betaFeedback

WorksMathematics

Finite element methods for multicomponent convection-diffusion

Abstract

We develop finite element methods for coupling the steady-state Onsager–Stefan–Maxwell equations to compressible Stokes flow. These equations describe multicomponent flow at low Reynolds number, where a mixture of different chemical species within a common thermodynamic phase is transported by convection and molecular diffusion. Developing a variational formulation for discretizing these equations is challenging: the formulation must balance physical relevance of the variables and boundary data, regularity assumptions, tractability of the analysis, enforcement of thermodynamic constraints, ease of discretization, and extensibility to the transient, anisothermal, and non-ideal settings. To resolve these competing goals, we employ two augmentations: the first enforces the mass-average constraint in the Onsager–Stefan–Maxwell equations, while its dual modifies the Stokes momentum equation to enforce symmetry. Remarkably, with these augmentations we achieve a Picard linearization of symmetric saddle point type, despite the equations not possessing a Lagrangian structure. Exploiting the structure of linear irreversible thermodynamics, we prove the inf-sup condition for this linearization, and identify finite element function spaces that automatically inherit well-posedness. We verify our error estimates with a numerical example, and illustrate the application of the method to non-ideal fluids with a simulation of the microfluidic mixing of hydrocarbons.

Introduction

Many fluids consist of mixtures; for example, air is a mixture of nitrogen, oxygen, carbon dioxide, and other species. In many situations, it is not necessary to resolve the motions of the individual species, such as when modelling the flow of air over an aircraft. However, in other contexts, detailed knowledge of the transport of individual species is required. Examples include biological applications, where one may be interested in the transport of oxygen and carbon dioxide in blood, in chemical engineering, where one may be interested in separating or combining the constituents of petroleum, or in electrochemistry, where the performance of a lithium-ion battery is often limited by the transport of lithium ions within an electrolyte. We describe this situation as a multicomponent flow, where a fluid is composed of 2≤n∈N+2 \le n \in\mathbb{N}_{+} distinct chemical species in a common thermodynamic phase. The primary contributions of our work are a novel variational formulation of a system of equations describing non-ideal miscible isothermal multicomponent flow, a Picard linearization of these equations that possesses symmetric saddle point structure (despite the equations not arising from a Lagrangian), and the numerical analysis of a structure-preserving finite element discretization. In particular, we identify the structural relationships required of the finite element spaces for the different variables that allow for the continuous well-posedness to be inherited automatically in the discretization.

More specifically, this work considers multicomponent flow in the concentrated (i.e. general) solution regime, as opposed to the simpler and more commonly studied dilute solution regime. We now provide an overview of this distinction, for self-containment. The dilute approximation applies when a single species called the solvent (conventionally assigned index i=ni=n) is taken to have a concentration very far in excess of the remaining species (i<ni<n), each of which is called a solute. Thus, if Ω⊂Rd\Omega\subset\mathbb{R}^{d} (d∈{2,3}d \in\{2,3\}) is the domain containing the mixture, the fluid density ρ:Ω→R\rho:\Omega\to\mathbb{R} varies negligibly with solute content in a dilute solution and approximately coincides with the mass density of the pure solvent. This decouples the flow, the material’s bulk motion, from the mass transport, the motion of individual molecular constituents comprising the material. One can thus solve for the flow velocity, and then employ this velocity in a system of independent advection-diffusion equations for the mass transport. A typical dilute solution problem at low Mach number (where the density is assumed constant) is to solve:

∂(ρv)∂t−div⁡(2ηε(v))+div⁡(ρv⊗v)+∇p=ρf,(1)\frac{\partial(\rho v)}{\partial t}-\operatorname{div}(2\eta\varepsilon(v))+\operatorname{div}(\rho v\otimes v)+\nabla p=\rho f, \tag*{(1)}
div⁡v=0,(2)\operatorname{div}v=0, \tag*{(2)}
∂ci∂t+div⁡(civ)+div⁡Ji=ri,i=1,…,n−1,(3)\frac{\partial c_i}{\partial t}+\operatorname{div}(c_i v)+\operatorname{div}J_i=r_i,\qquad i=1,\ldots,n-1, \tag*{(3)}
Ji=−Di∇ci,i=1,…,n−1,(4)J_i=-D_i\nabla c_i,\qquad i=1,\ldots,n-1, \tag*{(4)}

where Ω\Omega is bounded Lipschitz, v:Ω→Rdv:\Omega\to\mathbb{R}^{d} is the flow velocity, ε\varepsilon the symmetric gradient operator ε(v):=12(∇v+(∇v)T)\varepsilon(v):=\frac{1}{2}(\nabla v+(\nabla v)^{\mathsf{T}}), η>0\eta>0 the shear viscosity, p:Ω→Rp:\Omega\to\mathbb{R} the pressure, f:Ω→Rdf:\Omega\to\mathbb{R}^{d} the body acceleration induced within Ω\Omega due to the action of external fields, ci:Ω→Rc_i:\Omega\to\mathbb{R} the concentration of solute ii in the solvent, Ji:Ω→RdJ_i:\Omega\to\mathbb{R}^{d} its diffusive flux, ri:Ω→Rr_i:\Omega\to\mathbb{R} its volumetric rate of generation or depletion, and Di>0D_i>0 its Fickian diffusion coefficient. The velocity vi:Ω→Rdv_i:\Omega\to\mathbb{R}^{d} of each individual species is given by

civi=civ+Ji,(5)c_i v_i=c_i v+J_i, \tag*{(5)}

decomposing the transport of each species into a convective and a diffusive contribution. The dilute solution regime is characterized by the approximation v≈vnv\approx v_n, that the bulk motion of the fluid coincides with the motion of the solvent.

While the dilute solution approximation has been applied to great effect [11, 41, 23, 24], it fails starkly when no particular species is present in great excess, the concentrated solution regime of interest in this work. Several problems arise when attempting to relax the dilute solution approximation and formulate and discretize models for concentrated solutions. In concentrated solutions the very notion of ‘flow velocity’ becomes ambiguous, because the overall bulk motion of the fluid need not coincide with any particular species velocity and these species velocities will in general be distinct. One can still identify a natural composition-dependent definition of vv in the concentrated case, however. The density of the fluid is given by

ρ:=∑i=1nMici,(6)\rho:= \sum_{i=1}^{n} M_i c_i, \tag*{(6)}

in which Mi>0M_i > 0 represents the molar mass of species ii. Using (5), the continuity equations (3) may be rephrased in terms of species velocities as

∂ci∂t+div⁡(civi)=ri,i=1,…,n.(7)\frac{\partial c_i}{\partial t} + \operatorname{div}(c_i v_i) = r_i,\qquad i=1,\ldots,n. \tag*{(7)}

Time differentiation of (6), followed by elimination of the concentration derivatives with (7), yields

∂ρ∂t=∑iMi(ri−div⁡(civi))=−div⁡(∑iMicivi).(8)\frac{\partial\rho}{\partial t} = \sum_i M_i\left(r_i-\operatorname{div}(c_i v_i)\right) = -\operatorname{div}\left(\sum_i M_i c_i v_i\right). \tag*{(8)}

Here, the last equality incorporates the premise that homogeneous chemical reactions conserve atoms, which requires that ∑iMiri=0\sum_i M_i r_i = 0.1 Equation (8) is consistent with the common understanding of mass continuity if the flow velocity vv within a multicomponent fluid is identified as the so-called mass-average velocity, defined as [38 p. 454]

v:=∑iωivi,(9)v := \sum_i \omega_i v_i, \tag*{(9)}

where

ωi:=Miciρ(10)\omega_i := \frac{M_i c_i}{\rho} \tag*{(10)}

defines the mass fraction of species ii. Indeed, rewriting (8) in terms of the mass-average velocity yields

∂ρ∂t=−div⁡(ρv),(11)\frac{\partial\rho}{\partial t} = -\operatorname{div}(\rho v), \tag*{(11)}

thereby recovering the mass continuity equation familiar from fluid mechanics. In our formulation we will solve for both the mass-average velocity and the individual species velocities. The mass-average velocity is governed by a momentum balance, typically expressed in the form of the Cauchy equation

∂(ρv)∂t−div⁡τ+div⁡(ρv⊗v)+∇p=ρf,(12)\frac{\partial(\rho v)}{\partial t} - \operatorname{div}\tau+ \operatorname{div}(\rho v\otimes v) + \nabla p = \rho f, \tag*{(12)}

where τ:Ω→Rsymd×d\tau:\Omega\to\mathbb{R}_{\mathrm{sym}}^{d\times d} is the dissipative (viscous) internal stress tensor,2 to be specified with a constitutive law. If Newton’s law of viscosity is used, then (12) reduces to the Navier–Stokes momentum equation (1).

Another issue to address when moving to the concentrated solution regime is the choice of constitutive law for the diffusive fluxes. In dilute solutions each solute interacts at a molecular level almost solely with solvent molecules, and so the diffusive solute fluxes JiJ_i can each be modelled by Fick’s law [30]. In concentrated solutions the constitutive laws for mass transport become incomplete, because Fick’s law (4) fails to take into account all possible species-species interactions. Even in the case of simple diffusion (where v=0\mathbf{v}=0 uniformly), the diffusive flux of a given species can generally be driven by a concentration gradient of any other species in the solution—a phenomenon known as cross-diffusion. The theory of linear irreversible thermodynamics, pioneered by Onsager [46] (1945), enables the thermodynamically consistent generalization of Fick’s law (4) to the concentrated solution regime. This formalism is described in the next subsection.

Onsager–Stefan–Maxwell equations

Within a multi-species solution, the Onsager–Stefan–Maxwell (OSM) equations relate the diffusion driving forces di:Ω→Rd\mathbf{d}_i:\Omega\to\mathbb{R}^d to the species velocities vi\mathbf{v}_i via

di=∑jMijvj,i=1,…,n,(13)\mathbf{d}_i=\sum_j\mathbf{M}_{ij}\mathbf{v}_j,\qquad i=1,\ldots,n, \tag*{(13)}

where the diffusion driving forces di\mathbf{d}_i incorporate the effects of various state variable gradients and external forces [32 Eq. (2.5.4], and where M\mathbf{M} is the Onsager transport coefficient matrix with entries

Mij={−RTcicjDijcTif i≠j,∑k=1k≠inRTcickDikcTif i=j.(14)\mathbf{M}_{ij}= \begin{cases} -\dfrac{RTc_ic_j}{\mathcal{D}_{ij}c_T} & \text{if } i\ne j,\\ \sum_{\substack{k=1\\k\ne i}}^n\dfrac{RTc_ic_k}{\mathcal{D}_{ik}c_T} & \text{if } i=j. \end{cases} \tag*{(14)}

Here R>0R>0 is the ideal gas constant, T>0T>0 the ambient temperature, cTc_T denotes the total concentration defined as

cT:=∑ici,(15)c_T:=\sum_i c_i, \tag*{(15)}

and Dij∈R\mathcal{D}_{ij}\in\mathbb{R} represents the Stefan–Maxwell diffusivity of species ii through species j≠ij\ne i. The Stefan–Maxwell diffusivities are symmetric in the species indices, Dij=Dji\mathcal{D}_{ij}=\mathcal{D}_{ji}, and Dii\mathcal{D}_{ii} is not defined. In the present discussion we restrict attention to the case where every Dij\mathcal{D}_{ij} is constant, which in turn requires each to be positive [58], but in general the Stefan–Maxwell diffusivities may depend on the species concentrations, temperature, and pressure. Equation (13) is often presented as

di=∑j=1j≠inRTcicjDijcT(vi−vj),(16)\mathbf{d}_i=\sum_{\substack{j=1\\j\ne i}}^n\frac{RTc_ic_j}{\mathcal{D}_{ij}c_T}(\mathbf{v}_i-\mathbf{v}_j), \tag*{(16)}

which follows from (13) and (14).

In an isothermal but nonisobaric medium without external forces, the diffusion driving forces may be identified as [11 Eq. (24.1-8]

di:=−ci∇μi+ωi∇p,i=1,…,n,(17)\mathbf{d}_i:=-c_i\nabla\mu_i+\omega_i\nabla p,\qquad i=1,\ldots,n, \tag*{(17)}

in which μi:Ω→R\mu_i:\Omega\to\mathbb{R} is the chemical potential of species ii. The chemical potential represents the partial derivative of the Gibbs free energy (a quantity describing the total amount of work a system can deliver to isothermal, isobaric surroundings) with respect to the number of moles of a given species ii at constant temperature and pressure, and holding the molar contents of all other species fixed. The chemical potentials are related to the concentrations and pressure via a constitutive law, discussed below in section 1.4.

As a consequence of the statistical reciprocal relations developed by Onsager [46] and the second law of thermodynamics, the transport matrix M\mathbf{M} is everywhere symmetric positive semidefinite, with a single null eigenvalue associated with (1,…,1)T(1,\ldots,1)^{\mathsf{T}}:

∑jMij=0.(18)\sum_j M_{ij}=0. \tag*{(18)}

This nullspace is necessary to distinguish convection, which is non-dissipative, from diffusion. A consequence of this nullspace is that one may shift each viv_i in (13) by the mass-average velocity

di=∑jMij(vj−v),i=1,…,n,(19)d_i=\sum_j M_{ij}(v_j-v), \qquad i=1,\ldots,n, \tag*{(19)}

so that the transport matrix acts on terms proportional to the diffusive flux JiJ_i. Thus (13) can be understood as an implicit constitutive relation for the diffusive fluxes (Bulíček et al., [16]).

The symmetry of M\mathbf{M} combined with (18) allows one to show that

∑idi=∑i(−ci∇μi+ωi∇p)=0,(20)\sum_i d_i=\sum_i(-c_i\nabla\mu_i+\omega_i\nabla p)=0, \tag*{(20)}

an expression of the isothermal, nonisobaric Gibbs–Duhem equation from equilibrium thermodynamics. For further detail on the historical development and mathematical structure of the Onsager–Stefan–Maxwell equations, we refer the reader to Van-Brunt et al. [58].

Stokes equation

A further consequence of the nullspace of M\mathbf{M} (18) is that the species velocities cannot be determined from the diffusion driving forces alone. They can be recovered, however, by incorporating the Cauchy momentum equation (12) to specify the mass-average velocity (9), which is required for a complete description of the overall transport problem; this in turn requires a constitutive law for the viscous stress. For isothermal Newtonian fluids, the viscous stress τ\boldsymbol{\tau} relates to the linearized strain rate, the symmetric part of the velocity gradient, through

τ=2η(ε(v)−tr⁡ε(v)dI)+ζtr⁡(ε(v))I,(21)\boldsymbol{\tau}=2\eta\left(\boldsymbol{\varepsilon}(\boldsymbol{v})-\frac{\operatorname{tr}\boldsymbol{\varepsilon}(\boldsymbol{v})}{d}\mathbf{I}\right)+\zeta\operatorname{tr}(\boldsymbol{\varepsilon}(\boldsymbol{v}))\mathbf{I}, \tag*{(21)}

where ζ>0\zeta>0 is the bulk viscosity, or equivalently

ε(v)=12ητ+(1d2ζ−12ηd)(tr⁡τ)I=:Aτ,(22)\boldsymbol{\varepsilon}(\boldsymbol{v})=\frac{1}{2\eta}\boldsymbol{\tau}+\left(\frac{1}{d^{2}\zeta}-\frac{1}{2\eta d}\right)(\operatorname{tr}\boldsymbol{\tau})\mathbf{I}=:\mathcal{A}\boldsymbol{\tau}, \tag*{(22)}

where A:Rsymd×d→Rsymd×d\mathcal{A}:\mathbb{R}_{\mathrm{sym}}^{d\times d}\to\mathbb{R}_{\mathrm{sym}}^{d\times d} denotes the compliance tensor. The full Cauchy stress σ:Ω→Rsymd×d\boldsymbol{\sigma}:\Omega\to\mathbb{R}_{\mathrm{sym}}^{d\times d} may then be decomposed as

σ=τ−pI.(23)\boldsymbol{\sigma}=\boldsymbol{\tau}-p\mathbf{I}. \tag*{(23)}

We further consider steady-state creeping flow, under which assumptions Stokes’ equation,

div⁡τ−∇p=−ρf,(24)\operatorname{div}\boldsymbol{\tau}-\nabla p=-\rho\boldsymbol{f}, \tag*{(24)}

follows from the momentum balance (12).

The OSM equations (13) are written in force-explicit form: the equations express the species velocities (fluxes) as implicit variables. We choose also to write the Newtonian constitutive equation (22) in this manner, expressing the thermodynamic force (the linearized strain rate) in terms of the corresponding flux (the viscous stress) [38]. Typically in computational fluid dynamics, a flux-explicit formulation is obtained by using an explicit constitutive relation such as (21) to eliminate the Cauchy and viscous stresses in the first instance. For our overall coupled system (stated later in (1.29)), we do not eliminate the viscous stress, but include it as an implicit variable to be solved for. While this increases the computational cost, it has substantial benefits; the viscous stress plays a fundamental role in the calculation of local entropy production, but more significantly, we show in section 2 that the resulting system of equations can be cast as a symmetric perturbed saddle point-like system, which is conducive to both theoretical analysis and (we anticipate) efficient linear solvers.

Augmentation of the diffusion transport matrix and the Stokes momentum balance

The variational formulation of our equations must enforce the relation (9), between the bulk (mass-average) velocity and the species velocities. We employ the augmentation approach introduced by Helfand [37] and later used by Giovangigli [31]; Ern and Giovangigli [27]; Giovangigli [32]. We augment each OSM equation (16) by adding the mass-average velocity constraint (9) to both sides with a prefactor γ>0\gamma> 0:

di+γωiv=∑j≠iRTcicjDijcT(vi−vj)+γωi∑jωjvj.(25)d_i+\gamma\omega_i\mathbf{v}=\sum_{j\ne i}\frac{RTc_ic_j}{\mathcal{D}_{ij}c_{\mathrm{T}}}(\mathbf{v}_i-\mathbf{v}_j)+\gamma\omega_i\sum_j\omega_j\mathbf{v}_j. \tag*{(25)}

The units of γ\gamma are uniquely determined by (25). For convenience, we define the augmented transport matrix

Mijγ:=Mij+γωiωj,(26)\mathbf{M}_{ij}^{\gamma}:=\mathbf{M}_{ij}+\gamma\omega_i\omega_j, \tag*{(26)}

so that (25) can be stated as di+γωiv=∑j≠iMijγvjd_i+\gamma\omega_i\mathbf{v}=\sum_{j\ne i}\mathbf{M}_{ij}^{\gamma}\mathbf{v}_j. We may then compute directly

∑i,jvi⋅Mijγvj=12∑j≠iRTcicjDijcT∣vi−vj∣2+γ∣∑jωjvj∣2(27)\sum_{i,j}\mathbf{v}_i\cdot\mathbf{M}_{ij}^{\gamma}\mathbf{v}_j=\frac{1}{2}\sum_{j\ne i}\frac{RTc_ic_j}{\mathcal{D}_{ij}c_{\mathrm{T}}}\lvert\mathbf{v}_i-\mathbf{v}_j\rvert^2+\gamma\left\lvert\sum_j\omega_j\mathbf{v}_j\right\rvert^2 \tag*{(27)}

to show that the augmented transport matrix is symmetric positive definite [58]. This was used in Van-Brunt et al. [58] to construct coercive bilinear forms for the pure Stefan–Maxwell diffusion problem. With this augmentation, although the transport matrix is a priori singular, one can nevertheless recover the species velocities from the driving forces by coupling with the mass-average velocity constraint (9).

The augmentation (25) modifies a constitutive law of the system, which will induce coercivity of a certain bilinear form below. However, this comes at the cost of symmetry. To recover symmetry, we add a ‘dual’ augmentation to the Stokes equation (24):

div⁡τ−∇p=−ρf+γ∑jωj(v−vj).(28)\operatorname{div}\boldsymbol{\tau}-\nabla p=-\rho\mathbf{f}+\gamma\sum_j\omega_j(\mathbf{v}-\mathbf{v}_j). \tag*{(28)}

With these two augmentations (25) and (28), an important bilinear form defined later in (60) will be both symmetric and coercive on an appropriate kernel. This greatly aids the proofs of well-posedness for the continuous and discrete problems, as we will demonstrate in section 3.2 and section 4.1.

To close the equations, we must relate the chemical potentials to the concentrations via a thermodynamic equation of state, discussed next.

The chemical potential and equation of state

Our variational formulation will solve for the chemical potential μi\mu_i of each species ii. This has several advantages. First, this allows for a general formula for the diffusion driving forces (17), independent of the materials considered. If we were to make the (perhaps more obvious) choice of solving for concentrations cic_i as the primary variables instead, the form of the diffusion driving forces would change in a material-dependent manner. Second, our choice allows for a decoupling in the linearization we employ: the primary mixed system to solve only depends on the material via the diffusion coefficients and viscosities, with any non-ideality confined to the computation of concentrations and density postprocessed at every iteration using material-dependent thermodynamic constitutive relations discussed below. Third, together with the choice to solve for the viscous stress as described in § 1.2, this decoupling endows the equations to solve with a symmetric perturbed saddle point-like structure.

Generally each species concentration cic_i can be inferred from {μi}i=1n\{\mu_i\}_{i=1}^{n} and pp, given thermodynamically consistent constitutive laws for the chemical potential, and an equation of state which relates cTc_T to pressure and composition. Within an isothermal ideal gas, this relation is simply

ci=p⊖RTexp⁡(μi−μi⊖RT),(29)c_i = \frac{p^{\ominus}}{RT}\exp\left(\frac{\mu_i-\mu_i^{\ominus}}{RT}\right), \tag*{(29)}

for some known reference pressure p⊖p^{\ominus} and a set of reference chemical potentials {μi⊖}i\{\mu_i^{\ominus}\}_i. A general relation for non-ideal systems is

μi=μi⊖+RTln⁡(γixi),(30)\mu_i = \mu_i^{\ominus} + RT\ln(\gamma_i x_i), \tag*{(30)}

where xi:=ci/cTx_i := c_i/c_T is the mole fraction, and γi\gamma_i the activity coefficient, of species ii. (Within a system made up of nn species, specifying n−1n-1 mole fractions determines the composition referred to earlier.) In non-ideal solutions, the reference potentials μi⊖\mu_i^{\ominus} generally depend on the temperature and pressure [36, 5]; they determine the value of the molar Gibbs free energy of pure species ii at the TT and pp values of interest. Activity coefficients generally depend on temperature, pressure, and composition; the definition of the reference state further requires that they approach unity at infinite dilution, i.e. lim⁡xi→0γi=1\lim_{x_i\to0}\gamma_i=1. Constitutive laws (30) suffice to determine the mole fractions (but not concentrations) within non-ideal solutions. To obtain the concentrations, an additional equation of state for the system as a whole is required, which may be expressed in volumetric form as

cT=1∑iVixi,(31)c_T = \frac{1}{\sum_i V_i x_i}, \tag*{(31)}

in which Vi>0V_i>0 is the partial molar volume of species ii. Formally, the partial molar volume is a material property that quantifies the change in total fluid volume with respect to the number of moles of a species ii at constant temperature and pressure, holding all other species contents fixed. Maxwell relations derived from the Gibbs free energy also require that ViV_i quantifies the partial derivative of μi\mu_i with respect to pp [33 Eq. (28]. This dependence, part of which is embedded in the pressure dependence of μi⊖\mu_i^{\ominus}, may be regarded as given data that is experimentally measurable [25].

Our linearization below is designed so that the concentrations are calculated from the chemical potentials and pressure. This trivially guarantees positivity of the concentrations, but more significantly, the model is able to incorporate non-ideality by employing chemical potential constitutive laws more general than (29), such as (30).

Coupled problem statement

Our goal is to find and analyze a variational formulation and structure-preserving finite element discretization of the following problem: given data ff and {ri}i=1n\{r_i\}_{i=1}^n, find chemical potentials {μi}i=1n\{\mu_i\}_{i=1}^n, viscous stress τ\tau, pressure pp, species velocities {vi}i=1n\{v_i\}_{i=1}^n, and convective velocity vv satisfying

−ci∇μi+ωi∇p+γωiv=∑jMijγvj∀i,(32)-c_i\nabla\mu_i+\omega_i\nabla p+\gamma\omega_i v=\sum_j M_{ij}^{\gamma}v_j\quad\forall i, \tag*{(32)}

(augmented OSM equations)

ε(v)=Aτ,(33)\varepsilon(v)=\mathcal{A}\tau, \tag*{(33)}

(stress constitutive law)

div⁡τ−∇p−γ∑jωj(v−vj)=−ρf,(34)\operatorname{div}\tau-\nabla p-\gamma\sum_j\omega_j(v-v_j)=-\rho f, \tag*{(34)}

(augmented Stokes equation)

div⁡(civi)=ri∀i,(35)\operatorname{div}(c_iv_i)=r_i\quad\forall i, \tag*{(35)}

(species continuity equation)

div⁡(v)=div⁡(∑jωjvj),(36)\operatorname{div}(v)=\operatorname{div}\left(\sum_j\omega_jv_j\right), \tag*{(36)}

(mass-average velocity constraint) for an augmentation parameter γ≥0\gamma\ge0, where {ci,ωi}i=1n,ρ\{c_i,\omega_i\}_{i=1}^n,\rho are functions of the unknowns via chemical potential constitutive laws such as (30) and (31), and algebraic relations (6), (10). We shall introduce appropriate boundary conditions in section 2.2. We call the system (1.29) the (augmented) Stokes–Onsager–Stefan–Maxwell (SOSM) system. When the convection term div⁡(ρv⊗v)\operatorname{div}(\rho v\otimes v) is incorporated into (34), we call this the Navier–Stokes–Onsager–Stefan–Maxwell (NSOSM) system. Note in the system we only directly enforce the divergence of the mass-average velocity constraint (36), which may be interpreted as the compressible generalization of the standard divergence constraint (2); this choice gives rise to a saddle point-like structure, as we show in the next section. Nevertheless, the full constraint (9) is incorporated via the augmentations (25) and (28), as discussed further in Remark 2.

Relation to existing literature and outline

For dilute solutions with constant solvent concentration (and no volumetric generation or the depletion of the solvent, rn=0r_n=0), the NSOSM equations reduce to the incompressible Navier–Stokes equations, as well as convection-diffusion equations constituted by Fick’s law for each solute. These equations have been studied for many decades, and effective numerical techniques are available. We do not attempt a systematic review here, but mention [54]; [39]; [56]; [26] as gateways to this literature. In this regime, the momentum solve and the equation for the transport of concentration are decoupled using incompressibility.

Our formulation (1.29) solves for the viscous stress as an unknown variable. Of most relevance to this aspect of our approach is the work of Carstensen et al. [20], who discretized the stress in an incompressible stress-velocity Stokes system using the same stress elements of Arnold and Winther [3] that we shall employ.

Related systems of equations have been formulated and analyzed, including the coupling of the Stefan–Maxwell equations with the incompressible Navier–Stokes equations by Chen and Jüngel [22], the compressible Navier–Stokes equations by Bothe and Druet [15], the Darcy momentum equation by Ostrowski and Rohde [49], and the Cahn–Hilliard equations by Huo et al. [40]. The coupling of an anisothermal NSOSM system to surface phenomena by sorption was formulated by Souček et al. [51]; Bothe and Dreyer [14] proposed an extension to the Stefan–Maxwell equations in the presence of chemically reacting constituents.

Numerical methods for solving the NSOSM equations have received much less attention. The only works of which the authors are aware are those of Ern, Giovangigli, and coauthors, including a monograph [27] and a series of other works [32, 28, 29, 17] which apply multicomponent transport to combustion modelling for ideal gas mixtures. These schemes use sophisticated finite difference methods, with the important exception of Burman et al. [17], which uses a finite element method with additional least-squares terms to stabilize the formulation. The authors are unaware of any literature that addresses numerical methods for SOSM or NSOSM systems in the non-ideal case.

For OSM models of isobaric ideal gases under simple diffusion, several recent papers have addressed numerical approaches, including a finite element method proposed by McLeod and Bourgault [45], a finite volume method by Cancès et al. [18], and a finite difference scheme by Bondesan et al. [13].

Recently a finite element scheme for simple isobaric OSM diffusion in ideal gases was proposed by a subset of the current authors [58]. The approach solved the augmented OSM equations (25) combined with the species continuity equations (35). This paper builds on the foundation established in Van-Brunt et al. [58], but now fully incorporates momentum, non-ideality, and pressure-driven diffusion. In contrast to this prior work, we are able here to avoid a generalized saddle point formulation, and in § 3 will derive a symmetric perturbed saddle point system to be solved at each nonlinear iteration—a more classical linear algebraic structure for which many solvers have been developed [10]. However, due to the more complex form (17) of the driving force, and since we solve for the chemical potentials to allow for non-ideal fluids, we are not able to enforce the Gibbs–Duhem relation (20) to machine precision, as achieved in Van-Brunt et al. [58].

Remark 1. Many cross-diffusion systems, such as those describing multiagent systems in mathematical biology [19], arise from a gradient flow of an associated entropy functional. Unfortunately, although the OSM system admits an associated thermodynamic energy—the Gibbs free energy—we are not able to show equivalence of the (S)OSM system to the Euler–Lagrange stationarity condition of any energy or Lagrangian functional, and hence cannot exploit any gradient flow structure. Instead, our mathematical line of attack will be to exploit the positive definiteness of the augmented transport matrix Mγ\mathbf{M}^{\gamma} (26). With our augmentations of the equations, the Picard scheme we propose below in § 3 nevertheless gives rise to symmetric linearized problems to solve at each nonlinear iteration.

The remainder of this work is organized as follows. In § 2, we derive a novel variational formulation of the fully coupled nonlinear SOSM problem, incorporating boundary conditions and augmentation terms, as a nonlinear perturbed saddle point-like system, using a novel solution-dependent test space relating to the thermodynamic driving force; our principal discovery is the duality between the diffusion driving forces, and the combination of species continuity equations with the divergence of the mass-average velocity constraint. § 3 proposes a Picard-like linearization, which is proven to be well-posed under physically reasonable assumptions. § 4 identifies appropriate finite element spaces, and the structural relations which should hold between them, for a well-posed and convergent discretization of this linearization. We then validate our convergence results numerically. Finally, we illustrate our method by simulating the steady mixing of liquid benzene and cyclohexane in a two-dimensional microfluidic laminar-flow device.

Variational formulation

We employ standard notation for the Sobolev space Hk(Ω;X)H^k(\Omega;\mathsf{X}) (or L2(Ω;X)L^2(\Omega;\mathsf{X}) when k=0k=0) with domain Ω⊂Rd\Omega\subset\mathbb{R}^d and codomain X\mathsf{X}, and associated norm ∥⋅∥k\lVert\cdot\rVert_k and seminorm ∣⋅∣k\lvert\cdot\rvert_k. We denote by S=Rsymd×d\mathbb{S}=\mathbb{R}_{\mathrm{sym}}^{d\times d} the space of d×dd\times d symmetric tensors. The symbol ≲\lesssim denotes inequality up to a constant which may depend on mesh regularity but not mesh spacing hh. Let L02(Ω):={z∈L2(Ω)∣∫Ωz dx=0}L^2_0(\Omega):=\{z\in L^2(\Omega)\mid\int_\Omega z\,\mathrm{d}x=0\}. We use the notation q~=(q1,…,qn)\widetilde{q}=(q_1,\ldots,q_n) to denote an nn-tuple of functions. Let Γ=∂Ω\Gamma=\partial\Omega and let ⟨⋅,⋅⟩Γ\langle\cdot,\cdot\rangle_\Gamma denote the (H−1/2×H1/2)(Γ;R or Rd)(H^{-1/2}\times H^{1/2})(\Gamma;\mathbb{R}\text{ or }\mathbb{R}^d) dual pairing.

Integrability of pressure gradients

For isothermal, isobaric multicomponent diffusion as originally considered by Maxwell [44] and Stefan [52], it suffices to work with driving forces of the form

di=−RT∇ci.(37)d_i=-RT\nabla c_i. \tag*{(37)}

In a variational formulation of the nonisobaric case, one would like to integrate the pressure gradient term in our diffusion driving forces (17) by parts, to reduce the regularity requirement on pp. However, it is not obvious how to do so, since the mass fractions ωi\omega_i are spatially varying. In order to rigorously incorporate the effect of pressure-driven diffusion, we are therefore led to consider the somewhat unorthodox possibility of formulating the Stokes subproblem with pressure lying in H1(Ω)H^1(\Omega). Typically, the condition that p∈H1(Ω)p\in H^1(\Omega) may be provided by elliptic regularity results for the pressure field, but to the authors’ knowledge, the a priori square-integrability of pressure gradients (i.e. for which, we emphasize, pressure is defined to lie in H1(Ω)H^1(\Omega)) has not been considered for the Stokes system, except at the discrete level for the incompressible case in Stenberg [53]. This condition is also suggested by the case of pure Stefan–Maxwell diffusion for nonisobaric ideal gases. Here the driving forces are

di=−RT∇ci+ωi∇p,(38)d_i=-RT\nabla c_i+\omega_i\nabla p, \tag*{(38)}

which suggests considering each cic_i (and hence cTc_T) to lie in H1(Ω)H^1(\Omega), which forces the pressure to lie in the same space due to the ideal equation of state p=cTRTp=c_T RT.

In general, one must distinguish between the thermodynamic pressure pp, which we use throughout this paper, and the mechanical pressure pm:=−tr⁡σ/dp_m:=-\operatorname{tr}\sigma/d. The mechanical pressure is related to the spherical Cauchy stress by sph⁡σ:=tr⁡σdI=−pmI\operatorname{sph}\sigma:=\frac{\operatorname{tr}\sigma}{d}\mathbb{I}=-p_m\mathbb{I}, and to pp by

p=pm+ζdiv⁡v.(39)p=p_m+\zeta\operatorname{div}v. \tag*{(39)}

In the context of multicomponent flow, this decomposition is discussed in further detail by Bothe and Dreyer [14]. Even in the simpler incompressible limit where p=pmp=p_m, we cannot expect extra regularity of ∇p=−div⁡(sph⁡σ)\nabla p=-\operatorname{div}(\operatorname{sph}\sigma) because H(div⁡;S)H(\operatorname{div};\mathbb{S}), the natural space for σ\sigma, is not closed under taking spherical parts.3 Consequently, we do not take p∈H1(Ω)p\in H^1(\Omega), but as a compromise consider a weaker condition defined by the combined viscous stress-pressure space

{(τ,p)∈L2(Ω;S)×L2(Ω)∣div⁡τ−∇p∈L2(Ω;Rd)}⊇H(div⁡;S)×H1(Ω)(40)\left\{(\tau,p)\in L^{2}(\Omega;\mathcal{S})\times L^{2}(\Omega)\mid\operatorname{div}\tau-\nabla p\in L^{2}(\Omega;\mathbb{R}^{d})\right\}\supseteq H(\operatorname{div};\mathcal{S})\times H^{1}(\Omega) \tag*{(40)}

and assign to it the weaker norm ∥τ∥02+∥p∥02+∥div⁡τ−∇p∥02\lVert\tau\rVert_{0}^{2}+\lVert p\rVert_{0}^{2}+\lVert\operatorname{div}\tau-\nabla p\rVert_{0}^{2}. This space and norm were previously employed by Manouzi and Farhloul [43] in an analysis of a non-Newtonian incompressible Stokes flow where τ\tau was taken to be the deviatoric shear stress. Membership of the space (40) is naturally interpretable as the square-integrability of the divergence of the full Cauchy stress, i.e. that σ=τ−pI∈H(div⁡;S)\sigma=\tau-pI\in H(\operatorname{div};\mathcal{S}). Together with an analogous condition for the chemical potential gradient to be detailed next, this weaker condition will account for the pressure gradient in the driving forces.4

Fully coupled variational formulation

In this subsection, we derive a variational formulation for the stationary problem as a nonlinear perturbed saddle point-like system. We have found the following statement of the problem to be a feasible tradeoff between the (competing) goals of: physical relevance of variables and boundary data, regularity assumptions, numerical implementability and effectiveness, analytic tractability, enforcement of fundamental thermodynamic relations, and extensibility to the transient, anisothermal, and non-ideal settings.

For boundary data, we prescribe mass flux and molar fluxes:

ρv=gv∈H1/2(Γ;Rd)on Γ,(41)\rho v=g_{v}\in H^{1/2}(\Gamma;\mathbb{R}^{d})\quad\text{on }\Gamma, \tag*{(41)}
civi⋅n=gi∈H−1/2(Γ)on Γ,i=1,…,n.(42)c_{i}v_{i}\cdot\mathbf{n}=g_{i}\in H^{-1/2}(\Gamma)\quad\text{on }\Gamma,\quad i=1,\ldots,n. \tag*{(42)}

For consistency with the mass-average velocity constraint (9), we require

∑iMigi=gv⋅n,(43)\sum_{i}M_{i}g_{i}=g_{v}\cdot\mathbf{n}, \tag*{(43)}

with equality in H−1/2(Γ)H^{-1/2}(\Gamma). We further impose conditions

∫Ωp dx=∫Ωμi dx=0,i=1,…,n,(44)\int_{\Omega}p\,\mathrm{d}x=\int_{\Omega}\mu_{i}\,\mathrm{d}x=0,\quad i=1,\ldots,n, \tag*{(44)}

on the pressure and chemical potentials. Typically, the equation of state will require or imply strict positivity of pp everywhere, in which case this condition should be understood as ∫Ωp dx=p⊖>0\int_{\Omega}p\,\mathrm{d}x=p^{\ominus}>0 and that pp be shifted by the known value p⊖p^{\ominus} as a postprocessing step.

Let Q=L2(Ω;Rd)n×L2(Ω;Rd)Q=L^{2}(\Omega;\mathbb{R}^{d})^{n}\times L^{2}(\Omega;\mathbb{R}^{d}) with norm ∥(v~,v)∥Q2:=∥v~∥02+∥v∥02\lVert(\widetilde{v},v)\rVert_{Q}^{2}:=\lVert\widetilde{v}\rVert_{0}^{2}+\lVert v\rVert_{0}^{2}. For formal derivation of the weak form, we assume the solution tuple (μ~,τ,p,v~,v)(\widetilde{\mu},\tau,p,\widetilde{v},v) to be smooth on Ω\Omega, and consider choosing (w~,s,q)(\widetilde{w},s,q) from the solution-dependent potential-stress-pressure test space

Θ:={(w~,s,q)∈L02(Ω)n×L2(Ω;S)×L02(Ω) | div⁡s−∇q∈L2(Ω;Rd),−ci∇wi+ωi∇q∈L2(Ω;Rd)∀i}.(45)\Theta:=\left\{(\widetilde{w},s,q)\in L_{0}^{2}(\Omega)^{n}\times L^{2}(\Omega;\mathcal{S})\times L_{0}^{2}(\Omega)\ \middle|\ \begin{aligned} &\operatorname{div}s-\nabla q\in L^{2}(\Omega;\mathbb{R}^{d}),\\ &-c_{i}\nabla w_{i}+\omega_{i}\nabla q\in L^{2}(\Omega;\mathbb{R}^{d})\quad\forall i \end{aligned} \right\}. \tag*{(45)}

Here it is understood that the {ci,ωi}i\{c_{i},\omega_{i}\}_{i} are computed from the solution tuple. Multiplying the ithi^{\mathrm{th}} continuity equation (35) by wiw_{i}, the divergence of the mass-average velocity constraint (36) by qq, and contracting the stress constitutive law (1.29b) with ss, we obtain

∑i(div⁡(civi)−ri)wi+div⁡(v−∑iωivi)q+(Aτ−ε(v)):s=0,(46)\sum_i \left(\operatorname{div}(c_i\mathbf{v}_i)-r_i\right)w_i+\operatorname{div}\left(\mathbf{v}-\sum_i\omega_i\mathbf{v}_i\right)q+\left(\mathcal{A}\boldsymbol{\tau}-\varepsilon(\mathbf{v})\right):s=0, \tag*{(46)}

and hence

∫Ω∑i(div⁡(civi)wi−div⁡(ωivi)q)+Aτ:s−(s−qI):ε(v) dx=∫Ω∑iriwi dx.(47)\int_{\Omega}\sum_i\left(\operatorname{div}(c_i\mathbf{v}_i)w_i-\operatorname{div}(\omega_i\mathbf{v}_i)q\right)+\mathcal{A}\boldsymbol{\tau}:s-(s-qI):\varepsilon(\mathbf{v})\,\mathrm{d}x=\int_{\Omega}\sum_i r_iw_i\,\mathrm{d}x. \tag*{(47)}

Integrating by parts yields

∫ΩAτ:s+∑i(−ci∇wi+ωi∇q)⋅vi+(div⁡s−∇q)⋅v dx=⟨(s−qI)n,v⟩Γ+∑i⟨civi⋅n,−wi+ωiciq⟩Γ+∫Ωr~⋅w~ dx=⟨(s−qI)n,gvρ⟩Γ+∑i⟨gi,−wi+Miρq⟩Γ+∫Ωr~⋅w~ dx.(48)\begin{aligned} \int_{\Omega}\mathcal{A}\boldsymbol{\tau}:s+\sum_i(-c_i\nabla w_i+\omega_i\nabla q)\cdot\mathbf{v}_i+(\operatorname{div}s-\nabla q)\cdot\mathbf{v}\,\mathrm{d}x &=\langle(s-qI)\mathbf{n},\mathbf{v}\rangle_{\Gamma}+\sum_i\left\langle c_i\mathbf{v}_i\cdot\mathbf{n},-w_i+\frac{\omega_i}{c_i}q\right\rangle_{\Gamma}+\int_{\Omega}\widetilde{r}\cdot\widetilde{w}\,\mathrm{d}x\\ &=\left\langle(s-qI)\mathbf{n},\frac{\mathbf{g}_{v}}{\rho}\right\rangle_{\Gamma}+\sum_i\left\langle g_i,-w_i+\frac{M_i}{\rho}q\right\rangle_{\Gamma}+\int_{\Omega}\widetilde{r}\cdot\widetilde{w}\,\mathrm{d}x. \tag*{(48)} \end{aligned}

Now for each i=1,…,ni=1,\ldots,n, we take the scalar product of ui∈L2(Ω;Rd)u_i\in L^2(\Omega;\mathbb{R}^d) with the augmented OSM equation (1.29a) and integrate over Ω\Omega to obtain

∫Ω(−ci∇μi+ωi∇p)⋅ui−ui⋅∑jMijvj−γωi∑jωj(vj−v)⋅ui dx=0.(49)\int_{\Omega}(-c_i\nabla\mu_i+\omega_i\nabla p)\cdot u_i-u_i\cdot\sum_j M_{ij}v_j-\gamma\omega_i\sum_j\omega_j(v_j-v)\cdot u_i\,\mathrm{d}x=0. \tag*{(49)}

Taking the inner product of the augmented Cauchy momentum balance (1.29c) with u∈L2(Ω;Rd)u\in L^2(\Omega;\mathbb{R}^d) yields

∫Ω(div⁡τ−∇p)⋅u−γ(∑jωj(v−vj))⋅u dx=−∫Ωρf⋅u dx.(50)\int_{\Omega}(\operatorname{div}\tau-\nabla p)\cdot u-\gamma\left(\sum_j\omega_j(v-v_j)\right)\cdot u\,\mathrm{d}x=-\int_{\Omega}\rho f\cdot u\,\mathrm{d}x. \tag*{(50)}

We sum (49) over ii and add (50) to derive

∫Ω∑i(−ci∇μi+ωi∇p)⋅ui+(div⁡τ−∇p)⋅u−∑i,jui⋅Mijvj−γ(∑iωi(vi−v))⋅(∑jωj(uj−u)) dx=∫Ω−ρf⋅u dx.(51)\int_{\Omega}\sum_i(-c_i\nabla\mu_i+\omega_i\nabla p)\cdot u_i+(\operatorname{div}\tau-\nabla p)\cdot u-\sum_{i,j}u_i\cdot M_{ij}v_j-\gamma\left(\sum_i\omega_i(v_i-v)\right)\cdot\left(\sum_j\omega_j(u_j-u)\right)\,\mathrm{d}x=\int_{\Omega}-\rho f\cdot u\,\mathrm{d}x. \tag*{(51)}

Note that both augmentations (1.22) and (1.25) were involved in deriving this expression.

Finally, we observe that by definition, we have ωi∈L∞(Ω)\omega_i\in L^\infty(\Omega) with ∥ωi∥L∞(Ω)≤1\|\omega_i\|_{L^\infty(\Omega)}\le1. Moreover, we make the physically reasonable assumptions that the concentrations associated with the solution are uniformly bounded, ci∈L∞(Ω)c_i\in L^\infty(\Omega), with ci≥κ>0c_i\ge\kappa>0 a.e. (almost everywhere), as in Van-Brunt et al. [58] (which in turn implies Mij,ρ∈L∞(Ω)M_{ij},\rho\in L^\infty(\Omega), and ρ≥κ∑iMi>0\rho\ge\kappa\sum_i M_i>0 a.e.), and that the density gradient is uniformly bounded, ∇ρ∈L∞(Ω;Rd)\nabla\rho\in L^\infty(\Omega;\mathbb{R}^d).[5]

Definition 1 We define a weak solution to the augmented Stokes–Onsager–Stefan–Maxwell system to be a (2n+3)(2n+3)-tuple

({μi}i=1n,τ,p,{vi}i=1n,v)∈L02(Ω)n×L2(Ω;S)×L02(Ω)×L2(Ω;Rd)n×L2(Ω;Rd)⏟Q(52)\left(\{\mu_i\}_{i=1}^n,\tau,p,\{v_i\}_{i=1}^n,v\right)\in L_0^2(\Omega)^n\times L^2(\Omega;\mathbb{S})\times L_0^2(\Omega)\times\underbrace{L^2(\Omega;\mathbb{R}^d)^n\times L^2(\Omega;\mathbb{R}^d)}_{Q} \tag*{(52)}

inducing concentrations {ci}i=1n\{c_i\}_{i=1}^n through a constitutive law (such as (29)) implicitly defining ci=ci({μi}i=1n,p)≥κ>0c_i=c_i(\{\mu_i\}_{i=1}^n,p)\ge\kappa>0 a.e. for i=1,…,ni=1,\ldots,n, such that

∥ci∥L∞(Ω)<∞,i=1,…,n,(53)\lVert c_i\rVert_{L^\infty(\Omega)}<\infty,\quad i=1,\ldots,n, \tag*{(53)}
∥∇ρ∥L∞(Ω;Rd)<∞,(54)\lVert\nabla\rho\rVert_{L^\infty(\Omega;\mathbb{R}^d)}<\infty, \tag*{(54)}
∥div⁡τ−∇p∥0<∞,(55)\lVert\operatorname{div}\tau-\nabla p\rVert_0<\infty, \tag*{(55)}
∥−ci∇μi+ωi∇p∥02<∞,i=1,…,n,(56)\lVert-c_i\nabla\mu_i+\omega_i\nabla p\rVert_0^2<\infty,\quad i=1,\ldots,n, \tag*{(56)}

and satisfying (48), (51) for all test tuples ({wi}i=1n,s,q,{ui}i=1n,u)∈Θ×Q(\{w_i\}_{i=1}^n,s,q,\{u_i\}_{i=1}^n,u)\in\Theta\times Q, where Θ\Theta is defined in (45).

Observe that the solution tuple does not reside in any standard Sobolev space, but that the regularity assumptions placed on the solution tuple and test spaces ensure that the surface terms in (48) are well-defined. Recall that condition (55) is the square-integrability of the Cauchy stress σ\sigma (as in Manouzi and Farhloul [43]). The nonlinear integrability condition (56) is to our knowledge a novel requirement, but also has a natural interpretation, namely the square-integrability of the diffusion driving forces:6

di∈L2(Ω;Rd).(57)d_i\in L^2(\Omega;\mathbb{R}^d). \tag*{(57)}

Moreover, we emphasize that this unorthodox formulation allows the rigorous incorporation of pressure diffusion via the pressure gradient on the left side of (32), despite the fact that the pressure field is not a priori H1H^1-regular in the Stokes subsystem. Later in section 4.4 we observe convergence of the diffusion driving forces in L2L^2 and of the pressure in H1H^1, but otherwise leave this consideration, and further investigation into the optimal nonlinear formulation of the SOSM system, as intriguing open questions.

Remark 2. In the derivation of (46), we used the divergence of the mass-average velocity constraint (36), which ignores the curl component in the Helmholtz decomposition of the mass-average velocity relationship (9). This choice ensures that the number of equations matches the number of unknown variables. The full constraint is weakly incorporated, however, via the augmentations (25) and (28). Indeed, provided the constitutive equations relating concentrations and chemical potentials are thermodynamically rigorous in the sense that they arise from a Gibbs free energy functional, one can show using the first law of thermodynamics and the extensivity of the Gibbs free energy that the augmentations will enforce the full constraint (9). In the case of an ideal isobaric isothermal gas, a proof of this was given in Van-Brunt et al. [58]. As a consequence, although strict positivity of γ\gamma is required to make the system well-posed, the value chosen should not in principle affect the solution.

Linearization and well-posedness

Variational formulation of a generalized Picard scheme

In this section we derive a variational formulation of a generalized Picard linearization. Given a previous estimate for the potentials μ~k\tilde{\mu}^{k} and pressure pkp^{k} for k≥0k \ge0, we regard these as fixed quantities which determine the concentrations c~k\tilde{c}^{k} via chemical potential constitutive laws and an appropriate equation of state such as (29). This in turn determines the density ρk\rho^{k}, mass fractions ω~k\tilde{\omega}^{k}, total concentration cTkc_{\mathrm{T}}^{k}, and transport matrix Mk\mathbf{M}^{k} defined via (6), (10), (15), and (14), respectively. We then construct a linear system to solve for the next iterate (μ~k+1,c~k+1,pk+1,v~k+1)(\tilde{\mu}^{k+1},\tilde{c}^{k+1},p^{k+1},\tilde{v}^{k+1}). This update strategy is expected to be effective because the gradients of chemical potential, pressure, and mass-average velocity primarily drive the dynamics of multicomponent flow; the role of the species concentrations is mostly confined to the effect of altering the drag coefficients in the transport matrix. We make the following physically reasonable assumptions about each iterate, in analogy to Definition 1.

Assumption 3.1. For each k≥0k \ge0, we assume cik∈L∞(Ω),ρk∈W1,∞(Ω)c_i^k \in L^\infty(\Omega),\rho^k \in W^{1,\infty}(\Omega), and that cik≥κ>0c_i^k \ge\kappa> 0 a.e. for each ii.

This again implies ρk≥κ∑iMi>0\rho^k \ge\kappa\sum_i M_i > 0 a.e. We also assume henceforth that γ>0\gamma> 0.

Given c~k\tilde{c}^{k} and the corresponding ω~k\tilde{\omega}^{k}, we define the iteration-dependent weighted function space

Θk:={(w~,s,q)∈L02(Ω)n×L2(Ω;S)×L02(Ω) | div⁡s−∇q∈L2(Ω;Rd),−cik∇wi+ωik∇q∈L2(Ω;Rd) ∀i},(58)\Theta^k := \left\{(\tilde{w},s,q) \in L_0^2(\Omega)^n \times L^2(\Omega;\mathbb{S}) \times L_0^2(\Omega) \,\middle|\, \begin{array}{l} \operatorname{div}s-\nabla q \in L^2(\Omega;\mathbb{R}^d),\\ -c_i^k\nabla w_i+\omega_i^k\nabla q \in L^2(\Omega;\mathbb{R}^d)\ \forall i \end{array}\right\}, \tag*{(58)}

whose defining conditions linearize those in (45). This mixed space is Hilbertian with graph norm

∥(w~,s,q)∥Θk2:=∥s∥02+∥q∥02+∥div⁡s−∇q∥02+∑i(∥wi∥02+∥−cik∇wi+ωik∇q∥02).(59)\|(\tilde{w},s,q)\|_{\Theta^k}^2 := \|s\|_0^2+\|q\|_0^2+\|\operatorname{div}s-\nabla q\|_0^2+\sum_i\left(\|w_i\|_0^2+\|-c_i^k\nabla w_i+\omega_i^k\nabla q\|_0^2\right). \tag*{(59)}

We now formulate our linearized problem as a symmetric perturbed saddle point problem. Define Ak:Q→Q∗,Λ:Θk→(Θk)∗,Bk:Θk→Q∗A_k: Q \to Q^*,\Lambda:\Theta^k\to(\Theta^k)^*,B_k:\Theta^k\to Q^* by

Ak(v~,v;u~,u):=∫Ω∑i,jvi⋅Mijkuj dx+γ∫Ω(∑iωik(vi−v))⋅(∑jωjk(uj−u)) dx,(60)A_k(\tilde{v},v;\tilde{u},u):=\int_{\Omega}\sum_{i,j}v_i\cdot M_{ij}^k u_j\,\mathrm{d}x+\gamma\int_{\Omega}\left(\sum_i\omega_i^k(v_i-v)\right)\cdot\left(\sum_j\omega_j^k(u_j-u)\right)\,\mathrm{d}x, \tag*{(60)}
Λ(μ~,τ,p;w~,s,q):=∫ΩAτ:s dx,(61)\Lambda(\tilde{\mu},\tau,p;\tilde{w},s,q):=\int_{\Omega}\mathcal{A}\tau:s\,\mathrm{d}x, \tag*{(61)}
Bk(μ~,τ,p;u~,u):=∫Ω∑i(−cik∇μi+ωik∇p)⋅ui+(div⁡τ−∇p)⋅u dx,(62)B_k(\tilde{\mu},\tau,p;\tilde{u},u):=\int_{\Omega}\sum_i(-c_i^k\nabla\mu_i+\omega_i^k\nabla p)\cdot u_i+(\operatorname{div}\tau-\nabla p)\cdot u\,\mathrm{d}x, \tag*{(62)}

and the functionals

ℓk1(w~,s,q):=⟨(s−qI)n,gvρk⟩Γ+∑i⟨gi,−wi+Miρkq⟩Γ+∫Ωr~⋅w~ dx,\ell_k^1(\tilde{w},s,q):=\left\langle(s-qI)n,\frac{g_v}{\rho^k}\right\rangle_{\Gamma}+\sum_i\left\langle g_i,-w_i+\frac{M_i}{\rho^k}q\right\rangle_{\Gamma}+\int_{\Omega}\tilde{r}\cdot\tilde{w}\,\mathrm{d}x,
ℓk2(u~,u):=−∫Ωpkf⋅u dx.(63)\ell_k^2(\tilde{u},u):=-\int_{\Omega}p^k f\cdot u\,\mathrm{d}x. \tag*{(63)}

Note that under Assumption 3.1, each of the bilinear functionals is continuous; we will denote their norms as ∥Ak∥\|A_k\|, ∥Λ∥\|\Lambda\|, and ∥Bk∥\|B_k\|, respectively. Our linearized problem is posed as follows: find ((μ~k+1,τ~k+1,pk+1),(v~k+1,vk+1))∈Θk×Q((\tilde{\mu}^{k+1},\tilde{\tau}^{k+1},p^{k+1}),(\tilde{v}^{k+1},v^{k+1})) \in\Theta^k \times Q such that

Λ(μ~k+1,τ~k+1,pk+1;w~,s,q)+Bk(w~,s,q;v~k+1,vk+1)=ℓk1(w~,s,q)∀(w~,s,q)∈Θk,\Lambda(\tilde{\mu}^{k+1},\tilde{\tau}^{k+1},p^{k+1};\tilde{w},s,q)+B_k(\tilde{w},s,q;\tilde{v}^{k+1},v^{k+1})=\ell_k^1(\tilde{w},s,q)\quad\forall(\tilde{w},s,q)\in\Theta^k,
Bk(μ~k+1,τ~k+1,pk+1;u~,u)−Ak(v~k+1,vk+1;u~,u)=ℓk2(u~,u)∀(u~,u)∈Q,(64)B_k(\tilde{\mu}^{k+1},\tilde{\tau}^{k+1},p^{k+1};\tilde{u},u)-A_k(\tilde{v}^{k+1},v^{k+1};\tilde{u},u)=\ell_k^2(\tilde{u},u)\quad\forall(\tilde{u},u)\in Q, \tag*{(64)}

i.e., defining the transpose BkT:Q→(Θk)∗B_k^{\mathsf{T}}:Q\to(\Theta^k)^* in the canonical way,

\begin{bmatrix} \Lambda& B_k^{\mathsf{T}}\\ B_k & -A_k \end{bmatrix} \begin{bmatrix} \tilde{\mu}^{k+1}\\ \tilde{\tau}^{k+1}\\ p^{k+1}\\ \tilde{v}^{k+1}\\ v^{k+1} \end{bmatrix} = \begin{bmatrix} \ell_k^1\\ \ell_k^2 \end{bmatrix}. \tag*{(65)}\end{bmatrix}

We note that the variational terms involving chemical potential and pressure gradients are precisely of the same variational form as the species continuity equations and the divergence of the mass-average velocity constraint, which can be seen by inspecting (48) and (51). This key insight is what leads to a symmetric system.

Our nonlinear iteration scheme is as follows: for an initial estimate of the concentrations c~0\tilde{c}^0, we solve the system (64) for the updated variables ((μ~k+1,τ~k+1,pk+1),(v~k+1,vk+1))∈Θk×Q((\tilde{\mu}^{k+1},\tilde{\tau}^{k+1},p^{k+1}),(\tilde{v}^{k+1},v^{k+1}))\in\Theta^k\times Q, for k=0,1,2,…k=0,1,2,\ldots. By the relations detailed in § 1.4, these variables are used to calculate the updated concentrations c~k+1\tilde{c}^{k+1}. This is iterated until for some set tolerance ε>0\varepsilon>0,

(∥(μ~k+1,τ~k+1,pk+1)−(μ~k,τ~k,pk)∥Θk2+∥(v~k+1,vk+1)−(v~k,vk)∥Q2)1/2≤ε.(66)\left(\left\|(\tilde{\mu}^{k+1},\tilde{\tau}^{k+1},p^{k+1})-(\tilde{\mu}^{k},\tilde{\tau}^{k},p^{k})\right\|_{\Theta^k}^{2}+\left\|(\tilde{v}^{k+1},v^{k+1})-(\tilde{v}^{k},v^{k})\right\|_{Q}^{2}\right)^{1/2}\leq\varepsilon. \tag*{(66)}

Well-posedness of the linearized system

We will now prove that the saddle point system (64) is well-posed under Assumption 3.1. This will require a Poincaré-type inequality for the following seminorm on Θk\Theta^k:

∣(w~,s,q)∣Θk2:=∥s∥02+∥div⁡s−∇q∥02+∑i∥−cik∇wi+ωik∇q∥02.(67)|(\tilde{w},s,q)|_{\Theta^k}^{2}:=\|s\|_{0}^{2}+\|\operatorname{div}s-\nabla q\|_{0}^{2}+\sum_i\|-c_i^k\nabla w_i+\omega_i^k\nabla q\|_{0}^{2}. \tag*{(67)}

Lemma 1 Let Ω\Omega be a Lipschitz domain. Under Assumption 3.1, there exists K>0K>0 such that for each (μ~,τ,p)∈Θk(\tilde{\mu},\tau,p)\in\Theta^k,

∥(μ~,τ,p)∥Θk≤K∣(μ~,τ,p)∣Θk.(68)\|(\tilde{\mu},\tau,p)\|_{\Theta^k}\leq K|(\tilde{\mu},\tau,p)|_{\Theta^k}. \tag*{(68)}

This lemma will allow us to control the chemical potential and pressure in terms of the (linearized) driving forces, divergence of total stress, and the viscous stress. It should be thought of as the generalization to the OSM framework of Manouzi and Farhloul [43].

Proof of Lemma 1 The first stage of the proof is to show that

∥p∥0≲∥τ∥0+∥div⁡τ−∇p∥0,(69)\|p\|_0\lesssim\|\tau\|_0+\|\operatorname{div}\tau-\nabla p\|_0, \tag*{(69)}

following and mildly generalizing Manouzi and Farhloul [43]. Set θ=τ−pI−rI\theta=\tau-pI-rI where

r=1d∣Ω∣∫Ωtr⁡τ dx.r=\frac{1}{d|\Omega|}\int_{\Omega}\operatorname{tr}\tau\,\mathrm{d}x.

Then

∥τ−pI∥0≤∥θ∥0+∥rI∥0.(70)\|\tau-pI\|_0\leq\|\theta\|_0+\|rI\|_0. \tag*{(70)}

As ∫Ωtr⁡θ dx=0\int_{\Omega}\operatorname{tr}\theta\,\mathrm{d}x=0, we can use Boffi et al. [12] to derive

∥τ−pI∥0≲∥dev⁡θ∥0+∥div⁡θ∥0+∥rI∥0≲∥τ∥0+∥div⁡τ−∇p∥0.(71)\begin{aligned} \lVert\tau-p\mathbb{I}\rVert_0 &\lesssim\lVert\operatorname{dev}\theta\rVert_0+\lVert\operatorname{div}\theta\rVert_0+\lVert r\mathbb{I}\rVert_0\\ &\lesssim\lVert\tau\rVert_0+\lVert\operatorname{div}\tau-\nabla p\rVert_0. \tag*{(71)} \end{aligned}

where the deviator is defined as dev⁡θ:=θ−tr⁡θdI=dev⁡τ\operatorname{dev}\theta:=\theta-\frac{\operatorname{tr}\theta}{d}\mathbb{I}=\operatorname{dev}\tau. Now using

d∥pI∥0≤∥τ−pI∥0+∥τ∥0,(72)\sqrt{d}\lVert p\mathbb{I}\rVert_0\leq\lVert\tau-p\mathbb{I}\rVert_0+\lVert\tau\rVert_0, \tag*{(72)}

the result (69) follows. For the second stage of the proof, we will show that

∥μi∥0≲∥p∥0+∥−cik∇μi+ωik∇p∥0.(73)\lVert\mu_i\rVert_0\lesssim\lVert p\rVert_0+\lVert-c_i^k\nabla\mu_i+\omega_i^k\nabla p\rVert_0. \tag*{(73)}

This combined with (69) gives (68). To prove this second inequality, for each ii we take the unique zi∈H01(Ω;Rd)/ker⁡(div⁡)z_i\in H_0^1(\Omega;\mathbb{R}^d)/\ker(\operatorname{div}) such that div⁡zi=μi\operatorname{div}z_i=\mu_i. Then ui:=zi/cik∈L2(Ω;Rd)u_i:=z_i/c_i^k\in L^2(\Omega;\mathbb{R}^d) with div⁡(cikui)=μi\operatorname{div}(c_i^k u_i)=\mu_i. With integration by parts we deduce

∫Ω(−cik∇μi+ωik∇p)⋅ui dx=∫Ω(∣μi∣2−Mip(μiρk−∇pk(ρk)2⋅cikui)) dx.(74)\int_{\Omega}(-c_i^k\nabla\mu_i+\omega_i^k\nabla p)\cdot u_i\,\mathrm{d}x=\int_{\Omega}\left(|\mu_i|^2-M_i p\left(\frac{\mu_i}{\rho^k}-\frac{\nabla p^k}{(\rho^k)^2}\cdot c_i^k u_i\right)\right)\,\mathrm{d}x. \tag*{(74)}

Upon rearrangement, we can derive the inequality

∥μi∥02≤Mi∥p∥0(∥μi∥0κ∑jMj+∥ui∥0∥cik∥0∥∇ρk(ρk)2∥L∞(Ω;Rd))+∥−cik∇μi+ωi∇p∥0∥ui∥0≤κ−1∥p∥0(∥μi∥0+∥ui∥0∥cik∥L∞(Ω)∥∇ln⁡ρk∥L∞(Ω;Rd))+∥−cik∇μi+ωi∇p∥0∥ui∥0.(75)\begin{aligned} \lVert\mu_i\rVert_0^2\leq M_i\lVert p\rVert_0\left(\frac{\lVert\mu_i\rVert_0}{\kappa\sum_j M_j}+\lVert u_i\rVert_0\lVert c_i^k\rVert_0\left\lVert\frac{\nabla\rho^k}{(\rho^k)^2}\right\rVert_{L^\infty(\Omega;\mathbb{R}^d)}\right)+\lVert-c_i^k\nabla\mu_i+\omega_i\nabla p\rVert_0\lVert u_i\rVert_0\\ \leq\kappa^{-1}\lVert p\rVert_0\left(\lVert\mu_i\rVert_0+\lVert u_i\rVert_0\lVert c_i^k\rVert_{L^\infty(\Omega)}\lVert\nabla\ln\rho^k\rVert_{L^\infty(\Omega;\mathbb{R}^d)}\right)+\lVert-c_i^k\nabla\mu_i+\omega_i\nabla p\rVert_0\lVert u_i\rVert_0. \tag*{(75)} \end{aligned}

By the bounded inverse theorem, div⁡\operatorname{div} admits a bounded left inverse, so ∥zi∥1≲∥μi∥0\lVert z_i\rVert_1\lesssim\lVert\mu_i\rVert_0 and thus

∥ui∥0≤κ−1∥zi∥0≲κ−1∥μi∥0.(76)\lVert u_i\rVert_0\leq\kappa^{-1}\lVert z_i\rVert_0\lesssim\kappa^{-1}\lVert\mu_i\rVert_0. \tag*{(76)}

Combining this with (75), we can divide through by ∥μi∥0\lVert\mu_i\rVert_0 to derive

∥μi∥0≲κ−1∥p∥0(1+κ−1∥cik∥L∞(Ω)∥∇ln⁡pk∥L∞(Ω))+κ−1∥−cik∇μi+ωik∇p∥0.(77)\lVert\mu_i\rVert_0\lesssim\kappa^{-1}\lVert p\rVert_0\left(1+\kappa^{-1}\lVert c_i^k\rVert_{L^\infty(\Omega)}\lVert\nabla\ln p^k\rVert_{L^\infty(\Omega)}\right)+\kappa^{-1}\lVert-c_i^k\nabla\mu_i+\omega_i^k\nabla p\rVert_0. \tag*{(77)}

□\square

Remark 3. The constants appearing in (68) depend on two factors; the relative variation of the density and κ\kappa. Provided these two quantities are well-behaved across iterations, so will be the resulting constants.

A further intermediate lemma we need to prove well-posedness is the following.

Lemma 2 Under Assumption 3.1 there exists λκγ>0\lambda_\kappa^\gamma>0 depending on κ\kappa and γ\gamma such that for all (v~,v)∈Q(\widetilde v,v)\in Q,

(n+12)λκγ∥v∥02+Ak(v~,v;v~,v)≥λκγ2∥(v~,v)∥Q2.(78)\left(\frac{n+1}{2}\right)\lambda_\kappa^\gamma\lVert v\rVert_0^2+A_k(\widetilde v,v;\widetilde v,v)\geq\frac{\lambda_\kappa^\gamma}{2}\lVert(\widetilde v,v)\rVert_Q^2. \tag*{(78)}

Proof. Let us define δi=vi−v\delta_i = v_i-v for each ii. Then we can explicitly compute

Ak(v~,v;v~,v)=∫Ω∑i,jδi⋅Mijk,γδj dx,(79)A_k(\tilde{v},v;\tilde{v},v)=\int_{\Omega}\sum_{i,j}\delta_i\cdot M_{ij}^{k,\gamma}\delta_j\,\mathrm{d}x, \tag*{(79)}

where Mk,γM^{k,\gamma} is defined using c~k\tilde{c}^k via (26). It follows from Van-Brunt et al. [58]’s Lemma 4.1 that this is a coercive bilinear form in δ~\tilde{\delta}: for some λkγ\lambda_k^\gamma,

Ak(v~,v;v~,v)≥λkγ2∑i∥δi∥02=λkγ2∑i∥vi−v∥02,(80)A_k(\tilde{v},v;\tilde{v},v)\geq\frac{\lambda_k^\gamma}{2}\sum_i\|\delta_i\|_0^2=\frac{\lambda_k^\gamma}{2}\sum_i\|v_i-v\|_0^2, \tag*{(80)}

hence

(n+12)λkγ∥v∥02+Ak(v~,v;v~,v)≥λkγ2∑i(∥vi−v∥02+∥vi∥02)+λkγ2∥v∥02≥λkγ2∑i∥vi∥02+λkγ2∥v∥02=λkγ2∥(v~,v)∥Q2.(81)\begin{aligned} \left(\frac{n+1}{2}\right)\lambda_k^\gamma\|v\|_0^2+A_k(\tilde{v},v;\tilde{v},v)\geq\frac{\lambda_k^\gamma}{2}\sum_i\left(\|v_i-v\|_0^2+\|v_i\|_0^2\right)+\frac{\lambda_k^\gamma}{2}\|v\|_0^2 \\ &\geq\frac{\lambda_k^\gamma}{2}\sum_i\|v_i\|_0^2+\frac{\lambda_k^\gamma}{2}\|v\|_0^2=\frac{\lambda_k^\gamma}{2}\|(\tilde{v},v)\|_Q^2. \tag*{(81)} \end{aligned}

We now invoke standard Babuška theory for well-posedness (Babuška, [8]).

Theorem 1 Under Assumption 3.1, there exists a unique solution to the perturbed saddle point system (64).

Proof. For a given (p,q):=((μ~,p,τ),(v~,v))∈Θk×Q(p,q):=((\tilde{\mu},p,\tau),(\tilde{v},v))\in\Theta^k\times Q and (s,u):=((w~,s,q),(u~,u))∈Θk×Q(s,u):=((\tilde{w},s,q),(\tilde{u},u))\in\Theta^k\times Q we will define the bounded bilinear form G:(Θk×Q)×(Θk×Q)→R\mathcal{G}:(\Theta^k\times Q)\times(\Theta^k\times Q)\to\mathbb{R} as

G(p,q;s,u):=Λ(p;s)+Bk(s;q)+Bk(p;u)−Ak(q;u).(82)\mathcal{G}(p,q;s,u):=\Lambda(p;s)+B_k(s;q)+B_k(p;u)-A_k(q;u). \tag*{(82)}

We will prove Babuška’s inf-sup condition, namely that there exists a constant c>0c>0 such that for each (p,q)∈Θk×Q(p,q)\in\Theta^k\times Q there is (s,u)∈Θk×Q(s,u)\in\Theta^k\times Q such that

G(p,q;s,u)∥(s,u)∥Θk×Q≥c∥(p,q)∥Θk×Q,(83)\frac{\mathcal{G}(p,q;s,u)}{\|(s,u)\|_{\Theta^k\times Q}}\geq c\|(p,q)\|_{\Theta^k\times Q}, \tag*{(83)}

with product norm ∥(p,q)∥Θk×Q2:=∥p∥Θk2+∥q∥Q2\|(p,q)\|_{\Theta^k\times Q}^2:=\|p\|_{\Theta^k}^2+\|q\|_Q^2. Note that G\mathcal{G} is defined on the product of a space with itself and is symmetric, and so only the one inf-sup condition (83) needs to be verified. Proving (83) will be accomplished by showing that for a constant c>0c>0, for each (p,q)∈Θk×Q(p,q)\in\Theta^k\times Q there is (s,u)∈Θk×Q(s,u)\in\Theta^k\times Q such that G(p,q;s,u)≥c∥(p,q)∥Θk×Q2\mathcal{G}(p,q;s,u)\geq c\|(p,q)\|_{\Theta^k\times Q}^2, and for a C>0C>0 independent of (p,q)(p,q),

C∥(p,q)∥Θk×Q≥∥(s,u)∥Θk×Q.(84)C\|(p,q)\|_{\Theta^k\times Q}\geq\|(s,u)\|_{\Theta^k\times Q}. \tag*{(84)}

This combined with our Poincaré-type inequality will imply well-posedness. We do this by fixing (s,u)(s,u) as the ansatz

wi=C1μi,s=C1τ+C2sv,q=C1p,ui=C3(−cik∇μi+ωik∇p)−C1vi,u=−C1v+C4(div⁡τ−∇p).(85)\begin{aligned} w_i&=C_1\mu_i,\qquad s=C_1\tau+C_2s_v,\qquad q=C_1p,\\ u_i&=C_3(-c_i^k\nabla\mu_i+\omega_i^k\nabla p)-C_1v_i,\qquad u=-C_1v+C_4(\operatorname{div}\tau-\nabla p). \tag*{(85)} \end{aligned}

Here C1,…>0C_1,\ldots>0 are constants to be set later and sv∈H(div⁡;S)s_v\in H(\operatorname{div};\mathbb{S}) is chosen to satisfy div⁡sv=v\operatorname{div}s_v=v and ∥sv∥H(div⁡;S)≤CΣ∥v∥0\|s_v\|_{H(\operatorname{div};\mathbb{S})}\leq C_\Sigma\|v\|_0 for a constant CΣC_\Sigma independent of vv (its existence is guaranteed by Arnold [2], Theorem 8.1)). It is clear that (84) holds. With these choices of test functions we may compute

G(p,q;s,u)=∫ΩAτ:(C1τ+C2sv) dx+C3∑i∥−cik∇μi+ωik∇p∥02+C4∥div⁡τ−∇p∥02+C2∥v∥02−Ak(v~,v;u~,u),(86)\mathcal{G}(\mathbf{p},\mathbf{q};\mathbf{s},\mathbf{u})=\int_{\Omega}\mathcal{A}\boldsymbol{\tau}:(C_1\tau+C_2s_v)\,\mathrm{d}x+C_3\sum_i\left\|-c_i^k\nabla\mu_i+\omega_i^k\nabla p\right\|_0^2 +C_4\left\|\operatorname{div}\boldsymbol{\tau}-\nabla p\right\|_0^2+C_2\|v\|_0^2-A_k(\tilde{v},v;\tilde{u},u), \tag*{(86)}

and observe that the final term on the right may be written equivalently as

C1Ak(v~,v;v~,v)−Ak(v~,v;{C3(−cik∇μi+ωik∇p)}i,C4(div⁡τ−∇p)).(87)C_1A_k(\tilde{v},v;\tilde{v},v)-A_k\left(\tilde{v},v;\{C_3(-c_i^k\nabla\mu_i+\omega_i^k\nabla p)\}_i,C_4(\operatorname{div}\boldsymbol{\tau}-\nabla p)\right). \tag*{(87)}

With λkγ\lambda_k^\gamma given by lemma 2 we now choose C2=∥Ak∥2(n+1)λkγC_2=\|A_k\|^2(n+1)\lambda_k^\gamma and assume that C1≥2∥Ak∥2C_1\ge2\|A_k\|^2. Then

C2∥v∥02+C1Ak(v~,v;v~,v)≥∥Ak∥2λkγ∥(v~,v)∥Q2.(88)C_2\|v\|_0^2+C_1A_k(\tilde{v},v;\tilde{v},v)\ge\|A_k\|^2\lambda_k^\gamma\|(\tilde{v},v)\|_Q^2. \tag*{(88)}

Using uniform positive definiteness of the compliance tensor, the bound on svs_v, and boundedness of the operators Λ,Ak\Lambda,A_k, we can proceed to bound (86) from below as

G(p,q;s,u)≥αC1∥τ∥02+∥Ak∥2λkγ∥(v~,v)∥Q2+C3∑i∥−cik∇μi+ωik∇p∥02+C4∥div⁡τ−∇p∥02−λkγ(n+1)CΣ∥Ak∥2∥Λ∥∥v∥0∥τ∥0−∥Ak∥∥(v~,v)∥Q(C32∑i∥−cik∇μi+ωik∇p∥02+C42∥div⁡τ−∇p∥02)1/2.(89)\begin{aligned} \mathcal{G}(\mathbf{p},\mathbf{q};\mathbf{s},\mathbf{u})\ge{}&\alpha C_1\|\boldsymbol{\tau}\|_0^2+\|A_k\|^2\lambda_k^\gamma\|(\tilde{v},v)\|_Q^2+C_3\sum_i\left\|-c_i^k\nabla\mu_i+\omega_i^k\nabla p\right\|_0^2\\ &+C_4\|\operatorname{div}\boldsymbol{\tau}-\nabla p\|_0^2-\lambda_k^\gamma(n+1)C_\Sigma\|A_k\|^2\|\Lambda\|\|v\|_0\|\boldsymbol{\tau}\|_0\\ &-\|A_k\|\|(\tilde{v},v)\|_Q\left(C_3^2\sum_i\left\|-c_i^k\nabla\mu_i+\omega_i^k\nabla p\right\|_0^2+C_4^2\|\operatorname{div}\boldsymbol{\tau}-\nabla p\|_0^2\right)^{1/2}. \tag*{(89)} \end{aligned}

Here α>0\alpha>0 is such that ∫ΩAτ:τ dx≥α∥τ∥02\int_{\Omega}\mathcal{A}\boldsymbol{\tau}:\boldsymbol{\tau}\,\mathrm{d}x\ge\alpha\|\boldsymbol{\tau}\|_0^2 for all τ∈L2(Ω;S)\boldsymbol{\tau}\in L^2(\Omega;\mathbb{S}). The desired inequality may now be derived by judiciously selecting the constants C1,C3,C4C_1,C_3,C_4 (typically by choosing C1≫C2≫max⁡(C3,C4)C_1\gg C_2\gg\max(C_3,C_4)) and using the weighted Young inequality. For concreteness, constants we might pick are

C1=(λkγCΣ∥Λ∥2(n+1)2α+2)∥Ak∥2,C3=C4=λkγ.(90)C_1=\left(\frac{\lambda_k^\gamma C_\Sigma\|\Lambda\|^2(n+1)^2}{\alpha}+2\right)\|A_k\|^2,\qquad C_3=C_4=\lambda_k^\gamma. \tag*{(90)}

With this choice and our Poincaré-type inequality from Lemma 1, combined with the inequality (84) we may derive

G(p,q;s,u)≥2α∥Ak∥2∥τ∥02+λkγ6∥Ak∥2∥(v~,v)∥Q2+λkγ4(∑i∥−cik∇μi+ωik∇p∥02+∥div⁡τ−∇p∥02)≳∥p∥Θk2+∥q∥Q2≳∥(p,q)∥Θk×Q2≳∥(p,q)∥Θk×Q∥(s,u)∥Θk×Q.(91)\begin{aligned} \mathcal{G}(\mathbf{p},\mathbf{q};\mathbf{s},\mathbf{u})\ge{}&2\alpha\|A_k\|^2\|\boldsymbol{\tau}\|_0^2+\frac{\lambda_k^\gamma}{6}\|A_k\|^2\|(\tilde{v},v)\|_Q^2\\ &+\frac{\lambda_k^\gamma}{4}\left(\sum_i\left\|-c_i^k\nabla\mu_i+\omega_i^k\nabla p\right\|_0^2+\|\operatorname{div}\boldsymbol{\tau}-\nabla p\|_0^2\right)\\ &\gtrsim\|p\|_{\Theta_k}^2+\|q\|_Q^2\gtrsim\|(\mathbf{p},\mathbf{q})\|_{\Theta_k\times Q}^2\gtrsim\|(\mathbf{p},\mathbf{q})\|_{\Theta_k\times Q}\|(\mathbf{s},\mathbf{u})\|_{\Theta_k\times Q}. \tag*{(91)} \end{aligned}

which is the statement of the Babuška condition (83).

Discretization and numerical experiments

We now assume that Ω\Omega is polytopal, and admits a quasi-uniform triangulation Th\mathcal{T}_h with simplicial elements of maximal diameter hh. Denote conforming finite element spaces for the discrete solution tuple by

(Xhn×Σh×Ph)×(Whn×Vh)⏟=:Θhk×Qh⊂(L02(Ω)n×L2(Ω;S)×L02(Ω))×(L2(Ω;Rd)n×L2(Ω;Rd))⏟Θk×Q.(92)\underbrace{(X_h^n \times\Sigma_h \times P_h) \times(W_h^n \times V_h)}_{=: \Theta_h^k \times Q_h} \subset\underbrace{(L_0^2(\Omega)^n \times L^2(\Omega;\mathbb{S}) \times L_0^2(\Omega)) \times(L^2(\Omega;\mathbb{R}^d)^n \times L^2(\Omega;\mathbb{R}^d))}_{\Theta^k \times Q}. \tag*{(92)}

Here Θhk\Theta_h^k is independent of kk as a set, but inherits an iteration-dependent norm described below; QhQ_h inherits the norm of QQ. Our discretized linear problem after k≥0k \ge0 nonlinear iterations therefore reads: seek ((μ~h,τh,ph),(v~h,vh))∈Θhk×Qh((\widetilde{\mu}_h,\tau_h,p_h),(\widetilde{v}_h,v_h)) \in\Theta_h^k \times Q_h such that

Λ(μ~h,τh,ph;w~h,sh,qh)+Bk,h(w~h,sh,qh;v~h,vh)=ℓk,h1(w~h,sh,qh)∀(w~h,sh,qh)∈Θhk,\Lambda(\widetilde{\mu}_h,\tau_h,p_h;\widetilde{w}_h,s_h,q_h)+B_{k,h}(\widetilde{w}_h,s_h,q_h;\widetilde{v}_h,v_h)=\ell_{k,h}^1(\widetilde{w}_h,s_h,q_h) \quad\forall(\widetilde{w}_h,s_h,q_h)\in\Theta_h^k,
Bk,h(μ~h,τh,ph;u~h,uh)−Ak,h(v~h,vh;u~h,uh)=ℓk,h2(u~h,uh)∀(u~h,uh)∈Qh,(93)B_{k,h}(\widetilde{\mu}_h,\tau_h,p_h;\widetilde{u}_h,u_h)-A_{k,h}(\widetilde{v}_h,v_h;\widetilde{u}_h,u_h)=\ell_{k,h}^2(\widetilde{u}_h,u_h) \quad\forall(\widetilde{u}_h,u_h)\in Q_h, \tag*{(93)}

where Ak,h,Bk,hA_{k,h},B_{k,h} are obtained from Ak,BkA_k,B_k, and ℓk,h1,ℓk,h2\ell_{k,h}^1,\ell_{k,h}^2 from ℓk1,ℓk2\ell_k^1,\ell_k^2, respectively, by replacing the discretely computed concentrations cikc_i^k and inverse density (ρk)−1(\rho^k)^{-1} with discrete approximations; we use these to define a norm ∥⋅∥Θhk\|\cdot\|_{\Theta_h^k} for Θhk\Theta_h^k in analogy to (59). In block form, the linearized discrete problem reads

\begin{bmatrix} \Lambda& B_{k,h}^{\mathsf{T}} \\ B_{k,h} & -A_{k,h} \end{bmatrix} \begin{bmatrix} \widetilde{\mu}_h \\ \tau_h \\ p_h \\ \widetilde{v}_h \\ v_h \end{bmatrix} = \begin{bmatrix} \ell_{k,h}^1 \\ \ell_{k,h}^2 \end{bmatrix}. \tag*{(94)}\end{bmatrix}

Structure-preservation and well-posedness

We have already argued the need for pressure regularity greater than L2L^2. We therefore employ the continuous Lagrange element of degree t≥1t \ge1, Ph=CG⁡t(Th)P_h=\operatorname{CG}_t(\mathcal{T}_h), for the pressure. From the diffusion driving forces (17), it appears natural to take the chemical potential space XhX_h to be CG elements of at least the same degree r≥tr \ge t, Xh=CG⁡r(Th)X_h=\operatorname{CG}_r(\mathcal{T}_h).

The mass-average velocity constraint (9) suggests that the species velocity space be contained in the space used for convective velocity, Wh⊂VhW_h \subset V_h. For the Stokes subsystem, it is desirable that the pair (Σh×Ph,Vh)(\Sigma_h \times P_h,V_h) be inf-sup compatible, for which it is sufficient to have that the full divergence (τ,p)↦div⁡τ−∇p(\tau,p) \mapsto\operatorname{div}\tau-\nabla p is surjective onto VhV_h. By the regularity choice (40) for the pressure, it is thus natural to apply div-conforming tensor elements to discretize the viscous stress. By the decomposition (23), symmetry of the viscous stress is equivalent to the conservation of angular momentum; for now, we consider exactly symmetric stress elements (such as those proposed in Arnold and Winther [3] and Arnold et al. [4]) since this obviates the need for a symmetry-enforcing Lagrange multiplier which would add a further field to our (2n+3)(2n+3)-field formulation.

However, if one would like to repeat at the discrete level the proof of Theorem 1, it is necessary for div⁡:Σh→Vh\operatorname{div}:\Sigma_h \to V_h to be surjective, allowing us to construct the discrete analogue of the tensor field svs_v in the ansatz (85). This is stronger than surjectivity of (τ,p)↦div⁡τ−∇p,Σh×Ph→Vh(\tau,p)\mapsto\operatorname{div}\tau-\nabla p,\Sigma_h\times P_h\to V_h, but in practice is equivalent because many appropriate choices of Σh\Sigma_h are designed to discretize the full Cauchy stress.

Furthermore, the discrete analogue of the constant CΣC_{\Sigma} (and hence the resulting inf-sup constant) will a priori depend on hh; it is therefore necessary to assume that such sVs_{\mathrm{V}} can be constructed stably.

Assumption 4.1. There exists CΣC^{\Sigma} independent of hh such that for each uh∈Vhu_h \in V_h and for the unique sh∈Σh/ker⁡(div⁡)s_h \in\Sigma_h/\ker(\operatorname{div}) with div⁡sh=uh\operatorname{div}s_h=u_h, there holds ∥sh∥H(div⁡;S)≤CΣ∥uh∥0\lVert s_h\rVert_{H(\operatorname{div};\mathbb{S})} \le C^{\Sigma}\lVert u_h\rVert_0.

This is true for (for example) stress elements discretizing an elasticity complex which admits bounded commuting projections to the subcomplex, as is for example the case for the Arnold–Winther and Arnold–Awanou–Winther elements. The other relations are summarised below,7

Σh×Ph→(τ,p)↦div⁡τVhπ2↓↑⊂Xh Wh(95)\begin{CD} \Sigma_h \times P_h @>{(\tau,p)\mapsto\operatorname{div}\tau}>> V_h \\ @V{\pi_2}VV @AA{\subset}A \\ X_h @. W_h \tag*{(95)} \end{CD}

where →\to indicates surjectivity. The bottom row corresponds to the final segment of a discrete stress elasticity complex, with stress space refined for Stokes flow due to the decomposition (23). We will need the conditions of Lemma 1 to hold exactly in the discretization. This will in general require that we approximate the concentrations, cikc_i^k, and density reciprocal, (ρk)−1(\rho^k)^{-1}, in specific discrete function spaces. The interpolation of these terms will be denoted by ci,hkc_{i,h}^k and ρhk,−1\rho_h^{k,-1}, respectively.

Finally, to show well-posedness of the discrete problem, we require an additional condition which does not fit neatly onto (95).

Assumption 4.2. The operator given by

dhi,k(wh,qh):=−ci,hk∇wh+ωi,hk∇qh,(96)d_h^{i,k}(w_h,q_h) := -c_{i,h}^k\nabla w_h+\omega_{i,h}^k\nabla q_h, \tag*{(96)}

where ωi,hk:=Mci,hkρhk,−1\omega_{i,h}^k := M c_{i,h}^k\rho_h^{k,-1}, is well-defined as a map dhi,k:Xh×Ph→Whd_h^{i,k}:X_h\times P_h\to W_h, i.e., it takes values in WhW_h.

Remark 4. Note that Lemma 1 required the gradient of ρ−1\rho^{-1}, and so ρhk,−1\rho_h^{k,-1} should at least be a continuous piecewise linear function. In order to minimise the polynomial degree for WhW_h arising from Assumption 4.2, it is advantageous to interpolate ci,hkc_{i,h}^k onto the space DG0\mathrm{DG}_0. These coefficients do not affect the accuracy of the discretization, only the quality of the linearization, and nonlinear convergence appears robust regardless of this approximation.

Theorem 2 Under Assumptions 4.1 and 4.2 and the relations specified in (95), the the system (93) is well-posed, uniformly in hh.

Proof. Due to the structural conditions demanded in the Assumptions, by inspection the choices of test functions (85) are valid. As a consequence we may completely replicate the argument presented in the infinite-dimensional case, and derive the analogue of condition (83) with constant independent of hh. □\square

Error estimates

Following Xu and Zikatanov [60], for fixed kk we infer the abstract error estimate

∥(μ~k+1,τk+1,pk+1)−(μ~h,τh,ph)∥Θhk+∥(v~k+1,vk+1)−(v~h,vh)∥Q≲EΘhk+EQh,(97)\left\|\left(\widetilde{\mu}^{k+1},\tau^{k+1},p^{k+1}\right)-\left(\widetilde{\mu}_h,\tau_h,p_h\right)\right\|_{\Theta_h^k}+\left\|\left(\widetilde{v}^{k+1},v^{k+1}\right)-\left(\widetilde{v}_h,v_h\right)\right\|_{Q}\lesssim\mathbb{E}_{\Theta_h^k}+\mathbb{E}_{Q_h}, \tag*{(97)}

where

EΘhk:=inf⁡(w~h,sh,qh)∈Θhk∥(μ~k+1,τk+1,pk+1)−(w~h,sh,qh)∥Θhk(98)\mathbb{E}_{\Theta_h^k}:=\inf_{(\widetilde{w}_h,s_h,q_h)\in\Theta_h^k}\left\|\left(\widetilde{\mu}^{k+1},\tau^{k+1},p^{k+1}\right)-\left(\widetilde{w}_h,s_h,q_h\right)\right\|_{\Theta_h^k} \tag*{(98)}
EQh:=inf⁡(u~h,uh)∈Qh∥(v~k+1,vk+1)−(u~h,uh)∥Q.\mathbb{E}_{Q_h}:=\inf_{(\widetilde{u}_h,u_h)\in Q_h}\left\|\left(\widetilde{v}^{k+1},v^{k+1}\right)-\left(\widetilde{u}_h,u_h\right)\right\|_{Q}.

Here the tuple ((μ~k+1,τk+1,pk+1),(v~k+1,vk+1))\left(\left(\widetilde{\mu}^{k+1},\tau^{k+1},p^{k+1}\right),\left(\widetilde{v}^{k+1},v^{k+1}\right)\right) is defined as the exact solution to (64) but with Bk,Ak,ℓk1,ℓk2B_k,A_k,\ell_k^1,\ell_k^2 replaced with Bk,h,Ak,h,ℓk,h1,ℓk,h2B_{k,h},A_{k,h},\ell_{k,h}^1,\ell_{k,h}^2, respectively—that is, the solution of the system (64) with the estimates of the concentrations and inverse density replaced by ci,hkc_{i,h}^k and ρhk,−1\rho_h^{k,-1}, respectively.

To derive a practical error estimate, we will need to bound the quantities EΘhk\mathbb{E}_{\Theta_h^k} and EQh\mathbb{E}_{Q_h} by interpolation estimates specific to the choice of finite element spaces, by estimating ∥⋅∥Θhk,∥⋅∥Q\|\cdot\|_{\Theta_h^k},\|\cdot\|_Q in terms of standard Sobolev norms. To this end we can readily check that

EΘhk≲max⁡(1,∑i∥ci,hk∥L∞(Ω))inf⁡w~h∈Xhn∥μ~k+1−w~h∥1+max⁡(1,∑i∥ωi,hk∥L∞(Ω))inf⁡qh∈Ph∥pk+1−qh∥1+inf⁡sh∈Σh∥τk+1−sh∥H(div⁡;S),\begin{aligned} \mathbb{E}_{\Theta_h^k}\lesssim{}&\max\left(1,\sum_i\|c_{i,h}^k\|_{L^\infty(\Omega)}\right)\inf_{\widetilde{w}_h\in X_h^{n}}\|\widetilde{\mu}^{k+1}-\widetilde{w}_h\|_1\\ &+\max\left(1,\sum_i\|\omega_{i,h}^k\|_{L^\infty(\Omega)}\right)\inf_{q_h\in P_h}\|p^{k+1}-q_h\|_1+\inf_{s_h\in\Sigma_h}\|\tau^{k+1}-s_h\|_{H(\operatorname{div};S)}, \end{aligned}
EQh≲inf⁡u~h∈Whn∥v~k+1−u~h∥0+inf⁡uh∈Vh∥vk+1−uh∥0.(99)\mathbb{E}_{Q_h}\lesssim\inf_{\widetilde{u}_h\in W_h^n}\|\widetilde{v}^{k+1}-\widetilde{u}_h\|_0+\inf_{u_h\in V_h}\|v^{k+1}-u_h\|_0. \tag*{(99)}

Examples of suitable finite elements

Having derived abstract error estimates, we now consider two simple combinations of finite elements satisfying the structural conditions (95) and Assumptions 4.1 and 4.2.

The design and implementation of stress elements which exactly enforce symmetry and div-conformity is notoriously difficult; in 2D, one choice of such elements is the conforming Arnold–Winther element (Arnold and Winther, [3]), recently incorporated into the Firedrake finite element library ([50, 6]). In the lowest-order case we denote this element by AW3c\mathrm{AW}_3^c. Specifying

Xh=Ph=CG1(Th)∩L02(Ω),(100)X_h=P_h=\mathrm{CG}_1(\mathcal{T}_h)\cap L_0^2(\Omega), \tag*{(100)}
Σh=AW3c(Th),(101)\Sigma_h=\mathrm{AW}_3^c(\mathcal{T}_h), \tag*{(101)}
Wh=Vh=DG1(Th;Rd),(102)W_h=V_h=\mathrm{DG}_1(\mathcal{T}_h;\mathbb{R}^d), \tag*{(102)}

and further assuming that the discretely computed cikc_i^k and (ρk)−1(\rho^k)^{-1} have been interpolated into DG0\mathrm{DG}_0 and CG1\mathrm{CG}_1, respectively, then this discretization satisfies the structural properties (95) and Assumptions 4.1 and 4.2, hence we deduce the error estimate (97).

Let ΠhCG1:H2(Ω)→CG1(Th)\Pi_h^{\mathrm{CG}_1}:H^2(\Omega)\to\mathrm{CG}_1(\mathcal{T}_h), ΠhAW3c:H1(Ω;S)→AW3c(Th)\Pi_h^{\mathrm{AW}_3^c}:H^1(\Omega;S)\to\mathrm{AW}_3^c(\mathcal{T}_h), and ΠhDG1:H1(Ω;Rd)→DG1(Th;Rd)\Pi_h^{\mathrm{DG}_1}:H^1(\Omega;\mathbb{R}^d)\to\mathrm{DG}_1(\mathcal{T}_h;\mathbb{R}^d) be the associated interpolation operators. We then have the following estimates under sufficient regularity assumptions (for details we refer to Arnold and Winther [3], Boffi et al. [12] and Logg et al. [42]):

∥μ~−ΠhCG1μ~∥1≲h∣μ~∣2,(103)\lVert\tilde{\mu}-\Pi_h^{\mathrm{CG1}}\tilde{\mu}\rVert_1 \lesssim h\lvert\tilde{\mu}\rvert_2, \tag*{(103)}
∥p−ΠhCG1p∥1≲h∣p∣2,(104)\lVert p-\Pi_h^{\mathrm{CG1}}p\rVert_1 \lesssim h\lvert p\rvert_2, \tag*{(104)}
\lVert\tau-\Pi_h^{\mathrm{AW3}}_h\tau\rVert_0+h\lVert\operatorname{div}(\tau-\Pi_h^{\mathrm{AW3}}_h\tau)\rVert_0 \lesssim h^2\lvert\tau\rvert_2, \tag*{(105)}
\lVert(\tilde{v},v)-\Pi_h^{\mathrm{DG}d}_h(\tilde{v},v)\rVert_Q \lesssim h^2\lvert(\tilde{v},v)\rvert_1, \tag*{(106)}

where \Pi_h^{\mathrm{CG1}},\Pi_h^{\mathrm{DG}d}_h have been applied component-wise. Using these estimates for the interpolation operators and the error estimate (97), we can derive the error bound

∥(μ~k+1,τk+1,pk+1)−(μ~h,τh,ph)∥Θhk+∥(v~k+1,vk+1)−(v~h,vh)∥Q≲h.(107)\lVert(\tilde{\mu}^{k+1},\tau^{k+1},p^{k+1})-(\tilde{\mu}_h,\tau_h,p_h)\rVert_{\Theta_h^k}+\lVert(\tilde{v}^{k+1},v^{k+1})-(\tilde{v}_h,v_h)\rVert_Q \lesssim h. \tag*{(107)}

We will see in practice that the order of convergence for several fields is actually higher, but the error of the species velocities and the driving forces is O(h)O(h).

A second class of finite elements may be found by replacing (100) with

Xh=CG2⁡(Th)∩L02(Ω),Ph=CG1⁡(Th)∩L02(Ω),(108)X_h=\operatorname{CG2}(\mathcal{T}_h)\cap L_0^2(\Omega),\quad P_h=\operatorname{CG1}(\mathcal{T}_h)\cap L_0^2(\Omega), \tag*{(108)}

and leaving the others unchanged. Again the structural conditions are satisfied if cikc_i^k and (ρk)−1(\rho^k)^{-1} are interpolated into DG0⁡\operatorname{DG0} and CG1⁡\operatorname{CG1}, respectively. A similar error analysis again confers an error bound of O(h)O(h), though shortly we will see that this is actually higher in practice.

Remark 5. These estimates bound the error between the discrete solutions at iteration k+1k+1, ((μ~h,τh,ph),(v~h,vh))((\tilde{\mu}_h,\tau_h,p_h),(\tilde{v}_h,v_h)) and the continuous solution of the nonlinear scheme ((μ~k+1,τk+1,pk+1),(v~k+1,vk+1))((\tilde{\mu}^{k+1},\tau^{k+1},p^{k+1}),(\tilde{v}^{k+1},v^{k+1})) with the same (discrete) coefficients. In principle this is incomplete, as ideally we would derive error estimates against the continuous solution ((μ~k+1,τk+1,pk+1),(v~k+1,vk+1))((\tilde{\mu}^{k+1},\tau^{k+1},p^{k+1}),(\tilde{v}^{k+1},v^{k+1})) at iteration k+1k+1 with the exact (continuous) coefficients. Estimates on such consistency errors were analysed for a simpler system in Van-Brunt et al. [58] and some rationale was provided as to why in practice this is not an issue, based on the formal order of the consistency error being strictly less than the discretization error. We expect a similar (if laborious) analysis could be performed, following the strategy in Van-Brunt et al. [58]. In general the consistency errors are expected to be O(h2)O(h^2), which will be borne out in the numerical examples.

Remark 6. We emphasize that we have aimed to identify appropriate structural conditions between finite element spaces in order to preserve desirable properties of the SOSM system—in particular, conditions which allow mimicry of well-posedness proofs from the infinite-dimensional problem—rather than to advocate specifically for the elements used here. We expect it is possible to use Lagrange multipliers to weakly enforce the symmetry of the viscous stress, which would allow for the choice of higher polynomial degrees.

The system matrix of our discrete linearized system (94) has symmetric perturbed saddle point structure, and although indefinite, is such that both the blocks Λ,Ak,h\Lambda,A_{k,h} are symmetric positive semidefinite. These matrix properties hold independently of the particular material considered. We expect that this structure should be conducive to the development of fast preconditioners.

Validation with manufactured solutions

We now verify our scheme, implemented in Firedrake (Rathgeber et al., [50]). Firedrake currently only supports symmetry-enforcing stress elements in 2D, and we thus restrict ourselves to the case d=2d = 2. Throughout these experiments we chose the penalty parameter γ=0.1\gamma= 0.1, and the linear systems were solved using the sparse LU factorization of MUMPS (Amestoy et al., [1]) via PETSc (Balay et al., [9]).

To validate our error estimates, we construct a manufactured solution for an ideal gas on the unit square Ω=(0,1)2\Omega= (0,1)^2. If RT=1RT = 1, the diffusion coefficients are given by Dij=DiDj\mathcal{D}_{ij} = D_iD_j for Dj>0D_j > 0, and g:R2→Rg : \mathbb{R}^2 \to\mathbb{R} is smooth, then one can check that an analytical solution to the OSM subsystem (13) is given by

ci=exp⁡(gDi),vi=Di∇g,(109)c_i = \exp\left(\frac{g}{D_i}\right), \qquad v_i = D_i\nabla g, \tag*{(109)}

which together implicitly define every other quantity (for given shear and bulk viscosities) apart from the chemical potentials. We compute the latter by inverting the ideal gas relation (29) with p⊖=∫Ωp dxp^\ominus= \int_{\Omega} p\,\mathrm{d}x, μi⊖=∫Ωci dx\mu_i^\ominus= \int_{\Omega} c_i\,\mathrm{d}x ∀i\forall i. The molar mass of each species was set to 1, and ζ,η\zeta,\eta to 0.1. The initial guesses for the concentrations c~0\tilde{c}^0 were set as ci0=∫Ωci dxc_i^0 = \int_{\Omega} c_i\,\mathrm{d}x, i.e. as the average of the exact concentrations.

We used Di=12+i4,i=1,2,3D_i = \frac{1}{2} + \frac{i}{4}, i = 1,2,3, and g(x,y)=xy5(1−x)(1−y)g(x,y) = \frac{xy}{5}(1-x)(1-y) to generate Figure 1, the log-log error plot for the overall algorithm, which demonstrates the negligible effect of the consistency error induced by the discrete interpolations ci,hk,ci,hk−1c_{i,h}^k,c_{i,h}^{k-1}, and verifies the error estimate (107).

Image not extracted

Figure 1. Error plots for two finite element families: (4.9) (left), and (108) (right).

The tolerance in the outer solver was 10−710^{-7} in the ℓ2\ell^2 norm, and took 6 nonlinear iterations on the coarsest mesh of 4×44 \times4, to 7 iterations on finest mesh of 32×3232 \times32. We have denoted in fig. 1 di,hd_{i,h} as the discrete driving force defined by (96) at the final iteration, and σh:=τh−phI\sigma_h := \tau_h - p_hI. Note that we observe O(h2)O(h^2) convergence in the L2L^2 norms of the chemical potential, stress, and pressure. As in Van-Brunt et al. [58], this suggests that the error estimates could be improved, for example by duality arguments.

Due to our construction of the ‘linearized’ function spaces (58), it is the norm ∥⋅∥kh\lVert\cdot\rVert_k^h with respect to which we have proved convergence of the solution tuple. It is natural to wonder whether this is an artefact of our constructed function spaces. To answer this, we measure convergence of the chemical potential gradients ∇μi\nabla\mu_i, pressure gradient ∇p\nabla p, and divergence div⁡τ\operatorname{div}\tau of the non-equilibrium stress to their true values, compared to the convergence of the nonlinear diffusion driving forces and the divergence of the full Cauchy stress. For the former quantities, there is a priori no reason to expect any convergence. Remarkably, we observe in Figure 2 that not only do the components ∇μi,∇p,div⁡τ\nabla\mu_i,\nabla p,\operatorname{div}\tau converge, but in fact there is convergence of the nonlinear diffusion driving forces and of the divergence of the full Cauchy stress, and at a rate one order higher than these individual components; this suggests that, rather than being a mathematical artefact of our formulation, the conditions defining the Θk\Theta^k space capture the underlying thermodynamic quantities of interest. This also provides circumstantial evidence towards the physical relevance of our nonlinear formulation in Definition 1.

Image not extracted

Figure 2. Higher-order convergence in L2L^2 of the divergence of the full Cauchy stress, and driving forces, for two finite element families: (4.9) (left), and (108) (right).

Benzene-cyclohexane mixture

Cyclohexane (C6H6C_6H_6) is important in the petrochemical industry as it is used to synthesize a variety of products, such as nylon. It is primarily produced through the dehydrogenation of benzene (C6H12C_6H_{12}), resulting in a benzene-cyclohexane mixture. Separation of cyclohexane from this mixture is difficult due to their similar vaporization temperatures [59]. Since liquid mixtures of these components are important in the chemical industry, most of the required thermodynamic and dynamical property data are readily available in the literature. Because it provides a tractable non-ideal example for which a complete set of material properties is known, we consider here a microfluidic chamber in which Stokes flows of benzene and cyclohexane mix.

The required transport parameters (measured at 298.15 K) may be found in Guevara-Carrion et al. [35]. We observe from these data that the Stefan–Maxwell diffusivity and the shear viscosity are both approximately constant with respect to composition and pressure, and will be approximated as D12=2.1×10−9 m2/s\mathcal{D}_{12}=2.1\times10^{-9}\ \mathrm{m}^2/\mathrm{s} and 6×10−4 Pa⋅s6\times10^{-4}\ \mathrm{Pa}\cdot\mathrm{s}, respectively. Lacking accurate data for the bulk viscosities of either benzene or cyclohexane, we set them to be essentially zero, choosing ζ=10−7 Pa⋅s\zeta=10^{-7}\ \mathrm{Pa}\cdot\mathrm{s}. (Numerical experiments confirmed that a value of this order has no discernible impact on the output of the simulation.) The molar masses used in the simulation are 0.078 kg⋅mol−10.078\ \mathrm{kg}\cdot\mathrm{mol}^{-1} for benzene and 0.084 kg⋅mol−10.084\ \mathrm{kg}\cdot\mathrm{mol}^{-1} for cyclohexane. The ambient pressure was taken as p⊖=105 Pap^\ominus=10^5\ \mathrm{Pa}.

Although benzene and cyclohexane are fully miscible, they form a non-ideal solution. Information relating the chemical potentials to the mole fractions is therefore required. This is accomplished using a Margules model [34] for activity coefficients, the parameters of which were reported by Tasić et al. [55]. This well-known model parameterizes activity coefficients in terms of a minimal set of functions which maintain thermodynamic rigour.

To aid convergence, we use an under-relaxation scheme with respect to the concentrations in our nonlinear solver, with a relaxation parameter of 0.1. That is, we update the concentration as ci∗,k+1c_i^{*,k+1} where

ci∗,k+1=0.9cik+0.1cik+1.(110)c_i^{*,k+1}=0.9c_i^k+0.1c_i^{k+1}. \tag*{(110)}

This is necessary due to stiffness of the problem, which owes to the fact that the mixtures are essentially fully separated at the inlets to the apparatus.

To calculate the total concentration of the mixture we use

cT=cC6H6refcC6H12refxC6H6cC6H12ref+xC6H12cC6H6ref.(111)c_{\mathrm{T}}=\frac{c_{\mathrm{C_6H_6}}^{\mathrm{ref}}c_{\mathrm{C_6H_{12}}}^{\mathrm{ref}}}{x_{\mathrm{C_6H_6}}c_{\mathrm{C_6H_{12}}}^{\mathrm{ref}}+x_{\mathrm{C_6H_{12}}}c_{\mathrm{C_6H_6}}^{\mathrm{ref}}}. \tag*{(111)}

where crefc^{\mathrm{ref}} denotes the concentration (inverse molar volume) of the pure species at 10510^5 Pa and 298.15 K, approximately 9.20 mol L−1^{-1} and 11.23 mol L−1^{-1} for benzene and cyclohexane, respectively. Equation (4.15) is derived from (31) under the assumption that the partial molar volumes of the two components are independent of the solution’s composition.

We consider a two-dimensional pipe configuration where two inlets converge into a single outlet. At the top inlet, pure benzene enters and at the bottom pure cyclohexane, at speeds vC6H12refv_{\mathrm{C_6H_{12}}}^{\mathrm{ref}} and vC6H6refv_{\mathrm{C_6H_6}}^{\mathrm{ref}}, respectively. Rather than symmetrize these speeds, superior mixing results are obtained by symmetrizing the molar fluxes at the inlets. In other words, we impose the condition

cC6H6refvC6H6ref=cC6H12refvC6H12ref.(112)c_{\mathrm{C_6H_6}}^{\mathrm{ref}}v_{\mathrm{C_6H_6}}^{\mathrm{ref}}=c_{\mathrm{C_6H_{12}}}^{\mathrm{ref}}v_{\mathrm{C_6H_{12}}}^{\mathrm{ref}}. \tag*{(112)}

We will specify that vC6H12v_{\mathrm{C_6H_{12}}} enters at a speed of 4 μ\mum s−1^{-1}. This prescribes an inlet speed for benzene of approximately 3.28 μ\mum s−1^{-1}. A parabolic profile across each inlet and outlet is imposed, as this is consistent with the no-slip condition and the characteristics of a plane Poiseuille flow.

Results for the fields computed by the simulation are visualized in Figure 3 and Figure 4. We observe that the pressure profile is smooth, despite the nonconvex domain. We also note that, although the mass-average velocity exhibits rather simple predictable behaviour, the flow fields of the individual species are significantly more complex, and that these three flow profiles are cleanly distinguished. We see that both species develop convective rolls—behaviour markedly different from the convective velocity.

Image not extracted

Figure 3. Plot of change in pressure in the mixing chamber, with streamlines computed from the mass-average velocity.

Image not extracted

Figure 4. Concentrations of benzene (left) and cyclohexane (right), with streamlines computed from their velocities.

Code availability

For reproducibility, the exact software versions used to generate the numerical results in this paper are archived on Zenodo ([7]); our implementation employs a nondimensionalization of the SOSM system. The code, and scripts for the associated plots are available at https://bitbucket.org/FAznaran/sosm-numerics/.

Conclusions and outlook

We have proposed a formulation and numerical method for the Stokes–Onsager–Stefan–Maxwell equations of multicomponent flow, proving continuous and discrete inf-sup conditions for a linearization of the system with saddle point structure. We obtained error estimates in a norm corresponding to a space requiring square-integrable diffusion driving forces and total stress divergence, and verified these with numerical experiments.

This work represents a first step towards the simulation of non-ideal mixtures; further physical extensions will be required for realistic applications in chemical engineering. Of particular interest is the analysis of the transient problem, the incorporation of thermal effects based on the framework proposed in Van-Brunt et al. [57], the weak enforcement of symmetry for the viscous stress tensor to ease the extension of the method to three dimensions (Boffi et al. [12], Section 9.2), the consideration of the case of vanishing bulk viscosity as encountered in dilute monatomic gases (Hirschfelder et al. [38]), and the incorporation of advection into the formulation.

Rigorous investigation into a notion of weak solution more refined than Definition 1 incorporating (for example) integrability of thermodynamic pressure gradients, and weak form of linearization, would also be of significant interest. We also remark that a proof of convergence of the Picard iteration could be used to prove the existence of a solution tuple for the infinite-dimensional nonlinear SOSM system.

Acknowledgements

This work was supported by the Engineering and Physical Sciences Research Council Centre for Doctoral Training in Partial Differential Equations: Analysis and Applications (EP/L015811/1), the Engineering and Physical Sciences Research Council (grants EP/R029423/1 and EP/W026163/1), The MathWorks, Inc., the Clarendon fund scholarship, and the Faraday institution SOLBAT project and Multiscale Modelling projects (subawards FIRG007 and FIRG003 under grant EP/P003532/1). The authors are also grateful to J. Málek and M. P. Juniper for useful comments.

References

  1. [1]P. R. Amestoy, I. S. Duff, L. J.-Y., and J. Koster. A fully asynchronous multifrontal solver using distributed dynamic scheduling. SIAM Journal on Matrix Analysis and Applications, 23(1):15–41, Apr. 2001. doi: 10.1137/S0895479899358194.DOI
  2. [2]D. N. Arnold. Finite element exterior calculus. Number 93 in CBMS-NSF Regional Conference Series in Applied Mathematics. SIAM, Philadelphia, 2018. doi: 10.1137/1.9781611975543.DOI
  3. [3]D. N. Arnold and R. Winther. Mixed finite elements for elasticity. Numerische Mathematik, 92(3):401–419, Sept. 2002. doi: 10.1007/s002110100348.DOI
  4. [4]D. N. Arnold, G. Awanou, and R. Winther. Finite elements for symmetric tensors in three dimensions. Mathematics of Computation, 77(262):1229–1251, July 2008. doi: 10.1090/S0025-5718-08-02071-1.DOI
  5. [5]P. Atkins and J. de Paula. Atkins’ physical chemistry. Oxford University Press, 2010.
  6. [6]F. R. A. Aznaran, P. E. Farrell, and R. C. Kirby. Transformations for Piola-mapped elements. Oct. 2021.arxiv.org/abs/2110.13224
  7. [7]F. R. A. Aznaran, P. E. Farrell, C. W. Monroe, and A. J. Van-Brunt. Software used in ‘Finite element methods for multicomponent convection-diffusion’. Aug. 2022. doi: 10.5281/zenodo.7017917. URL https://doi.org/10.5281/zenodo.7017917. https://doi.org/10.5281/zenodo.7017917.
  8. [8]I. Babuška. Error-bounds for finite element method. Numerische Mathematik, 16:322–333, Jan. 1971. doi: 10.1007/BF02165003.DOI
  9. [9]S. Balay, S. Abhyankar, M. F. Adams, J. Brown, P. Brune, K. Buschelman, L. Dalcin, V. Eijkhout, W. D. Gropp, D. Karpeyev, D. Kaushik, M. G. Knepley, D. May, L. C. McInnes, R. T. Mills, T. Munson, K. Rupp, P. Sanan, B. F. Smith, S. Zampini, H. Zhang, and H. Zhang. PETSc Users Manual. Technical Report ANL-95/11 - Revision 3.11, Argonne National Laboratory, Mar. 2019.
  10. [10]M. Benzi, G. H. Golub, and J. Liesen. Numerical solution of saddle point problems. Acta Numerica, 14:1–137, Apr. 2005. doi: 10.1017/S0962492904000212.
  11. [11]R. B. Bird, W. E. Stewart, and E. N. Lightfoot. Transport phenomena. Number 1 in Transport Phenomena. Wiley, 2nd edition, 2002.
  12. [12]D. Boffi, F. Brezzi, and M. Fortin. Mixed finite element methods and applications, volume 44 of Springer Series in Computational Mathematics. Springer, Berlin, 2013. doi: 10.1007/978-3-642-36519-5.
  13. [13]A. Bondesan, L. Boudin, and B. Grec. A numerical scheme for a kinetic model for mixtures in the diffusive limit using the moment method. Numerical Methods for Partial Differential Equations, 35(3):1184–1205, Jan. 2019. doi: 10.1002/num.22345.DOI
  14. [14]D. Bothe and W. Dreyer. Continuum thermodynamics of chemically reacting fluid mixtures. Acta Mechanica, 226 (6):1757–1805, June 2015.DOI
  15. [15]D. Bothe and P.-E. Druet. Mass transport in multicomponent compressible fluids: Local and global well-posedness in classes of strong solutions for general class-one models. Nonlinear Analysis, 210:112389, Sept. 2021. doi: 10.1016/j.na.2021.112389.DOI
  16. [16]M. Buliček, E. Maringová, and J. Málek. On nonlinear problems of parabolic type with implicit constitutive equations involving flux. Mathematical Models & Methods in Applied Sciences, 31(10):2039–2090, Aug. 2021. doi: 10.1142/S0218202521500457.DOI
  17. [17]E. Burman, A. Ern, and V. Giovangigli. Bunsen flame simulation by finite elements on adaptively refined, unstructured triangulations. Combustion Theory and Modelling, 8(1):65–84, Mar. 2004. doi: 10.1088/1364-7830/8/1/004.DOI
  18. [18]C. Cancès, V. Ehrlacher, and L. Monasse. Finite volumes for the Stefan–Maxwell cross-diffusion system. July 2020.arxiv.org/abs/2007.09951
  19. [19]J. A. Carrillo, Y. Huang, and M. Schmidtchen. Zoology of a nonlocal cross-diffusion model for two species. SIAM Journal on Applied Mathematics, 78(2):1078–1104, Apr. 2018. doi: 10.1137/17M1128782.DOI
  20. [20]C. Carstensen, J. Gedicke, and E.-J. Park. Numerical experiments for the Arnold–Winther mixed finite elements for the Stokes problem. SIAM Journal on Scientific Computing, 34(4):A2267–A2287, Aug. 2012. doi: 10.1137/10082906.DOI
  21. [22]X. Chen and A. Jüngel. Analysis of an incompressible Navier–Stokes–Maxwell–Stefan system. Communications in Mathematical Physics, 340(2):471–497, Sept. 2015. doi: 10.1007/s00220-015-2472-z.DOI
  22. [23]E. L. Cussler. Diffusion: Mass transfer in fluid systems. Cambridge Series in Chemical Engineering. Cambridge University Press, 2009. doi: 10.1017/CBO9780511805134.DOI
  23. [24]W. M. Deen. Introduction to chemical engineering fluid mechanics. Cambridge Series in Chemical Engineering. Cambridge University Press, Cambridge, 2016. doi: 10.1017/CBO9781316403464.DOI
  24. [25]M. Doyle and J. Newman. Analysis of capacity rate data for lithium batteries using simplified models of the discharge process. Journal of Applied Electrochemistry, 27:846–856, July 1997. doi: 10.1023/A:1018481030499.
  25. [26]H. C. Elman, D. J. Silvester, and A. J. Wathen. Finite elements and fast iterative solvers: With applications in incompressible fluid dynamics. Numerical Mathematics and Scientific Computation. Oxford University Press, 2nd edition, 2014. doi: 10.1093/acprof:oso/9780199678792.001.0001.DOI
  26. [27]A. Ern and V. Giovangigli. Multicomponent transport algorithms, volume 24 of Lecture Notes in Artificial Intelligence. Springer-Verlag, 1994. doi: 10.1007/978-3-540-48650-3.DOI
  27. [28]A. Ern and V. Giovangigli. Thermal diffusion effects in hydrogen-air and methane-air flames. Combustion Theory and Modelling, 2(4):349–372, Dec. 1998. doi: 10.1088/1364-7830/2/4/001.DOI
  28. [29]A. Ern and V. Giovangigli. Impact of detailed multicomponent transport on planar and counterflow hydrogen/air and methane/air flames. Combustion Science and Technology, 149(1-6):157–181, Jan. 1999. doi: 10.1080/00102209908952104.
  29. [30]A. Fick. Über Diffusion. Annalen der Physik, 170(1):59–86, 1855. doi: 10.1002/andp.18551700105.
  30. [31]V. Giovangigli. Mass conservation and singular multicomponent diffusion algorithms. IMPACT of Computing in Science and Engineering, 2(1):73–97, Mar. 1990. doi: 10.1016/0899-8248(90)90004-T.DOI
  31. [32]V. Giovangigli. Multicomponent flow modeling. Modeling and Simulation in Science, Engineering and Technology. Birkhäuser Basel, 1999. doi: 10.1007/978-1-4612-1580-6.DOI
  32. [33]P. Goyal and C. W. Monroe. New foundations of Newman’s theory for solid electrolytes: Thermodynamics and transient balances. Journal of The Electrochemical Society, 164(11):E3647–E3660, Aug. 2017. doi: 10.1149/2.0617111jes.
  33. [34]D. W. Green and R. H. Perry. Perry’s chemical engineers’ handbook. McGraw Hill Professional. McGraw-Hill Education, 8th edition, 2007.DOI
  34. [35]G. Guevara-Carrion, T. Janzen, Y. Muñoz-Muñoz, and J. Vrabeс. Mutual diffusion of binary liquid mixtures containing methanol, ethanol, acetone, benzene, cyclohexane, toluene, and carbon tetrachloride. The Journal of Chemical Physics, 144:124501, Mar. 2016. doi: 10.1063/1.4943395.
  35. [36]E. A. Guggenheim. Thermodynamics: An advanced treatment for chemists and physicists. North-Holland Books. North-Holland, 1985.
  36. [37]E. Helfand. On inversion of the linear laws of irreversible thermodynamics. The Journal of Chemical Physics, 33(2):319–322, Aug. 1960. doi: 10.1063/1.1731144.DOI
  37. [38]J. O. Hirschfelder, C. F. Curtiss, and R. B. Bird. The molecular theory of gases and liquids. John Wiley & Sons, New York, 1954. doi: 10.1002/pol.1955.120178311.
  38. [39]W. Hundsdorfer and J. G. Verwer. Numerical solution of time-dependent advection-diffusion-reaction equations. Springer Series in Computational Mathematics. Springer Berlin Heidelberg, 2013. doi: 10.1007/978-3-662-09017-6.
  39. [40]X. Huo, A. Jüngel, and A. E. Tzavaras. Existence and weak-strong uniqueness for Maxwell–Stefan–Cahn–Hilliard systems. May 2022.arxiv.org/abs/2205.06478
  40. [41]V. G. Levich. Physicochemical hydrodynamics. Prentice-Hall International Series in the Physical and Chemical Engineering Sciences. Prentice-Hall, Englewood Cliffs, New Jersey, 1962.
  41. [42]A. Logg, K.-A. Mardal, and G. N. Wells, editors. Automated solution of differential equations by the finite element method: The FEniCS book, volume 84 of Lecture Notes in Computational Science and Engineering. Springer, 2012. doi: 10.1007/978-3-642-23099-8.DOI
  42. [43]H. Manouzi and M. Farhloul. Mixed finite element analysis of a non-linear three-fields Stokes model. IMA Journal of Numerical Analysis, 21(1):143–164, Jan. 2001. doi: 10.1093/imanum/21.1.143.DOI
  43. [44]J. C. Maxwell. IV. On the dynamical theory of gases. Philosophical Transactions of the Royal Society of London, 157:49–88, Jan. 1867. doi: 10.1098/rstl.1867.0004.
  44. [45]M. McLeod and Y. Bourgault. Mixed finite element methods for addressing multi-species diffusion using the Maxwell–Stefan equations. Computer Methods in Applied Mechanics and Engineering, 279:515–535, Sept. 2014. doi: 10.1016/j.cma.2014.07.010.DOI
  45. [46]L. Onsager. Reciprocal relations in irreversible processes. I. Physical Review, 37:405–426, Feb. 1931a. doi: 10.1103/PhysRev.37.405.DOI
  46. [49]L. Ostrowski and C. Rohde. Compressible multicomponent flow in porous media with Maxwell–Stefan diffusion. Mathematical Methods in the Applied Sciences, 43(7):4200–4221, Jan. 2020. doi: 10.1002/mma.6185.DOI
  47. [50]F. Rathgeber, D. A. Ham, L. Mitchell, M. Lange, F. Luporini, A. T. T. McRae, G.-T. Bercea, G. R. Markall, and P. H. J. Kelly. Firedrake: Automating the finite element method by composing abstractions. ACM Transactions on Mathematical Software, 43(3):1–27, Dec. 2016. doi: 10.1145/2998441.DOI
  48. [51]O. Souček, V. Orava, J. Málek, and D. Bothe. A continuum model of heterogeneous catalysis: Thermodynamic framework for multicomponent bulk and surface phenomena coupled by sorption. International Journal of Engineering Science, 138:82–117, May 2019. doi: 10.1016/j.ijengsci.2019.01.001.DOI
  49. [52]J. Stefan. Über das Gleichgewicht und die Bewegung, insbesondere die Diffusion von Gasgemengen. Sitzungsberichte der Mathematisch-Naturwissenschaftlichen Classe der Kaiserlichen Akademie der Wissenschaften Wien, 2te Abteilung, 63:63–124, 1871.
  50. [53]R. Stenberg. Some new families of finite elements for the Stokes equations. Numerische Mathematik, 56(8):827–838, Aug. 1989. doi: 10.1007/BF01405291.DOI
  51. [54]M. Stynes and D. Stynes. Convection-diffusion problems. Graduate Studies in Mathematics. American Mathematical Society, 2018. doi: 10.1090/gsm/196.DOI
  52. [55]A. Tasić, B. Djordjević, D. Grozdaníć, N. Afgan, and D. Malić. Vapour-liquid equilibria of the systems acetone–benzene, benzene–cyclohexane and acetone–cyclohexane at 25°C. Chemical Engineering Science, 33(2):189–197, 1978. doi: 10.1016/0009-2509(78)85053-2.DOI
  53. [56]V. Thomée. Galerkin finite element methods for parabolic problems, volume 25 of Springer Series in Computational Mathematics. Springer Berlin Heidelberg, Berlin, 2nd edition, 2006. doi: 10.1007/3-540-33122-0.
  54. [57]A. Van-Brunt, P. E. Farrell, and C. W. Monroe. Consolidated theory of fluid thermodiffusion. AIChE Journal, 68(5):e17599, Jan. 2022. doi: 10.1002/aic.17599.
  55. [58]A. J. Van-Brunt, P. E. Farrell, and C. W. Monroe. Augmented saddle point formulation of the steady-state Stefan–Maxwell diffusion problem. IMA Journal of Numerical Analysis, Oct. 2021. doi: 10.1093/imanum/drab067.
  56. [59]J. Villaluenga and A. Tabe-Mohammadi. A review on the separation of benzene/cyclohexane mixtures by pervaporation process. Journal of Membrane Science, 169:159–174, May 2000. doi: 10.1016/S0376-7388(99)00337-3.DOI
  57. [60]J. Xu and L. Zikatanov. Some observations on Babuška and Brezzi theories. Numerische Mathematik, 94:195–202, Mar. 2003. doi: 10.1007/s002110100308.DOI

Paper details

vlnc:W_5hs8ua8r
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.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents