Three-dimensional Simulations of Geometrically Complex Subduction with Large Viscosity Variations ∗
Margarete A. Jadamec
Department of Geological Sciences Brown University 324 Brook Street, Box 1846 Providence, RI 02912
Magali I. Billen
Department of Geology University of California, Davis One Shields Avenue Davis, CA 95616
[email protected] [email protected] ABSTRACT The incorporation of geologic realism into numerical models of subduction is becoming increasingly necessary as observational and experimental constraints indicate plate boundaries are inherently three-dimensional (3D) in nature and contain large viscosity variations. However, large viscosity variations occurring over short distances pose a challenge for computational codes, and models with complex 3D geometries require substantially greater numbers of elements, increasing the computational demands. We modified a community mantle convection code, CitcomCU, to model realistic subduction zones that use an arbitrarily shaped 3D plate boundary interface and incorporate the effects of a strainrate dependent viscosity based on an experimentally derived flow law for olivine aggregates. Tests of this implementation on 3D models with a simple subduction zone geometry indicate that limiting the overall viscosity range in the model, as well as limiting the viscosity jump across an element, improves model runtime and convergence behavior, consistent with what has been shown previously. In addition, the choice of method and averaging scheme used to transfer the viscosity structure to the different levels in the multigrid solver can significantly improve model performance. These optimizations can improve model runtime by over 20%. 3D models of a subduction zone with a complex plate boundary geometry were then constructed, containing over 100 million finite element nodes with a local resolution of up to 2.35 km, and run on XSEDE. These complex 3D models, representative of the Alaska subduction zone-transform plate boundary, contain viscosity variations of up to seven orders of magnitude. The optimizations in solver parameters determined from the simple 3D models of subduction applied to the much larger and more complex models of an actual subduction zone im∗Corresponding author.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. XSEDE12 Chicago, IL Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$15.00.
Oliver Kreylos
Institute for Data Analysis and Visualization University of California, Davis 2306 Academic Surge Building Davis, CA 95616
[email protected]
proved model convergence behavior and reduced runtimes by on the order of 25%. One scientific result from 3D models of Alaska is that a laterally variable mantle viscosity emerges in the mantle as a consequence of variations in the flow field, with localized velocities of greater than 80 cm/yr occurring close to the subduction zone where the negative buoyancy of the slab drives the flow. These results are a significant departure from the paradigm of two-dimensional (2D) models of subduction where the slab velocity is often fixed to surface plate motion. While the solver parameter optimization can improve model performance, the results also demonstrate the need for new solvers to keep pace with the demands for increasingly complex numerical simulations in mantle convection.
Categories and Subject Descriptors J.2 [Physical Sciences and Engineering]: Earth and atmospheric sciences; I.6.5 [Simulation and Modeling]: Model Development—Modeling methodologies
General Terms Experimentation, performance
Keywords ACM proceedings, subduction, mantle convection, rheology, multigrid
1.
INTRODUCTION
The rheology of the earth contains large viscosity variations occurring over short distances, and this poses a challenge for computational codes [25, 24, 30, 22, 10, 6, 16, 9]. For example, experimental studies and observations of seismic anisotropy suggest that dislocation creep is the dominant deformation mechanism in the upper mantle [14, 21, 16]. The non-linear relationship between stress and strainrate for deformation by dislocation creep leads to a faster rate of deformation (higher strain-rates) for a given stress, which can be expressed as a reduction in the effective viscosity (i.e., ηeff = σ/2ε). ˙ This suggests that where strain rates are high in the mantle wedge, the region between the overriding plate and the subducting plate sinking into the mantle, the viscosity can be reduced to as low as 1017 Pa s. Therefore, in subduction zones, large viscosity variations
can occur between the relatively stronger lithosphere and weaker mantle wedge, with a transition from 1024 Pa s in the lithosphere to 1017 Pa s in the center of the mantle wedge occurring over distances of less than 100 km [14, 16]. In addition, at subduction zones, a weak interface is required to allow the adjacent higher viscosity plates to slide past one another. Numerical models of subduction zone dynamics that incorporate both a subducting plate and an overriding plate typically treat the plate interface as either a discrete fault [18, 34] or as a localized low viscosity zone [12, 19, 3, 16]. Depending on the coupling between the plates, the rheological contrasts across this interface can be significant. Previous numerical models have shown that viscosity contrasts on the order of 103 to 105 are needed to generate plate-like behavior [33]. Contributing to the numerical challenge is the transition from modeling plate boundaries in two dimensions and with simple geometries to modeling subduction in three dimensions and with realistic complex geometries [23, 2, 16, 29, 17]. For example, depending on the shape of the plate interface, 3D models of the fault or weak zone can have a complex geometry with a variable dip that changes rapidly along the length of the plate boundary, and the large viscosity contrasts are then expected to occur throughout significant portions of the model domain. This can lead to regional models with well over 100 million finite element nodes to obtain local resolutions of on the order of 3 km in the critical portions of the subduction zone [16, 29]. Constraints from 2D numerical models of mantle convection benchmarked against analytical solutions provide guidelines for choices of model parameters and acceptable ranges in viscosity contrasts in 3D models of subduction [25]. However, in regional 3D models of realistic subduction zones, that contain both an overriding and downgoing plate and where the geometry of the narrow low viscosity zone along the plate interface contains a complex geometry, a balance must be achieved between accurately representing the real geometry, using a tractable number of elements, and obtaining an acceptable model solution. Because decreasing element size can result in mesh sizes that require extensive computational resources, it is worthwhile to consider model parameters, besides increasing mesh resolution, that can improve model convergence. We show the results from two phases of a numerical modeling study: Phase 1 – 3D models of subduction with a simple geometry run on a local cluster and Phase 2 – 3D models of subduction with a complex geometry based on the Alaska subduction-transform plate boundary and run on XSEDE. Phase 1 models implement a method for prescribing the plate interface into a community mantle convection code and test the effects of the 3D localized low viscosity zone on model runtime and convergence behavior. These models vary parameters that have been shown to have significant effects on convergence for 2D models with large viscosity variations. The results give insights and place quantitative constraints on which sets of parameters are most appropriate for the more complex 3D models. Phase 2 models use the weak zone implementation developed in Phase 1 and demonstrate how the incorporation of a realistic plate boundary geome-
try and dislocation creep flow law, allowing for a strain-rate dependent viscosity, can lead to paradigm shifting results.
2.
METHODS
The plate boundary geometry and temperature structure for the simple 3D models of subduction are generated using MATLAB. To generate the complex 3D plate boundary geometry and temperature field for the models of the Alaska subduction-transform boundary, a C/C++ code, referred to as SlabGenerator, was written [15, 16, 17]. Because of the complexity in the Alaska subduction-transform boundary, SlabGenerator was written to take in generic shapes and therefore is easily portable to make input for other subduction zones. The initial 3D plate boundary configuration and thermal structure are then used as input to CitcomCU, a community mantle convection code developed by [25, 23, 32]. CitcomCU simulates viscous flow in the mantle given the driving forces imposed, in this case, by the thermal structure generated from SlabGenerator. We first describe the finite element code CitcomCU, and then changes that were made to CitcomCU to incorporate the experimentally based flow law for olivine viscosity and the 3D plate boundary shear zone. We then briefly describe the model set-up for the simple and complex 3D subduction models.
2.1
Community mantle convection code
The 3D viscous models of subduction were run with the thermo-chemical mantle convection code, CitcomCU, which solves the conservation of mass (Eq. 1), momentum (Eq. 2), and energy (Eq. 3) equations for the flow velocity and pressure, assuming an incompressible fluid with a high Prandlt number ∇·u=0
(1)
∇ · σ + ρo α(T − To )gδrr = 0
(2)
∂T = −u · ∇T + κ∇2 T (3) ∂t where u, σ, ρo , α, T, To , g, δ, κ are the flow velocity, stress tensor, density, coefficient of thermal expansion, temperature, reference temperature, acceleration due to gravity, Kronecker delta, and thermal diffusivity, respectively [32]. The constitutive relation is defined by σij = −P δij + ηef f ε˙ij
(4)
where σ ij , P, ηef f , and ε˙ij are the stress tensor, the pressure, the effective viscosity, and the strain-rate tensor, respectively. CitcomCU is based on the cartesian finite-element code, CITCOM, which utilizes an Uzawa iteration scheme to solve the discretized momentum equation and uses the multigrid method (MG) with Gauss-Seidel relaxation [25, 23]. The discretized conservation of mass equation is used as a penalty constraint on the pressure. CitcomCU implements the full multigrid method (FMG) to reduce the velocity residual and accelerate convergence [32]. Finite element models of 2D mantle convection, run during the initial benchmarking of CITCOM indicate the numerical solvers are sensitive to large viscosity contrasts [25, 24]. 2D models with viscosity contrasts of up to 1030 across
Figure 1: Assignment of the low viscosity zone. a) Cross section through weak zone field, Awk . An Awk value of 1 corresponds to the fully weakened region. An Awk of 0 corresponds to no weakening. Reference surface is located at node 11. b) Cross section through η f in models with a sigma shaped low viscosity zone, for a range in background viscosities. For comparison, η f is also plotted for a model with a linear shaped low viscosity zone against a background viscosity of 1 ×1025 Pa·s (solid line). c) Elements spanning the viscosity contrast versus the factor of viscosity change per element in the low viscosity zone, plotted for a range in viscosity contrasts in models that use the linear weak zone field. Horizontal gray line marks fηe for models with low viscosity zone spanning 21 nodes. Horizontal red line denotes the fηe (2.37) recommended by [24]. the model domain, and a viscosity field that varies both exponentially and as a step function, find that the error increases with increasing viscosity contrast within an element and that to limit the error to 1% or less, the factor of viscosity change across an element, fηe , should be no more than 2.37 [24]. In addition, the implementation of the multigrid method accelerates convergence by more rapidly reducing the velocity residual, and convergence rates are further improved with an error scaling correction and the appropriate choice in the number of cycles in the multigrid solver [25]. The models also showed the appropriate use and choice of the penalty parameter on the incompressibility can improve model runtime by more rapid reduction of error in the pressure [24]. It is therefore expected that 3D models of realistic subduction zones with large viscosity variations across the plate interface and a strain-rate dependent viscosity will pose challenges for the numerical solvers and optimizing the solver parameters for a particular problem can save valuable compute time.
2.2
deformation for olivine aggregates [14] such that � �1 � � n 1−n dp E + Pl V n ηdf,ds = ε ˙ exp , r ACOH nR(T + Tad )
where Pl is the lithostatic pressure, R is the universal gas constant, T is non-adiabatic temperature, Tad is the adiabatic temperature (with an imposed gradient of 0.3 K/km), and A, n, d, p, COH , r, E, and V are the pre-exponential factor, stress exponent, grain size, grain size exponent, water content in H/106 Si, exponent for water term, activation energy and activation volume, respectively [14]. Where strainrates are high, the ηds deformation mechanism leads to a local reduction in the viscosity, such that both Equations 4 and 7 are satisfied. We use a stress exponent of 3.5 [14]. To incorporate the effects of plastic yielding, the stresses calculated in the model are limited by a depth-dependent yield stress, assuming a gradient of 15 MPa per km. We thus define the effective viscosity, η eff , ηeff = {ηcom , if σII < σy } ,
Viscosity from experimentally derived flowlaw for olivine aggregates
We employ a formulation for the composite viscosity, η com , in CitcomCU [16, 17] following the implementation in CitcomT [3] which assumes the total strain rate is a sum of the contributions from the diffusion and dislocation creep deformation mechanisms [13] ε˙com = ε˙df + ε˙ds
(5)
where ε˙ without the subscripts i-j refers to the second invariant of the strain-rate tensor, ε˙II . This leads to the following form of the composite viscosity, ηdf ηds ηcom = . (6) ηdf + ηds The viscosity components, ηdf and ηds , are defined assuming the experimentally-determined viscous flow law governing
(7)
=
��
σy ε˙II
�
, if σII > σy
(8) �
(9)
In CitcomCU, the effective viscosity is solved for as an additional loop that iterates until the global difference between the velocity field of consecutive solutions is less than a specified value, typically 1% [3].
2.3
Formulation of plate boundary interface
To model plate boundary systems comprised of an upper and lower plate separated by an arbitrarily shaped plate interface, we modified CitcomCU to read in a scalar weak zone field, Awk , constructed by SlabGenerator (Figure 1). Awk is a non-dimensional field defined at every mesh node where 1 ≥ Awk ≥ 0. Values of 1 indicate full weakening, whereas, values of 0 indicate no weakening (Figure 1). The
Figure 2: Set-up for 3D model of subduction zone with simple geometry. a) 2D slice through temperature field. Gaussian thermal anomaly defines vertically dipping slab. b) Distribution of nodes across plate interface defined by a linear and sigma functions. c) Vertical slice through viscosity. Inset shows plate interface. Depression in isosurface of temperature at 1300◦ C delineates vertically dipping slab. form of the transition from 1 to 0 can be constructed using any smoothly varying function. We test both a linear and sigma function, where the sigma function is defined by, S(x) = 1 − x2 (3 − 2x)
(10)
where 0 ≤ x ≤ 1, and x is a non-dimensional distance scaled by the thickness of the plate interface. Within CitcomCU, we define the non-dimensional low viscosity zone, η wk , as a function of Awk such that η wk is smoothly blended into the background viscosity, η eff , ηwk = ηo 10(1−Awk ) log10 (ηeff /ηo )
(11)
where η eff is as defined in Eq. 8 and η o is the reference viscosity equal to 1×1020 Pa s [15, 16]. The η wk value serves as an upper bound on the viscosity in the low viscosity zone and will be overwritten if the viscosity calculated by Eq. 8 is lower. Therefore, after implementing the low viscosity zone formulation, the final value of the viscosity, η f , becomes ηf = min (ηeff , ηwk ) .
2.4
(12)
Multigrid parameter tests on 3D model of subduction with simple geometry
For the simple 3D model of subduction, we test the effect of several multigrid parameters on model convergence and runtime. In the FMG, the discretized problem is defined on a series of meshes of successively increasing resolution. The coarsest mesh corresponds to the lowest level; the finest mesh corresponds to the uppermost level. An approximation to the solution is first calculated on the coarsest mesh. This initial approximation is then used as a starting point for a series of telescoping iterations, referred to as V (or W) cycles, that work sequentially upwards to solve the problem on the successively finer meshes [5, 31]. This method is commonly used to accelerate convergence of the discretized form of an elliptic partial differential equation [5, 31]). We test the effects of using the V and W cycle in the FMG as well as the number of smoothing cycles (Gauss-seidel relaxations) at
the different stages within the multilevel correction scheme. During code execution in CitcomCU, the viscosity structure, η f , is calculated on the uppermost level of the FMG. Therefore, the viscosity structure must be represented in some way on the coarser meshes that correspond to the lower levels of the FMG. It was determined for Citcom that rather than use the rheological law to calculate the viscosity on the coarser meshes of the multigrid, it was more effective to transfer the viscosity calculated on the finest mesh down to the lower levels [25]. In CitcomCU, there are four predefined options for how to transfer the viscosity from an upper level down to a lower level. We test the effect of each of these viscosity transfer options on the convergence behavior and runtime. For all of the viscosity transfer options, before the viscosity, η f , is transferred from an upper to lower multigrid level, the viscosity is averaged over the element assuming an arithmetic mean averaging scheme. In addition, the options that project, as opposed to inject, the viscosity use an additional averaging step that uses the arithmetic mean by default. In the arithmetic mean averaging scheme, the larger order of magnitude numbers will dominate. Thus, smaller order of magnitude values within a localized low viscosity zone may not be preserved. In order to improve convergence behavior, it may be worthwhile to preserve the smaller viscosity values of the low viscosity zone on the coarser meshes. Therefore, we test the effect of the geometric mean, which weights the small and large magnitudes evenly, and the harmonic mean, which weights smaller magnitudes more.
2.5
Set-up for 3D model of subduction zone with simple geometry
To test the sensitivity of model runtime and convergence behavior to the strain-rate dependent viscosity and the large viscosity variations across a plate boundary interface, we ran a series of numerical experiments of a vertically dipping slab with vertically dipping plate interface on one side (Figure 2). The model domain spans 44◦ x 44◦ x 2867 km, in longitude, latitude, and depth, respectively (Figure 2), with a lateral resolution of 0.1 to 1.0◦ and a depth resolution of 3 to 32
Figure 3: Set-up for 3D model of subduction zone with complex geometry based on Alaska plate boundary. a) Region of earth’s surface represented by model. b) Diagram of model bounds and components. NAM – North American upper plate; PAC – Pacific plate; Aleutian slab – subducted part of Pacific plate; PBSZ – plate boundary shear zone; FQC – Fairweather-Queen Charlotte transform fault; SMSZ – southern mesh boundary shear zone; and JdFR – Juan de Fuca Ridge. km. The corresponding grid is 161 × 161 × 129 elements. The vertically dipping slab is defined by a negative thermal anomaly with a Gaussian distribution (Figures 2a and 2c). The thermal anomaly extends from 21◦ E to 23.5◦ E, 8◦ S to 8◦ N, and from 0 km to 600 km depth. The models are instantaneous; the temperature field is not advected. The plate boundary interface is located on the west side of the slab and extends to a depth of 150 km (Figure 2). Reflecting boundary conditions are used. The effects of chemical buoyancy and melt are not included. These 3D models of a subduction zone with a simple geometry were run on 32 processors on the Beowulf Cluster, Amala, at UC Davis.
2.6
Set-up for complex 3D model of Alaska subduction zone
To test the role of rheology in decoupling the mantle from the surface plates, as indicated by seismic anisotropy, we constructed 3D regional models of the subduction-transform plate boundary system in southern Alaska (Figure 3). By constructing a regional model based on an actual plate boundary, we can directly compare model predictions to geophysical observables from that region, and thus constrain the flow dynamics predicted by the models [16, 17]. The 3D regional models of the subduction-transform plate boundary system in southern Alaska contain an overriding plate (the North American plate), a subducting plate (the Pacific plate), and the underlying mantle (Figure 3). The subducting plate geometry is based on Wadati-Benioff zone seismicity, seismic reflection and refraction, and seismic tomography [27, 11, 28, 8]. The temperatures for the subducting and overriding plates are based on geologic and geophysical observables, thereby capturing the regional variability specific to this particular plate boundary system, as in [26, 20, 4, 16, 17]. The driving forces in the system are the negative thermal buoyancy of the subducting slab (slab pull) and the positive thermal buoyancy of the Juan de Fuca ridge (ridge push) (Figure 3). The resisting forces are the viscous stresses in the mantle, the plate boundary shear zone, and
within the interior of the slab. There are no driving velocities applied anywhere in the model, rather the models predict flow velocities for the plates and mantle. The 3D regional models are run on XSEDE, using 360 processors on the cluster Lonestar, at the Texas Advanced Computing Center, for approximately 48 hours (17,000 SUs) per job for models with the composite viscosity. The 3D regional model domain spans from 185◦ to 240◦ longitude, 45◦ N to 72◦ N latitude, and 0 to 1500 km in depth (Figure 3). The finite-element mesh varies in resolution from 0.04◦ to 0.255◦ in the longitudinal direction, 0.0211◦ to 0.18◦ in the latitudinal direction, and 2.35 km to 25 km in the radial direction, with the highest resolution centered on the plate boundary in south central Alaska. The mesh contains 960 x 648 x 160 elements in the longitudinal, latitudinal, and radial directions, respectively, giving over 100 million finite element nodes. Reflecting (free-slip) boundary conditions are used on all boundaries. 2D tests were also used to determine the necessary box depth and width in order to minimize boundary condition effects on the flow in the subduction zone. These are instantaneous flow simulations designed to explore the present-day balance of forces, lithosphere and mantle structure.
3. 3.1
RESULTS 3D models of subduction zone with simple geometry
Figure 4a shows the runtimes for the fifty-two 3D models with the simple geometry. The models are geo-referenced with local mesh refinement corresponding to regions of tectonic complexity, thus we present the results in the form of raw runtime rather than weak scaling. The results demonstrate that optimizing the multigrid parameters for a particular problem can enable convergence and can significantly reduce model runtime (Figure 4). In general, decreasing the factor of viscosity change per element, fηe reduces model runtime, but for a constant fηe the model runtime varies as
Figure 4: Summary of model runtimes as a function of multigrid options for 3D model of simple subduction zone. a) Model runtimes plotted as a function of the factor of viscosity change per element in the plate interface, shown for all 52 models using the simple 3D subduction set-up. b) Comparison for selected models of variations in V versus W cycle, projection option (0-3), and maximum viscosity (non-dimensional). M2 refers to down heavy = 9, up heavy = 3, vlowstep = 100, vhighstep = 15, versus the default values of 9, 9, 400, and 9, respectively. c) Runtimes for selected models that vary the viscosity transfer option and viscosity averaging schemes. Option 0 – a node-to-node injection scheme; Option 1 – a node-to-node projection scheme; Option 2 – an element-to-element injection scheme; and Option 3 – an element-to-element projection scheme. a function of parameters related to the multigrid solver (Figure 4a). We find that limiting the maximum viscosity contrast in the model to 104 instead of 105 has the biggest effect on runtime, without changing the model dynamics (Figure 4b). Keeping the viscosity jump across on element small, < 10, is the next most important factor in insuring convergence and reducing the runtime (Figure 4a). Note this is less restrictive then the 2.37 value required to achieve 1% accuracy as found by [24] for models of 2D convection, and will likely lead to larger errors in the 3D models. For the multigrid solver, a V-cycle is usually sufficient to reach convergence if the viscosity field is sufficiently smooth. However, if a model is still not converging, then a W-cycle may help (Figure 4b). Previous studies found that the Wcycle yields faster runtimes in models with large viscosity contrasts but that after the residual is reduced to < 1%, the rate of convergence is similar for V- and W-cycles [25]. Lastly, choosing a viscosity projection option that preserves a smooth representation of the low viscosity zone at the coarsest level, i.e., a node-to-node transfer option, usually leads to the best convergence behavior (Figure 4c). In general, a viscosity averaging scheme that limits the magnitude of the viscosity jumps on the coarser levels should be used (geometric or arithmetic mean) (Figure 4c). In some cases, choosing an option that completely smooths over the low viscosity zone on the lower-most multigrid levels may also be successful, but is less predictable and will depend on the number of nodes spanned by the low viscosity zone as well as the number of multigrid levels. Prior to the first iterative step on the composite rheology, solutions for these models were also calculated using the Newtonian rheology only. All of the solutions using only the Newtonian rheology were less sensitive to the parameters varied and converged in less than 2 hrs. The strain-rate de-
pendent rheology results in higher viscosity contrasts than in the Newtonian-only models, and thus in a more challenging problem for the numerical solvers. However, as shown in Figure 4 the proper choice of solver related parameters can improve runtimes by over 20% for the models with the more complicated viscosity structure.
3.2
3D models of Alaska subduction zone
After isolating the optimal multigrid parameters for the 3D model of subduction with a simple geometry, we applied these parameters to the very large 3D model of the southern Alaska subduction zone-transform plate boundary system run on XSEDE [16, 17]. Compared to our original attempt to run this model, we found a greater than 25% speed-up, which is similar to that found for the simpler 3D models. The scientific results from the 3D models of the Alaska subduction-transform system demonstrate the importance of including a realistic plate boundary structure and rheologic flow law [16, 17] (Figure 5). For models using the composite viscosity, a laterally variable mantle viscosity emerges in the mantle as a consequence of the lateral variations in the mantle flow field; this does not occur in models that use a Newtonian only viscosity. Spatially variable mantle velocity magnitudes are predicted, with localized velocities of greater than 80 cm/yr occurring in this emergent low viscosity region close to the slab where the negative buoyancy of the slab is driving the flow [16, 17] (Figure 5b). The same models produce surface plate motions of less than 10 cm/yr, comparable to observed plate motions (Figure 5a). These results are a significant departure from the paradigm of 2D models of subduction where the slab velocity is often fixed to surface plate motion. Separate tests on a simple slab structure with an imposed low viscosity wedge and compar-
Figure 5: Results from 3D model of Alaska plate boundary run on XSEDE [16, 17]. Predicted velocity on (a) model surface and (b) at 100 km depth. (c) Azimuths from calculated infinite strain axis (ISA). Fast seismic SKS directions (blue) from [7]. Viscosity isosurface of subducting plate with (d) oblique velocity slice, (e) radial slice through velocity, and (c) cross sections through velocity. Upper plate not shown in b-f. ison to analytic solutions verify that such fast velocities are expected as a consequence of the decrease in viscosity caused by the non-Newtonian rheology [1]. Thus, these results show that a power law rheology, i.e., one that includes the effects of the dislocation creep deformation mechanism, can explain both observations of seismic anisotropy and the decoupling of mantle flow from surface plate motion [16, 17].
4.
CONCLUSIONS
We modified the community mantle convection code, CitcomCU, to model realistic subduction zones that use an arbitrarily shaped 3D plate boundary interface and incorporate the effects of a strain-rate dependent viscosity based on an experimentally derived flow law for olivine aggregates. To test the effects of the 3D plate boundary interface on model runtime and convergence behavior, we ran a series of numerical experiments on a simple 3D model of subduction. We varied parameters that have been shown to have significant effects on convergence for 2D models with large viscosity variations. The results give insights and place quantitative constraints on which sets of parameters are most appropriate for more complicated models. The 3D models of a complex subduction system, using the Alaska plate boundary as a case study and run on XSEDE, demonstrate how modeling specific plate boundary systems allows for direct comparison with observational constraints and can lead to new discover-
ies in plate tectonic processes. The results also demonstrate the need for improved solvers to keep pace with the demands for increasingly complex numerical simulations.
5.
ACKNOWLEDGMENTS
We thank the Computational Infrastructure for Geodynamics for the CitcomCU source code and the UC Davis KeckCAVES for the 3D visualization. High resolution models were run on the XSEDE site, Lonestar, at the Texas Advanced Computing Center through TG-EAR080015N. This work was supported by National Science Foundation grants EAR-0537995 and EAR-1049545.
6.
REFERENCES
[1] M. Billen and M. A. Jadamec. Origin of localized fast mantle flow velocity in numerical models of subduction. Geochemistry Geophysics Geosystems, 13:Q01016, 2012. [2] M. I. Billen, M. Gurnis, and M. Simons. Multiscale dynamics of the Tonga-Kermadec subduction zone. Geophysical Journal International, 153:359–388, 2003. [3] M. I. Billen and G. Hirth. Rheological controls on slab dynamics. Geochemistry Geophysics Geosystems, 8(Q08012), 2007. [4] D. D. Blackwell and M. Richards. Geothermal map of
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12] [13]
[14]
[15]
[16]
[17]
[18]
[19]
North America. scale 1:6,500,000, American Association of Petroleum Geologists, 2004. W. Briggs, V. E. Henson, and S. McCormick. A Multigrid Tutorial. Society for Industrial and Applied Mathematics, 2000. C. Burstedde, O. Ghattas, G. Stadler, T. Tu, and L. Wilcox. Parallel scalable adjoint-based adaptive solution of variable-viscosity stokes flow problems. Computer Methods in Applied Mechanics and Engineering, 198(21-26):1691–1700, 2009. D. H. Christensen and G. A. Abers. Seismic anisotropy under central Alaska from SKS splitting observations. Journal of Geophysical Research, 115(B04315), 2010. D. Eberhart-Phillips, D. H. Christensen, T. M. Brocher, R. Hansen, N. A. Ruppert, P. J. Haeussler, and G. A. Abers. Imaging the transition from Aleutian subduction to Yakutat collision in central Alaska, with local earthquakes and active source data. Journal of Geophysical Research, 111(B11303), 2006. M. Furuichi, D. May, and P. Tackley. Development of a stokes flow solver robust to large viscosity jumps using a schur complement approach with mixed precision arithmetic. Journal of Computational Physics, 2011. T. Geenen, M. ur Rehman, S. MacLachlan, G. Segal, C. Vuik, A. van den Berg, and W. Spakman. Scalable robust solvers for unstructured fe geodynamic modeling applications: Solving the stokes equation for models with large localized viscosity contrasts. Geoch. Geoph. Geosyst, 10:Q09002, 2009. O. Gudmundsson and M. Sambridge. A regionalized upper mantle (RUM) seismic model. Journal of Geophysical Research, 103(B4):7121–7136, 1998. M. Gurnis and B. H. Hager. Controls of the structure of subducted slabs. Nature, 335(22):317–321, 1988. C. E. Hall and E. M. Parmentier. Influence of grain size evolution on convective instability. Geochemistry Geophysics Geosystems, 4(3), March 2003. G. Hirth and D. Kohlstedt. Rheology of the upper mantle and the mantle wedge: A view from the experimentalists. In J. Eiler, editor, Inside the Subduction Factory, volume 138 of Geophysical Monograph, pages 83–105. American Geophysical Union, Washington, D. C., 2003. M. A. Jadamec. Three-dimensional lithosphere and mantle dynamics: Models of the subduction-transform plate boundary system in southern Alaska. PhD thesis, University of California, Davis, March 2009. M. A. Jadamec and M. I. Billen. Reconciling surface plate motions and rapid three-dimensional flow around a slab edge. Nature, 465:338–342, 2010. M. A. Jadamec and M. I. Billen. The role of rheology and slab shape on rapid mantle flow: Three-dimensional numerical models of the Alaska slab edge. Journal of Geophysical Research, 117(B02304), 2012. C. Kincaid and I. S. Sacks. Thermal and dynamical evolution of the upper mantle in subduction zones. Journal of Geophysical Research, 102(B6):12295–12315, 1997. M. Kukacka and C. Matyska. Influence of the zone of weakness on dip angle and shear heating of subducted slabs. Physics of the Earth and Planetary Interiors,
141:243–252, 2004. [20] T. J. Lewis, R. D. Hyndman, and P. Fluck. Heat flow, heat generation, and crustal temperatures in the northern Canandian Cordillera: Thermal controls of tectonics. Journal of Geophysical Research, 108(B6):2316, 2003. [21] M. D. Long and P. G. Silver. The subduction zone flow field from seismic anisotropy: A global view. Science, 319:315–318, 2008. [22] D. May and L. Moresi. Preconditioned iterative methods for stokes flow problems arising in computational geodynamics. Physics of the Earth and Planetary Interiors, 171(1-4):33–47, 2008. [23] L. Moresi and M. Gurnis. Constraints on the lateral strength of slabs form three-dimensional dynamic flow models. Earth and Planetary Science Letters, 138:15–28, 1996. [24] L. Moresi, S. Zhong, and M. Gurnis. The accuracy of finite element solutions of Stoke’s flow with strongly varying viscosity. Physics of the Earth and Planetary Interiors, 97:83–94, 1996. [25] L. N. Moresi and V. S. Solomatov. Numerical investigation of 2D convection with extremely large viscosity variations. Physics of Fluids, 7(9):2154–2162, 1995. [26] R. D. Muller, W. R. Roest, J. Y. Royer, L. M. Gahagan, and J. G. Sclater. Digital isochrons of the world’s ocean floor. Journal of Geophysical Research, 102:3211–3214, 1997. [27] R. A. Page, C. D. Stephens, and J. C. Lahr. Seismicity of the Wrangell and Aleutian Wadati-Benioff zones and the North American plate along the Trans-Alaska Crustal Transect, Chugach Mountains and Copper River Basin, southern Alaska. Journal of Geophysical Research, 94(B11):16059–16082, 1989. [28] N. A. Ratchkovski and R. A. Hansen. New evidence for segmentation of the Alaska subduction zone. Bulletin of the Seismological Society of America, 92(5):1754–1765, 2002. [29] G. Stadler, M. Gurnis, C. Burstedde, L. C. Wilcox, L. Alisic, and O. Ghattas. The dynamics of plate tectonics and mantle flow: From local to global scales. Science, 329:1033–1038, 2010. [30] P. Tackley. Effects of strongly variable viscosity on three-dimensional compressible convection in planetary mantles. Journal of geophysical research, 101(B2):3311–3332, 1996. [31] P. Wesseling. An Introduction to Multigrid Methods. R.T. Edwards, 2004. [32] S. Zhong. Constraints on thermochemical convection of the mantle from plume heat flux, plume excess temperature, and upper mantle temperature. Journal of Geophysical Research, 111(B04409), 2006. [33] S. Zhong and M. Gurnis. Interaction of weak faults and non-Newtonian rheology produces plate tectonics in a 3D model of mantle flow. Nature, 383:245–247, 1996. [34] S. Zhong, M. Gurnis, and L. Moresi. Role of faults, nonlinear–rheology, and viscosity structure in generating plates from instantaneous mantle flow models. Journal of Geophysical Research, 103(B7):15255–15268, 1998.
