Paleoclimate proxies are being used in conjunction with ice sheet modeling
experiments to determine how the Greenland ice sheet responded to past
changes, particularly during the last deglaciation. Although these
comparisons have been a critical component in our understanding of the
Greenland ice sheet sensitivity to past warming, they often rely on modeling
experiments that favor minimizing computational expense over increased model
physics. Over Paleoclimate timescales, simulating the thermal structure of
the ice sheet has large implications on the modeled ice viscosity, which can
feedback onto the basal sliding and ice flow. To accurately capture the
thermal field, models often require a high number of vertical layers. This is
not the case for the stress balance computation, however, where a high
vertical resolution is not necessary. Consequently, since stress balance and
thermal equations are generally performed on the same mesh, more time is
spent on the stress balance computation than is otherwise necessary. For
these reasons, running a higher-order ice sheet model (e.g., Blatter-Pattyn)
over timescales equivalent to the paleoclimate record has not been possible
without incurring a large computational expense. To mitigate this issue, we
propose a method that can be implemented within ice sheet models, whereby the
vertical interpolation along the
Although the future trajectory of the Greenland ice sheet (GrIS) trends toward continued mass loss under elevated surface temperature into the future, the speed and magnitude of these changes remain unknown (Church et al., 2013). To provide clues as to how past surface forcings influenced change over the GrIS, researchers have often relied on the paleoclimate record to serve as an analog for potential future changes (Alley et al., 2010). These records allow scientists to gain crucial insights into the evolution of the ice sheet during different climatic settings and are often corroborated by multiple lines of proxy evidence highlighting ice sheet change (e.g., ice core records, marine sediment records, terrestrial records). With respect to the GrIS, a wealth of data has been produced highlighting these changes since the beginning of the Holocene (e.g., Alley et al., 2010; Briner et al., 2016). These datasets have the potential to provide invaluable constraints for ice sheet modeling efforts aimed at exploring the sensitivity of the GrIS to past climate changes. For example, using relative sea level records throughout Greenland, Tarasov and Peltier (2002) were able to constrain an ice sheet model of the GrIS over the last deglaciation. This approach was improved through increased data coverage during later studies (Simpson et al., 2009; Lecavalier et al., 2014), highlighting the practical usage of paleoclimate proxies in ice sheet modeling efforts. Recently, ice sheet modeling results of the last deglaciation and Holocene have been compared with terrestrial records that capture changes in the ice sheet margin position (Larsen et al., 2015; Young and Briner, 2015; Sinclair et al., 2016). Because these comparisons are still relatively nascent, large model–data discrepancies do exist in some locations between the modeled margin and the margin derived from the proxy evidence, particularly in areas along the ice sheet margin where fast flow dominates. Some reasons for the model–data discrepancies include the use of a relatively coarse (10 km or greater) grid and use of the shallow-ice approximation (SIA; Hutter, 1983; Sinclair et al., 2016). Because the SIA was mainly developed for modeling the interior flow of ice sheets where the ice flow is dominated by vertical shear, it ignores membrane stresses (longitudinal and lateral drag) that are predominant closer to the GrIS margin (Hutter, 1983) and can lead to large thickness errors in these regions (Bueler et al., 2005). Both of these limitations have the impact of restricting how well an ice sheet model can simulate the behavior of an ice sheet near the margin, which is where the majority of paleoclimate evidence exists (Kirchner et al., 2011; Seddik et al., 2012, 2017).
Nevertheless, to improve the simulation speed needed for long paleoclimate
spin-ups, ice flow models of reduced complexity often utilizing the SIA with
a horizontal resolution of 10 km or greater are used to decrease
computational cost, ultimately allowing for more efficient modeling over
time intervals equivalent to a glacial cycle (
With model–data comparisons of past ice sheet changes becoming more common, however, some applications may benefit from using an ice sheet model of increased complexity, particularly when comparisons of past margin behavior are of interest. Ideally, full Stokes (FS) models provide a comprehensive 3-D solution to the diagnostic. FS models, however, are prohibitively expensive computationally and are mainly relegated to modeling experiments of no more than a few hundred years. As described above, SIA models represent the highest degree of simplification of the full Stokes equations, in which the vertical shear stress is the only non-zero stress component in the force balance equations. Although advantageous due to its computational efficiency, SIA models cannot simulate ice streams, grounding line dynamics, and floating ice shelves. On the contrary, shelfy-stream or shallow-shelf approximation models (SSA; MacAyeal, 1989) were developed to be implemented in ice shelf regions where longitudinal stresses dominate. However, these models cannot represent slow flow in the interior of the ice sheet where vertical shear is non-negligible. Higher-order models (Blatter, 1995; Pattyn, 2003; herein referred to as BP for Blatter-Pattyn), on the other hand, that include membrane stresses and elements of the vertical shear stress have been a hallmark in the ice sheet modeling community over the past decade, being favored for their ability to model both the fast and slow areas of ice flow while being computationally cheaper than full Stokes models. The majority of the computational demand for an ice sheet model resides within the stress balance computation. Although the thermal model requires many vertical layers in order to capture sharp thermal gradients near the base of the ice, stress balance tests performed with ISSM (Ice Sheet System Model) (not shown here) on models with 25 layers and 5 layers show the area-averaged differences in the surface and basal velocities to be 0.22 and 0.012 %, respectively. Therefore, for the purposes of the experiments outlined in this study, we consider that the stress balance computation does not require a high vertical resolution. As a consequence of the high number of vertical layers needed for the thermal computation, however, more runtime is needed during the stress balance computation than is necessary. Because of the increased model complexity in BP models they have therefore not been run over paleoclimate timescales due to the large computational expenses needed to complete the runs. To utilize BP models in paleoclimate simulations, methods to improve runtime speed without sacrificing the models precision need to be addressed.
Here we present a method which builds upon the thermomechanical ice flow model ISSM to improve model speed within the BP ice sheet model simulations. While our implementation and analysis are done with ISSM, the methods can be applied to a wide range of finite-element ice sheet models. The main component of this development focuses on the vertical extrusion of layers within ISSM and the type of finite elements used to create the vertical interpolation. The aim of this method is to allow the user to perform model simulations that have a smaller number of vertical layers than typically used, while still being able to more precisely capture the thermal state of the ice sheet than would otherwise be captured using traditional means of linear vertical interpolation. We begin by first describing the methodology associated with the implementation of higher-order vertical elements in Sect. 2, followed by a description of the model experiment setup for an idealized single-dome ice sheet and a realistic GrIS configuration in Sect. 3. The results are accompanied by a discussion in Sect. 4 and conclusions in Sect. 5.
Panel
Like many finite-element ice sheet models, ISSM relies on prismatic
elements, which are the result of a vertical extrusion of a two-dimensional
triangular mesh. The interpolation used in these elements is decomposed into
a horizontal interpolation and a vertical interpolation. A P2
Since the nodal functions are taken as a product of horizontal and vertical
polynomials, they can be written in the following terms:
In the vertical direction, we use a reference element that goes from
For a quadratic finite element in the vertical direction (herein noted as
P2), we have nine nodes per element (Fig. 1): one per vertex and one in the
center of each vertical segment. We have the following functions in the
vertical direction:
For a cubic finite element in the vertical direction (herein noted as P3),
one needs 12 nodes per element (Fig. 1): 1 per vertex and 2 located at one-third and two-thirds of each vertical segment. The vertical components of
the nodal functions are
Panel (
Increasing the degree of finite elements along the
For the following model experiments we use the ISSM (Larour et al., 2012), a finite-element, thermomechanical ice sheet model. The tests performed in this study can be split into two experiments. We first test the precision of the higher-order vertical interpolation using a simplified single-dome ice sheet experiment that uses the SIA, following Experiment A of the European Ice Sheet Modeling INiTiative (EISMINT 2) experiments (Payne et al., 2000). We then apply a similar setup to a GrIS-wide model, where the steady-state thermal solution is computed using the higher-order BP model. Specifics regarding model setup and the relevant experiments are discussed below.
To test the performance of the higher-order vertical interpolation, we adopt
a setup similar to the EISMINT 2 experiments (Payne et al., 2000), which were
targeted for the assessment of thermomechanical shallow-ice models. We
perform all of our single-ice-dome experiments using the SIA on models with
a horizontal grid resolution of 20 km
Rather than perform all of the experiments associated with the EISMINT 2 benchmarks, we choose to limit the analysis to only Experiment A, where models begin from the same initial state. Experiments begin with zero ice over a bed with flat topography and are run to relaxation for 100 000 years. To compare the differences between the vertical interpolations, we run 24 simulations in total. These simulations use the P1, P2, and P3 vertical interpolation for models that have a minimum of 3 nonuniform layers to a maximum of 10 nonuniform layers. Each model uses an extrusion exponent of 1.2, indicating that the layers are not equally spaced but rather modestly biased towards thinner layers at the bed. Aside from comparison of the results to EISMINT 2, we run a simulation using the P1 vertical interpolation on a model with 25 layers. This model will serve as the benchmark to compare the other simulations to, with a 25-layer P1 model being representative of what is typically used in the setup for GrIS-wide simulations in ISSM (Seroussi et al., 2013). We note that for the stress balance computation, we use the P1 vertical interpolation, while the thermal computation makes use of the higher-order vertical elements.
In addition to comparison with the EISMINT 2 Experiment A, thermal
steady-state computations are performed for a GrIS-wide model to determine
how well the vertical interpolations can capture thermal profiles and basal
temperatures throughout the ice sheet. The three-dimensional higher-order
model (i.e., BP) of Blatter (1995) and Pattyn (2003) is used for the momentum
balance equations. The nonlinear effective ice viscosity results from Glen's
flow law (Glen, 1955) and is given in Eq. (5):
The GrIS-wide model relies on anisotropic mesh adaptation, whereby the
element size is refined as a function of surface elevation (Howat et al.,
2014) and InSAR (interferometric synthetic-aperture radar) surface velocities from Rignot and Mouginot (2012), becoming
finer in areas where the second derivative of these two quantities is
higher. The model mesh has a horizontal resolution ranging from 3 km in
areas of ice streams to 20 km over the interior regions where the ice flow
is slow, corresponding to a two-dimensional model with
The models are initialized with bed topography from BedMachine Greenland v3
(Morlighem et al., 2017) and ice surface elevation from the GMIP
DEM (Greenland Mapping Project Digital Elevation Model) of
Howat et al. (2014). The surface mass balance and surface temperatures are
taken from Ettema et al. (2009), and the geothermal heat flux relies on a
setup identical to Seroussi et al. (2013). The underlying geothermal heat
flux from Shapiro and Ritzwoller (2004) is used; however, values of
20 and 60 mWm
The thermal model for both the single-dome and steady-state experiments use an enthalpy formulation derived from Aschwanden et al. (2012), which includes both temperate and cold ice. At the ice surface, air temperature is imposed, while the geothermal heat flux is applied at the base. For full details outlining the thermal model used in ISSM, we direct the reader to Seroussi et al. (2013) and Larour et al. (2012). Lastly, the spatially varying basal drag coefficient is determined using inverse methods (Morlighem et al., 2010; Larour et al., 2012), providing the best match between modeled and InSAR surface velocities from Rignot and Mouginot (2012).
The percent difference in ice volume from the 25-layer P1 model
for models using the P1
Ice volume, ice divide basal temperature, and ice divide thickness for each individual simulation after a 100 kyr relaxation. Also shown are the corresponding mean values for the EISMINT 2 (Payne et al., 2000) Experiment A simulation and the standard deviation. Those values that fall within 1 standard deviation from the EISMINT 2 Experiment A mean values are given in italics, those within 2 standard deviations in bold italics and those within 3 standard deviations in bold.
Each individual model is relaxed for 100 000 years to bring the ice sheet into steady state with respect to both the ice thickness and temperature. In Fig. 3, the ice volume for each particular simulation is shown as a percent difference from the 25-layer P1 simulation with the shading corresponding to the zone where models fall within 2 % of the ending ice volume simulated by the 25-layer P1 model. Although all models simulate the same relative trend for the ice volume relaxation, they do not all converge on the ice volume simulated by the 25-layer P1 model. For the models where the linear (P1) interpolation (Fig. 3a) is used in the thermal model, only those models with at least 8 layers fall within the 2 % range of ending ice volume for the 25-layer P1 simulation. When using a higher-order vertical interpolation (P2 and P3), however, models with four layers and above fall within the 2 % range (Fig. 3b and c).
The percent difference in simulated ice volume after the 100 000-year relaxation for the single-ice-dome experiment compared to the 25-layer P1 model. Each model run is shown as a function of the vertical interpolation and the number of layers used.
To further compare the performance of each model, the corresponding ice
volume, ice divide basal temperature, and ice divide thickness are shown in
Table 1 for each model simulation and are compared to the mean values
derived from the EISMINT 2 Experiment A results (Payne et al., 2000). It is
important to note that no known analytic solution was provided in the
EISMINT 2 Experiment A comparison. Similar to Rutt et al. (2009), however,
we compare our simulated values to the mean and the standard deviation of
the values for Experiment A in the EISMINT 2 experiments to assess the
relative spread. In general, models using the higher-order vertical
interpolation tend to better match the EISMINT 2 results. Models with four
layers or more using the P2 or P3 vertical interpolation fall within 1
standard deviation (
Models with five or more layers using the P2 or P3 vertical interpolation fall
within 2
The absolute value of the percent difference between each individual model run and the 25-layer P1 simulation at the end of the 100 000-year relaxation for ice volume, ice divide basal temperature, and ice divide thickness. Italics denote models that fall within 1 % of the variables simulated by the 25-layer P1 model at the end of the relaxation.
Because of the potential difficulties in assessing differences between our
results and those derived from the EISMINT 2 Experiment A, we also compare
our results to the model simulation using the 25-layer P1 vertical
interpolation. Because this model is representative of what is
characteristically used for three-dimensional, thermomechanical modeling in
ISSM (Seroussi et al., 2013), further comparisons can be made to those
models that agree well with simulated ice volume, ice divide basal
temperature, and ice divide thickness from the 25-layer P1 model. In Table 2,
the absolute value of the percent difference is shown between each
individual model simulation and that using the 25-layer P1 model. Following
from the comparison with the EISMINT 2 Experiment A results, the higher-order
vertical interpolation allows models with fewer layers to capture changes
simulated by the 25-layer P1 model with a higher precision. In Table 2, italics denote those model simulations where the simulated ice
volume, ice divide basal temperature, or ice divide thickness is within
1 % of the 25-layer P1 model. Generally, models with at least 4 (P3) and 5
(P2) layers capture the simulated ice volume within 1 % of that simulated
by the 25-layer P1 model. Using the linear vertical interpolation, 10 layers
are needed before simulating ice volume within 1 % of the 25-layer P1
model. This is better illustrated in Fig. 4, where the percent difference in
ice volume from the 25-layer P1 model is shown as a function of the number
of layers in each model. Those models using the P2 and P3 vertical
interpolation converge significantly faster to
The resulting temperature profiles at the ice divide after the 100 000-year single-ice-dome relaxation for models with 3, 5, and 7 layers, compared to the temperature profile from the 25-layer P1 model.
Differences between the linear vertical interpolation and the P2 or P3 interpolation become more apparent when analyzing ice temperature profiles. In Fig. 5, ice temperature profiles are plotted at the ice divide for models with three, five, and seven layers. With only three layers, models with the P1, P2, and P3 vertical interpolation simulate a temperature profile that is too warm between 500 and 1500 m and too cold approaching the base. Despite the vertical interpolation used, the profile is not well captured, although improvements to the shape of the temperature profile in the transition between 500 and 1500 m can be seen in models using the higher-order vertical interpolation. Adding more layers to each model improves the overall fit to the 25-layer P1 model, although the models using the P2 and P3 vertical interpolation capture the shape of the temperature profile much better than the linear interpolation. The overall fit is improved not only at the base but also in the transition between 500 and 1500 m where the ice begins to warm more rapidly approaching the base. We also find that the differences between the P2 and P3 vertical interpolation are marginal in this example, indicating that using a quadratic vertical interpolation (P2) is suitable when given the choice to use a cubic vertical interpolation (P3).
Although much of the success regarding the higher-order vertical interpolation resides in the model's ability to capture the vertical structure of temperature in the ice using fewer layers than is needed from the traditional linear vertical interpolation, improvements to model speed are the main motivation for its implementation, particularly in BP models. To test how model speed is improved when implementing the higher-order vertical interpolation, we begin by using the relaxed model simulations that have thus far only used the SIA for the single-dome experiments in Sect. 4.1. From the relaxed model states, each simulation is run for 100 years using the BP ice flow model in ISSM and using the same boundary conditions from the relaxation with a fixed time step of 0.2 years.
Run times for the 100-year higher-order simulation of the single ice dome for each individual model based upon the number of layers and the vertical interpolation scheme used.
Since we assume that the horizontal mesh accurately captures variations in the model fields, running a higher-order vertical interpolation reduces the number of layers used in the stress balance computation, which is the most computationally demanding part of transient simulations. Comparing the simulation time for each individual model compared to the 25-layer P1 model, all models, despite the vertical interpolation used, complete the 100-year run anywhere between 241 (3P1) and 9 (10P3) times faster (Fig. 6). To determine how models perform based upon the vertical interpolation, a criterion is established based upon Table 2, such that each model's simulated ice volume must be within 1 % of those values simulated by the 25-layer P1 model, which represents the relative uncertainty associated with the present-day ice volume of the GrIS (Morlighem et al., 2017). Based upon these criteria, models using the P1 vertical interpolation must have 10 layers or more, while models using the P2 and P3 vertical interpolation can use at least 5 or 4 layers, respectively. When applying these criteria, runtime is 5 times faster for a 5-layer P2 model versus a 10-layer P1 model. If we assume a seven-layer P1 model is adequate, the runtime for a five-layer P2 model is 2 times faster. When compared with the 25-layer P1 model, the 5-layer P2 model completes the relaxation 57 times faster.
The thermal steady-state simulation is compared with the GRIP ice core
record (Dahl-Jensen et al., 1998) in Fig. 7 for models with 3, 5, and 7
layers as well as the 25-layer model with the P1 vertical interpolation. The
simulated thermal structure for the 25-layer P1 model is similar to the
thermal profile presented in Seroussi et al. (2013). Temperature differences
of 2–5
The resulting temperature profiles for the higher-order steady-state thermal computation at the GRIP ice core site location for models with 3, 5, and 7 layers, compared to the temperature profile from the 25-layer P1 model and the measured GRIP temperature profile (Dahl-Jensen et al., 1998).
This study aims at addressing the current computational limitation in using higher-order stress balance ice sheet models for paleoclimate studies. Currently, analysis of ice sheet modeling experiments focusing on the past behavior of the GrIS is being complemented with rich paleoclimate data-constraining features of the past ice sheet behavior (Larsen et al., 2015; Young and Briner, 2015; Sinclair et al., 2016). Where shallow-ice models might be limited in their ability to simulate the marginal behavior of the GrIS through the exclusion of higher-order stress terms and an inability to run on a high-resolution mesh, BP models may become more appropriate for such comparisons in the future. To help alleviate the computational expense in using a BP model, we implement higher-order vertical elements. As shown in Sect. 4.1 of this study, increasing the degree of the vertical interpolation allows the model to capture gradients in the thermal profile of the ice with more precision than would otherwise be captured using a model with a linear vertical interpolation, despite having the same number of vertical layers. Models with correspondingly fewer layers that used the higher-order vertical interpolation were able to capture the transient behavior consistent with the EISMINT 2 Experiment A results (Payne et al., 2000) and also performed well when compared to a model similar to those that are used for modeling studies in ISSM (Seroussi et al., 2013).
The biggest attraction for using higher-order vertical elements is that they not only decrease the computational burden for the thermal model but also for the stress balance computation, due to a decrease in the number of vertical layers needed. Overall, this leads to a large reduction in computational time, particularly when a BP model is used. Models using the higher-order vertical interpolation were shown to shorten runtime anywhere between 2 and 5 times for a 5-layer model compared to models with 7 and 10 layers, respectively, using a linear vertical interpolation. When compared to the 25-layer model using the linear vertical interpolation, models with 5 to 10 layers using the higher-order vertical interpolation had anywhere between a 57 to 10 times faster runtime, with minimal impacts on the precision of the simulated ice volume and thermal state. When the higher-order vertical elements were applied to a three-dimensional, BP model of the GrIS, experiments showed the thermal state of the ice sheet can be captured as precisely as our 25-layer P1 model when at least 5 layers are used for a quadratic (P2) vertical interpolation and at least 4 layers for a cubic (P3) vertical interpolation. When comparing the quadratic and cubic vertical interpolation, the benefits of using a cubic vertical interpolation are slight, although it may be useful when modeling in areas of complex thermal regimes.
In the context of paleoclimate simulations, using a higher-order vertical interpolation improves simulation speed, particularly for BP ice sheet models. BP models using this will still likely be too computationally intensive for simulations which sample parameter space and thus require multiple independent simulations (Applegate et al., 2012; Robinson et al., 2011). However, in experiments where BP models may offer improvements in model data comparison versus using shallow-ice models, higher-order vertical elements can be used as a means to improve model speed while still being able to capture the qualities simulated in a model with many more layers but at the fraction of the speed. In this respect, future studies will use these higher-order vertical elements to enhance computational speed while maintaining mechanical complexity for ice sheet modeling experiments over various paleoclimate timescales.
The higher-order finite elements are currently implemented in the ISSM code,
which can be compiled following the instructions on the ISSM website
(
The supplement related to this article is available online at:
The authors declare that they have no conflict of interest. Edited by: Jeremy Fyke Reviewed by: two anonymous referees