Subglacial sedimentary basins focus key vulnerabilities of the Antarctic ice-sheet

Antarctica preserves Earth’s largest ice sheet which, in response to climate warming, may lose ice mass and raise sea level by several metres. The ice-sheet bed exerts critical controls on dynamic mass loss through feedbacks between water and heat uxes, topographic forcing and basal sliding. Here we show that through hydrogeological processes, sedimentary basins amplify critical feedbacks that are known to impact ice-sheet retreat dynamics. We create a high-resolution subglacial bedrock classication for Antarctica by applying a supervised machine learning method to geophysical data, revealing the distribution of sedimentary basins. Sedimentary basins are found in the upper reaches of Antarctica’s most rapidly changing ice streams, including Thwaites and Pine Island Glaciers. Hydro-mechanical numerical modelling reveals that where sedimentary basins exist, water discharge rate scales with the rate of ice unloading and the resulting hydrological instabilities are likely to amplify further retreat and unloading. These results indicate that the presence of a sedimentary bed in the catchment focuses instabilities that increase the vulnerability of the ice streams to rapid retreat and enhanced dynamic mass loss. or collapse scenario with rates > 650 m a -1 the groundwater ux exceeds 10 mm a -1 , potentially sustaining a signicant source of subglacial water even in high basal melt rate environments. The enhanced ux caused by retreat feeds back into dynamic processes affecting basal sliding including a weakened till layer, enhanced heat ux and enhanced basal melt rate and also potential destabilisation of ice shelves 17 . We propose that the instabilities caused by enhanced subglacial water supply from sedimentary aquifers, triggered by rapid ice-sheet retreat, are a crucial mechanism controlling the vulnerability of the Antarctic ice-sheet.

With 99% of the continent covered by thick ice, the understanding of subglacial geology in Antarctica relies on geophysical data. Fig. 1 shows the current understanding of Antarctic bedrock type, which is driven by the interpretation of data from numerous individual studies ( Supplementary Information 1.2).
Direct constraints do not exist everywhere, and so the extents of sedimentary basins are unde ned in many areas. A systematic understanding of basins has also been limited by the wide diversity of data and methods used and variable mapping criteria.
Here we develop the rst sedimentary basin likelihood map for Antarctica using the supervised machine learning method Random Forest (RF) 10 . RF has proven to be a valid tool in lithology classi cation, which has high predictive performance with limited training information 11 . We apply this method to generate a likelihood model based on the current understanding of Antarctic bedrock type distribution. Evidence layers are sourced from the available continental-scale geophysical datasets including bedrock topography 9 , gravity eld 12 , magnetic eld 13

East Antarctic basins
In East Antarctica, major coast-perpendicular fault systems bound the sedimentary basin distribution in Wilkes Subglacial Basin 16 . We infer the presence of an extensive sedimentary basin preserved in the southern Wilkes Subglacial Basin, transitioning to less extensive cover in the north. This is consistent with a more complex history in the north, combining differential subsidence and glacial erosion 16 . For the major ice streams, our map shows sedimentary basins preserved beneath the Cook and Ninnis Glacier catchments, in contrast to a crystalline bed beneath the Mertz Glacier catchment (Fig. 3a).
The coast in Western Wilkes land is dominated by crystalline bedrock, however, the inland region contains broadly-distributed sedimentary basins including Aurora, Vincennes, Sabrina and Knox Subglacial Basins ( Fig. 2). Ice mass loss in this region is concentrated on Denman, Totten and Vanderford Glaciers 20 , with the sedimentary basins located in the upper catchments of these ice streams (Extended Data Fig. 1). The sedimentary basin distribution for the Totten Glacier catchment has been interpreted to represent repeated instability driven retreat over the region now possessing sedimentary basins 21 . As well as reverse bed slopes 21 , fast retreat and readvance here may be exacerbated by hydrogeological feedbacks 6 .
In the Recovery region, major basins are de ned in the upper catchments of the Bailey, Slessor and Recovery Ice Streams, linking to coastal basins (Fig. 3b). Airborne radar observations are consistent with the presence of a subglacial hydrological system that originates at lakes in the upper Recovery Ice Stream, controlling its dynamics 22 . The broad basins we resolve are located beneath this active lake system, and could support the latter's interaction with a deep groundwater system. We also map two distinct sedimentary basins near the South Pole: The Pensacola-Pole Basin extending into the Foundation Ice Stream catchment, and a separate basin extending to the Wilkes Subglacial Basin (Fig. 2).
The basins are associated with large numbers of subglacial lakes 18  In the Siple Coast region (Fig. 3c), the map shows a sedimentary basin interspersed with extensive volcanoes 15 , re ecting the volcanic-sedimentary nature of the West Antarctic Rift System 26 . The overlying ice streams are characterised by rapid changes in the ice-sheet state controled by the till property variations coupled with subglacial hydrology 27,28 . In the Bellingshausen Sea sector, a sedimentary basin is preserved in the Ferrigno Rift, interpreted to enhance the dynamics of the ice stream above 29 . In the Amundsen Sea region (Fig. 3d), the map shows a transition from sedimentary basin to crystalline bedrock on the inner continental shelf, recording paleo-ice stream history 30 . For Pine Island Glacier, the paleo-grounding line has undergone rapid retreat through the sedimentary basin and crystalline bed transition region due to high meltwater ux 30 . Our map shows moderate basin likelihood (~0.5) at the grounding line, but the upper catchment is clearly associated with a sedimentary basin. In Thwaites Glacier, we nd the downstream side is dominated by a crystalline bed, while the upper ice stream preserves a broad sedimentary basin. The transition from sedimentary basin to a crystalline bed is coincident with a transition from a distributed subglacial hydrological system to one dominated by a basal network of channels 19  Till conditions are essential in facilitating basal sliding, with many feedbacks with sediment sources, hydrology and ice-loading processes recognised 32 . In many cases, a weak till layer is thought to reduce the basal friction facilitating enhanced ice-sheet ow 33 , whereas the lack of a weak till layer is commonly associated with much slower ow 34 . An important factor is a consistent supply of till in the so-called "till conveyor" which depends on the upstream erosion of bedrock 35 . In contrast to a crystalline bedrock, a sedimentary basin is mechanically weaker due to reduced competency and may also have layered structures leading to higher erodibility 36 . Consequently, the existence of sedimentary basins in the upper portions of ice streams may favour a sustained supply of subglacial till 8 .
Subglacial hydrology is a critical factor in ice-sheet dynamics, with wet-based glaciers showing faster ow and enhanced basal sliding overall, compared to cold-based glaciers 37 . Subglacial hydrology also strongly impacts till conditions: The input of basal water commonly weakens a till layer, while the extraction of pore water consolidates it 38 . In addition, freshwater that ows across the grounding line in uences ice shelf stability with feedbacks that affect grounded ice dynamics 17 . Finally, the circulation of water in subglacial aquifers may transport heat from depth to the ice-sheet bed, with impacts on till and ice rheology and hydrology 6 .
The presence of permeable rocks at the ice-sheet bed therefore introduces the potential for interaction between the basal water system and deep groundwater aquifer systems [5][6][7] . In particular, ice-sheet retreat and unloading may lead to increased water ux at the ice-sheet bed 6 . We use a 2D Control Volume Finite Element Model 39 to investigate the potential impact of a permeable bed on water ux for retreat scenarios. We test the hypothesis that where a permeable bed exists in the upstream catchment, ice-sheet retreat causes a signi cant increase in basal water discharge rate relative to an entirely crystalline bed. Secondly we investigate the scaling between the grounding line retreat rate and the basal water discharge rate (Extended Data Fig. 2).
Our simulations indicate that during ice-sheet retreat water discharges into the basal water system due to ice-unloading. The subglacial vertical water ux is controlled by the permeability and thickness of the sedimentary strata. Our base scenario has a 3 km thickness sedimentary basin with vertical permeability of crystalline basement, con ned unit (clay and shale) and aquifer (sandstone) at 10 -19 , 10 -17 , 10 -15 m 2 respectively. Unloading over 10 ka with a grounding-line retreat rate of 130 m a -1 causes an additional mean subglacial water ux of up to 1.96 mm a -1 compare with crystalline basement only (Fig. 4a).
Considering the mean basal melt rate for grounded ice sheet is 5.3 mm a -1 40 , this additional water ux from groundwater has a signi cant contribution to the hydrologic budget, as indicated previously for the Siple Coast ice streams, where up to 45% of ux may be groundwater derived 28 . We nd increasing basal aquifer permeability and thickness facilitates higher groundwater ux rates during ice-sheet unloading, although extremely high permeability (e.g. in gravels or fractured rock aquifers) may sink major basal water into the groundwater system ( Supplementary Information 2.3.1).
Increased water ux, as we indicate above, has further implications for both heat ux and till conditions. Water circulating in the sedimentary basin advects the geothermal heat from deeper aquifers to the icesheet bed. With a 60 mW m -2 bottom ux boundary condition, we model 2-5 mW m -2 higher heat ux due to groundwater discharge during the retreat phase ( Supplementary Information 2.3.3), leading to a potential enhancement of basal melting (0.2-0.5 mm a -1 ). In the presence of till, both water from enhanced basal melting, and water discharged from groundwater may weaken the till strength at the bed 38,41 . Finally, upward water ux may balance the gravitational weight of sediment causing sediment liquefaction 42 . Combined, these processes will reduce the basal friction to enhance the ice ow, promoting dynamic ice mass loss.
Crucially, high ice stream retreat rates may have even more marked effects on the magnitude of water ux ( Fig. 4c-d). For a retreat rate of 300 m a -1 , approximately that observed for Thwaites Glacier 43 , enhanced basal water ux of 5 mm a -1 is modelled, comparable to modelled basal melt rates in the upper Thwaites catchment 44 . A hypothetical fast retreat of the ice sheet (1,300 m a -1 ) enhances basal water ux substantially (up to 20 mm a -1 ) which may be substantial even in the lower catchments of fast retreating glaciers.
Our sedimentary basin map for Antarctica indicates that the fastest-changing ice catchments in East and West Antarctica possess sedimentary beds in their upper portions (Fig. 3), indicating that processes linked to sedimentary basins contribute signi cantly to catchment-scale ice dynamics. We model enhanced groundwater discharge into the ice-sheet bed where ice retreat is coupled with permeable aquifers upstream, so increasing subglacial water ux. Under moderate retreat scenarios with rates < 300 m a -1 , groundwater ux may be enhanced by up to 5 mm a -1 , comparable with average basal melt rates driven by geothermal heat ux for Antarctica 44 . While not insigni cant, this is small in comparison to basal melt rates in fast-owing glaciers, which may reach 20-100 mm a -144 . However, in a fast retreat or collapse scenario with rates > 650 m a -1 the groundwater ux exceeds 10 mm a -1 , potentially sustaining a signi cant source of subglacial water even in high basal melt rate environments. The enhanced ux caused by retreat feeds back into dynamic processes affecting basal sliding including a weakened till layer, enhanced heat ux and enhanced basal melt rate and also potential destabilisation of ice shelves 17 . We propose that the instabilities caused by enhanced subglacial water supply from sedimentary aquifers, triggered by rapid ice-sheet retreat, are a crucial mechanism controlling the vulnerability of the Antarctic ice-sheet.

Sedimentary basin mapping with Random Forest prediction
The RF prediction algorithm 10 is built based on an ensemble of decision trees 45 . The randomness of RF is guaranteed by the uniqueness of each decision tree. Every tree in RF is constructed by a bootstrapped sample method (select nearly 2/3 of total training data with replacement), and grew by a random subset of evidence layer at each split. For bedrock type classi cation, each tree "votes" for the bedrock class, with the nal result is assembled by the result of all uncorrelated trees. In each location, the averaged vote represents the probability of each bedrock class being present. For the binary bedrock type classi cation problem in this study, a 0.5 likelihood is a natural boundary representing sedimentary basin or the crystalline bedrock. The R core 4.0.2 and package randomForest 46      Groundwater discharge through time for moderate to fast retreat rates. a-d, Vertical water ux at the icesheet bed associated with various ice-sheet retreat rates. a, 130 m a-1. b, 260 m a-1. c, 650 m a-1. d, 1,300 m a-1. Black dashed line shows the location of the ice-sheet margin with subglacial water ux imaged below the line. Positive vertical water ux indicates groundwater discharge while negative vertical water ux indicates recharge. We show sedimentary basin geometry at the right. Aquifer unit A has horizontal permeability κx =10-14 m2 and vertical permeability κz =10-15 m2; the con ned unit C has κx =10-16 m2 and κz =10-17 m2; the crystalline bedrock basement B has κx = κz =10-19 m2. The ice expands over 20 ka to glacial maximum at 1300 km, before initiating retreat. The grey area shows the parabolic ice-sheet geometry at glacial maximum. e, Mean subglacial water ux changes relative to a crystalline bedrock for different retreat rates. The permeable bed promotes enhanced discharge from the upstream basin during ice-sheet retreat and the mean discharge-rate scales with retreat rate.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.