Volume 45, Issue 20 p. 11,206-11,218
Research Letter
Free Access

Subglacial Conduit Roughness: Insights From Computational Fluid Dynamics Models

Yunxiang Chen

Yunxiang Chen

Department of Civil and Environmental Engineering, Pennsylvania State University, University Park, PA, USA

Search for more papers by this author
Xiaofeng Liu

Corresponding Author

Xiaofeng Liu

Department of Civil and Environmental Engineering, Pennsylvania State University, University Park, PA, USA

Correspondence to: X. Liu,

[email protected]

Search for more papers by this author
Jason D. Gulley

Jason D. Gulley

School of Geosciences, University of South Florida, Tampa, FL, USA

Search for more papers by this author
Kenneth D. Mankoff

Kenneth D. Mankoff

Department of Glaciology and Climate, Geological Survey of Denmark and Greenland (GEUS), Copenhagen, Denmark

Search for more papers by this author
First published: 09 October 2018
Citations: 8


Flow resistance in subglacial conduits regulates the basal water pressure and sliding speeds of glaciers by controlling drainage efficiency and conduit enlargement and closure. Flow dynamics within subglacial conduits, however, remain poorly understood due to limited accessibility. Here we report the results of the first computational fluid dynamics simulations of flow within a realistic subglacial conduit beneath Hansbreen, a polythermal glacier in Svalbard, Norway. The simulated friction factor is 2.34 ± 0.05, which is around 5 to 230 times greater than values (0.01–0.5) commonly used in glacier hydrological modeling studies. Head losses from sinuosity and cross-sectional variations dominate flow resistance (∼ 94%), whereas surface roughness from rocks and ice features contributes only a small portion (∼6%). Most glacier hydrology models neglect head losses due to sinuosity and cross-sectional variations and thus severely underestimate flow resistance, overestimating the conduit peak effective pressure by 2 times and underestimating the conduit enlargement area by 3.4 times, respectively.

Key Points

  • For the first time, flow resistance in a real subglacial conduit is quantified using CFD and structure-from-motion photogrammetry
  • Flow resistance in subglacial conduits is dominated by sinuosity and cross-sectional variations, not surface roughness from rocks and ice
  • Most glacier hydrology models severely underestimate flow resistance and significantly misrepresent subglacial flow and ice dynamics

Plain Language Summary

Subglacial conduits drain meltwater from polar ice sheets, thus directly regulating the ice sheet sliding speed through basal flow resistance and water pressure inside the conduits. Despite their importance, our understanding of subglacial conduits is extremely limited due to difficulties of observing them and their interiors with either remote sensing or in situ exploration. Simplified models have been proposed for the hydraulics inside these conduits. A key problem in these models is the lack of scientific support in parameterizing the flow resistance. Currently, the resistance is parameterized but has not been validated due to the accessibility issues. To narrow this knowledge gap, we performed three-dimensional computational fluid dynamics simulations based on a millimeter-scale resolution model of an actual subglacial conduit in the Arctic. For the first time we give a direct and physics-based estimation of the flow resistance in an actual subglacial conduit and highlight the important contributions of the cross-sectional variations and longitudinal sinuosity. We further demonstrate the impacts of our simulated flow resistance on subglacial hydrodynamics and ice sheet dynamics.

1 Introduction

Subglacial water pressure exerts an important but poorly understood control on the velocity of glaciers and ice sheets. A growing body of literature has theorized about (Clarke, 2005; Creyts & Schoof, 2009; Röthlisberger, 1972; Weertman, 1972), remotely observed (Hock & Hooke, 1993; Hooke, 1989; Iken & Bindschadler, 1986; Nienow et al., 1998), and modeled (Covington et al., 2012; Creyts & Schoof, 2009; Hewitt, 2013; Mankoff & Tulaczyk, 2017; Schoof, 2010; Werder et al., 2013) glacier hydrological systems, but difficulties in acquiring data in subglacial conduits have limited the understanding of subglacial drainage efficiency and its impacts on ice motion. Consequently, much of our understanding of subglacial hydrological systems has been informed by the results of numerical models. A persistent problem in models of subglacial hydrology is how to parameterize water flow resistance in conduits. Resistance to water flow is related to hydraulic roughness, a loosely defined term encapsulating energy that is dissipated as heat due to turbulence and viscous forces (i.e., headlosses). Because the heat generated by these headlosses is directly responsible for conduit enlargement which further affects the hydrodynamics (e.g., water velocity and water pressure) in subglacial conduits and associated ice sheet dynamics (Bell, 2008), accurate parameterization of roughness and flow resistance is necessary to accurately simulate the evolution of subglacial hydrological systems.

Direct simulations of energy dissipation require resolving scales ranging from the conduit cross-section scale (of the order 1–2 m; Mankoff et al., 2017) to the Kolmogorov scale (of the order 1–2 μm) and require a computational grid number of the order 1016 which exceeds the memory capacity of most current computers (Choi & Moin, 2012). Headlosses are therefore commonly parameterized through physical features that contribute to hydraulic roughness. In glacier hydrological models, roughness is usually defined as the difference between the original conduit and a simplified straight pipe, which characterizes the effects of boulders or bedrock protruding into the flow field (defined here as surface roughness). If the height of surface roughness and the hydraulic diameter of conduits are known, hydraulic roughness coefficients such as the Darcy-Weisbach friction factor can be determined from empirical relationships. Relating friction factor to surface roughness height allows friction factors to be calculated as the hydraulic diameters of conduits fluctuate due to melt and creep. In studies of rivers, bedforms and sinuosity, rather than surface roughness, have been shown to dominate flow resistance (Francalanci et al., 2012; Hey, 1988; Millar, 1999; Parker & Peterson, 1980). However, due to lack of data, the impacts of sinuosity and cross-sectional variations on the resistance of subglacial conduits have rarely been considered, which partially explains the discrepancy between the friction factors commonly used in glaciology models and those obtained in field measurements. For example, numerical models of glacier hydrology are typically parameterized with Darcy-Weisbach friction factors in the range of 0.01–0.5 (Boulton et al., 2007; Clarke, 1996; Colgan et al., 2011; Covington et al., 2012; Fowler, 2009; Schoof, 2010; Spring & Hutter, 1981). In contrast, dye-tracing studies suggest that hydraulic roughness values are orders of magnitude higher, such as one dye tracing study in Svalbard found that Darcy-Weisbach friction factors fluctuated between 0.97 and 75 in response to changes of water level in a subglacial conduit (Gulley et al., 2014).

To better quantify the bulk flow resistance in a real subglacial conduit and the contributions of complex surface features (Figure 1), we perform computational fluid dynamics (CFD) simulations over a 10-m-long millimeter-scale resolution 3-D model of an actual subglacial conduit (Figure 1d) beneath Hansbreen (a polythermal glacier in the archipelago of Svalbard; Gulley et al., 2012; Mankoff et al., 2017) and synthetic models (Figure 2a) derived from the actual model. The quantified bulk flow resistance is then fed to a one-dimensional subglacial model (Spring & Hutter, 1981) to evaluate its impacts on subglacial hydrodynamics.

Details are in the caption following the image
Field site and the conduit beneath Hansbreen, Svalbard. (a) The reconstructed 125-m-long 3-D conduit model and the 10-m millimeter-scale subsection (box near A10). (b) A photograph from inside showing the unique composition of the conduit with ice roof and rocky bottom. (c) A conceptualized cross section of the conduit and the associated dynamic processes (solid red arrows). (d) The centerline (the longest red line), mean centerline (the dashed blue line), and cross-section segments (lines perpendicular to the centerline).
Details are in the caption following the image
Geometric processing, computational domain, mesh, simulated flow field, flow resistance, and sinuosity. (a) The definition of different surfaces: S0 is the original conduit with all surface features, S1 is a synthetic surface containing cross-sectional variations and surface roughness, S2 contains only surface roughness, P1 is a pipe with equivalent volume and length as S1, and P2 is diameter-variable pipe. (b) Computational fluid dynamics simulation domain (cyan) and the extended region (yellow) for generating turbulent inflow. (c) A cutaway showing the internal mesh of the simulation domain. (d) Instantaneous streamwise velocity field. (e) The time history of friction factor for simulations in Groups 0–2. The time is nondimensionalized by flow through time defined by T = L0/U0 or T = L1/U0 (see U0 in Table 1). The final friction factor is defined as the time-averaged friction factor after reaching the quasi-steady status (the region to the right side of the vertical dashed line). (f) Streamwise variations of three sinuosity metrics of original conduit S0.

2 Methods

To untangle the effects of different surface features, we processed the conduit surface as follows (Figure 2a). To isolate the effects of sinuosity on flow resistance, we removed the sinuosity from the initial 3-D model surface S0, by straightening its centerline and creating surface S1. To isolate the effects of cross-sectional variations on flow resistance, we removed the cross-sectional variations from the straightened surface S1 by subtracting the cross-sectional variations from the surface S1 and creating surface S2. Here the cross-sectional variations are defined as the difference between a variable diameter pipe (P2) and a constant diameter pipe (P1), which characterize both the longitudinal shrinking trend and small variations. The diameter of P1 is D1= urn:x-wiley:grl:media:grl58127:grl58127-math-0001 with V1 and L1 denoting the volume and length of S1. The diameter of P2 is calculated by D(x)=2 urn:x-wiley:grl:media:grl58127:grl58127-math-0002 with A(x) denoting the area of a cross section at location x on surface S1 (see Text S1 in the supporting information for details of surface decomposition and the definition of surface subtracting). We categorized the simulations into four groups (0–3) corresponding to surfaces S0, S1, S2, and P1 for convenience (see Table 1 for configurations and main results).

Table 1. The Simulation Cases, Setup, and Results Summary
Group Case Surface σr urn:x-wiley:grl:media:grl58127:grl58127-math-0003 U0 RD δ ft fc±σfc Fp/Ft
name name name (m) (m/s) 106 (cm) (%)
0 S0a S0 NA NA 1 1.34 2.19 NA 2.46 ± 0.06 99.4
S0b S0 NA NA 1 1.34 1.70 NA 2.34 ± 0.04 99.3
S0c S0 NA NA 1 1.34 1.32 NA 2.34 ± 0.05 99.3
S0d S0 NA NA 0.5 0.67 1.94 NA 2.42 ± 0.04 99.2
S0e S0 NA NA 1 1.34 1.94 NA 2.41 ± 0.05 99.3
S0f S0 NA NA 2 2.67 1.94 NA 2.40 ± 0.05 99.4
1 S1a S1 0.21 0.16 0.5 0.66 1.65 0.135 1.95 ± 0.04 99.1
S1b S1 0.21 0.16 1 1.33 1.65 0.135 1.94 ± 0.04 99.2
S1c S1 0.21 0.16 2 2.65 1.65 0.135 1.94 ± 0.05 99.3
2 S2a S2 0.19 0.14 1 1.33 1.48 0.123 0.142 ± 0.003 96.4
3 S3a P1 0 0 1 1.33 1.27 0.0111 0.0103 ± 0.00004 0
S3b P1 0.027 0.02 1 1.33 1.57 0.0486 0.0467 ± 0.00033 0
  • Note. Symbols denote surface roughness (σr), transverse scale of a pipe (De), mean velocity over the pipe inlet (U0), relative roughness (σr/De), Reynolds number (RD), and average grid size (δ). Reynolds number is defined as RD=U0De/ν with ν denoting the kinematic viscosity of water. Average grid size is defined as δ = (Vi/Nc)1/3 with Vi and Nc denoting the total volume and number of grid cells of each computational domain. ft and fc denote the friction factors calculated from existing rough pipe theories and time-averaged friction factor from CFD simulations, respectively. urn:x-wiley:grl:media:grl58127:grl58127-math-0004 is the 1 standard deviation of the instantaneous friction factor away from the time-averaged value. Fp/Ft is the ratio of pressure force to the total resistance force. (See values of the above variables in section 2 and Appendices Appendix A and Appendix B; see values Vi, Nc, and the maximum and minimal values of fc in Table S1). CFD = computational fluid dynamics; NA = not applicable.
With the original conduit S0 and synthetic conduits (S1, S2, and P1), we solve the Navier-Stokes equations for incompressible flows with the large eddy simulation (LES) turbulence model due to its better physical representation of turbulence than Reynolds-averaged models and less computational costs compared to direct numerical simulation (Argyropoulos & Markatos, 2015). The LES resolves the large-scale turbulent flow which is quantified by filtered velocity and pressure and parameterizes the effects of unresolved small-scale turbulence by a turbulence closure model with the following form:
where urn:x-wiley:grl:media:grl58127:grl58127-math-0007 is the filtered velocity, urn:x-wiley:grl:media:grl58127:grl58127-math-0008 is the pressure divided by the constant water density, t is time, xi and xj are spatial coordinates, i and j are both integer indexes, and νsgs is the modeled subgrid scale (SGS) eddy viscosity for turbulence closure, which accounts for the influence of the unresolved turbulent motions (see Appendix Appendix A for the details of SGS eddy viscosity model). For Groups 0–2, boundary conditions for the mapped inflow, the wall of the pipe, and the outflow (Figure 2b) are determined as follows: we assumed a fully developed turbulent flow for the mapped inflow, which was numerically achieved by an internal mapping between the inflow and the mapped inflow (de Villiers, 2006; Lund et al., 1998). This internal mapping contains two operations: interpolating the flow variables (velocity and turbulent kinetic energy) on the mapped inflow boundary to the inflow boundary and adjusting these interpolated variables to satisfy a target mean velocity U0 over the inflow boundary (see details of flow adjusting in Appendix Appendix A). As the estimated mean flow velocity in Greenland subglacial conduits is in the range of 0.1 to 10 m/s (Chikita et al., 2010; Fountain & Walder, 1998), we set U0 as 1 m/s for most of the simulations. U0 values of 0.5 and 2 m/s were also simulated to investigate the effect of Reynolds number. The wall of the pipe was represented by a body-fitted mesh, and thus, we set it as a no-slip condition with a zero velocity and a zero gradient for pressure. The Spalding wall function for SGS eddy viscosity νsgs is also used to avoid resolving the viscous sublayer (Spalding, 1961; see Appendix Appendix A for details of implementing a wall boundary and see Text S2 for uncertainty analysis of using a wall function; Celik et al., 2008; Choi & Moin, 2012; Napoli et al., 2008). For the outflow, a zero-gradient boundary condition was applied to all flow variables except the pressure, which had a fixed value of zero. For Group 3 where the flow in a straight pipe P1 was simulated, we are able to generate a mesh with identical internal mesh structure on the inflow and outflow boundaries. This enables the application of cyclic boundary conditions on the two boundaries with a specific mean velocity of 1 m/s. We also impose a rough wall function on pipe P1 with a given relative roughness 2% which corresponds to a friction factor of 0.05 commonly used in glaciology models (Colgan et al., 2011; Fowler, 2009), to show how an estimated surface roughness in real subglacial conduits can be applied in future CFD models of large-scale conduits. Simulations in Group 3 are also used to validate our computational models because theories exist to predict their friction factor ft.
The total flow resistance (Ft) is composed of wall shear-induced friction force (Ff) and pressure-induced form drag (Fp). With available CFD data of wall shear and pressure, the simulated bulk friction factor (fc), nondimensionalized by the characteristic force 1/8 urn:x-wiley:grl:media:grl58127:grl58127-math-0009 (Munson et al., 2009), can be calculated by equation 3.
where Ae is a characteristic conduit surface area. For surface S0, Ae=πD0L0; for surfaces S1, S2, and P1, Ae=πD1L1. Sw denotes the wall surface; urn:x-wiley:grl:media:grl58127:grl58127-math-0011 and urn:x-wiley:grl:media:grl58127:grl58127-math-0012 denote the water density divided pressure and wall shear stress, respectively. urn:x-wiley:grl:media:grl58127:grl58127-math-0013 is the unit vector along the flow direction. The semiempirical formulas for friction factor estimation (Schlichting, 1979) used in this paper are listed in Appendix Appendix B.

Flow resistance greatly impacts the physical properties of subglacial conduit systems, such as melt opening and creep closure rates, discharge, water velocity, and effective pressure. To evaluate the impacts of our simulated friction factor (fc) on conduit dynamics relative to values commonly used, such as 0.1 (Schoof, 2010), we solved the governing equations of the mass, momentum, and energy balance for a simplified one-dimensional circular straight conduit (Spring & Hutter, 1981; see Appendix Appendix C for details of governing equations and parameters). Other parameters typically used in this model include conduit slope angle θ (=3°; Schoof, 2010), ice sheet thickness Hi (=1,000 m), the inflow discharge to moulin/lake from upstream rivers Q0 (= 3 m3/s; Swift et al., 2005), moulin/lake area AL(=50 m2), conduit length l (=50 km), and initial conduit size S0 (=1.41 m2).

3 Results

We show results in the order of Groups 3, 2, 1, and 0 which have increasingly complex surface features. Starting from the two simulations in Group 3 with pipe surface P1, the simulated Darcy-Weisbach friction factor (fc) is around 0.01 and 0.05 for the relative roughness of 0% and 2%, respectively. These values compare well with empirical formulas for pipe flows (see equations B1-B3 and Table 1), which validates our computational model.

In Group 2, one simulation was performed with the surface S2, where both sinuosity and cross-sectional variations have been removed, and only surface roughness remains. The simulated friction factor has a value of 0.142, as illustrated by the time history of case S2a in Figure 2e. This value is about 14 times larger than those for smooth pipe in Group 3. Following the definition of roughness in this paper, we quantified the roughness height on rough surface as the standard deviation of the radius of each cross section (Mankoff et al., 2017). The global roughness height (averaged over all cross sections) was calculated to be σr = 0.19 m. Assuming empirical formulas is applicable to subglacial conduits; an estimated friction factor ft has a value of 0.123, which is close to the simulated value of 0.142. The good agreement between the two indicates that the roughness defined in this work is able to predict the portion of flow resistance due to surface roughness.

In Group 1 simulations with the surface S1, where only the sinuosity is removed while the surface roughness and cross-sectional variations remain on the surface, the simulated friction factor increased to around 1.94, as illustrated by the time history of case S1b in Figure 2e. This is about 200 times larger than the value with surface P1 and about 14 times larger than the value with surface S2. If using the linear superposition principle, which assumes that contributions of different sources of flow resistance are independent (Einstein & Banks, 1950; Leopold et al., 1960), the friction factor due to cross-sectional variations is around 1.8 which is 13 (=1.8/0.142) times of that due to surface roughness.

In Group 0 with the original surface S0, where surface roughness, cross-sectional variations, and sinuosity are all included, the simulated Darcy-Weisbach friction factor is about 2.34 (case S0c) in Figure 2e, which is about 230 and 16 times of the resistance of cases S3a and S2a, respectively. The contribution of sinuosity to flow resistance is about 0.4, calculated by taking the difference between cases S0c and S1b with holding the linear superposition principle.

4 Discussion

4.1 Important Surface Features

With the CFD simulations, we quantify the bulk flow resistance in a realistic subglacial conduit as around 2.34, with cross-sectional variations, sinuosity, and surface roughness contributing around 1.8, 0.4, and 0.14, respectively. Cross-sectional variations therefore contribute the biggest portion in the bulk flow resistance (77% of the total resistance), with sinuosity contributing about 17% and surface roughness only contributing about 6%. This conclusion is made based on the linear superposition principle which assumes no interactions among all surface features. There are other studies only assuming no interactions between surface roughness and cross-sectional variations but considering sinuosity as an enhanced effect (Cowan, 1956; Lane, 2005; Powell, 2014; Yen, 2002). Following this assumption, the total resistance can be partitioned as f=(fR+fC)(1+fS) with fR, fC, and fS denoting the resistance attributed to surface roughness, cross-sectional variations, and sinuosity. The friction factor calculated from the original surface S0 is around 2.34, while that from surface S1 (sinuosity removed) is around 1.94, which means that the enhanced effect is about 21% (=2.34/1.94-1). Removing the sinuosity effect, surface roughness contributes to 6% and cross-sectional variations contribute to 73%. Here the relative importance of surface roughness to cross-sectional variations is unchanged because we still assume no interactions between these two features. While these percentages are only rough estimations, they likely explain the big discrepancy between field-measured friction factors and those typically used in models (Gulley et al., 2014), as models of subglacial conduits have relied on numerical simulators that do not simulate cross-sectional variations or sinuosity impacts on flow resistance which contribute around 94% (=17%+77% or 21%+73%) of the total resistance. As assumptions of interactions between different surface features were made in the resistance partition methods, we performed one simulation using a surface generated by only removing surface roughness, to consider the effects of these assumptions (see Text S3).

It is important to note that there is little sinuosity in this mostly straight 10-m section. As seen in Figure 1a, prior to and after this section there are approximately 90° bends in the subglacial conduit. However, even this minor sinuosity in this section contributes to a flow resistance 3 times of that due to surface roughness. We observe that such flow resistance is attributed to three metrics in sinuosity: transverse oscillation angle θxy, vertical oscillation angle θxz, and oscillation distance rxy (see Figure 1d for definitions). For a straight pipe, all these metrics have a value of zero. For a realistic conduit, however, all have nonzero values and contribute to flow resistance. Figure 2f shows how these metrics for this real conduit vary in the longitudinal direction. For conduits with large sinuosity over a longer distance in natural subglacial systems, the flow resistance due to these metrics likely becomes more important. In addition, the cross-sectional variations in this 10-m section contain both a long-term shrinking trend and small-scale local variations. These local variations are more likely uniform across the conduit, but the long-term trend varies a lot along flow direction. Considering the dominant contributions of cross-sectional variations and sinuosity, our work suggests that the next generation glaciology models need to explicitly consider the effects of the long-term trend of the conduit size and the sinuosity metrics. Our CFD results also show that the simulated friction factor is independent of Reynolds number (see Text S4).

4.2 Impacts on Hydrodynamics and Ice Dynamics

By solving the governing equations (Appendix Appendix C) for a simplified one-dimensional straight circular subglacial conduit with our simulated friction factor 2.34 and the commonly used value 0.1, we highlight the impacts of parameterized flow resistance on conduit enlargement, subglacial hydrodynamics, and associated ice sheet dynamics. For the case simulated here, the results show that the conduit evolution can be divided into two stages: the initial adjusting stage (from the beginning to the first peak of conduit properties; see vertical dashed lines in Figure 3) and the quasi-equilibrium stage (the time after the first peak). During the initial adjusting stage, the conduit is less developed and cannot accommodate all inflow to the moulin (Q/Q0<1; see Figure 3c), resulting in water overspilling the moulin and creating a constant pressure head system. Constant pressure head conditions remain until enlargement of the subglacial conduit begins lowering moulin water levels. The timescale for initial lowering of moulin water levels is 6.7 and 1.7 days for the friction factor value of 2.34 and 0.1, respectively, which means that the use of a smaller friction factor (0.1) underestimates the initial adjusting stage by about 4 times (=6.7/1.7).

Details are in the caption following the image
The comparison of the time evolution of subglacial conduit properties by using the commonly used friction factor 0.1 and computational fluid dynamics simulated friction factor 2.34 at fixed input discharge rate 3 m3/s, ice thickness 1,000 m, and conduit slope 3°. (a) The moulin/lake water depth. (b) Ice melting rate and creep closure rate. (c) Conduit discharge. (d) Conduit cross-sectional size. The dashed lines are fitted curves with β1=0.098/day and β2=0.032/day. (e) Water velocity. (f) Effective pressure.

At the quasi-equilibrium stage (t> 6.7 days), the conduit has fully developed and its drainage is mainly constrained by inflow discharge to the moulin (Figure 3c). At this stage, all conduit properties are approximately constant despite small amplitude oscillations and the mean ice melting rate due to viscous dissipation is approximately constant (Figure 3b). Simulations parameterized with the smaller friction factor 0.1 underestimated the average conduit cross-sectional area by 3.4 times (=3.4/1.0, Figure 3d) and overestimated the water velocity by 3.5 times (=3.2/0.92, Figure 3e) compared to those with the friction factor 2.34. The use of the smaller friction factor 0.1 also underestimates the moulin water level and effective pressure fluctuation period by about 1.6 times (1.6 days with f=2.34 vs. 1.0 day with f=0.1; Figures 3a and 3b) and overestimates the effective pressure peak value by about 2 times (Figure 3f).

While our CFD modeling provides new insights on subglacial conduit roughness, flow resistance in real conduits is probably even higher than what we have calculated here and likely much higher than values commonly employed in subglacial hydrological models (Boulton et al., 2007; Clarke, 1996; Colgan et al., 2011; Covington et al., 2012; Fowler, 2009; Schoof, 2010; Spring & Hutter, 1981). For a longer subglacial conduit with larger sinuosity and more frequent cross-sectional size changes, the friction factor is expected to increase many times than the friction factor of 2.34 used in our models. Indeed, friction factors determined from dye-tracing studies indicate that friction factors can be as high as 75 (Gulley et al., 2014). While we assumed a fixed hydraulic roughness value in our modeling, friction factors in real subglacial conduits are likely to be time varying. Friction factors could be especially high during the early stages of conduit enlargement when the ice roof is much closer to the conduit floor.

Regardless of the cause, our results demonstrate that diurnal fluctuating amplitudes in moulin water level, and hence effective pressure, are attenuated with elevated conduit flow resistance (Figures 3a and 3f). Because ice sliding speeds are inversely proportional to effective pressure (Bindschadler, 1983; Fowler, 1986), lower effective pressures in rougher conduits suggest that the ability of channelization to decrease sliding speeds of glaciers and ice sheets may be diminished in natural systems. The influence of hydraulically rough conduits on effective pressure and ice sliding speeds should be explored further, as increased channelization of subglacial drainage is widely regarded to act as a buffer against future meltwater-based acceleration of the Greenland Ice Sheet (Nienow et al., 2017; Sole et al., 2013). Further research should be conducted when larger-scale 3-D models of subglacial conduits become available.


This research was supported by the National Science Foundation Award OPP 1503928. The CFD simulations were performed on the HPCs at the Institute for CyberScience at the Pennsylvania State University. Conduit geometric data and CFD simulation cases are available at https://doi.org/10.18739/a23z3w.

    Appendix A: Computation Model

    Accurate quantification of the flow resistance in subglacial conduits depends on the accurate calculation of wall shear stress and form drag. In the CFD simulations, LES were performed, which resolved large turbulent structures within the conduit and parameterized the effects of unresolved scales on the large scales. These parameterized effects are represented by an eddy viscosity approximation which has the form of urn:x-wiley:grl:media:grl58127:grl58127-math-0014 with urn:x-wiley:grl:media:grl58127:grl58127-math-0015, νsgs, and urn:x-wiley:grl:media:grl58127:grl58127-math-0016 denoting the unresolved SGS stresses, SGS eddy viscosity, and the resolved strain rate, respectively (Wilcox, 2006). With this eddy viscosity approximation, the filtered Navier-Stokes equations for incompressible fluids have the form of equations 1 and 2. There are many models for the eddy viscosity in the literature. In this work, the localized dynamic one-equation model was chosen due to its advantages in dealing with backscatter, numerical instabilities, history, and nonlocal effects, which are important for modeling complex turbulent structures generated by complex geometries (Kim & Menon, 1995). The SGS model has the form of the following:
    where ksgs is the subgrid kinetic energy. 〈Ck〉 is a constant used to guarantee numerical stability and can be computed as urn:x-wiley:grl:media:grl58127:grl58127-math-0019 with 〈〉 denoting spatial average over the computational domain. Δ is the filter grid size. The value of Ck can be dynamically calculated as in standard dynamic models (Davidson, 1997; Germano et al., 1991; Kim & Menon, 1995).

    The simulation cases for each group are designed as follows. Group 0 has six cases with S0aS0c designed to test mesh independence and S0dS0f designed to test the effects of Reynolds number. Group 1 also contains three cases to test the effects of Reynolds number. The change of the Reynolds number is through the changing of mean velocity U0. In Group 2, only one simulation was performed to calculate flow resistance purely due to roughness as defined in this work. For Group 3, two simulations, one with a smooth wall and another with a rough wall with a prescribed roughness height, were performed based on the volume equivalent pipe P1. The computational domain and mesh generation are shown in Figures 2b and 2c (case S2a is chosen as an example). The turbulent inflow condition was generated using an internal mapping approach where the flow variables on the mapped inflow boundary were first interpolated to the inflow boundary (denoted by urn:x-wiley:grl:media:grl58127:grl58127-math-0020 and urn:x-wiley:grl:media:grl58127:grl58127-math-0021) and then were adjusted to satisfy the target mean velocity U0. If we denote the average values of the interpolated flow variables over the inflow boundary by urn:x-wiley:grl:media:grl58127:grl58127-math-0022 and urn:x-wiley:grl:media:grl58127:grl58127-math-0023, then we adjusted the flow variables as follows: if urn:x-wiley:grl:media:grl58127:grl58127-math-0024, urn:x-wiley:grl:media:grl58127:grl58127-math-0025, otherwise, urn:x-wiley:grl:media:grl58127:grl58127-math-0026. This adjusting is also applied to the turbulent kinetic energy. The boundary conditions for turbulent properties on the inflow and outflow are zero gradient except for the turbulent kinetic energy. We set a no-slip condition on the wall of the pipe, which numerically imposes a velocity urn:x-wiley:grl:media:grl58127:grl58127-math-0027 = 0 and a pressure urn:x-wiley:grl:media:grl58127:grl58127-math-0028 = urn:x-wiley:grl:media:grl58127:grl58127-math-0029 on the wall. Here urn:x-wiley:grl:media:grl58127:grl58127-math-0030 denotes the pressure at the center of the first cell away from the wall boundary at the previous time step. When a wall function is also applied, a water density divided wall shear stress τb is also imposed in the tangential direction of the wall boundary. Therefore, a no-slip boundary constrains the solution by adding two extra force terms, a pressure force due to urn:x-wiley:grl:media:grl58127:grl58127-math-0031 and wall force urn:x-wiley:grl:media:grl58127:grl58127-math-0032. Here Sb, urn:x-wiley:grl:media:grl58127:grl58127-math-0033, and urn:x-wiley:grl:media:grl58127:grl58127-math-0034 denote the area and the normal and tangential directions of a boundary face. If we use y0 and urn:x-wiley:grl:media:grl58127:grl58127-math-0035 to denote the distance from this boundary face to the center of the nearest cell and its tangential velocity component, the wall shear stress can be determined by a logarithmic law urn:x-wiley:grl:media:grl58127:grl58127-math-0036 with urn:x-wiley:grl:media:grl58127:grl58127-math-0037, urn:x-wiley:grl:media:grl58127:grl58127-math-0038, and urn:x-wiley:grl:media:grl58127:grl58127-math-0039 (see the form of G in Spalding, 1961). In this way, the logarithmic law of the wall is imposed as a boundary condition.

    The open source CFD platform OpenFOAM was used for the simulations (CFDDirect, 2017). In OpenFOAM, the governing equations were solved using the finite volume method. The unsteady (temporal) terms were discretized with a backward (implicit) Euler scheme; the convection terms were discretized with a central differencing scheme (termed as Gauss linear scheme in OpenFOAM). A corrected version of the Gauss linear scheme was chosen to discretize the diffusion terms. The linearizations of velocity-velocity (advection term in equation 2) and velocity-pressure coupling were achieved through the Pressure-Implicit with Splitting of Operators (PISO) algorithms (Issa, 1985). PISO is composed of one predictor step and multiple corrector steps. At each time step, the predictor is first implemented by solving the momentum equation with a guessed pressure field (usually the pressure at an old time) and computing predicted mass fluxes at cell faces. The subsequent corrector step contains two types of corrections: a pressure correction achieved by solving a Poisson equation (taking a divergence over equation 2) with the previously updated velocity and fluxes and a velocity correction achieved by solving the momentum equation with the corrected pressure. Such a corrector step was repeated multiple times until the continuity equation is satisfied, followed by solving the turbulent kinetic energy equation (equation A1) and then the time step marches to the next one. It is worthy to note that the velocity-velocity coupling is linearized by replacing one velocity by its value at previous iteration step, and the velocity-pressure coupling is decoupled by solving the momentum equation and Poisson equation separately. The momentum prediction and correction and the turbulent kinetic energy equation were solved with a preconditioned biconjugate gradient method, while the Poisson equation was solved with a generalized geometric-algebraic multigrid method. The computational meshes were generated using the snappyHexMesh tool in OpenFOAM. The initial time step was set as 0.0001 s and allowed to adjust during simulation to not exceed the max Courant number of 0.4. The most time-consuming part of the simulations is the pressure correction. Each time step has about 15–20 corrections, and each correction solves about 1–3 million (the same as the computational cell number) linear equations. To demonstrate the simulated flow field, Figure 2d shows the longitudinal velocity component in half of the S2a case domain. A grid convergence study is discussed in Text S2 and Figure S2.

    In terms of the non-dimensional scales, it is convenient to define the conduit length and the diameter of a pipe whose volume equating to that of the original conduit V0 or straight-up conduit V1. With this method, for Group 0, the longitudinal length scale is L0 = 10 m, the transverse scale (diameter) is D0 = 2R0 = urn:x-wiley:grl:media:grl58127:grl58127-math-0040 = 1.34 m, where V0 = 14.13 m3. For Groups 1 and 2, the longitudinal length scale is L1 = 10.45 m due to straightening, and thus the transverse scale is D1 = 2R1 = urn:x-wiley:grl:media:grl58127:grl58127-math-0041 = 1.33 m, where V1 = 14.53 m3. Here we see that the straightening affects little in the conduit non-dimensional scales for this 10 m section.

    Appendix B: Semiempirical Resistance Formulas

    For pipe flows, there are also semiempirical formulas for friction factor estimation (Schlichting, 1979). The resistance law for smooth pipes is given by the Prandtl's universal law in equation B1. For completely rough pipes, it is given by the von Karman's law in equation B2. Based on the available data on smooth and rough pipes, a correlating formula was also proposed for the whole transition region from hydraulically smooth to completely rough flow (Colebrook & White, 1937), as shown in equation B3.
    where ft is the theoretical friction factor, RD is the Reynolds number, De is the equivalent pipe diameter, and ks is an equivalent sand-grain roughness value, or termed as hydraulic roughness. Another useful predictive tool is the Moody diagram which is based on extensive experiments on commercial rough pipes (ks/De≤ 5%; Moody, 1944).

    The key for applying equations B2 and B3 to rough conduits is to define a proper hydraulic roughness height ks. However, for the roughness fields in subglacial conduits, the relative roughness is likely to be higher than 5% (Mankoff et al., 2017). In addition, the roughness distribution is very different from those for the derivation of empirical formulas and the Moody diagram. For a subglacial conduit, the roughness is not even well defined. Using the detailed CFD results, we found that the applicability of the empirical formulas and the Moody diagram is limited to estimate the friction contribution to small-scale roughness and sinuosity. They are not applicable to subglacial conduits with large cross-sectional variations.

    Appendix C: One-Dimensional Subglacial Conduit Model

    Without considering the ice melting due to ice sheet sliding and the temperature effect and assuming a circular or semicircular cross-sectional conduit shape, the one-dimensional mass balance for the ice and water and the momentum equation for water are listed in equations C1-C3 (Spring & Hutter, 1981).
    where S, U, and Pw are time-dependent (t) conduit cross-sectional area, water velocity, and water pressure, respectively. Mf and Mp denote the ice melting rate due to viscous friction and the due to viscous creep, respectively. They are defined as follows:

    The ice melting rate Mf is determined by an energy balance between the latent heat of fusion of ice and the heat generated by viscous friction. The friction heat generation is proportional to the wall shear stress urn:x-wiley:grl:media:grl58127:grl58127-math-0050. Here ρw (=1,000 kg/m3), ρi (=917 kg/m3), and L (=3.34 × 105 J/kg) are the water density, ice density, and the latent heat of fusion per unit mass of ice, respectively. a0 = urn:x-wiley:grl:media:grl58127:grl58127-math-0051 is a constant for circular shape. K0 has a value of 4.5 × 10−25 Pa−3 · s−1 in this work. Pi is the ice pressure calculated as Pi=ρigHi with g = 9.81 m2/s. Hi is the ice sheet thickness. The difference between ice pressure and water pressure is commonly termed as effective pressure N = PiPw. θ is the conduit slope angle. Dimensional analysis has shown that the second term in the right-hand side of equation C2 is less than 5% of the other terms (Fowler, 2009), and thus, we dropped this term for simplicity and stability of our numerical algorithms, while keeping all other terms to capture the necessary physics.

    The conduit system is similar to a pipe flow which is driven by the pressure difference between the two ends. Thus, we use a time-varying pressure boundary condition at the conduit entrance (x = 0) and the conduit outlet (x = l). We assume a zero effective pressure at outlet, and thus, the fixed water pressure at the outlet is urn:x-wiley:grl:media:grl58127:grl58127-math-0052. The pressure at the entrance is determined by a lake conduit or moulin conduit system where the entrance pressure is calculated as urn:x-wiley:grl:media:grl58127:grl58127-math-0053 where V0 denotes the volume at the beginning of each time step, AL is the lake area (assumed constant), and Q0 is the upstream discharge to the lake or moulin conduit system. Δt is the time step size, and Qout is calculated as the product of the conduit cross-sectional area and velocity at the entrance. The boundary conditions for conduit size and water velocity are set as zero gradient at the conduit entrance and outlet. Initially, the water pressure at the entrance is set as Pw0=ρwgH0 with H0=V0/A0. We set a fixed area A0 = 50 m2 and 1×104 m2 to represent typical moulin size and lake size, respectively. To investigate the effects of ice thickness, we consider ice thickness Hi=1,000 m and initial entrance water depth H0=900 m. The measured diurnal upstream discharge to moulins and lakes in Greenland Ice Sheet is in the range 0–10 m3/s (Palmer et al., 2015; Smith et al., 2017); we choose an input discharge, Q0 = 3 m3/s, in this range (Swift et al., 2005). We set the initial water pressure as a linear distribution with Pw0 and Pi at the two boundaries. The initial condition for cross-sectional area and velocity is uniformly distributed with the value of 1.41 m2 and 1 m/s, respectively. This initial conduit size for the smaller friction factor 0.1 is too big so that the conduit quickly adjusted to a proper size of 0.55 m2 (see Figure 3d) after a short time (t<0.3 day; see Figures 3b and 3d).

    With these boundary and initial conditions, we first solve the mass and momentum conservation equations (equations C2 and C3) with the known value of cross-sectional area from the previous time step. The solution of the two coupled equations for velocity and pressure is through the PISO scheme in CFD (Issa, 1985). Then we update the mass balance equation for ice (equation C1). We solve the above equations with the finite volume method, and the time derivatives are discretized with a second-order Crank-Nicolson method. The spatial derivatives for pressure and cross-sectional area are discretized with the second-order central differencing scheme (termed as Gauss linear scheme in OpenFOAM), while the advection term of flow velocity is discretized by a linearized upwind scheme. The initial time step is set as Δt = 0.001 s but allowed to automatically adjust up to Δt = 1 s. The grid resolution is set as Δx = 50 m. A solver named conduitFoam for solving the governing equations with these boundary/initial conditions has been implemented in the open source platform OpenFOAM (CFD Direct, 2017).