betaFeedback

WorksPhysics and Astronomy

Modeling sea ice in the marginal ice zone as a dense granular flow with rheology inferred from discrete element model data

Abstract

The marginal ice zone (MIZ) represents the periphery of the sea ice cover. In this region, the macroscale behavior of the sea ice results from collisions and enduring contact between ice floes. This configuration closely resembles that of dense granular flows, which have been modeled successfully with the μ(I)μ(I) rheology. Here, we present a continuum model based on the μ(I)μ(I) rheology which treats sea ice as a compressible fluid, with the local sea ice concentration given by a dilatancy function Φ(I)Φ(I). We infer expressions for μ(I)μ(I) and Φ(I)Φ(I) by nonlinear regression using data produced with a discrete element method (DEM) which considers polygonal-shaped ice floes. We do this by driving the sea ice with a one-dimensional shearing ocean current. The resulting continuum model is a nonlinear system of equations with the sea ice velocity, local concentration, and pressure as unknowns. The rheology is given by the sum of a plastic and a viscous term. In the context of a periodic patch of ocean, which is effectively a one dimensional problem, and under steady conditions, we prove this system to be well-posed, present a numerical algorithm for solving it, and compare its solutions to those of the DEM. These comparisons demonstrate the continuum model's ability to capture most of the DEM's results accurately. The continuum model is particularly accurate for ocean currents faster than 0.25 m/s; however, for low concentrations and slow ocean currents, the continuum model is less effective in capturing the DEM results. In the latter case, the lack of accuracy of the continuum model is found to be accompanied by the breakdown of a balance between the average shear stress and the integrated ocean drag extracted from the DEM.

Introduction

The periphery of the ice cover is known as the marginal ice zone (MIZ) and consists of relatively small, typically polygon-shaped ice floes. It is often defined as the region where ocean waves play an important role in shaping the morphological properties of the ice [6]. On large scales, sea ice dynamics are typically described with Hibler’s model [17], which treats ice as a viscoplastic fluid whose yield strength depends on the sea ice concentration and thickness. This model was developed for the central ice pack, where ice floes are closely interlocked and deformation is mostly due to the opening of leads or the formation of ridges. Recently, elasto-brittle rheologies have also been used to model the evolution of the central ice pack. This class of rheological models, which were first proposed by Girard et al. (2011), appear to be superior to Hibler’s model in capturing the ice deformation field (Rampal et al. 2019).

In the MIZ, however, it is the collisions and enduring contact between ice floes that give rise to the macroscale dynamical properties of the ice cover [8, 15, 16]. This configuration closely resembles that of dense granular flows, albeit at different spatial scales, since practically all studies for granular materials consider e.g. polystyrene beads, glass beads, and sand, whose particles’ diameters are in the order of 0.1 and 1 mm [23]. Dense granular flows have been successfully modeled with the so-called μ(I)\mu(I) rheology (Da Cruz et al. 2005). The dense granular flow regime is understood as a transition between the quasi-static and dilute flow regimes. Whenever grain inertia is negligible, a quasi-static regime emerges which is often modeled as an elastoplastic solid [24]. The critical state at which plastic deformation occurs is characterized with a Coulomb-like criterion dependent on a so-called internal angle of friction [35]. Conversely, under great agitation and/or dilute concentrations of grains, particles interact only through binary, instantaneous, uncorrelated collisions. As a result, ideas from kinetic theory become applicable in this dilute regime [19]. However, in dense granular flows, grains interact via collisions and enduring contacts, such that inertial effects are important yet the collisions may no longer be assumed to be binary, instantaneous, or uncorrelated in general. This transitional regime is characterized in terms of the inertial number II and an effective friction coefficient μ\mu dependent on II (Da Cruz et al. 2005).

Existing models for the MIZ recognize the importance of both collisions and plastic deformation, and derive rheological models based on first principles (Shen et al. 1987; Gutfraind & Savage 1997; Feltham 2005). Recently, Herman (2022) suggests the use of the μ(I)\mu(I) rheology for modeling sea ice in the MIZ and derives a μ(I)\mu(I) function from computations performed with a discrete element method (DEM). In these computations, disk-shaped ice floes are sheared by a moving wall in the classical manner of rheological studies. Unlike the previous models for the MIZ, Herman (2022) infers the rheological properties from data generated by a DEM. In particular, Herman [16] fits a μ(I)\mu(I) function to the DEM’s data, although the resulting continuum model and its accuracy in replicating the DEM’s results is not examined.

This work represents an advance in the development of a continuum model for the MIZ that could improve the accuracy of Hibler’s model, which is currently used in large-scale climate models over the MIZ, see for example (Danabasoglu et al. 2020). A comparison of Hibler’s model with the DEM data is presented in § 6.1, where it is demonstrated that it cannot capture the DEM results accurately. We extend the investigation initiated in Herman [16] and explore the μ(I)\mu(I) rheology’s accuracy in modeling sea ice dynamics in the MIZ. We infer a μ(I)\mu(I) function from data produced with the DEM implemented in SubZero [22]. This DEM considers polygon-shaped ice floes that are driven by oceanic currents in an open patch of ocean, a setup which we believe to be more natural for studying sea ice than the classical shearing test with a moving wall and disk-shaped ice floes. This inference results in a continuum viscous fluid model whose rheology is given by the sum of a viscous and a plastic term. Moreover, for this system to be well-posed, the emerging model problem requires the continuum to be compressible and complemented with a constraint on the global sea ice concentration. Assuming the continuum to be compressible requires the inference of a dilatancy function Φ(I)\Phi(I) from the DEM computations which establishes a relationship between local sea ice concentration and the inertial number II.

The contributions of this paper can be summarized as follows: (1) Inference of the μ(I)\mu(I) and Φ(I)\Phi(I) constitutive functions for sea ice in the MIZ from data produced with the DEM. These computations are performed in an open ocean configuration where the sea ice is sheared by ocean currents. (2) Analysis of the resulting continuum model, establishing the existence and uniqueness of solutions. (3) Determination of the continuum model’s range of validity by comparing its numerical solutions to those of the DEM.

We remark that the analysis of the continuum model and its comparisons with the DEM are restricted to a steady one dimensional setup. The model can be extended to unsteady two dimensional problems as explained in § 2.1, although we expect new complications will arise with these extensions. For example, Barker et al. [2] demonstrate the emergence of time-dependent instabilities in μ(I)\mu(I) models, which Schaeffer et al. [32] remedy with further modifications of the model. These potential complications should be studied carefully in future investigations.

This paper is structured as follows. In § 2, we formulate the continuum model, first in a general two-dimensional unsteady setting, then in the one dimensional steady configuration considered in this paper. In this formulation, two functions, μ(I)\mu(I) and Φ(I)\Phi(I), are to be inferred from DEM data. This inference is presented in § 3. § 4 contains a detailed analysis of the continuum model resulting from this inference. This analysis examines several properties of the momentum equation, the numerical solution of the continuum model, and its well-posedness. Then, in § 5, we compare the continuum model and the DEM. This comparison allows us to establish the range of validity of the continuum model and its limitations. In § 6, we discuss the similarities and differences between our continuum model and other sea ice models, such as Hibler’s model. We then end this paper with § 7, where we recommend potential extensions of this work to be explored in the future.

Mathematical formulation of the continuum problem

The dense flow regime represents a transition between the quasi-static and dilute flow regimes [4]. This transitional regime is characterized in terms of the inertial number II, an effective friction coefficient μ(I)\mu(I), and, whenever the continuum is assumed to be compressible, a dilatancy function Φ(I)\Phi(I). Below, we define these three terms and present a general formulation of the μ(I)\mu(I) rheology in two dimensions and in the one dimensional steady configuration considered in the subsequent sections of this paper.

The two-dimensional setting

Although the problems presented in this paper are effectively one dimensional, we first present the general form of the flow model in two dimensions for completeness. We denote the ice velocity, concentration, and Cauchy stress tensor by u\mathbf{u}, AA, and σ\boldsymbol{\sigma}, respectively. We write the components of the Cauchy stress tensor and the velocity vector field as

σ=[σxxσxyσxyσyy]andu=(u,v),(1)\boldsymbol{\sigma} = \begin{bmatrix} \sigma_{xx} & \sigma_{xy} \\ \sigma_{xy} & \sigma_{yy} \end{bmatrix} \quad\text{and} \quad \mathbf{u} = (u,v), \tag*{(1)}

respectively. We assume that the morphology of the ice floes remains invariant by neglecting all thermomechanical effects, such as fracturing, melting, or ridging, that can change the shape of a floe. For simplicity, we also neglect the Coriolis force, ocean tilting and the atmospheric drag (we assume low-wind conditions). Under these conditions, conservation of momentum and mass lead to the following system of equations:

ρHDuDt=∇⋅σ+to,(2)\rho H \frac{\mathrm{D}\mathbf{u}}{\mathrm{D}t} = \nabla\cdot\boldsymbol{\sigma} + \mathbf{t}_o, \tag*{(2)}
DADt=−A∇⋅u.(3)\frac{\mathrm{D}A}{\mathrm{D}t} = -A \nabla\cdot\mathbf{u}. \tag*{(3)}

see Hibler III [17] or Gray & Morland [12]. For any scalar or vector-valued function ff, the material derivative is given by Df/Dt=∂f/∂t+(u⋅∇)f\mathrm{D}f/\mathrm{D}t = \partial f/\partial t + (\mathbf{u} \cdot\nabla)f. Here, we assume the ice thickness HH to be spatially uniform for simplicity, although in general we require an additional equation, analogous to (3) but in terms of HH, for mass to be conserved. Equation (2) is a depth-averaged statement of conservation of momentum of the sea ice layer; here, to\mathbf{t}_o is the drag force exerted by the ocean on the sea ice. Given the surface ocean velocity field uo\mathbf{u}_o, this drag force is generally parameterized in terms of the drag coefficient CoC_o and the ocean water density ρo\rho_o by

to:=ρoCo∥uo−u∥(uo−u),(4)\mathbf{t}_o := \rho_o C_o \lVert\mathbf{u}_o - \mathbf{u} \rVert(\mathbf{u}_o - \mathbf{u}), \tag*{(4)}

with ∥⋅∥\lVert\cdot\rVert denoting the Euclidean norm of a vector. The conservation laws (2.2) must be accompanied by constitutive relations. To write these, we first decompose the Cauchy stress tensor into a pressure term pp and its deviatoric component τ\boldsymbol{\tau},

σ=τ−pI,(5)\boldsymbol{\sigma} = \boldsymbol{\tau} - p\mathbf{I}, \tag*{(5)}

where I\mathbf{I} is the identity tensor, and define the strain rate tensor D\mathbf{D} and its deviatoric component S\mathbf{S} as

D:=12(∇u+∇uT)andS:=D−12(∇⋅u)I.(6)\mathbf{D} := \frac{1}{2}(\nabla\mathbf{u} + \nabla\mathbf{u}^{\mathsf{T}}) \quad\text{and} \quad\mathbf{S} := \mathbf{D} - \frac{1}{2}(\nabla\cdot\mathbf{u})\mathbf{I}. \tag*{(6)}

In the following, for a given tensor T\mathbf{T}, its second invariant is denoted by

∥T∥=12tr⁡(T2).(7)\lVert\mathbf{T} \rVert= \sqrt{\frac{1}{2}\operatorname{tr}(\mathbf{T}^2)}. \tag*{(7)}

The fundamental idea behind the μ(I)\mu(I) rheology is that the constitutive relation for a granular flow depends on the inertial number, a dimensionless quantity defined as

I:=dˉHρip∥S∥,(8)I := \bar{d}\sqrt{\frac{H\rho_i}{p}}\lVert\mathbf{S}\rVert, \tag*{(8)}

where ρi\rho_i is the ice density, dˉ\bar{d} denotes an average ice floe size, and pp is the pressure emerging in (5) ([4, 20, 26]). Throughout this document, we set dˉ\bar{d} to be spatially constant over the whole domain, avoiding the need to consider the transport of this quantity. Savage [31] interprets the quantity I2I^2 as the ratio between collisional (i.e. inertial) stresses and the total shear stress exerted on the material. Accordingly, for low values of II, the inertial effects of grains become negligible and the flow approaches a quasi-static regime; conversely, as II increases, collisional forces become increasingly important relative to the external forces exerted on the material. The functional relationship that establishes the material’s rheology is written in terms of an effective friction μ(I)\mu(I), defined as

μ(I):=∥τ∥p.(9)\mu(I) := \frac{\lVert\boldsymbol{\tau} \rVert}{p}. \tag*{(9)}

We remark that the effective friction μ\mu is defined in analogy with Coulomb’s model of friction as the ratio between the shear (tangential) stress and the pressure (normal stress). Moreover, it is also helpful to think of the pressure pp as a quantification of the material’s strength and its resistance to viscous and plastic deformation, as made clear in § 4. It should be noted that the μ(I)\mu(I) model is a phenomenological model that has been found to work well with granular media, yet it is unclear if it represents some kind of limit for a large particle system.

To obtain a relationship between stress and strain, we need an additional constitutive law. Jop et al. [20] propose the following equality that aligns S\mathbf{S} with τ\boldsymbol{\tau}:

S∥S∥=τ∥τ∥.(10)\frac{\mathbf{S}}{\lVert\mathbf{S}\rVert} = \frac{\boldsymbol{\tau}}{\lVert\boldsymbol{\tau}\rVert}. \tag*{(10)}

Combining (9) and (10), the relationship between deviatoric components of the stress tensor and the shear strain can be written as

τ=μ(I)pS∥S∥.(11)\boldsymbol{\tau} = \mu(I)p\frac{\mathbf{S}}{\lVert\mathbf{S}\rVert}. \tag*{(11)}

Compressible granular flows require a dilatancy law which relates the concentration AA with the inertial number II,

A=Φ(I),(12)A = \Phi(I), \tag*{(12)}

see Da Cruz et al. [4]. In general, Φ\Phi is found to be a strictly decreasing function of II, in such a way that the concentration AA decreases with the rate of shearing ∥S∥\lVert\mathbf{S}\rVert, a phenomenon know as dilatancy. Moreover, if Φ\Phi is strictly decreasing, it is invertible, and one can write an expression for the pressure pp analogous to an equation of state in thermodynamics,

p=ρidˉ2H(∥S∥Φ−1(A))2,(13)p = \rho_i\bar{d}^{2}H\left(\frac{\lVert\mathbf{S}\rVert}{\Phi^{-1}(A)}\right)^{2}, \tag*{(13)}

where we have combined equations (8) and (12). In the problems considered in this paper, we find the spatial variations in sea ice concentration to be small. Although this would make the assumption of incompressibility reasonable, the periodic one dimensional nature of these problems renders the dilatancy law (12) necessary for the model to be well-posed. This point is explained below in section 2.22.2.

The steady one dimensional periodic ocean problem

The model problem considered in this paper consists of a square patch of ocean of length LL with periodic boundaries in both the xx and yy-directions. The ice floes floating on this patch are driven by an ocean velocity field that only varies in the yy-direction, as depicted in Figure 1. We neglect time-dependent effects and only consider steady conditions in the forcing i.e. uo(x,y,t)=(uo(y),0)\mathbf{u}_o(x,y,t)=(u_o(y),0).

Image not extracted

Figure 1. The periodic ocean setup. The domain is a square patch of ocean, periodic in the horizontal and vertical directions. The ice floes are driven by the ocean velocity field (19) (in blue), which does not vary in the xx-direction.

This configuration renders the continuum problem one dimensional and independent of time, such that u(x,y,t)=(u(y),0)\mathbf{u}(x,y,t)=(u(y),0), A(x,y,t)=A(y)A(x,y,t)=A(y), and p(x,y,t)=p(y)p(x,y,t)=p(y). In this setting, the equations for conservation of momentum (2), together with the constitutive equation (11), simplify to the following system on (0,L)(0,L):

−dσxydy=ρoCo∣uo−u∣(uo−u)(14)-\frac{\mathrm{d}\sigma_{xy}}{\mathrm{d}y} = \rho_o C_o |u_o-u|(u_o-u) \tag*{(14)}
dpdy=0,(15)\frac{\mathrm{d}p}{\mathrm{d}y}=0, \tag*{(15)}
σxy=μ(I)pdu/dy∣du/dy∣,(16)\sigma_{xy}=\mu(I)p\frac{\mathrm{d}u/\mathrm{d}y}{|\mathrm{d}u/\mathrm{d}y|}, \tag*{(16)}
I=dˉHρip∣du/dy∣.(17)I=\bar{d}\sqrt{\frac{H\rho_i}{p}}|\mathrm{d}u/\mathrm{d}y|. \tag*{(17)}

Due to (15), which represents the balance of momentum in the yy-direction, the pressure is a constant (but unknown) over the domain. Da Cruz et al. [4] and Herman [16] find pp by enforcing normal stress boundary conditions along a boundary of the domain, but we cannot do the same because the domain is periodic. In the DEM computations, which we introduce in section 3, we set a global ice concentration A0∈[0,1]A_0 \in[0,1] which equals the domain averaged value of the local concentration AA, such that

1L∫0LA dy=A0.(18)\frac{1}{L}\int_0^L A\,\mathrm{d}y=A_0. \tag*{(18)}

Therefore, if we assume the sea ice to behave like a compressible fluid, condition (18) and the dilatancy law (12) close the system of equations. In sections 3 and 5, we justify this modeling choice by demonstrating that dilatancy emerges in the DEM computations and that our model is capable of capturing it accurately.

In this paper we only consider the following ocean velocity profile for simplicity:

uo(y)=uo,max⁡(1−∣1−2yL∣),(19)u_o(y)=u_{o,\max}\left(1-\left|1-\frac{2y}{L}\right|\right), \tag*{(19)}

for some maximum velocity uo,max⁡>0u_{o,\max}>0. Figure 1 contains a plot of this velocity profile.

Inferring the constitutive equations of the system from the DEM

The system of equations presented in section 2.2 is incomplete because we need additional expressions for μ(I)\mu(I) and Φ(I)\Phi(I). In this paper, we infer these additional equations from data generated with SubZero, a DEM developed by Manucharyan & Montemuro [22] and used for modeling sea ice dynamics with polygonal-shaped ice floes. Following the setup presented in § 2.2, we perform runs with n=2000n = 2000 floes over a square patch of ocean of length L=100 kmL = 100\,\mathrm{km}, driven by the ocean velocity field (19). This means that the μ(I)\mu(I) and Φ(I)\Phi(I) functions are inferred from DEM solutions with a constant number of floes (and therefore constant average floe size). In theory, the effects of floe size are included in the non-dimensional parameter II. In practice, the effects of nn are subtle and not too well captured with our continuum model, see Figure 12 in § 5 below.

Figure 12. Solutions to the non-dimensional continuum model (4.2) (lines) compared with results from the DEM (markers). The solutions to the continuum model are plotted in red for n=500n = 500, blue for n=2000n = 2000 and green for n=5000n = 5000; the tone is set to light if A0=0.8A_0 = 0.8 and dark if A0=0.9A_0 = 0.9 (see legend in (e)). For each pair of panels in the first row, we fix the sea ice thickness HH and vary the maximum ocean velocity uo,max⁡u_{o,\max} (the ocean velocity is plotted in blue in panels (a) and (c)). In panel (e), we plot the pressure pp in terms of nn.

For a given number of floes and a global sea ice concentration A0A_0, the initial configuration of ice floes is generated with SubZero’s packing algorithm, which is based on a Voronoi tessellation of the domain [22] (see panel (a) of Figure 2 below for an example of the outcome of this packing algorithm). Defining the floe size dd as the square root of the floe area, this generates a polydisperse floe size distribution whose histogram we can see in Figure 3 for three values of A0A_0; we find that d/Ld/L is approximately between 0.01 and 0.04. Although we do not study the effects of polydispersity on the rheology here, we remark that Herman [16] finds the dilatancy law Φ\Phi to vary visibly with the degree of polydispersity, while the effective friction μ\mu presents much smaller variations.

Image not extracted

Figure 2. Emergence of dilatancy in the DEM. Here, A0=0.8A_0 = 0.8 and uo=1 m/su_o = 1\,\mathrm{m}/\mathrm{s}. Ten runs to steady state are performed from randomized initial conditions, with (a) and (b) being examples of initial and final states, respectively. (c) ocean and sea ice velocities, (d) initial (red) and final (black) mean local sea ice concentration with the standard deviation (bars). Clearly, AA decreases in the areas of maximum shear strain, compare (c) and (d).

Image not extracted

Figure 3. Floe size distributions for different global sea ice concentrations A0A_0 and a total number of ice floes n=2000n = 2000. The floe size dd is defined as the square root of the floe area.

We run the DEM simulations for 2×1042 \times10^4 time steps, each of 5 s (the total running time is approximately equal to 27.8 hours). In general, we find that the velocity, stress, and sea ice concentration, averaged over the last 25% of the time steps, remain relatively unchanged when a longer computation is performed, and hence we consider that a steady state has been reached. This temporal averaging is performed over data which, at each time step, has been averaged spatially over a grid as described in appendix A.§ 2. For these simulations, the material parameters which determine the effects of the ocean drag and the collisions among floes are presented in Table 1. Collisional forces and the resultant stresses, which determine the fields σ\sigma and pp, are computed as explained in appendix A.§ 1.

CoC_oρi\rho_iρo\rho_oEEν\nuμ∗\mu^*
3×10−33 \times10^{-3}900 kg/m3900\,\mathrm{kg}/\mathrm{m}^31026 kg/m31026\,\mathrm{kg}/\mathrm{m}^36×106 Pa6 \times10^6\,\mathrm{Pa}0.30.30.20.2

Table 1. Values for material parameters used throughout this document. Here, CoC_o is the drag coefficient for the ocean current and ρi\rho_i and ρo\rho_o are the ice and ocean water densities, respectively. The Young’s modulus EE, Poisson’s ratio ν\nu, and inter-floe friction coefficient μ∗\mu^* are used in the calculation of collisional forces, as described in Appendix A.§ 1.

In Figure 4 we plot the values of the friction μ=∣σxy∣/p\mu= \lvert\sigma_{xy} \rvert/p and local concentration AA against II for different global concentrations A0A_0 between 0.7 and 0.95 and different maximum ocean velocities uo,max⁡u_{o,\max} between 0.1 and 1 m/s. The mean ocean velocities in the ocean patch are therefore between 0.05 and 0.5 m/s, values that are consistent with real observations [34]. In all of these computations, we set the ice thickness to H=2H = 2 m. We find an increase in the friction μ\mu and a decrease in the local concentration AA as II increases. The decrease in the local concentration of sea ice AA with an increase in II is due to dilatancy. Figure 2 presents an example of how dilatancy emerges in the DEM computations: given a random initial distribution of ice floes, when a steady state is reached the concentration AA decreases in the areas where the largest shearing occurs (y=1/4y = 1/4 and y=3/4y = 3/4), and increases elsewhere. Since the global sea ice concentration A0A_0 is constant, in this context dilatancy represents a reorganization of the local concentration profile A(y)A(y).

Image not extracted

Figure 4. (a) friction μ=∣σxy∣/p\mu= \lvert\sigma_{xy}\rvert/p and (b) concentration AA against II. Ten friction and concentration values are extracted from each DEM run by averaging along the grid cells plotted in Figure 15. The black lines correspond to the functions (20) fitted to the data by minimizing the least-squares misfit.

The trends found in the data in Figure 4 are well fitted with the following family of functions:

μ(I)=μ0+μ1IandΦ(I)=1−ϕ0Iα.(20)\mu(I) = \mu_0 + \mu_1 I \qquad\text{and} \qquad\Phi(I) = 1 - \phi_0 I^\alpha. \tag*{(20)}

A linear behavior is also found for μ\mu in Da Cruz et al. [4]. The four parameters (μ0,μ1,ϕ0,α)(\mu_0, \mu_1, \phi_0, \alpha) are calculated by minimizing the least-squares misfit problem between the points in Figure 4 and the functions in (20). The resulting values are shown in Table 2. For the remainder of the document, any numerical solution of the one dimensional system (2.13) is solved by setting the parameters in functions (20) to the values given in Table 2.

μ0\mu_0μ1\mu_1ϕ0\phi_0α\alpha
0.264.930.530.24

Table 2. Parameters for the functions μ(I)\mu(I) and Φ(I)\Phi(I) in (20) obtained by minimizing the least-squares misfit with the data plotted in Figure 4. These are the numerical values used for computing solutions to the continuum model (2.13) throughout this paper.

Departure from the fitting curves are most visible when the ocean velocities and sea ice concentrations are low, see the case where uo,max⁡=0.1u_{o,\max} = 0.1 m/s and A0=0.7A_0 = 0.7. Unsurprisingly, in § 5 below, we also find the greatest misfit between the DEM and the continuum model precisely in this setting, when uo,max⁡=0.1u_{o,\max} = 0.1 m/s and A0=0.7A_0 = 0.7, see panel (m) in Figure 9 below. In particular, in this setting, the fundamental balance between shear stress and ocean drag in the DEM is found to no longer hold, see § 5.

Image not extracted

Figure 9. Solutions to the non-dimensional continuum model (4.2) (black lines) compared with results from the DEM (markers), with the ocean velocity in blue. For each pair of panels in the first three rows, we fix the global sea ice concentration A0A_0, vary the maximum ocean velocity uo,max⁡u_{o,\max} for the DEM, and plot the velocities in the panels to the left ((a), (c), (e), (g), (i), and (k)) and the local concentration to the right ((b), (d), (f), (h), (j), and (l)). In panel (m), we plot the pressure pp in terms of A0A_0. Due to the normalization in terms of [u]=uo,max⁡[u] = u_{o,\max}, the solutions to the continuum model are indifferent to a change in uo,max⁡u_{o,\max}.

The constitutive equation in 2D resulting from functions (20) is the sum of a plastic and a viscous term:

τ=μ0pS∥S∥+μ1dˉ2ρiHˉp S(21)\boldsymbol{\tau} = \mu_{0}p\frac{\boldsymbol{S}}{\lVert\boldsymbol{S}\rVert} + \mu_{1}\bar{d}^{2}\sqrt{\rho_{i}\bar{H}p}\,\boldsymbol{S} \tag*{(21)}

A consequence of this linear behavior is that μ\mu is approximately constant for small values of II, and this is precisely what we see for I<10−2I < 10^{-2} in Figure 4. For I≪1I \ll1, we have that μ(I)≈μ0\mu(I) \approx\mu_{0}, and therefore it is the plastic term that dominates the rheology. This is essentially the quasi-static regime, where collisions are negligible. This plastic term follows from a Mohr-Coulomb yield criterion with an internal angle of friction tan⁡−1(μ0)\tan^{-1}(\mu_{0}); examples of the Mohr-Coulomb yield criterion used for sea ice modeling can be found in Ip et al. (1991), Gutfraind & Savage [13] and Ringeisen et al. [29]. The viscous term, which becomes increasingly important as the inertial number II increases, can be associated with the collisional component of the rheology. A viscous rheology is derived in Shen et al. [33] for modeling the rheological effects of collisions in sea ice, which, as we explain in § 6, is very similar to the viscous component in (21).

Analysis of the inferred continuum model

This section focuses on the one dimensional system of equations presented in § 2.2, with the functions μ\mu and Φ\Phi taking the form (20). Firstly, in § 4.1 we non-dimensionalize the system of equations to understand the relative importance of the different terms involved. The remaining two sections explore the existence and uniqueness of solutions to this system, a regularization technique that facilitates its numerical solution, and the behavior of solutions under different limits. In particular, in § 4.2, we focus only on the momentum equation (15). In § 4.2.1, we show that the momentum equation can be rewritten as a minimization problem. This allows us to establish that solutions to this equation exist and are unique, and it allows us to make sense of the plastic component (see (21)) in a rigorous sense (as a variational inequality). Moreover, this equivalence also motivates a regularization of the plastic term that simplifies its numerical solution considerably, as described in § 4.3. We end § 4.2 with an analysis of how the sea-ice velocity behaves under different limits in parameter values: in § 4.4, we explore the behavior of the sea-ice velocity for small and large pressures pp; then, in § 4.5, we derive an analytical solution for the sea-ice velocity in the purely plastic limit. Understanding these limit solutions for the velocity is helpful for interpreting the DEM results we present in § 5. It is also useful for the analysis we present in § 4.6, which considers the complete system of equations (2.13). We begin by presenting a numerical method for solving the complete system in section 4.3.1. Then, in § 4.7, we sketch out a demonstration of the existence and uniqueness of solutions to the complete system.

We remark that § 4.2.1, § 4.3, and § 4.6 are mostly concerned with questions of a mathematical and numerical nature. Although we believe these to be important topics in establishing the suitability of our model for modeling purposes, they are not required for understanding the remainder of the paper. We also note that, for simplicity, the solutions presented throughout this section result from driving the ice floes with the ocean velocity profile (19), although the analysis can be extended to more general ocean velocity profiles by following the same steps.

Non-dimensionalization of the system

For the non-dimensionalization, we set the characteristic magnitudes

[y]=L,[H]=H,[u]=uo,max⁡,and[σ]=ρi[u]2[H](22)[y] = L,\qquad[H] = H,\qquad[u] = u_{o,\max},\qquad\text{and}\qquad[\sigma] = \frac{\rho_i [u]^2}{[H]} \tag*{(22)}

for the length, thickness, velocity, and stress, respectively. We scale the velocities uu and uou_o with [u][u], the spatial variables yy and dˉ\bar{d} with [y][y], the thickness HH with [H][H], and σxy\sigma_{xy} and pp with [σ][\sigma].

From this point onward, all variables considered are non-dimensional unless the contrary is made explicitly clear or units are specified. Keeping the same notation as used for dimensional variables, the following normalized system of equations is derived for uu, II, and AA, and for the constant p>0p > 0:

−ϵddy(μ0pdu/dy∣du/dy∣+μ1A0npdudy)=β0∣uo−u∣(uo−u)(23)-\epsilon\frac{\mathrm{d}}{\mathrm{d}y}\left(\mu_0 p \frac{\mathrm{d}u/\mathrm{d}y}{\lvert\mathrm{d}u/\mathrm{d}y\rvert}+\mu_1\sqrt{\frac{A_0}{n}p}\frac{\mathrm{d}u}{\mathrm{d}y}\right)=\beta_0\lvert u_o-u\rvert(u_o-u) \tag*{(23)}
I=A0ρn∣du/dy∣(24)I=\sqrt{\frac{A_0}{\rho n}}\lvert\mathrm{d}u/\mathrm{d}y\rvert \tag*{(24)}
A=1−ϕ0Iα(25)A=1-\phi_0 I^\alpha \tag*{(25)}
∫01A dy=A0(26)\int_0^1 A\,\mathrm{d}y=A_0 \tag*{(26)}

Here, ϵ=H/L\epsilon=H/L and β0=ρo/ρiCo\beta_0=\rho_o/\rho_i C_o; the non-dimensional average floe size dˉ\bar{d} is set to A0/n\sqrt{A_0/n}. The system (4.2) is closed by enforcing periodic boundary conditions for uu, II, and AA. Following our findings in Table 2, we assume that the parameters μ0\mu_0, ϕ0\phi_0, and α\alpha are strictly greater than zero. We also assume μ1>0\mu_1>0 in all but § 4.5, where we study the case when μ1=0\mu_1=0 with the intention of understanding the plastic component of the momentum equation (23).

All numerical results computed in this section take (19) as the ocean velocity, which is written as

uo(y)=1−∣1−2y∣(27)u_o(y) = 1 - \lvert1 - 2y \rvert \tag*{(27)}

for y∈(0,1)y \in(0,1) when non-dimensionalized.

The momentum equation

In order to understand the system of equations (4.2), we first focus on the momentum equation (23). When considering the entire system (4.2), the pressure p∈Rp \in\mathbb{R} is one of the unknowns. However, it is useful to first assume it to be known, in which case we can solve the momentum equation (23) for uu and study the effect of pp on uu. Here, we show that (23) can be understood as a minimization problem. This reformulation of the momentum equation allows us to establish the existence and uniqueness of solutions. Moreover, the optimality conditions for the minimization problem result in a different formulation of the plasticity component of the rheology which avoids the singularity, present in (23), at du/dy=0\mathrm{d}u/\mathrm{d}y = 0. With this reformulation of the plastic term, we are able to find analytical solutions to the purely plastic problem which arises when I≪1I \ll1, near the quasi-static regime.

Reformulation of (4.2a) as a minimization problem

Given a pressure p>0p > 0, solutions uu to (23) minimize the following functional,

J(u):=ϵμ0p∫01∣dudy∣ dy+ϵμ12pA0n∫01∣dudy∣2 dy+βo3∫01∣uo−u∣3 dy(28)\mathcal{J}(u) := \epsilon\mu_0p\int_0^1 \left|\frac{\mathrm{d}u}{\mathrm{d}y}\right|\,\mathrm{d}y + \frac{\epsilon\mu_1}{2}\sqrt{p\frac{A_0}{n}}\int_0^1 \left|\frac{\mathrm{d}u}{\mathrm{d}y}\right|^2\,\mathrm{d}y + \frac{\beta_o}{3}\int_0^1 |u_o-u|^3\,\mathrm{d}y \tag*{(28)}

over the space of velocity profiles

V:={v∈H1((0,1)):v is a periodic function}.(29)V := \left\{v \in H^1((0,1)) : v\text{ is a periodic function}\right\}. \tag*{(29)}

In the definition of VV, the space H1((0,1))H^1((0,1)) denotes the Sobolev space of weakly differentiable and periodic functions on the unit interval [1]. As explained in appendix B, the functional J\mathcal{J} is strictly convex and coercive over VV and therefore admits a unique minimizer. In this sense, one can state that the momentum equation (23) also has a unique solution.

To derive (23) for the minimizer uu of J\mathcal{J}, we first note that, if uu minimizes J\mathcal{J}, then

1t(J(u+t(v−u))−J(u))≥0∀v∈V, ∀t∈(0,1).(30)\frac{1}{t}\left(\mathcal{J}(u+t(v-u))-\mathcal{J}(u)\right) \ge0 \quad\forall v \in V,\ \forall t \in(0,1). \tag*{(30)}

The three terms in the right hand size of (28) are convex, with the two last ones differentiable over all VV. By exploiting the convexity of the first term (the L1L^1 norm) and the differentiability of the other two terms, we find that

ϵμ0p(∣dvdy∣−∣dudy∣)+ϵμ1pA0n∫01dudyd(v−u)dy dx−βo∫01∣uo−u∣(uo−u)(v−u) dx≥0∀v∈V.(31)\begin{aligned} &\epsilon\mu_0p\left(\left|\frac{\mathrm{d}v}{\mathrm{d}y}\right|-\left|\frac{\mathrm{d}u}{\mathrm{d}y}\right|\right) +\epsilon\mu_1\sqrt{p\frac{A_0}{n}}\int_0^1\frac{\mathrm{d}u}{\mathrm{d}y}\frac{\mathrm{d}(v-u)}{\mathrm{d}y}\,\mathrm{d}x\\ &\qquad-\beta_o\int_0^1|u_o-u|(u_o-u)(v-u)\,\mathrm{d}x \ge0 \quad\forall v\in V. \tag*{(31)} \end{aligned}

A variational statement as in (31) is known as a variational inequality. Under the assumption that the solution uu is not only in VV but is twice continuously differentiable, we may deduce that

−ϵddy(σxyP+μ1pA0ndudy)=β0∣uo−u∣(uo−u),(32)-\epsilon\frac{\mathrm{d}}{\mathrm{d}y}\left(\sigma^{P}_{xy}+\mu_1\sqrt{p\frac{A_0}{n}}\frac{\mathrm{d}u}{\mathrm{d}y}\right)=\beta_0\lvert u_o-u\rvert(u_o-u), \tag*{(32)}
∣σxyP∣≤μ0p,(33)\lvert\sigma^{P}_{xy}\rvert\leq\mu_0p, \tag*{(33)}
σxyP∣dudy∣=μ0pdudy.(34)\sigma^{P}_{xy}\left\lvert\frac{\mathrm{d}u}{\mathrm{d}y}\right\rvert=\mu_0p\frac{\mathrm{d}u}{\mathrm{d}y}. \tag*{(34)}

A similar derivation to that of (4.8) from (31) can be found in Glowinski et al. [11], section 1.3). In (4.8), we have introduced σxyP\sigma^{P}_{xy}, the purely plastic component of the shear stress. Introducing this new variable allows us to reformulate (23) such that the singularity at du/dy=0\mathrm{d}u/\mathrm{d}y=0 is removed. Indeed, if du/dy≠0\mathrm{d}u/\mathrm{d}y\ne0, it is clear that (4.8) is equivalent to (23). In this case, we have that ∣σxyP∣=μ0p\lvert\sigma^{P}_{xy}\rvert=\mu_0p and we say that the material has reached its plastic yield strength μ0p\mu_0p. Conversely, when du/dy=0\mathrm{d}u/\mathrm{d}y=0, (32) remains well defined, unlike (23). We remark that du/dy=0\mathrm{d}u/\mathrm{d}y=0 must follow from (34) whenever ∣σxyP∣<μ0p\lvert\sigma^{P}_{xy}\rvert<\mu_0p (the material has not reached its plastic yield strength). Below, in section 4.2.4, we provide further insight into the plastic component of the shear stress by computing purely plastic solutions to the momentum equation analytically.

Regularization of the plastic term to facilitate its numerical solution

The first order optimality condition for the minimization of J\mathcal{J} is a variational inequality (rather than a variational equality) because the first term to the right hand side of (28) (the L1L^1 norm) is non differentiable when du/dy=0\mathrm{d}u/\mathrm{d}y=0. We can make J\mathcal{J} differentiable by regularizing it as follows:

JΔ(u):=ϵμ0p∫01∣dudy∣2+Δ2 dy+ϵμ12pA0n∫01∣dudy∣2 dy+β03∫01∣uo−u∣3 dy.(35)\mathcal{J}_{\Delta}(u):=\epsilon\mu_0p\int_0^1\sqrt{\left\lvert\frac{\mathrm{d}u}{\mathrm{d}y}\right\rvert^2+\Delta^2}\,\mathrm{d}y+\frac{\epsilon\mu_1}{2}\sqrt{p\frac{A_0}{n}}\int_0^1\left\lvert\frac{\mathrm{d}u}{\mathrm{d}y}\right\rvert^2\,\mathrm{d}y+\frac{\beta_0}{3}\int_0^1\lvert u_o-u\rvert^3\,\mathrm{d}y. \tag*{(35)}

where Δ>0\Delta>0 is a small parameter. The first order optimality conditions for the minimization of JΔ\mathcal{J}_{\Delta} over VV corresponds with the following equation:

−ϵddy(μ0pdu/dy∣du/dy∣2+Δ2+μ1pA0ndudy)=β0∣uo−u∣(uo−u).(36)-\epsilon\frac{\mathrm{d}}{\mathrm{d}y}\left(\mu_0p\frac{\mathrm{d}u/\mathrm{d}y}{\sqrt{\lvert\mathrm{d}u/\mathrm{d}y\rvert^2+\Delta^2}}+\mu_1\sqrt{p\frac{A_0}{n}}\frac{\mathrm{d}u}{\mathrm{d}y}\right)=\beta_0\lvert u_o-u\rvert(u_o-u). \tag*{(36)}

Although the system (4.8) can be solved numerically by e.g. introducing a Lagrange multiplier (Glowinski et al. [11]), it is easier to solve equation (36). This is the strategy we take for solving the momentum equation and, as we explain below in section 4.3.1, the complete system (4.2). To do so, we use the finite element method (FEM) implemented in Firedrake (Ham et al. [14]). In particular, we approximate the velocity profile uu with continuous piece-wise linear functions. In Figure 5, we plot solutions to (36) for two different values of pp and a range of Δ>0\Delta>0. Convergence of the velocity profiles as Δ→0\Delta\to0 is clearly visible in these figures; in fact, for Δ≤10−2\Delta\leq10^{-2}, the solutions become indistinguishable. We remark that, if we remove the viscous component of the rheology in the regularized equation (36), we essentially arrive at Hibler’s model in one dimension, given below by (59).

Image not extracted

Figure 5. Solutions uu to the momentum equation (36) for different values of the ice pressure pp and the numerical regularization parameter Δ\Delta. We set uou_o equal to (19), normalized with [u]=uo,max⁡[u] = u_{o,\max}, and ϵ=2×105\epsilon= 2 \times10^{5}, A0=0.8A_0 = 0.8 and n=2000n = 2000. We observe convergence to a solution as Δ→0\Delta\to0 and, for small pp, u→uou \to u_o in (a) and, for large pp, u→0.5u \to0.5 in (b), as expected from the theory.

Velocity profiles in the limit of small and large pressures

The pressure or ice strength pp is a fundamental variable in the continuum model; understanding its effect on uu is fundamental for making sense of our sea ice model. Figure 5 suggests that, for small pp, the velocity profile uu approaches the ocean’s uou_o, and for large pp, uu flattens and comes close to a constant-valued velocity profile. We can deduce this behavior from the functional J\mathcal{J}. For small values of pp,

lim⁡p→0J(u)=β03∫01∣uo−u∣3 dy,(37)\lim_{p \to0} \mathcal{J}(u) = \frac{\beta_0}{3}\int_0^1 \lvert u_o-u\rvert^3\,\mathrm{d}y, \tag*{(37)}

and therefore, since uu minimizes J\mathcal{J}, it must follow that u→uou \to u_o. On the other hand, for large values of pp, we see that

J(u)≈ϵμ0p∫01∣dudy∣ dyas p→∞(38)\mathcal{J}(u) \approx\epsilon\mu_0p\int_0^1\left\lvert\frac{\mathrm{d}u}{\mathrm{d}y}\right\rvert\,\mathrm{d}y \quad\text{as } p \to\infty \tag*{(38)}

and, in principle, any constant velocity profile uc∈Ru_c \in\mathbb{R} minimizes (38). However, this constant velocity field is constrained by the total force balance of the system. That is, due to the periodic boundary conditions, if we integrate (32) along (0,1)(0,1), we must have that

β0∫01∣uo−u∣(uo−u) dy=0.(39)\beta_0\int_0^1\lvert u_o-u\rvert(u_o-u)\,\mathrm{d}y = 0. \tag*{(39)}

Therefore, the constant value to which uu tends as p→∞p \to\infty will be a solution to (39) with u=uc∈Ru = u_c \in\mathbb{R}. Below, in section 4.2.4, we show that a critical pressure pcp_c can be found such that uu is constant whenever p>pcp > p_c and I≪1I \ll1.

Purely plastic solutions to the momentum equation

In figure 4 we can see that μ(I)\mu(I) roughly becomes constant for small inertial numbers I≪1I \ll1, such that μ(I)≈μ0\mu(I) \approx\mu_0 and the flow rheology is plastic. This regime is closely related to the quasi-static regime for granular media, with the material behaving like a purely plastic flow characterized by a critical state at which plastic deformation occurs [35].

The momentum equation to the purely plastic problem where μ(I)=μ0\mu(I) = \mu_0 is given by

ϵdσxyPdy=−β0∣uo−u∣(uo−u).(40)\epsilon\frac{\mathrm{d}\sigma_{xy}^{P}}{\mathrm{d}y}=-\beta_0\lvert u_o-u\rvert(u_o-u). \tag*{(40)}

Equation (40) must be complemented with (33) and (34). Here, we present a method for calculating solutions to (4.20). Additionally, we find a critical pressure pcp_c such that, for p>pcp > p_c, the velocity profiles u\mathbf{u} that solve (40) remain constant and no shear strain occurs in the sea ice. In § 5 below, we show that this critical pressure approximates the pressure computed from the DEM when the global sea ice concentration is high. When following the derivation of purely plastic solutions, it is helpful to look at their plots in Figure 6 below.

Image not extracted

Figure 6. Solutions uu to the purely plastic momentum equation (40). We use the same parameter values as those used for Figure 5 and Δ=10−2\Delta= 10^{-2}. The numerical solutions (black) to the purely plastic problem are indistinguishable from the analytical solution given by (4.20) (red, dotted). As the ice pressure pp increases, the velocity profiles flattens at the critical pressure pcp_c given by (45) and pc≈13.2p_c \approx13.2 in this case.

Conditions (33) and (34) for the plastic stress tensor indicate that there exist two distinct regions of the flow field: a region where the sea ice has yielded and ∣σxyP∣=pμ0\lvert\sigma^{P}_{xy} \rvert= p\mu_0, and another region where the ice has not yielded and ∣σxyP∣≤pμ0\lvert\sigma^{P}_{xy} \rvert\le p\mu_0 and du/dy=0\mathrm{d}u/\mathrm{d}y = 0. By working with this distinction, we can find a purely plastic solution to (40). Due to the symmetry of the problem, we assume that σxyP=0\sigma^{P}_{xy} = 0 at y=0y = 0, 1/21/2, and 11. Then, integrating (40) along the interval (0,y)(0,y) for some y∈(0,1)y \in(0,1), we find that

σxyP(y)=−βoϵ∫0y∣uo−u∣(uo−u) dy.(41)\sigma^{P}_{xy}(y) = -\frac{\beta_o}{\epsilon} \int_0^y \lvert u_o-u\rvert(u_o-u)\,\mathrm{d}y. \tag*{(41)}

Since σxyP(0)=0\sigma^{P}_{xy}(0) = 0, we must necessarily have an interval (0,y1)(0,y_1) where the ice has not yielded and the velocity equals a constant u1u_1. In the context of the ocean velocity profile (27), it makes sense to assume that u1>0u_1 > 0, and therefore u1≥uou_1 \ge u_o near y=0y = 0, so that

σxyP(y)=βo6ϵ((2y−u1)3+u1).(42)\sigma^{P}_{xy}(y) = \frac{\beta_o}{6\epsilon}\left((2y-u_1)^3+u_1\right). \tag*{(42)}

Since the material has not yielded for y∈(0,y1)y \in(0,y_1), we have that ∣σxyP(y)∣<μ0p\lvert\sigma^{P}_{xy}(y)\rvert< \mu_0p. Equation (42) tells us that σxyP(y)\sigma^{P}_{xy}(y) increases with yy over this interval; this means that

lim⁡y→y1y<y1σxyP(y)=μ0p,(43)\lim_{\substack{y\to y_1\\y<y_1}}\sigma^{P}_{xy}(y)=\mu_0p, \tag*{(43)}

and we find that

u1=(6ϵμ0pβo)1/3.(44)u_1=\left(\frac{6\epsilon\mu_0p}{\beta_o}\right)^{1/3}. \tag*{(44)}

Clearly, an upper bound is needed for u1u_1 in (44), because it grows indefinitely with pp, yet it is senseless for the ice to circulate at speeds greater than the maximum ocean velocity when a steady state has been reached. We can make sense of this paradox by firstly assuming the existence of an interval (y1,y2)(y_1,y_2), where y1<y2<1/2y_1<y_2<1/2, in which the sea ice has yielded and σxyP=μ0p\sigma^{P}_{xy}=\mu_0p. In this interval we must have that u=uou=u_o because σxyP\sigma^{P}_{xy} is constant and therefore the ocean drag is zero. This means that y1=u1/2y_1=u_1/2. Moreover, repeating the same argument as that used for deriving (44), we assume that u=u2u=u_2 for some constant u2<1u_2<1 on (y2,1/2)(y_2,1/2) and find that u2=1−u1u_2=1-u_1 and y2=u2/2y_2=u_2/2. Now, the assumption that y1<y2<1/2y_1<y_2<1/2 will only hold for values of pp for which u1≤u2u_1\le u_2; that is, u1≤1/2u_1\le1/2, and this upper bound on u1u_1 defines a critical pressure pcp_c given by

pc=βo48ϵμ0.(45)p_c=\frac{\beta_o}{48\epsilon\mu_0}. \tag*{(45)}

For p≥pcp\ge p_c, the integral force balance along the domain must hold, see (39); as a result, u1u_1 can be at most equal to 1/21/2. Putting these results together, we may write the analytical solution to (40) as

u(y)={u10<y<u12,uo(y)y1<y<12−y1,1−u112−y1<y<12,if p<pc,(46)u(y)= \begin{cases} u_1 & 0<y<\dfrac{u_1}{2},\\ u_o(y) & y_1<y<\dfrac{1}{2}-y_1,\\ 1-u_1 & \dfrac{1}{2}-y_1<y<\dfrac{1}{2}, \end{cases} \qquad\text{if }p<p_c, \tag*{(46)}
u(y)=12if p≥pc.(47)u(y)=\frac{1}{2} \qquad\text{if }p\ge p_c. \tag*{(47)}

for y∈[0,1]y \in[0,1]. We test the validity of (4.20) by showing that it is indistinct from its numerical counterpart in Figure 6. We compute this numerical solution by regularizing the shear stress σxyα\sigma_{xy}^{\alpha} as in (36) and setting Δ=10−3\Delta= 10^{-3}.

Solutions to the complete continuum model

In section 4.2, we have seen that, given a value for the pressure pp, we can solve the momentum equation and find a velocity profile uu for the sea ice. We also prove that solutions to the momentum equation must exist and be unique. However, in general, the pressure pp is one of the unknowns in the system of equations (4.2), together with the sea ice concentration AA and the inertial number II. Here, we first present a numerical method for solving (4.2) and show that solutions to this system always exist and, under some circumstance, are unique.

A numerical method for the complete model

In order to solve the system (4.2), we follow the approach discussed in section 4.2.2 for solving the momentum equation. There, the regularized equation (36) is solved numerically with the FEM. When solving the complete system (4.2), we find that also regularizing the inertial number II improves the convergence properties of our numerical solver substantially. Therefore, we solve for

IΔ≔A0pn[∣dudy∣2+Δ2]andA=1−ϕ0IΔα(48)I_{\Delta} \coloneqq\sqrt{\frac{A_0}{p n}\left[\left\lvert\frac{\mathrm{d}u}{\mathrm{d}y}\right\rvert^2+\Delta^2\right]} \qquad\text{and}\qquad A = 1-\phi_0 I_{\Delta}^{\alpha} \tag*{(48)}

instead of (24) and (25). Then, we use the FEM to solve the system of equations given by the regularized momentum equation (36), (48), and the constraint for global concentration (26). We approximate uu with continuous piece-wise linear functions and IΔI_{\Delta} and AA with piece-wise constant functions. Our solver is implemented in Firedrake ([14]) in such a way that the complete nonlinear system is solved with Newton’s method. For small values of Δ\Delta, Newton’s method tends to fail unless a very good initial guess for the solution (u,IΔ,A,p)(u,I_{\Delta},A,p) is given. For this reason, we find that solving for a sequence of decreasing values of Δ\Delta, using the solution for the previous Δ\Delta as the initial guess for the next Δ\Delta, yields a robust solution method.

We test the sensitivity of solutions (u,IΔ,A,p)(u,I_{\Delta},A,p) to changes in Δ\Delta by solving the system for values of Δ\Delta between 10−310^{-3} and 10 and for A0=0.75A_0=0.75 and A0=0.9A_0=0.9. The numerical results, which are plotted in figure 7, indicate that the solutions become more sensitive to Δ\Delta as A0A_0 increases; this is natural, since II decreases with A0A_0 and the plasticity term becomes more important. It is also clear from these plots that, although the velocity profiles uu come very close to convergence for the smallest values of Δ\Delta, the local concentration profiles still experience visible changes around the symmetry points y=0y=0, 0.5, and 1. In these points the shear strain drops to 0 and, by the definition of II, we expect A=1A=1 there. However, as we argue below in section 5, when comparing the continuum model and the DEM, we consider such drastic changes in the local concentration to be artificial. This argument motivates the use of Δ\Delta not just as a numerical parameter that helps us solve the system numerically, but as a parameter that improves the model and may have a physical significance.

Existence and uniqueness of solutions

We conclude the analysis of the continuum model by showing that at least one solution to (2.13) must exist and, whenever u0u_0 is given by (19), is unique. To do so, we first reformulate (2.13) solely in terms of uu and pp by eliminating AA and II as follows. Substituting (25) into (26) yields

∫01Iα dy=1−A0ϕ0.(49)\int_0^1 I^\alpha\, \mathrm{d}y = \frac{1-A_0}{\phi_0}. \tag*{(49)}

Then, by substituting the definition of II, equation (24), into the integrand in the expression above, we arrive at

∫01∣dudy∣α dy=1−A0ϕ0(pnA0)α2.(50)\int_0^1 \left|\frac{\mathrm{d}u}{\mathrm{d}y}\right|^\alpha\, \mathrm{d}y = \frac{1-A_0}{\phi_0}\left(p\frac{n}{A_0}\right)^{\frac{\alpha}{2}}. \tag*{(50)}

Therefore, the system of equations (4.2) is equivalent to solving the momentum equation (23) together with the constraint (50) for uu and pp. We can interpret this problem as the minimization of the functional (28) over a set of velocity profiles subject to the constraint (50). Next, we define F:R+→R+\mathcal{F}:\mathbb{R}_{+}\to\mathbb{R}_{+} by

F(p):=∫01∣dudy∣α dy,(51)\mathcal{F}(p) := \int_{0}^{1}\left|\frac{\mathrm{d}u}{\mathrm{d}y}\right|^{\alpha}\,\mathrm{d}y, \tag*{(51)}

with uu denoting the solution to (23) with the pressure set to pp; this operator is well-defined because, given a pressure p>0p>0, a unique velocity profile uu solves the momentum equation (23). We also define C:R+→R+C:\mathbb{R}_{+}\to\mathbb{R}_{+}, given by the left hand side of (50), that is,

C(p):=1−A0ϕ0(pnA0)α2.(52)C(p) := \frac{1-A_0}{\phi_0}\left(p\frac{n}{A_0}\right)^{\frac{\alpha}{2}}. \tag*{(52)}

It is then clear that a pressure pp is part of the solution to the system of equations (4.2) if and only if

C(p)=F(p).(53)C(p)=\mathcal{F}(p). \tag*{(53)}

In § 4.4 we show that the velocity profile uu has two distinct limit behaviors. We find that uu approaches uou_o as p→0p\to0 and that du/dy\mathrm{d}u/\mathrm{d}y tends to 0 as p→∞p\to\infty. This means that

lim⁡p→0F(p)=∫01∣duody∣α dyandlim⁡p→∞F(p)=0.(54)\lim_{p\to0}\mathcal{F}(p)=\int_{0}^{1}\left|\frac{\mathrm{d}u_o}{\mathrm{d}y}\right|^{\alpha}\,\mathrm{d}y \quad\text{and}\quad \lim_{p\to\infty}\mathcal{F}(p)=0. \tag*{(54)}

On the other hand, the function CC is strictly increasing, with C(0)=0C(0)=0. Therefore, whenever ∫01∣duo/dy∣α dy>0\int_{0}^{1}\left|\mathrm{d}u_o/\mathrm{d}y\right|^{\alpha}\,\mathrm{d}y>0 (that is, uou_o is not a flat velocity profile) and F\mathcal{F} is a continuous function, we must have at least one solution pp to (53). Moreover, if F\mathcal{F} is a decreasing function, this pressure must be unique.

In Figure 8, we plot the functions CC and F\mathcal{F} for several values of A0A_0 and an ocean velocity profile given by (27). In this case, we have that

Image not extracted

Figure 8. Functions F(p)\mathcal{F}(p) and C(p)C(p), defined in (51) and (52), respectively, for different values of A0A_0. The lighter tones of blue and red are associated with lower values of A0A_0, although F\mathcal{F} hardly changes with A0A_0. A pressure pp solves (2.13) whenever C(p)=F(p)C(p)=\mathcal{F}(p) (circles). To generate these curves numerically, the same problem setting as that of Figure 5 is used.

∫01∣duody∣α dy=2α(55)\int_{0}^{1}\left|\frac{\mathrm{d}u_o}{\mathrm{d}y}\right|^{\alpha}\,\mathrm{d}y=2^{\alpha} \tag*{(55)}

and the function F\mathcal{F} appears to be strictly decreasing. This is expected, because an increase in pressure is accompanied by a flattening of the sea ice velocity profile. This means that there is a unique pp for which (53) holds; since there is only one velocity profile uu that solves the momentum equation (23), it follows that the solution to the continuum model (4.2) must be unique.

Comparison of the DEM with the continuum model

The continuum model (4.2) is designed with the objective of capturing the averaged behavior of the sea ice simulations carried out with the DEM. Here, we verify that, in the one dimensional setting of the steady ocean periodic problem, the continuum model is indeed capable of replicating most of the results of the DEM. We remind the reader that the DEM solutions can be considered steady and one-dimensional only in the sense that, for spatially and temporally-averaged data, we can expect variations in time and in the xx-direction to be negligible. At a small scale, we certainly expect the data to present variations in time and in the xx-direction, and the vertical velocity of the sea-ice to be non-zero. Moreover, DEM computations are initialized using a random floe initialization; however, as displayed in Figure 2, the (spatially and temporally-averaged) steady states that the DEM computations reach from different floe initializations are almost indistinguishable.

Throughout this section, we use the parameters in tables 1 and 2 when solving the continuum model. The continuum model is solved with the FEM as explained in section 4.3.1, setting Δ=10−3\Delta= 10^{-3} and using a uniform mesh of 300 cells.

Variation in global concentration A0A_0 and ocean speed uo,max⁡u_{o,\max}

We first evaluate the continuum model’s accuracy in replicating the DEM results used for fitting the functions μ(I)\mu(I) and Φ(I)\Phi(I) in § 3. These results are computed for the ocean velocity profile (19) and an ice thickness of H=2H = 2 m. We consider six global sea ice concentrations A0A_0 between 0.7 and 0.95 and four ocean velocities uo,max⁡u_{o,\max} between 0.1 and 1 m/s. For each of these cases, we run the DEM until a steady state is reached and then we extract 10 values of the sea ice velocity and concentration along the yy-direction, as explained in § 3, and one value for the pressure pp, averaged over the whole square patch of ocean.

Figure 9 displays the velocity and sea ice concentration profiles uu and AA and the pressure pp for both the continuum model and the DEM. The velocity profile uu and the pressure pp are normalized as explained in § 4.1 using [u]=uo,max⁡[u] = u_{o,\max}. A consequence of this normalization is that the non-dimensional solutions uu, AA, and pp to the continuum model (4.2) are indifferent to the value of uo,max⁡u_{o,\max}. For most of the DEM results, this is also the case. For each value of A0A_0, almost all of the normalized velocity profiles (panels (a), (c), (e), (g), (i), and (k)) and pressure points (panel (m)) appear to collapse onto a single curve or point, which is well approximated by the continuum model.

Departures from the other normalized DEM results are most visible for slow ocean currents and low concentrations. This becomes particularly clear when looking at the pressure in panel (m) when uo,max⁡=0.1u_{o,\max} = 0.1 m/s and A0≤0.8A_0 \le0.8, where the pressure values from the DEM (circles) depart substantially from the prediction of the continuum model (black solid line). These mismatches signal the continuum model’s limitations to capture the DEM results for the range of regimes considered. When fitting the dilatancy function Φ(I)\Phi(I), the largest misfit is also found for points of slow ocean currents and low concentration (uo,max⁡=0.1u_{o,\max} = 0.1 m/s and A0≤0.75A_0 \le0.75), see panel (b) in Figure 4. If the DEM results are to approximate a continuum model as in (4.2), we expect the equality

σxy(y)=−∫0ytox dy,(56)\sigma_{xy}(y) = - \int_{0}^{y} t_{ox}\,\mathrm{d}y, \tag*{(56)}

to hold approximately for the DEM results, with toxt_{ox} denoting the horizontal component of the ocean drag. If we denote by σxy,i\sigma_{xy,i} and tox,it_{ox,i} the values extracted from the DEM at the grid cell of width Δy\Delta y located at yiy_i (see appendix A.§ 2 for an overview of how these quantities are obtained), a discrete balance analogous to (56) is

σxy,i=−Δy∑j=1itox,j.(57)\sigma_{xy,i} = -\Delta y \sum_{j=1}^{i} t_{ox,j}. \tag*{(57)}

We remark that the term tox,jt_{ox,j} results from averaging the ocean drag on each ice floe, as opposed to introducing the averaged sea ice velocity into (4). We also note that (57) ignores any choice of rheology and should hold independently of our choice of functions μ(I)\mu(I) and Φ(I)\Phi(I). A surprising result we find when investigating the DEM data is that the terms to the left and right hand side of the equality in (57) differ in orders of magnitude whenever the sea ice concentration is low and the ocean currents are slow, see panel (a) of Figure 10. Conversely, for faster ocean currents and denser concentrations, these terms become approximately equal (see panels (c)–(f) in Figure 10). Inertial effects are found to be negligible in all of the cases we consider; moreover, since DEM quantities are averaged along the whole length of the domain in the xx-direction, the term corresponding to ∂σxx/∂x\partial\sigma_{xx}/\partial x, which should be considered in a two-dimensional setting, becomes zero. Therefore, this finding raises the questions of whether a fundamentally different continuum model should be used for low sea ice concentrations and slow ocean currents or if a continuum model is valid at all.

Figure 10. Shear stress σxy,i\sigma_{xy,i} and integrated ocean drag ∫tox,i:=Δy∑j=1itox,j\int t_{ox,i} := \Delta y \sum_{j=1}^{i} t_{ox,j} extracted from the DEM for A0=0.7A_0=0.7 and 0.90.9 and uo,max⁡=0.1u_{o,\max}=0.1, 0.50.5 and 11 m/s.

The DEM’s local concentration AA is accurately captured by the continuum model for low sea ice concentrations (panels (b), (d), (f), and (h) in Figure 9). For A0=0.90A_0=0.90 and 0.950.95 (panels (j) and (l)), the continuum model overestimates the degree of dilatancy, although the general trends are visibly similar, with regions of higher concentration around y=0y=0, 1/21/2, and 11, where the strain rates are lowest. Since an increase in Δ\Delta diminishes the local variations in AA (see Figure 7), this suggests that Δ\Delta could be adjusted to improve the fit with the data. In this case, it would enter the model as a physical parameter whose interpretation should be explored further.

Image not extracted

Figure 7. Solutions uu and AA to the continuum model (4.2) with A0=0.75A_0=0.75 ((a),(b)) and A0=0.9A_0=0.9 ((c),(d)), regularized with a parameter Δ\Delta as explained in section 4.3.1. In (e) we plot the variation of the pressure pp with Δ\Delta. We use the same parameter values as those used for figure 5.

In panel (m) of Figure 9, we also test two limit approximations of the pressure which we find in § 4 when uou_o is given by 4.3. When p≪1p \ll1, which occurs for the smaller values of A0A_0, we expect C(p)≈2αC(p) \approx2^{\alpha} because u≈uou \approx u_o. This implies that

p≈4A0n(ϕ01−A0)2/αas A0→0.(58)p \approx4\frac{A_0}{n}\left(\frac{\phi_0}{1-A_0}\right)^{2/\alpha}\quad\text{as }A_0\to0. \tag*{(58)}

On the other hand, high values of A0A_0 result in I≪1I \ll1, see Figure 4, and therefore μ(I)≈μ0\mu(I) \approx\mu_0, leading to the purely plastic regime studied in § 4.5. Panel (k) shows that the velocity profiles are mostly flat in this region, suggesting that the critical pressure pcp_c calculated in (4.19) is a good approximation of the pp. By plotting (5.3) and pcp_c in panel (m), we find that these are indeed good approximations in their respective limits.

Variation in ice thickness and number of floes

Next, we test the effectiveness of the continuum model in capturing the DEM results for different ice thicknesses and floe sizes. In Figure 11, we plot results for the DEM and the continuum model for steady states with H=0.5H = 0.5 and H=4H = 4 m, different ocean velocities uo,max⁡u_{o,\max}, and a global sea ice concentration of A0=0.8A_0 = 0.8. The ice thickness HH enters the continuum model via the parameter ϵ=H/L\epsilon= H/L in the momentum equation (23). An increase in HH is accompanied by an increase in ϵ\epsilon, which effectively acts as a decrease in the normalized external forcing. We therefore expect an increase in HH to result in a decrease in the shear strain and pressure. As observed in panels (a) and (c) of Figure 11, this is the case for the velocity profiles of both the continuum model and the DEM; an increase in HH is accompanied by a flattening of the normalized velocity profiles. The continuum model is once again accurate in capturing the averaged velocities of the DEM and the general trends in the variation of the sea ice concentration. The pressure resulting from the continuum model and the DEM also decreases as the ice thickness increases, see panel (f); for the pressure, we find a decent fit between the DEM and the continuum model.

Figure 11. Solutions to the non-dimensional continuum model (4.2) (black lines) compared with results from the DEM (markers). For each pair of panels in the first row, we fix the sea ice thickness HH and vary the maximum ocean velocity uo,max⁡u_{o,\max} (the ocean velocity is plotted in blue in panels (a) and (c)). In panel (e), we plot the pressure pp in terms of HH. Due to the normalization in terms of [u]=uo,max⁡[u] = u_{o,\max}, the solutions to the continuum model are indifferent to a change in uo,max⁡u_{o,\max}.

Finally, Figure 12 contains a comparison between the DEM and the continuum model for different numbers of floes nn. In particular, we present results with n=500n = 500, 20002000 and 50005000 for A0=0.8A_0 = 0.8 and 0.90.9 and uo,max⁡=0.5u_{o,\max} = 0.5 m/s. An increase in nn implies a decrease in the effective viscosity of the model, and we therefore expect a decrease in the ice strength or pressure pp; in the limit where n→∞n \to\infty, the rheology becomes purely plastic, this can be seen by taking this limit in (23). In Figure 12, one can see that the pressure in continuum model indeed decreases with nn, but this is not the case with the DEM. In fact, the results of the DEM appear to change little with nn, especially for A0=0.9A_0 = 0.9. Despite this, the results from the DEM do not depart from those of the continuum model excessively.

Comparisons with existing models for sea ice

The continuum model studied in this paper shares features with existing models for sea ice. Here, we examine those similarities and also establish differences with our continuum model. Due to its ubiquity in sea ice modeling, we first consider Hibler’s model in § 6.1, before considering other models in § 6.2.

Hibler's model

The most widely used model for sea ice is Hibler’s model, which was first presented in Hibler III [17] and treats sea ice like a viscoplastic material. Under the one dimensional conditions of the steady square ocean patch and the non-dimensionalization in § 4, Hibler’s model reduces to the following form:

−ϵddy(p2edu/dy∣du/dy∣2+δ2)=βo∣uo−u∣(uo−u)(59)-\epsilon\frac{\mathrm{d}}{\mathrm{d}y}\left(\frac{p}{2e}\frac{\mathrm{d}u/\mathrm{d}y}{\sqrt{\lvert\mathrm{d}u/\mathrm{d}y\rvert^2+\delta^2}}\right)=\beta_o\lvert u_o-u\rvert(u_o-u) \tag*{(59)}
p=P∗ρiuo,max⁡2exp⁡(−20(1−A)).(60)p=\frac{P^*}{\rho_i u_{o,\max}^2}\exp\left(-20(1-A)\right). \tag*{(60)}

In (6.1), the parameter P∗P^* is an empirical constant, δ\delta is a regularization parameter, and e=2e = 2 represents the eccentricity of the elliptical yield curve from which Hibler’s model is derived. That is, Hibler’s model assumes a plastic rheology based on an elliptical yield curve which is then regularized. Inside the plane of principal stresses, the yield curve is set in such a way that the ice only resists compression, not pure extension. This geometrical configuration is motivated by the observation that, in pack ice, deformation occurs through ridging (compression) and the opening of leads (extension); of these two mechanisms, only ridging requires a non-negligible amount of plastic work [30].

In the steady one dimensional setting, equation (59) can be recovered from the regularized momentum equation (36) by setting μ0=1/(2e)\mu_0 = 1/(2e) and μ1=0\mu_1 = 0 (pure plasticity). In fact, 1/(2e)=0.251/(2e) = 0.25, which comes very close to the value μ0=0.26\mu_0 = 0.26 which we infer from the DEM. This means that close to the quasi-static regime, when I≪1I \ll1 and μ(I)≈μ0\mu(I) \approx\mu_0, we essentially work with the momentum equation from Hibler’s model. This is a very reasonable coincidence, because Hibler’s model was designed for the central ice pack where the sea ice concentration is very high.

The main departure between our model and Hibler’s is the expression for the pressure (60): as uo,max⁡u_{o,\max} increases, pp decreases in Hibler’s model, unlike our continuum model where pp is independent of uo,max⁡u_{o,\max}. This makes it impossible for Hibler’s model to capture the invariance of the non-dimensional sea ice velocity and pressure with uo,max⁡u_{o,\max} which we find in most of the DEM results in § 5. This is made clear in Figure 13, where we show solutions to Hibler’s model and compare them with the DEM for different values of A0A_0 and uo,max⁡u_{o,\max}; an increase in uo,max⁡u_{o,\max} is accompanied by excessively large changes in the velocity profile. We set P∗=5×104 N m−1P^* = 5 \times10^4\,\mathrm{N}\,\mathrm{m}^{-1} and δ=0.1\delta= 0.1 in order to get a good fit with the DEM for some values of uo,max⁡u_{o,\max}. We remark that Hibler’s model is designed for ice that fractures and ridges; in this setting, an increase in the ocean drag weakens the ice through this mechanical deformation. Below, in § 7, we state that future extensions of our continuum model should include the effects of fracturing and ridging, available in SubZero. These future investigations should study the validity of (60) when fracturing and ridging are accounted for.

Figure 13. Solutions to Hibler’s model for different global concentrations A0A_0 and ocean velocities uo,max⁡u_{o,\max}. We also plot the results from the DEM as a reference. Here, we set H=2 mH = 2\,\mathrm{m} and n=2000n = 2000.

The choice of an elliptical yield curve in Hibler’s model is motivated by the simplicity of the resulting rheological formulation, yet other possibilities consistent with the requirement of null resistance to pure extension are available, such as the parabolic lens and the teardrop yield curves [36, 28] and the Mohr-Coulomb criterion (Ip et al. 1991). In fact, as mentioned at the end of § 3, the purely plastic rheology in the μ(I)\mu(I) formulation considered here (given by (3.2) with $\mu_1 = 0) can be written as the Mohr-Coulomb rheology presented in Ip et al. (1991) and [13]. This becomes clear if we note that ∥S∥=D1−D2\lVert\mathbf{S} \rVert= D_1 - D_2, where D1D_1 and D2D_2 denote the principal components (eigenvalues) of D\mathbf{D}, and we compare the plastic component in (21) with expression (6) in [13]. A significant difference between the two expressions is that in (6) from Gutfraind & Savage [13] the viscosity is capped to a maximum value in order to avoid the inevitable blow-up that occurs with a purely plastic rheology; this is another form of regularization. This difference of plastic yield curves between Hibler’s model and the purely plastic version of our model is another point of departure between both models in a two dimensional setting.

Other models

In section 3, we explain that the viscous component in (21) can be interpreted as the rheological effects of collisions, which become increasingly important as II increases. As mentioned in that section, a collisional rheology is derived in Shen et al. [33] which is also of a linearly viscous nature. In fact, this collisional rheology establishes a viscosity which is linearly proportional to dˉρiHp\bar{d}\sqrt{\rho_i H p}, as in (21) (see Herman [16] for a clear description of the collisional rheology). For this, the model proposed in Feltham [8] comes close to (21), since it takes the ice rheology to be the sum of Hibler’s rheology and the collisional rheology from Shen et al. [33].

To the knowledge of the authors, the only other study using the μ(I)\mu(I) rheology to model sea ice is due to Herman [16]. There, the floes are driven by a moving wall (as opposed to an ocean or atmospheric current, as in our case), and the DEM is based on disk shaped floes, with a much more severe polydispersity. Two main differences can be found between the function μ(I)\mu(I) that we infer (Figure 4) and the one found in Herman [16]. First, although both cases consider a very similar range of II values, the magnitude of the friction μ\mu differs considerably, although it is of the same order of magnitude. Second, the μ(I)\mu(I) curve in Herman [16] plateaus for I>10−1I > 10^{-1}, something that we do not observe. It remains unclear what may cause these differences. Regarding the second point, we justify our use of a linear function for μ\mu, as in (20), by remarking that it simplifies the resulting constitutive equation, enabling a more thorough analysis of the model, and resembles the μ\mu function proposed in Herman [16] over a large range of II values.

Conclusions and future work

In this paper, we have presented a novel mathematical model for sea ice in the MIZ based on the μ(I)\mu(I) rheology. We have inferred the form of this rheology in section 3 using data produced with DEM computations. With the analysis in section 4, we prove that the steady one dimensional formulation of this model, given by (4.2), is well-posed in the sense that it has a unique solution. The numerical results presented in section 5 demonstrate that this model is capable of replicating most of the results of the DEM in the context of steady one dimensional problems. The most visible departure from the continuum model occurs for low ocean velocities and sea ice concentrations, with uo,max⁡=0.1u_{o,\max} = 0.1 m/s and A0≤0.85A_0 \le0.85; in this case, the DEM results indicate that the shear stress no longer balances the integrated ocean drag, signaling a breakdown of the underlying grid-averaged momentum balance equation. That is, since inertial effects are found to be negligible in the steady states we consider and the extensional stresses cancel out when integrating across the xx-direction of the domain to perform the averaging of DEM quantities, our basic continuum model, prior to any choice of rheology, establishes a balance between shear stresses and ocean drag. In Figure 10, this balance is found to approximately hold for all cases but those of slow ocean currents and low sea ice concentrations.

The lack of validity of the continuum model for slow ocean currents and low concentrations found in section 5 should be explored further, since this is a regime we expect to find in parts of the MIZ ([34]). As explained in section 5, this lack of accuracy is accompanied by the breakdown of the balance between the averaged shear stress and the integrated ocean drag extracted from the DEM. This balance precedes the choice of rheology and therefore indicates that either a fundamentally different continuum model should be used or some assumption necessary for a continuum model to even hold is no longer valid.

Mechanical processes like fracturing, ridging, and welding are fundamental processes in sea ice dynamics [9]. Therefore, future improvements of the model presented here should consider the inclusion of these effects. Since ridging and fracturing are fundamental processes in the central ice pack [30], our continuum model could also be valid in this area when these effects are included, yielding a unified sea ice model. Since SubZero includes these mechanical interactions between ice floes [22], it is a promising tool for exploring their macroscopic effects on the rheology. A first step could be to explore the consequences of including floe fracturing. We expect this would set an upper bound on the pressure and shear stress that the ice cover can sustain. This bound will probably be closely related to the fracture criterion used at the ice floe level. Moreover, floe fracturing will create smaller floes that lead to higher degrees of polydispersity.

The comparisons between the continuum model and the DEM in § 5 are confined to the context of the steady periodic ocean problem, which is essentially one dimensional. Any future studies must examine the accuracy of the μ(I)\mu(I) rheology in modeling sea ice dynamics under unsteady conditions and in a two-dimensional configuration. Barker et al. [2] discovered the μ(I)\mu(I) formulation presented in § 2.1 to be mathematically ill-posed in time-dependent problems, and Schaeffer et al. [32] has proposed modifications that avoid these instabilities while leaving the steady equations unchanged. A future investigation in the context of sea ice should take these studies into account.

The rheology we propose in (21) is local in the sense that the viscosity and the pressure at a certain point of the domain only depend on other quantities and their derivatives at that same point. Yet, granular materials create complex contact networks that enable the interaction of grains set far apart. This leads to non-local effects, and extensions of the μ(I)\mu(I) rheology that include these effects have been proposed in the context of granular flows [21, 3]. Such effects will probably also arise in sea ice modeling and should be considered in future extensions of the model we propose here.

Appendix A

Computing the stress tensor for an ice floe

At a given instant in time, the floe ii is in contact with floes whose indices are given by the set CiC_i. For each floe j∈Cij \in C_i, there are ni,jcn_{i,j}^{c} contact points (there can be several contact points between two floes if one of them is concave). The stress tensor σi\boldsymbol{\sigma}_i of floe ii is given by

σi=1ai∑j∈Ci∑k=1ni,jcfi,jk⊗ri,jk,(61)\boldsymbol{\sigma}_i = \frac{1}{a_i}\sum_{j \in C_i}\sum_{k=1}^{n_{i,j}^{c}} \boldsymbol{f}_{i,j}^{k} \otimes\boldsymbol{r}_{i,j}^{k}, \tag*{(61)}

where aia_i is the area of floe ii, fi,jk\boldsymbol{f}_{i,j}^{k} is the force at the kkth contact point exerted by floe jj on floe ii, and ri,jk\boldsymbol{r}_{i,j}^{k} is the vector connecting the center of mass of floe ii with the kk-th contact point. Expression (A 1) corresponds with the Love–Weber formula and, in general, additional terms corresponding to dynamics effects must also be accounted for [25]. However, we find these dynamic terms to be negligible in all of our computations. We also remark that (A 1) differs from its counterpart in [22 equation (9] in two points: (1) we divide by the floe area aia_i rather than its volume aiHia_i H_i to obtain the right units and account for the fact that we are working with depth-integrated stresses. (2) We avoid forcing σi\boldsymbol{\sigma}_i to be symmetric and simply use the Love-Weber formula. In general, we find that σi,xy≈σi,yx\sigma_{i,xy} \approx\sigma_{i,yx} in all of our DEM computations.

For the convenience of the reader, we now summarize the calculation of contact forces between two colliding floes in SubZero. A complete account is given in [22]. The contact force fi,jk\boldsymbol{f}_{i,j}^{k} is the sum of its normal and tangential components,

fi,jk=fi,jN,k+fi,jT,k.(62)\boldsymbol{f}_{i,j}^{k}=\boldsymbol{f}_{i,j}^{N,k}+\boldsymbol{f}_{i,j}^{T,k}. \tag*{(62)}

In Figure 14 we represent the parameters and vectors involved in the collision of two floes. Two colliding floes intersect; this intersection results in an overlap polygon, represented in grey in Figure 14, of area A\mathcal{A} and center of mass cc. The point cc is then considered the contact point between floes ii and jj. The normal direction ni,jk\boldsymbol{n}_{i,j}^{k} from floe jj to ii is defined perpendicular to the chord uniting the two intersection points between both floes, as in Figure 14. The normal force is then defined as

Figure 14. Collision between two floes. The contact forces are calculated in terms of geometric properties of the overlap area, shown in gray. Here, cc represents the center of mass of the overlap area and the normal direction ni,jk\boldsymbol{n}_{i,j}^{k} is taken normal to the line connecting the two intersection points between the floes.

fi,jN,k=κAni,jk,(63)\boldsymbol{f}_{i,j}^{N,k}=\kappa\mathcal{A}\boldsymbol{n}_{i,j}^{k}, \tag*{(63)}

where

κ=EHiHjHidi+Hjdj.(64)\kappa=E\frac{H_iH_j}{H_id_i+H_jd_j}. \tag*{(64)}

In (A 4), EE is Young’s modulus, HiH_i the floe thickness, and di=aid_i=\sqrt{a_i} a measure of the floe size. Normal forces are thus elastic and do not dissipate energy.

The tangential force is given by

fi,jT,k=ci,jkG Δt vi,jk∣fi,jN,k∣ti,jk,(65)\boldsymbol{f}_{i,j}^{T,k}=c_{i,j}^{k}G\,\Delta t\,v_{i,j}^{k}\lvert\boldsymbol{f}_{i,j}^{N,k}\rvert\boldsymbol{t}_{i,j}^{k}, \tag*{(65)}

where ci,jkc_{i,j}^{k} is the length of the chord uniting the two intersection points between floes ii and jj and GG is the shear modulus G=E/(2(1+ν))G=E/(2(1+\nu)), defined in terms of Poisson’s ratio ν\nu. The parameter Δt\Delta t is the simulation’s time step, vi,jkv_{i,j}^{k} the tangential velocity difference between both floes, and ti,jk\boldsymbol{t}_{i,j}^{k} the tangential direction. The tangential force is limited to the following upper bound:

∣fi,jT,k∣≤μ∗∣fi,jN,k∣,(66)\lvert\boldsymbol{f}_{i,j}^{T,k}\rvert\leq\mu^{*}\lvert\boldsymbol{f}_{i,j}^{N,k}\rvert, \tag*{(66)}

where μ∗\mu^{*} is the inter-floe friction coefficient.

Spatial averaging of data

To average the DEM data in space, we must grid the square domain. Taking into account the one dimensional nature of the problem, we divide the domain into N=10N = 10 cells that stretch in the xx-direction, such that the edges separating these cells are defined along N+1N + 1 equispaced points (y0,y1,…,yN)(y_0,y_1,\ldots,y_N) in the yy-direction. Hence, the resulting grid consists in the cells [0,L]×[yi−1,yi][0,L] \times[y_{i-1},y_i] for i=1,…,Ni = 1,\ldots,N, as depicted in Figure 15. At each time step, we compute the horizontal velocity uu and the components of the stress tensor σ\boldsymbol{\sigma} by spatially averaging the velocities and stresses of the ice floes contained in each region (the manner in which the stress tensor σ\boldsymbol{\sigma} is computed for each ice floe is explained in above in appendix A.§ 1. In particular, for each cell of the grid, we perform a mass-weighted averaging such that, for ice floes that are only partially contained in the cell, only the mass of the floe inside the cell is considered. More information about the averaging can be found in Manucharyan & Montemuro [22]. In order to be consistent with (15), for each run we extract a single pressure p=12(σxx+σyy)p = \frac{1}{2}(\sigma_{xx} + \sigma_{yy}) by spatially averaging this quantity over the whole domain.

Figure 15. The values of the horizontal velocity uu and the stress tensor σ\boldsymbol{\sigma} are extracted from the DEM data for each time step by averaging spatially over the 10 regions, elongated in the xx-direction, shown here. In this figure, n=2000n = 2000, uo,max⁡=2u_{o,\max} = 2 m/s, and A0=0.8A_0 = 0.8.

By performing the spatial and temporal averaging, for each DEM computation we obtain a pressure p∈Rp \in\mathbb{R} and the vectors (ui)(u_i), (σxy,i)(\sigma_{xy,i}), and (Ai)(A_i) of data points in RN\mathbb{R}^N. To compute the inertial numbers II, we first find the strain rate vector (dui)(\mathrm{d}u_i) via central finite differences, such that dui=(ui+1−ui−1)/(2(yi+1−yi−1))\mathrm{d}u_i = (u_{i+1} - u_{i-1})/(2(y_{i+1} - y_{i-1})). Then, Ii=dˉHρi/p∣dui∣I_i = \bar{d}\sqrt{H\rho_i}/p|\mathrm{d}u_i|, with dˉ=A0L2/n\bar{d} = \sqrt{A_0L^2/n}.

Appendix B

Since the momentum equation (4.8) of the continuum model is equivalent to the minimization of the functional J\mathcal{J} over the space VV, defined in (29), it suffices to show that J\mathcal{J} admits a unique minimizer to prove the existence and uniqueness of solutions to the momentum equation. To do so, we first write the functional J:V→R\mathcal{J}: V \to\mathbb{R} as follows for simplicity:

J(v)=γ1∣v∣1+γ2∣v∣22+γ3∥uo−v∥33(67)\mathcal{J}(v) = \gamma_1 |v|_1 + \gamma_2 |v|_2^2 + \gamma_3 \lVert u_o - v \rVert_3^3 \tag*{(67)}

where γi>0\gamma_i > 0 for i=1,2,3i = 1,2,3 are constants and, for q≥1q \ge1, ∥⋅∥q\lVert\cdot\rVert_q and ∣⋅∣q|\cdot|_q are the Sobolev norm and semi-norms, respectively, for the Lq((0,1))L^q((0,1)) and W1,q((0,1))W^{1,q}((0,1)) spaces; more precisely, these are defined by

∥v∥q=(∫01∣v∣q dx)1/qand∣v∣q=(∫01∣dvdx∣q dx)1/q.(68)\lVert v \rVert_q = \left(\int_0^1 |v|^q\,\mathrm{d}x\right)^{1/q} \quad\text{and}\quad |v|_q = \left(\int_0^1 \left|\frac{\mathrm{d}v}{\mathrm{d}x}\right|^q\,\mathrm{d}x\right)^{1/q}. \tag*{(68)}

We first remark that J\mathcal{J} is well defined for all v∈Vv \in V by the Sobolev embedding theorem, which implies that ∥v∥3<∞\lVert v\rVert_{3} < \infty. Moreover, for the sake of rigor, we must also assume that uo∈L3((0,1))u_o \in L^{3}((0,1)).

According to Evans [7], at least one function VV exists that minimizes J\mathcal{J} if the functional is convex and coercive. It is straightforward to check that J\mathcal{J} is convex. The functional J\mathcal{J} is said to be coercive in VV if

∥v∥V→∞  ⟹  J(v)→∞,(69)\lVert v\rVert_{V} \to\infty\implies\mathcal{J}(v) \to\infty, \tag*{(69)}

where ∥v∥V2=∥v∥22+∣v∣22\lVert v\rVert_{V}^{2} = \lVert v\rVert_{2}^{2} + |v|_{2}^{2}. We first note that, by the triangle inequality and Young’s inequality, it follows that

∥v∥33≤4(∥uo−v∥33+∥uo∥33).(70)\lVert v\rVert_{3}^{3} \leq4\left(\lVert u_o-v\rVert_{3}^{3}+\lVert u_o\rVert_{3}^{3}\right). \tag*{(70)}

Therefore, for any v∈Vv \in V,

J(v)≥γ1∣v∣1+γ2∣v∣22+γ34∥v∥33−γ3∥uo∥33(71)\mathcal{J}(v) \geq\gamma_1 |v|_1+\gamma_2 |v|_2^{2}+\frac{\gamma_3}{4}\lVert v\rVert_{3}^{3}-\gamma_3\lVert u_o\rVert_{3}^{3} \tag*{(71)}

Using Hölder’s inequality we can show that ∥v∥3≥∥v∥2\lVert v\rVert_{3} \geq\lVert v\rVert_{2}, such that

J(v)≥γ2∣v∣22+γ34∥v∥33−γ3∥uo∥33,(72)\mathcal{J}(v) \geq\gamma_2 |v|_2^{2}+\frac{\gamma_3}{4}\lVert v\rVert_{3}^{3}-\gamma_3\lVert u_o\rVert_{3}^{3}, \tag*{(72)}

from where coercivity follows. We note that this argument fails whenever γ2=0\gamma_2=0, which corresponds with the purely plastic regime, because ∣v∣1→∞|v|_1 \to\infty does not necessarily follow from ∥v∥V→∞\lVert v\rVert_{V} \to\infty.

To prove that there is only one function that minimizes J\mathcal{J}, we assume by contradiction that both u1u_1 and u2u_2 minimize J\mathcal{J} and u1≠u2u_1 \ne u_2. In this case, we must have that J(u1)=J(u2)\mathcal{J}(u_1)=\mathcal{J}(u_2). We define w=1/2(u1+u2)w=1/2(u_1+u_2) and note that

∥uo−w∥33<12∥uo−u1∥33+12∥uo−u2∥33,(73)\lVert u_o-w\rVert_{3}^{3}<\frac{1}{2}\lVert u_o-u_1\rVert_{3}^{3}+\frac{1}{2}\lVert u_o-u_2\rVert_{3}^{3}, \tag*{(73)}

due to the strict convexity of the function ∥⋅∥33\lVert\cdot\rVert_{3}^{3} in R\mathbb{R} and the injectivity of ∥uo−⋅∥33\lVert u_o-\cdot\rVert_{3}^{3} in VV. As a result, by appealing to the convexity of the seminorm ∥⋅∥q\lVert\cdot\rVert_q for all q≥1q \geq1,

J(w)<12J(u1)+12J(u2)=J(u1),(74)\mathcal{J}(w)<\frac{1}{2}\mathcal{J}(u_1)+\frac{1}{2}\mathcal{J}(u_2)=\mathcal{J}(u_1), \tag*{(74)}

a contradiction because J(u1)≤J(v)\mathcal{J}(u_1)\leq\mathcal{J}(v) for all v∈Vv \in V.

Acknowledgements. We thank Andrew Thompson for many insightful comments and suggestions shared with us during the preparation of the manuscript.

Funding. This work has been supported by the Multidisciplinary University Research Initiatives (MURI) Program, Office of Naval Research (ONR) grant number N00014-19-1-242.

Declaration of Interests. The authors report no conflict of interest.

Author contributions. All authors contributed to conceiving and designing the study. GD wrote the paper, designed the figures, performed the simulations and mathematical analysis. MG, SG and GS revised the paper. MG, RH and SG contributed to the simulations, and SG contributed simulation tools. MG, GS and RH provided input to the analysis.

References

  1. [1]Adams, Robert A & Fournier, John JF 2003 Sobolev Spaces. Pure and Applied Mathematics 140. Academic Press, Elsevier.
  2. [2]Barker, Thomas, Schaeffer, David G, Bohórquez, Patricio & Gray, JMNT 2015 Well-posed and ill-posed behaviour of the μ(I)-rheology for granular flow. Journal of Fluid Mechanics 779, 794–818.DOI
  3. [3]Bouzid, Mehdi, Trulsson, Martin, Claudin, Philippe, Clément, Eric & Andreotti, Bruno 2013 Nonlocal rheology of granular flows across yield conditions. Physical review letters 111 (23), 238301.pubmed.ncbi.nlm.nih.gov/24476308
  4. [4]Da Cruz, Frédéric, Emam, Sacha, Prochnow, Michaël, Roux, Jean-Noël & Chevoir, François 2005 Rheophysics of dense granular materials: Discrete simulation of plane shear flows. Physical Review E 72 (2), 021309.DOI
  5. [6]Dumont, Dany 2022 Marginal ice zone dynamics: history, definitions and research perspectives. Philosophical Transactions of the Royal Society A 380 (2235), 20210253.DOI
  6. [7]Evans, Lawrence C 2022 Partial differential equations, , vol. 19. American Mathematical Society.
  7. [8]Feltham, Daniel L 2005 Granular flow in the marginal ice zone. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 363 (1832), 1677–1700.pubmed.ncbi.nlm.nih.gov/16011939
  8. [9]Feltham, Daniel L 2008 Sea ice rheology. Annu. Rev. Fluid Mech. 40, 91–112.DOI
  9. [11]Glowinski, R., Lions, J. L. & Trémolières, R. 1981 Numerical Analysis of Variational Inequalities. North Holland.
  10. [12]Gray, J. M. N. T & Morland, L. W. 1994 A two-dimensional model for the dynamics of sea ice. Philosophical Transactions of the Royal Society of London. Series A: Physical and Engineering Sciences 347 (1682), 219–290.DOI
  11. [13]Gutfraind, Ricardo & Savage, Stuart B 1997 Marginal ice zone rheology: Comparison of results from continuum-plastic models and discrete-particle simulations. Journal of Geophysical Research: Oceans 102 (C6), 12647–12661.
  12. [14]Ham, David A., Kelly, Paul H. J., Mitchell, Lawrence, Cotter, Colin J., Kirby, Robert C., Sagiyama, Koki, Bouziani, Nacime, Vorderwuelbecke, Sophia, Gregory, Thomas J., Betteridge, Jack, Shapero, Daniel R., Nixon-Hill, Reuben W., Ward, Connor J., Farrell, Patrick E., Brubeck, Pablo D., Marsden, India, Gibson, Thomas H., Homolya, Miklós, Sun, Tianjiao, McRae, Andrew T. T., Luporini, Fabio, Gregory, Alastair, Lange, Michael, Funke, Simon W., Rathgeber, Florian, Bercea, Gheorghe-Teodor & Markall, Graham R. 2023 Firedrake User Manual, first edition edn. Imperial College London and University of Oxford and Baylor University and University of Washington.DOI
  13. [15]Herman, Agnieszka 2011 Molecular-dynamics simulation of clustering processes in sea-ice floes. Physical Review E 84 (5), 056104.DOI
  14. [16]Herman, Agnieszka 2022 Granular effects in sea ice rheology in the marginal ice zone. Philosophical Transactions of the Royal Society A 380 (2235), 20210260.DOI
  15. [17]Hibler III, WD 1979 A dynamic thermodynamic sea ice model. Journal of physical oceanography 9 (4), 815–846.doi.org/10.1175/1520-0485(1979)009<0815:adtsim>2.0.co;2
  16. [19]Jenkins, James T & Savage, Stuart B 1983 A theory for the rapid flow of identical, smooth, nearly elastic, spherical particles. Journal of fluid mechanics 130, 187–202.
  17. [20]Jop, Pierre, Forterre, Yoël & Pouliquen, Olivier 2006 A constitutive law for dense granular flows. Nature 441 (7094), 727–730.DOI
  18. [21]Kamrin, Ken & Koval, Georg 2012 Nonlocal constitutive relation for steady granular flow. Physical review letters 108 (17), 178301.pubmed.ncbi.nlm.nih.gov/22680912
  19. [22]Manucharyan, Georgy E & Montemuro, Brandon P 2022 Subzero: A sea ice model with an explicit representation of the floe life cycle. Journal of Advances in Modeling Earth Systems 14 (12), e2022MS003247.
  20. [23]Midi, GDR 2004 On dense granular flows. The European Physical Journal E 14, 341–365.
  21. [24]Nedderman, Ronald Midgley 1992 Statics and kinematics of granular materials. Cambridge University Press.DOI
  22. [25]Nicot, François, Hadda, Nejib, Guessasma, Mohamed, Fortin, Jerome & Millet, Olivier 2013 On the definition of the stress tensor in granular media. International Journal of Solids and Structures 50 (14-15), 2508–2517.DOI
  23. [26]Pouliquen, Olivier, Cassar, Cyril, Jop, Pierre, Forterre, Yoel & Nicolas, Maxime 2006 Flow of dense granular material: towards simple constitutive laws. Journal of Statistical Mechanics: Theory and Experiment 2006 (07), P07020.DOI
  24. [28]Ringeisen, Damien, Losch, Martin & Tremblay, L Bruno 2023 Teardrop and parabolic lens yield curves for viscous-plastic sea ice models: New constitutive equations and failure angles. Journal of Advances in Modeling Earth Systems 15 (9), e2023MS003613.DOI
  25. [29]Ringeisen, Damien, Losch, Martin, Tremblay, L Bruno & Hutter, Nils 2019 Simulating intersection angles between conjugate faults in sea ice with different viscous-plastic rheologies. The Cryosphere 13 (4), 1167–1186.DOI
  26. [30]Rothrock, D. A. 1975 The energetics of the plastic deformation of pack ice by ridging. Journal of Geophysical Research 80 (33), 4514–4519.DOI
  27. [31]Savage, Stuart B 1984 The mechanics of rapid granular flows. Advances in applied mechanics 24, 289–366.DOI
  28. [32]Schaeffer, DG, Barker, Thomas, Tsuji, Daisuke, Gremaud, Pierre, Shearer, Michael & Gray, JMNT 2019 Constitutive relations for compressible granular flow in the inertial regime. Journal of Fluid Mechanics 874, 926–951.DOI
  29. [33]Shen, Hayley H, Hibler III, William D & Leppäranta, Matti 1987 The role of floe collisions in sea ice rheology. Journal of Geophysical Research: Oceans 92 (C7), 7085–7096.DOI
  30. [34]Stewart, Andrew L, Klocker, Andreas & Menemenlis, Dimitris 2019 Acceleration and overturning of the Antarctic Slope Current by winds, eddies, and tides. Journal of Physical Oceanography 49 (8), 2043–2074.DOI
  31. [35]Wood, David Muir 1990 Soil behaviour and critical state soil mechanics. Cambridge university press.DOI
  32. [36]Zhang, Jinlun & Rothrock, D. A. 2005 Effect of sea ice rheology in numerical investigations of climate. Journal of Geophysical Research: Oceans 110 (C8).DOI

Paper details

vlnc:W_7ktfdc2n
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents