cavity components32,33, exists beneath the glacier at the time of lake drainage.

To test our interpretation that the velocity pulse results from a pressure pulse in the subglacial drainage system, and that a well-developed drainage system is necessary for transporting this pulse down-glacier at the speeds observed in the GPS data, we simulate supraglacial lake drainage in a numerical model43,44 of subglacial hydrology at Helheim Glacier (Supplementary Fig. S6; “Methods”). The model consists of a continuous subglacial sheet connected to discrete channels at every model node43. Following a simulation forced with a fixed basal-melt rate and daily estimates of surface runoff from a regional climate model45 from 2007/1–229, 0.009 km3 of water is injected to the bed at the L1 location over 9.6 h beginning at 2007/229.5. This set-up and forcing simulates the L1 drainage at the time it occurred in the latter half of the 2007 melt season, when the drainage system has reached an evolved state after ~80 days of meltwater input (Fig. 3a). We refer to this simulation as M229 for the day of year corresponding to the modeled lake drainage. Drainage-system response time to runoff input is affected by the choice of values for drainage-system sheet permeability \({K}_{s}\) and ice englacial void fraction \(\sigma\). We test a range of values for these model parameters (Supplementary Fig. S7) and consider acceptable values to be those that produce results consistent with the observed GPS velocity-pulse duration and uplift.

Fig. 3: Modeled effective-pressure response to simulated lake drainage.
figure 3

a (left) Discharge \(q\) and (right) effective pressure \(N\) prior to simulated rapid lake drainage. The yellow circle shows the discharge outlet location along the glacier terminus. White diamond shows the location of simulated lake drainage. bg Difference between modeled values of \(q\) and \(N\) at six time points during the simulated lake drainage and the model values shown in a at 2007/229.00 and prior to the start of the simulated lake drainage. Triangles show GPS station locations. h (gray shading) Prescribed lake discharge \({Q}_{{lake}}\) and (curves) modeled effective pressure at the location of each GPS station, plotted as \({N/N}_{228}\), where \({N}_{228}\) are individual average values of \(N\) at each GPS station location on 2007/228. Black triangles mark time slices shown in (ag). i Modeled effective pressure from 2007/230–232 at the location of each GPS station, plotted as \({N/N}_{228}\), over a finer range in \({N/N}_{228}\) than shown in (h). GPS station colors as in Figs. 1 and 2. Model simulation uses parameter values \({K}_{s}\) = 1 Pa−1 s−1 and \(\sigma\) = 10−6. Source data are provided as a source data file.

Model results show a pulse in reduced effective pressure that transits down-glacier following the onset of the simulated lake drainage (Fig. 3). We present modeled effective pressure \(N\) at the location of each GPS station (Fig. 3h), as glacier sliding depends strongly on effective pressure, with low effective pressures thought to correspond to higher sliding speeds9,10,11,32,33. Interpreting effective pressure as a proxy for changes in glacier sliding is justified for the purposes of this study because, although the relationship between \(N\) and sliding speed can be nonlinear, modeled values of \(N\) at the location of each GPS station remain positive during the simulated lake drainage (Fig. 3h), indicating that we have not reached the limiting case where \(N\) approaches zero and the relationship between \(N\) and sliding speed breaks down46,47.

Like the positive velocity pulse observed at individual GPS stations, the duration of reduced effective pressure increases from the lake-proximal to terminus-proximal stations (Fig. 3h). In addition, modeled effective pressure shows a post-drainage increase similar in duration to the post-drainage slowdown observed in the GPS data. Modeled effective pressure shows a lower-amplitude response in the near-terminus region than in the region near the lake (Fig. 3), likely due to the lower fractional contribution to water flow from the lake drainage in the near-terminus region—lake, surface runoff, and basal-melt inputs integrate along the bed moving towards the terminus (Fig. 3a)—and because the near-hydrostatic pressure condition at the ocean boundary holds effective pressures closer to flotation in the near-terminus region24,48 (Fig. 3a). Multiple M229 model simulations using a range of reasonable parameter values are able to match GPS pulse durations to within a few hours and to within observed GPS uplift magnitudes (Supplementary Fig. S7). In Fig. 3, we present one of the three best-fitting M229 simulations, which reproduces the observed network-wide pulse duration to within 0.7 h and shows 0.02 m of uplift. Thus, our GPS observations and modeling results indicate that an efficient drainage system exists beneath one of Greenland’s fastest tidewater glaciers, and that such a system modulates flow while accommodating lake-drainage events.

To verify the importance of the well-developed drainage system predicted by our model in the late melt season for accommodating lake drainage, we create simulations of hypothetical drainages at times when the system is comparatively inefficient32. We simulate an equivalent L1 drainage prior to the start of the melt season, when the drainage system receives inputs from basal melt alone and is less developed at the L1 location (Supplementary Fig. S8a). We refer to this simulation as Mwinter and the simulated L1 drainage begins at 2007/69.5. Though both M229 and Mwinter simulations exhibit broad regions of decreased effective pressure during the drainage events (Fig. 3b–d and Supplementary Fig. S8b–d), the Mwinter simulation does not exhibit a post-drainage increase in effective pressure and results in a longer pulse duration, with a misfit to our observations of +66.9 h. In agreement with past idealized modeling33, our results imply that rapid lake drainage into a less-developed drainage system at Helheim, or similar glaciers, would primarily result in reduced effective pressure along the flowline, persisting for a longer time, and would produce net positive glacier advection.

The descriptors efficient and inefficient are often equated with channel-dominated and cavity-dominated drainage systems, respectively9,10,32,49,50. Our model allows water to flow through both discrete channels and the continuous cavity sheet layer at all points of space and time; the partitioning of the flow between them evolves dynamically over time43. We find that flow occurs through a mixture of channels and the sheet layer both during and outside of times of prescribed lake discharge, with spatial variation evident along the glacier (Fig. 4). Near the lake, we observe a decrease in effective pressure and an increase in the flow carried in the channels immediately after the lake discharge begins in both Mwinter and M229 simulations (Fig. 4c, e). Importantly, for the M229 simulation, the initial decrease in effective pressure is lower, and is followed by a post-drainage increase in effective pressure (Fig. 4c). By contrast, the Mwinter simulation shows larger decreases in effective pressure and weak or absent post-drainage increases in effective pressure in both the lake-proximal and near-terminus regions (Fig. 4c, d and Supplementary Fig. S8h). Near the terminus, the flow is dominantly carried by channels, even before lake drainage, in both Mwinter and M229 simulations; the effect of the lake drainage near the terminus is to temporarily reduce the effective pressure by a small amount, and a slightly larger proportion of the water flows in the sheet layer (Fig. 4e, f).

Fig. 4: Modeled effective pressure and proportion of discharge carried by channels during simulated lake drainages.
figure 4

a Discharge \(q\) on 2007/69.00 and (gray lines) location of across-flow lake-proximal and near-terminus transects in the model domain. White diamond shows location of simulated lake drainage. Triangles show GPS station locations, with station colors as in Figs. 13. b Equivalent plot for 2007/229.00. c (gray shading; right axis) Prescribed lake discharge \({Q}_{{lake}}\) and (left axis) modeled effective pressure \(N\) across the lake-proximal transect during (red) Mwinter and (blue) M229 simulated lake drainages. d Equivalent plot for the near-terminus transect with different axis limits for modeled effective pressure. e (gray shading; right axis) Prescribed lake discharge \({Q}_{{lake}}\) and (left axis) the proportion \(f\) of modeled discharge carried by channels across the lake-proximal transect during (red) Mwinter and (blue) M229 simulated lake drainages. f Equivalent plot for the near-terminus transect. Model simulations use parameter values \({K}_{s}\) = 1 Pa−1 s−1 and \(\sigma\) = 10−6. Prescribed lake discharge in the Mwinter and M229 simulations begins on 2007/69.50 and 2007/229.50, respectively.

Together, these results indicate that, for Helheim Glacier, an efficient drainage system cannot simply be equated with a channel-dominated system. The system is able to rapidly adjust to the extra lake water input, even though a significant proportion of the pre-lake-drainage flow is carried by the cavity sheet (Fig. 4e, f). Our observations (Fig. 2) and model results (Figs. 3 and 4) suggest a better test for efficient drainage beneath glaciers like Helheim Glacier—lightly grounded2, fast-flowing tidewater glaciers with high meltwater throughput—would be whether increased meltwater input leads to net neutral or reduced ice advection.

At Helheim, when lakes drain during the melt season30, the effect of a single lake-drainage event on ice advection is small compared to background flow speeds. With a daily, lower-terminus ablation rate of ~0.04 m d−1 water equivalent (w.e.) over a four-month melt season34, the basal drainage system receives high meltwater volumes from abundant surface crevasses30. This setting facilitates a well-developed drainage system that allows rapid response to and recovery from a sudden lake-drainage event. The relative insensitivity of net glacier advection to lake drainage observed at Helheim, however, is unlikely to extend to tidewater glaciers that experience lower seasonal meltwater forcing. For example, remote-sensing observations show that sporadic austral summer-melt events (<0.005 m d−1 w.e. for 1-week duration) at multiple Antarctic Peninsula tidewater glaciers coincide with ice-flow accelerations of up to 100% above annual speeds20. These flow accelerations are not followed by sustained periods of reduced velocities and result in net positive glacier advection, likely because water pressure spikes into an inefficient drainage system20, consistent with our model results for simulated winter-season lake drainage at Helheim Glacier. Thus, while these Antarctic Peninsula20 and other, currently colder, High Arctic tidewater glaciers51 may initially undergo a net speed-up as the amount of surface meltwater reaching the bed increases, their flow response to individual melt events may eventually decrease as higher seasonal melt magnitudes are attained as the climate continues to warm52.

Observational and theoretical advances are needed along the surface-melt continuum to test this response-evolution hypothesis and improve projections of global tidewater-glacier response15,16 to expected changes in surface-melt production on timescales longer than individual melt or lake-drainage events. In particular, we have argued that the Helheim Glacier flow response to lake drainage provides strong evidence that the lake drains into an efficient drainage system. Helheim Glacier’s near-terminus seasonal velocity pattern has, by contrast, been invoked to infer that an inefficient drainage system persists throughout the melt season18. Our observations suggest that near-terminus seasonal velocity patterns at glaciers of this type may not be well correlated with drainage-system efficiency, or the corresponding capacity to accommodate meltwater input with minimal impact on glacier discharge. At fast-flowing tidewater glaciers, where a marine margin and high surface melt rates keep subglacial water pressure high24, observations of seasonal velocity patterns17,18,19 may be misinterpreted to indicate inefficient drainage, where, in fact, an efficient system exists and is capable of exporting large volumes of meltwater rapidly. The widely used binary framework for interpreting subglacial drainage system efficiency from seasonal velocity patterns3,7,17,18,19,20, though observationally justified for inland ice-sheet regions12,13, does not explain the full range of behaviors observed at lightly grounded2, fast-flowing tidewater glaciers with high meltwater throughput.


GPS data

A network of geodetic-quality, dual-frequency GPS receivers was deployed from late June to late August 2007. During the time of the L1 lake drainage, the network consisted of eleven receivers that spanned an along-flow distance of 2–24 km from the calving front. Additional fixed stations were located at bedrock sites. GPS data were processed in kinematic mode using the TRACK software package53 to yield position estimates every 15 s21,22,23. Here, we eliminate position estimates with unfixed biases, and rotate the timeseries to obtain position estimates in the local along-flow and vertical directions at each station. We use a stochastic filter to estimate a horizontal along-flow velocity \(v\), a principally semi-diurnal horizontal ocean-tide response, and a diurnal component of horizontal glacier flow23,24. Step changes in along-flow velocity are allowed at times of glacial earthquakes, which represent large calving events21,23,24 (Supplementary Fig. 3). Eight of the eleven receivers in operation recorded data of sufficient quality for high-resolution, stochastic-filter analysis (Supplementary Methods). We focus our analysis on the along-flow velocity \(v\), which is largely free of glacier flow responses to tidal22 and diurnal23,24 forcing (Supplementary Figs. 917).

We characterize the velocity response to lake drainage using three time points: (1) \({t}_{0}\), the time of an initial increase in velocity; (2) \({t}_{{peak}}\), the time of maximum velocity response; and, (3) \({t}_{{node}}\), the end of the positive velocity response (Supplementary Fig. 3). The time of \({t}_{0}\) is picked by eye to be immediately prior to the acceleration in \(v\) beginning after 2007/229.5. The time of \({t}_{{peak}}\) is the time of the maximum value of \(v\) after \({t}_{0}\). The time of \({t}_{{node}}\) is the time when \(v\) first crosses back below \({v}_{228}\), the average value of \(v\) on 2007/228 at each station, following \({t}_{{peak}}\). To facilitate interstation comparison of the ice-flow response to lake drainage, we calculate the fractional change in \(v\) over the lake-drainage event at each station. We present the ratio of \((v-{v}_{228})/{v}_{228}\) in Fig. 2a. Values of \({v}_{228}\) range from 11–24 m d−1, with larger \({v}_{228}\) observed at near-terminus stations (Supplementary Fig. 3). To produce the averaged velocity-pulse curves in Fig. 2b, the \((v-{v}_{228})/{v}_{228}\) ratios are aligned in time based on the pulse onset time \({t}_{0}\) at each station, and then averaged across stations within 2 km of each other.

We calculate the speed of the down-glacier propagation of the velocity pulse by differencing \({t}_{{peak}}\) between adjacent stations and then dividing by interstation distance measured along the flowline. Uncertainties in this estimate are calculated using a ±0.02 m uncertainty in GPS station position equivalent to 1σ uncertainty in horizontal position associated with the TRACK position solutions, and an uncertainty in \({t}_{{peak}}\) estimated at ±1 h for all stations except IS35. A data gap at station IS35 during the drainage event requires \(v\) to be interpolated across the time of \({t}_{{peak}}\), resulting in an uncertainty in \({t}_{{peak}}\) of ±2.16 h for this station (Supplementary Methods; Fig. 2a and Supplementary Fig. S17).

Additional observations supporting the velocity-pulse interpretation

In addition to the primary datasets of the 2007 GPS data and satellite imagery, four independent sets of observations are consistent with and further support the interpretation of the glacier velocity pulse as a response to the rapid drainage of lake L1. First, contemporaneous temperature and net short-wave radiative flux observations at an on-ice Automatic Weather Station (AWS) show no anomalously high values in the days leading up to the drainage, suggesting that the velocity pulse is not related to a period of above-average surface melt (Supplementary Fig. 5). Second, contemporaneous observations at the Tasiilaq off-ice AWS are consistent with those from the glacier, and show no anomalous temperature, insolation, relative humidity, or precipitation in the days prior to the event. The 6-hr precipitation observations in Tasiilaq record a rain event on 2007/231–233; however, this event occurs after elevated along-flow velocities have subsided, suggesting that the velocity pulse is not due to a rain event. Third, an L1 rapid drainage was recorded in 2009 by a water-pressure transducer deployed in the lake (Supplementary Fig. 2). Though no GPS velocities are available during the 2009 event, the pressure transducer records a 9-hr drainage duration, providing evidence that the lake can drain rapidly, with a duration similar to that of other rapid supraglacial lake drainages in Greenland26,27,54,55. Finally, we observe no change in other supraglacial water bodies or the structure of the proglacial ice mélange between sequential satellite images spanning the event (Supplementary Fig. 1). The velocity pulse is thus not likely due to the drainage of a supraglacial water body other than L1 or to a change in back pressure on the glacier terminus. We do observe changes in water levels of ice-dammed lakes along the eastern side of the northern tributary of Helheim Glacier at the time of the L1 drainage (Supplementary Fig. 1); however, the subglacial drainage pathways for these ice-dammed lake drainages would not transit beneath stations IS27–29 based on the subglacial topography (Supplementary Fig. 6a) and modeled discharge routes (Fig. 3a) for the northern tributary.

Automatic weather station data

We use AWS data recorded on34 and off56 the glacier to evaluate meteorological conditions during the time of lake drainage. The on-ice AWS was located near IS28 (66.46˚N, 38.44˚W) (Fig. 1a), and recorded a standard suite of meteorological parameters at an hourly sampling rate, including temperature, relative humidity, and incoming and reflected short-wave radiative fluxes (Supplementary Fig. 5). The net short-wave radiative flux (insolation) is the AWS variable most closely correlated with the total energy flux available for melting the glacier surface34. We compare the on-ice AWS observations to an off-ice AWS located ~102 km to the southeast in Tasiilaq56 (65.60˚N, 37.62˚W). The off-ice AWS recorded hourly temperature and relative humidity and a record of precipitation every 6 h.

Water-pressure-transducer data

An internally logging HOBO water-pressure transducer was deployed within the L1 lake basin (66.46˚N, 38.46˚W) from late June to late August 2009, and recorded measurements of water depth at an hourly sampling rate (Supplementary Fig. 2). From 2009/205–234, the raw water-level record shows five, 0.5–2.5 m step offsets that we attribute to instrument malfunction. We estimate the magnitude of each step offset by eye and shift water-depth measurements up after each offset such that measurements are continuous with the relatively cleaner record from 2009/192–204. We then use a moving-average filter 0.1-d in width to identify and remove observations more than 3σ from the mean to produce the relative water-depth record shown in Supplementary Fig. 2.

Subglacial-hydrology model

We simulate supraglacial lake drainage in a numerical model of subglacial hydrology at Helheim Glacier. We use a two-dimensional model of subglacial drainage43 most recently employed to investigate seasonal drainage-system dynamics of land- and marine-terminating regions of the western Greenland Ice Sheet margin44,57. Full model equations are given in ref. 44.

The model, which is similar to the Glacier Drainage System model50, routes surface meltwater input into a continuous sheet connected to discrete channels melted upward into the base of the ice sheet43. Water moves between a continuous sheet, channels, and englacial storage to maintain a continuous hydraulic potential. The continuous sheet has a thickness \(h\), which is the sum of a cavity sheet layer with thickness \({h}_{{cav}}\) and an elastic sheet layer with thickness \({h}_{{el}}\), which is included to represent elastic uplift of the glacier when \(N\) becomes negative (\({P}_{w}\; > \; {P}_{i}\)). The thickness of the cavity sheet evolves due to the combined effects of basal ice melt, cavity opening by basal sliding, and cavity closing by ice creep58,59. Water flux through the sheet \({{{{{{\bf{q}}}}}}}_{{{{{{\bf{s}}}}}}}\) is dependent on the coefficient \({K}_{s}\) controlling the sheet permeability, and sheet thickness \(h\). Though \({K}_{s}\) is a spatially uniform parameter, the effective hydraulic transmissivity of the sheet layer \({K}_{s}{h}^{3}\) varies in both space and time in response to water input.

Water flux in the sheet \({{{{{{\bf{q}}}}}}}_{{{{{{\bf{s}}}}}}}\) is connected to discrete channels with discharge \({q}_{Q}\) at every model node. The growth and decay of channel cross-sectional area is a competition between melt opening and creep closure of channel walls60. The incipient sheet width contributing to channel melting \({\lambda }_{c}\) sets the length scale over which ice melting contributes to channel formation. Mass conservation is a balance between the sheet, channel, and englacial storage components. Englacial storage is dependent on the connected void fraction of the ice \(\sigma\) and the cross-sectional area of moulins \({A}_{m}\) (Supplementary Table 1).

For comparison with the horizontal and vertical GPS data, we focus our analysis on model predictions of \(N\), total flux \(q\) (the combined flux from the channel and sheet layers), and sheet thickness \(h\), which are solved at every model grid point spaced at 150 m. The ice-flow model in ref. 43 is not used here, as the exact form of the sliding law relating \(N\) to basal sliding velocities is uncertain, can be highly nonlinear, and would introduce an additional set of unknown model parameters. Instead, we interpret modeled \(N\) as a proxy for changes in glacier velocity during simulated supraglacial lake drainages. The use of the ref. 43 hydrology model alone prevents a direct coupling between basal sliding and rates of cavity opening, which can result in a negative feedback when additional cavity space is opened during times of faster sliding61. However, due to the small magnitude (<5%; Fig. 2b), short-duration (1-day) velocity changes we observe in response to the L1 drainage, we expect increases in basal sliding to have a minor impact on rates of cavity opening.

Model domain and boundary conditions

The model domain is an ~800 km2 region that extends over the three main tributary branches of Helheim Glacier (Supplementary Fig. 6), and consists of a regularly spaced, rectangular mesh of model nodes with 150-m spacing. We use the 150-m resolution IceBridge BedMachine Greenland v3 bedmap62 to represent bedrock topography, and the 150-m resolution Greenland Ice Mapping Project digital elevation model63 to represent glacier-surface elevation. The basal sliding speed \({U}_{b}\) at each node is set to the value of the corresponding surface speed from the MEaSUREs Greenland Ice Sheet Velocity compilation64,65 for July 2007. Bedrock topography, glacier-surface elevation, and basal sliding speed are assumed to be constant.

The model is forced by a constant basal-melt rate and daily surface runoff. The basal-melt rate \(m\) is prescribed everywhere to be 0.0262 m yr−1 based on an average geothermal heat flux of 0.063 W m−2 beneath Greenland66 and a contribution to basal melt from frictional heating due to sliding estimated using a basal shear stress \({\tau }_{b}\) of 60 kPa and a basal sliding velocity of 500 m yr−1. Surface runoff is derived from downscaled 1-km-resolution Regional Atmospheric Climate Model v2.3 (refs. 45, 67) runoff estimates, and interpolated to each of the model grid nodes44. Because surface meltwater enters the basal drainage system through abundant crevasses on Helheim Glacier30,34, we input surface runoff to the glacier bed at every model node. There is no surface storage term in the model, and surface runoff transits immediately to the bed. For boundary conditions, a small upstream basal-melt flux is applied at the eastern boundary on the subglacial drainage system, equivalent to the basal-melt rate (0.0262 m yr−1; Supplementary Table 1) integrated over a 50-km-distance inland from the eastern boundary. We prescribe \(N\) to be zero at both marine- and land-terminating margins of the glacier. This boundary condition represents the approximately hydrostatic pressure expected at the marine terminus, given observations of glacier surface height and bed elevation that indicate the glacier is near flotation on 2007/23624 and interpretations of a transiently floating terminus in more recent years31 when the terminus has retreated back to near its location on 2007/236 (Fig. 1a).

Simulated supraglacial lake drainage

Daily changes in runoff forcing prevent the model from reaching a steady state during the melt season43,44. In order to isolate the effect of a supraglacial lake-drainage event on the drainage system, we first force the model with daily runoff from January 1, 2007 to December 31, 2007. We then explore the effect of L1 drainages by beginning the lake-drainage events on different days of the year. For M229 scenarios, we initiate the lake drainage at approximately the same time as \({t}_{0}\) inferred from the GPS data: 2007/229.50. This date falls late in the melt season, when the model receives inputs from both surface runoff and basal melt. For Mwinter scenarios, we initiate the lake drainage on 2007/69.50, when the drainage system receives inputs from basal melt, but surface-melt input is negligible. At 2007/69.00, discharge is largely accommodated in the sheet layer in the lake-proximal region (IS27–29) (Supplementary Fig. S8a). Thus, initiating a lake drainage on 2007/69.50 simulates injecting a lake into an inefficient drainage system at the L1 location.

The simulated lake-drainage volume, injection location, and duration is equivalent for M229 and Mwinter scenarios. Because the exact drainage volume and duration are unknown for the 2007/229 L1 drainage, these values are approximated based on observations of L1 drainage duration in 2009 (Supplementary Fig. 2) and L1 maximum volume in 2007 (ref. 30). The volume estimate given in ref. 30 was obtained from a digital elevation model of the post-drainage L1 basin on 2007/205. Our lake-drainage simulation injects 0.009 km3 of water in total over 9.6 h and across four neighboring model nodes located within the extent of L1 on the glacier surface. Prescribed lake discharge \({Q}_{{lake}}\) increases in a step-wise fashion over the 9.6 h to a maximum of 434 m3 s−1 (Fig. 3h), which is within the range of discharge rates for rapidly draining supraglacial lakes26,54. The full model run is 7 days in length, extending in time from 2007/228–235 for M229 scenarios and 2007/68–75 for Mwinter scenarios. Basal-melt and surface-runoff inputs across the entire model domain are held fixed during the simulated lake-drainage event at 2007/227 values of 0.0004 km3 d−1 and 0.0147 km3 d−1, respectively, summed across the model domain for M229 scenarios and at 2007/67 values of 0.0004 km3 d−1 and 0 km3 d−1, respectively, summed across the model domain for Mwinter scenarios. This is done to isolate the effective-pressure response to the lake-drainage event from daily changes in effective pressure at day boundaries due to daily changes in surface-runoff forcing.

Model parameter-space calibration

Model parameter values are chosen from previous work applying this model to the western margin of the Greenland Ice Sheet44,57 (Supplementary Table 1). The bed roughness height scale \({h}_{r}\) and length scale \({l}_{r}\) are equivalent to those used in refs. 44, 57 (Supplementary Table 1). The primary justifications for using these values are that they keep us in a previously examined region of this model’s parameter space44,57 and that they are similar to length scales for bed roughness used by previous studies10,43,68. We vary sheet permeability \({K}_{s}\) by factors of 10 across the range 10−4–101 Pa−1 s−1, and englacial void fraction \(\sigma\) by factors of 10 across the range 10−8–10−2. This results in 42 different parameter combinations (Supplementary Fig. 7). The model converged for all 42 parameter combinations. At high englacial void fractions, the volume of englacial storage space (0.01 km3) is greater than the volume of water in the simulated lake drainage (0.009 km3), which results in lake-drainage inputs being mostly accommodated within the englacial void space. In these cases, a response in \(N\) is observed at the lake-proximal stations (IS27–29), but no response is observed at the terminus-proximal stations (IS35–39) by the end of the model run, which extends 5.5 days after the onset of lake drainage.

We assess parameter-space fitness using two criteria. First, we compare the observed pulse duration for the GPS network to modeled values of \(N\) at the location of the GPS stations in the model domain (Supplementary Fig. 7a, b). The pulse duration in the GPS data is calculated as the time between \({t}_{0}\) at station IS27 and \({t}_{{node}}\) averaged across stations IS35–39 to be 28.16 h. The pulse duration in the model is calculated as the time between \({t}_{0}\) of \(N\) at station IS27 and \({t}_{{node}}\) of \(N\) averaged across stations IS35–39, where \({t}_{0}\) of \(N\) is the time when \({N/N}_{228}\) < 0.98 and \({t}_{{node}}\) of N is the time when \({N/N}_{228}\) > 1.00 following \({t}_{{peak}}\) (Fig. 3h). We calculate the misfit between the observed and modeled pulse duration by differencing the two values, and present the misfit in hours (Supplementary Fig. 7a, b).

Second, we compare the amount of observed vertical uplift at stations IS27–29 following \({t}_{0}\) to modeled changes in drainage-system height (Supplementary Fig. 7c, d). In the vertical GPS data, we observe ice-sheet surface uplift of 0.06 m over 12 h beginning at 2007/229.5 at stations IS27–29 (Supplementary Fig. 4). We do not observe similar uplift at other GPS stations. The 0.06 m of uplift observed at stations IS27–29 is equivalent to 2 standard deviations in uncertainty in vertical height associated with the TRACK position solutions, and we would have observed larger vertical uplifts had they occurred. We, therefore, exclude models where a change in the height of the sheet layer \(h\) during the course of the simulated lake drainage exceeds 0.09 m, equivalent to 3 standard deviations in vertical position uncertainty, at the location of stations IS27–29. This uplift criterion is satisfied for 24 (57%) of the scenarios (Supplementary Fig. 7c, d). When \({K}_{s}\) ≤ 10−2 Pa−1 s−1 and \(\sigma\) ≤ 10−3, changes in \(h\) during the simulated drainage range from 0.15 to 0.86 m, and thus do not satisfy the uplift criterion.

Overall, both pulse duration and uplift are more strongly controlled by sheet permeability than by englacial void fraction. The pulse-duration criterion provides a finer model-space calibration than the uplift criterion. One of the three best-fitting M229 simulations has a pulse-duration misfit of –0.7 h (42 min), with \({K}_{s}\) = 1 Pa−1 s−1 and \(\sigma\) = 10−6 (Supplementary Fig. 7a). Model outputs from this M229 scenario and an Mwinter scenario with equivalent parameter values are shown in Fig. 3 and Supplementary Fig. 8, respectively.