1. Introduction
The pressing need to have detailed knowledge of local features of climate systems and the high costs of obtaining such knowledge has motivated research on regional climate modeling. A new global-to-regional multiresolution approach has been proposed for which only one grid, with variable resolutions and with smooth transitions between fine and coarse grid regions, is involved (Ringler et al. 2011). This approach gives rise to a new set of challenges. One challenge is how to produce suitable meshes with local resolutions (see, e.g., Ringler et al. 2008). Another challenge is related to the implementation of parameterizations on a variable-resolution grid. This work is part of a series of efforts to address the second issue.
Developing scale-aware parameterizations for the atmosphere and ocean has been a difficult and largely unmet challenge. While the closures for clouds in the atmosphere (Arakawa and Schubert 1974) and eddies in the ocean (Gent and McWilliams 1990) have clearly been successful, neither has been generalized across spatial and/or temporal scales. The long-term success of models that operate on meshes with multiple resolutions will depend on access to closure parameterizations that act appropriately across a wide range of spatial and temporal scales with little or no ad hoc tuning. Such closures, assuming that they exist, would demonstrate an understanding of the modeled physical process(es) far beyond our current capability.
Two of the most important components of the climate system, the ocean and the atmosphere, can both be considered as thin layers of fluids surrounding the earth. In other words, the large-scale motions of the ocean and atmosphere are nearly two-dimensional flows and, thus, the fundamental principles of two-dimensional turbulent flows provide guidance into the dynamics of the ocean and atmosphere. The most striking features of such turbulent flows are the inverse cascade of the kinetic energy and the cascade of the enstrophy (Batchelor 1969; Kraichnan 1967; Lilly 1971, 1969; Maltrud and Vallis 1991). For two-dimensional turbulent flows, the enstrophy plays the role that the kinetic energy does for three-dimensional turbulent flows. In adiabatic systems, such as the shallow-water equations or the primitive equations cast in isopycnal coordinates, this role is played by the potential enstrophy. The presence of the upscale (inverse) energy cascade in two-dimensional turbulence acts to move kinetic energy away from the grid scale; thus requiring very little, if any, grid-scale dissipation of kinetic energy in numerical simulations. At the same time, the downscale cascade of potential enstrophy necessitates the removal of grid-scale variance in potential vorticity in order to maintain robust simulations. The fundamentally different behavior of energy and enstrophy in two-dimensional turbulence has led to the development of energy-conserving, enstrophy-dissipating numerical schemes (see, e.g., Arakawa and Hsu 1990; Sadourny and Basdevant 1985).
As we have mentioned, we intend to study parameterizations on variable-resolution grids. Thus, we are pursuing approaches for the development of scale-aware parameterizations. As the first step, here we focus on the APVM (1)–(2) on a sequence of quasi-uniform grids with varying resolution. We ask the following questions: what is the optimal form–value of the coefficient γ and how does the optimal parameter change in response to changes in the grid resolution and time step?
Clearly, the primary challenge in the development of any scale-aware parameterization will be the identification of parameter(s) that are largely insensitive to the spatial and temporal resolution of the numerical model. So whereas the form of γ in (3) is appealing because of its clear physical interpretation, it is clearly sensitive to the choice of model time step. There is no analysis available to support the invariance of the parameter σ when using the APVM; in fact, our experience shows that the optimal value of σ is indeed influenced by the time-step size and the grid resolution of each particular simulation. In our opinion, this puts a severe constraint on the applicability of the form in (3) for γ because, for each particular simulation, without extensive fine-tuning and comparisons, it is not clear what the optimal value of σ one should use.
If we can cast the APVM in terms of a scale-insensitive parameter, then we can use basic parameter optimization techniques to find the appropriate value of this parameter through comparisons to high-resolution reference solutions. Parameter optimization is really only practical for scale-insensitive parameters.
To overcome the difficulty associated with the form in (3) for γ, we endeavor to find a form of the APVM parameterization that is a function of a single, largely scale-invariant parameter, which is denoted as α in our analysis below. Our main tools are a scale analysis and the phenomenological theories of two-dimensional turbulent flows. The nonlinear advective term plays an important role in the phenomenological theories of turbulence (see, e.g., Frisch 1995). However, the nonlinear advective term usually does not participate in the scale analysis carried out in designing subgrid eddy closure schemes (Berselli et al. 2006). The dissipation term always has a key place in scale analysis because the dissipation rate is assumed to be representative of the eddy fluxes throughout the inertial range. A nonlinear dissipation term, such as the one with APVM [see (17) as well as (A10) and (A11)] will lead to nonlinear convoluted terms in the contributions to the (potential) enstrophy dissipation at a given scale [see (A15) for the two-dimensional incompressible case]. We deal with the difficulty associated with these terms by analyzing the scale interactions and singling out the dominant term, which is then taken as an approximation to the (potential) enstrophy dissipation rate at that scale. Details are presented in section 2 for shallow-water flows and in the appendix for two-dimensional incompressible flows.
In section 3, we present the results of numerical experiments conducted to determine the optimal value of α that appears in the new form of γ. One role of those experiments is to show that the scale analysis is valid, despite the assumptions made. In fact, it is seen from the experiments that α is invariant to the size of the time step and has a very weak dependence on the grid resolution. In section 4 we provide the concluding remarks.
2. Subgrid eddy closure based on the APVM
In this section, we perform a scale analysis on the APVM applied to the shallow-water equations, and derive a new formulation that depends on a single parameter that is formally independent of the time-step size and the grid resolution and therefore is suitable for parameter optimization. The scale analysis is an extension of the analysis given in the appendix for two-dimensional incompressible turbulent flows for which the phenomenological (Batchelor 1969; Kraichnan 1967; Leith 1968) and mathematical (Bardos 1972; Constantin 2007) theories are mature. This extension is possible if we assume that the shallow-water flow is predominantly two-dimensional and variations in the fluid thickness variable h are negligible compared to the mean fluid thickness. For many realistic cases, this is a reasonable assumption.
3. Numerical experiments
a. The spectra and the basic optimization technique
Our purpose here is to evaluate the effectiveness of the scale-aware APVM closure developed above. In particular, several reasonable sounding assumptions were made in carrying out the scale analysis, so that numerical testing of the scale invariance of α is useful in verifying that result of the analysis. As with the evaluation of most closures, the measure of effectiveness is based on the ability of a low-resolution simulation with the closure to reproduce certain important aspects of a high-resolution, reference simulation. In the simulations below, we evaluate the closure based in its ability to reproduce the spectrum of the enstrophy obtained in the high-resolution simulation.
It has been conjectured (see, e.g., Batchelor 1969; Kraichnan 1967) and approximately verified in the literature (see, e.g., Lilly 1969, 1971) that for two-dimensional incompressible turbulent flows, the spectrum of the enstrophy satisfies a −1 power law. Equivalently, the spectrum of the kinetic energy satisfies a −3 power law, in contrast to the famous −5/3 power law for three-dimensional turbulent flows. Our reference solution is computed on a very fine grid, with the least damping that is still capable of producing a potential enstrophy spectrum that approximates the −1 power law well. The damping mechanism used to produce the reference solution is provided by conventional hyperviscosity that has been widely used in geophysical fluid dynamics.
b. The optimal parameters
The numerical experiments are conducted with the one-layer shallow-water equations on the whole sphere. The numerical model is based on the numerical scheme presented in Thuburn et al. (2009) and Ringler et al. (2011). This approach is being used to develop global atmosphere and ocean models that are a part of the Model for Prediction Across Scale (MPAS) project. The MPAS modeling approach is particularly attractive for this study since it solves the vector-invariant form of the momentum equation from which the APVM is derived in (4), conserves potential vorticity in analogy to the continuous system in (9), and guarantees that the APVM correction term does not create or destroy kinetic energy. While the MPAS system can be used with a wide class of meshes, for our experiments we use spherical centroidal Voronoi tessellations (SCVTs) because of their global uniformity and isotropy (Du et al. 1999). The remapping from SCVTs to regular Gaussian grids on the sphere (as input to the spherical harmonics package; Adams and Swarztrauber 1997) is done with the Spherical Coordinate Remapping and Integration Package (SCRIP; Jones 1999).
The reference solution is computed on an SCVT mesh with 655 362 cells (dx ≈ 30 km), with the traditional ∇4 dissipation of 109 m4 s−1. The shallow-water test case 5 evolves into turbulence in about 20–25 days. A snapshot of the potential vorticity field on day 50 is shown in Fig. 1. The axis of the globe is slightly tilted in order to better display the structure of the potential vorticity field on the Northern Hemisphere. We plot the spectrum of the potential enstrophy of the reference solution on day 150 in Fig. 2. An inertial range of width approximately 1 decade appears between wavenumbers 20 and 120, which approximately verifies the −1 power law for potential enstrophy spectra.
Before we present comparison and optimization results, we remark on the choice of the starting and ending wavenumbers of the subrange used in the comparisons. The subrange for comparison should be part of the inertial ranges of the reference and approximate solutions. The reason for this is that the high noise level in the scales larger than the inertial range will render the comparison results largely stochastic, and hence noninformative, whereas the spectra at scales smaller than the inertial range heavily depend on the damping mechanism used in each particular simulation, and hence cannot give reliable results. Because the approximate solutions are computed using the APVM on a coarser grid to measure the effectiveness of the closure, these simulation display a spectral range that is narrower than that of the reference solution. Their inertial range will also be narrower than that of the reference solution. This fact should also be taken into consideration when choosing the starting and ending wavenumbers of the subrange used for comparisons.
Approximate solutions are computed on a 10 242-cell (approximate resolution of 240 km) SCVT grid, with the APVM parameter α taking values from 0 to 0.01, with an increment of 0.0001. The APVM in (15) is the only closure used in these low-resolution simulations (i.e., the ∇4 dissipation is turned off). Hence, there are 101 simulations. The spectrum of the potential enstrophy for each of these simulations is compared to that of the reference solution and a distance between the spectra is calculated using (30). The approximate solutions on the 10 242-cell grid have an inertial range between wavenumbers 20 and 80. Hence, we take k0 = 20 and k1 = 80 in (30), per the discussion in the preceding paragraph. Then, we plot the spectral distance against the APVM parameter α in Fig. 3. It is seen that the minimum distance is obtained at α = 0.0013. Just to show what happens if α is too small or too large, we plot in Fig. 4 the potential enstrophy spectra of the reference solution and of the simulations with α = 0, 0.0013, 0.0080. With α = 0, the potential enstrophy is transferred down scales by the nonlinearity and then, because of the lack of dissipation, it piles up at high wavenumbers. On the other hand, with α = 0.008, the overdamping is obvious as it pulls down the tail of the spectrum significantly below that of the reference solution leading to an inertial range with a slope significantly steeper than the −1 slope.
The APVM, as it was formulated in Sadourny and Basdevant (1985) and used in the literature, is not suitable for parameter optimization because the way it is formulated, the parameter will depend on the time-step size as well as the grid resolution. The advantage of our formulation is that the parameter is invariant with regard to the time-step size, up to time truncation error. It is also formally independent of the grid resolution, though we suspect that the parameter may have a weak dependence on the grid resolution. In what follows, we demonstrate the time-step invariance aspect of our formulation and we explore whether and how the parameter depends on the spatial grid resolution.
To show the invariance of the parameter with respect to the time-step size dt, we redo the comparison study, but with dt varying over {691.2 s, 345.6 s, 172.8 s, 86.4 s, 43.2 s, 21.6 s}. For each of these step sizes, a curve is plotted in Fig. 5 showing how the distance in the spectra varies with respect to the APVM parameter α. These curves for different time-step sizes agree very well with each other. In particular, they all point to the same minimizing parameter α = 0.0013. This is not surprising, as our modified subgrid eddy closure model consists of partial differential equations that are independent of the time-step size, and hence the solutions of the model are independent of the time-step size as well.
To explore whether and how the parameter α depends on the grid resolution, we fix the time-step size at dt = 172.8 s and perform the optimization on α on 2 additional grids, one coarser with 2562 cells, and the other finer with 40 962 cells. For each of these grids (nCells = 2562, 10 242, and 40 962) we plot, in Fig. 6, a curve depicting how the spectral distance changes with respect to the APVM parameter α. The curve for the coarser grid (nCells = 2562) attains its minimum at α = 0.0010 whereas the curve for nCells = 10 242 attains its minimum at α = 0.0013, as we have already seen. The curve for the finer grid (nCells = 40 962) attains its minimum at α = 0.0018. Through this experiment, we see that the minimizing parameter α tends to increase as the grid is refined. The dependence of the minimizing parameter α is weak because, in this experiment, the grid resolution changes by a factor of 2, but the change in the minimizing α is much slower. As a matter of fact, the minimizing α is approximately proportional to nCells0.212, with the exponent 0.212 being much smaller than the exponent 3 over ℓ in (27).
There are a few other noteworthy features about Fig. 6. We see that for the same parameter α, finer grids usually produce better results (smaller spectral distance when compared to the reference solution). We also note that as the grid gets finer, the spectral distance curve becomes less sensitive to the increment in the parameter α. In fact, even though the curve for nCells = 40 962 attains its minimum at α = 0.0018, its value at α = 0.0013 is very close to the minimum value. To corroborate on this point, we compare, in Fig. 7, the spectra of the results on the finer grid (nCells = 40 962) with α = 0.0010, 0.0013, 0.0018, and 0.0050. The spectrum of the reference solution is also plotted for comparison. We see that the spectra for α = 0.0010, 0.0013, and 0.0018 stick together and are close to the spectrum of the reference solution in the inertial range. On the other hand, the spectrum for α = 0.0050 is pulled down below the reference solution due to overdamping. These results indicate that, although the minimizing value of the parameter α is weakly dependent on the grid resolution, there is some robustness present: using optimal values for the parameter determined using one grid resolution can seemingly be safely used for other grid resolutions.
4. Conclusions
The original form of the coefficient γ for the anticipated potential vorticity method, as suggested in Sadourny and Basdevant (1985), is not suitable for parameter optimization because it involves a parameter that is not invariant with respect to the time-step size and the spatial grid resolution. Using a scale analysis technique and the phenomenological theories of two-dimensional turbulence, we propose a new form for the APVM coefficient γ such that the new parameter involved is formally invariant with respect to the time-step size and spatial grid resolution.
Numerical experiments are conducted on the whole sphere with different grids, each of which is a quasi-uniform spherical centroidal Voronoi tessellation of the sphere. Two test cases have been used. One is the standard shallow-water test case 5 involving a mountain topography, and the other is a flow evolving from arbitrarily imposed initial data. Our basic optimization technique is to compare the potential enstrophy spectra of the APVM solutions to that of a reference solution that is calculated on a very fine grid with ∇4 hyperviscosity. The numerical results demonstrate the time-step invariance aspect of our formulation for γ. Over a sequence of grids having the number of cells varying from 2562 to 40 962, the optimal APVM parameter is found to be within the range 0.001 to 0.002. Because the measure function is relatively flat near the optimal value of α, this factor of 2 change in the optimal parameter is not significant.
One issue that we have not touched upon in this article is the performance of the APVM compared to the traditional hyperviscosity (iterated Δ) method. This issue was discussed in Sadourny and Basdevant (1985) for the APVM in its original formulation; it was shown that, for the same spatial grid resolution, the APVM produces more realistic results than the traditional hyperviscosity method. A comparison between the APVM in the new scale-invariant formulation and the traditional hyperviscosity method is then in order. With the standard shallow-water test case 5, the reference solution is computed on a 655 362-cell (approximate resolution of 30 km) grid with a ∇4 dissipation of 1.0 × 109 m4 s−1, which is the result of extensive fine tuning. Basic scaling arguments say that the ∇4 parameter goes like ℓ4. Thus, the optimal ∇4 parameter on a 10 242-cell grid should be 4.0 × 1012 m4 s−1. For comparison, we plot, in Fig. 10, the spectra between the spherical wavenumbers 10 and 100, of the high-resolution reference solution, the 4.0 × 1012 m4 s−1 ∇4 solution, and the APVM solution with the optimal parameter α = 0.0013. It is seen that the solution of the APVM and the solution of the ∇4 method match very well at low wavenumbers; at high wavenumbers the APVM solution has a more extensive inertial range that matches well with that of the reference solution. The spectrum curve of the ∇4 solution also appears slightly steeper than that of the reference solution in the inertial range. This test shows that the APVM in the new formulation developed in this article is at least as good as the traditional ∇4 method at producing solutions that conform to the phenomenological theories of two-dimensional turbulence.
The full three-dimensional primitive system will require some type of ∇4 closure to dissipate the downscale cascade of kinetic energy. Even in this situation, we expect that the APVM is a valuable component of the overall model closure by allowing the ∇4 parameters to be smaller, and therefore less dissipative, than they would be otherwise.
The closure presented here includes the assumption of small deviations in fluid thickness. Namely, before starting the scaling analysis in (17), we assume that the fluid thickness h that is included on both sides of (16) cancels out. For the simulations conducted above and, likely, for the broad class of shallow-water systems where this closure might be utilized, this will, in general, be a valid assumption. As we apply the closure to more realistic systems, such as the isopycnal model of the ocean circulation, this assumption will have to be carefully reevaluated.
This work is the first in a series of efforts to address the issue of subgrid eddy parameterizations on variable-resolution grids in global models. The experiments in this work are conducted on quasi-uniform grids and we address the question about how the resolution of the quasi-uniform grid affects the optimizing parameter for the APVM. Naturally, our next step is to address the issue on a variable-resolution grid. There, the question becomes: how the resolution of each region affects the optimal parameter α. The form of (27) for γ involves the grid resolution ℓ and therefore the regional resolution has been accounted for in the coefficient γ. Hence, we believe the form of (27) can be the starting point for parameter optimizations on variable-resolution grids. However, we also expect new challenges (e.g., the interactions between the flows from different regions).
Our first step in doing parameter optimizations is to identify a parameter that is invariant with respect to other configurations of the model (e.g., the time-step size and the spatial grid resolution). We believe that this approach applies to and can be taken toward other parameterizations or subgrid eddy closure schemes. It is our intention to apply this approach to certain other important parameterizations (e.g., the Gent–McWilliams parameterization of turbulent transport on variable-resolution grids; Gent and McWilliams 1990; Gent et al. 1995; Ringler and Gent 2011).
Acknowledgments
The authors owe thanks to Mat Maltrud for a careful reading of a draft of this paper and for helpful comments. The authors also thank the anonymous referees, whose comments and suggestions helped to improve the manuscript. Q. Chen and M. Gunzburger were supported by the U.S. Department of Energy Grant DE-SC0002624 as part of the Climate Modeling: Simulating Climate at Regional Scale program. T. Ringler was supported by the DOE Office of Science’s Climate Change Prediction Program Grant DOE 07SCPF152.
APPENDIX
Scale Analysis for the 2D Incompressible Flows
In this appendix, we give details about the scale analysis for two-dimensional inviscid incompressible flows, which leads to a scale-aware formulation for the anticipated potential vorticity method (APVM). Such an analysis serves as a base for the extension to shallow-water flows given in section 2. We also hope that it will provide guidelines for the analysis of other geophysical flows that are predominantly two-dimensional, such as the primitive equations cast in isopycnal coordinates.
REFERENCES
Adams, J. C., and P. N. Swarztrauber, 1997: Spherepack 2.0: A model development facility. NCAR Tech. Note NCAR/TN-436-STR, 59 pp.
Arakawa, A., and W. Schubert, 1974: Interaction of a cumulus cloud ensemble with the large-scale environment, Part I. J. Atmos. Sci., 31, 674–701.
Arakawa, A., and Y.-J. G. Hsu, 1990: Energy conserving and potential-enstrophy dissipating schemes for the shallow water equations. Mon. Wea. Rev., 118, 1960–1969.
Bardos, C., 1972: Existence et unicité de la solution de l’équation d’Euler en dimension deux (Existence and uniqueness of the solution of the two dimensional Euler equations). J. Math. Anal. Appl., 40, 769–790.
Batchelor, G. K., 1969: Computation of the energy spectrum in homogeneous two-dimensional turbulence. Phys. Fluids, 12 (12), 233–239.
Berselli, L. C., T. Iliescu, and W. J. Layton, 2006: Mathematics of Large Eddy Simulation of Turbulent Flows. Springer, 366 pp.
Constantin, P., 2007: On the Euler equations of incompressible fluids. Bull. Amer. Math. Soc., 44, 603–621.
Courant, R., and D. Hilbert, 1953: Methods of Mathematical Physics. Vol. I. Interscience Publishers, xv + 561 pp.
Du, Q., V. Faber, and M. Gunzburger, 1999: Centroidal Voronoi tessellations: Applications and algorithms. SIAM Rev., 41 (4), 637–676.
Frisch, U., 1995: Turbulence: The Legacy of A.N. Kolmogorov. Cambridge University Press, 312 pp.
Gent, P. R., and J. C. McWilliams, 1990: Isopycnal mixing in ocean circulation models. J. Phys. Oceanogr., 20, 150–155.
Gent, P. R., J. Willebrand, T. J. McDougall, and J. C. McWilliams, 1995: Parameterizing eddy-induced tracer transports in ocean circulation models. J. Phys. Oceanogr., 25, 463–474.
Jones, P. W., 1999: First- and second-order conservative remapping schemes for grids in spherical coordinates. Mon. Wea. Rev., 127, 2204–2210.
Kolmogorov, A. N., 1941a: Dissipation of energy in locally isotropic turbulence. Proc. U.S.S.R. Acad. Sci. (Atmos. Ocean. Phys.), 32, 16–18.
Kolmogorov, A. N., 1941b: The local structure of turbulence in incompressible viscous fluid for very large Reynolds numbers. Proc. U.S.S.R. Acad. Sci. (Atmos. Oceanic Phys.), 30, 299–303.
Kraichnan, R. H., 1967: Inertial ranges in two-dimensional turbulence. Phys. Fluids, 10 (7), 1417–1423.
Landau, L. D., and E. M. Lifshitz, 1987: Course of Theoretical Physics. Vol. 6, Fluid Mechanics, 2nd ed. Pergamon Press, xiv + 539 pp.
Leith, C. E., 1968: Diffusion approximation for two-dimensional turbulence. Phys. Fluids, 11 (3), 671–672.
Lilly, D. K., 1969: Numerical simulation of two-dimensional turbulence. Phys. Fluids, 12 (12), 240–249.
Lilly, D. K., 1971: Numerical simulation of developing and decaying two-dimensional turbulence. J. Fluid Mech., 45 (2), 395–415.
Maltrud, M. E., and G. K. Vallis, 1991: Energy spectra and coherent structures in forced two-dimensional and beta-plane turbulence. J. Fluid Mech., 228, 321–342.
McWilliams, J. C., 1989: Statistical properties of decaying geostrophic turbulence. J. Fluid Mech., 198, 199–230.
Pedlosky, J., 1987: Geophysical Fluid Dynamics. 2nd ed. Springer, 728 pp.
Rhines, P. B., 1975: Waves and turbulence on a beta-plane. J. Fluid Mech., 69, 417–443.
Ringler, T., and P. Gent, 2011: An eddy closure for potential vorticity. Ocean Modell., in press.
Ringler, T., L. Ju, and M. Gunzburger, 2008: A multiresolution method for climate system modeling: Application of spherical centroidal Voronoi tessellations. Ocean Dyn., 58, 475–498.
Ringler, T., J. Thuburn, J. B. Klemp, and W. C. Skamarock, 2010: A unified approach to energy conservation and potential vorticity dynamics for arbitrarily-structured c-grids. J. Comput. Phys., 229 (9), 3065–3090.
Ringler, T., D. Jacobsen, M. Gunzburger, L. Ju, M. Duda, and W. Skamarock, 2011: Exploring a multiresolution modeling approach within the shallow-water equations. Mon. Wea. Rev., in press.
Sadourny, R., and C. Basdevant, 1985: Parameterization of subgrid-scale barotropic and baroclinic eddies in quasi-geostrophic models: Anticipated potential vorticity method. J. Atmos. Sci., 42, 1353–1363.
Smagorinsky, J., 1963: General circulation experiments with the primitive equations. Mon. Wea. Rev., 91, 99–164.
Thuburn, J., T. Ringler, W. Skamarock, and J. Klemp, 2009: Numerical representation of geostrophic modes on arbitrarily structured c-grids. J. Comput. Phys., 228, 8321–8335.
Williamson, D. L., J. B. Drake, J. J. Hack, R. Jakob, and P. N. Swarztrauber, 1992: A standard test set for numerical approximations to the shallow water equations in spherical geometry. J. Comput. Phys., 102, 211–224.
In the foregoing reference, θ instead of γ is used for the APVM coefficient. We have switched to γ to avoid confusion with the spherical coordinate θ introduced later.