Introduction
Fluid–poroelastic structure interaction couples free fluid motion with deformation and filtration in a saturated porous material. Such interactions arise in tissue perfusion, arterial-wall mechanics, filtration devices, and subsurface flow. A standard mathematical description combines the incompressible Stokes or Navier–Stokes equations with Biot poroelasticity [2 , 3 ] . The interface conditions enforce conservation of normal fluid flux, accounting for both the porous skeleton motion and Darcy filtration, together with traction of balance and the Beavers–Joseph–Saffman tangential slip law. These conditions connect the fluid velocity and stress to both the skeleton motion and the pore-fluid response. Early computational work on the Navier–Stokes–Biot coupling includes the monolithic and partitioned formulations of Badia, Quaini, and Quarteroni [4 ] . Since then, numerical methods for Stokes–Biot problems with various formulations and finite element discretizations have been developed. We refer to [5 , 6 , 7 , 8 , 9 , 10 , 11 ] for related formulations and discretizations. The numerical treatment depends in part on the variational formulation chosen for each component. In a primal Darcy formulation , the filtration velocity is eliminated through Darcy’s law and the pore pressure is sought in an H 1 H^1 H 1 space. Coupled with displacement-based elasticity, this leads to a poroelastic displacement–pressure formulation, with structure velocity introduced when inertia is retained. In a dual-mixed Darcy formulation , filtration velocity and pore pressure are independent unknowns, typically in H ( div ) H(\operatorname{div}) H ( div ) and L 2 L^2 L 2 , respectively. The normal filtration flux is then a natural trace variable, and the normal-flux transmission condition is of essential type for the mixed Darcy formulation; it may be imposed weakly, for example through a Lagrange multiplier or Nitsche-type interface operators.
Several methods have been developed for the mixed setting. Bukač, Yotov, Zakerzadeh, and Zunino [5 ] used a mixed Darcy formulation and Nitsche interface operators to construct and analyze partitioned Stokes–Biot discretizations. Ambartsumyan, Khattatov, Yotov, and Zunino [7 ] imposed normal-flux continuity through a Lagrange multiplier and established stability and error estimates for mixed finite element approximations. Mixed formulations can also be applied to the mechanical variables: Li and Yotov [8 ] combined weakly symmetric stress–displacement–rotation elasticity with mixed Darcy flow and a velocity–pressure Stokes formulation. Caucao, Li, and Yotov [9 ] developed a fully mixed Stokes–Biot formulation and a multipoint stress–flux method, using interface multipliers and local elimination of stress, rotation, and filtration variables. Thus, the terms primal and mixed must be interpreted componentwise, rather than as interchangeable descriptions of the entire coupled system.
Other formulations address complementary discretization issues. Ruiz-Baier, Taffetani, Westermeyer, and Yotov [10 ] introduced total poroelastic pressure as an additional unknown and developed a mixed-primal Biot–Stokes formulation using Stokes-stable pairs for the fluid and mechanical variables. Their transmission conditions are incorporated without interface Lagrange multipliers. Cesmelioglu, Lee, and Rhebergen [11 ] proposed hybridizable discontinuous Galerkin methods with H ( div ) H(\operatorname{div}) H ( div ) -conforming velocity and displacement approximations, elementwise satisfaction of the compressibility equations, and a locking-free error analysis. These developments concern the spatial formulation and interface enforcement; the design of a time-splitting method requires an additional analysis of the lagged interface exchange.
The use of Robin transmission conditions in partitioned Stokes–Biot methods is closely related to developments in incompressible fluid–structure interaction. Fernández, Mullaert, and Vidrascu [12 ] developed explicit Robin–Neumann schemes for incompressible fluids coupled to thin-walled elastic structures, incorporating fluid–solid inertia into the interface treatment. Burman, Durst, Fernández, and Guzmán [13 ] established fully discrete stability and error estimates for a non-iterative Robin–Robin fluid–structure scheme, with interfacial stresses evaluated variationally. The same authors subsequently provided a unified Robin–Robin splitting analysis for parabolic–parabolic and parabolic–hyperbolic interface problems [14 ] . These works provide foundations for explicit interface coupling, while their models differ from Stokes–Biot interaction, where filtration and pore pressure introduce additional transmission terms.
For fluid–poroelastic interaction itself, Bukač, Yotov, and Zunino [6 ] developed a Lie operator-splitting method for a fluid interacting with a multilayered structure containing a thin elastic membrane and a thick poroelastic layer. The Nitsche-based method in [5 ] instead separates the Stokes, Darcy, and elasticity equations through time lagging and stabilization of the interface operators. Dalal, Durst, Quaini, and Yotov [15 ] developed a Robin–Robin splitting with an auxiliary interface variable representing the Robin data and one fluid and one Biot solve per time step, together with stability and time-discretization error analysis. Guo, Sun, Wang, Yue, and Zheng [16 ] proposed a fully parallelizable loosely coupled method for moving-domain Navier–Stokes–Biot interaction, using previous-time-level interface data to advance the fluid and poroelastic subproblems independently. The partition of the Biot equations, the interface unknowns, and the order in which data are exchanged differ among these approaches.
Second-order temporal accuracy introduces further requirements on the interface approximation. Oyekole and Bukač [17 ] developed loosely coupled methods based on BDF2 with extrapolated coupling terms and on a Crank–Nicolson/leap-frog construction. Kunwar, Lee, and Seelman [18 ] studied second-order time discretizations for a coupled quasi-Newtonian fluid–poroelastic system. Parrow and Bukač [19 ] developed a strongly coupled Robin–Robin partitioned method based on a refactorized Cauchy θ \theta θ -like scheme. Subsequently, they analyzed a loosely coupled Robin–Robin method with conditional stability and a suboptimal temporal error estimate of Δ t \sqrt{\Delta t} Δ t , and investigated an iterative higher-order alternative exhibiting second-order convergence in numerical experiments [20 ] .
The immediate starting point of the present work is the second-order explicit splitting of Wang, Lee, and Čanić [1 ] . That method uses a primal Darcy formulation, displacement-based elasticity, BDF2 time stepping, and second-order extrapolation of the interface data. The Stokes and Biot blocks are independent at each time step and can be solved concurrently. Its mechanical normal condition on the Biot side contains an auxiliary Robin term involving the normal structure velocity. This term cancels from the continuous transmission condition, but produces a second time difference of normal structure velocity when one occurrence is imposed at the new time and the other is extrapolated. The earlier fully discrete analysis controls this contribution under a parabolic restriction, Δ t ≲ C ∗ h 2 \Delta t \lesssim C_{*}h^{2} Δ t ≲ C ∗ h 2 .
In this paper, we retain the same two-block partition and second-order time discretization, but replace the Biot mechanical normal Robin condition by the physical Neumann-type condition
n p ⋅ σ p ( η , ϕ ) n p + ϕ = 0. \mathbf{n}_{p}\cdot\boldsymbol{\sigma}_{p}(\boldsymbol{\eta},\phi)\mathbf{n}_{p}+\phi=0. n p ⋅ σ p ( η , ϕ ) n p + ϕ = 0.
Both stress and pore pressure are evaluated within the current Biot solve. Since pore pressure is already a Biot unknown, this treatment preserves the independence of the fluid and poroelastic solves. The fluid normal condition, Darcy interface condition, and tangential Robin/BJS conditions retain their Robin form. We refer to the resulting algorithm as a Neumann–Robin splitting , listing the poroelastic mechanical normal condition first. The modification concerns the normal mechanical coupling, while filtration and tangential slip remain coupled through their respective transmission conditions.
The principal analytical benefit is the removal of the extrapolated normal structure-velocity residual. In the earlier energy argument, this residual is paired with the current normal structure velocity, and estimating both traces through bulk L 2 L^{2} L 2 norms generates the factor associated with the parabolic restriction. In the modified scheme, each remaining explicit interface term can be estimated using at most one inverse trace, together with fluid viscous dissipation, Darcy dissipation, or tangential slip control. For fixed positive material and interface coefficients, we establish a closed homogeneous fully discrete energy estimate under the linear restriction Δ t ≤ C ∗ h \Delta t\leq C_{*}h Δ t ≤ C ∗ h . The extension to nonhomogeneous source terms is discussed separately. Our comparison here only concerns the sufficient conditions obtained for this scheme and the specific second-order method in [1 ] , and thus it does not imply that all Robin–Robin formulations require a parabolic restriction.
We establish unique solvability of the two independent subproblems and derive an exact BDF2 discrete energy identity for the proposed Neumann–Robin algorithm. For the homogeneous problem, we prove a fully discrete energy stability estimate under a linear mesh-dependent time-step restriction. Under stated regularity, finite element approximation, and initialization assumptions, we further establish a bulk energy-norm error estimate of order h k + Δ t 2 h^{k}+\Delta t^{2} h k + Δ t 2 for the fluid velocity, structure velocity, pore pressure, and displacement. The error analysis uses a common elasticity projection for displacement and structure velocity, together with a discrete summation-by-parts argument to control the pore-pressure projection traction term without requiring normal structure-velocity dissipation. Finally, we construct a smooth polynomial manufactured solution satisfying the physical interface conditions without additional interface loads and present spatial and temporal convergence studies. These experiments assess the accuracy of the proposed splitting scheme, while the improvement in the sufficient time-step restriction is established through the theoretical stability analysis.
The paper is organized as follows. Section 2 introduces the continuous problem and Section 3 presents the Neumann–Robin algorithm. Stability is established in Section 4, and the error estimate is proved in Section 5. Section 6 discusses the improvement in comparison with the Robin–Robin splitting, and Section 7 presents spatial and temporal convergence studies using manufactured solutions, compares the Neumann-Robin and Robin-Robin splitting schemes, and investigates their numerical stability behavior in a 3D cylindrical pressure-pulse benchmark.
The continuous Stokes–Biot problem
Domains and equations
Let Ω f , Ω p ⊂ R d \Omega_f,\Omega_p \subset\mathbb{R}^d Ω f , Ω p ⊂ R d , d ∈ { 2 , 3 } d \in\{2,3\} d ∈ { 2 , 3 } , be bounded Lipschitz domains with common interface Γ \Gamma Γ . The remaining boundaries are denoted by Σ f \Sigma_f Σ f and Σ p \Sigma_p Σ p . Both subdomains and the interface are fixed. Their outward normals satisfy n p = − n f n_p=-n_f n p = − n f on Γ \Gamma Γ , and
P v : = v − ( v ⋅ n f ) n f = v − ( v ⋅ n p ) n p P v := v - (v \cdot n_f)n_f = v - (v \cdot n_p)n_p P v := v − ( v ⋅ n f ) n f = v − ( v ⋅ n p ) n p
is the common tangential projector. We assume that the exterior Dirichlet portions have positive measure and provide the Korn and Poincaré inequalities used below.
The fluid velocity and pressure are ( u , p f ) (u,p_f) ( u , p f ) , and the poroelastic displacement, velocity, and pore pressure are ( η , ξ , ϕ ) (\eta,\xi,\phi) ( η , ξ , ϕ ) . With D ( v ) = ( ∇ v + ∇ v T ) / 2 D(v)=(\nabla v+\nabla v^{\mathsf T})/2 D ( v ) = ( ∇ v + ∇ v T ) /2 , define
σ f ( u , p f ) = 2 μ f D ( u ) − p f I , σ p ( η , ϕ ) = 2 μ p D ( η ) + λ p ( ∇ ⋅ η ) I − α ϕ I , (1) \begin{aligned}
\sigma_f(u,p_f) &= 2\mu_f D(u)-p_f I,\\
\sigma_p(\eta,\phi) &= 2\mu_p D(\eta)+\lambda_p(\nabla\cdot\eta)I-\alpha\phi I, \tag*{(1)}
\end{aligned} σ f ( u , p f ) σ p ( η , ϕ ) = 2 μ f D ( u ) − p f I , = 2 μ p D ( η ) + λ p ( ∇ ⋅ η ) I − α ϕ I , (1)
where I I I is the d × d d \times d d × d identity matrix. The governing equations are
{ ρ f ∂ t u − ∇ ⋅ σ f ( u , p f ) = f f , in Ω f , ∇ ⋅ u = 0 , in Ω f , ρ p ∂ t ξ − ∇ ⋅ σ p ( η , ϕ ) = f p , in Ω p , ∂ t η = ξ , in Ω p , C 0 ∂ t ϕ + α ∇ ⋅ ξ − ∇ ⋅ ( K ∇ ϕ ) = g p , in Ω p , u p = − K ∇ ϕ , in Ω p . (2) \left\{
\begin{aligned}
\rho_f\partial_t u-\nabla\cdot\sigma_f(u,p_f)&=f_f, &&\text{in }\Omega_f,\\
\nabla\cdot u&=0, &&\text{in }\Omega_f,\\
\rho_p\partial_t\xi-\nabla\cdot\sigma_p(\eta,\phi)&=f_p, &&\text{in }\Omega_p,\\
\partial_t\eta&=\xi, &&\text{in }\Omega_p,\\
C_0\partial_t\phi+\alpha\nabla\cdot\xi-\nabla\cdot(K\nabla\phi)&=g_p, &&\text{in }\Omega_p,\\
u_p&=-K\nabla\phi, &&\text{in }\Omega_p.
\end{aligned}
\right. \tag*{(2)} ⎩ ⎨ ⎧ ρ f ∂ t u − ∇ ⋅ σ f ( u , p f ) ∇ ⋅ u ρ p ∂ t ξ − ∇ ⋅ σ p ( η , ϕ ) ∂ t η C 0 ∂ t ϕ + α ∇ ⋅ ξ − ∇ ⋅ ( K ∇ ϕ ) u p = f f , = 0 , = f p , = ξ , = g p , = − K ∇ ϕ , in Ω f , in Ω f , in Ω p , in Ω p , in Ω p , in Ω p . (2)
Here ρ f , ρ p , μ f , μ p , C 0 > 0 \rho_f,\rho_p,\mu_f,\mu_p,C_0>0 ρ f , ρ p , μ f , μ p , C 0 > 0 , λ p ≥ 0 \lambda_p\ge0 λ p ≥ 0 , and the Biot–Willis coefficient α ∈ R \alpha\in\mathbb{R} α ∈ R is fixed [1 ] . The time-independent conductivity K ∈ L ∞ ( Ω p ) d × d K \in L^\infty(\Omega_p)^{d\times d} K ∈ L ∞ ( Ω p ) d × d is symmetric and uniformly positive definite. We impose the homogeneous Dirichlet boundary conditions, namely u = 0 u=0 u = 0 on Σ f \Sigma_f Σ f , η = ξ = 0 \eta=\xi=0 η = ξ = 0 on Σ p \Sigma_p Σ p , and ϕ = 0 \phi=0 ϕ = 0 on Σ p \Sigma_p Σ p , and prescribe compatible initial data. We note that these boundary conditions are part of the analytical setting, rather than a claim covering every application boundary condition.
Coupling conditions and energy
On the fluid structure interface Γ \Gamma Γ , we consider the following physical coupling conditions, as reported in (cf. [1 ] )
u ⋅ n f = ( ξ − K ∇ ϕ ) ⋅ n f , (3) \mathbf{u}\cdot\mathbf{n}_{f}=(\boldsymbol{\xi}-K\nabla\phi)\cdot\mathbf{n}_{f}, \tag*{(3)} u ⋅ n f = ( ξ − K ∇ ϕ ) ⋅ n f , (3)
P σ f n f = − γ P ( u − ξ ) , (4) P\boldsymbol{\sigma}_{f}\mathbf{n}_{f}=-\gamma P(\mathbf{u}-\boldsymbol{\xi}), \tag*{(4)} P σ f n f = − γ P ( u − ξ ) , (4)
n f ⋅ σ f n f = − ϕ , (5) \mathbf{n}_{f}\cdot\boldsymbol{\sigma}_{f}\mathbf{n}_{f}=-\phi, \tag*{(5)} n f ⋅ σ f n f = − ϕ , (5)
σ f n f = − σ p n p . (6) \boldsymbol{\sigma}_{f}\mathbf{n}_{f}=-\boldsymbol{\sigma}_{p}\mathbf{n}_{p}. \tag*{(6)} σ f n f = − σ p n p . (6)
Where γ > 0 \gamma>0 γ > 0 is the Beavers–Joseph–Saffman coefficient. In particular, we can derive the mechanical normal condition on the porous side from (5) and (6) , satisfying
n p ⋅ σ p ( η , ϕ ) n p + ϕ = 0. (7) \mathbf{n}_{p}\cdot\boldsymbol{\sigma}_{p}(\boldsymbol{\eta},\phi)\mathbf{n}_{p}+\phi=0. \tag*{(7)} n p ⋅ σ p ( η , ϕ ) n p + ϕ = 0. (7)
It is Neumann-type with respect to the mechanics equation and remains coupled to the unknown pore pressure.
We consider the following function spaces
V f = { v ∈ H 1 ( Ω f ) d : v ∣ Σ f = 0 } , Q f = L 2 ( Ω f ) , \mathbf{V}_{f}=\{\mathbf{v}\in H^{1}(\Omega_{f})^{d}:\mathbf{v}|_{\Sigma_{f}}=0\},\qquad Q_{f}=L^{2}(\Omega_{f}), V f = { v ∈ H 1 ( Ω f ) d : v ∣ Σ f = 0 } , Q f = L 2 ( Ω f ) ,
V p = { ζ ∈ H 1 ( Ω p ) d : ζ ∣ Σ p = 0 } , Q p = { ψ ∈ H 1 ( Ω p ) : ψ ∣ Σ p = 0 } . \mathbf{V}_{p}=\{\boldsymbol{\zeta}\in H^{1}(\Omega_{p})^{d}:\boldsymbol{\zeta}|_{\Sigma_{p}}=0\},\qquad Q_{p}=\{\psi\in H^{1}(\Omega_{p}):\psi|_{\Sigma_{p}}=0\}. V p = { ζ ∈ H 1 ( Ω p ) d : ζ ∣ Σ p = 0 } , Q p = { ψ ∈ H 1 ( Ω p ) : ψ ∣ Σ p = 0 } .
For inner products ( ⋅ , ⋅ ) i (\cdot,\cdot)_{i} ( ⋅ , ⋅ ) i on L 2 ( Ω i ) L^{2}(\Omega_{i}) L 2 ( Ω i ) ( i = f , p ) (i=f,p) ( i = f , p ) and interface pairing ⟨ ⋅ , ⋅ ⟩ Γ \langle\cdot,\cdot\rangle_{\Gamma} ⟨ ⋅ , ⋅ ⟩ Γ , we introduce the following bilinear forms and norm:
a f ( u , v ) = 2 μ f ( D ( u ) , D ( v ) ) f , a_{f}(\mathbf{u},\mathbf{v})=2\mu_{f}(D(\mathbf{u}),D(\mathbf{v}))_{f}, a f ( u , v ) = 2 μ f ( D ( u ) , D ( v ) ) f ,
a e ( η , ζ ) = 2 μ p ( D ( η ) , D ( ζ ) ) p + λ p ( ∇ ⋅ η , ∇ ⋅ ζ ) p , a_{e}(\boldsymbol{\eta},\boldsymbol{\zeta})=2\mu_{p}(D(\boldsymbol{\eta}),D(\boldsymbol{\zeta}))_{p}+\lambda_{p}(\nabla\cdot\boldsymbol{\eta},\nabla\cdot\boldsymbol{\zeta})_{p}, a e ( η , ζ ) = 2 μ p ( D ( η ) , D ( ζ ) ) p + λ p ( ∇ ⋅ η , ∇ ⋅ ζ ) p ,
d ( ϕ , ψ ) = ( K ∇ ϕ , ∇ ψ ) p , ∥ η ∥ a e 2 = a e ( η , η ) . d(\phi,\psi)=(K\nabla\phi,\nabla\psi)_{p},\qquad\|\boldsymbol{\eta}\|_{a_{e}}^{2}=a_{e}(\boldsymbol{\eta},\boldsymbol{\eta}). d ( ϕ , ψ ) = ( K ∇ ϕ , ∇ ψ ) p , ∥ η ∥ a e 2 = a e ( η , η ) .
The coupled weak formulation of (2) , with ∂ t η = ξ \partial_{t}\boldsymbol{\eta}=\boldsymbol{\xi} ∂ t η = ξ , is
ρ f ( ∂ t u , v ) f + a f ( u , v ) − ( p f , ∇ ⋅ v ) f + ( ∇ ⋅ u , q ) f + ⟨ ϕ , v ⋅ n f ⟩ Γ + γ ⟨ P ( u − ξ ) , P v ⟩ Γ = ( f f , v ) f , (8) \rho_{f}(\partial_{t}\mathbf{u},\mathbf{v})_{f}+a_{f}(\mathbf{u},\mathbf{v})-(p_{f},\nabla\cdot\mathbf{v})_{f}+(\nabla\cdot\mathbf{u},q)_{f}+\langle\phi,\mathbf{v}\cdot\mathbf{n}_{f}\rangle_{\Gamma}+\gamma\langle P(\mathbf{u}-\boldsymbol{\xi}),P\mathbf{v}\rangle_{\Gamma}=(\mathbf{f}_{f},\mathbf{v})_{f}, \tag*{(8)} ρ f ( ∂ t u , v ) f + a f ( u , v ) − ( p f , ∇ ⋅ v ) f + ( ∇ ⋅ u , q ) f + ⟨ ϕ , v ⋅ n f ⟩ Γ + γ ⟨ P ( u − ξ ) , P v ⟩ Γ = ( f f , v ) f , (8)
ρ p ( ∂ t ξ , ζ ) p + a e ( η , ζ ) − α ( ϕ , ∇ ⋅ ζ ) p + ⟨ ϕ , ζ ⋅ n p ⟩ Γ + γ ⟨ P ( ξ − u ) , P ζ ⟩ Γ = ( f p , ζ ) p , (9) \rho_{p}(\partial_{t}\boldsymbol{\xi},\boldsymbol{\zeta})_{p}+a_{e}(\boldsymbol{\eta},\boldsymbol{\zeta})-\alpha(\phi,\nabla\cdot\boldsymbol{\zeta})_{p}+\langle\phi,\boldsymbol{\zeta}\cdot\mathbf{n}_{p}\rangle_{\Gamma}+\gamma\langle P(\boldsymbol{\xi}-\mathbf{u}),P\boldsymbol{\zeta}\rangle_{\Gamma}=(\mathbf{f}_{p},\boldsymbol{\zeta})_{p}, \tag*{(9)} ρ p ( ∂ t ξ , ζ ) p + a e ( η , ζ ) − α ( ϕ , ∇ ⋅ ζ ) p + ⟨ ϕ , ζ ⋅ n p ⟩ Γ + γ ⟨ P ( ξ − u ) , P ζ ⟩ Γ = ( f p , ζ ) p , (9)
C 0 ( ∂ t ϕ , ψ ) p + α ( ∇ ⋅ ξ , ψ ) p + d ( ϕ , ψ ) − ⟨ ξ ⋅ n p , ψ ⟩ Γ + ⟨ u ⋅ n p , ψ ⟩ Γ = ( g p , ψ ) p . (10) C_{0}(\partial_{t}\phi,\psi)_{p}+\alpha(\nabla\cdot\boldsymbol{\xi},\psi)_{p}+d(\phi,\psi)-\langle\boldsymbol{\xi}\cdot\mathbf{n}_{p},\psi\rangle_{\Gamma}+\langle\mathbf{u}\cdot\mathbf{n}_{p},\psi\rangle_{\Gamma}=(g_{p},\psi)_{p}. \tag*{(10)} C 0 ( ∂ t ϕ , ψ ) p + α ( ∇ ⋅ ξ , ψ ) p + d ( ϕ , ψ ) − ⟨ ξ ⋅ n p , ψ ⟩ Γ + ⟨ u ⋅ n p , ψ ⟩ Γ = ( g p , ψ ) p . (10)
Taking ( v , q , ζ , ψ ) = ( u , p f , ξ , ϕ ) (\mathbf{v},q,\boldsymbol{\zeta},\psi)=(\mathbf{u},p_{f},\boldsymbol{\xi},\phi) ( v , q , ζ , ψ ) = ( u , p f , ξ , ϕ ) as the test functions cancels both the volume Biot coupling terms and the paired normal pore-pressure and mechanical terms. The normal fluid–Darcy exchange terms also cancel because the interface normals are oppositely oriented. Hence, for smooth solutions,
d d t 1 2 ( ρ f ∥ u ∥ f 2 + ρ p ∥ ξ ∥ p 2 + ∥ η ∥ a e 2 + C 0 ∥ ϕ ∥ p 2 ) + 2 μ f ∥ D ( u ) ∥ f 2 + ∥ K 1 / 2 ∇ ϕ ∥ p 2 + γ ∥ P ( u − ξ ) ∥ Γ 2 = ( f f , u ) f + ( f p , ξ ) p + ( g p , ϕ ) p . (11) \frac{d}{dt}\frac{1}{2}\left(\rho_{f}\|\mathbf{u}\|_{f}^{2}+\rho_{p}\|\boldsymbol{\xi}\|_{p}^{2}+\|\boldsymbol{\eta}\|_{a_{e}}^{2}+C_{0}\|\phi\|_{p}^{2}\right)+2\mu_{f}\|D(\mathbf{u})\|_{f}^{2}+\|K^{1/2}\nabla\phi\|_{p}^{2}+\gamma\|P(\mathbf{u}-\boldsymbol{\xi})\|_{\Gamma}^{2}=(\mathbf{f}_{f},\mathbf{u})_{f}+(\mathbf{f}_{p},\boldsymbol{\xi})_{p}+(g_{p},\phi)_{p}. \tag*{(11)} d t d 2 1 ( ρ f ∥ u ∥ f 2 + ρ p ∥ ξ ∥ p 2 + ∥ η ∥ a e 2 + C 0 ∥ ϕ ∥ p 2 ) + 2 μ f ∥ D ( u ) ∥ f 2 + ∥ K 1/2 ∇ ϕ ∥ p 2 + γ ∥ P ( u − ξ ) ∥ Γ 2 = ( f f , u ) f + ( f p , ξ ) p + ( g p , ϕ ) p . (11)
Throughout, ∥ ⋅ ∥ i \|\cdot\|_{i} ∥ ⋅ ∥ i , i = f , p , Γ i=f,p,\Gamma i = f , p , Γ denotes the L 2 L^{2} L 2 norm on the corresponding domain Ω i \Omega_{i} Ω i or interface Γ \Gamma Γ .
A Neumann–Robin splitting scheme
Modified interface conditions
For a fixed L > 0 L > 0 L > 0 , the fluid-side conditions can be written as
n f ⋅ σ f n f + L u ⋅ n f = L u ⋅ n f − ϕ , (12) \mathbf{n}_f \cdot\boldsymbol{\sigma}_f \mathbf{n}_f + L\mathbf{u} \cdot\mathbf{n}_f = L\mathbf{u} \cdot\mathbf{n}_f - \phi, \tag*{(12)} n f ⋅ σ f n f + L u ⋅ n f = L u ⋅ n f − ϕ , (12)
P σ f n f + γ P u = γ P ξ . (13) \mathbf{P}\boldsymbol{\sigma}_f\mathbf{n}_f + \gamma\mathbf{P}\mathbf{u} = \gamma\mathbf{P}\boldsymbol{\xi}. \tag*{(13)} P σ f n f + γ Pu = γ P ξ . (13)
For the same L > 0 L > 0 L > 0 , the poroelastic-side conditions can be written as
n p ⋅ σ p n p + ϕ = 0 , (14) \mathbf{n}_p \cdot\boldsymbol{\sigma}_p\mathbf{n}_p + \phi= 0, \tag*{(14)} n p ⋅ σ p n p + ϕ = 0 , (14)
K ∇ ϕ ⋅ n p + L − 1 ϕ − ξ ⋅ n p = L − 1 ϕ − u ⋅ n p , (15) K\nabla\phi\cdot\mathbf{n}_p + L^{-1}\phi- \boldsymbol{\xi} \cdot\mathbf{n}_p = L^{-1}\phi- \mathbf{u} \cdot\mathbf{n}_p, \tag*{(15)} K ∇ ϕ ⋅ n p + L − 1 ϕ − ξ ⋅ n p = L − 1 ϕ − u ⋅ n p , (15)
P σ p n p + γ P ξ = γ P u . (16) \mathbf{P}\boldsymbol{\sigma}_p\mathbf{n}_p + \gamma\mathbf{P}\boldsymbol{\xi} = \gamma\mathbf{P}\mathbf{u}. \tag*{(16)} P σ p n p + γ P ξ = γ Pu . (16)
The nonzero right-hand sides will be extrapolated in splitting algorithms. In (14) , both the stress and pore pressure belong to the current Biot solve, and thus no current fluid unknown appears.
Fully discrete algorithm
For finite element discretization we consider conforming, shape-regular, quasi-uniform meshes of Ω f \Omega_f Ω f and Ω p \Omega_p Ω p which are fitted with the interface Γ \Gamma Γ . We also assume mesh size h ≤ 1 h \le1 h ≤ 1 throughout this paper. We choose conforming finite element spaces V f , h × Q f , h ⊂ V f × Q f V_{f,h} \times Q_{f,h} \subset V_f \times Q_f V f , h × Q f , h ⊂ V f × Q f , V p , h ⊂ V p V_{p,h} \subset V_p V p , h ⊂ V p , and Q p , h ⊂ Q p Q_{p,h} \subset Q_p Q p , h ⊂ Q p . Displacement and structure velocity use the same space V p , h V_{p,h} V p , h for discretization. We assume that there exists β f > 0 \beta_f > 0 β f > 0 independent of h h h such that
inf 0 ≠ q h ∈ Q f , h sup 0 ≠ v h ∈ V f , h ( ∇ ⋅ v h , q h ) f ∥ v h ∥ H 1 ( Ω f ) ∥ q h ∥ f ≥ β f > 0. (17) \inf_{0 \ne q_h \in Q_{f,h}}\sup_{0 \ne v_h \in V_{f,h}} \frac{(\nabla\cdot v_h,q_h)_f}{\lVert v_h\rVert_{H^1(\Omega_f)}\lVert q_h\rVert_f} \ge\beta_f > 0. \tag*{(17)} 0 = q h ∈ Q f , h inf 0 = v h ∈ V f , h sup ∥ v h ∥ H 1 ( Ω f ) ∥ q h ∥ f ( ∇ ⋅ v h , q h ) f ≥ β f > 0. (17)
There are well-known finite elements satisfying this condition, such as the MINI element [21 ] and the Taylor–Hood element [22 ] ; their inf–sup stability under mild mesh conditions is analyzed in [23 Chapter 8] . These results are usually stated for velocities vanishing on the whole boundary and mean-zero pressures. Since the functions in V f , h V_{f,h} V f , h vanish only on Σ f \Sigma_f Σ f , (17) is required for the full pressure space Q f , h Q_{f,h} Q f , h . The extension to the full pressure space follows from the mean-zero inf–sup condition and a family of discrete velocities whose H 1 H^1 H 1 norms are uniformly bounded and whose fluxes through Γ \Gamma Γ are bounded away from zero. Such velocities can be obtained, for sufficiently small h h h , by boundary-preserving interpolation of a fixed velocity with nonzero interface flux.
Let T > 0 T > 0 T > 0 be the final time, let N N N be a positive integer, and set Δ t = T / N \Delta t = T/N Δ t = T / N . For a uniform temporal grid t n = n Δ t t_n = n\Delta t t n = n Δ t , n = 0 , … , N n = 0,\ldots,N n = 0 , … , N , define the BDF2 approximation and second-order linear extrapolation (EX2) as
D t a n + 1 = 3 a n + 1 − 4 a n + a n − 1 2 Δ t , a ∗ , n + 1 = 2 a n − a n − 1 , n ≥ 1. (18) D_t a^{n+1} = \frac{3a^{n+1} - 4a^n + a^{n-1}}{2\Delta t}, \qquad a^{*,n+1} = 2a^n - a^{n-1}, \qquad n \ge1. \tag*{(18)} D t a n + 1 = 2Δ t 3 a n + 1 − 4 a n + a n − 1 , a ∗ , n + 1 = 2 a n − a n − 1 , n ≥ 1. (18)
For a time-dependent function w w w , we write w n = w ( t n ) w^n = w(t_n) w n = w ( t n ) . In particular, f f n + 1 = f f ( t n + 1 ) \mathbf{f}_f^{n+1} = \mathbf{f}_f(t_{n+1}) f f n + 1 = f f ( t n + 1 ) , f p n + 1 = f p ( t n + 1 ) \mathbf{f}_p^{n+1} = \mathbf{f}_p(t_{n+1}) f p n + 1 = f p ( t n + 1 ) , and g p n + 1 = g p ( t n + 1 ) g_p^{n+1} = g_p(t_{n+1}) g p n + 1 = g p ( t n + 1 ) . The operators in (18) are applied in the same way to sampled exact and projected fields.
Let
a f R ( u , v ) = a f ( u , v ) + γ ⟨ P u , P v ⟩ Γ + L ⟨ u ⋅ n f , v ⋅ n f ⟩ Γ , (19) a_f^R(\mathbf{u},\mathbf{v}) = a_f(\mathbf{u},\mathbf{v}) + \gamma\langle\mathbf{P}\mathbf{u},\mathbf{P}\mathbf{v}\rangle_\Gamma+ L\langle\mathbf{u}\cdot\mathbf{n}_f,\mathbf{v}\cdot\mathbf{n}_f\rangle_\Gamma, \tag*{(19)} a f R ( u , v ) = a f ( u , v ) + γ ⟨ Pu , Pv ⟩ Γ + L ⟨ u ⋅ n f , v ⋅ n f ⟩ Γ , (19)
d R ( ϕ , ψ ) = d ( ϕ , ψ ) + L − 1 ⟨ ϕ , ψ ⟩ Γ . (20) d^R(\phi,\psi) = d(\phi,\psi) + L^{-1}\langle\phi,\psi\rangle_\Gamma. \tag*{(20)} d R ( ϕ , ψ ) = d ( ϕ , ψ ) + L − 1 ⟨ ϕ , ψ ⟩ Γ . (20)
Given levels n n n and n − 1 n-1 n − 1 , perform the following two solves independently.
Fluid solve. Find ( u h n + 1 , p f , h n + 1 ) ∈ V f , h × Q f , h (u_h^{n+1},p_{f,h}^{n+1}) \in V_{f,h} \times Q_{f,h} ( u h n + 1 , p f , h n + 1 ) ∈ V f , h × Q f , h satisfying
ρ f ( D t u h n + 1 , v h ) f + a f R ( u h n + 1 , v h ) − ( p f , h n + 1 , ∇ ⋅ v h ) f + ( ∇ ⋅ u h n + 1 , q h ) f = ( f f n + 1 , v h ) f + γ ⟨ P ξ h ∗ , n + 1 , P v h ⟩ Γ + ⟨ L u h ∗ , n + 1 ⋅ n f − ϕ h ∗ , n + 1 , v h ⋅ n f ⟩ Γ (21) \rho_f(D_tu_h^{n+1},v_h)_f+a_f^R(u_h^{n+1},v_h)-(p_{f,h}^{n+1},\nabla\cdot v_h)_f+(\nabla\cdot u_h^{n+1},q_h)_f
= (f_f^{n+1},v_h)_f+\gamma\langle P\xi_h^{*,n+1},Pv_h\rangle_\Gamma+\langle Lu_h^{*,n+1}\cdot n_f-\phi_h^{*,n+1},v_h\cdot n_f\rangle_\Gamma \tag*{(21)} ρ f ( D t u h n + 1 , v h ) f + a f R ( u h n + 1 , v h ) − ( p f , h n + 1 , ∇ ⋅ v h ) f + ( ∇ ⋅ u h n + 1 , q h ) f = ( f f n + 1 , v h ) f + γ ⟨ P ξ h ∗ , n + 1 , P v h ⟩ Γ + ⟨ L u h ∗ , n + 1 ⋅ n f − ϕ h ∗ , n + 1 , v h ⋅ n f ⟩ Γ (21)
for ( v h , q h ) ∈ V f , h × Q f , h (v_h,q_h)\in V_{f,h}\times Q_{f,h} ( v h , q h ) ∈ V f , h × Q f , h .
Biot solve. Find ( η h n + 1 , ξ h n + 1 , ϕ h n + 1 ) ∈ V p , h × V p , h × Q p , h (\eta_h^{n+1},\xi_h^{n+1},\phi_h^{n+1})\in V_{p,h}\times V_{p,h}\times Q_{p,h} ( η h n + 1 , ξ h n + 1 , ϕ h n + 1 ) ∈ V p , h × V p , h × Q p , h such that
( D t η h n + 1 , χ h ) p = ( ξ h n + 1 , χ h ) p , (22) (D_t\eta_h^{n+1},\chi_h)_p=(\xi_h^{n+1},\chi_h)_p, \tag*{(22)} ( D t η h n + 1 , χ h ) p = ( ξ h n + 1 , χ h ) p , (22)
ρ p ( D t ξ h n + 1 , ζ h ) p + a e ( η h n + 1 , ζ h ) − α ( ϕ h n + 1 , ∇ ⋅ ζ h ) p + γ ⟨ P ξ h n + 1 , P ζ h ⟩ Γ + ⟨ ϕ h n + 1 , ζ h ⋅ n p ⟩ Γ = ( f p n + 1 , ζ h ) p + γ ⟨ P u h ∗ , n + 1 , P ζ h ⟩ Γ , (23) \rho_p(D_t\xi_h^{n+1},\zeta_h)_p+a_e(\eta_h^{n+1},\zeta_h)-\alpha(\phi_h^{n+1},\nabla\cdot\zeta_h)_p+\gamma\langle P\xi_h^{n+1},P\zeta_h\rangle_\Gamma+\langle\phi_h^{n+1},\zeta_h\cdot n_p\rangle_\Gamma
= (f_p^{n+1},\zeta_h)_p+\gamma\langle Pu_h^{*,n+1},P\zeta_h\rangle_\Gamma, \tag*{(23)} ρ p ( D t ξ h n + 1 , ζ h ) p + a e ( η h n + 1 , ζ h ) − α ( ϕ h n + 1 , ∇ ⋅ ζ h ) p + γ ⟨ P ξ h n + 1 , P ζ h ⟩ Γ + ⟨ ϕ h n + 1 , ζ h ⋅ n p ⟩ Γ = ( f p n + 1 , ζ h ) p + γ ⟨ P u h ∗ , n + 1 , P ζ h ⟩ Γ , (23)
C 0 ( D t ϕ h n + 1 , ψ h ) p + α ( ∇ ⋅ ξ h n + 1 , ψ h ) p + d R ( ϕ h n + 1 , ψ h ) − ⟨ ξ h n + 1 ⋅ n p , ψ h ⟩ Γ = ( g p n + 1 , ψ h ) p + ⟨ − u h ∗ , n + 1 ⋅ n p + L − 1 ϕ h ∗ , n + 1 , ψ h ⟩ Γ (24) C_0(D_t\phi_h^{n+1},\psi_h)_p+\alpha(\nabla\cdot\xi_h^{n+1},\psi_h)_p+d^R(\phi_h^{n+1},\psi_h)-\langle\xi_h^{n+1}\cdot n_p,\psi_h\rangle_\Gamma
= (g_p^{n+1},\psi_h)_p+\langle-u_h^{*,n+1}\cdot n_p+L^{-1}\phi_h^{*,n+1},\psi_h\rangle_\Gamma \tag*{(24)} C 0 ( D t ϕ h n + 1 , ψ h ) p + α ( ∇ ⋅ ξ h n + 1 , ψ h ) p + d R ( ϕ h n + 1 , ψ h ) − ⟨ ξ h n + 1 ⋅ n p , ψ h ⟩ Γ = ( g p n + 1 , ψ h ) p + ⟨ − u h ∗ , n + 1 ⋅ n p + L − 1 ϕ h ∗ , n + 1 , ψ h ⟩ Γ (24)
for ( χ h , ζ h , ψ h ) ∈ V p , h × V p , h × Q p , h (\chi_h,\zeta_h,\psi_h)\in V_{p,h}\times V_{p,h}\times Q_{p,h} ( χ h , ζ h , ψ h ) ∈ V p , h × V p , h × Q p , h .
The method exchanges data explicitly between the fluid and Biot subproblems during each time step, while solving each subproblem implicitly. For error analysis the starting values at n = 0 n=0 n = 0 and n = 1 n=1 n = 1 need to be good approximations of the continuous solutions. The error conditions will be explicitly given in § 5 .
Proposition 3.1 (Subproblem solvability). Assuming the discrete inf–sup condition (17) is true, each step of (21) –(3.8) has a unique solution for every Δ t > 0 \Delta t>0 Δ t > 0 , given previous numerical solutions and the right-hand sides.
Proof. For fixed previous numerical solutions and right-hand sides, both subproblems are square finite-dimensional linear systems, so it suffices to prove uniqueness.
For the difference of two fluid solutions, say δ u h \delta u_h δ u h , testing the momentum equation (21) with δ u h \delta u_h δ u h and using the incompressibility constraint yields
3 ρ f 2 Δ t ∥ δ u h ∥ f 2 + a f R ( δ u h , δ u h ) = 0. \frac{3\rho_f}{2\Delta t}\lVert\delta u_h\rVert_f^2+a_f^R(\delta u_h,\delta u_h)=0. 2Δ t 3 ρ f ∥ δ u h ∥ f 2 + a f R ( δ u h , δ u h ) = 0.
The positive mass, viscous, and Robin contributions on the boundary imply that δ u h = 0 \delta u_h=0 δ u h = 0 . The discrete inf–sup condition (17) then implies that the pressure difference also vanishes.
For the Biot subproblem, denote the differences of two solutions at time t n + 1 t_{n+1} t n + 1 by δ η h \delta\eta_h δ η h , δ ξ h \delta\xi_h δ ξ h , and δ ϕ h \delta\phi_h δ ϕ h . Since the two solutions have identical histories, subtraction of their BDF2 kinematic equation (22) gives 3 2 Δ t δ η h = δ ξ h \frac{3}{2\Delta t}\delta\eta_h=\delta\xi_h 2Δ t 3 δ η h = δ ξ h . Test the difference equation of (23) with δ ξ h \delta\xi_h δ ξ h and the difference equation of (24) with δ ϕ h \delta\phi_h δ ϕ h , and add them. Using the relation 3 2 Δ t δ η h = δ ξ h \frac{3}{2\Delta t}\delta\eta_h=\delta\xi_h 2Δ t 3 δ η h = δ ξ h in the elastic term then yields
0 = 3 ρ p 2 Δ t ∥ δ ξ h ∥ p 2 + 2 Δ t 3 ∥ δ ξ h ∥ a e 2 + 3 C 0 2 Δ t ∥ δ ϕ h ∥ p 2 + d R ( δ ϕ h , δ ϕ h ) + γ ∥ P δ ξ h ∥ Γ 2 . 0=\frac{3\rho_p}{2\Delta t}\lVert\delta\xi_h\rVert_p^2+\frac{2\Delta t}{3}\lVert\delta\xi_h\rVert_{a_e}^2+\frac{3C_0}{2\Delta t}\lVert\delta\phi_h\rVert_p^2+d^R(\delta\phi_h,\delta\phi_h)+\gamma\lVert P\delta\xi_h\rVert_\Gamma^2. 0 = 2Δ t 3 ρ p ∥ δ ξ h ∥ p 2 + 3 2Δ t ∥ δ ξ h ∥ a e 2 + 2Δ t 3 C 0 ∥ δ ϕ h ∥ p 2 + d R ( δ ϕ h , δ ϕ h ) + γ ∥ P δ ξ h ∥ Γ 2 .
Since ρ p > 0 \rho_p>0 ρ p > 0 and C 0 > 0 C_0>0 C 0 > 0 , this identity implies δ ξ h = 0 \delta\xi_h=0 δ ξ h = 0 , δ ϕ h = 0 \delta\phi_h=0 δ ϕ h = 0 , and therefore δ η h = 0 \delta\eta_h=0 δ η h = 0 . □
Discrete energy stability under a linear CFL condition
Since the fluid and porous structure exchange energy through the interface and our proposed scheme uses explicit predicted interface values (EX2), that exchange is not perfectly balanced at the discrete level. In this section, we show that the resulting imbalance can be controlled when the time step is sufficiently small relative to the mesh size.
Trace inequality and BDF2 energy
It is known that there are constants C f , C p C_f,C_p C f , C p independent of h h h such that
∥ v h ∥ Γ 2 ≤ C f ∥ D ( v h ) ∥ f 2 , ∥ ψ h ∥ Γ 2 ≤ C p ∥ K 1 / 2 ∇ ψ h ∥ p 2 (25) \lVert v_h\rVert_{\Gamma}^{2} \le C_f\lVert D(v_h)\rVert_f^{2},\qquad\lVert\psi_h\rVert_{\Gamma}^{2} \le C_p\lVert K^{1/2}\nabla\psi_h\rVert_p^{2} \tag*{(25)} ∥ v h ∥ Γ 2 ≤ C f ∥ D ( v h ) ∥ f 2 , ∥ ψ h ∥ Γ 2 ≤ C p ∥ K 1/2 ∇ ψ h ∥ p 2 (25)
which are obtained from continuous trace inequalities, Korn’s inequality, and the Poincaré inequality (see, e.g., [24 , 25 ] ). For discrete scalar or vector fields on either subdomain, we have an inverse trace inequality
∥ w h ∥ Γ 2 ≤ C I h − 1 ∥ w h ∥ i 2 , i ∈ { f , p } , (26) \lVert w_h\rVert_{\Gamma}^{2} \le C_I h^{-1}\lVert w_h\rVert_i^{2},\qquad i\in\{f,p\}, \tag*{(26)} ∥ w h ∥ Γ 2 ≤ C I h − 1 ∥ w h ∥ i 2 , i ∈ { f , p } , (26)
with C I > 0 C_I>0 C I > 0 independent of h h h .
For any Hilbert inner product, the BDF2 approximation has the following identity
( D t a n + 1 , a n + 1 ) = 1 4 Δ t ( ∥ a n + 1 ∥ 2 + ∥ 2 a n + 1 − a n ∥ 2 − ∥ a n ∥ 2 − ∥ 2 a n − a n − 1 ∥ 2 ) + 1 4 Δ t ∥ a n + 1 − 2 a n + a n − 1 ∥ 2 . (27) (D_t a^{n+1},a^{n+1})=\frac{1}{4\Delta t}\left(\lVert a^{n+1}\rVert^2+\lVert2a^{n+1}-a^n\rVert^2-\lVert a^n\rVert^2-\lVert2a^n-a^{n-1}\rVert^2\right)+\frac{1}{4\Delta t}\lVert a^{n+1}-2a^n+a^{n-1}\rVert^2. \tag*{(27)} ( D t a n + 1 , a n + 1 ) = 4Δ t 1 ( ∥ a n + 1 ∥ 2 + ∥ 2 a n + 1 − a n ∥ 2 − ∥ a n ∥ 2 − ∥ 2 a n − a n − 1 ∥ 2 ) + 4Δ t 1 ∥ a n + 1 − 2 a n + a n − 1 ∥ 2 . (27)
The identity applies also to a e a_e a e . Define
E n : = 1 4 [ ρ f ( ∥ u h n ∥ f 2 + ∥ 2 u h n − u h n − 1 ∥ f 2 ) + ρ p ( ∥ ξ h n ∥ p 2 + ∥ 2 ξ h n − ξ h n − 1 ∥ p 2 ) + C 0 ( ∥ ϕ h n ∥ p 2 + ∥ 2 ϕ h n − ϕ h n − 1 ∥ p 2 ) + ∥ η h n ∥ a e 2 + ∥ 2 η h n − η h n − 1 ∥ a e 2 ] . (28) \mathcal{E}^n:=\frac{1}{4}\left[\rho_f\left(\lVert u_h^n\rVert_f^2+\lVert2u_h^n-u_h^{n-1}\rVert_f^2\right)+\rho_p\left(\lVert\xi_h^n\rVert_p^2+\lVert2\xi_h^n-\xi_h^{n-1}\rVert_p^2\right)+C_0\left(\lVert\phi_h^n\rVert_p^2+\lVert2\phi_h^n-\phi_h^{n-1}\rVert_p^2\right)+\lVert\eta_h^n\rVert_{a_e}^2+\lVert2\eta_h^n-\eta_h^{n-1}\rVert_{a_e}^2\right]. \tag*{(28)} E n := 4 1 [ ρ f ( ∥ u h n ∥ f 2 + ∥ 2 u h n − u h n − 1 ∥ f 2 ) + ρ p ( ∥ ξ h n ∥ p 2 + ∥ 2 ξ h n − ξ h n − 1 ∥ p 2 ) + C 0 ( ∥ ϕ h n ∥ p 2 + ∥ 2 ϕ h n − ϕ h n − 1 ∥ p 2 ) + ∥ η h n ∥ a e 2 + ∥ 2 η h n − η h n − 1 ∥ a e 2 ] . (28)
Let r a n + 1 : = a h n + 1 − 2 a h n + a h n − 1 r_a^{n+1}:=a_h^{n+1}-2a_h^n+a_h^{n-1} r a n + 1 := a h n + 1 − 2 a h n + a h n − 1 and set
N n + 1 : = 1 4 Δ t [ ρ f ∥ r u n + 1 ∥ f 2 + ρ p ∥ r ξ n + 1 ∥ p 2 + C 0 ∥ r ϕ n + 1 ∥ p 2 + ∥ r η n + 1 ∥ a e 2 ] , (29) \mathcal{N}^{n+1}:=\frac{1}{4\Delta t}\left[\rho_f\lVert r_u^{n+1}\rVert_f^2+\rho_p\lVert r_\xi^{n+1}\rVert_p^2+C_0\lVert r_\phi^{n+1}\rVert_p^2+\lVert r_\eta^{n+1}\rVert_{a_e}^2\right], \tag*{(29)} N n + 1 := 4Δ t 1 [ ρ f ∥ r u n + 1 ∥ f 2 + ρ p ∥ r ξ n + 1 ∥ p 2 + C 0 ∥ r ϕ n + 1 ∥ p 2 + ∥ r η n + 1 ∥ a e 2 ] , (29)
which corresponds to numerical dissipation.
Discrete stability
Define the constants used in the stability estimate. Choose
q Γ : = min { γ 4 , μ f 4 C f } > 0 , (30) q_{\Gamma}:=\min\left\{\frac{\gamma}{4},\frac{\mu_f}{4C_f}\right\}>0, \tag*{(30)} q Γ := min { 4 γ , 4 C f μ f } > 0 , (30)
so that
2 q Γ ≤ γ 2 , 2 C f q Γ ≤ μ f 2 . 2q_{\Gamma}\le\frac{\gamma}{2},\qquad2C_fq_{\Gamma}\le\frac{\mu_f}{2}. 2 q Γ ≤ 2 γ , 2 C f q Γ ≤ 2 μ f .
Define
A : = 8 C I C f ( γ 2 ρ p + L 2 ρ f + 1 C 0 ) , B : = 8 C I C p ( 1 ρ f + 1 L 2 C 0 ) , (31) A:=8C_IC_f\left(\frac{\gamma^2}{\rho_p}+\frac{L^2}{\rho_f}+\frac{1}{C_0}\right),\qquad B:=8C_IC_p\left(\frac{1}{\rho_f}+\frac{1}{L^2C_0}\right), \tag*{(31)} A := 8 C I C f ( ρ p γ 2 + ρ f L 2 + C 0 1 ) , B := 8 C I C p ( ρ f 1 + L 2 C 0 1 ) , (31)
and choose
C ∗ : = min { ρ f q Γ 16 γ 2 C I ′ , μ f 2 A , 1 2 B } . (32) C_*:=\min\left\{\frac{\rho_fq_{\Gamma}}{16\gamma^2C_I'},\frac{\mu_f}{2A},\frac{1}{2B}\right\}. \tag*{(32)} C ∗ := min { 16 γ 2 C I ′ ρ f q Γ , 2 A μ f , 2 B 1 } . (32)
These constants are independent of h h h and Δ t \Delta t Δ t , but may depend on the material parameters, Robin boundary coefficient L L L , the geometry, and mesh regularity.
Theorem 4.1 (Discrete energy stability). Let the sampled sources satisfy f f j ∈ V f ′ f_f^j \in V_f' f f j ∈ V f ′ , f p j ∈ L 2 ( Ω p ) d f_p^j \in L^2(\Omega_p)^d f p j ∈ L 2 ( Ω p ) d , and g p j ∈ Q p ′ g_p^j \in Q_p' g p j ∈ Q p ′ for 2 ≤ j ≤ N 2 \le j \le N 2 ≤ j ≤ N , where V f ′ V_f' V f ′ and Q p ′ Q_p' Q p ′ are the dual spaces of V f V_f V f and Q p Q_p Q p equipped with the H 1 H^1 H 1 norms, and the source terms in (21) and (3.8) are understood as duality pairings. If Δ t ≤ C ∗ h \Delta t \le C_*h Δ t ≤ C ∗ h for C ∗ C_* C ∗ in (32) , then, for 2 ≤ M ≤ N 2 \le M \le N 2 ≤ M ≤ N , the solution of (21) –(3.8) satisfies
E M + Δ t 2 ∑ j = 2 M ( N j + D ∗ j ) ≤ 2 E 1 + C Δ t ∑ j = 2 M ( ∥ f f j ∥ V f ′ 2 + ∥ g p j ∥ Q p ′ 2 ) + 4 ρ p ( Δ t ∑ j = 2 M ∥ f p j ∥ p ) 2 . (33) \mathcal{E}^{M}+\frac{\Delta t}{2}\sum_{j=2}^{M}\left(\mathcal{N}^{j}+\mathcal{D}_{*}^{j}\right)\le2\mathcal{E}^{1}+C\Delta t\sum_{j=2}^{M}\left(\left\|f_f^j\right\|_{V_f'}^2+\left\|g_p^j\right\|_{Q_p'}^2\right)+\frac{4}{\rho_p}\left(\Delta t\sum_{j=2}^{M}\left\|f_p^j\right\|_p\right)^2 . \tag*{(33)} E M + 2 Δ t j = 2 ∑ M ( N j + D ∗ j ) ≤ 2 E 1 + C Δ t j = 2 ∑ M ( f f j V f ′ 2 + g p j Q p ′ 2 ) + ρ p 4 ( Δ t j = 2 ∑ M f p j p ) 2 . (33) with E n \mathcal{E}^n E n , N n \mathcal{N}^n N n defined in (28) , (29) , and
D ∗ j : = μ f ∥ D ( u h j ) ∥ f 2 + 1 2 ∥ K 1 / 2 ∇ ϕ h j ∥ p 2 + γ Γ 2 ∥ P ξ h j ∥ Γ 2 . (34) \mathcal{D}_{*}^{j}:=\mu_f\left\|\mathsf{D}(u_h^j)\right\|_f^2+\frac{1}{2}\left\|K^{1/2}\nabla\phi_h^j\right\|_p^2+\frac{\gamma\Gamma}{2}\left\|\mathsf{P}\xi_h^j\right\|_\Gamma^2 . \tag*{(34)} D ∗ j := μ f D ( u h j ) f 2 + 2 1 K 1/2 ∇ ϕ h j p 2 + 2 γ Γ P ξ h j Γ 2 . (34) Here C C C depends only on μ f \mu_f μ f , the lower bound of the permeability K K K , and the Korn and Poincaré constants of V f V_f V f and Q p Q_p Q p but is independent of h h h , Δ t \Delta t Δ t , and M M M . If the sources vanish, then
E M + Δ t ∑ j = 2 M ( 1 2 N j + D ∗ j ) ≤ E 1 . (35) \mathcal{E}^{M}+\Delta t\sum_{j=2}^{M}\left(\frac{1}{2}\mathcal{N}^{j}+\mathcal{D}_{*}^{j}\right)\le\mathcal{E}^{1}. \tag*{(35)} E M + Δ t j = 2 ∑ M ( 2 1 N j + D ∗ j ) ≤ E 1 . (35)
Proof. Steps 1–3 bound the interface residuals and prove (35) ; Step 4 treats nonzero sources and proves (33) .
Step 1: Discrete energy identity. Since D t η h n + 1 − ξ h n + 1 ∈ V p , h D_t\eta_h^{n+1}-\xi_h^{n+1}\in V_{p,h} D t η h n + 1 − ξ h n + 1 ∈ V p , h , taking χ h = D t η h n + 1 − ξ h n + 1 \chi_h=D_t\eta_h^{n+1}-\xi_h^{n+1} χ h = D t η h n + 1 − ξ h n + 1 in (22) gives ξ h n + 1 = D t η h n + 1 \xi_h^{n+1}=D_t\eta_h^{n+1} ξ h n + 1 = D t η h n + 1 . We test the fluid equation (21) with ( v h , q h ) = ( u h n + 1 , p f , h n + 1 ) (v_h,q_h)=(u_h^{n+1},p_{f,h}^{n+1}) ( v h , q h ) = ( u h n + 1 , p f , h n + 1 ) , the equations (23) and (24) with ζ h = ξ h n + 1 \zeta_h=\xi_h^{n+1} ζ h = ξ h n + 1 and ψ h = ϕ h n + 1 \psi_h=\phi_h^{n+1} ψ h = ϕ h n + 1 , and add them. By the previous identity, the elastic term is a e ( η h n + 1 , ξ h n + 1 ) = a e ( η h n + 1 , D t η h n + 1 ) a_e(\eta_h^{n+1},\xi_h^{n+1})=a_e(\eta_h^{n+1},D_t\eta_h^{n+1}) a e ( η h n + 1 , ξ h n + 1 ) = a e ( η h n + 1 , D t η h n + 1 ) . The volume coupling terms in (21) , (3.8), and the mechanical normal-interface terms in (3.8) cancel. The extrapolation in (18) satisfies
a h n + 1 − a h ∗ , n + 1 = a h n + 1 − 2 a h n + a h n − 1 = r a n + 1 for a = u , ξ , ϕ . a_h^{n+1}-a_h^{*,n+1}=a_h^{n+1}-2a_h^n+a_h^{n-1}=r_a^{n+1}\qquad\text{for }a=u,\xi,\phi. a h n + 1 − a h ∗ , n + 1 = a h n + 1 − 2 a h n + a h n − 1 = r a n + 1 for a = u , ξ , ϕ .
Using the identity (27) , and replacing each predictor by a h ∗ , n + 1 = a h n + 1 − r a n + 1 a_h^{*,n+1}=a_h^{n+1}-r_a^{n+1} a h ∗ , n + 1 = a h n + 1 − r a n + 1 , gives
E n + 1 − E n Δ t + N n + 1 + 2 μ f ∥ D ( u h n + 1 ) ∥ f 2 + ∥ K 1 / 2 ∇ ϕ h n + 1 ∥ p 2 + γ ∥ P ( u h n + 1 − ξ h n + 1 ) ∥ Γ 2 = R Γ n + 1 + W n + 1 . (36) \frac{\mathcal{E}^{n+1}-\mathcal{E}^{n}}{\Delta t}+\mathcal{N}^{n+1}+2\mu_f\left\|\mathsf{D}(u_h^{n+1})\right\|_f^2+\left\|K^{1/2}\nabla\phi_h^{n+1}\right\|_p^2+\gamma\left\|\mathsf{P}(u_h^{n+1}-\xi_h^{n+1})\right\|_\Gamma^2=\mathcal{R}_\Gamma^{n+1}+\mathcal{W}^{n+1}. \tag*{(36)} Δ t E n + 1 − E n + N n + 1 + 2 μ f D ( u h n + 1 ) f 2 + K 1/2 ∇ ϕ h n + 1 p 2 + γ P ( u h n + 1 − ξ h n + 1 ) Γ 2 = R Γ n + 1 + W n + 1 . (36)
where
R Γ n + 1 : = − γ ⟨ P r ξ n + 1 , P u h n + 1 ⟩ Γ − γ ⟨ P r u n + 1 , P ξ h n + 1 ⟩ Γ − L ⟨ r u n + 1 ⋅ n f , u h n + 1 ⋅ n f ⟩ Γ + ⟨ r ϕ n + 1 , u h n + 1 ⋅ n f ⟩ Γ + ⟨ r u n + 1 ⋅ n p , ϕ h n + 1 ⟩ Γ − L − 1 ⟨ r ϕ n + 1 , ϕ h n + 1 ⟩ Γ , (37) \begin{aligned}
\mathcal{R}_\Gamma^{n+1}:={}&-\gamma\left\langle\mathsf{P}r_\xi^{n+1},\mathsf{P}u_h^{n+1}\right\rangle_\Gamma-\gamma\left\langle\mathsf{P}r_u^{n+1},\mathsf{P}\xi_h^{n+1}\right\rangle_\Gamma-L\left\langle r_u^{n+1}\cdot n_f,u_h^{n+1}\cdot n_f\right\rangle_\Gamma\\
&+\left\langle r_\phi^{n+1},u_h^{n+1}\cdot n_f\right\rangle_\Gamma+\left\langle r_u^{n+1}\cdot n_p,\phi_h^{n+1}\right\rangle_\Gamma-L^{-1}\left\langle r_\phi^{n+1},\phi_h^{n+1}\right\rangle_\Gamma, \tag*{(37)}
\end{aligned} R Γ n + 1 := − γ ⟨ P r ξ n + 1 , P u h n + 1 ⟩ Γ − γ ⟨ P r u n + 1 , P ξ h n + 1 ⟩ Γ − L ⟨ r u n + 1 ⋅ n f , u h n + 1 ⋅ n f ⟩ Γ + ⟨ r ϕ n + 1 , u h n + 1 ⋅ n f ⟩ Γ + ⟨ r u n + 1 ⋅ n p , ϕ h n + 1 ⟩ Γ − L − 1 ⟨ r ϕ n + 1 , ϕ h n + 1 ⟩ Γ , (37)
W n + 1 : = ( f f n + 1 , u h n + 1 ) f + ( f p n + 1 , ξ h n + 1 ) p + ( g p n + 1 , ϕ h n + 1 ) p . (38) \mathcal{W}^{n+1}:=(f_f^{n+1},u_h^{n+1})_f+(f_p^{n+1},\xi_h^{n+1})_p+(g_p^{n+1},\phi_h^{n+1})_p. \tag*{(38)} W n + 1 := ( f f n + 1 , u h n + 1 ) f + ( f p n + 1 , ξ h n + 1 ) p + ( g p n + 1 , ϕ h n + 1 ) p . (38)
Step 2: Bounds for the residual terms. By the Cauchy–Schwarz inequality, the inverse trace inequality (26) for r ξ n + 1 ∈ V p , h r_\xi^{n+1}\in V_{p,h} r ξ n + 1 ∈ V p , h , the trace estimate (25) for u h n + 1 u_h^{n+1} u h n + 1 , and Young’s inequality a b ≤ ε a 2 + b 2 / ( 4 ε ) ab\le\varepsilon a^2+b^2/(4\varepsilon) ab ≤ ε a 2 + b 2 / ( 4 ε ) with ε = ρ p / ( 32 Δ t ) \varepsilon=\rho_p/(32\Delta t) ε = ρ p / ( 32Δ t ) , we obtain
γ ∣ ⟨ P r ξ n + 1 , P u h n + 1 ⟩ Γ ∣ ≤ ρ p 32 Δ t ∥ r ξ n + 1 ∥ p 2 + 8 γ 2 C I C f Δ t ρ p h ∥ D ( u h n + 1 ) ∥ f 2 . (39) \gamma\left|\left\langle\mathsf{P}r_\xi^{n+1},\mathsf{P}u_h^{n+1}\right\rangle_\Gamma\right|\le\frac{\rho_p}{32\Delta t}\left\|r_\xi^{n+1}\right\|_p^2+8\gamma^2 C_I C_f\frac{\Delta t}{\rho_p h}\left\|\mathsf{D}(u_h^{n+1})\right\|_f^2. \tag*{(39)} γ ⟨ P r ξ n + 1 , P u h n + 1 ⟩ Γ ≤ 32Δ t ρ p r ξ n + 1 p 2 + 8 γ 2 C I C f ρ p h Δ t D ( u h n + 1 ) f 2 . (39)
Similarly, applying (25) –(26) and Young’s inequality yields
L ∣ ⟨ r u n + 1 ⋅ n f , u h n + 1 ⋅ n f ⟩ Γ ∣ ≤ ρ f 32 Δ t ∥ r u n + 1 ∥ f 2 + 8 L 2 C I C f Δ t ρ f h ∥ D ( u h n + 1 ) ∥ f 2 , (40) L\left|\left\langle r_u^{n+1}\cdot n_f,u_h^{n+1}\cdot n_f\right\rangle_\Gamma\right|
\leq\frac{\rho_f}{32\Delta t}\left\|r_u^{n+1}\right\|_f^2+8L^2C_IC_f\frac{\Delta t}{\rho_f h}\left\|D(u_h^{n+1})\right\|_f^2, \tag*{(40)} L ⟨ r u n + 1 ⋅ n f , u h n + 1 ⋅ n f ⟩ Γ ≤ 32Δ t ρ f r u n + 1 f 2 + 8 L 2 C I C f ρ f h Δ t D ( u h n + 1 ) f 2 , (40)
∣ ⟨ r ϕ n + 1 , u h n + 1 ⋅ n f ⟩ Γ ∣ ≤ C 0 32 Δ t ∥ r ϕ n + 1 ∥ p 2 + 8 C I C f Δ t C 0 h ∥ D ( u h n + 1 ) ∥ f 2 , (41) \left|\left\langle r_\phi^{n+1},u_h^{n+1}\cdot n_f\right\rangle_\Gamma\right|
\leq\frac{C_0}{32\Delta t}\left\|r_\phi^{n+1}\right\|_p^2+8C_IC_f\frac{\Delta t}{C_0h}\left\|D(u_h^{n+1})\right\|_f^2, \tag*{(41)} ⟨ r ϕ n + 1 , u h n + 1 ⋅ n f ⟩ Γ ≤ 32Δ t C 0 r ϕ n + 1 p 2 + 8 C I C f C 0 h Δ t D ( u h n + 1 ) f 2 , (41)
∣ ⟨ r u n + 1 ⋅ n p , ϕ h n + 1 ⟩ Γ ∣ ≤ ρ f 32 Δ t ∥ r u n + 1 ∥ f 2 + 8 C I C p Δ t ρ f h ∥ K 1 / 2 ∇ ϕ h n + 1 ∥ p 2 , (42) \left|\left\langle r_u^{n+1}\cdot n_p,\phi_h^{n+1}\right\rangle_\Gamma\right|
\leq\frac{\rho_f}{32\Delta t}\left\|r_u^{n+1}\right\|_f^2+8C_IC_p\frac{\Delta t}{\rho_f h}\left\|K^{1/2}\nabla\phi_h^{n+1}\right\|_p^2, \tag*{(42)} ⟨ r u n + 1 ⋅ n p , ϕ h n + 1 ⟩ Γ ≤ 32Δ t ρ f r u n + 1 f 2 + 8 C I C p ρ f h Δ t K 1/2 ∇ ϕ h n + 1 p 2 , (42)
L − 1 ∣ ⟨ r ϕ n + 1 , ϕ h n + 1 ⟩ Γ ∣ ≤ C 0 32 Δ t ∥ r ϕ n + 1 ∥ p 2 + 8 C I C p Δ t L 2 C 0 h ∥ K 1 / 2 ∇ ϕ h n + 1 ∥ p 2 . (43) L^{-1}\left|\left\langle r_\phi^{n+1},\phi_h^{n+1}\right\rangle_\Gamma\right|
\leq\frac{C_0}{32\Delta t}\left\|r_\phi^{n+1}\right\|_p^2+8C_IC_p\frac{\Delta t}{L^2C_0h}\left\|K^{1/2}\nabla\phi_h^{n+1}\right\|_p^2. \tag*{(43)} L − 1 ⟨ r ϕ n + 1 , ϕ h n + 1 ⟩ Γ ≤ 32Δ t C 0 r ϕ n + 1 p 2 + 8 C I C p L 2 C 0 h Δ t K 1/2 ∇ ϕ h n + 1 p 2 . (43)
For the tangential structure-velocity product, the inverse trace inequality (26) for r u n + 1 r_u^{n+1} r u n + 1 and Young’s inequality give
γ ∣ ⟨ P r u n + 1 , P ξ h n + 1 ⟩ Γ ∣ ≤ γ Γ 2 ∥ P ξ h n + 1 ∥ Γ 2 + γ 2 C I 2 q Γ h ∥ r u n + 1 ∥ f 2 . (44) \gamma\left|\left\langle P r_u^{n+1},P\xi_h^{n+1}\right\rangle_\Gamma\right|
\leq\frac{\gamma\Gamma}{2}\left\|P\xi_h^{n+1}\right\|_\Gamma^2+\frac{\gamma^2C_I}{2q\Gamma h}\left\|r_u^{n+1}\right\|_f^2. \tag*{(44)} γ ⟨ P r u n + 1 , P ξ h n + 1 ⟩ Γ ≤ 2 γ Γ P ξ h n + 1 Γ 2 + 2 q Γ h γ 2 C I r u n + 1 f 2 . (44)
If Δ t / h ≤ ρ f q Γ / ( 16 γ 2 C I ) \Delta t/h\leq\rho_f q\Gamma/(16\gamma^2C_I) Δ t / h ≤ ρ f q Γ/ ( 16 γ 2 C I ) , the last term is at most ρ f ∥ r u n + 1 ∥ f 2 / ( 32 Δ t ) \rho_f\left\|r_u^{n+1}\right\|_f^2/(32\Delta t) ρ f r u n + 1 f 2 / ( 32Δ t ) . Writing P ξ h n + 1 = P ( ξ h n + 1 − u h n + 1 ) + P u h n + 1 P\xi_h^{n+1}=P(\xi_h^{n+1}-u_h^{n+1})+Pu_h^{n+1} P ξ h n + 1 = P ( ξ h n + 1 − u h n + 1 ) + P u h n + 1 , using ∥ a + b ∥ 2 ≤ 2 ∥ a ∥ 2 + 2 ∥ b ∥ 2 \|a+b\|^2\leq2\|a\|^2+2\|b\|^2 ∥ a + b ∥ 2 ≤ 2∥ a ∥ 2 + 2∥ b ∥ 2 , and applying (25) to u h n + 1 u_h^{n+1} u h n + 1 , we obtain from (30)
q Γ ∥ P ξ h n + 1 ∥ Γ 2 ≤ γ 2 ∥ P ( u h n + 1 − ξ h n + 1 ) ∥ Γ 2 + μ f 2 ∥ D ( u h n + 1 ) ∥ f 2 . (45) q\Gamma\left\|P\xi_h^{n+1}\right\|_\Gamma^2\leq\frac{\gamma}{2}\left\|P(u_h^{n+1}-\xi_h^{n+1})\right\|_\Gamma^2+\frac{\mu_f}{2}\left\|D(u_h^{n+1})\right\|_f^2. \tag*{(45)} q Γ P ξ h n + 1 Γ 2 ≤ 2 γ P ( u h n + 1 − ξ h n + 1 ) Γ 2 + 2 μ f D ( u h n + 1 ) f 2 . (45)
Notice that this estimate controls the tangential structure velocity through the slip term and fluid viscous dissipation, without requiring a bulk H 1 H^1 H 1 estimate for the structure velocity. Step 3: Absorption and summation. By the first restriction in (32) , the condition at the end of Step 2 holds. Thus the total coefficients assigned to ∥ r u n + 1 ∥ f 2 \left\|r_u^{n+1}\right\|_f^2 r u n + 1 f 2 , ∥ r ξ n + 1 ∥ p 2 \left\|r_\xi^{n+1}\right\|_p^2 r ξ n + 1 p 2 , and ∥ r ϕ n + 1 ∥ p 2 \left\|r_\phi^{n+1}\right\|_p^2 r ϕ n + 1 p 2 in (39) –(44) are at most
3 ρ f 32 Δ t , ρ p 32 Δ t , 2 C 0 32 Δ t , \frac{3\rho_f}{32\Delta t},\qquad\frac{\rho_p}{32\Delta t},\qquad\frac{2C_0}{32\Delta t}, 32Δ t 3 ρ f , 32Δ t ρ p , 32Δ t 2 C 0 ,
respectively. These contributions are bounded by N n + 1 / 2 \mathcal{N}^{n+1/2} N n + 1/2 .
Recalling A A A , B B B in (31) the remaining restrictions in (32) imply
A Δ t h ≤ μ f 2 , B Δ t h ≤ 1 2 . A\frac{\Delta t}{h}\leq\frac{\mu_f}{2},\qquad B\frac{\Delta t}{h}\leq\frac{1}{2}. A h Δ t ≤ 2 μ f , B h Δ t ≤ 2 1 .
Combining the residual estimates therefore yields
∣ R Γ n + 1 ∣ ≤ 1 2 N n + 1 + μ f 2 ∥ D ( u h n + 1 ) ∥ f 2 + 1 2 ∥ K 1 / 2 ∇ ϕ h n + 1 ∥ p 2 + q Γ 2 ∥ P ξ h n + 1 ∥ Γ 2 . (46) \left|\mathcal{R}_\Gamma^{n+1}\right|\leq\frac{1}{2}\mathcal{N}^{n+1}+\frac{\mu_f}{2}\left\|D(u_h^{n+1})\right\|_f^2+\frac{1}{2}\left\|K^{1/2}\nabla\phi_h^{n+1}\right\|_p^2+\frac{q\Gamma}{2}\left\|P\xi_h^{n+1}\right\|_\Gamma^2. \tag*{(46)} R Γ n + 1 ≤ 2 1 N n + 1 + 2 μ f D ( u h n + 1 ) f 2 + 2 1 K 1/2 ∇ ϕ h n + 1 p 2 + 2 q Γ P ξ h n + 1 Γ 2 . (46)
Substituting this bound into (36) gives
E n + 1 − E n Δ t + 1 2 N n + 1 + 3 μ f 2 ∥ D ( u h n + 1 ) ∥ f 2 + 1 2 ∥ K 1 / 2 ∇ ϕ h n + 1 ∥ p 2 + γ ∥ P ( u h n + 1 − ξ h n + 1 ) ∥ Γ 2 − q Γ 2 ∥ P ξ h n + 1 ∥ Γ 2 ≤ W n + 1 . \frac{\mathcal{E}^{n+1}-\mathcal{E}^n}{\Delta t}+\frac{1}{2}\mathcal{N}^{n+1}+\frac{3\mu_f}{2}\left\|D(u_h^{n+1})\right\|_f^2+\frac{1}{2}\left\|K^{1/2}\nabla\phi_h^{n+1}\right\|_p^2+\gamma\left\|P(u_h^{n+1}-\xi_h^{n+1})\right\|_\Gamma^2-\frac{q\Gamma}{2}\left\|P\xi_h^{n+1}\right\|_\Gamma^2\leq\mathcal{W}^{n+1}. Δ t E n + 1 − E n + 2 1 N n + 1 + 2 3 μ f D ( u h n + 1 ) f 2 + 2 1 K 1/2 ∇ ϕ h n + 1 p 2 + γ P ( u h n + 1 − ξ h n + 1 ) Γ 2 − 2 q Γ P ξ h n + 1 Γ 2 ≤ W n + 1 .
By (45) ,
3 μ f 2 ∥ D ( u h n + 1 ) ∥ f 2 + γ ∥ P ( u h n + 1 − ξ h n + 1 ) ∥ Γ 2 − q Γ 2 ∥ P ξ h n + 1 ∥ Γ 2 ≥ μ f ∥ D ( u h n + 1 ) ∥ f 2 + q Γ 2 ∥ P ξ h n + 1 ∥ Γ 2 + γ 2 ∥ P ( u h n + 1 − ξ h n + 1 ) ∥ Γ 2 . \frac{3\mu_f}{2}\left\|D(u_h^{n+1})\right\|_f^2+\gamma\left\|P(u_h^{n+1}-\xi_h^{n+1})\right\|_\Gamma^2-\frac{q\Gamma}{2}\left\|P\xi_h^{n+1}\right\|_\Gamma^2
\geq\mu_f\left\|D(u_h^{n+1})\right\|_f^2+\frac{q\Gamma}{2}\left\|P\xi_h^{n+1}\right\|_\Gamma^2+\frac{\gamma}{2}\left\|P(u_h^{n+1}-\xi_h^{n+1})\right\|_\Gamma^2. 2 3 μ f D ( u h n + 1 ) f 2 + γ P ( u h n + 1 − ξ h n + 1 ) Γ 2 − 2 q Γ P ξ h n + 1 Γ 2 ≥ μ f D ( u h n + 1 ) f 2 + 2 q Γ P ξ h n + 1 Γ 2 + 2 γ P ( u h n + 1 − ξ h n + 1 ) Γ 2 .
Inserting this lower bound and discarding the remaining nonnegative slip term gives
E n + 1 − E n Δ t + 1 2 N n + 1 + μ f ∥ D ( u h n + 1 ) ∥ f 2 + 1 2 ∥ K 1 / 2 ∇ ϕ h n + 1 ∥ p 2 + q F 2 ∥ P ξ h n + 1 ∥ Γ 2 ≤ W n + 1 . (47) \frac{\mathcal{E}^{n+1}-\mathcal{E}^{n}}{\Delta t}+\frac{1}{2}\mathcal{N}^{n+1}+\mu_f\left\|D(u_h^{n+1})\right\|_f^2+\frac{1}{2}\left\|K^{1/2}\nabla\phi_h^{n+1}\right\|_p^2+\frac{q_F}{2}\left\|P\xi_h^{n+1}\right\|_{\Gamma}^2\leq\mathcal{W}^{n+1}. \tag*{(47)} Δ t E n + 1 − E n + 2 1 N n + 1 + μ f D ( u h n + 1 ) f 2 + 2 1 K 1/2 ∇ ϕ h n + 1 p 2 + 2 q F P ξ h n + 1 Γ 2 ≤ W n + 1 . (47)
If the sources vanish, then W n + 1 = 0 \mathcal{W}^{n+1}=0 W n + 1 = 0 , and multiplying (47) by Δ t \Delta t Δ t and summing over n = 1 , … , M − 1 n=1,\ldots,M-1 n = 1 , … , M − 1 proves (35) .
Step 4: Nonzero sources. Since u h n + 1 ∈ V f u_h^{n+1}\in V_f u h n + 1 ∈ V f and ϕ h n + 1 ∈ Q p \phi_h^{n+1}\in Q_p ϕ h n + 1 ∈ Q p , Korn’s and Poincaré’s inequalities and Young’s inequality give
∣ ( f f n + 1 , u h n + 1 ) f ∣ ≤ μ f 2 ∥ D ( u h n + 1 ) ∥ f 2 + C ∥ f f n + 1 ∥ V f ′ , \left|(f_f^{n+1},u_h^{n+1})_f\right|\leq\frac{\mu_f}{2}\left\|D(u_h^{n+1})\right\|_f^2+C\left\|f_f^{n+1}\right\|_{V_f'}, ( f f n + 1 , u h n + 1 ) f ≤ 2 μ f D ( u h n + 1 ) f 2 + C f f n + 1 V f ′ ,
∣ ( g p n + 1 , ϕ h n + 1 ) p ∣ ≤ 1 4 ∥ K 1 / 2 ∇ ϕ h n + 1 ∥ p 2 + C ∥ g p n + 1 ∥ Q p ′ 2 . \left|(g_p^{n+1},\phi_h^{n+1})_p\right|\leq\frac{1}{4}\left\|K^{1/2}\nabla\phi_h^{n+1}\right\|_p^2+C\left\|g_p^{n+1}\right\|_{Q_p'}^2. ( g p n + 1 , ϕ h n + 1 ) p ≤ 4 1 K 1/2 ∇ ϕ h n + 1 p 2 + C g p n + 1 Q p ′ 2 .
These absorptions use one half of the fluid and pore-pressure parts of D ∗ n + 1 \mathcal{D}_*^{n+1} D ∗ n + 1 , so (47) yields
E n + 1 − E n Δ t + 1 2 N n + 1 + 1 2 D ∗ n + 1 ≤ ( f p n + 1 , ξ h n + 1 ) p + C ( ∥ f f n + 1 ∥ V f ′ 2 + ∥ g p n + 1 ∥ Q p ′ 2 ) . \frac{\mathcal{E}^{n+1}-\mathcal{E}^{n}}{\Delta t}+\frac{1}{2}\mathcal{N}^{n+1}+\frac{1}{2}\mathcal{D}_*^{n+1}\leq(f_p^{n+1},\xi_h^{n+1})_p+C\left(\left\|f_f^{n+1}\right\|_{V_f'}^2+\left\|g_p^{n+1}\right\|_{Q_p'}^2\right). Δ t E n + 1 − E n + 2 1 N n + 1 + 2 1 D ∗ n + 1 ≤ ( f p n + 1 , ξ h n + 1 ) p + C ( f f n + 1 V f ′ 2 + g p n + 1 Q p ′ 2 ) .
The mechanical source is tested with the bulk structure velocity, which is not directly controlled by the dissipation, so we bound its time sum through the maximum of the energy. Let 2 ≤ m ≤ M 2\leq m\leq M 2 ≤ m ≤ M and E max M : = max 1 ≤ j ≤ M E j \mathcal{E}_{\max}^M:=\max_{1\leq j\leq M}\mathcal{E}^j E m a x M := max 1 ≤ j ≤ M E j . Since ρ p ∥ ξ h j ∥ p 2 ≤ 4 E j \rho_p\left\|\xi_h^j\right\|_p^2\leq4\mathcal{E}^j ρ p ξ h j p 2 ≤ 4 E j for j ≥ 1 j\geq1 j ≥ 1 , Young’s inequality gives
Δ t ∑ j = 2 m ∣ ( f p j , ξ h j ) p ∣ ≤ 2 ρ p 1 / 2 ( E max M ) 1 / 2 Δ t ∑ j = 2 M ∥ f p j ∥ p ≤ 1 2 E max M + 2 ρ p ( Δ t ∑ j = 2 M ∥ f p j ∥ p ) 2 . \Delta t\sum_{j=2}^{m}\left|(f_p^j,\xi_h^j)_p\right|\leq\frac{2}{\rho_p^{1/2}}\left(\mathcal{E}_{\max}^M\right)^{1/2}\Delta t\sum_{j=2}^{M}\left\|f_p^j\right\|_p\leq\frac{1}{2}\mathcal{E}_{\max}^M+\frac{2}{\rho_p}\left(\Delta t\sum_{j=2}^{M}\left\|f_p^j\right\|_p\right)^2. Δ t j = 2 ∑ m ( f p j , ξ h j ) p ≤ ρ p 1/2 2 ( E m a x M ) 1/2 Δ t j = 2 ∑ M f p j p ≤ 2 1 E m a x M + ρ p 2 ( Δ t j = 2 ∑ M f p j p ) 2 .
Multiplying the previous per-step inequality by Δ t \Delta t Δ t , summing over n = 1 , … , m − 1 n=1,\ldots,m-1 n = 1 , … , m − 1 , and using this bound, we obtain
E m + Δ t 2 ∑ j = 2 m ( N j + D ∗ j ) ≤ E 1 + 1 2 E max M + S M , (48) \mathcal{E}^m+\frac{\Delta t}{2}\sum_{j=2}^{m}\left(\mathcal{N}^j+\mathcal{D}_*^j\right)\leq\mathcal{E}^1+\frac{1}{2}\mathcal{E}_{\max}^M+\mathcal{S}^M, \tag*{(48)} E m + 2 Δ t j = 2 ∑ m ( N j + D ∗ j ) ≤ E 1 + 2 1 E m a x M + S M , (48)
where S M : = C Δ t ∑ j = 2 M ( ∥ f f j ∥ V f ′ 2 + ∥ g p j ∥ Q p ′ 2 ) + 2 ρ p ( Δ t ∑ j = 2 M ∥ f p j ∥ p ) 2 \mathcal{S}^M:=C\Delta t\sum_{j=2}^{M}\left(\left\|f_f^j\right\|_{V_f'}^2+\left\|g_p^j\right\|_{Q_p'}^2\right)+\frac{2}{\rho_p}\left(\Delta t\sum_{j=2}^{M}\left\|f_p^j\right\|_p\right)^2 S M := C Δ t ∑ j = 2 M ( f f j V f ′ 2 + g p j Q p ′ 2 ) + ρ p 2 ( Δ t ∑ j = 2 M f p j p ) 2 . Inequality (48) also holds for m = 1 m=1 m = 1 , where the sum on the left is empty. Dropping the nonnegative sum and taking the maximum over 1 ≤ m ≤ M 1\leq m\leq M 1 ≤ m ≤ M gives E max M ≤ E 1 + 1 2 E max M + S M \mathcal{E}_{\max}^M\leq\mathcal{E}^1+\frac{1}{2}\mathcal{E}_{\max}^M+\mathcal{S}^M E m a x M ≤ E 1 + 2 1 E m a x M + S M , that is, E max M ≤ 2 E 1 + 2 S M \mathcal{E}_{\max}^M\leq2\mathcal{E}^1+2\mathcal{S}^M E m a x M ≤ 2 E 1 + 2 S M . Inserting this into (48) with m = M m=M m = M gives E M + Δ t 2 ∑ j = 2 M ( N j + D ∗ j ) ≤ 2 E 1 + 2 S M \mathcal{E}^M+\frac{\Delta t}{2}\sum_{j=2}^{M}\left(\mathcal{N}^j+\mathcal{D}_*^j\right)\leq2\mathcal{E}^1+2\mathcal{S}^M E M + 2 Δ t ∑ j = 2 M ( N j + D ∗ j ) ≤ 2 E 1 + 2 S M , which is (33) after renaming C C C .
Remark 4.2. The restriction Δ t ≤ C ∗ h \Delta t\leq C_*h Δ t ≤ C ∗ h is sufficient for the absorption argument above. The optimal admissible time-step constant is not determined here.
A priori error analysis
We estimate the difference between the exact solution ( u , p f , η , ξ , ϕ ) (u,p_f,\eta,\xi,\phi) ( u , p f , η , ξ , ϕ ) and the fully discrete solution ( u h n , p f , h n , η h n , ξ h n , ϕ h n ) (u_h^n,p_{f,h}^n,\eta_h^n,\xi_h^n,\phi_h^n) ( u h n , p f , h n , η h n , ξ h n , ϕ h n ) . The goal is to prove an error bound of order h k + Δ t 2 h^k+\Delta t^2 h k + Δ t 2 in the bulk energy norms under the linear CFL restriction. We use the operators in (18) and the sampling convention w n = w ( t n ) w^n=w(t_n) w n = w ( t n ) of Section 3.
The analysis is organized as follows. In Section 5.1 we introduce the projections of the exact solution and split the errors into projection and discrete parts. In Section 5.2 we derive the equations satisfied by the discrete errors, whose right-hand sides consist of consistency residuals. These residuals are estimated in Section 5.3. Section 5.4 states the error estimate and proves it by an energy argument, summing in time without a discrete Gronwall inequality.
Regularity assumptions and projection operators
Let k ≥ 1 k \ge1 k ≥ 1 be an integer. We assume the following regularity
u , ξ , η , ϕ ∈ W 3 , ∞ ( 0 , T ; H k + 1 ) , p f ∈ W 3 , ∞ ( 0 , T ; H k ) , ∂ t η = ξ , (49) \mathbf{u},\boldsymbol{\xi},\boldsymbol{\eta},\phi\in W^{3,\infty}(0,T;H^{k+1}),\qquad p_f \in W^{3,\infty}(0,T;H^k),\qquad\partial_t\boldsymbol{\eta}=\boldsymbol{\xi}, \tag*{(49)} u , ξ , η , ϕ ∈ W 3 , ∞ ( 0 , T ; H k + 1 ) , p f ∈ W 3 , ∞ ( 0 , T ; H k ) , ∂ t η = ξ , (49)
Here each field is defined on its corresponding subdomain, and vector-valued spaces are understood. We define
M : = ∥ u ∥ W 3 , ∞ ( 0 , T ; H k + 1 ( Ω f ) ) + ∥ p f ∥ W 3 , ∞ ( 0 , T ; H k ( Ω f ) ) + ∥ η ∥ W 3 , ∞ ( 0 , T ; H k + 1 ( Ω p ) ) + ∥ ξ ∥ W 3 , ∞ ( 0 , T ; H k + 1 ( Ω p ) ) + ∥ ϕ ∥ W 3 , ∞ ( 0 , T ; H k + 1 ( Ω p ) ) , (50) \mathcal{M}:=\|\mathbf{u}\|_{W^{3,\infty}(0,T;H^{k+1}(\Omega_f))}+\|p_f\|_{W^{3,\infty}(0,T;H^k(\Omega_f))}+\|\boldsymbol{\eta}\|_{W^{3,\infty}(0,T;H^{k+1}(\Omega_p))}
+\|\boldsymbol{\xi}\|_{W^{3,\infty}(0,T;H^{k+1}(\Omega_p))}+\|\phi\|_{W^{3,\infty}(0,T;H^{k+1}(\Omega_p))}, \tag*{(50)} M := ∥ u ∥ W 3 , ∞ ( 0 , T ; H k + 1 ( Ω f )) + ∥ p f ∥ W 3 , ∞ ( 0 , T ; H k ( Ω f )) + ∥ η ∥ W 3 , ∞ ( 0 , T ; H k + 1 ( Ω p )) + ∥ ξ ∥ W 3 , ∞ ( 0 , T ; H k + 1 ( Ω p )) + ∥ ϕ ∥ W 3 , ∞ ( 0 , T ; H k + 1 ( Ω p )) , (50)
where ∥ w ∥ W 3 , ∞ ( 0 , T ; X ) : = max 0 ≤ j ≤ 3 ess sup 0 < t < T ∥ ∂ t j w ( t ) ∥ X \|w\|_{W^{3,\infty}(0,T;X)}:=\max_{0\le j\le3}\operatorname{ess\,sup}_{0<t<T}\left\|\partial_t^j w(t)\right\|_X ∥ w ∥ W 3 , ∞ ( 0 , T ; X ) := max 0 ≤ j ≤ 3 ess sup 0 < t < T ∂ t j w ( t ) X for a Banach space X X X . These assumptions provide the spatial derivatives needed for an h k h^k h k energy estimate and the time derivatives needed for second-order BDF2 and EX2 consistency. The identity η t = ξ \boldsymbol{\eta}_t=\boldsymbol{\xi} η t = ξ also supplies the corresponding additional displacement time regularity. All constants below are independent of mesh size h h h and time step size Δ t \Delta t Δ t , although they may depend on the fixed coefficients and the stated regularity bounds.
The projections are chosen to cancel the principal spatial terms when we subtract the numerical equations. We use a mixed projection for the fluid and Ritz projections for elasticity and pore pressure.
At each time, define the Robin–Stokes projection ( Π u u , Π p p f ) ∈ V f , h × Q f , h (\Pi_{\mathbf{u}}\mathbf{u},\Pi_{p}p_f)\in V_{f,h}\times Q_{f,h} ( Π u u , Π p p f ) ∈ V f , h × Q f , h by
a f R ( Π u u − u , v h ) − ( Π p p f − p f , ∇ ⋅ v h ) f = 0 , ( ∇ ⋅ ( Π u u − u ) , q h ) f = 0. (51) a_f^R(\Pi_{\mathbf{u}}\mathbf{u}-\mathbf{u},\mathbf{v}_h)-(\Pi_p p_f-p_f,\nabla\cdot\mathbf{v}_h)_f=0,\qquad(\nabla\cdot(\Pi_{\mathbf{u}}\mathbf{u}-\mathbf{u}),q_h)_f=0. \tag*{(51)} a f R ( Π u u − u , v h ) − ( Π p p f − p f , ∇ ⋅ v h ) f = 0 , ( ∇ ⋅ ( Π u u − u ) , q h ) f = 0. (51)
Here a f R a_f^R a f R is the fluid bilinear form including the Robin contributions defined in (19) . The pair ( Π u u , Π p p f ) (\Pi_{\mathbf{u}}\mathbf{u},\Pi_p p_f) ( Π u u , Π p p f ) is defined jointly by (51) , so Π u u \Pi_{\mathbf{u}}\mathbf{u} Π u u and Π p p f \Pi_p p_f Π p p f in general depend on both u \mathbf{u} u and p f p_f p f ; we keep the componentwise notation for brevity. The divergence condition preserves discrete incompressibility.
For projections for the poroelastic subproblem define the elasticity Ritz map Π η : V p → V p , h \Pi_\eta:V_p\to V_{p,h} Π η : V p → V p , h and the Darcy Ritz map Π ϕ : Q p → Q p , h \Pi_\phi:Q_p\to Q_{p,h} Π ϕ : Q p → Q p , h by
a e ( Π η ζ − ζ , w h ) = 0 , d R ( Π ϕ ψ − ψ , χ h ) = 0 (52) a_e(\Pi_\eta\boldsymbol{\zeta}-\boldsymbol{\zeta},\mathbf{w}_h)=0,\qquad d^R(\Pi_\phi\psi-\psi,\chi_h)=0 \tag*{(52)} a e ( Π η ζ − ζ , w h ) = 0 , d R ( Π ϕ ψ − ψ , χ h ) = 0 (52)
for all discrete tests. The form d R d^R d R is defined in (20) . The projected displacement, structure velocity, and pore pressure are Π η η \Pi_\eta\boldsymbol{\eta} Π η η , Π ξ ξ \Pi_\xi\boldsymbol{\xi} Π ξ ξ , and Π ϕ ϕ \Pi_\phi\phi Π ϕ ϕ , where the velocity projection is the elasticity Ritz map itself:
Π ξ : = Π η . (53) \Pi_\xi:=\Pi_\eta. \tag*{(53)} Π ξ := Π η . (53)
All spaces, coefficients, and maps are time independent. In particular, the same elasticity map is used for displacement and velocity, so that
∂ t ( Π η η ) = Π η ∂ t η = Π η ξ = Π ξ ξ . \partial_t(\Pi_\eta\boldsymbol{\eta})=\Pi_\eta\partial_t\boldsymbol{\eta}=\Pi_\eta\boldsymbol{\xi}=\Pi_\xi\boldsymbol{\xi}. ∂ t ( Π η η ) = Π η ∂ t η = Π η ξ = Π ξ ξ .
This choice preserves the continuous kinematic relation after spatial projection and avoids an additional spatial defect in that relation.
We first state the approximation requirements on the finite element spaces. The spaces used in § 7 satisfy them with k = 2 k=2 k = 2 ; see the comment after Assumption 5.1.
Assumption 5.1 (Finite element approximation). In addition to the stability assumptions above, we assume the following best-approximation properties, with constants independent of h h h :
inf v h ∈ V f , h ∥ w − v h ∥ H 1 ( Ω f ) + inf q h ∈ Q f , h ∥ r − q h ∥ f ≤ C h k ( ∥ w ∥ H k + 1 ( Ω f ) + ∥ r ∥ H k ( Ω f ) ) , inf ζ h ∈ V p , h ∥ ζ − ζ h ∥ H 1 ( Ω p ) ≤ C h k ∥ ζ ∥ H k + 1 ( Ω p ) , inf ψ h ∈ Q p , h ∥ ψ − ψ h ∥ H 1 ( Ω p ) ≤ C h k ∥ ψ ∥ H k + 1 ( Ω p ) . (54) \begin{aligned}
\inf_{v_h\in V_{f,h}}\lVert\mathbf{w}-\mathbf{v}_h\rVert_{H^1(\Omega_f)}+\inf_{q_h\in Q_{f,h}}\lVert r-q_h\rVert_f&\le Ch^k\left(\lVert\mathbf{w}\rVert_{H^{k+1}(\Omega_f)}+\lVert r\rVert_{H^k(\Omega_f)}\right),\\
\inf_{\boldsymbol{\zeta}_h\in V_{p,h}}\lVert\boldsymbol{\zeta}-\boldsymbol{\zeta}_h\rVert_{H^1(\Omega_p)}&\le Ch^k\lVert\boldsymbol{\zeta}\rVert_{H^{k+1}(\Omega_p)},\\
\inf_{\psi_h\in Q_{p,h}}\lVert\psi-\psi_h\rVert_{H^1(\Omega_p)}&\le Ch^k\lVert\psi\rVert_{H^{k+1}(\Omega_p)}. \tag*{(54)}
\end{aligned} v h ∈ V f , h inf ∥ w − v h ∥ H 1 ( Ω f ) + q h ∈ Q f , h inf ∥ r − q h ∥ f ζ h ∈ V p , h inf ∥ ζ − ζ h ∥ H 1 ( Ω p ) ψ h ∈ Q p , h inf ∥ ψ − ψ h ∥ H 1 ( Ω p ) ≤ C h k ( ∥ w ∥ H k + 1 ( Ω f ) + ∥ r ∥ H k ( Ω f ) ) , ≤ C h k ∥ ζ ∥ H k + 1 ( Ω p ) , ≤ C h k ∥ ψ ∥ H k + 1 ( Ω p ) . (54) The fields in these inequalities belong to the indicated regularity spaces and satisfy their essential boundary conditions.
For the spaces used in § 7 , namely continuous piecewise quadratic elements for the fluid velocity, displacement, structure velocity, and pore pressure and continuous piecewise linear elements for the fluid pressure, (54) holds with k = 2 k=2 k = 2 by standard interpolation estimates that preserve the homogeneous boundary conditions on Σ f \Sigma_f Σ f and Σ p \Sigma_p Σ p [25 ] .
We split each total error into a spatial projection error and a discrete error. The former is bounded in Lemma 5.2 below; the latter will be estimated using the error equations. The projection error, denoted by the superscript I I I , is projected minus exact, and the discrete error, denoted by the superscript h h h , is projected minus discrete. The total errors therefore split as follows:
e u n : = u n − u h n = ( u n − Π u u n ) + ( Π u u n − u h n ) = : − e u I , n + e u h , n , (55) e_{\mathbf{u}}^n:=\mathbf{u}^n-\mathbf{u}_h^n=(\mathbf{u}^n-\Pi_{\mathbf{u}}\mathbf{u}^n)+(\Pi_{\mathbf{u}}\mathbf{u}^n-\mathbf{u}_h^n)=:-e_{\mathbf{u}}^{I,n}+e_{\mathbf{u}}^{h,n}, \tag*{(55)} e u n := u n − u h n = ( u n − Π u u n ) + ( Π u u n − u h n ) =: − e u I , n + e u h , n , (55)
e p n : = p f n − p f , h n = ( p f n − Π p p f n ) + ( Π p p f n − p f , h n ) = : − e p I , n + e p h , n , (56) e_p^n:=p_f^n-p_{f,h}^n=(p_f^n-\Pi_p p_f^n)+(\Pi_p p_f^n-p_{f,h}^n)=:-e_p^{I,n}+e_p^{h,n}, \tag*{(56)} e p n := p f n − p f , h n = ( p f n − Π p p f n ) + ( Π p p f n − p f , h n ) =: − e p I , n + e p h , n , (56)
e η n : = η n − η h n = ( η n − Π η η n ) + ( Π η η n − η h n ) = : − e η I , n + e η h , n , (57) e_{\eta}^n:=\eta^n-\eta_h^n=(\eta^n-\Pi_{\eta}\eta^n)+(\Pi_{\eta}\eta^n-\eta_h^n)=:-e_{\eta}^{I,n}+e_{\eta}^{h,n}, \tag*{(57)} e η n := η n − η h n = ( η n − Π η η n ) + ( Π η η n − η h n ) =: − e η I , n + e η h , n , (57)
e ξ n : = ξ n − ξ h n = ( ξ n − Π ξ ξ n ) + ( Π ξ ξ n − ξ h n ) = : − e ξ I , n + e ξ h , n , (58) e_{\xi}^n:=\xi^n-\xi_h^n=(\xi^n-\Pi_{\xi}\xi^n)+(\Pi_{\xi}\xi^n-\xi_h^n)=:-e_{\xi}^{I,n}+e_{\xi}^{h,n}, \tag*{(58)} e ξ n := ξ n − ξ h n = ( ξ n − Π ξ ξ n ) + ( Π ξ ξ n − ξ h n ) =: − e ξ I , n + e ξ h , n , (58)
e ϕ n : = ϕ n − ϕ h n = ( ϕ n − Π ϕ ϕ n ) + ( Π ϕ ϕ n − ϕ h n ) = : − e ϕ I , n + e ϕ h , n . (59) e_{\phi}^n:=\phi^n-\phi_h^n=(\phi^n-\Pi_{\phi}\phi^n)+(\Pi_{\phi}\phi^n-\phi_h^n)=:-e_{\phi}^{I,n}+e_{\phi}^{h,n}. \tag*{(59)} e ϕ n := ϕ n − ϕ h n = ( ϕ n − Π ϕ ϕ n ) + ( Π ϕ ϕ n − ϕ h n ) =: − e ϕ I , n + e ϕ h , n . (59)
The same definitions apply to the time-dependent projected fields before sampling at t n t_n t n ; for instance, e u I = Π u u − u e_{\mathbf{u}}^I=\Pi_{\mathbf{u}}\mathbf{u}-\mathbf{u} e u I = Π u u − u and e ϕ I = Π ϕ ϕ − ϕ e_{\phi}^I=\Pi_{\phi}\phi-\phi e ϕ I = Π ϕ ϕ − ϕ . The following lemma bounds the projection errors.
Lemma 5.2 (Projection estimates sufficient for the energy bound). Under Assumption 5.1, the positivity and boundary hypotheses, and (5.1), the projections are well defined. For the projection errors in (5.7) and j = 0 , 1 j=0,1 j = 0 , 1 ,
∥ ∂ t j e u I ∥ H 1 ( Ω f ) + ∥ ∂ t j e p I ∥ f + ∥ ∂ t j e η I ∥ H 1 ( Ω p ) + ∥ ∂ t j e ξ I ∥ H 1 ( Ω p ) + ∥ ∂ t j e ϕ I ∥ H 1 ( Ω p ) ≤ C h k M , (60) \left\lVert\partial_t^j e_{\mathbf{u}}^I\right\rVert_{H^1(\Omega_f)}+\left\lVert\partial_t^j e_p^I\right\rVert_f+\left\lVert\partial_t^j e_{\eta}^I\right\rVert_{H^1(\Omega_p)}+\left\lVert\partial_t^j e_{\xi}^I\right\rVert_{H^1(\Omega_p)}+\left\lVert\partial_t^j e_{\phi}^I\right\rVert_{H^1(\Omega_p)}\le Ch^k\mathcal{M}, \tag*{(60)} ∂ t j e u I H 1 ( Ω f ) + ∂ t j e p I f + ∂ t j e η I H 1 ( Ω p ) + ∂ t j e ξ I H 1 ( Ω p ) + ∂ t j e ϕ I H 1 ( Ω p ) ≤ C h k M , (60) ∥ ∂ t j e u I ∥ Γ + ∥ ∂ t j e η I ∥ Γ + ∥ ∂ t j e ξ I ∥ Γ + ∥ ∂ t j e ϕ I ∥ Γ ≤ C h k M . (61) \left\lVert\partial_t^j e_{\mathbf{u}}^I\right\rVert_{\Gamma}+\left\lVert\partial_t^j e_{\eta}^I\right\rVert_{\Gamma}+\left\lVert\partial_t^j e_{\xi}^I\right\rVert_{\Gamma}+\left\lVert\partial_t^j e_{\phi}^I\right\rVert_{\Gamma}\le Ch^k\mathcal{M}. \tag*{(61)} ∂ t j e u I Γ + ∂ t j e η I Γ + ∂ t j e ξ I Γ + ∂ t j e ϕ I Γ ≤ C h k M . (61) Moreover, for 0 ≤ j ≤ 3 0\le j\le3 0 ≤ j ≤ 3 and almost every t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) ,
∥ ∂ t j Π u u ( t ) ∥ H 1 ( Ω f ) ≤ C ( ∥ ∂ t j u ( t ) ∥ H 1 ( Ω f ) + ∥ ∂ t j p f ( t ) ∥ f ) ≤ C M , (62) \left\lVert\partial_t^j\Pi_{\mathbf{u}}\mathbf{u}(t)\right\rVert_{H^1(\Omega_f)}\le C\left(\left\lVert\partial_t^j\mathbf{u}(t)\right\rVert_{H^1(\Omega_f)}+\left\lVert\partial_t^j p_f(t)\right\rVert_f\right)\le C\mathcal{M}, \tag*{(62)} ∂ t j Π u u ( t ) H 1 ( Ω f ) ≤ C ( ∂ t j u ( t ) H 1 ( Ω f ) + ∂ t j p f ( t ) f ) ≤ C M , (62) ∥ ∂ t j Π η η ( t ) ∥ H 1 ( Ω p ) + ∥ ∂ t j Π ξ ξ ( t ) ∥ H 1 ( Ω p ) ≤ C M , ∥ ∂ t j Π ϕ ϕ ( t ) ∥ H 1 ( Ω p ) ≤ C M . \left\lVert\partial_t^j\Pi_{\eta}\eta(t)\right\rVert_{H^1(\Omega_p)}+\left\lVert\partial_t^j\Pi_{\xi}\xi(t)\right\rVert_{H^1(\Omega_p)}\le C\mathcal{M},\qquad
\left\lVert\partial_t^j\Pi_{\phi}\phi(t)\right\rVert_{H^1(\Omega_p)}\le C\mathcal{M}. ∂ t j Π η η ( t ) H 1 ( Ω p ) + ∂ t j Π ξ ξ ( t ) H 1 ( Ω p ) ≤ C M , ∂ t j Π ϕ ϕ ( t ) H 1 ( Ω p ) ≤ C M . Both Ritz maps are H 1 H^1 H 1 stable.
Proof. Spatial estimates. The forms a f R a_f^R a f R , a e a_e a e , and d R d^R d R are continuous and coercive in their respective H 1 H^1 H 1 spaces by the continuous trace, Korn, and Poincaré inequalities under the stated coefficient and boundary assumptions. The discrete mixed stability estimate for (51) therefore gives
∥ Π u u − u ∥ H 1 ( Ω f ) + ∥ Π p p f − p f ∥ f ≤ C ( inf v h ∈ V f , h ∥ u − v h ∥ H 1 ( Ω f ) + inf q h ∈ Q f , h ∥ p f − q h ∥ f ) . \lVert\Pi_{\mathbf{u}}\mathbf{u}-\mathbf{u}\rVert_{H^1(\Omega_f)}+\lVert\Pi_p p_f-p_f\rVert_f\le C\left(\inf_{v_h\in V_{f,h}}\lVert\mathbf{u}-\mathbf{v}_h\rVert_{H^1(\Omega_f)}+\inf_{q_h\in Q_{f,h}}\lVert p_f-q_h\rVert_f\right). ∥ Π u u − u ∥ H 1 ( Ω f ) + ∥ Π p p f − p f ∥ f ≤ C ( v h ∈ V f , h inf ∥ u − v h ∥ H 1 ( Ω f ) + q h ∈ Q f , h inf ∥ p f − q h ∥ f ) .
Céa’s lemma gives the analogous best-approximation estimates for Π η \Pi_\eta Π η and Π ϕ \Pi_\phi Π ϕ ; see [23 ] . Combining these estimates with (5.6) proves (5.8) for j = 0 j=0 j = 0 .
Time derivatives and interface estimates. Since the maps are linear, bounded, and time independent, they commute with time differentiation, because bounded linear maps commute with Bochner integrals. For the Robin–Stokes projection, ( ∂ t j Π u u , ∂ t j Π p p f ) (\partial_t^j\Pi_u u,\partial_t^j\Pi_p p_f) ( ∂ t j Π u u , ∂ t j Π p p f ) is the projection of the pair ( ∂ t j u , ∂ t j p f ) (\partial_t^j u,\partial_t^j p_f) ( ∂ t j u , ∂ t j p f ) . Applying the same estimates to the first time derivatives proves the case j = 1 j=1 j = 1 . The continuous trace theorem then gives (5.9).
Stability of the projected fields. Coercivity also gives ∥ Π η ζ ∥ H 1 ≤ C ∥ ζ ∥ H 1 \|\Pi_\eta\zeta\|_{H^1}\le C\|\zeta\|_{H^1} ∥ Π η ζ ∥ H 1 ≤ C ∥ ζ ∥ H 1 and ∥ Π ϕ ψ ∥ H 1 ≤ C ∥ ψ ∥ H 1 \|\Pi_\phi\psi\|_{H^1}\le C\|\psi\|_{H^1} ∥ Π ϕ ψ ∥ H 1 ≤ C ∥ ψ ∥ H 1 . Mixed stability yields ∥ Π u u ∥ H 1 + ∥ Π p p f ∥ f ≤ C ( ∥ u ∥ H 1 + ∥ p f ∥ f ) \|\Pi_u u\|_{H^1}+\|\Pi_p p_f\|_f\le C(\|u\|_{H^1}+\|p_f\|_f) ∥ Π u u ∥ H 1 + ∥ Π p p f ∥ f ≤ C ( ∥ u ∥ H 1 + ∥ p f ∥ f ) . Apply these stability bounds to time derivatives through order three to obtain (5.10). The bound for Π u u \Pi_u u Π u u in (5.10) involves the time derivatives of p f p_f p f ; this is why M \mathcal{M} M contains the W 3 , ∞ W^{3,\infty} W 3 , ∞ norm of p f p_f p f .
Error equations with residuals
In this subsection we derive the equations satisfied by the discrete errors. Their right-hand sides consist of consistency defects, which we define first, and of residual functionals built from these defects and the projection errors.
Consistency defects. Although the exact fields satisfy ξ = ∂ t η \xi=\partial_t\eta ξ = ∂ t η , replacing the time derivative by BDF2 introduces a mismatch between the projected velocity and the discrete derivative of projected displacement. Define
κ n + 1 : = Π ξ ξ n + 1 − D t Π η η n + 1 . (63) \kappa^{n+1}:=\Pi_\xi\xi^{n+1}-D_t\Pi_\eta\eta^{n+1}. \tag*{(63)} κ n + 1 := Π ξ ξ n + 1 − D t Π η η n + 1 . (63)
The remaining consistency errors arise from approximating time derivatives and predicting interface values. Define
m u n + 1 = D t Π u u n + 1 − ∂ t u ( t n + 1 ) , m ξ n + 1 = D t Π ξ ξ n + 1 − ∂ t ξ ( t n + 1 ) , m ϕ n + 1 = D t Π ϕ ϕ n + 1 − ∂ t ϕ ( t n + 1 ) , z u n + 1 = Π u u ∗ , n + 1 − u ( t n + 1 ) , z ξ n + 1 = Π ξ ξ ∗ , n + 1 − ξ ( t n + 1 ) , z ϕ n + 1 = Π ϕ ϕ ∗ , n + 1 − ϕ ( t n + 1 ) . \begin{aligned}
m_u^{n+1}&=D_t\Pi_u u^{n+1}-\partial_tu(t_{n+1}),&
m_\xi^{n+1}&=D_t\Pi_\xi\xi^{n+1}-\partial_t\xi(t_{n+1}),\\
m_\phi^{n+1}&=D_t\Pi_\phi\phi^{n+1}-\partial_t\phi(t_{n+1}),&
z_u^{n+1}&=\Pi_u u^{*,n+1}-u(t_{n+1}),\\
z_\xi^{n+1}&=\Pi_\xi\xi^{*,n+1}-\xi(t_{n+1}),&
z_\phi^{n+1}&=\Pi_\phi\phi^{*,n+1}-\phi(t_{n+1}).
\end{aligned} m u n + 1 m ϕ n + 1 z ξ n + 1 = D t Π u u n + 1 − ∂ t u ( t n + 1 ) , = D t Π ϕ ϕ n + 1 − ∂ t ϕ ( t n + 1 ) , = Π ξ ξ ∗ , n + 1 − ξ ( t n + 1 ) , m ξ n + 1 z u n + 1 z ϕ n + 1 = D t Π ξ ξ n + 1 − ∂ t ξ ( t n + 1 ) , = Π u u ∗ , n + 1 − u ( t n + 1 ) , = Π ϕ ϕ ∗ , n + 1 − ϕ ( t n + 1 ) .
Residual functionals. The error equations have the same principal operators as the numerical scheme. Their additional right-hand sides record the defects defined above. We group these terms by the equation in which they occur: F 1 F_1 F 1 for the fluid, F 2 F_2 F 2 for elastic mechanics, and F 4 F_4 F 4 for pore pressure. The pore-pressure projection load in mechanics is denoted separately by F 3 F_3 F 3 , since it requires summation by parts in time. For compactness, all defects on the right-hand sides below are evaluated at t n + 1 t_{n+1} t n + 1 unless another time is displayed. For v h ∈ V f , h v_h\in V_{f,h} v h ∈ V f , h , ζ h ∈ V p , h \zeta_h\in V_{p,h} ζ h ∈ V p , h , ζ ∈ V p \zeta\in V_p ζ ∈ V p , and ψ h ∈ Q p , h \psi_h\in Q_{p,h} ψ h ∈ Q p , h , the residuals are
F 1 n + 1 ( v h ) = ρ f ( m u , v h ) f − γ ⟨ P z ξ , P v h ⟩ Γ − L ⟨ z u ⋅ n f , v h ⋅ n f ⟩ Γ + ⟨ z ϕ , v h ⋅ n f ⟩ Γ , (64) F_1^{n+1}(v_h)=\rho_f(m_u,v_h)_f-\gamma\langle Pz_\xi,Pv_h\rangle_\Gamma-L\langle z_u\cdot n_f,v_h\cdot n_f\rangle_\Gamma+\langle z_\phi,v_h\cdot n_f\rangle_\Gamma, \tag*{(64)} F 1 n + 1 ( v h ) = ρ f ( m u , v h ) f − γ ⟨ P z ξ , P v h ⟩ Γ − L ⟨ z u ⋅ n f , v h ⋅ n f ⟩ Γ + ⟨ z ϕ , v h ⋅ n f ⟩ Γ , (64)
F 2 n + 1 ( ζ h ) = ρ p ( m ξ , ζ h ) p + γ ⟨ P ( e ξ I − z u ) , P ζ h ⟩ Γ , (65) F_2^{n+1}(\zeta_h)=\rho_p(m_\xi,\zeta_h)_p+\gamma\left\langle P(e_\xi^I-z_u),P\zeta_h\right\rangle_\Gamma, \tag*{(65)} F 2 n + 1 ( ζ h ) = ρ p ( m ξ , ζ h ) p + γ ⟨ P ( e ξ I − z u ) , P ζ h ⟩ Γ , (65)
F 3 ( t ; ζ ) = − α ( e ϕ I ( t ) , ∇ ⋅ ζ ) p + ⟨ e ϕ I ( t ) , ζ ⋅ n p ⟩ Γ (66) F_3(t;\zeta)=-\alpha(e_\phi^I(t),\nabla\cdot\zeta)_p+\left\langle e_\phi^I(t),\zeta\cdot n_p\right\rangle_\Gamma \tag*{(66)} F 3 ( t ; ζ ) = − α ( e ϕ I ( t ) , ∇ ⋅ ζ ) p + ⟨ e ϕ I ( t ) , ζ ⋅ n p ⟩ Γ (66)
= ( 1 − α ) ( e ϕ I ( t ) , ∇ ⋅ ζ ) p + ( ∇ e ϕ I ( t ) , ζ ) p , (67) =(1-\alpha)(e_\phi^I(t),\nabla\cdot\zeta)_p+(\nabla e_\phi^I(t),\zeta)_p, \tag*{(67)} = ( 1 − α ) ( e ϕ I ( t ) , ∇ ⋅ ζ ) p + ( ∇ e ϕ I ( t ) , ζ ) p , (67)
F 4 n + 1 ( ψ h ) = C 0 ( m ϕ , ψ h ) p + α ( ∇ ⋅ e ξ , ψ h ) p − ⟨ e ξ I ⋅ n p , ψ h ⟩ Γ + ⟨ z u ⋅ n p − L − 1 z ϕ , ψ h ⟩ Γ . (68) F_4^{n+1}(\psi_h)=C_0(m_\phi,\psi_h)_p+\alpha(\nabla\cdot e_\xi,\psi_h)_p-\left\langle e_\xi^I\cdot n_p,\psi_h\right\rangle_\Gamma+\left\langle z_u\cdot n_p-L^{-1}z_\phi,\psi_h\right\rangle_\Gamma. \tag*{(68)} F 4 n + 1 ( ψ h ) = C 0 ( m ϕ , ψ h ) p + α ( ∇ ⋅ e ξ , ψ h ) p − ⟨ e ξ I ⋅ n p , ψ h ⟩ Γ + ⟨ z u ⋅ n p − L − 1 z ϕ , ψ h ⟩ Γ . (68)
Lemma 5.3 (Error equations). For 1 ≤ n ≤ N − 1 1\le n\le N-1 1 ≤ n ≤ N − 1 , the discrete errors satisfy the kinematic relation
e ξ h , n + 1 = D t e η h , n + 1 + κ n + 1 , (69) e_\xi^{h,n+1}=D_t e_\eta^{h,n+1}+\kappa^{n+1}, \tag*{(69)} e ξ h , n + 1 = D t e η h , n + 1 + κ n + 1 , (69)
the fluid error equation
ρ f ( D t e u h , n + 1 , v h ) f + a f R ( e u h , n + 1 , v h ) − ( e p h , n + 1 , ∇ ⋅ v h ) f + ( ∇ ⋅ e u h , n + 1 , q h ) f = γ ⟨ P e ξ h , ∗ , n + 1 , P v h ⟩ Γ + ⟨ L e u h , ∗ , n + 1 ⋅ n f − e ϕ h , ∗ , n + 1 , v h ⋅ n f ⟩ Γ + F 1 n + 1 ( v h ) . (70) \rho_f(D_t e_u^{h,n+1},v_h)_f+a_f^R(e_u^{h,n+1},v_h)-(e_p^{h,n+1},\nabla\cdot v_h)_f+(\nabla\cdot e_u^{h,n+1},q_h)_f
=
\gamma\left\langle P e_\xi^{h,*,n+1},P v_h\right\rangle_\Gamma+\left\langle L e_u^{h,*,n+1}\cdot n_f-e_\phi^{h,*,n+1},v_h\cdot n_f\right\rangle_\Gamma+F_1^{n+1}(v_h). \tag*{(70)} ρ f ( D t e u h , n + 1 , v h ) f + a f R ( e u h , n + 1 , v h ) − ( e p h , n + 1 , ∇ ⋅ v h ) f + ( ∇ ⋅ e u h , n + 1 , q h ) f = γ ⟨ P e ξ h , ∗ , n + 1 , P v h ⟩ Γ + ⟨ L e u h , ∗ , n + 1 ⋅ n f − e ϕ h , ∗ , n + 1 , v h ⋅ n f ⟩ Γ + F 1 n + 1 ( v h ) . (70)
for all ( v h , q h ) ∈ V f , h × Q f , h (v_h,q_h)\in V_{f,h}\times Q_{f,h} ( v h , q h ) ∈ V f , h × Q f , h , and the Biot error equations
ρ p ( D t e ξ h , n + 1 , ζ h ) p + a e ( e η h , n + 1 , ζ h ) − α ( e ϕ h , n + 1 , ∇ ⋅ ζ h ) p + γ ⟨ P e ξ h , n + 1 , P ζ h ⟩ Γ + ⟨ e ϕ h , n + 1 , ζ h ⋅ n p ⟩ Γ = γ ⟨ P e u h , ∗ , n + 1 , P ζ h ⟩ Γ + F 2 n + 1 ( ζ h ) + F 3 ( t n + 1 ; ζ h ) , (71) \rho_p(D_t e_\xi^{h,n+1},\zeta_h)_p+a_e(e_\eta^{h,n+1},\zeta_h)-\alpha(e_\phi^{h,n+1},\nabla\cdot\zeta_h)_p+\gamma\left\langle P e_\xi^{h,n+1},P\zeta_h\right\rangle_\Gamma+\left\langle e_\phi^{h,n+1},\zeta_h\cdot n_p\right\rangle_\Gamma
=
\gamma\left\langle P e_u^{h,*,n+1},P\zeta_h\right\rangle_\Gamma+F_2^{n+1}(\zeta_h)+F_3(t_{n+1};\zeta_h), \tag*{(71)} ρ p ( D t e ξ h , n + 1 , ζ h ) p + a e ( e η h , n + 1 , ζ h ) − α ( e ϕ h , n + 1 , ∇ ⋅ ζ h ) p + γ ⟨ P e ξ h , n + 1 , P ζ h ⟩ Γ + ⟨ e ϕ h , n + 1 , ζ h ⋅ n p ⟩ Γ = γ ⟨ P e u h , ∗ , n + 1 , P ζ h ⟩ Γ + F 2 n + 1 ( ζ h ) + F 3 ( t n + 1 ; ζ h ) , (71)
C 0 ( D t e ϕ h , n + 1 , ψ h ) p + α ( ∇ ⋅ e ξ h , n + 1 , ψ h ) p + d R ( e ϕ h , n + 1 , ψ h ) − ⟨ e ξ h , n + 1 ⋅ n p , ψ h ⟩ Γ = ⟨ − e u h , ∗ , n + 1 ⋅ n p + L − 1 e ϕ h , ∗ , n + 1 , ψ h ⟩ Γ + F 4 n + 1 ( ψ h ) . (72) C_0(D_t e_\phi^{h,n+1},\psi_h)_p+\alpha(\nabla\cdot e_\xi^{h,n+1},\psi_h)_p+d^R(e_\phi^{h,n+1},\psi_h)-\left\langle e_\xi^{h,n+1}\cdot n_p,\psi_h\right\rangle_\Gamma
=
\left\langle-e_u^{h,*,n+1}\cdot n_p+L^{-1}e_\phi^{h,*,n+1},\psi_h\right\rangle_\Gamma+F_4^{n+1}(\psi_h). \tag*{(72)} C 0 ( D t e ϕ h , n + 1 , ψ h ) p + α ( ∇ ⋅ e ξ h , n + 1 , ψ h ) p + d R ( e ϕ h , n + 1 , ψ h ) − ⟨ e ξ h , n + 1 ⋅ n p , ψ h ⟩ Γ = ⟨ − e u h , ∗ , n + 1 ⋅ n p + L − 1 e ϕ h , ∗ , n + 1 , ψ h ⟩ Γ + F 4 n + 1 ( ψ h ) . (72)
for all ( ζ h , ψ h ) ∈ V p , h × Q p , h (\zeta_h,\psi_h)\in V_{p,h}\times Q_{p,h} ( ζ h , ψ h ) ∈ V p , h × Q p , h .
Proof. Since D t η h n + 1 − ξ h n + 1 ∈ V p , h D_t\eta_h^{n+1}-\xi_h^{n+1}\in V_{p,h} D t η h n + 1 − ξ h n + 1 ∈ V p , h , the kinematic equation (22) gives ξ h n + 1 = D t η h n + 1 \xi_h^{n+1}=D_t\eta_h^{n+1} ξ h n + 1 = D t η h n + 1 . Subtracting this identity from (63) gives (69) .
For the fluid equation we evaluate (8) at t n + 1 t_{n+1} t n + 1 and test it with ( v h , q h ) (v_h,q_h) ( v h , q h ) . Adding L ⟨ u ⋅ n f , v h ⋅ n f ⟩ Γ L\langle u\cdot n_f,v_h\cdot n_f\rangle_\Gamma L ⟨ u ⋅ n f , v h ⋅ n f ⟩ Γ to both sides and moving the terms containing ξ \xi ξ and ϕ \phi ϕ to the right, the left-hand side contains a f R ( u , v h ) a_f^R(u,v_h) a f R ( u , v h ) from (19) . By (51) , the elliptic, pressure, and divergence terms are unchanged when ( u , p f ) (u,p_f) ( u , p f ) is replaced by ( Π u u , Π p p f ) (\Pi_u u,\Pi_p p_f) ( Π u u , Π p p f ) . Inserting ∂ t u ( t n + 1 ) = D t Π u u n + 1 − m u n + 1 \partial_tu(t_{n+1})=D_t\Pi_u u^{n+1}-m_u^{n+1} ∂ t u ( t n + 1 ) = D t Π u u n + 1 − m u n + 1 and, in the interface terms, u ( t n + 1 ) = Π u u ∗ , n + 1 − z u n + 1 u(t_{n+1})=\Pi_u u^{*,n+1}-z_u^{n+1} u ( t n + 1 ) = Π u u ∗ , n + 1 − z u n + 1 , ξ ( t n + 1 ) = Π ξ ξ ∗ , n + 1 − z ξ n + 1 \xi(t_{n+1})=\Pi_\xi\xi^{*,n+1}-z_\xi^{n+1} ξ ( t n + 1 ) = Π ξ ξ ∗ , n + 1 − z ξ n + 1 , and ϕ ( t n + 1 ) = Π ϕ ϕ ∗ , n + 1 − z ϕ n + 1 \phi(t_{n+1})=\Pi_\phi\phi^{*,n+1}-z_\phi^{n+1} ϕ ( t n + 1 ) = Π ϕ ϕ ∗ , n + 1 − z ϕ n + 1 , we obtain
ρ f ( D t Π u u n + 1 , v h ) f + a f R ( Π u u n + 1 , v h ) − ( Π p p f n + 1 , ∇ ⋅ v h ) f + ( ∇ ⋅ Π u u n + 1 , q h ) f = ( f n + 1 , v h ) f + γ ⟨ P Π ξ ξ ∗ , n + 1 , P v h ⟩ Γ + ⟨ L Π u u ∗ , n + 1 ⋅ n f − Π ϕ ϕ ∗ , n + 1 , v h ⋅ n f ⟩ Γ + F 1 n + 1 ( v h ) . \rho_f(D_t\Pi_u u^{n+1},v_h)_f+a_f^R(\Pi_u u^{n+1},v_h)-(\Pi_p p_f^{n+1},\nabla\cdot v_h)_f+(\nabla\cdot\Pi_u u^{n+1},q_h)_f
=
(f^{n+1},v_h)_f+\gamma\left\langle P\Pi_\xi\xi^{*,n+1},P v_h\right\rangle_\Gamma+\left\langle L\Pi_u u^{*,n+1}\cdot n_f-\Pi_\phi\phi^{*,n+1},v_h\cdot n_f\right\rangle_\Gamma+F_1^{n+1}(v_h). ρ f ( D t Π u u n + 1 , v h ) f + a f R ( Π u u n + 1 , v h ) − ( Π p p f n + 1 , ∇ ⋅ v h ) f + ( ∇ ⋅ Π u u n + 1 , q h ) f = ( f n + 1 , v h ) f + γ ⟨ P Π ξ ξ ∗ , n + 1 , P v h ⟩ Γ + ⟨ L Π u u ∗ , n + 1 ⋅ n f − Π ϕ ϕ ∗ , n + 1 , v h ⋅ n f ⟩ Γ + F 1 n + 1 ( v h ) .
Subtracting (21) and using the linearity of the extrapolation, for instance Π u u ∗ , n + 1 − u h ∗ , n + 1 = e u h , ∗ , n + 1 \Pi_u u^{*,n+1}-u_h^{*,n+1}=e_u^{h,*,n+1} Π u u ∗ , n + 1 − u h ∗ , n + 1 = e u h , ∗ , n + 1 , gives (70) .
For the Biot equation, by (52) , a e ( η , ζ h ) = a e ( Π η η , ζ h ) a_e(\eta,\zeta_h)=a_e(\Pi_\eta\eta,\zeta_h) a e ( η , ζ h ) = a e ( Π η η , ζ h ) and d R ( ϕ , ψ h ) = d R ( Π ϕ ϕ , ψ h ) d^R(\phi,\psi_h)=d^R(\Pi_\phi\phi,\psi_h) d R ( ϕ , ψ h ) = d R ( Π ϕ ϕ , ψ h ) . We evaluate (9) at t n + 1 t_{n+1} t n + 1 , test it with ζ h \zeta_h ζ h , and insert ∂ t ξ ( t n + 1 ) = D t Π ξ ξ n + 1 − m ξ n + 1 \partial_t\xi(t_{n+1})=D_t\Pi_\xi\xi^{n+1}-m_\xi^{n+1} ∂ t ξ ( t n + 1 ) = D t Π ξ ξ n + 1 − m ξ n + 1 , ξ n + 1 = Π ξ ξ n + 1 − e ξ I , n + 1 \xi^{n+1}=\Pi_\xi\xi^{n+1}-e_\xi^{I,n+1} ξ n + 1 = Π ξ ξ n + 1 − e ξ I , n + 1 , ϕ n + 1 = Π ϕ ϕ n + 1 − e ϕ I , n + 1 \phi^{n+1}=\Pi_\phi\phi^{n+1}-e_\phi^{I,n+1} ϕ n + 1 = Π ϕ ϕ n + 1 − e ϕ I , n + 1 , and, in the tangential coupling term, u ( t n + 1 ) = Π u u ∗ , n + 1 − z u n + 1 u(t_{n+1})=\Pi_u u^{*,n+1}-z_u^{n+1} u ( t n + 1 ) = Π u u ∗ , n + 1 − z u n + 1 . The result is (23) written for the projected fields, with the additional terms F 2 n + 1 ( ζ h ) + F 3 ( t n + 1 ; ζ h ) F_2^{n+1}(\zeta_h)+F_3(t_{n+1};\zeta_h) F 2 n + 1 ( ζ h ) + F 3 ( t n + 1 ; ζ h ) on the right-hand side; subtracting (23) gives (71) . Similarly, we evaluate (10) at t n + 1 t_{n+1} t n + 1 , test it with ψ h \psi_h ψ h , and add L − 1 ⟨ ϕ , ψ h ⟩ Γ L^{-1}\langle\phi,\psi_h\rangle_\Gamma L − 1 ⟨ ϕ , ψ h ⟩ Γ to both sides, so that d R d^R d R from (20) appears on the left. Inserting ∂ t ϕ ( t n + 1 ) = D t Π ϕ ϕ n + 1 − m ϕ n + 1 \partial_t\phi(t_{n+1})=D_t\Pi_\phi\phi^{n+1}-m_\phi^{n+1} ∂ t ϕ ( t n + 1 ) = D t Π ϕ ϕ n + 1 − m ϕ n + 1 and ξ n + 1 = Π ξ ξ n + 1 − e ξ I , n + 1 \xi^{n+1}=\Pi_\xi\xi^{n+1}-e_\xi^{I,n+1} ξ n + 1 = Π ξ ξ n + 1 − e ξ I , n + 1 on the left, and u ( t n + 1 ) = Π u u ∗ , n + 1 − z u n + 1 u(t_{n+1})=\Pi_u u^{*,n+1}-z_u^{n+1} u ( t n + 1 ) = Π u u ∗ , n + 1 − z u n + 1 and ϕ ( t n + 1 ) = Π ϕ ϕ ∗ , n + 1 − z ϕ n + 1 \phi(t_{n+1})=\Pi_\phi\phi^{*,n+1}-z_\phi^{n+1} ϕ ( t n + 1 ) = Π ϕ ϕ ∗ , n + 1 − z ϕ n + 1 on the right, gives (24) written for the projected fields, with the additional term F 4 n + 1 ( ψ h ) F_4^{n+1}(\psi_h) F 4 n + 1 ( ψ h ) . Subtracting (24) gives (72) .
Estimates of the residuals
We now estimate the defects and residual functionals appearing in Lemma 5.3.
Lemma 5.4 (Consistency defects). For 1 ≤ n ≤ N − 1 1\le n\le N-1 1 ≤ n ≤ N − 1 ,
∥ r n + 1 ∥ H 1 ( Ω p ) ≤ C Δ t 2 M . (73) \left\|r^{n+1}\right\|_{H^1(\Omega_p)}\le C\Delta t^2\mathcal{M}. \tag*{(73)} r n + 1 H 1 ( Ω p ) ≤ C Δ t 2 M . (73)
Moreover, the derivative defects are measured in bulk L 2 L^2 L 2 norms, while the extrapolation defects are measured on the interface:
∥ m u n + 1 ∥ f + ∥ m ξ n + 1 ∥ p + ∥ m ϕ n + 1 ∥ p ≤ C ( h k + Δ t 2 ) M , (74) \left\|m_u^{n+1}\right\|_f+\left\|m_\xi^{n+1}\right\|_p+\left\|m_\phi^{n+1}\right\|_p\le C(h^k+\Delta t^2)\mathcal{M}, \tag*{(74)} m u n + 1 f + m ξ n + 1 p + m ϕ n + 1 p ≤ C ( h k + Δ t 2 ) M , (74)
∥ z u n + 1 ∥ Γ + ∥ z ξ n + 1 ∥ Γ + ∥ z ϕ n + 1 ∥ Γ ≤ C ( h k + Δ t 2 ) M . (75) \left\|z_u^{n+1}\right\|_\Gamma+\left\|z_\xi^{n+1}\right\|_\Gamma+\left\|z_\phi^{n+1}\right\|_\Gamma\le C(h^k+\Delta t^2)\mathcal{M}. \tag*{(75)} z u n + 1 Γ + z ξ n + 1 Γ + z ϕ n + 1 Γ ≤ C ( h k + Δ t 2 ) M . (75)
Proof. We first show two consistency error estimates using Taylor expansions. Let X X X be a Banach space, w ∈ W 3 , ∞ ( 0 , T ; X ) w\in W^{3,\infty}(0,T;X) w ∈ W 3 , ∞ ( 0 , T ; X ) , and I n = ( t n − 1 , t n + 1 ) I_n=(t_{n-1},t_{n+1}) I n = ( t n − 1 , t n + 1 ) . Taylor’s formula with integral remainder at t n + 1 t_{n+1} t n + 1 gives, for s ∈ { Δ t , 2 Δ t } s\in\{\Delta t,2\Delta t\} s ∈ { Δ t , 2Δ t } ,
w ( t n + 1 − s ) = w ( t n + 1 ) − s ∂ t w ( t n + 1 ) + s 2 2 ∂ t 2 w ( t n + 1 ) − R ( s ) , w(t_{n+1}-s)=w(t_{n+1})-s\partial_t w(t_{n+1})+\frac{s^2}{2}\partial_t^2w(t_{n+1})-R(s), w ( t n + 1 − s ) = w ( t n + 1 ) − s ∂ t w ( t n + 1 ) + 2 s 2 ∂ t 2 w ( t n + 1 ) − R ( s ) ,
R ( s ) = 1 2 ∫ t n + 1 − s t n + 1 ( τ − t n + 1 + s ) 2 ∂ t 3 w ( τ ) d τ . R(s)=\frac{1}{2}\int_{t_{n+1}-s}^{t_{n+1}}(\tau-t_{n+1}+s)^2\partial_t^3w(\tau)\,\mathrm{d}\tau. R ( s ) = 2 1 ∫ t n + 1 − s t n + 1 ( τ − t n + 1 + s ) 2 ∂ t 3 w ( τ ) d τ .
with ∥ R ( s ) ∥ X ≤ s 3 6 ∥ ∂ t 3 w ∥ L ∞ ( I n ; X ) \left\|R(s)\right\|_X\le\frac{s^3}{6}\left\|\partial_t^3w\right\|_{L^\infty(I_n;X)} ∥ R ( s ) ∥ X ≤ 6 s 3 ∂ t 3 w L ∞ ( I n ; X ) . Hence 3 w n + 1 − 4 w n + w n − 1 = 2 Δ t ∂ t w ( t n + 1 ) + 4 R ( Δ t ) − R ( 2 Δ t ) 3w^{n+1}-4w^n+w^{n-1}=2\Delta t\,\partial_t w(t_{n+1})+4R(\Delta t)-R(2\Delta t) 3 w n + 1 − 4 w n + w n − 1 = 2Δ t ∂ t w ( t n + 1 ) + 4 R ( Δ t ) − R ( 2Δ t ) . Moreover, w n + 1 − 2 w n + w n − 1 = ∫ t n − 1 t n + 1 ( Δ t − ∣ τ − t n ∣ ) ∂ t 2 w ( τ ) d τ w^{n+1}-2w^n+w^{n-1}=\int_{t_{n-1}}^{t_{n+1}}(\Delta t-\lvert\tau-t_n\rvert)\partial_t^2w(\tau)\,\mathrm{d}\tau w n + 1 − 2 w n + w n − 1 = ∫ t n − 1 t n + 1 ( Δ t − ∣ τ − t n ∣) ∂ t 2 w ( τ ) d τ , and the kernel has integral Δ t 2 \Delta t^2 Δ t 2 . Therefore
∥ D t w n + 1 − ∂ t w ( t n + 1 ) ∥ X ≤ Δ t 2 ∥ ∂ t 3 w ∥ L ∞ ( I n ; X ) , ∥ w ∗ , n + 1 − w ( t n + 1 ) ∥ X ≤ Δ t 2 ∥ ∂ t 2 w ∥ L ∞ ( I n ; X ) . (76) \left\|D_tw^{n+1}-\partial_tw(t_{n+1})\right\|_X\le\Delta t^2\left\|\partial_t^3w\right\|_{L^\infty(I_n;X)},\qquad\left\|w^{*,n+1}-w(t_{n+1})\right\|_X\le\Delta t^2\left\|\partial_t^2w\right\|_{L^\infty(I_n;X)}. \tag*{(76)} D t w n + 1 − ∂ t w ( t n + 1 ) X ≤ Δ t 2 ∂ t 3 w L ∞ ( I n ; X ) , w ∗ , n + 1 − w ( t n + 1 ) X ≤ Δ t 2 ∂ t 2 w L ∞ ( I n ; X ) . (76)
We now show the main estimates. Since Π ξ = Π η \Pi_\xi=\Pi_\eta Π ξ = Π η is linear and ξ = ∂ t η \xi=\partial_t\eta ξ = ∂ t η , the common elasticity map gives the exact identity
κ n + 1 = Π η ( ∂ t η ( t n + 1 ) − D t η n + 1 ) . \kappa^{n+1}=\Pi_\eta\left(\partial_t\eta(t_{n+1})-D_t\eta^{n+1}\right). κ n + 1 = Π η ( ∂ t η ( t n + 1 ) − D t η n + 1 ) .
The first estimate in (76) with X = H 1 ( Ω p ) d X=H^1(\Omega_p)^d X = H 1 ( Ω p ) d and w = η w=\eta w = η gives
∥ η ( t n + 1 ) − D t η n + 1 ∥ H 1 ( Ω p ) ≤ C Δ t 2 ∥ ∂ t 3 η ∥ L ∞ ( t n − 1 , t n + 1 ; H 1 ( Ω p ) ) . \left\|\eta(t_{n+1})-D_t\eta^{n+1}\right\|_{H^1(\Omega_p)}\le C\Delta t^2\left\|\partial_t^3\eta\right\|_{L^\infty(t_{n-1},t_{n+1};H^1(\Omega_p))}. η ( t n + 1 ) − D t η n + 1 H 1 ( Ω p ) ≤ C Δ t 2 ∂ t 3 η L ∞ ( t n − 1 , t n + 1 ; H 1 ( Ω p )) .
The H 1 H^1 H 1 stability of Π η \Pi_\eta Π η therefore yields (5.20).
To show (74) , we split m u n + 1 m_u^{n+1} m u n + 1 into a BDF2 defect and a projection error:
m u n + 1 = D t Π u u n + 1 − ∂ t Π u u ( t n + 1 ) + ∂ t e u I ( t n + 1 ) . m_u^{n+1}=D_t\Pi_u u^{n+1}-\partial_t\Pi_u u(t_{n+1})+\partial_t e_u^I(t_{n+1}). m u n + 1 = D t Π u u n + 1 − ∂ t Π u u ( t n + 1 ) + ∂ t e u I ( t n + 1 ) .
By the first estimate in (76) with X = L 2 ( Ω f ) d X=L^2(\Omega_f)^d X = L 2 ( Ω f ) d and w = Π u u w=\Pi_u u w = Π u u , and by (62) with j = 3 j=3 j = 3 ,
∥ D t Π u u n + 1 − ∂ t Π u u ( t n + 1 ) ∥ f ≤ Δ t 2 ∥ ∂ t 3 Π u u ∥ L ∞ ( I n ; H 1 ( Ω f ) ) ≤ C Δ t 2 ( ∥ ∂ t 3 u ∥ L ∞ ( I n ; H 1 ( Ω f ) ) + ∥ ∂ t 3 p f ∥ L ∞ ( I n ; L 2 ( Ω f ) ) ) ≤ C Δ t 2 M . \begin{aligned}
\left\|D_t\Pi_u u^{n+1}-\partial_t\Pi_u u(t_{n+1})\right\|_f\le\Delta t^2\left\|\partial_t^3\Pi_u u\right\|_{L^\infty(I_n;H^1(\Omega_f))} \\
&\le C\Delta t^2\left(\left\|\partial_t^3u\right\|_{L^\infty(I_n;H^1(\Omega_f))}+\left\|\partial_t^3p_f\right\|_{L^\infty(I_n;L^2(\Omega_f))}\right)\le C\Delta t^2\mathcal{M}.
\end{aligned} D t Π u u n + 1 − ∂ t Π u u ( t n + 1 ) f ≤ Δ t 2 ∂ t 3 Π u u L ∞ ( I n ; H 1 ( Ω f )) ≤ C Δ t 2 ( ∂ t 3 u L ∞ ( I n ; H 1 ( Ω f )) + ∂ t 3 p f L ∞ ( I n ; L 2 ( Ω f )) ) ≤ C Δ t 2 M .
The projection error satisfies ∥ ∂ t e u I ( t n + 1 ) ∥ f ≤ C h k M \left\|\partial_t e_u^I(t_{n+1})\right\|_f\le Ch^k\mathcal{M} ∂ t e u I ( t n + 1 ) f ≤ C h k M by (60) with j = 1 j=1 j = 1 . For m ξ m_\xi m ξ and m ϕ m_\phi m ϕ we use the same splitting with Π ξ ξ \Pi_\xi\xi Π ξ ξ and Π ϕ ϕ \Pi_\phi\phi Π ϕ ϕ in place of Π u u \Pi_u u Π u u . By (76) and (62) with j = 3 j=3 j = 3 , their BDF2 defects are bounded by C Δ t 2 M C\Delta t^2\mathcal{M} C Δ t 2 M in L 2 ( Ω p ) L^2(\Omega_p) L 2 ( Ω p ) , and their projection errors ∂ t e ξ I ( t n + 1 ) \partial_t e_\xi^I(t_{n+1}) ∂ t e ξ I ( t n + 1 ) and ∂ t e ϕ I ( t n + 1 ) \partial_t e_\phi^I(t_{n+1}) ∂ t e ϕ I ( t n + 1 ) are bounded by C h k M Ch^k\mathcal{M} C h k M by (60) with j = 1 j=1 j = 1 . This proves (5.21).
For (75) we split z u n + 1 z_u^{n+1} z u n + 1 into an EX2 defect and a projection error:
z u n + 1 = Π u u ∗ , n + 1 − Π u u n + 1 + e u I , n + 1 . z_u^{n+1}=\Pi_u u^{*,n+1}-\Pi_u u^{n+1}+e_u^{I,n+1}. z u n + 1 = Π u u ∗ , n + 1 − Π u u n + 1 + e u I , n + 1 .
By the second estimate in (76) with X = H 1 ( Ω f ) d X=H^{1}(\Omega_f)^d X = H 1 ( Ω f ) d and w = Π u u w=\Pi_u u w = Π u u , and by (62) with j = 2 j=2 j = 2 , the EX2 defect is bounded in H 1 ( Ω f ) H^{1}(\Omega_f) H 1 ( Ω f ) by
C Δ t 2 ( ∥ ∂ t 2 u ∥ L ∞ ( I n ; H 1 ( Ω f ) ) + ∥ ∂ t 2 p f ∥ L ∞ ( I n ; L 2 ( Ω f ) ) ) ≤ C Δ t 2 M ; C\Delta t^2\left(\|\partial_t^2u\|_{L^\infty(I_n;H^1(\Omega_f))}+\|\partial_t^2p_f\|_{L^\infty(I_n;L^2(\Omega_f))}\right)\le C\Delta t^2\mathcal{M}; C Δ t 2 ( ∥ ∂ t 2 u ∥ L ∞ ( I n ; H 1 ( Ω f )) + ∥ ∂ t 2 p f ∥ L ∞ ( I n ; L 2 ( Ω f )) ) ≤ C Δ t 2 M ;
the continuous trace theorem gives the same bound in L 2 ( Γ ) L^2(\Gamma) L 2 ( Γ ) . The projection error satisfies ∥ e u I , n + 1 ∥ ≤ C h k M \|e_u^{I,n+1}\|\le Ch^k\mathcal{M} ∥ e u I , n + 1 ∥ ≤ C h k M by (61) . The bounds for z ξ z_\xi z ξ and z ϕ z_\phi z ϕ follow in the same way, with the bounds for Π ξ ξ \Pi_\xi\xi Π ξ ξ and Π ϕ ϕ \Pi_\phi\phi Π ϕ ϕ in (62) and the trace theorem on Ω p \Omega_p Ω p . Adding these estimates gives (75) .
To measure the pressure-projection load F 3 F_3 F 3 , we give V p V_p V p the equivalent elasticity norm ∥ ζ ∥ a e = a e ( ζ , ζ ) 1 / 2 \|\zeta\|_{a_e}=a_e(\zeta,\zeta)^{1/2} ∥ ζ ∥ a e = a e ( ζ , ζ ) 1/2 . Its dual norm is
∥ F 3 ( t ) ∥ a e ′ = sup 0 ≠ ζ ∈ V p ∣ F 3 ( t ; ζ ) ∣ ∥ ζ ∥ a e . \|F_3(t)\|_{a_e'}=\sup_{0\ne\zeta\in V_p}\frac{|F_3(t;\zeta)|}{\|\zeta\|_{a_e}}. ∥ F 3 ( t ) ∥ a e ′ = 0 = ζ ∈ V p sup ∥ ζ ∥ a e ∣ F 3 ( t ; ζ ) ∣ .
Lemma 5.5 (Residual bounds). For 1 ≤ n ≤ N − 1 1\le n\le N-1 1 ≤ n ≤ N − 1 and all v h ∈ V f , h v_h\in V_{f,h} v h ∈ V f , h , ζ h ∈ V p , h \zeta_h\in V_{p,h} ζ h ∈ V p , h , and ψ h ∈ Q p , h \psi_h\in Q_{p,h} ψ h ∈ Q p , h ,
∣ F 1 n + 1 ( v h ) ∣ ≤ C ( h k + Δ t 2 ) M ∥ v h ∥ H 1 ( Ω f ) , |F_1^{n+1}(v_h)|\le C(h^k+\Delta t^2)\mathcal{M}\|v_h\|_{H^1(\Omega_f)}, ∣ F 1 n + 1 ( v h ) ∣ ≤ C ( h k + Δ t 2 ) M ∥ v h ∥ H 1 ( Ω f ) , ∣ F 2 n + 1 ( ζ h ) ∣ ≤ C ( h k + Δ t 2 ) M ( ∥ ζ h ∥ p + ∥ P ζ h ∥ Γ ) , |F_2^{n+1}(\zeta_h)|\le C(h^k+\Delta t^2)\mathcal{M}\left(\|\zeta_h\|_p+\|P\zeta_h\|_\Gamma\right), ∣ F 2 n + 1 ( ζ h ) ∣ ≤ C ( h k + Δ t 2 ) M ( ∥ ζ h ∥ p + ∥ P ζ h ∥ Γ ) , ∣ F 4 n + 1 ( ψ h ) ∣ ≤ C ( h k + Δ t 2 ) M ∥ ψ h ∥ H 1 ( Ω p ) . (77) |F_4^{n+1}(\psi_h)|\le C(h^k+\Delta t^2)\mathcal{M}\|\psi_h\|_{H^1(\Omega_p)}. \tag*{(77)} ∣ F 4 n + 1 ( ψ h ) ∣ ≤ C ( h k + Δ t 2 ) M ∥ ψ h ∥ H 1 ( Ω p ) . (77) Moreover, the pressure-projection load satisfies
sup 0 ≤ t ≤ T ∥ F 3 ( t ) ∥ a e ′ + ∫ 0 T ∥ ∂ t F 3 ( t ) ∥ a e ′ d t ≤ C T h k M . (78) \sup_{0\le t\le T}\|F_3(t)\|_{a_e'}+\int_0^T\|\partial_tF_3(t)\|_{a_e'}\,dt\le C_T h^k\mathcal{M}. \tag*{(78)} 0 ≤ t ≤ T sup ∥ F 3 ( t ) ∥ a e ′ + ∫ 0 T ∥ ∂ t F 3 ( t ) ∥ a e ′ d t ≤ C T h k M . (78)
Proof. Applying Cauchy–Schwarz and the continuous trace theorem to the individual terms of (64) , (65) , and (68) gives (77) . Here we used (60) –(61) and (74) –(75) . For example,
∣ F 2 n + 1 ( ζ h ) ∣ ≤ ρ p ∥ m ξ ∥ p ∥ ζ h ∥ p + γ ( ∥ P e ξ I ∥ Γ + ∥ P z u ∥ Γ ) ∥ P ζ h ∥ Γ . |F_2^{n+1}(\zeta_h)|\le\rho_p\|m_\xi\|_p\|\zeta_h\|_p+\gamma\left(\|Pe_\xi^I\|_\Gamma+\|Pz_u\|_\Gamma\right)\|P\zeta_h\|_\Gamma. ∣ F 2 n + 1 ( ζ h ) ∣ ≤ ρ p ∥ m ξ ∥ p ∥ ζ h ∥ p + γ ( ∥ P e ξ I ∥ Γ + ∥ P z u ∥ Γ ) ∥ P ζ h ∥ Γ .
Similarly, since ∥ v h ∥ Γ ≤ C ∥ v h ∥ H 1 ( Ω f ) \|v_h\|_\Gamma\le C\|v_h\|_{H^1(\Omega_f)} ∥ v h ∥ Γ ≤ C ∥ v h ∥ H 1 ( Ω f ) and ∥ ψ h ∥ Γ ≤ C ∥ ψ h ∥ H 1 ( Ω p ) \|\psi_h\|_\Gamma\le C\|\psi_h\|_{H^1(\Omega_p)} ∥ ψ h ∥ Γ ≤ C ∥ ψ h ∥ H 1 ( Ω p ) ,
∣ F 1 n + 1 ( v h ) ∣ ≤ ρ f ∥ m u ∥ f ∥ v h ∥ f + ( γ ∥ P z ξ ∥ Γ + L ∥ z u ∥ Γ + ∥ z ϕ ∥ Γ ) ∥ v h ∥ Γ , |F_1^{n+1}(v_h)|\le\rho_f\|m_u\|_f\|v_h\|_f+\left(\gamma\|Pz_\xi\|_\Gamma+L\|z_u\|_\Gamma+\|z_\phi\|_\Gamma\right)\|v_h\|_\Gamma, ∣ F 1 n + 1 ( v h ) ∣ ≤ ρ f ∥ m u ∥ f ∥ v h ∥ f + ( γ ∥ P z ξ ∥ Γ + L ∥ z u ∥ Γ + ∥ z ϕ ∥ Γ ) ∥ v h ∥ Γ ,
∣ F 4 n + 1 ( ψ h ) ∣ ≤ ( C 0 ∥ m ϕ ∥ p + ∣ α ∣ ∥ ∇ ⋅ e ξ I ∥ p ) ∥ ψ h ∥ p + ( ∥ e ξ I ∥ Γ + ∥ z u ∥ Γ + L − 1 ∥ z ϕ ∥ Γ ) ∥ ψ h ∥ Γ . |F_4^{n+1}(\psi_h)|\le\left(C_0\|m_\phi\|_p+|\alpha|\|\nabla\cdot e_\xi^I\|_p\right)\|\psi_h\|_p+\left(\|e_\xi^I\|_\Gamma+\|z_u\|_\Gamma+L^{-1}\|z_\phi\|_\Gamma\right)\|\psi_h\|_\Gamma. ∣ F 4 n + 1 ( ψ h ) ∣ ≤ ( C 0 ∥ m ϕ ∥ p + ∣ α ∣∥∇ ⋅ e ξ I ∥ p ) ∥ ψ h ∥ p + ( ∥ e ξ I ∥ Γ + ∥ z u ∥ Γ + L − 1 ∥ z ϕ ∥ Γ ) ∥ ψ h ∥ Γ .
It remains to prove (78) . Indeed, by (66) –(67) and Korn’s inequality, ∣ F 3 ( t ; ζ ) ∣ ≤ C ∥ e ϕ I ( t ) ∥ H 1 ( Ω p ) ∥ ζ ∥ a e |F_3(t;\zeta)|\le C\|e_\phi^I(t)\|_{H^1(\Omega_p)}\|\zeta\|_{a_e} ∣ F 3 ( t ; ζ ) ∣ ≤ C ∥ e ϕ I ( t ) ∥ H 1 ( Ω p ) ∥ ζ ∥ a e for all ζ ∈ V p \zeta\in V_p ζ ∈ V p , and ∂ t F 3 \partial_tF_3 ∂ t F 3 satisfies the same bound with ∂ t e ϕ I \partial_te_\phi^I ∂ t e ϕ I in place of e ϕ I e_\phi^I e ϕ I ; (60) with j = 0 , 1 j=0,1 j = 0 , 1 then gives (78) .
The bound for F 2 F_2 F 2 thus requires only a bulk L 2 L^2 L 2 norm and a tangential trace norm. A direct estimate of F 3 ( e ξ h ) F_3(e_\xi^h) F 3 ( e ξ h ) would instead require spatial control of the structure-velocity error that is not provided by the energy estimate. We therefore estimate its time sum. The key observation is that e ξ h = D t e η h + κ e_\xi^h=D_te_\eta^h+\kappa e ξ h = D t e η h + κ . We can therefore transfer the time difference from the displacement error to the known load F 3 F_3 F 3 . This replaces a velocity estimate by a displacement-energy estimate. The following lemma transfers the BDF2 difference from a sequence to its coefficient, leaving only endpoint values and first differences of that coefficient.
Lemma 5.6 (BDF2 summation by parts). Let y j y^j y j belong to a normed space Y Y Y and ℓ j ∈ Y ′ \ell^j\in Y' ℓ j ∈ Y ′ . For M ≥ 2 M\ge2 M ≥ 2 ,
∣ Δ t ∑ j = 2 M ℓ j ( D t y j ) ∣ ≤ C ( max 2 ≤ j ≤ M ∥ ℓ j ∥ Y ′ + ∑ j = 2 M − 1 ∥ ℓ j + 1 − ℓ j ∥ Y ′ ) max 0 ≤ j ≤ M ∥ y j ∥ Y . (79) \left|\Delta t\sum_{j=2}^{M}\ell^j(D_ty^j)\right|
\le C\left(\max_{2\le j\le M}\|\ell^j\|_{Y'}+\sum_{j=2}^{M-1}\|\ell^{j+1}-\ell^j\|_{Y'}\right)\max_{0\le j\le M}\|y^j\|_Y. \tag*{(79)} Δ t j = 2 ∑ M ℓ j ( D t y j ) ≤ C ( 2 ≤ j ≤ M max ∥ ℓ j ∥ Y ′ + j = 2 ∑ M − 1 ∥ ℓ j + 1 − ℓ j ∥ Y ′ ) 0 ≤ j ≤ M max ∥ y j ∥ Y . (79) with an absolute constant independent of Δ t \Delta t Δ t and $M.
Proof. For j ≥ 2 j \ge2 j ≥ 2 , 2 Δ t D t y j = ( 3 y j − y j − 1 ) − ( 3 y j − 1 − y j − 2 ) 2\Delta t\,D_t y^j=(3y^j-y^{j-1})-(3y^{j-1}-y^{j-2}) 2Δ t D t y j = ( 3 y j − y j − 1 ) − ( 3 y j − 1 − y j − 2 ) . Summation by parts therefore gives, for every M ≥ 2 M \ge2 M ≥ 2 ,
Δ t ∑ j = 2 M ℓ j ( D t y j ) = 1 2 ℓ M ( 3 y M − y M − 1 ) − 1 2 ℓ 2 ( 3 y 1 − y 0 ) − 1 2 ∑ j = 2 M − 1 ( ℓ j + 1 − ℓ j ) ( 3 y j − y j − 1 ) . (80) \Delta t\sum_{j=2}^{M}\ell^j(D_t y^j)=\frac{1}{2}\ell^M(3y^M-y^{M-1})-\frac{1}{2}\ell^2(3y^1-y^0)-\frac{1}{2}\sum_{j=2}^{M-1}(\ell^{j+1}-\ell^j)(3y^j-y^{j-1}). \tag*{(80)} Δ t j = 2 ∑ M ℓ j ( D t y j ) = 2 1 ℓ M ( 3 y M − y M − 1 ) − 2 1 ℓ 2 ( 3 y 1 − y 0 ) − 2 1 j = 2 ∑ M − 1 ( ℓ j + 1 − ℓ j ) ( 3 y j − y j − 1 ) . (80)
where the last sum is empty for M = 2 M=2 M = 2 . Since ∥ 3 y j − y j − 1 ∥ Y ≤ 4 max 0 ≤ i ≤ M ∥ y i ∥ Y \lVert3y^j-y^{j-1}\rVert_Y\le4\max_{0\le i\le M}\lVert y^i\rVert_Y ∥ 3 y j − y j − 1 ∥ Y ≤ 4 max 0 ≤ i ≤ M ∥ y i ∥ Y , the triangle inequality yields (5.26) with C = 4 C=4 C = 4 .
Lemma 5.7 (Control of the pressure-projection load). For 2 ≤ M ≤ N 2\le M\le N 2 ≤ M ≤ N and every ϵ > 0 \epsilon>0 ϵ > 0 ,
∣ Δ t ∑ j = 2 M F 3 ( t j ; e ξ h , j ) ∣ ≤ ϵ max 0 ≤ j ≤ M ∥ e η h , j ∥ a e 2 + C T , ϵ ( h 2 k + Δ t 4 ) M 2 . (81) \left|\Delta t\sum_{j=2}^{M}F_3(t_j;e_\xi^{h,j})\right|\le\epsilon\max_{0\le j\le M}\lVert e_\eta^{h,j}\rVert_{a_e}^2+C_{T,\epsilon}(h^{2k}+\Delta t^4)M^2. \tag*{(81)} Δ t j = 2 ∑ M F 3 ( t j ; e ξ h , j ) ≤ ϵ 0 ≤ j ≤ M max ∥ e η h , j ∥ a e 2 + C T , ϵ ( h 2 k + Δ t 4 ) M 2 . (81) where C T , ϵ C_{T,\epsilon} C T , ϵ is independent of h h h , Δ t \Delta t Δ t , and M M M .
Proof. Using (5.17), write
Δ t ∑ j = 2 M F 3 ( t j ; e ξ h , j ) = Δ t ∑ j = 2 M F 3 ( t j ; D t e η h , j ) + Δ t ∑ j = 2 M F 3 ( t j ; κ j ) . (82) \Delta t\sum_{j=2}^{M}F_3(t_j;e_\xi^{h,j})=\Delta t\sum_{j=2}^{M}F_3(t_j;D_te_\eta^{h,j})+\Delta t\sum_{j=2}^{M}F_3(t_j;\kappa^j). \tag*{(82)} Δ t j = 2 ∑ M F 3 ( t j ; e ξ h , j ) = Δ t j = 2 ∑ M F 3 ( t j ; D t e η h , j ) + Δ t j = 2 ∑ M F 3 ( t j ; κ j ) . (82)
The fundamental theorem of calculus gives
∑ j = 2 M − 1 ∥ F 3 ( t j + 1 ) − F 3 ( t j ) ∥ a e ≤ ∫ t 2 t M ∥ ∂ t F 3 ( t ) ∥ a e d t . \sum_{j=2}^{M-1}\lVert F_3(t_{j+1})-F_3(t_j)\rVert_{a_e}\le\int_{t_2}^{t_M}\lVert\partial_tF_3(t)\rVert_{a_e}\,dt. j = 2 ∑ M − 1 ∥ F 3 ( t j + 1 ) − F 3 ( t j ) ∥ a e ≤ ∫ t 2 t M ∥ ∂ t F 3 ( t ) ∥ a e d t .
Thus the variation is bounded by (5.25), independently of the number of time steps. Applying Lemma 5.6 to the first sum in (5.29), with Y = V p Y=V_p Y = V p normed by ∥ ⋅ ∥ a e \lVert\cdot\rVert_{a_e} ∥ ⋅ ∥ a e , ℓ j = F 3 ( t j ) \ell^j=F_3(t_j) ℓ j = F 3 ( t j ) , and y j = e η h , j y^j=e_\eta^{h,j} y j = e η h , j , and using (5.20) for the second, gives
∣ Δ t ∑ j = 2 M F 3 ( t j ; e ξ h , j ) ∣ ≤ C T h k M max 0 ≤ j ≤ M ∥ e η h , j ∥ a e + C T h k Δ t 2 M 2 . \left|\Delta t\sum_{j=2}^{M}F_3(t_j;e_\xi^{h,j})\right|\le C_T h^k M\max_{0\le j\le M}\lVert e_\eta^{h,j}\rVert_{a_e}+C_T h^k\Delta t^2M^2. Δ t j = 2 ∑ M F 3 ( t j ; e ξ h , j ) ≤ C T h k M 0 ≤ j ≤ M max ∥ e η h , j ∥ a e + C T h k Δ t 2 M 2 .
Young’s inequality and h k Δ t 2 ≤ 1 2 ( h 2 k + Δ t 4 ) h^k\Delta t^2\le\frac{1}{2}(h^{2k}+\Delta t^4) h k Δ t 2 ≤ 2 1 ( h 2 k + Δ t 4 ) give (5.28).
Energy estimate and convergence theorem
Let E h j \mathcal{E}_h^j E h j , N h j \mathcal{N}_h^j N h j , and D ∗ , h j \mathcal{D}_{*,h}^j D ∗ , h j be (4.4), (4.5), and (4.10) with all variables replaced by their discrete errors.
Theorem 5.8 (Bulk energy error). Assume the model, mesh, trace, regularity, and finite element approximation hypotheses above. Suppose the initialization satisfies
E h 1 ≤ C i n i t ( h 2 k + Δ t 4 ) . (83) \mathcal{E}_h^1\le C_{\mathrm{init}}(h^{2k}+\Delta t^4). \tag*{(83)} E h 1 ≤ C init ( h 2 k + Δ t 4 ) . (83) If Δ t ≤ C ∗ h \Delta t\le C_*h Δ t ≤ C ∗ h , then
max 1 ≤ j ≤ N ( ∥ e u h , j ∥ f 2 + ∥ e ξ h , j ∥ p 2 + ∥ e ϕ h , j ∥ p 2 + ∥ e η h , j ∥ a e 2 ) + Δ t ∑ j = 2 N ( ∥ e u h , j ∥ H 1 ( Ω f ) 2 + ∥ e ϕ h , j ∥ H 1 ( Ω p ) 2 + ∥ P e ξ h , j ∥ Γ 2 ) ≤ C T ( h 2 k + Δ t 4 ) . (84) \max_{1\le j\le N}\left(\lVert e_u^{h,j}\rVert_f^2+\lVert e_\xi^{h,j}\rVert_p^2+\lVert e_\phi^{h,j}\rVert_p^2+\lVert e_\eta^{h,j}\rVert_{a_e}^2\right)
+\Delta t\sum_{j=2}^{N}\left(\lVert e_u^{h,j}\rVert_{H^1(\Omega_f)}^2+\lVert e_\phi^{h,j}\rVert_{H^1(\Omega_p)}^2+\lVert Pe_\xi^{h,j}\rVert_{\Gamma}^2\right)\le C_T(h^{2k}+\Delta t^4). \tag*{(84)} 1 ≤ j ≤ N max ( ∥ e u h , j ∥ f 2 + ∥ e ξ h , j ∥ p 2 + ∥ e ϕ h , j ∥ p 2 + ∥ e η h , j ∥ a e 2 ) + Δ t j = 2 ∑ N ( ∥ e u h , j ∥ H 1 ( Ω f ) 2 + ∥ e ϕ h , j ∥ H 1 ( Ω p ) 2 + ∥ P e ξ h , j ∥ Γ 2 ) ≤ C T ( h 2 k + Δ t 4 ) . (84)
The same bound holds for the total errors u − u h \mathbf{u}-\mathbf{u}_h u − u h , ξ − ξ h \boldsymbol{\xi}-\boldsymbol{\xi}_h ξ − ξ h , ϕ − ϕ h \phi-\phi_h ϕ − ϕ h , and η − η h \boldsymbol{\eta}-\boldsymbol{\eta}_h η − η h , with the indicated norms. In particular,
max 1 ≤ j ≤ N ( ∥ u ( t j ) − u h j ∥ f + ∥ ξ ( t j ) − ξ h j ∥ p + ∥ ϕ ( t j ) − ϕ h j ∥ p + ∥ η ( t j ) − η h j ∥ H 1 ( Ω p ) ) ≤ C T ( h k + Δ t 2 ) . (85) \max_{1\le j\le N}\left(\left\|\mathbf{u}(t_j)-\mathbf{u}_h^j\right\|_f+\left\|\boldsymbol{\xi}(t_j)-\boldsymbol{\xi}_h^j\right\|_p+\left\|\phi(t_j)-\phi_h^j\right\|_p+\left\|\boldsymbol{\eta}(t_j)-\boldsymbol{\eta}_h^j\right\|_{H^1(\Omega_p)}\right)\le C_T(h^k+\Delta t^2). \tag*{(85)} 1 ≤ j ≤ N max ( u ( t j ) − u h j f + ξ ( t j ) − ξ h j p + ϕ ( t j ) − ϕ h j p + η ( t j ) − η h j H 1 ( Ω p ) ) ≤ C T ( h k + Δ t 2 ) . (85)
The constant can depend on T T T , fixed coefficients, projection constants, M \mathcal{M} M , and initialization quality, but is independent of h h h , Δ t \Delta t Δ t .
Proof. We now combine the preceding bounds with the interface stability estimates. The argument is organized into the energy identity, residual control, and a summation step in which the maximum of the error energy is absorbed; no discrete Gronwall inequality is used.
Step 1: The error energy identity. Testing (70) –(5.19) with ( e u h , n + 1 , e h n + 1 , e ξ h , n + 1 , e ϕ h , n + 1 ) (e_u^{h,n+1},e_h^{n+1},e_\xi^{h,n+1},e_\phi^{h,n+1}) ( e u h , n + 1 , e h n + 1 , e ξ h , n + 1 , e ϕ h , n + 1 ) gives
E h n + 1 − E h n Δ t + N h n + 1 + 2 μ f ∥ D ( e u h , n + 1 ) ∥ f 2 + ∥ K 1 / 2 ∇ e ϕ h , n + 1 ∥ p 2 + γ ∥ P ( e u h , n + 1 − e ξ h , n + 1 ) ∥ Γ 2 = R Γ n + 1 ( e h ) + F 1 n + 1 ( e u h , n + 1 ) + F 2 n + 1 ( e ξ h , n + 1 ) + F 4 n + 1 ( e ϕ h , n + 1 ) + F 3 ( t n + 1 ; e ξ h , n + 1 ) − a e ( e η h , n + 1 , κ n + 1 ) . (86) \begin{aligned}
\frac{\mathcal{E}_h^{n+1}-\mathcal{E}_h^n}{\Delta t}+\mathcal{N}_h^{n+1}+2\mu_f\left\|\mathbf{D}(e_u^{h,n+1})\right\|_f^2+\left\|K^{1/2}\nabla e_\phi^{h,n+1}\right\|_p^2+\gamma\left\|\mathbf{P}(e_u^{h,n+1}-e_\xi^{h,n+1})\right\|_\Gamma^2 \\
&=\mathcal{R}_\Gamma^{n+1}(e^h)+\mathcal{F}_1^{n+1}(e_u^{h,n+1})+\mathcal{F}_2^{n+1}(e_\xi^{h,n+1})+\mathcal{F}_4^{n+1}(e_\phi^{h,n+1})+\mathcal{F}_3(t_{n+1};e_\xi^{h,n+1})-a_e(e_\eta^{h,n+1},\kappa^{n+1}). \tag*{(86)}
\end{aligned} Δ t E h n + 1 − E h n + N h n + 1 + 2 μ f D ( e u h , n + 1 ) f 2 + K 1/2 ∇ e ϕ h , n + 1 p 2 + γ P ( e u h , n + 1 − e ξ h , n + 1 ) Γ 2 = R Γ n + 1 ( e h ) + F 1 n + 1 ( e u h , n + 1 ) + F 2 n + 1 ( e ξ h , n + 1 ) + F 4 n + 1 ( e ϕ h , n + 1 ) + F 3 ( t n + 1 ; e ξ h , n + 1 ) − a e ( e η h , n + 1 , κ n + 1 ) . (86)
The last term follows from a e ( e η h , n + 1 , e ξ h ) = a e ( e η h , n + 1 , D t e η h ) + a e ( e η h , n + 1 , κ ) a_e(e_\eta^{h,n+1},e_\xi^{h})=a_e(e_\eta^{h,n+1},D_t e_\eta^{h})+a_e(e_\eta^{h,n+1},\kappa) a e ( e η h , n + 1 , e ξ h ) = a e ( e η h , n + 1 , D t e η h ) + a e ( e η h , n + 1 , κ ) ; there is no additional kinematic testing term.
Step 2: The per-step error inequality. All arguments of R Γ n + 1 ( e h ) \mathcal{R}_\Gamma^{n+1}(e^h) R Γ n + 1 ( e h ) are finite element functions, so Steps 2 and 3 of the proof of Theorem 4.1 apply verbatim under Δ t ≤ C ∗ h \Delta t\le C_*h Δ t ≤ C ∗ h : R Γ n + 1 ( e h ) \mathcal{R}_\Gamma^{n+1}(e^h) R Γ n + 1 ( e h ) satisfies (46) , and (45) holds for ( e u h , n + 1 , e ξ h , n + 1 ) (e_u^{h,n+1},e_\xi^{h,n+1}) ( e u h , n + 1 , e ξ h , n + 1 ) . Hence (86) gives
E h n + 1 − E h n Δ t + 1 2 N h n + 1 + D ∗ n + 1 , h ≤ F 1 n + 1 ( e u h , n + 1 ) + F 2 n + 1 ( e ξ h , n + 1 ) + F 4 n + 1 ( e ϕ h , n + 1 ) + F 3 ( t n + 1 ; e ξ h , n + 1 ) − a e ( e η h , n + 1 , κ n + 1 ) . \frac{\mathcal{E}_h^{n+1}-\mathcal{E}_h^n}{\Delta t}+\frac{1}{2}\mathcal{N}_h^{n+1}+\mathcal{D}_*^{n+1,h}\le\mathcal{F}_1^{n+1}(e_u^{h,n+1})+\mathcal{F}_2^{n+1}(e_\xi^{h,n+1})+\mathcal{F}_4^{n+1}(e_\phi^{h,n+1})+\mathcal{F}_3(t_{n+1};e_\xi^{h,n+1})-a_e(e_\eta^{h,n+1},\kappa^{n+1}). Δ t E h n + 1 − E h n + 2 1 N h n + 1 + D ∗ n + 1 , h ≤ F 1 n + 1 ( e u h , n + 1 ) + F 2 n + 1 ( e ξ h , n + 1 ) + F 4 n + 1 ( e ϕ h , n + 1 ) + F 3 ( t n + 1 ; e ξ h , n + 1 ) − a e ( e η h , n + 1 , κ n + 1 ) .
By (77) , Korn’s and Poincaré’s inequalities, and Young’s inequality, using ( h k + Δ t 2 ) 2 ≤ 2 ( h 2 k + Δ t 4 ) (h^k+\Delta t^2)^2\le2(h^{2k}+\Delta t^4) ( h k + Δ t 2 ) 2 ≤ 2 ( h 2 k + Δ t 4 ) ,
∣ F 1 n + 1 ( e u h , n + 1 ) ∣ ≤ μ f 2 ∥ D ( e u h , n + 1 ) ∥ f 2 + C ( h 2 k + Δ t 4 ) M 2 , \left|\mathcal{F}_1^{n+1}(e_u^{h,n+1})\right|\le\frac{\mu_f}{2}\left\|\mathbf{D}(e_u^{h,n+1})\right\|_f^2+C(h^{2k}+\Delta t^4)\mathcal{M}^2, F 1 n + 1 ( e u h , n + 1 ) ≤ 2 μ f D ( e u h , n + 1 ) f 2 + C ( h 2 k + Δ t 4 ) M 2 ,
∣ F 4 n + 1 ( e ϕ h , n + 1 ) ∣ ≤ 1 4 ∥ K 1 / 2 ∇ e ϕ h , n + 1 ∥ p 2 + C ( h 2 k + Δ t 4 ) M 2 . \left|\mathcal{F}_4^{n+1}(e_\phi^{h,n+1})\right|\le\frac{1}{4}\left\|K^{1/2}\nabla e_\phi^{h,n+1}\right\|_p^2+C(h^{2k}+\Delta t^4)\mathcal{M}^2. F 4 n + 1 ( e ϕ h , n + 1 ) ≤ 4 1 K 1/2 ∇ e ϕ h , n + 1 p 2 + C ( h 2 k + Δ t 4 ) M 2 .
We split F 2 n + 1 ( e ξ h , n + 1 ) \mathcal{F}_2^{n+1}(e_\xi^{h,n+1}) F 2 n + 1 ( e ξ h , n + 1 ) into its bulk part ρ p ( m ξ n + 1 , e ξ h , n + 1 ) p \rho_p(\mathbf{m}_\xi^{n+1},e_\xi^{h,n+1})_p ρ p ( m ξ n + 1 , e ξ h , n + 1 ) p and its tangential part; by (61) and (75) ,
γ ∣ ⟨ P ( e ξ I , n + 1 − z u n + 1 ) , P e ξ h , n + 1 ⟩ Γ ∣ ≤ γ Γ 4 ∥ P e ξ h , n + 1 ∥ Γ 2 + C ( h 2 k + Δ t 4 ) M 2 . \gamma\left|\left\langle\mathbf{P}(e_\xi^{I,n+1}-\mathbf{z}_u^{n+1}),\mathbf{P}e_\xi^{h,n+1}\right\rangle_\Gamma\right|\le\frac{\gamma\Gamma}{4}\left\|\mathbf{P}e_\xi^{h,n+1}\right\|_\Gamma^2+C(h^{2k}+\Delta t^4)\mathcal{M}^2. γ ⟨ P ( e ξ I , n + 1 − z u n + 1 ) , P e ξ h , n + 1 ⟩ Γ ≤ 4 γ Γ P e ξ h , n + 1 Γ 2 + C ( h 2 k + Δ t 4 ) M 2 .
These three absorptions use one half of D ∗ n + 1 , h \mathcal{D}_*^{n+1,h} D ∗ n + 1 , h . Hence, for n ≥ 1 n\ge1 n ≥ 1 ,
E h n + 1 − E h n Δ t + 1 2 N h n + 1 + 1 2 D ∗ n + 1 , h ≤ ρ p ( m ξ n + 1 , e ξ h , n + 1 ) p − a e ( e η h , n + 1 , κ n + 1 ) + F 3 ( t n + 1 ; e ξ h , n + 1 ) + C 1 ( h 2 k + Δ t 4 ) M 2 . (87) \frac{\mathcal{E}_h^{n+1}-\mathcal{E}_h^n}{\Delta t}+\frac{1}{2}\mathcal{N}_h^{n+1}+\frac{1}{2}\mathcal{D}_*^{n+1,h}\le\rho_p(\mathbf{m}_\xi^{n+1},e_\xi^{h,n+1})_p-a_e(e_\eta^{h,n+1},\kappa^{n+1})+\mathcal{F}_3(t_{n+1};e_\xi^{h,n+1})+C_1(h^{2k}+\Delta t^4)\mathcal{M}^2. \tag*{(87)} Δ t E h n + 1 − E h n + 2 1 N h n + 1 + 2 1 D ∗ n + 1 , h ≤ ρ p ( m ξ n + 1 , e ξ h , n + 1 ) p − a e ( e η h , n + 1 , κ n + 1 ) + F 3 ( t n + 1 ; e ξ h , n + 1 ) + C 1 ( h 2 k + Δ t 4 ) M 2 . (87)
The first two terms on the right are tested with the bulk structure-velocity and displacement errors, which are not directly controlled by the dissipation. We do not bound them by the current energy, which would lead to a Gronwall argument; instead, we sum them in time and use the maximum of the error energy.
Step 3: Summation in time. Let $2 \le m \le M \le N$. Multiplying (5.34) by Δ t \Delta t Δ t and summing over n = 1 , … , m − 1 n=1,\ldots,m-1 n = 1 , … , m − 1 gives, since ( m − 1 ) Δ t ≤ T (m-1)\Delta t \le T ( m − 1 ) Δ t ≤ T ,
E h m + Δ t 2 ∑ j = 2 m ( N h j + D ∗ , h j ) ≤ E h 1 + C 1 T ( h 2 k + Δ t 4 ) M 2 + ∣ Δ t ∑ j = 2 m F 3 ( t j ; e ξ h , j ) ∣ + Δ t ∑ j = 2 m ( ρ p ∥ m ξ j ∥ p ∥ e ξ h , j ∥ p + ∥ e η h , j ∥ a e ∥ k j ∥ a e ) . (88) \mathcal{E}_h^m+\frac{\Delta t}{2}\sum_{j=2}^{m}\left(\mathcal{N}_h^j+\mathcal{D}_{*,h}^j\right)\le\mathcal{E}_h^1+C_1T\left(h^{2k}+\Delta t^4\right)\mathcal{M}^2+\left|\Delta t\sum_{j=2}^{m}F_3\left(t_j;e_\xi^{h,j}\right)\right|
+\Delta t\sum_{j=2}^{m}\left(\rho_p\left\|m_\xi^j\right\|_p\left\|e_\xi^{h,j}\right\|_p+\left\|e_\eta^{h,j}\right\|_{a_e}\left\|k^j\right\|_{a_e}\right). \tag*{(88)} E h m + 2 Δ t j = 2 ∑ m ( N h j + D ∗ , h j ) ≤ E h 1 + C 1 T ( h 2 k + Δ t 4 ) M 2 + Δ t j = 2 ∑ m F 3 ( t j ; e ξ h , j ) + Δ t j = 2 ∑ m ( ρ p m ξ j p e ξ h , j p + e η h , j a e k j a e ) . (88)
Let S M = max 1 ≤ j ≤ M E h j S_M=\max_{1\le j\le M}\mathcal{E}_h^j S M = max 1 ≤ j ≤ M E h j . By the definition of the BDF2 energy, ρ p ∥ e ξ h , j ∥ p 2 ≤ 4 E h j \rho_p\left\|e_\xi^{h,j}\right\|_p^2\le4\mathcal{E}_h^j ρ p e ξ h , j p 2 ≤ 4 E h j and ∥ e η h , j ∥ a e 2 ≤ 4 E h j \left\|e_\eta^{h,j}\right\|_{a_e}^2\le4\mathcal{E}_h^j e η h , j a e 2 ≤ 4 E h j for j ≥ 1 j\ge1 j ≥ 1 . Moreover, the identity e η h , 0 = 2 e η h , 1 − ( 2 e η h , 1 − e η h , 0 ) e_\eta^{h,0}=2e_\eta^{h,1}-(2e_\eta^{h,1}-e_\eta^{h,0}) e η h , 0 = 2 e η h , 1 − ( 2 e η h , 1 − e η h , 0 ) gives ∥ e η h , 0 ∥ a e 2 ≤ 5 ( ∥ e η h , 1 ∥ a e 2 + ∥ 2 e η h , 1 − e η h , 0 ∥ a e 2 ) ≤ 20 E h 1 \left\|e_\eta^{h,0}\right\|_{a_e}^2\le5\left(\left\|e_\eta^{h,1}\right\|_{a_e}^2+\left\|2e_\eta^{h,1}-e_\eta^{h,0}\right\|_{a_e}^2\right)\le20\mathcal{E}_h^1 e η h , 0 a e 2 ≤ 5 ( e η h , 1 a e 2 + 2 e η h , 1 − e η h , 0 a e 2 ) ≤ 20 E h 1 . Hence
max 2 ≤ j ≤ M ∥ e ξ h , j ∥ p ≤ 2 ρ p − 1 / 2 S M 1 / 2 , max 0 ≤ j ≤ M ∥ e η h , j ∥ a e 2 ≤ 20 S M . (89) \max_{2\le j\le M}\left\|e_\xi^{h,j}\right\|_p\le2\rho_p^{-1/2}S_M^{1/2},\qquad\max_{0\le j\le M}\left\|e_\eta^{h,j}\right\|_{a_e}^2\le20S_M. \tag*{(89)} 2 ≤ j ≤ M max e ξ h , j p ≤ 2 ρ p − 1/2 S M 1/2 , 0 ≤ j ≤ M max e η h , j a e 2 ≤ 20 S M . (89)
We bound the last three terms of (5.35). By (5.21), (5.36), and Young’s inequality,
Δ t ∑ j = 2 m ρ p ∥ m ξ j ∥ p ∥ e ξ h , j ∥ p ≤ C T ( h k + Δ t 2 ) M S M 1 / 2 ≤ 1 8 S M + C T 2 ( h 2 k + Δ t 4 ) M 2 . \Delta t\sum_{j=2}^{m}\rho_p\left\|m_\xi^j\right\|_p\left\|e_\xi^{h,j}\right\|_p\le C_T\left(h^k+\Delta t^2\right)\mathcal{M}S_M^{1/2}\le\frac{1}{8}S_M+C_T^2\left(h^{2k}+\Delta t^4\right)\mathcal{M}^2. Δ t j = 2 ∑ m ρ p m ξ j p e ξ h , j p ≤ C T ( h k + Δ t 2 ) M S M 1/2 ≤ 8 1 S M + C T 2 ( h 2 k + Δ t 4 ) M 2 .
By the continuity of a e a_e a e , (5.20), and (5.36),
Δ t ∑ j = 2 m ∥ e η h , j ∥ a e ∥ k j ∥ a e ≤ C T Δ t 2 M S M 1 / 2 ≤ 1 8 S M + C T 2 Δ t 4 M 2 . \Delta t\sum_{j=2}^{m}\left\|e_\eta^{h,j}\right\|_{a_e}\left\|k^j\right\|_{a_e}\le C_T\Delta t^2\mathcal{M}S_M^{1/2}\le\frac{1}{8}S_M+C_T^2\Delta t^4\mathcal{M}^2. Δ t j = 2 ∑ m e η h , j a e k j a e ≤ C T Δ t 2 M S M 1/2 ≤ 8 1 S M + C T 2 Δ t 4 M 2 .
Finally, (5.28) with M M M replaced by m m m and ϵ = 1 / 80 \epsilon=1/80 ϵ = 1/80 , together with (5.36), bounds the F 3 F_3 F 3 -term by 1 4 S M + C T ( h 2 k + Δ t 4 ) M 2 \frac{1}{4}S_M+C_T\left(h^{2k}+\Delta t^4\right)\mathcal{M}^2 4 1 S M + C T ( h 2 k + Δ t 4 ) M 2 . Combining these bounds with (5.35) yields, for every 2 ≤ m ≤ M 2\le m\le M 2 ≤ m ≤ M ,
E h m + Δ t 2 ∑ j = 2 m ( N h j + D ∗ , h j ) ≤ E h 1 + 1 2 S M + C T ( h 2 k + Δ t 4 ) M 2 , (90) \mathcal{E}_h^m+\frac{\Delta t}{2}\sum_{j=2}^{m}\left(\mathcal{N}_h^j+\mathcal{D}_{*,h}^j\right)\le\mathcal{E}_h^1+\frac{1}{2}S_M+C_T\left(h^{2k}+\Delta t^4\right)\mathcal{M}^2, \tag*{(90)} E h m + 2 Δ t j = 2 ∑ m ( N h j + D ∗ , h j ) ≤ E h 1 + 2 1 S M + C T ( h 2 k + Δ t 4 ) M 2 , (90)
where C T C_T C T is independent of h h h , Δ t \Delta t Δ t , m m m , and M M M .
Step 4: Absorption of the maximum. Write K T = C T ( h 2 k + Δ t 4 ) M 2 K_T=C_T\left(h^{2k}+\Delta t^4\right)\mathcal{M}^2 K T = C T ( h 2 k + Δ t 4 ) M 2 for the last term in (5.37). Inequality (5.37) also holds for m = 1 m=1 m = 1 , where the sum is empty. Dropping the nonnegative dissipation and taking the maximum over 1 ≤ m ≤ M 1\le m\le M 1 ≤ m ≤ M gives S M ≤ E h 1 + 1 2 S M + K T S_M\le\mathcal{E}_h^1+\frac{1}{2}S_M+K_T S M ≤ E h 1 + 2 1 S M + K T , that is, S M ≤ 2 E h 1 + 2 K T S_M\le2\mathcal{E}_h^1+2K_T S M ≤ 2 E h 1 + 2 K T . Inserting this into (5.37) with m = M m=M m = M gives
E h M + Δ t 2 ∑ j = 2 M ( N h j + D ∗ , h j ) ≤ 2 E h 1 + 2 K T , 2 ≤ M ≤ N . \mathcal{E}_h^M+\frac{\Delta t}{2}\sum_{j=2}^{M}\left(\mathcal{N}_h^j+\mathcal{D}_{*,h}^j\right)\le2\mathcal{E}_h^1+2K_T,\qquad2\le M\le N. E h M + 2 Δ t j = 2 ∑ M ( N h j + D ∗ , h j ) ≤ 2 E h 1 + 2 K T , 2 ≤ M ≤ N .
Step 5: Dissipation and total errors. The bound for S N S_N S N controls the maximum in (5.31), and the last display controls the summed numerical and physical dissipation. Korn and Poincaré inequalities convert the fluid and scalar parts of D ∗ , h \mathcal{D}_{*,h} D ∗ , h into the stated H 1 H^1 H 1 norms. Finally, use (5.30), the approximation assumptions, the triangle inequality, and h ≤ 1 h\le1 h ≤ 1 to obtain the total-error estimate. ▫
Comparison with Robin-Robin splitting scheme
In this section we identify the residual that leads to the sufficient restriction Δ t ≲ h 2 \Delta t \lesssim h^2 Δ t ≲ h 2 in the analysis of the Robin–Robin splitting in [1 ] . The mechanical normal condition in [1 ] has the form
n p ⋅ σ p n + 1 n p + ϕ n + 1 + β ξ n + 1 ⋅ n p = β ξ ∗ , n + 1 ⋅ n p , (91) \boldsymbol{n}_p \cdot\boldsymbol{\sigma}_p^{n+1}\boldsymbol{n}_p + \phi^{n+1} + \beta\boldsymbol{\xi}^{n+1} \cdot\boldsymbol{n}_p = \beta\boldsymbol{\xi}^{*,n+1} \cdot\boldsymbol{n}_p, \tag*{(91)} n p ⋅ σ p n + 1 n p + ϕ n + 1 + β ξ n + 1 ⋅ n p = β ξ ∗ , n + 1 ⋅ n p , (91)
with a fixed Robin parameter β > 0 \beta> 0 β > 0 . Compared with (23) , its weak equation contains two additional terms, namely
β ⟨ ξ h n + 1 ⋅ n p , ζ h ⋅ n p ⟩ Γ \beta\left\langle\boldsymbol{\xi}_h^{n+1} \cdot\boldsymbol{n}_p,\boldsymbol{\zeta}_h \cdot\boldsymbol{n}_p \right\rangle_{\Gamma} β ⟨ ξ h n + 1 ⋅ n p , ζ h ⋅ n p ⟩ Γ
on the left,
β ⟨ ξ h ∗ , n + 1 ⋅ n p , ζ h ⋅ n p ⟩ Γ \beta\left\langle\boldsymbol{\xi}_h^{*,n+1} \cdot\boldsymbol{n}_p,\boldsymbol{\zeta}_h \cdot\boldsymbol{n}_p \right\rangle_{\Gamma} β ⟨ ξ h ∗ , n + 1 ⋅ n p , ζ h ⋅ n p ⟩ Γ
on the right. (6.2)
In residual form, the difference is
β ⟨ ( ξ h n + 1 − 2 ξ h n + ξ h n − 1 ) ⋅ n p , ζ h ⋅ n p ⟩ Γ . (92) \beta\left\langle\left(\boldsymbol{\xi}_h^{n+1} - 2\boldsymbol{\xi}_h^n + \boldsymbol{\xi}_h^{n-1}\right) \cdot\boldsymbol{n}_p,\boldsymbol{\zeta}_h \cdot\boldsymbol{n}_p \right\rangle_{\Gamma}. \tag*{(92)} β ⟨ ( ξ h n + 1 − 2 ξ h n + ξ h n − 1 ) ⋅ n p , ζ h ⋅ n p ⟩ Γ . (92)
For a smooth exact solution this is O ( Δ t 2 ) O(\Delta t^2) O ( Δ t 2 ) , so removing it preserves formal second-order consistency. It nevertheless changes the discrete method. Thus the two schemes differ only by the normal structure-velocity residual (6.3), with r ξ n + 1 = ξ h n + 1 − 2 ξ h n + ξ h n − 1 \boldsymbol{r}_{\xi}^{n+1} = \boldsymbol{\xi}_h^{n+1} - 2\boldsymbol{\xi}_h^n + \boldsymbol{\xi}_h^{n-1} r ξ n + 1 = ξ h n + 1 − 2 ξ h n + ξ h n − 1 . The additional normal-velocity term contributes
− β ⟨ r ξ n + 1 ⋅ n p , ξ h n + 1 ⋅ n p ⟩ Γ (93) -\beta\left\langle\boldsymbol{r}_{\xi}^{n+1} \cdot\boldsymbol{n}_p,\boldsymbol{\xi}_h^{n+1} \cdot\boldsymbol{n}_p \right\rangle_{\Gamma} \tag*{(93)} − β ⟨ r ξ n + 1 ⋅ n p , ξ h n + 1 ⋅ n p ⟩ Γ (93)
to the energy identity. If both interface traces are estimated through bulk L 2 L^2 L 2 norms, each inverse trace inequality contributes a factor h − 1 / 2 h^{-1/2} h − 1/2 . Cauchy–Schwarz and Young’s inequality then give
β ∣ ⟨ r ξ ⋅ n p , ξ h ⋅ n p ⟩ Γ ∣ ≤ C β h − 1 ∥ r ξ ∥ p ∥ ξ h ∥ p ≤ ρ p 32 Δ t ∥ r ξ ∥ p 2 + C β 2 Δ t ρ p h 2 ∥ ξ h ∥ p 2 . (94) \begin{aligned}
\beta\left|\left\langle\boldsymbol{r}_{\xi} \cdot\boldsymbol{n}_p,\boldsymbol{\xi}_h \cdot\boldsymbol{n}_p \right\rangle_{\Gamma}\right| \le C\beta h^{-1}\|\boldsymbol{r}_{\xi}\|_p\|\boldsymbol{\xi}_h\|_p \\
&\le\frac{\rho_p}{32\Delta t}\|\boldsymbol{r}_{\xi}\|_p^2 + C\frac{\beta^2\Delta t}{\rho_p h^2}\|\boldsymbol{\xi}_h\|_p^2. \tag*{(94)}
\end{aligned} β ⟨ r ξ ⋅ n p , ξ h ⋅ n p ⟩ Γ ≤ C β h − 1 ∥ r ξ ∥ p ∥ ξ h ∥ p ≤ 32Δ t ρ p ∥ r ξ ∥ p 2 + C ρ p h 2 β 2 Δ t ∥ ξ h ∥ p 2 . (94)
The first term is absorbed by the BDF2 numerical dissipation. The second is proportional to the current structure kinetic energy, with a coefficient of order Δ t / h 2 \Delta t/h^2 Δ t / h 2 .
To track this contribution, let E R R j \mathcal{E}_{\mathrm{RR}}^j E RR j , N R R j \mathcal{N}_{\mathrm{RR}}^j N RR j , and D ∗ , R R j \mathcal{D}_{*,\mathrm{RR}}^j D ∗ , RR j be the quantities in (28) , (29) , and (34) , evaluated on the Robin–Robin variant retaining (6.2). Including (94) , the total coefficient assigned to ∥ r ξ n + 1 ∥ p 2 \|\boldsymbol{r}_{\xi}^{n+1}\|_p^2 ∥ r ξ n + 1 ∥ p 2 is 2 ρ p / ( 32 Δ t ) 2\rho_p/(32\Delta t) 2 ρ p / ( 32Δ t ) , which still fits within N R R n + 1 / 2 \mathcal{N}_{\mathrm{RR}}^{n+1}/2 N RR n + 1 /2 . Since
E R R n + 1 ≥ ρ p 4 ∥ ξ h n + 1 ∥ p 2 , \mathcal{E}_{\mathrm{RR}}^{n+1} \ge\frac{\rho_p}{4}\|\boldsymbol{\xi}_h^{n+1}\|_p^2, E RR n + 1 ≥ 4 ρ p ∥ ξ h n + 1 ∥ p 2 ,
the remaining term yields, for zero sources and Δ t ≤ C ∗ h \Delta t \le C_*h Δ t ≤ C ∗ h ,
E R R n + 1 − E R R n Δ t + 1 2 N R R n + 1 + D ∗ , R R n + 1 ≤ a R R Δ t h 2 E R R n + 1 . (95) \frac{\mathcal{E}_{\mathrm{RR}}^{n+1} - \mathcal{E}_{\mathrm{RR}}^n}{\Delta t} + \frac{1}{2}\mathcal{N}_{\mathrm{RR}}^{n+1} + \mathcal{D}_{*,\mathrm{RR}}^{n+1} \le a_{\mathrm{RR}}\frac{\Delta t}{h^2}\mathcal{E}_{\mathrm{RR}}^{n+1}. \tag*{(95)} Δ t E RR n + 1 − E RR n + 2 1 N RR n + 1 + D ∗ , RR n + 1 ≤ a RR h 2 Δ t E RR n + 1 . (95)
Here a R R > 0 a_{\mathrm{RR}} > 0 a RR > 0 is independent of h h h and Δ t \Delta t Δ t and contains the fixed factor β 2 / ρ p 2 \beta^2/\rho_p^2 β 2 / ρ p 2 .
Accumulation over time. Set z = a R R Δ t 2 / h 2 z = a_{\mathrm{RR}}\Delta t^2/h^2 z = a RR Δ t 2 / h 2 and suppose z ≤ 1 / 2 z \le1/2 z ≤ 1/2 . Multiplying (95) by Δ t \Delta t Δ t gives
( 1 − z ) E R R n + 1 + Δ t ( 1 2 N R R n + 1 + D ∗ , R R n + 1 ) ≤ E R R n . (96) (1-z)\mathcal{E}_{\mathrm{RR}}^{n+1} + \Delta t\left(\frac{1}{2}\mathcal{N}_{\mathrm{RR}}^{n+1} + \mathcal{D}_{*,\mathrm{RR}}^{n+1}\right) \le\mathcal{E}_{\mathrm{RR}}^n. \tag*{(96)} ( 1 − z ) E RR n + 1 + Δ t ( 2 1 N RR n + 1 + D ∗ , RR n + 1 ) ≤ E RR n . (96)
Table 1. Comparison with the second-order Robin–Robin scheme in [1 ] . The time-step conditions are sufficient analytical bounds for fixed coefficients.
Thus each step can increase the energy by a factor ( 1 − z ) − 1 (1-z)^{-1} ( 1 − z ) − 1 . Iteration, retaining the accumulated nonnegative dissipation, gives
E R R M + Δ t ∑ j = 2 M ( 1 2 N R R j + D ∗ , R R j ) ≤ ( 1 − z ) − ( M − 1 ) E R R 1 ≤ exp ( 2 a R R Δ t h 2 ) E R R 1 , t M ≤ T . (97) \mathcal{E}_{\mathrm{RR}}^{M}+\Delta t\sum_{j=2}^{M}\left(\frac{1}{2}\mathcal{N}_{\mathrm{RR}}^{j}+\mathcal{D}_{*,\mathrm{RR}}^{j}\right)
\leq(1-z)^{-(M-1)}\mathcal{E}_{\mathrm{RR}}^{1}
\leq\exp\left(2a_{\mathrm{RR}}\frac{\Delta t}{h^2}\right)\mathcal{E}_{\mathrm{RR}}^{1},\qquad t_M\leq T. \tag*{(97)} E RR M + Δ t j = 2 ∑ M ( 2 1 N RR j + D ∗ , RR j ) ≤ ( 1 − z ) − ( M − 1 ) E RR 1 ≤ exp ( 2 a RR h 2 Δ t ) E RR 1 , t M ≤ T . (97)
The last inequality uses ( 1 − z ) − 1 ≤ e 2 z (1-z)^{-1}\leq e^{2z} ( 1 − z ) − 1 ≤ e 2 z and ( M − 1 ) Δ t ≤ T (M-1)\Delta t\leq T ( M − 1 ) Δ t ≤ T . The exponential is bounded uniformly under refinement if Δ t / h 2 \Delta t/h^2 Δ t / h 2 remains bounded. This is why this argument gives a parabolic time-step restriction for the earlier scheme.
More precisely, because h ≤ 1 h\leq1 h ≤ 1 , the choice
Δ t ≤ c R R h 2 , 0 < c R R ≤ min { C ∗ , ( 2 a R R ) − 1 / 2 } \Delta t\leq c_{\mathrm{RR}}h^2,\qquad0<c_{\mathrm{RR}}\leq\min\left\{C_*,(2a_{\mathrm{RR}})^{-1/2}\right\} Δ t ≤ c RR h 2 , 0 < c RR ≤ min { C ∗ , ( 2 a RR ) − 1/2 }
ensures both Δ t ≤ C ∗ h \Delta t\leq C_*h Δ t ≤ C ∗ h and z ≤ 1 / 2 z\leq1/2 z ≤ 1/2 . It also bounds the exponential in (97) by exp ( 2 a R R c R R T ) \exp(2a_{\mathrm{RR}}c_{\mathrm{RR}}T) exp ( 2 a RR c RR T ) , independently of the mesh and time step.
In the Neumann–Robin splitting scheme, (93) does not appear, so it removes the growth contribution on the right of (95) . The remaining interface residuals are controlled under Δ t ≤ C ∗ h \Delta t\leq C_*h Δ t ≤ C ∗ h , giving the closed energy estimate of Theorem 4.1.
Numerical experiments
Test 1: spatial and temporal convergence study
We assess the accuracy of the second-order Neumann–Robin splitting scheme using a polynomial manufactured solution. The exact solutions determine the volume sources and satisfy all physical interface conditions without additional interface loads. Separate spatial and temporal refinement studies are performed to examine the corresponding convergence.
Consider the following domains for fluid, poroelastic structure and the interface
Ω f = ( 0 , 1 ) × ( 0 , 1 ) , Ω p = ( 0 , 1 ) × ( − 1 , 0 ) , Γ = ( 0 , 1 ) × { 0 } , \Omega_f=(0,1)\times(0,1),\qquad\Omega_p=(0,1)\times(-1,0),\qquad\Gamma=(0,1)\times\{0\}, Ω f = ( 0 , 1 ) × ( 0 , 1 ) , Ω p = ( 0 , 1 ) × ( − 1 , 0 ) , Γ = ( 0 , 1 ) × { 0 } ,
with outward interface normals n f = ( 0 , − 1 ) T \mathbf{n}_f=(0,-1)^T n f = ( 0 , − 1 ) T and n p = ( 0 , 1 ) T \mathbf{n}_p=(0,1)^T n p = ( 0 , 1 ) T . We set the final time T = 1 T=1 T = 1 and choose the parameters ρ f = ρ p = μ f = μ p = λ p = α = C 0 = γ = L = 1 \rho_f=\rho_p=\mu_f=\mu_p=\lambda_p=\alpha=C_0=\gamma=L=1 ρ f = ρ p = μ f = μ p = λ p = α = C 0 = γ = L = 1 , K = I K=I K = I .
Define the spatial polynomials
F ( x ) = x 2 ( 1 − x ) 2 , A ( y ) = ( 1 + y ) 2 ( 1 − 2 y ) , B ( y ) = y 2 ( 1 + y ) , F(x)=x^2(1-x)^2,\qquad A(y)=(1+y)^2(1-2y),\qquad B(y)=y^2(1+y), F ( x ) = x 2 ( 1 − x ) 2 , A ( y ) = ( 1 + y ) 2 ( 1 − 2 y ) , B ( y ) = y 2 ( 1 + y ) ,
C ( y ) = y − 3 y 3 + 2 y 4 , D ( y ) = y ( 1 + y ) 2 . C(y)=y-3y^3+2y^4,\qquad D(y)=y(1+y)^2. C ( y ) = y − 3 y 3 + 2 y 4 , D ( y ) = y ( 1 + y ) 2 .
Using the stream function
Ψ ( t , x , y ) = 2 t F ( x ) C ( y ) , (98) \Psi(t,x,y)=2tF(x)C(y), \tag*{(98)} Ψ ( t , x , y ) = 2 tF ( x ) C ( y ) , (98)
we prescribe the exact fields
u ( t , x , y ) = 2 t ( F ( x ) C ′ ( y ) , − F ′ ( x ) C ( y ) ) T , p f ( t , x , y ) = t 2 F ( x ) − 4 t F ′ ( x ) , η ( t , x , y ) = t 2 ( F ( x ) A ( y ) , − 1 3 F ′ ( x ) D ( y ) ) T , ξ ( t , x , y ) = 2 t ( F ( x ) A ( y ) , − 1 3 F ′ ( x ) D ( y ) ) T , ϕ ( t , x , y ) = F ( x ) ( B ( y ) + t 2 A ( y ) ) . (99) \begin{aligned}
\mathbf{u}(t,x,y)&=2t\left(F(x)C'(y),-F'(x)C(y)\right)^T,\\
p_f(t,x,y)&=t^2F(x)-4tF'(x),\\
\boldsymbol{\eta}(t,x,y)&=t^2\left(F(x)A(y),-\frac{1}{3}F'(x)D(y)\right)^T,\\
\boldsymbol{\xi}(t,x,y)&=2t\left(F(x)A(y),-\frac{1}{3}F'(x)D(y)\right)^T,\\
\phi(t,x,y)&=F(x)\left(B(y)+t^2A(y)\right). \tag*{(99)}
\end{aligned} u ( t , x , y ) p f ( t , x , y ) η ( t , x , y ) ξ ( t , x , y ) ϕ ( t , x , y ) = 2 t ( F ( x ) C ′ ( y ) , − F ′ ( x ) C ( y ) ) T , = t 2 F ( x ) − 4 t F ′ ( x ) , = t 2 ( F ( x ) A ( y ) , − 3 1 F ′ ( x ) D ( y ) ) T , = 2 t ( F ( x ) A ( y ) , − 3 1 F ′ ( x ) D ( y ) ) T , = F ( x ) ( B ( y ) + t 2 A ( y ) ) . (99)
In particular, the stream-function construction gives u = ( ∂ y Ψ , − ∂ x Ψ ) T \mathbf{u}=(\partial_y\Psi,-\partial_x\Psi)^T u = ( ∂ y Ψ , − ∂ x Ψ ) T and ∇ ⋅ u = 0 \nabla\cdot\mathbf{u}=0 ∇ ⋅ u = 0 . The kinematic relation ∂ t η = ξ \partial_t\boldsymbol{\eta}=\boldsymbol{\xi} ∂ t η = ξ also holds identically. Moreover, the manufactured solutions also ensure the homogeneous exterior conditions as follows
u = 0 on Σ f , η = ξ = 0 , ϕ = 0 on Σ p . \mathbf{u}=0\quad\text{on }\Sigma_f,\qquad\boldsymbol{\eta}=\boldsymbol{\xi}=0,\quad\phi=0\quad\text{on }\Sigma_p. u = 0 on Σ f , η = ξ = 0 , ϕ = 0 on Σ p .
No Dirichlet condition is imposed on the fluid pressure, and its level is determined by the interface traction condition. At the interface Γ \Gamma Γ , the interface traces satisfy that
u ∣ Γ = ξ ∣ Γ = ( 2 t F ( x ) , 0 ) T , η ∣ Γ = ( t 2 F ( x ) , 0 ) T , ϕ ∣ Γ = t 2 F ( x ) , ∂ y ϕ ∣ Γ = 0. (100) \mathbf{u}|_\Gamma=\boldsymbol{\xi}|_\Gamma=(2tF(x),0)^T,\qquad\boldsymbol{\eta}|_\Gamma=(t^2F(x),0)^T,\qquad\phi|_\Gamma=t^2F(x),\quad\partial_y\phi|_\Gamma=0. \tag*{(100)} u ∣ Γ = ξ ∣ Γ = ( 2 tF ( x ) , 0 ) T , η ∣ Γ = ( t 2 F ( x ) , 0 ) T , ϕ ∣ Γ = t 2 F ( x ) , ∂ y ϕ ∣ Γ = 0. (100)
We note that the manufactured solution has polynomial temporal dependence of degree at most two. Consequently, the BDF2 approximation is exact when applied to the corresponding exact solution values. The second-order extrapolation of the pore pressure, however, introduces a nonzero temporal consistency defect of order ( Δ t 2 ) (\Delta t^2) ( Δ t 2 ) . Moreover, the exact normal structure velocity vanishes on the interface, namely ξ ⋅ n p = 0 \boldsymbol{\xi}\cdot\mathbf{n}_p=0 ξ ⋅ n p = 0 . Thus, the present example provides a convergence test for the coupled numerical method, but does not independently test the normal structure-velocity residual removed by the proposed Neumann–Robin modification.
The volume sources are obtained by substituting the exact fields into the continuous equations:
f f = ρ f ∂ t u − ∇ ⋅ σ f ( u , p f ) , f p = ρ p ∂ t ξ − ∇ ⋅ σ p ( η , ϕ ) , g p = C 0 ∂ t ϕ + α ∇ ⋅ ξ − ∇ ⋅ ( K ∇ ϕ ) . \begin{aligned}
\mathbf{f}_f&=\rho_f\partial_t\mathbf{u}-\nabla\cdot\boldsymbol{\sigma}_f(\mathbf{u},p_f),\\
\mathbf{f}_p&=\rho_p\partial_t\boldsymbol{\xi}-\nabla\cdot\boldsymbol{\sigma}_p(\boldsymbol{\eta},\phi),\\
g_p&=C_0\partial_t\phi+\alpha\nabla\cdot\boldsymbol{\xi}-\nabla\cdot(K\nabla\phi).
\end{aligned} f f f p g p = ρ f ∂ t u − ∇ ⋅ σ f ( u , p f ) , = ρ p ∂ t ξ − ∇ ⋅ σ p ( η , ϕ ) , = C 0 ∂ t ϕ + α ∇ ⋅ ξ − ∇ ⋅ ( K ∇ ϕ ) .
And the initial data are
u ( 0 ) = 0 , p f ( 0 ) = 0 , η ( 0 ) = ξ ( 0 ) = 0 , ϕ ( 0 , x , y ) = F ( x ) B ( y ) . \mathbf{u}(0)=0,\qquad p_f(0)=0,\qquad\boldsymbol{\eta}(0)=\boldsymbol{\xi}(0)=0,\qquad\phi(0,x,y)=F(x)B(y). u ( 0 ) = 0 , p f ( 0 ) = 0 , η ( 0 ) = ξ ( 0 ) = 0 , ϕ ( 0 , x , y ) = F ( x ) B ( y ) .
In particular, the initial pore pressure is nonzero in the interior of Ω p \Omega_p Ω p . The two BDF2 starting levels, at t 0 = 0 t_0=0 t 0 = 0 and t 1 = Δ t t_1=\Delta t t 1 = Δ t , are prescribed by projecting the exact fields. We use a joint Robin–Stokes projection for the fluid pair, a common elasticity Ritz projection for displacement and structure velocity, and the Darcy Ritz projection defined in (52) for pore pressure.
In both refinement studies, we use continuous piecewise quadratic (P 2 P_2 P 2 ) elements for the fluid velocity, poroelastic displacement, structure velocity, and pore pressure, and continuous piecewise linear (P 1 P_1 P 1 ) elements for the fluid pressure. Both subdomains are discretized using uniform triangular meshes with n n n intervals in each coordinate direction. We report the mesh spacing as Δ x = 1 / n \Delta x = 1/n Δ x = 1/ n .
Let t j = j Δ t t_j = j\Delta t t j = j Δ t and N = T / Δ t N = T/\Delta t N = T /Δ t . The displacement error is measured in the elastic energy norm:
E η = max 2 ≤ j ≤ N ( 2 μ p ∥ D ( η ( t j ) − η h j ) ∥ L 2 ( Ω p ) 2 + λ p ∥ ∇ ⋅ ( η ( t j ) − η h j ) ∥ L 2 ( Ω p ) 2 ) 1 / 2 . (101) E_\eta= \max_{2 \le j \le N} \left(2\mu_p \lVert D(\eta(t_j)-\eta_h^j)\rVert_{L^2(\Omega_p)}^2 + \lambda_p \lVert\nabla\cdot(\eta(t_j)-\eta_h^j)\rVert_{L^2(\Omega_p)}^2\right)^{1/2}. \tag*{(101)} E η = 2 ≤ j ≤ N max ( 2 μ p ∥ D ( η ( t j ) − η h j ) ∥ L 2 ( Ω p ) 2 + λ p ∥ ∇ ⋅ ( η ( t j ) − η h j ) ∥ L 2 ( Ω p ) 2 ) 1/2 . (101)
Under the homogeneous displacement boundary conditions, this norm is equivalent to the H 1 ( Ω p ) H^1(\Omega_p) H 1 ( Ω p ) norm. For the remaining variables, we use maximum-in-time L 2 L^2 L 2 errors:
E z = max 2 ≤ j ≤ N ∥ z ( t j ) − z h j ∥ L 2 ( Ω ∗ ) , z ∈ { u , p f , ξ , ϕ } , E_z = \max_{2 \le j \le N} \lVert z(t_j)-z_h^j\rVert_{L^2(\Omega_*)}, \qquad z \in\{u,p_f,\xi,\phi\}, E z = 2 ≤ j ≤ N max ∥ z ( t j ) − z h j ∥ L 2 ( Ω ∗ ) , z ∈ { u , p f , ξ , ϕ } ,
where Ω ∗ = Ω f \Omega_* = \Omega_f Ω ∗ = Ω f for u , p f u,p_f u , p f and Ω ∗ = Ω p \Omega_* = \Omega_p Ω ∗ = Ω p for ξ , ϕ \xi,\phi ξ , ϕ . All maxima are taken over the computed BDF2 time levels.
For spatial refinement, we choose Δ t = 10 − 3 \Delta t = 10^{-3} Δ t = 1 0 − 3 and T = 1 T = 1 T = 1 and use n = 4 , 8 , 16 , 32 , 64 , 128 n = 4, 8, 16, 32, 64, 128 n = 4 , 8 , 16 , 32 , 64 , 128 . The observed rates are calculated as
rate = log ( E Δ x / E Δ x / 2 ) log 2 . \operatorname{rate} = \frac{\log(E_{\Delta x}/E_{\Delta x/2})}{\log2}. rate = log 2 log ( E Δ x / E Δ x /2 ) .
Table 2 shows approximately second-order spatial convergence for the displacement in the elastic energy norm and the fluid pressure in the L 2 L^2 L 2 norm. The fluid velocity, structure velocity, and pore pressure exhibit approximately third-order convergence in the L 2 L^2 L 2 norm over most of the refinement range. These observed rates are higher than the O ( h k ) O(h^k) O ( h k ) spatial convergence rate guaranteed by the bulk energy-norm error estimate. For the P 2 P_2 P 2 approximation (k = 2 k = 2 k = 2 ), the theorem establishes an O ( h 2 + Δ t 2 ) O(h^{2}+\Delta t^{2}) O ( h 2 + Δ t 2 ) bound, whereas the numerical results exhibit approximately O ( h 3 ) O(h^{3}) O ( h 3 ) spatial convergence for several variables in the L 2 L^{2} L 2 norm. Such higher-order behavior is consistent with the approximation properties of quadratic finite elements under additional regularity assumptions. However, establishing an optimal L 2 L^{2} L 2 error estimate for the fully discrete splitting scheme would require further analysis of the interface and splitting errors and is beyond the scope of the present work.
For temporal refinement, we fix the mesh at Δ x = 1 / 250 \Delta x = 1/250 Δ x = 1/250 intervals in each coordinate direction, and T = 1 T = 1 T = 1 . The time step is set to be halved successively:
Δ t ∈ { 1 4 , 1 8 , 1 16 , 1 32 , 1 64 } . \Delta t \in\left\{ \frac{1}{4}, \frac{1}{8}, \frac{1}{16}, \frac{1}{32}, \frac{1}{64} \right\}. Δ t ∈ { 4 1 , 8 1 , 16 1 , 32 1 , 64 1 } .
The observed temporal rates are calculated as
rate = log ( E Δ t / E Δ t / 2 ) log 2 . \operatorname{rate} = \frac{\log(E_{\Delta t}/E_{\Delta t/2})}{\log2}. rate = log 2 log ( E Δ t / E Δ t /2 ) .
Table 3: Temporal convergence of the Neumann–Robin (NR, β = 0 \beta= 0 β = 0 ) scheme for the old polynomial manufactured solution with T = 1 T = 1 T = 1 on the fixed mesh n = 250 n = 250 n = 250 , Δ x = 1 / 250 \Delta x = 1/250 Δ x = 1/250 . The displacement error E η E_{\eta} E η is measured in the H 1 H^{1} H 1 -equivalent elastic energy norm, and all other errors are measured in L 2 L^{2} L 2 . All errors are maxima over the computed BDF2 time levels.
Table 3.
Table 3 shows rates close to two for the fluid velocity, structure velocity, and pore pressure throughout the finer temporal refinements. On the finest refinement, the observed convergence rates for E u E_u E u , E ξ E_{\xi} E ξ , and E ϕ E_{\phi} E ϕ are 2.00, 1.99, and 1.99, respectively, supporting the second-order temporal behavior of the BDF2–EX2 discretization for these variables. In contrast, the fluid-pressure convergence rate decreases to 1.78, while the displacement convergence rate decreases to 1.13, suggesting that spatial discretization errors may influence the finest temporal refinements.
To examine these reductions variable by variable, we compare the finest temporal errors with spatial errors estimated from Table 2. Scaling the errors for n = 128 n = 128 n = 128 to n = 250 n = 250 n = 250 with the approximate observed spatial orders, two for η \eta η and p f p_f p f and three for u \mathbf{u} u and ϕ \phi ϕ , gives
E η , s p a c e ( Δ x = 1 / 250 ) ≈ 2.05 × 10 − 5 ( 128 250 ) 2 ≈ 5.37 × 10 − 6 , E_{\eta,\mathrm{space}}(\Delta x = 1/250) \approx2.05 \times10^{-5}\left(\frac{128}{250}\right)^2 \approx5.37 \times10^{-6}, E η , space ( Δ x = 1/250 ) ≈ 2.05 × 1 0 − 5 ( 250 128 ) 2 ≈ 5.37 × 1 0 − 6 ,
E p f , s p a c e ( Δ x = 1 / 250 ) ≈ 6.31 × 10 − 5 ( 128 250 ) 2 ≈ 1.65 × 10 − 5 , E_{p_f,\mathrm{space}}(\Delta x = 1/250) \approx6.31 \times10^{-5}\left(\frac{128}{250}\right)^2 \approx1.65 \times10^{-5}, E p f , space ( Δ x = 1/250 ) ≈ 6.31 × 1 0 − 5 ( 250 128 ) 2 ≈ 1.65 × 1 0 − 5 ,
E u , s p a c e ( Δ x = 1 / 250 ) ≈ 6.76 × 10 − 8 ( 128 250 ) 3 ≈ 9.07 × 10 − 9 , E_{u,\mathrm{space}}(\Delta x = 1/250) \approx6.76 \times10^{-8}\left(\frac{128}{250}\right)^3 \approx9.07 \times10^{-9}, E u , space ( Δ x = 1/250 ) ≈ 6.76 × 1 0 − 8 ( 250 128 ) 3 ≈ 9.07 × 1 0 − 9 ,
E ϕ , s p a c e ( Δ x = 1 / 250 ) ≈ 1.69 × 10 − 8 ( 128 250 ) 3 ≈ 2.27 × 10 − 9 . E_{\phi,\mathrm{space}}(\Delta x = 1/250) \approx1.69 \times10^{-8}\left(\frac{128}{250}\right)^3 \approx2.27 \times10^{-9}. E ϕ , space ( Δ x = 1/250 ) ≈ 1.69 × 1 0 − 8 ( 250 128 ) 3 ≈ 2.27 × 1 0 − 9 .
Since Table 2 was computed with Δ t = 10 − 3 \Delta t = 10^{-3} Δ t = 1 0 − 3 , these values approximate, but do not isolate, the spatial error. The estimated displacement spatial error, 5.37 × 10 − 6 5.37 \times10^{-6} 5.37 × 1 0 − 6 , is close to the finest temporal error, 6.17 × 10 − 6 6.17 \times10^{-6} 6.17 × 1 0 − 6 , supporting the interpretation that spatial error limits the final reduction of the displacement error. Similarly, the finest fluid-pressure error, 3.13 × 10 − 5 3.13 \times10^{-5} 3.13 × 1 0 − 5 , is approximately 1.89 times the estimated spatial error, suggesting that spatial discretization contributes to its reduced convergence rate. In contrast, the finest fluid-velocity error, 2.70 × 10 − 7 2.70 \times10^{-7} 2.70 × 1 0 − 7 , is approximately 30 times the estimated spatial error, while the finest pore-pressure error, 2.41 × 10 − 6 2.41 \times10^{-6} 2.41 × 1 0 − 6 , is approximately 1060 times the corresponding spatial-error estimate. These differences are consistent with the nearly second-order temporal convergence observed for fluid velocity and pore pressure, whose errors do not appear to be limited by the estimated spatial-error levels over the tested range.
Table 2: Spatial convergence of the Neumann–Robin scheme for T = 1 T = 1 T = 1 and fixed Δ t = 10 − 3 \Delta t = 10^{-3} Δ t = 1 0 − 3 . The displacement error E η E_\eta E η is measured in the H 1 H^1 H 1 -equivalent elastic energy norm, and all other errors are measured in L 2 L^2 L 2 . All errors are maxima over the computed BDF2 time levels.
Table 2.
In the temporal study, Δ t / Δ x \Delta t/\Delta x Δ t /Δ x ranges from 3.91 to 62.5. Since the constant C ∗ C_* C ∗ in (32) has not been evaluated for this configuration, we do not claim that these runs satisfy the sufficient condition Δ t ≤ C ∗ h \Delta t \le C_*h Δ t ≤ C ∗ h of Theorem 4.1 . Thus, the observed convergence is empirical and does not test the sharpness of that condition.
Test 2
We compare the second-order Neumann-Robin (NR) splitting scheme with the Robin-Robin (RR) variant in [1 ] using the same nontrivial polynomial manufactured solution. The exact fields satisfy the physical interface conditions without additional interface loads. Unlike Test 1 in 7.1, the present solution has nonzero normal skeleton motion, Darcy filtration, and tangential slip, as well as nonzero temporal consistency defects. Both schemes are tested with the same meshes, forcing terms, starting values, and finite element spaces, and they differ only in their poroelastic mechanical normal condition.
We consider the same domains and parameter values as depicted in 7.1. For the Robin-Robin scheme we introduce a parameter β = 1 \beta= 1 β = 1 to denote the additional normal mechanical component. Then, the Neumann-Robin scheme corresponds to the same mechanical interface equation with β = 0 \beta= 0 β = 0 . Namely, the normal condition used for the comparative tests is
n p ⋅ σ p ( η h n + 1 , ϕ h n + 1 ) n p + ϕ h n + 1 + β ξ h n + 1 ⋅ n p = β ( 2 ξ h n − ξ h n − 1 ) ⋅ n p . \mathbf{n}_p \cdot\boldsymbol{\sigma}_p(\eta_h^{n+1},\phi_h^{n+1})\mathbf{n}_p+\phi_h^{n+1}+\beta\xi_h^{n+1}\cdot\mathbf{n}_p=\beta(2\xi_h^n-\xi_h^{n-1})\cdot\mathbf{n}_p. n p ⋅ σ p ( η h n + 1 , ϕ h n + 1 ) n p + ϕ h n + 1 + β ξ h n + 1 ⋅ n p = β ( 2 ξ h n − ξ h n − 1 ) ⋅ n p .
While all other variational terms are identical in both runs.
Define the time-dependent polynomials
a ( t ) = t + t 4 , s ( t ) = t + t 3 , b ( t ) = 1 + t 3 , (102) a(t)=t+t^4,\qquad s(t)=t+t^3,\qquad b(t)=1+t^3, \tag*{(102)} a ( t ) = t + t 4 , s ( t ) = t + t 3 , b ( t ) = 1 + t 3 , (102)
and the spatial polynomials
F ( x ) = 256 x 4 ( 1 − x ) 4 , A ( y ) = y ( 1 + y ) , B ( y ) = ( 1 + y ) 2 ( 1 − 2 y ) , C ( y ) = 1 − 4 y 3 + 3 y 4 , D ( y ) = y − 3 y 3 + 2 y 4 , H ( y ) = 1 2 y 2 ( 1 − y ) 2 . (103) \begin{aligned}
F(x)&=256x^4(1-x)^4,\qquad A(y)=y(1+y),\qquad B(y)=(1+y)^2(1-2y),\\
C(y)&=1-4y^3+3y^4,\qquad D(y)=y-3y^3+2y^4,\qquad H(y)=\tfrac{1}{2}y^2(1-y)^2. \tag*{(103)}
\end{aligned} F ( x ) C ( y ) = 256 x 4 ( 1 − x ) 4 , A ( y ) = y ( 1 + y ) , B ( y ) = ( 1 + y ) 2 ( 1 − 2 y ) , = 1 − 4 y 3 + 3 y 4 , D ( y ) = y − 3 y 3 + 2 y 4 , H ( y ) = 2 1 y 2 ( 1 − y ) 2 . (103)
Using the stream function
Ψ ( t , x , y ) = s ( t ) F ( x ) C ( y ) + a ( t ) ( F ( x ) + F ′ ( x ) ) D ( y ) + [ a ( t ) ( F ( x ) + F ′ ( x ) ) + s ( t ) F ′ ′ ( x ) ] H ( y ) . (104) \Psi(t,x,y)=s(t)F(x)C(y)+a(t)\left(F(x)+F'(x)\right)D(y)+\left[a(t)\left(F(x)+F'(x)\right)+s(t)F''(x)\right]H(y). \tag*{(104)} Ψ ( t , x , y ) = s ( t ) F ( x ) C ( y ) + a ( t ) ( F ( x ) + F ′ ( x ) ) D ( y ) + [ a ( t ) ( F ( x ) + F ′ ( x ) ) + s ( t ) F ′′ ( x ) ] H ( y ) . (104)
we can prescribe the following exact fields
u ( t , x , y ) = ( ∂ y Ψ , − ∂ x Ψ ) T , p f ( t , x , y ) = b ( t ) F ( x ) − 2 a ( t ) ( F ′ ( x ) + F ′ ′ ( x ) ) , η ( t , x , y ) = a ( t ) ( F ( x ) A ( y ) , F ( x ) B ( y ) ) T , ξ ( t , x , y ) = a ′ ( t ) ( F ( x ) A ( y ) , F ( x ) B ( y ) ) T , ϕ ( t , x , y ) = b ( t ) F ( x ) B ( y ) + [ a ′ ( t ) F ( x ) + s ( t ) F ′ ( x ) ] A ( y ) . (105) \begin{aligned}
u(t,x,y)&=\left(\partial_y\Psi,-\partial_x\Psi\right)^T,\\
p_f(t,x,y)&=b(t)F(x)-2a(t)\left(F'(x)+F''(x)\right),\\
\eta(t,x,y)&=a(t)\left(F(x)A(y),F(x)B(y)\right)^T,\\
\xi(t,x,y)&=a'(t)\left(F(x)A(y),F(x)B(y)\right)^T,\\
\phi(t,x,y)&=b(t)F(x)B(y)+\left[a'(t)F(x)+s(t)F'(x)\right]A(y). \tag*{(105)}
\end{aligned} u ( t , x , y ) p f ( t , x , y ) η ( t , x , y ) ξ ( t , x , y ) ϕ ( t , x , y ) = ( ∂ y Ψ , − ∂ x Ψ ) T , = b ( t ) F ( x ) − 2 a ( t ) ( F ′ ( x ) + F ′′ ( x ) ) , = a ( t ) ( F ( x ) A ( y ) , F ( x ) B ( y ) ) T , = a ′ ( t ) ( F ( x ) A ( y ) , F ( x ) B ( y ) ) T , = b ( t ) F ( x ) B ( y ) + [ a ′ ( t ) F ( x ) + s ( t ) F ′ ( x ) ] A ( y ) . (105)
In particular, the stream-function representation gives fluid velocity divergence free ∇ ⋅ u = 0 \nabla\cdot u=0 ∇ ⋅ u = 0 , and the kinematic relation ∂ t η = ξ \partial_t\eta=\xi ∂ t η = ξ holds identically. The factors in (7.6) also yield the following homogeneous exterior conditions:
u = 0 on Σ f , η = ξ = 0 , ϕ = 0 on Σ p . u=0\quad\text{on }\Sigma_f,\qquad\eta=\xi=0,\qquad\phi=0\quad\text{on }\Sigma_p. u = 0 on Σ f , η = ξ = 0 , ϕ = 0 on Σ p .
At the interface Γ \Gamma Γ , these fields are nonzero and satisfy
u ∣ Γ = ( a ( t ) ( F + F ′ ) , − s ( t ) F ′ ) T , ξ ∣ Γ = ( 0 , a ′ ( t ) F ) T , η ∣ Γ = ( 0 , a ( t ) F ) T , ϕ ∣ Γ = b ( t ) F , ∂ y ϕ ∣ Γ = a ′ ( t ) F + s ( t ) F ′ , u p ∣ Γ ⋅ n p = − a ′ ( t ) F − s ( t ) F ′ , (106) \begin{aligned}
u|_\Gamma&=\left(a(t)(F+F'),-s(t)F'\right)^T,\qquad
\xi|_\Gamma=\left(0,a'(t)F\right)^T,\qquad
\eta|_\Gamma=\left(0,a(t)F\right)^T,\\
\phi|_\Gamma&=b(t)F,\qquad
\partial_y\phi|_\Gamma=a'(t)F+s(t)F',\qquad
u_p|_\Gamma\cdot n_p=-a'(t)F-s(t)F', \tag*{(106)}
\end{aligned} u ∣ Γ ϕ ∣ Γ = ( a ( t ) ( F + F ′ ) , − s ( t ) F ′ ) T , ξ ∣ Γ = ( 0 , a ′ ( t ) F ) T , η ∣ Γ = ( 0 , a ( t ) F ) T , = b ( t ) F , ∂ y ϕ ∣ Γ = a ′ ( t ) F + s ( t ) F ′ , u p ∣ Γ ⋅ n p = − a ′ ( t ) F − s ( t ) F ′ , (106)
where F F F , F ′ F' F ′ , and F ′ ′ F'' F ′′ are evaluated at x x x . Therefore, they satisfy the corresponding coupling conditions as listed in (2.3a)-(2.3d):
u ⋅ n f = ( ξ + u p ) ⋅ n f , u p = − K ∇ ϕ . u\cdot n_f=(\xi+u_p)\cdot n_f,\qquad u_p=-K\nabla\phi. u ⋅ n f = ( ξ + u p ) ⋅ n f , u p = − K ∇ ϕ .
( σ f ( u , p f ) ) x y ∣ Γ = ( σ p ( η , ϕ ) ) x y ∣ Γ = a ( t ) ( F + F ′ ) , ( σ f ( u , p f ) ) y y ∣ Γ = ( σ p ( η , ϕ ) ) y y ∣ Γ = − b ( t ) F = − ϕ ∣ Γ . (107) \begin{aligned}
\left(\sigma_f(u,p_f)\right)_{xy}|_\Gamma&=\left(\sigma_p(\eta,\phi)\right)_{xy}|_\Gamma=a(t)(F+F'),\\
\left(\sigma_f(u,p_f)\right)_{yy}|_\Gamma&=\left(\sigma_p(\eta,\phi)\right)_{yy}|_\Gamma=-b(t)F=-\phi|_\Gamma. \tag*{(107)}
\end{aligned} ( σ f ( u , p f ) ) x y ∣ Γ ( σ f ( u , p f ) ) y y ∣ Γ = ( σ p ( η , ϕ ) ) x y ∣ Γ = a ( t ) ( F + F ′ ) , = ( σ p ( η , ϕ ) ) y y ∣ Γ = − b ( t ) F = − ϕ ∣ Γ . (107)
In particular, ξ ⋅ n p = a ′ ( t ) F ( x ) ≠ 0 \xi\cdot n_p=a'(t)F(x)\ne0 ξ ⋅ n p = a ′ ( t ) F ( x ) = 0 , and at the exact time levels
( ξ n + 1 − 2 ξ n + ξ n − 1 ) ⋅ n p = 24 t n ( Δ t ) 2 F ( x ) . (108) \left(\xi^{n+1}-2\xi^n+\xi^{n-1}\right)\cdot n_p=24t_n(\Delta t)^2F(x). \tag*{(108)} ( ξ n + 1 − 2 ξ n + ξ n − 1 ) ⋅ n p = 24 t n ( Δ t ) 2 F ( x ) . (108)
Although the exact fields satisfy the continuous physical interface conditions, their substitution into the time-discrete Robin–Robin mechanical condition produces the nonzero second-order extrapolation residual in (7.11). Therefore the normal structure-velocity extrapolation term retained by the Robin–Robin algorithm in [1 ] is nonzero for this benchmark. Moreover, the second-order backward-difference approximation is not exact on these higher-degree temporal polynomials. For example, when applied to the exact structure velocity, the backward differentiation defect will be
D t ξ ( t n + 1 ) − ∂ t ξ ( t n + 1 ) = − 8 ( Δ t ) 2 ( F ( x ) A ( y ) , F ( x ) B ( y ) ) T . D_t\xi(t_{n+1})-\partial_t\xi(t_{n+1})=-8(\Delta t)^2\left(F(x)A(y),F(x)B(y)\right)^T. D t ξ ( t n + 1 ) − ∂ t ξ ( t n + 1 ) = − 8 ( Δ t ) 2 ( F ( x ) A ( y ) , F ( x ) B ( y ) ) T .
Thus this example tests both the interface extrapolation and backward differentiation errors, comparing to 7.1.
The volume sources are obtained by substituting the exact fields into the continuous equations:
f f = ∂ t u − Δ u + ∇ p f , f p = ∂ t ξ − Δ η − 2 ∇ ( ∇ ⋅ η ) + ∇ ϕ , g p = ∂ t ϕ + ∇ ⋅ ξ − Δ ϕ . f_f = \partial_t u - \Delta u + \nabla p_f,\qquad f_p = \partial_t \xi- \Delta\eta- 2\nabla(\nabla\cdot\eta) + \nabla\phi,\qquad g_p = \partial_t \phi+ \nabla\cdot\xi- \Delta\phi. f f = ∂ t u − Δ u + ∇ p f , f p = ∂ t ξ − Δ η − 2∇ ( ∇ ⋅ η ) + ∇ ϕ , g p = ∂ t ϕ + ∇ ⋅ ξ − Δ ϕ .
The initial data are
u ( 0 ) = 0 , p f ( 0 , x , y ) = F ( x ) , η ( 0 ) = 0 , u(0) = 0,\qquad p_f(0,x,y) = F(x),\qquad\eta(0) = 0, u ( 0 ) = 0 , p f ( 0 , x , y ) = F ( x ) , η ( 0 ) = 0 ,
ξ ( 0 , x , y ) = ( F ( x ) A ( y ) , F ( x ) B ( y ) ) T , ϕ ( 0 , x , y ) = F ( x ) ( A ( y ) + B ( y ) ) . \xi(0,x,y) = \left(F(x)A(y),F(x)B(y)\right)^{T},\qquad\phi(0,x,y) = F(x)\left(A(y)+B(y)\right). ξ ( 0 , x , y ) = ( F ( x ) A ( y ) , F ( x ) B ( y ) ) T , ϕ ( 0 , x , y ) = F ( x ) ( A ( y ) + B ( y ) ) .
The two initialization levels, namely t 0 = 0 t_0 = 0 t 0 = 0 and t 1 = Δ t t_1 = \Delta t t 1 = Δ t , are prescribed by projecting the exact fields.
Same as in 7.1, in both refinement studies, we use continuous piecewise quadratic (P 2 P_2 P 2 ) elements for fluid velocity, poroelastic displacement, structure velocity, and pore pressure, and continuous piecewise linear (P 1 P_1 P 1 ) elements for fluid pressure. Both subdomains are discretized using uniform triangular meshes with n n n intervals in each coordinate direction.
Table 4 : Spatial convergence of the Neumann–Robin (NR, β = 0 \beta= 0 β = 0 ) and Robin–Robin (RR, β = 1 \beta= 1 β = 1 ) schemes for T = 1 T = 1 T = 1 and fixed Δ t = 10 − 3 \Delta t = 10^{-3} Δ t = 1 0 − 3 . The displacement error E η E_\eta E η is measured in the H 1 H^1 H 1 -equivalent elastic energy norm, and all other errors are measured in L 2 L^2 L 2 . All errors are maxima over the computed BDF2 time levels.
Table 4. Spatial convergence of the Neumann–Robin and Robin–Robin schemes.
Table 4 shows approximately second-order spatial convergence of displacement in the elastic energy norm and fluid pressure in L 2 L^2 L 2 for both schemes, and approximately third-order L 2 L^2 L 2 convergence of fluid velocity. Structure velocity and pore pressure exhibit rates near three over much of the refinement range. On the finest refinement, the structure-velocity rates decrease to 2.28 for NR and 1.74 for RR, while the pore-pressure rates are 2.91 and 2.88, respectively. As in the preceding manufactured-solution test, the approximately third-order spatial convergence observed for several variables in the L 2 L^2 L 2 norm exceeds the order guaranteed by the bulk energy-norm error theorem. These numerical observations are consistent with the higher approximation order of quadratic finite elements, but they do not constitute a proof of an optimal L 2 L^2 L 2 error bound for either splitting scheme.
Table 5 reports the temporal convergence results on the fixed mesh n = 250 n = 250 n = 250 , with T = 1 T = 1 T = 1 and Δ t = 1 / 4 , 1 / 8 , 1 / 16 , 1 / 32 , 1 / 64 \Delta t = 1/4, 1/8, 1/16, 1/32, 1/64 Δ t = 1/4 , 1/8 , 1/16 , 1/32 , 1/64 . For both schemes, the convergence rates for displacement, structure velocity, and pore pressure approach two as the time step decreases, supporting the second-order temporal accuracy of the BDF2–EX2 discretization for these variables. On the finest refinement, the observed rates for ( E η , E ξ , E ϕ ) (E_\eta, E_\xi, E_\phi) ( E η , E ξ , E ϕ ) are ( 1.95 , 1.96 , 1.97 ) (1.95, 1.96, 1.97) ( 1.95 , 1.96 , 1.97 ) for the NR scheme and ( 1.95 , 1.99 , 1.98 ) (1.95, 1.99, 1.98) ( 1.95 , 1.99 , 1.98 ) for the RR scheme, respectively. The fluid-velocity errors also exhibit nearly second-order convergence, with final observed rates of 1.97 for both schemes. In contrast, the fluid-pressure errors decrease initially but gradually level off near 5.04 × 10 − 3 5.04 \times10^{-3} 5.04 × 1 0 − 3 for the NR scheme and 5.06 × 10 − 3 5.06 \times10^{-3} 5.06 × 1 0 − 3 for the RR scheme. The final observed convergence rates decrease to 0.39 and 0.42, respectively. This reduction in the observed temporal convergence rate of fluid pressure at smaller time steps is consistent with the influence of the fixed spatial discretization error, which becomes dominant as the temporal discretization error decreases.
Table 5. Temporal convergence of the Neumann–Robin (NR, β = 0 \beta= 0 β = 0 ) and Robin–Robin (RR, β = 1 \beta= 1 β = 1 ) schemes for T = 1 T = 1 T = 1 on the fixed mesh n = 250 n = 250 n = 250 , Δ x = 1 / 250 \Delta x = 1/250 Δ x = 1/250 . The displacement error E η E_\eta E η is measured in the H 1 H^1 H 1 -equivalent elastic energy norm, and all other errors are measured in L 2 L^2 L 2 . All errors are maxima over the computed BDF2 time levels. (a) Fluid variables. (b) Poroelastic variables.
In summary, for identical spatial meshes and time steps, the NR scheme consistently yields smaller errors in the structure velocity and pore pressure across all temporal refinement cases, while the RR scheme achieves smaller displacement errors. The fluid-velocity errors are comparable between the two methods. Overall, neither scheme demonstrates a uniform accuracy advantage across all solution variables.
Numerical stability study based on 3D cylindrical pressure-pulse benchmark
We want to investigate the stability of the NR and RR schemes by fixing the mesh and increasing the time step. The theoretical sufficient restrictions have the forms as follows:
Δ t ≤ C N R h , Δ t ≤ C R R h 2 . \Delta t \le C_{\mathrm{NR}}h,\qquad\Delta t \le C_{\mathrm{RR}}h^2. Δ t ≤ C NR h , Δ t ≤ C RR h 2 .
We compare the monolithic, second-order NR, and RR methods in a 3D setting by considering a cylindrical fluid lumen surrounded by a poroelastic wall and subjected to a inlet-pressure pulse. All three methods use the same reference mesh, physical parameters, loading, and exterior boundary conditions. The wall is modeled using small-deformation poroelasticity on a fixed reference geometry: its displacement is computed, but the fluid mesh is not moved.
Figure 1. 3D computational meshes for the cylindrical Stokes–Biot problem. Left: tetrahedral mesh of the fluid lumen Ω f \Omega_f Ω f . Right: tetrahedral mesh of the surrounding poroelastic wall Ω p \Omega_p Ω p , consisting of three radial layers. The fluid and structure meshes share a conforming interface at r = 0.5 r = 0.5 r = 0.5 cm. The fluid mesh contains 20,892 tetrahedra and the wall mesh contains 40,518 tetrahedra, giving 61,410 tetrahedra and 11,668 unique mesh vertices in total. The interface has 4,502 matching triangular facets.
The cylinder axis is aligned with the x x x -direction and r = y 2 + z 2 r = \sqrt{y^2 + z^2} r = y 2 + z 2 . The fluid and poroelastic domains are
Ω f = { ( x , y , z ) : 0 < x < 6 , r < 0.5 } , Ω p = { ( x , y , z ) : 0 < x < 6 , 0.5 < r < 0.6 } , (109) \Omega_f = \{(x,y,z) : 0 < x < 6,\ r < 0.5\}, \qquad\Omega_p = \{(x,y,z) : 0 < x < 6,\ 0.5 < r < 0.6\}, \tag*{(109)} Ω f = {( x , y , z ) : 0 < x < 6 , r < 0.5 } , Ω p = {( x , y , z ) : 0 < x < 6 , 0.5 < r < 0.6 } , (109)
where all lengths are in centimeters. Their common interface is Γ = { ( x , y , z ) : 0 < x < 6 , r = 0.5 } \Gamma= \{(x,y,z) : 0 < x < 6,\ r = 0.5\} Γ = {( x , y , z ) : 0 < x < 6 , r = 0.5 } ; the inlet and outlet are the cross-sections at x = 0 x = 0 x = 0 and x = 6 x = 6 x = 6 , respectively. All quantities are expressed in cgs units. The parameters used in the simulations are given in Table 6 . These values are taken from the numerical model and are not presented as a calibration to experimental vessel data.
Table 6. Physical and numerical parameters of the cylindrical pressure-pulse benchmark.
A spatially uniform inlet pressure is applied at the inlet Γ i n = { x = 0 , r < 0.5 } \Gamma_{\mathrm{in}} = \{x = 0,\ r < 0.5\} Γ in = { x = 0 , r < 0.5 } :
p i n ( t ) = { 13333 2 [ 1 − cos ( 2 π t 0.003 ) ] , 0 ≤ t ≤ 0.003 s , 0 , t > 0.003 s . (110) p_{\mathrm{in}}(t) =
\begin{cases}
\frac{13333}{2}\left[1-\cos\left(\frac{2\pi t}{0.003}\right)\right], & 0 \le t \le0.003\ \mathrm{s},\\
0, & t > 0.003\ \mathrm{s}.
\end{cases} \tag*{(110)} p in ( t ) = { 2 13333 [ 1 − cos ( 0.003 2 π t ) ] , 0 , 0 ≤ t ≤ 0.003 s , t > 0.003 s . (110)
and we have the following boundary conditions:
σ f n f = − p i n ( t ) n f on the fluid inlet , σ f n f = 0 on the fluid outlet , η = ξ = 0 on the two wall-end annuli , σ p n p = 0 on the outer wall r = 0.6 , q ⋅ n p = 0 on all exterior porous boundaries . (111) \begin{aligned}
\boldsymbol{\sigma}_f\boldsymbol{n}_f &= -p_{\mathrm{in}}(t)\boldsymbol{n}_f && \text{on the fluid inlet},\\
\boldsymbol{\sigma}_f\boldsymbol{n}_f &= 0 && \text{on the fluid outlet},\\
\boldsymbol{\eta} &= \boldsymbol{\xi} = 0 && \text{on the two wall-end annuli},\\
\boldsymbol{\sigma}_p\boldsymbol{n}_p &= 0 && \text{on the outer wall } r=0.6,\\
\boldsymbol{q}\cdot\boldsymbol{n}_p &= 0 && \text{on all exterior porous boundaries}. \tag*{(111)}
\end{aligned} σ f n f σ f n f η σ p n p q ⋅ n p = − p in ( t ) n f = 0 = ξ = 0 = 0 = 0 on the fluid inlet , on the fluid outlet , on the two wall-end annuli , on the outer wall r = 0.6 , on all exterior porous boundaries . (111)
Figure 2 shows the overlaid pressure fields at t = 6 t=6 t = 6 , 12 12 12 , 18 18 18 , and 24 m s 24\,\mathrm{ms} 24 ms , obtained from the NR scheme. The fluid pressure is displayed in the cylindrical lumen, while the pore pressure is shown in the surrounding poroelastic wall.
(a) t = 6 t = 6 t = 6 ms
(b) t = 12 t = 12 t = 12 ms
(c) t = 18 t = 18 t = 18 ms
(d) t = 24 t = 24 t = 24 ms
Figure 2. Snapshots of the fluid pressure p f p_f p f and pore pressure ϕ \phi ϕ at (a) t = 6 t = 6 t = 6 ms, (b) t = 12 t = 12 t = 12 ms, (c) t = 18 t = 18 t = 18 ms, and (d) t = 24 t = 24 t = 24 ms, obtained from the NR scheme. The pressure fields are overlaid to illustrate the coupled response of the fluid lumen and surrounding poroelastic wall. Color scales vary between snapshots.
Remark 7.1 (Scope of the stability analysis). The 3D benchmark employs boundary conditions that differ from the homogeneous exterior Dirichlet conditions assumed in the theoretical stability analysis. In particular, the benchmark includes nonhomogeneous fluid traction loading, a traction-free mechanical boundary, and an impermeable exterior porous boundary. Consequently, the stability theorem does not directly apply to this configuration. The numerical experiment is intended to assess the behavior of the proposed Neumann-Robin splitting under these more general boundary conditions, rather than to provide a direct verification of the theoretical time-step restriction.
Numerical stability under time-step refinement. We investigate the stability of the NR and RR splitting schemes by comparing their physical energy evolution for different time-step sizes on the same 3D cylindrical mesh. All simulations use continuous P 2 P_2 P 2 finite elements for the fluid velocity u \boldsymbol{u} u , displacement η \boldsymbol{\eta} η , structure velocity ξ \boldsymbol{\xi} ξ , and pore pressure ϕ \phi ϕ , and continuous P 1 P_1 P 1 elements for the fluid pressure p f p_f p f . The time steps are
Δ t ∈ { 1 × 10 − 5 , 2 × 10 − 5 , 2.5 × 10 − 5 , 5 × 10 − 5 } s , T = 0.024 s . (112) \Delta t \in\left\{1 \times10^{-5},\,2 \times10^{-5},\,2.5 \times10^{-5},\,5 \times10^{-5}\right\}\,\mathrm{s}, \qquad T=0.024\,\mathrm{s}. \tag*{(112)} Δ t ∈ { 1 × 1 0 − 5 , 2 × 1 0 − 5 , 2.5 × 1 0 − 5 , 5 × 1 0 − 5 } s , T = 0.024 s . (112)
Both splitting schemes use identical spatial meshes, physical parameters, loading, and boundary conditions. All initial fields and volume sources are zero. The monolithic solution computed with Δ t = 10 − 5 s \Delta t=10^{-5}\,\mathrm{s} Δ t = 1 0 − 5 s is included as a reference.
Since the inlet-pressure pulse supplies mechanical work during 0 < t < 0.003 s 0<t<0.003\,\mathrm{s} 0 < t < 0.003 s , an increase in energy within this interval does not necessarily indicate numerical instability. The instantaneous inlet power is given by
P i n ( t ) = − ∫ Γ i n p i n ( t ) u ⋅ n f d S . (113) P_{\mathrm{in}}(t)=-\int_{\Gamma_{\mathrm{in}}}p_{\mathrm{in}}(t)\,\boldsymbol{u}\cdot\boldsymbol{n}_f\,\mathrm{d}S. \tag*{(113)} P in ( t ) = − ∫ Γ in p in ( t ) u ⋅ n f d S . (113)
We therefore focus on the energy behavior after the pulse and compare the splitting solutions with the monolithic reference.
As shown in Figure 3 , both NR and RR exhibit rapid energy growth for Δ t = 5 × 10 − 5 \Delta t = 5 \times10^{-5} Δ t = 5 × 1 0 − 5 s, leading to early termination. For Δ t = 2.5 × 10 − 5 \Delta t = 2.5 \times10^{-5} Δ t = 2.5 × 1 0 − 5 s, both schemes complete the simulations but develop substantial nonphysical energy amplification after the inlet-pressure pulse. In contrast, the solution obtained from a smaller step of Δ t = 2 × 10 − 5 \Delta t = 2 \times10^{-5} Δ t = 2 × 1 0 − 5 s remains bounded throughout the simulation. The NR and RR energy curves nearly coincide at each tested time step, indicating no discernible difference in their numerical stability behavior on the present mesh.
Figure 3. Physical energy evolution of the NR and RR schemes for different time-step sizes on the same 3D cylindrical mesh. The dashed black curve represents the small-step monolithic reference solution, and the vertical dotted line marks the end of the inlet-pressure pulse at t = 3 t = 3 t = 3 ms. The two splitting schemes exhibit nearly identical energy behavior for the tested time steps.
These findings highlight the distinction between theoretical sufficient stability conditions and actual numerical stability thresholds. The NR analysis establishes stability under a linear restriction Δ t = O ( h ) \Delta t = O(h) Δ t = O ( h ) , improving upon the parabolic restriction Δ t = O ( h 2 ) \Delta t = O(h^2) Δ t = O ( h 2 ) obtained in the previous RR analysis. However, neither condition is known to be necessary or sharp. The nearly identical numerical behavior suggests that the more restrictive theoretical condition for RR may be conservative and does not necessarily reflect its practical stability limit. Thus, although NR admits a less restrictive rigorous stability guarantee, the present experiments do not demonstrate a corresponding numerical stability advantage over RR.
Conclusion
We have formulated and analyzed a Neumann–Robin modification of the second-order Robin-based splitting in [1 ] . The fluid normal Robin condition, Darcy Robin condition, and tangential Robin/BJS treatment are retained, while the Biot mechanical normal condition is imposed with the current pore pressure and without the auxiliary normal structure-velocity Robin term. The fluid and Biot subproblems remain independent at each time step and can be solved in parallel using extrapolated interface data.
The removal of the extrapolated normal structure-velocity residual permits a fully discrete energy stability estimate under the linear time-step restriction Δ t ≤ C ∗ h \Delta t \le C_*h Δ t ≤ C ∗ h for fixed coefficients satisfying the stated assumptions. The stability result covers both homogeneous and nonhomogeneous source terms. Under appropriate regularity, finite element approximation, and initialization assumptions, we also establish a bulk energy-norm error estimate of order h k + Δ t 2 h^k+\Delta t^2 h k + Δ t 2 for the fluid velocity, structure velocity, pore pressure, and displacement. The error analysis employs a common elasticity projection and a BDF2 summation-by-parts argument to control the pore-pressure projection traction term without requiring normal structure-velocity dissipation.
The manufactured-solution experiments exhibit the expected spatial convergence behavior and temporal convergence trends consistent with second-order accuracy. Reduced convergence rates on the finest temporal refinements are partly attributable to spatial-error saturation. In the 3D cylindrical benchmark, the Neumann-Robin and Robin-Robin schemes exhibit nearly identical numerical stability behavior for the tested time steps. The improvement from a parabolic to a linear mesh-dependent sufficient stability condition is therefore established analytically, while the numerical experiments do not demonstrate a corresponding stability advantage or determine the sharp stability threshold of either scheme.
Acknowledgments
Wang’s research has been supported in part by the National Science Foundation under grant DMS-2247001 and by a Simons Foundation Travel Award.