Introduction
Fluid-structure interaction between an incompressible fluid and a thin elastic wall plays an important role in modeling flow through compliant vessels. In this paper, we consider a viscous incompressible fluid coupled to a Naghdi-type shell with three Cartesian displacement components and three independent infinitesimal rotation components. Unlike the conventional Koiter shell model, the Naghdi shell accounts for membrane, transverse-shear, and bending deformations using only first spatial derivatives of the unknowns. The fluid-Naghdi shell model builds on our earlier work [1 , 2 ] , where we developed a sequential, non-iterative loosely coupled scheme that incorporates shell inertia through a Robin boundary condition for the fluid subproblem. In the present study, we retain the same Naghdi shell formulation but introduce a different interface partitioning strategy that allows the fluid and shell subproblems to be solved concurrently at each time step. Our main focus is the development and analysis of this partitioned scheme on a fixed domain.
Fluid-structure interaction problems are commonly solved using either monolithic or partitioned methods. Monolithic methods solve the fluid and structural unknowns simultaneously, while partitioned methods treat the two subproblems separately, allowing existing fluid and structural solvers to be used. However, partitioned methods can suffer from the added-mass effect, particularly when an incompressible fluid interacts with a thin, light elastic wall. As shown by Causin, Gerbeau, and Nobile [3 ] , conventional explicit Dirichlet-Neumann coupling can become unstable in this setting. Strongly coupled partitioned methods can overcome this difficulty, but they generally require multiple subiterations within each time step, increasing the computational cost. This motivates the development of stable, loosely coupled partitioned schemes that advance the fluid and structural subproblems separately without requiring subiterations. Several approaches have been developed to improve the stability of partitioned schemes through the choice of interface conditions. The kinematically coupled β \beta β -scheme provides a non-iterative method for fluid–thin-wall interaction, including models that account for longitudinal wall displacement [4 ] . For the case β = 1 \beta= 1 β = 1 , Bukač and Muha [5 ] proved unconditional stability and first-order convergence in time for a linear model. Other approaches include stabilized explicit coupling using Nitsche’s method [6 ] and generalized Robin–Neumann schemes for thick elastic structures [7 ] . Robin-Robin coupling has also been analyzed, with stability and error estimates established at the time-semidiscrete level [8 ] and for fully discrete schemes [9 ] . These studies demonstrate how the choice of interface conditions affects the stability and accuracy of partitioned methods.
In this work, we extend the parallelizable Robin-Robin splitting scheme developed by [10 ] for an incompressible fluid interacting with a thick elastic structure to a reduced-dimensional Naghdi shell model. The key feature of the proposed scheme is that both Robin interface conditions depend only on previous-time data, allowing the fluid and shell subproblems to be solved concurrently and in parallel without subiterations. The fluid interacts with the shell through its translational velocity and the corresponding interface force per unit reference area, while the independent shell rotations and angular velocities are governed by the shell equations. At each time step, the Robin conditions use the average of the previous fluid and shell velocities, together with the corresponding average interface loads entering with opposite signs. After the separate subproblems are solved, the updated interface loads are recovered from the Robin conditions and used in the next time step. Although each subproblem is solved implicitly, the interface coupling is explicit. Consequently, the fluid and shell velocities need not match exactly, and the interface forces need not be in exact balance at each time step. The main contribution of this work is the extension of the parallelizable Robin-Robin splitting scheme for fluid thick structure interaction to the fluid-Naghdi shell formulation, together with stability and error analysis that accounts for membrane, transverse-shear, and bending effects, as well as rotational inertia. Our theoretical analysis focuses on the fixed-domain Stokes-Naghdi shell problem. We derive a discrete energy identity and establish unconditional stability for homogeneous external data with any positive time step and fixed positive common Robin coefficient. We also derive stability estimates for nonzero external forces and boundary traction data under appropriate assumptions. Furthermore, we establish first-order temporal consistency and an O ( Δ t ) O(\sqrt{\Delta t}) O ( Δ t ) error estimate in the mechanical energy norm under suitable regularity and initialization assumptions. Finally, we assess the proposed scheme for moving-domain fluid-structure interaction through numerical experiments, comparing its results with those obtained from the kinematically coupled β \beta β scheme and a monolithic solver.
The paper is organized as follows. § 2 introduces the moving-domain fluid-Naghdi shell model. § 3 presents the explicit Robin-Robin splitting scheme and its numerical implementation. § 4 describes the fixed-domain problem and the assumptions used in the later analysis. § 5 , § 6 , and § 7 establish the discrete energy identity, consistency estimates, and temporal error bounds, respectively. Finally, § 8 presents numerical comparisons of the proposed Robin-Robin splitting scheme with the loosely coupled β \beta β scheme and the monolithic method. It also demonstrates the application of the Fluid-Naghdi shell model and the proposed explicit Robin-Robin splitting scheme to simulate pressure wave propagation through a stented artery.
Mathematical model
We consider the interaction between an incompressible Newtonian fluid and a deformable Naghdi shell. The fluid occupies a time-dependent 3D domain, while the shell is described on a fixed 2D parameter domain, with 3D displacement and infinitesimal rotation fields. We assume that the shell midsurface coincides with the fluid-structure interface, neglecting the offset between the physical inner wall and the shell midsurface. The shell is modeled using a linear constitutive law defined on the reference surface, while the fluid domain deforms according to the displacement of the shell midsurface.
Reference domains and shell geometry
Let T > 0 T > 0 T > 0 be the final time. Denote the reference fluid domain by Ω ^ f ⊂ R 3 \widehat{\Omega}_f \subset\mathbb{R}^3 Ω f ⊂ R 3 and its elastic lateral boundary by Γ ^ s \widehat{\Gamma}_s Γ s . The remaining boundary is decomposed into a homogeneous Dirichlet portion and inlet/outlet portions:
∂ Ω ^ f = Γ ^ s ∪ Σ ^ f D ∪ Γ ^ f i n ∪ Γ ^ f o u t . (1) \partial\widehat{\Omega}_f = \widehat{\Gamma}_s \cup\widehat{\Sigma}_f^D \cup\widehat{\Gamma}_f^{\mathrm{in}} \cup\widehat{\Gamma}_f^{\mathrm{out}}. \tag*{(1)} ∂ Ω f = Γ s ∪ Σ f D ∪ Γ f in ∪ Γ f out . (1)
Boundary portions have disjoint relative interiors.
We consider a cylindrical reference shell of axial length L > 0 L > 0 L > 0 , midsurface radius R > 0 R > 0 R > 0 , and thickness ε > 0 \varepsilon> 0 ε > 0 . The reference midsurface Γ ^ s \widehat{\Gamma}_s Γ s is parameterized by the axial coordinate x x x and the circumferential arc-length coordinate y a r c y_{\mathrm{arc}} y arc as
ω = ( 0 , L ) × ( 0 , 2 π R ) , Γ ^ s = φ ( ω ) = φ ( x , y a r c ) = ( x , R cos ( y a r c / R ) , R sin ( y a r c / R ) ) . (2) \omega= (0,L) \times(0,2\pi R), \qquad\widehat{\Gamma}_s = \varphi(\omega) = \varphi(x,y_{\mathrm{arc}}) = (x,R\cos(y_{\mathrm{arc}}/R),R\sin(y_{\mathrm{arc}}/R)). \tag*{(2)} ω = ( 0 , L ) × ( 0 , 2 π R ) , Γ s = φ ( ω ) = φ ( x , y arc ) = ( x , R cos ( y arc / R ) , R sin ( y arc / R )) . (2)
Here, y a r c y_{\mathrm{arc}} y arc denotes the circumferential arc-length coordinate, with periodic identification of the boundaries y a r c = 0 y_{\mathrm{arc}} = 0 y arc = 0 and y a r c = 2 π R y_{\mathrm{arc}} = 2\pi R y arc = 2 π R . The parameterization is defined on the resulting periodic domain, and its inverse is understood locally or on the cylindrical surface with this identification.
Letting θ = y a r c / R \theta= y_{\mathrm{arc}}/R θ = y arc / R , we define the covariant tangent vectors and the unit normal to the reference midsurface as
a 1 = ∂ 1 φ = ( 1 , 0 , 0 ) , a 2 = ∂ 2 φ = ( 0 , − sin θ , cos θ ) , a 3 = a 1 × a 2 = ( 0 , − cos θ , − sin θ ) , (3) \mathbf{a}_1 = \partial_1\varphi= (1,0,0), \qquad\mathbf{a}_2 = \partial_2\varphi= (0,-\sin\theta,\cos\theta), \qquad\mathbf{a}_3 = \mathbf{a}_1 \times\mathbf{a}_2 = (0,-\cos\theta,-\sin\theta), \tag*{(3)} a 1 = ∂ 1 φ = ( 1 , 0 , 0 ) , a 2 = ∂ 2 φ = ( 0 , − sin θ , cos θ ) , a 3 = a 1 × a 2 = ( 0 , − cos θ , − sin θ ) , (3)
where ∂ 1 = ∂ x \partial_1 = \partial_x ∂ 1 = ∂ x and ∂ 2 = ∂ y a r c \partial_2 = \partial_{y_{\mathrm{arc}}} ∂ 2 = ∂ y arc . With this parameterization, a 3 \mathbf{a}_3 a 3 points inward and is opposite to the outward fluid normal on the undeformed lateral boundary. Throughout the paper, fluid tractions are defined using the outward fluid normal n f = − a 3 \mathbf{n}_f = -\mathbf{a}_3 n f = − a 3 . With a α β = a α ⋅ a β a_{\alpha\beta} = \mathbf{a}_\alpha\cdot\mathbf{a}_\beta a α β = a α ⋅ a β , a c = [ a α β ] a_c = [a_{\alpha\beta}] a c = [ a α β ] , a c = a c − 1 a^c = a_c^{-1} a c = a c − 1 , and b α β = a 3 ⋅ ∂ α β φ b_{\alpha\beta} = \mathbf{a}_3 \cdot\partial_{\alpha\beta}\varphi b α β = a 3 ⋅ ∂ α β φ , the reference metric and curvature are
a c = a c = I 2 , [ b α β ] = ( 0 0 0 1 / R ) , d S 0 = d x d y a r c . (4) a_c = a^c = \mathbf{I}_2, \qquad[b_{\alpha\beta}] = \begin{pmatrix} 0 & 0 \\ 0 & 1/R \end{pmatrix}, \qquad\mathrm{d}S_0 = \mathrm{d}x\,\mathrm{d}y_{\mathrm{arc}}. \tag*{(4)} a c = a c = I 2 , [ b α β ] = ( 0 0 0 1/ R ) , d S 0 = d x d y arc . (4)
In particular, no additional factor R R R belongs in reference shell integrals when this arc-length parameter is used.
The shell displacement and infinitesimal rotation, both expressed in Cartesian coordinates, are given by
η = ( η x , η y , η z ) , ω = ( ω 1 , ω 2 , ω 3 ) . (5) \boldsymbol{\eta}=(\eta_x,\eta_y,\eta_z),\qquad\boldsymbol{\omega}=(\omega_1,\omega_2,\omega_3). \tag*{(5)} η = ( η x , η y , η z ) , ω = ( ω 1 , ω 2 , ω 3 ) . (5)
The deformed midsurface is described by
Φ s ( x , y a r c , t ) = φ ( x , y a r c ) + η ( x , y a r c , t ) , Γ s ( t ) = Φ s ( ω , t ) . (6) \Phi_s(x,y_{\mathrm{arc}},t)=\varphi(x,y_{\mathrm{arc}})+\boldsymbol{\eta}(x,y_{\mathrm{arc}},t),\qquad\Gamma_s(t)=\Phi_s(\omega,t). \tag*{(6)} Φ s ( x , y arc , t ) = φ ( x , y arc ) + η ( x , y arc , t ) , Γ s ( t ) = Φ s ( ω , t ) . (6)
where Γ s ( t ) \Gamma_s(t) Γ s ( t ) is the moving fluid-shell interface. Assuming that Φ s \Phi_s Φ s is sufficiently regular and injective, we define the surface Jacobian as
J s ( x , y a r c , t ) = ∣ ∂ 1 Φ s ( x , y a r c , t ) × ∂ 2 Φ s ( x , y a r c , t ) ∣ > 0. (7) J_s(x,y_{\mathrm{arc}},t)=\left\lvert\partial_1\Phi_s(x,y_{\mathrm{arc}},t)\times\partial_2\Phi_s(x,y_{\mathrm{arc}},t)\right\rvert>0. \tag*{(7)} J s ( x , y arc , t ) = ∣ ∂ 1 Φ s ( x , y arc , t ) × ∂ 2 Φ s ( x , y arc , t ) ∣ > 0. (7)
For the moving-domain formulation, we assume that the displacement is sufficiently regular, for example, η ∈ W 1 , ∞ \boldsymbol{\eta}\in W^{1,\infty} η ∈ W 1 , ∞ in space, and that Φ s \Phi_s Φ s is injective with J s J_s J s bounded away from zero. These assumptions ensure a well-defined deformed interface. In contrast, the H 1 H^1 H 1 shell energy space introduced below specifies the weak solution and test spaces but does not, by itself, guarantee an admissible moving fluid domain. Under these assumptions, integration over the moving interface can be expressed on the fixed parameter domain through
∫ Γ s ( t ) h ( x , t ) d S = ∫ ω h ( Φ s ( x , y a r c , t ) , t ) J s ( x , y a r c , t ) d x d y a r c . (8) \int_{\Gamma_s(t)}h(x,t)\,\mathrm{d}S=\int_{\omega}h(\Phi_s(x,y_{\mathrm{arc}},t),t)J_s(x,y_{\mathrm{arc}},t)\,\mathrm{d}x\,\mathrm{d}y_{\mathrm{arc}}. \tag*{(8)} ∫ Γ s ( t ) h ( x , t ) d S = ∫ ω h ( Φ s ( x , y arc , t ) , t ) J s ( x , y arc , t ) d x d y arc . (8)
Arbitrary Lagrangian–Eulerian mapping
To describe the moving fluid domain, we introduce the arbitrary Lagrangian-Eulerian (ALE) mapping from the fixed reference domain Ω ^ f \widehat{\Omega}_f Ω f to the current fluid domain Ω f ( t ) \Omega_f(t) Ω f ( t ) :
φ ^ f ( x ^ , t ) = x ^ + η ^ f ( x ^ , t ) , Ω f ( t ) = φ ^ f ( Ω ^ f , t ) . (9) \widehat{\varphi}_f(\widehat{x},t)=\widehat{x}+\widehat{\boldsymbol{\eta}}_f(\widehat{x},t),\qquad\Omega_f(t)=\widehat{\varphi}_f(\widehat{\Omega}_f,t). \tag*{(9)} φ f ( x , t ) = x + η f ( x , t ) , Ω f ( t ) = φ f ( Ω f , t ) . (9)
where η ^ f \widehat{\boldsymbol{\eta}}_f η f denotes the fluid mesh displacement. We assume that φ ^ f \widehat{\varphi}_f φ f is an orientation-preserving diffeomorphism, with deformation gradient and Jacobian given by
F ^ f = ∇ ^ φ ^ f , J ^ f = det F ^ f > 0. (10) \widehat{\mathbf{F}}_f=\widehat{\nabla}\widehat{\varphi}_f,\qquad\widehat{J}_f=\det\widehat{\mathbf{F}}_f>0. \tag*{(10)} F f = ∇ φ f , J f = det F f > 0. (10)
Here, J ^ f \widehat{J}_f J f is the volume Jacobian of the ALE mapping, which is distinct from the surface Jacobian J s J_s J s defined in (7) . Throughout the paper, we use the vector-gradient convention ( ∇ v ) i j = ∂ j v i (\nabla v)_{ij}=\partial_jv_i ( ∇ v ) ij = ∂ j v i . Under the ALE mapping, the spatial gradient transforms according to
∇ v = ( ∇ ^ v ^ ) F ^ f − 1 , v ^ = v ∘ φ ^ f . (11) \nabla v=(\widehat{\nabla}\widehat{v})\widehat{\mathbf{F}}_f^{-1},\qquad\widehat{v}=v\circ\widehat{\varphi}_f. \tag*{(11)} ∇ v = ( ∇ v ) F f − 1 , v = v ∘ φ f . (11)
where the gradients are evaluated at corresponding points in the current and reference domains.
At the fluid-shell interface, the mesh displacement matches the shell displacement, giving
η ^ f ( φ ( x , y a r c ) , t ) ) = η ( x , y a r c , t ) , φ ^ f ( φ ( x , y a r c ) , t ) ) = Φ s ( x , y a r c , t ) . (12) \widehat{\boldsymbol{\eta}}_f(\varphi(x,y_{\mathrm{arc}),t))=\boldsymbol{\eta}(x,y_{\mathrm{arc}},t),\qquad\widehat{\varphi}_f(\varphi(x,y_{\mathrm{arc}),t))=\Phi_s(x,y_{\mathrm{arc},t).}}} \tag*{(12)} η f ( φ ( x , y arc ) , t )) = η ( x , y arc , t ) , φ f ( φ ( x , y arc ) , t )) = Φ s ( x , y arc , t ) . (12)
The displacement is then extended into the fluid domain to determine the interior mesh motion. For fixed external boundaries, one possible choice is the harmonic extension
{ − Δ ^ η ^ f = 0 in Ω ^ f , η ^ f ∘ φ = η on ω , η ^ f = 0 on ∂ Ω ^ f ∖ Γ ^ s . (13) \begin{cases}
-\widehat{\Delta}\widehat{\boldsymbol{\eta}}_f=0 & \text{in }\widehat{\Omega}_f,\\
\widehat{\boldsymbol{\eta}}_f\circ\varphi=\boldsymbol{\eta} & \text{on }\omega,\\
\widehat{\boldsymbol{\eta}}_f=0 & \text{on }\partial\widehat{\Omega}_f\setminus\widehat{\Gamma}_s. \tag*{(13)}
\end{cases} ⎩ ⎨ ⎧ − Δ η f = 0 η f ∘ φ = η η f = 0 in Ω f , on ω , on ∂ Ω f ∖ Γ s . (13)
Other mesh extension methods may also be used, provided that the interface displacement is preserved and the ALE mapping remains admissible.
The fluid mesh velocity and the pullback of the fluid velocity to the reference domain are defined by
w f ( x , t ) = ∂ t ϕ ^ f ( ϕ ^ f − 1 ( x , t ) , t ) , u ^ ( x ^ , t ) = u ( ϕ ^ f ( x ^ , t ) , t ) . (14) w_f(x,t)=\partial_t\widehat{\phi}_f\left(\widehat{\phi}_f^{-1}(x,t),t\right),\qquad\widehat{u}(\widehat{x},t)=u\left(\widehat{\phi}_f(\widehat{x},t),t\right). \tag*{(14)} w f ( x , t ) = ∂ t ϕ f ( ϕ f − 1 ( x , t ) , t ) , u ( x , t ) = u ( ϕ f ( x , t ) , t ) . (14)
Applying the chain rule gives the ALE time derivative
∂ t u ∣ Ω f = ∂ t u + ( w f ⋅ ∇ ) u . (15) \left.\partial_tu\right|_{\Omega_f}=\partial_tu+(w_f\cdot\nabla)u. \tag*{(15)} ∂ t u ∣ Ω f = ∂ t u + ( w f ⋅ ∇ ) u . (15)
Fluid equations
Let ρ f > 0 \rho_f>0 ρ f > 0 and μ f > 0 \mu_f>0 μ f > 0 denote the fluid density and dynamic viscosity, respectively. The fluid velocity u \mathbf{u} u and pressure p p p satisfy the incompressible Navier-Stokes equations in the ALE formulation:
{ ρ f ( ∂ t u ∣ Ω f + ( ( u − w f ) ⋅ ∇ ) u ) − ∇ ⋅ σ f ( u , p ) = F f in Ω f ( t ) , ∇ ⋅ u = 0 in Ω f ( t ) , (16) \begin{cases}
\rho_f\left(\left.\partial_tu\right|_{\Omega_f}+\left((u-w_f)\cdot\nabla\right)u\right)-\nabla\cdot\sigma_f(u,p)=F_f & \text{in }\Omega_f(t),\\
\nabla\cdot u=0 & \text{in }\Omega_f(t), \tag*{(16)}
\end{cases} { ρ f ( ∂ t u ∣ Ω f + ( ( u − w f ) ⋅ ∇ ) u ) − ∇ ⋅ σ f ( u , p ) = F f ∇ ⋅ u = 0 in Ω f ( t ) , in Ω f ( t ) , (16)
where
σ f ( u , p ) = − p I 3 + 2 μ f D ( u ) , D ( u ) = ∇ u + ( ∇ u ) T 2 . (17) \sigma_f(u,p)=-pI_3+2\mu_f\mathbb{D}(u),\qquad\mathbb{D}(u)=\frac{\nabla u+(\nabla u)^{\mathsf{T}}}{2}. \tag*{(17)} σ f ( u , p ) = − p I 3 + 2 μ f D ( u ) , D ( u ) = 2 ∇ u + ( ∇ u ) T . (17)
Here F f F_f F f is the external force.
Let Σ f D ( t ) \Sigma_f^D(t) Σ f D ( t ) , Γ f i n ( t ) \Gamma_f^{\mathrm{in}}(t) Γ f in ( t ) , and Γ f o u t ( t ) \Gamma_f^{\mathrm{out}}(t) Γ f out ( t ) denote the current Dirichlet, inlet, and outlet boundaries, respectively. We impose the following boundary conditions:
u = 0 on Σ f D ( t ) , σ f ( u , p ) n f = − p i n n f on Γ f i n ( t ) , σ f ( u , p ) n f = − p o u t n f on Γ f o u t ( t ) . (18) \begin{aligned}
u&=0 &&\text{on }\Sigma_f^D(t),\\
\sigma_f(u,p)n_f&=-p_{\mathrm{in}}n_f &&\text{on }\Gamma_f^{\mathrm{in}}(t),\\
\sigma_f(u,p)n_f&=-p_{\mathrm{out}}n_f &&\text{on }\Gamma_f^{\mathrm{out}}(t). \tag*{(18)}
\end{aligned} u σ f ( u , p ) n f σ f ( u , p ) n f = 0 = − p in n f = − p out n f on Σ f D ( t ) , on Γ f in ( t ) , on Γ f out ( t ) . (18)
Here, p i n p_{\mathrm{in}} p in and p o u t p_{\mathrm{out}} p out specify the normal traction at the inlet and outlet, rather than Dirichlet conditions on the fluid pressure. Alternatively, a prescribed inlet velocity may be used in place of the inlet traction condition, with appropriate modifications to the fluid trial and test spaces. In the subsequent analysis, we use the boundary conditions given here.
Naghdi shell equation
In the Naghdi shell model, the infinitesimal rotation ω \omega ω is treated as an independent variable rather than being determined by the displacement gradient. We express the shell strains in the local orthonormal frame ( a 1 , a 2 , a 3 ) (a_1,a_2,a_3) ( a 1 , a 2 , a 3 ) , represented by
Q = ( 1 0 0 0 − sin θ cos θ 0 − cos θ − sin θ ) . (19) \mathbf{Q}=
\begin{pmatrix}
1 & 0 & 0\\
0 & -\sin\theta& \cos\theta\\
0 & -\cos\theta& -\sin\theta
\end{pmatrix}. \tag*{(19)} Q = 1 0 0 0 − sin θ − cos θ 0 cos θ − sin θ . (19)
The displacement and rotation enter the strain formulation through the vectors ∂ i η + a i × ω \partial_i\eta+a_i\times\omega ∂ i η + a i × ω , i = 1 , 2 i=1,2 i = 1 , 2 . Projecting these vectors onto the tangential directions of the reference midsurface gives the membrane-shear strain block
A ( η , ω ) = ( a 1 ⋅ ∂ 1 η a 1 ⋅ ∂ 2 η + ω ⋅ a 3 a 2 ⋅ ∂ 1 η − ω ⋅ a 3 a 2 ⋅ ∂ 2 η ) . (20) A(\eta,\omega)=
\begin{pmatrix}
a_1\cdot\partial_1\eta& a_1\cdot\partial_2\eta+\omega\cdot a_3\\
a_2\cdot\partial_1\eta-\omega\cdot a_3 & a_2\cdot\partial_2\eta
\end{pmatrix}. \tag*{(20)} A ( η , ω ) = ( a 1 ⋅ ∂ 1 η a 2 ⋅ ∂ 1 η − ω ⋅ a 3 a 1 ⋅ ∂ 2 η + ω ⋅ a 3 a 2 ⋅ ∂ 2 η ) . (20)
The transverse shear components are obtained by projecting the same vectors onto the reference normal a 3 \mathbf a_3 a 3 :
e ( η , ω ) T = ( a 3 ⋅ ∂ 1 η + ω ⋅ a 2 a 3 ⋅ ∂ 2 η − ω ⋅ a 1 ) . (21) \mathbf e(\boldsymbol\eta,\boldsymbol\omega)^{T}=\left(\mathbf a_3\cdot\partial_1\boldsymbol\eta+\boldsymbol\omega\cdot\mathbf a_2\quad\mathbf a_3\cdot\partial_2\boldsymbol\eta-\boldsymbol\omega\cdot\mathbf a_1\right). \tag*{(21)} e ( η , ω ) T = ( a 3 ⋅ ∂ 1 η + ω ⋅ a 2 a 3 ⋅ ∂ 2 η − ω ⋅ a 1 ) . (21)
Thus, combining the tangential and transverse shear components, we obtain the membrane-shear strain tensor
E ^ m e m ( η , ω ) = ( A ( η , ω ) e ( η , ω ) T ) ∈ R 3 × 2 . (22) \widehat{\mathbf E}_{\mathrm{mem}}(\boldsymbol\eta,\boldsymbol\omega)=
\begin{pmatrix}
\mathbf A(\boldsymbol\eta,\boldsymbol\omega)\\
\mathbf e(\boldsymbol\eta,\boldsymbol\omega)^{T}
\end{pmatrix}\in\mathbb R^{3\times2}. \tag*{(22)} E mem ( η , ω ) = ( A ( η , ω ) e ( η , ω ) T ) ∈ R 3 × 2 . (22)
The full tangential block and transverse shear components are retained in the constitutive law, accounting for the contributions of the independent rotation and its coupling with the displacement gradients.
For test functions ( v , w ) (\mathbf v,\mathbf w) ( v , w ) , the corresponding membrane-shear strain components are defined in the same way:
D ( v , w ) = ( a 1 ⋅ ∂ 1 v a 1 ⋅ ∂ 2 v + w ⋅ a 3 a 2 ⋅ ∂ 1 v − w ⋅ a 3 a 2 ⋅ ∂ 2 v ) , (23) \mathbf D(\mathbf v,\mathbf w)=
\begin{pmatrix}
\mathbf a_1\cdot\partial_1\mathbf v & \mathbf a_1\cdot\partial_2\mathbf v+\mathbf w\cdot\mathbf a_3\\
\mathbf a_2\cdot\partial_1\mathbf v-\mathbf w\cdot\mathbf a_3 & \mathbf a_2\cdot\partial_2\mathbf v
\end{pmatrix}, \tag*{(23)} D ( v , w ) = ( a 1 ⋅ ∂ 1 v a 2 ⋅ ∂ 1 v − w ⋅ a 3 a 1 ⋅ ∂ 2 v + w ⋅ a 3 a 2 ⋅ ∂ 2 v ) , (23)
d ( v , w ) T = ( a 3 ⋅ ∂ 1 v + w ⋅ a 2 a 3 ⋅ ∂ 2 v − w ⋅ a 1 ) , D ^ m e m ( v , w ) = ( D ( v , w ) d ( v , w ) T ) . (24) \mathbf d(\mathbf v,\mathbf w)^{T}=\left(\mathbf a_3\cdot\partial_1\mathbf v+\mathbf w\cdot\mathbf a_2\quad\mathbf a_3\cdot\partial_2\mathbf v-\mathbf w\cdot\mathbf a_1\right),\qquad
\widehat{\mathbf D}_{\mathrm{mem}}(\mathbf v,\mathbf w)=
\begin{pmatrix}
\mathbf D(\mathbf v,\mathbf w)\\
\mathbf d(\mathbf v,\mathbf w)^{T}
\end{pmatrix}. \tag*{(24)} d ( v , w ) T = ( a 3 ⋅ ∂ 1 v + w ⋅ a 2 a 3 ⋅ ∂ 2 v − w ⋅ a 1 ) , D mem ( v , w ) = ( D ( v , w ) d ( v , w ) T ) . (24)
The bending strain and its corresponding test strain are obtained by projecting the derivatives of the rotation and test functions onto the local reference frame:
E ^ b e n d ( ω ) = ( a 1 ⋅ ∂ 1 ω a 1 ⋅ ∂ 2 ω a 2 ⋅ ∂ 1 ω a 2 ⋅ ∂ 2 ω a 3 ⋅ ∂ 1 ω a 3 ⋅ ∂ 2 ω ) , D ^ b e n d ( w ) = ( a 1 ⋅ ∂ 1 w a 1 ⋅ ∂ 2 w a 2 ⋅ ∂ 1 w a 2 ⋅ ∂ 2 w a 3 ⋅ ∂ 1 w a 3 ⋅ ∂ 2 w ) . (25) \widehat{\mathbf E}_{\mathrm{bend}}(\boldsymbol\omega)=
\begin{pmatrix}
\mathbf a_1\cdot\partial_1\boldsymbol\omega& \mathbf a_1\cdot\partial_2\boldsymbol\omega\\
\mathbf a_2\cdot\partial_1\boldsymbol\omega& \mathbf a_2\cdot\partial_2\boldsymbol\omega\\
\mathbf a_3\cdot\partial_1\boldsymbol\omega& \mathbf a_3\cdot\partial_2\boldsymbol\omega
\end{pmatrix},
\qquad
\widehat{\mathbf D}_{\mathrm{bend}}(\mathbf w)=
\begin{pmatrix}
\mathbf a_1\cdot\partial_1\mathbf w & \mathbf a_1\cdot\partial_2\mathbf w\\
\mathbf a_2\cdot\partial_1\mathbf w & \mathbf a_2\cdot\partial_2\mathbf w\\
\mathbf a_3\cdot\partial_1\mathbf w & \mathbf a_3\cdot\partial_2\mathbf w
\end{pmatrix}. \tag*{(25)} E bend ( ω ) = a 1 ⋅ ∂ 1 ω a 2 ⋅ ∂ 1 ω a 3 ⋅ ∂ 1 ω a 1 ⋅ ∂ 2 ω a 2 ⋅ ∂ 2 ω a 3 ⋅ ∂ 2 ω , D bend ( w ) = a 1 ⋅ ∂ 1 w a 2 ⋅ ∂ 1 w a 3 ⋅ ∂ 1 w a 1 ⋅ ∂ 2 w a 2 ⋅ ∂ 2 w a 3 ⋅ ∂ 2 w . (25)
For use in the constitutive law, we separate the tangential blocks and normal rows:
E ^ b e n d ( ω ) = ( E b ( ω ) e b ( ω ) T ) , D ^ b e n d ( w ) = ( D b ( w ) d b ( w ) T ) . (26) \widehat{\mathbf E}_{\mathrm{bend}}(\boldsymbol\omega)=
\begin{pmatrix}
\mathbf E_b(\boldsymbol\omega)\\
\mathbf e_b(\boldsymbol\omega)^{T}
\end{pmatrix},
\qquad
\widehat{\mathbf D}_{\mathrm{bend}}(\mathbf w)=
\begin{pmatrix}
\mathbf D_b(\mathbf w)\\
\mathbf d_b(\mathbf w)^{T}
\end{pmatrix}. \tag*{(26)} E bend ( ω ) = ( E b ( ω ) e b ( ω ) T ) , D bend ( w ) = ( D b ( w ) d b ( w ) T ) . (26)
where E b , D b ∈ R 2 × 2 \mathbf E_b,\mathbf D_b\in\mathbb R^{2\times2} E b , D b ∈ R 2 × 2 and e b , d b ∈ R 2 \mathbf e_b,\mathbf d_b\in\mathbb R^2 e b , d b ∈ R 2 . Since the rotation field is expressed in Cartesian coordinates, its derivatives are taken before projection onto the local reference frame. This distinction is important because differentiating the local rotation components directly would require additional terms arising from variations of the reference frame.
Let C m \mathcal C_m C m and C b \mathcal C_b C b denote the elasticity tensors defined on the reference midsurface. Using the matrix contraction A : B = ∑ i , j A i j B i j \mathbf A:\mathbf B=\sum_{i,j}A_{ij}B_{ij} A : B = ∑ i , j A ij B ij , we define
B m ( ( η , ω ) , ( v , w ) ) = ε ∫ ω C m E ^ m e m ( η , ω ) : D ^ m e m ( v , w ) d x d y a r c , (27) B_m((\boldsymbol\eta,\boldsymbol\omega),(\mathbf v,\mathbf w))
=\varepsilon\int_{\omega}\mathcal C_m\widehat{\mathbf E}_{\mathrm{mem}}(\boldsymbol\eta,\boldsymbol\omega):\widehat{\mathbf D}_{\mathrm{mem}}(\mathbf v,\mathbf w)\,dx\,dy_{\mathrm{arc}}, \tag*{(27)} B m (( η , ω ) , ( v , w )) = ε ∫ ω C m E mem ( η , ω ) : D mem ( v , w ) d x d y arc , (27)
B b ( ( η , ω ) , ( v , w ) ) = ε 3 12 ∫ ω C b E ^ b e n d ( ω ) : D ^ b e n d ( w ) d x d y a r c . (28) B_b((\boldsymbol\eta,\boldsymbol\omega),(\mathbf v,\mathbf w))
=\frac{\varepsilon^3}{12}\int_{\omega}\mathcal C_b\widehat{\mathbf E}_{\mathrm{bend}}(\boldsymbol\omega):\widehat{\mathbf D}_{\mathrm{bend}}(\mathbf w)\,dx\,dy_{\mathrm{arc}}. \tag*{(28)} B b (( η , ω ) , ( v , w )) = 12 ε 3 ∫ ω C b E bend ( ω ) : D bend ( w ) d x d y arc . (28)
For the cylindrical shell, let λ \lambda λ and μ \mu μ denote the Lamé parameters of the shell material. We assume an isotropic material with Young’s modulus E s > 0 E_s>0 E s > 0 and Poisson’s ratio − 1 < ν s ≤ 0.5 -1<\nu_s\le0.5 − 1 < ν s ≤ 0.5 . The reduced elasticity coefficients are
2 λ μ λ + 2 μ = E s ν s 1 − ν s 2 , 2 μ = E s 1 + ν s . (29) \frac{2\lambda\mu}{\lambda+2\mu}=\frac{E_s\nu_s}{1-\nu_s^2},\qquad2\mu=\frac{E_s}{1+\nu_s}. \tag*{(29)} λ + 2 μ 2 λ μ = 1 − ν s 2 E s ν s , 2 μ = 1 + ν s E s . (29)
At ν s = 0.5 \nu_s = 0.5 ν s = 0.5 , these coefficients are defined by continuity and remain finite, so the singular 3D Lamé parameter λ \lambda λ need not be evaluated. The membrane-shear constitutive law is given by
C m E ^ m e m ( η , ω ) : D ^ m e m ( v , w ) = 2 λ μ λ + 2 μ tr A ( η , ω ) tr D ( v , w ) + 2 μ A ( η , ω ) : D ( v , w ) + B m e ( η , ω ) ⋅ d ( v , w ) , (30) C_m\widehat{\mathbf{E}}_{\mathrm{mem}}(\boldsymbol{\eta},\boldsymbol{\omega}) : \widehat{\mathbf{D}}_{\mathrm{mem}}(\mathbf{v},\mathbf{w}) = \frac{2\lambda\mu}{\lambda+ 2\mu} \operatorname{tr}\mathbf{A}(\boldsymbol{\eta},\boldsymbol{\omega})\operatorname{tr}\mathbf{D}(\mathbf{v},\mathbf{w}) + 2\mu\mathbf{A}(\boldsymbol{\eta},\boldsymbol{\omega}) : \mathbf{D}(\mathbf{v},\mathbf{w}) + \mathbf{B}_m\mathbf{e}(\boldsymbol{\eta},\boldsymbol{\omega}) \cdot\mathbf{d}(\mathbf{v},\mathbf{w}), \tag*{(30)} C m E mem ( η , ω ) : D mem ( v , w ) = λ + 2 μ 2 λ μ tr A ( η , ω ) tr D ( v , w ) + 2 μ A ( η , ω ) : D ( v , w ) + B m e ( η , ω ) ⋅ d ( v , w ) , (30)
where B m = μ I 2 \mathbf{B}_m = \mu\mathbf{I}_2 B m = μ I 2 is the transverse shear coefficient for the arc-length parameterized cylinder.
For the bending constitutive law, we introduce
J = ( 0 − 1 1 0 ) . (31) \mathbf{J} = \begin{pmatrix} 0 & -1 \\ 1 & 0 \end{pmatrix}. \tag*{(31)} J = ( 0 1 − 1 0 ) . (31)
The bending contribution is then defined by
C b E ^ b e n d ( ω ) : D ^ b e n d ( w ) = 2 λ μ λ + 2 μ tr ( J E b ( ω ) ) tr ( J D b ( w ) ) + 2 μ ( J E b ( ω ) ) : ( J D b ( w ) ) + μ e b ( ω ) ⋅ d b ( w ) . (32) C_b\widehat{\mathbf{E}}_{\mathrm{bend}}(\boldsymbol{\omega}) : \widehat{\mathbf{D}}_{\mathrm{bend}}(\mathbf{w}) = \frac{2\lambda\mu}{\lambda+ 2\mu} \operatorname{tr}\left(\mathbf{J}\mathbf{E}_b(\boldsymbol{\omega})\right)\operatorname{tr}\left(\mathbf{J}\mathbf{D}_b(\mathbf{w})\right) + 2\mu\left(\mathbf{J}\mathbf{E}_b(\boldsymbol{\omega})\right) : \left(\mathbf{J}\mathbf{D}_b(\mathbf{w})\right) + \mu\mathbf{e}_b(\boldsymbol{\omega}) \cdot\mathbf{d}_b(\mathbf{w}). \tag*{(32)} C b E bend ( ω ) : D bend ( w ) = λ + 2 μ 2 λ μ tr ( J E b ( ω ) ) tr ( J D b ( w ) ) + 2 μ ( J E b ( ω ) ) : ( J D b ( w ) ) + μ e b ( ω ) ⋅ d b ( w ) . (32)
This formulation includes both the tangential and normal components of the rotation derivatives, with the normal-row coefficient B b = μ I 2 \mathbf{B}_b = \mu\mathbf{I}_2 B b = μ I 2 . The bending thickness factor ε 3 / 12 \varepsilon^3/12 ε 3 /12 is included separately in (28) . Both constitutive laws use the reduced coefficients in (29) , including their limiting values at ν s = 0.5 \nu_s = 0.5 ν s = 0.5 . The stability and error analysis below relies on the symmetry and nonnegativity of the resulting bilinear forms. We also note that replacing J \mathbf{J} J by − J -\mathbf{J} − J does not change the bending constitutive law, since J \mathbf{J} J appears in both factors of each tangential term.
We assume that the bilinear forms B m B_m B m and B b B_b B b are continuous, time-independent, symmetric, and nonnegative:
B m ( ( η , ω ) , ( v , w ) ) = B m ( ( v , w ) , ( η , ω ) ) , B m ( ( η , ω ) , ( η , ω ) ) ≥ 0 , B b ( ( η , ω ) , ( v , w ) ) = B b ( ( v , w ) , ( η , ω ) ) , B b ( ( η , ω ) , ( η , ω ) ) ≥ 0. (33) \begin{aligned}
B_m((\boldsymbol{\eta},\boldsymbol{\omega}),(\mathbf{v},\mathbf{w})) &= B_m((\mathbf{v},\mathbf{w}),(\boldsymbol{\eta},\boldsymbol{\omega})), \qquad B_m((\boldsymbol{\eta},\boldsymbol{\omega}),(\boldsymbol{\eta},\boldsymbol{\omega})) \ge0,\\
B_b((\boldsymbol{\eta},\boldsymbol{\omega}),(\mathbf{v},\mathbf{w})) &= B_b((\mathbf{v},\mathbf{w}),(\boldsymbol{\eta},\boldsymbol{\omega})), \qquad B_b((\boldsymbol{\eta},\boldsymbol{\omega}),(\boldsymbol{\eta},\boldsymbol{\omega})) \ge0. \tag*{(33)}
\end{aligned} B m (( η , ω ) , ( v , w )) B b (( η , ω ) , ( v , w )) = B m (( v , w ) , ( η , ω )) , B m (( η , ω ) , ( η , ω )) ≥ 0 , = B b (( v , w ) , ( η , ω )) , B b (( η , ω ) , ( η , ω )) ≥ 0. (33)
These properties are sufficient for the discrete energy identities derived below. However, controlling the full shell norm requires an additional coercivity estimate on the chosen shell space. Such coercivity does not follow from the positivity of the material coefficients alone.
We impose clamped boundary conditions at both ends of the shell and periodic conditions in the circumferential direction:
η ( 0 , y a r c , t ) = η ( L , y a r c , t ) = 0 , ω ( 0 , y a r c , t ) = ω ( L , y a r c , t ) = 0 , η ( x , 0 , t ) = η ( x , 2 π R , t ) , ω ( x , 0 , t ) = ω ( x , 2 π R , t ) . (34) \begin{aligned}
\boldsymbol{\eta}(0,y_{\mathrm{arc}},t) &= \boldsymbol{\eta}(L,y_{\mathrm{arc}},t) = 0, \qquad\boldsymbol{\omega}(0,y_{\mathrm{arc}},t) = \boldsymbol{\omega}(L,y_{\mathrm{arc}},t) = 0,\\
\boldsymbol{\eta}(x,0,t) &= \boldsymbol{\eta}(x,2\pi R,t), \qquad\boldsymbol{\omega}(x,0,t) = \boldsymbol{\omega}(x,2\pi R,t). \tag*{(34)}
\end{aligned} η ( 0 , y arc , t ) η ( x , 0 , t ) = η ( L , y arc , t ) = 0 , ω ( 0 , y arc , t ) = ω ( L , y arc , t ) = 0 , = η ( x , 2 π R , t ) , ω ( x , 0 , t ) = ω ( x , 2 π R , t ) . (34)
The corresponding shell space is
V N ( ω ) = { ( v , w ) ∈ H 1 ( ω ; R 3 ) × H 1 ( ω ; R 3 ) : v = w = 0 at x = 0 , L , v , w have periodic circumferential traces } . (35) V_N(\omega) = \left\{(\mathbf{v},\mathbf{w}) \in H^1(\omega;\mathbb{R}^3) \times H^1(\omega;\mathbb{R}^3) : \mathbf{v} = \mathbf{w} = 0 \text{ at } x = 0,L,\ \mathbf{v},\mathbf{w}\text{ have periodic circumferential traces}\right\}. \tag*{(35)} V N ( ω ) = { ( v , w ) ∈ H 1 ( ω ; R 3 ) × H 1 ( ω ; R 3 ) : v = w = 0 at x = 0 , L , v , w have periodic circumferential traces } . (35)
The boundary conditions in V N ( ω ) V_N(\omega) V N ( ω ) are imposed in the trace sense. In particular, periodicity requires matching traces at the circumferential boundaries, without additional conditions on their derivatives.
Let ρ s > 0 \rho_s > 0 ρ s > 0 denote the shell density. We introduce the translational and angular velocities through ∂ t η = η ˙ \partial_t\boldsymbol{\eta} = \dot{\boldsymbol{\eta}} ∂ t η = η ˙ , ∂ t ω = ω ˙ \partial_t\boldsymbol{\omega} = \dot{\boldsymbol{\omega}} ∂ t ω = ω ˙ . The weak form of the shell equations is: find ( η , ω ) (\boldsymbol{\eta},\boldsymbol{\omega}) ( η , ω ) such that, for every ( v , w ) ∈ V N ( ω ) (v,w) \in V_N(\omega) ( v , w ) ∈ V N ( ω ) ,
ε ρ s ∫ ω ∂ t η ˙ ⋅ v d x d y a r c + ε 3 ρ s 12 ∫ ω ∂ t ω ˙ ⋅ w d x d y a r c + B m ( ( η , ω ) , ( v , w ) ) + B b ( ( η , ω ) , ( v , w ) ) = ∫ ω f s r e f ⋅ v d x d y a r c . (36) \varepsilon\rho_s \int_{\omega} \partial_t \dot{\eta} \cdot v \, dx\,dy_{\mathrm{arc}} + \frac{\varepsilon^3 \rho_s}{12} \int_{\omega} \partial_t \dot{\omega} \cdot w \, dx\,dy_{\mathrm{arc}} + B_m((\eta,\omega),(v,w)) + B_b((\eta,\omega),(v,w)) = \int_{\omega} f_s^{\mathrm{ref}} \cdot v \, dx\,dy_{\mathrm{arc}}. \tag*{(36)} ε ρ s ∫ ω ∂ t η ˙ ⋅ v d x d y arc + 12 ε 3 ρ s ∫ ω ∂ t ω ˙ ⋅ w d x d y arc + B m (( η , ω ) , ( v , w )) + B b (( η , ω ) , ( v , w )) = ∫ ω f s ref ⋅ v d x d y arc . (36)
Here, ε ρ s \varepsilon\rho_s ε ρ s and ε 3 ρ s / 12 \varepsilon^3 \rho_s/12 ε 3 ρ s /12 represent the translational mass and rotational inertia per unit reference area, respectively. The term f s r e f f_s^{\mathrm{ref}} f s ref denotes the applied force per unit reference parameter area. No external couple is applied to the rotational equations.
Interface coupling conditions
The fluid and shell are coupled through the kinematic condition
u ( Φ s ( x , y a r c , t ) , t ) = η ˙ ( x , y a r c , t ) , (37) u(\Phi_s(x,y_{\mathrm{arc}},t),t) = \dot{\eta}(x,y_{\mathrm{arc}},t), \tag*{(37)} u ( Φ s ( x , y arc , t ) , t ) = η ˙ ( x , y arc , t ) , (37)
which enforces no-slip and no-penetration at the moving interface. Differentiating (12) with respect to time gives
w f ∘ Φ s = η ˙ , u = w f on Γ s ( t ) . (38) w_f \circ\Phi_s = \dot{\eta}, \qquad u = w_f \quad\text{on } \Gamma_s(t). \tag*{(38)} w f ∘ Φ s = η ˙ , u = w f on Γ s ( t ) . (38)
Thus, the fluid velocity and mesh velocity coincide at the interface. Since ω \omega ω represents an independent infinitesimal shell rotation, no additional condition relating ω ˙ \dot{\omega} ω ˙ to the fluid velocity is imposed.
The fluid traction on the current interface is defined by
t f ( x , t ) = σ f ( u ( x , t ) , p ( x , t ) ) n f ( x , t ) , x ∈ Γ s ( t ) . (39) t_f(x,t) = \sigma_f(u(x,t),p(x,t))n_f(x,t), \qquad x \in\Gamma_s(t). \tag*{(39)} t f ( x , t ) = σ f ( u ( x , t ) , p ( x , t )) n f ( x , t ) , x ∈ Γ s ( t ) . (39)
Using the surface change of variables, we express this traction as a force density on the reference parameter domain:
t ^ f ( x , y a r c , t ) = J s ( x , y a r c , t ) σ f ( u , p ) ( Φ s ( x , y a r c , t ) , t ) n f ( Φ s ( x , y a r c , t ) , t ) . (40) \hat{t}_f(x,y_{\mathrm{arc}},t) = J_s(x,y_{\mathrm{arc}},t)\sigma_f(u,p)(\Phi_s(x,y_{\mathrm{arc}},t),t)n_f(\Phi_s(x,y_{\mathrm{arc}},t),t). \tag*{(40)} t ^ f ( x , y arc , t ) = J s ( x , y arc , t ) σ f ( u , p ) ( Φ s ( x , y arc , t ) , t ) n f ( Φ s ( x , y arc , t ) , t ) . (40)
Here, t ^ f \hat{t}_f t ^ f includes the surface Jacobian J s J_s J s and represents force per unit reference parameter area, rather than simply the pullback of the current traction.
Let f s Γ f_s^\Gamma f s Γ denote the force acting on the shell per unit current surface area. By the balance of interface forces,
f s Γ = − t f , f s r e f = ( f s Γ ∘ Φ s ) J s = − t ^ f . (41) f_s^\Gamma= -t_f, \qquad f_s^{\mathrm{ref}} = (f_s^\Gamma\circ\Phi_s)J_s = -\hat{t}_f. \tag*{(41)} f s Γ = − t f , f s ref = ( f s Γ ∘ Φ s ) J s = − t ^ f . (41)
In the Naghdi shell model, the internal structural forces are represented by the membrane and bending bilinear forms (27) and (28) . Therefore, no separate 3D structural traction tensor is introduced.
For compatible fluid and shell test functions satisfying v ~ ∘ Φ s = v \tilde{v}\circ\Phi_s = v v ~ ∘ Φ s = v , the interface work contributions cancel:
∫ Γ s ( t ) t f ⋅ v ~ d S + ∫ ω f s r e f ⋅ v d x d y a r c = ∫ ω ( t ^ f + f s r e f ) ⋅ v d x d y a r c = 0. (42) \int_{\Gamma_s(t)} t_f \cdot\tilde{v}\,dS + \int_{\omega} f_s^{\mathrm{ref}} \cdot v\,dx\,dy_{\mathrm{arc}} = \int_{\omega}(\hat{t}_f + f_s^{\mathrm{ref}})\cdot v\,dx\,dy_{\mathrm{arc}} = 0. \tag*{(42)} ∫ Γ s ( t ) t f ⋅ v ~ d S + ∫ ω f s ref ⋅ v d x d y arc = ∫ ω ( t ^ f + f s ref ) ⋅ v d x d y arc = 0. (42)
This identity establishes the sign convention and reference surface measure used in the Robin-Robin splitting scheme.
Initial conditions and compatibility conditions
We prescribe the initial fluid velocity, shell displacement, translational velocity, rotation, and angular velocity:
u ( x , 0 ) = u 0 ( x ) , η ( x , y a r c , 0 ) = η 0 ( x , y a r c ) , η ˙ ( x , y a r c , 0 ) = η ˙ 0 ( x , y a r c ) , \boldsymbol{u}(\boldsymbol{x},0)=\boldsymbol{u}^{0}(\boldsymbol{x}),\qquad\boldsymbol{\eta}(x,y_{\mathrm{arc}},0)=\boldsymbol{\eta}^{0}(x,y_{\mathrm{arc}}),\qquad\dot{\boldsymbol{\eta}}(x,y_{\mathrm{arc}},0)=\dot{\boldsymbol{\eta}}^{0}(x,y_{\mathrm{arc}}), u ( x , 0 ) = u 0 ( x ) , η ( x , y arc , 0 ) = η 0 ( x , y arc ) , η ˙ ( x , y arc , 0 ) = η ˙ 0 ( x , y arc ) ,
ω ( x , y a r c , 0 ) = ω 0 ( x , y a r c ) , ω ˙ ( x , y a r c , 0 ) = ω ˙ 0 ( x , y a r c ) . (43) \boldsymbol{\omega}(x,y_{\mathrm{arc}},0)=\boldsymbol{\omega}^{0}(x,y_{\mathrm{arc}}),\qquad\dot{\boldsymbol{\omega}}(x,y_{\mathrm{arc}},0)=\dot{\boldsymbol{\omega}}^{0}(x,y_{\mathrm{arc}}). \tag*{(43)} ω ( x , y arc , 0 ) = ω 0 ( x , y arc ) , ω ˙ ( x , y arc , 0 ) = ω ˙ 0 ( x , y arc ) . (43)
The initial fluid domain is determined by the initial shell displacement through the ALE mapping (9) and the interface condition (2.12. In particular, when the reference configuration is chosen to coincide with the initial configuration, we take η 0 = 0 \boldsymbol{\eta}^{0}=\boldsymbol{0} η 0 = 0 and η ^ f ( ⋅ , 0 ) = 0 \widehat{\boldsymbol{\eta}}_{f}(\cdot,0)=\boldsymbol{0} η f ( ⋅ , 0 ) = 0 .
For sufficiently regular solutions, the initial data must satisfy the incompressibility and kinematic compatibility conditions
∇ ⋅ u 0 = 0 in Ω f ( 0 ) , u 0 ∘ Φ s ( ⋅ , 0 ) = η ˙ 0 on ω . (44) \nabla\cdot\boldsymbol{u}^{0}=0\quad\text{in }\Omega_{f}(0),\qquad\boldsymbol{u}^{0}\circ\boldsymbol{\Phi}_{s}(\cdot,0)=\dot{\boldsymbol{\eta}}^{0}\quad\text{on }\omega. \tag*{(44)} ∇ ⋅ u 0 = 0 in Ω f ( 0 ) , u 0 ∘ Φ s ( ⋅ , 0 ) = η ˙ 0 on ω . (44)
The initial data must also satisfy the prescribed external velocity conditions, clamped shell boundary conditions, and circumferential periodicity. The fluid pressure serves as a Lagrange multiplier enforcing incompressibility and is not prescribed independently as an initial evolution variable. The initial interface loads required by the partitioned scheme are introduced together with the time discretization.
The weak formulation of the coupled system
For each current domain, we define
V f ( t ) = { v ~ ∈ H 1 ( Ω f ( t ) ; R 3 ) : v ~ = 0 on Σ f D ( t ) } , Q f ( t ) = L 2 ( Ω f ( t ) ) . (45) V_{f}(t)=\left\{\widetilde{\boldsymbol{v}}\in H^{1}(\Omega_{f}(t);\mathbb{R}^{3}):\widetilde{\boldsymbol{v}}=\boldsymbol{0}\text{ on }\Sigma_{f}^{D}(t)\right\},\qquad Q_{f}(t)=L^{2}(\Omega_{f}(t)). \tag*{(45)} V f ( t ) = { v ∈ H 1 ( Ω f ( t ) ; R 3 ) : v = 0 on Σ f D ( t ) } , Q f ( t ) = L 2 ( Ω f ( t )) . (45)
For the stated boundary conditions, the inlet and outlet traction boundaries have positive surface measure and fix the pressure level. Accordingly, Q f ( t ) = L 2 ( Ω f ( t ) ) Q_{f}(t)=L^{2}(\Omega_{f}(t)) Q f ( t ) = L 2 ( Ω f ( t )) is used without an additional zero-mean constraint. A different boundary specification that leaves a pressure nullspace would require a corresponding pressure normalization.
Find ( η , ω ) ∈ V N ( ω ) (\boldsymbol{\eta},\boldsymbol{\omega})\in V_{N}(\omega) ( η , ω ) ∈ V N ( ω ) , their velocities, and ( u , p ) ∈ V f ( t ) × Q f ( t ) (\boldsymbol{u},p)\in V_{f}(t)\times Q_{f}(t) ( u , p ) ∈ V f ( t ) × Q f ( t ) , satisfying the geometry, kinematics, and initial data above and u ∘ Φ s = η ˙ \boldsymbol{u}\circ\boldsymbol{\Phi}_{s}=\dot{\boldsymbol{\eta}} u ∘ Φ s = η ˙ , such that for all ( v ~ , q , v , w ) (\widetilde{\boldsymbol{v}},q,\boldsymbol{v},\boldsymbol{w}) ( v , q , v , w ) with ( v ~ , q ) ∈ V f ( t ) × Q f ( t ) (\widetilde{\boldsymbol{v}},q)\in V_{f}(t)\times Q_{f}(t) ( v , q ) ∈ V f ( t ) × Q f ( t ) , ( v , w ) ∈ V N ( ω ) (\boldsymbol{v},\boldsymbol{w})\in V_{N}(\omega) ( v , w ) ∈ V N ( ω ) , and v ~ ∘ Φ s = v \widetilde{\boldsymbol{v}}\circ\boldsymbol{\Phi}_{s}=\boldsymbol{v} v ∘ Φ s = v , one has
ρ f ∫ Ω f ( t ) ∂ t u ∣ Ω ^ f ⋅ v ~ d x + ρ f ∫ Ω f ( t ) [ ( ( u − w f ) ⋅ ∇ ) u ] ⋅ v ~ d x + 2 μ f ∫ Ω f ( t ) D ( u ) : D ( v ~ ) d x − ∫ Ω f ( t ) p ∇ ⋅ v ~ d x + ∫ Ω f ( t ) q ∇ ⋅ u d x + ε ρ s ∫ ω ∂ t η ⋅ v d x d y a r c + ε 3 ρ s 12 ∫ ω ∂ t ω ⋅ w d x d y a r c + B m ( η , ω ) , ( v , w ) + B b ( η , ω ) , ( v , w ) = ∫ Ω f ( t ) F f ⋅ v ~ d x − ∫ Γ f i n ( t ) p i n n f ⋅ v ~ d S t − ∫ Γ f o u t ( t ) p o u t n f ⋅ v ~ d S t . (46) \begin{aligned}
&\rho_{f}\int_{\Omega_{f}(t)}\left.\partial_{t}\boldsymbol{u}\right|_{\widehat{\Omega}_{f}}\cdot\widetilde{\boldsymbol{v}}\,dx+\rho_{f}\int_{\Omega_{f}(t)}[((\boldsymbol{u}-\boldsymbol{w}_{f})\cdot\nabla)\boldsymbol{u}]\cdot\widetilde{\boldsymbol{v}}\,dx+2\mu_{f}\int_{\Omega_{f}(t)}\mathbb{D}(\boldsymbol{u}):\mathbb{D}(\widetilde{\boldsymbol{v}})\,dx\\
&\quad-\int_{\Omega_{f}(t)}p\nabla\cdot\widetilde{\boldsymbol{v}}\,dx+\int_{\Omega_{f}(t)}q\nabla\cdot\boldsymbol{u}\,dx+\varepsilon\rho_{s}\int_{\omega}\partial_{t}\boldsymbol{\eta}\cdot\boldsymbol{v}\,dx\,dy_{\mathrm{arc}}+\frac{\varepsilon^{3}\rho_{s}}{12}\int_{\omega}\partial_{t}\boldsymbol{\omega}\cdot\boldsymbol{w}\,dx\,dy_{\mathrm{arc}}\\
&\quad+B_{m}(\boldsymbol{\eta},\boldsymbol{\omega}),(\boldsymbol{v},\boldsymbol{w})+B_{b}(\boldsymbol{\eta},\boldsymbol{\omega}),(\boldsymbol{v},\boldsymbol{w})\\
&=\int_{\Omega_{f}(t)}\boldsymbol{F}_{f}\cdot\widetilde{\boldsymbol{v}}\,dx-\int_{\Gamma_{f}^{\mathrm{in}}(t)}p_{\mathrm{in}}\boldsymbol{n}_{f}\cdot\widetilde{\boldsymbol{v}}\,dS_{t}-\int_{\Gamma_{f}^{\mathrm{out}}(t)}p_{\mathrm{out}}\boldsymbol{n}_{f}\cdot\widetilde{\boldsymbol{v}}\,dS_{t}. \tag*{(46)}
\end{aligned} ρ f ∫ Ω f ( t ) ∂ t u ∣ Ω f ⋅ v d x + ρ f ∫ Ω f ( t ) [(( u − w f ) ⋅ ∇ ) u ] ⋅ v d x + 2 μ f ∫ Ω f ( t ) D ( u ) : D ( v ) d x − ∫ Ω f ( t ) p ∇ ⋅ v d x + ∫ Ω f ( t ) q ∇ ⋅ u d x + ε ρ s ∫ ω ∂ t η ⋅ v d x d y arc + 12 ε 3 ρ s ∫ ω ∂ t ω ⋅ w d x d y arc + B m ( η , ω ) , ( v , w ) + B b ( η , ω ) , ( v , w ) = ∫ Ω f ( t ) F f ⋅ v d x − ∫ Γ f in ( t ) p in n f ⋅ v d S t − ∫ Γ f out ( t ) p out n f ⋅ v d S t . (46)
This formulation follows by integration by parts in the fluid momentum equation and addition of Naghdi shell equation (36) , while the interface terms vanish by (42) . For lower regularity, boundary traction work is interpreted in the appropriate trace duality. Its representation as an L 2 L^{2} L 2 surface integral requires additional traction regularity, which will be stated explicitly for the splitting analysis.
Remark 2.1 (Scope of the subsequent analysis). The model in this section retains the moving fluid domain and nonlinear fluid convection. Section 3 first states a fully implicit fluid formulation with the physical traction conditions, then describes the implemented lagged skew-convection variant separately. The discrete energy and temporal error proofs will then be developed for the fixed-domain Stokes–Naghdi linearization. Freezing the domain does not suppress the shell displacement or rotation unknowns; it removes their feedback on the fluid geometry. The resulting theorems do not, by themselves, assert stability or convergence for the moving-domain problem.
Explicit Robin–Robin splitting scheme
We now discretize the moving-domain problem (46) in time. During each step, the fluid and shell exchange only data from the previous time level. This allows the two subproblems to be solved independently, without subiterations. The word explicit refers to this interface exchange. In this section, we will first present the fully implicit fluid formulation, then describe the lagged skew-convection form used in the implementation.
Let N ≥ 1 N \ge1 N ≥ 1 be an integer, Δ t = T / N \Delta t = T/N Δ t = T / N , and t n = n Δ t t^n = n\Delta t t n = n Δ t . At time t n t^n t n , given that
u ^ n , p ^ n , η ^ n , η ^ ˙ n , ω ^ n , ω ^ ˙ n , t ^ f n , f s r e f , n (47) \widehat{u}^n,\quad\widehat{p}^n,\quad\widehat{\eta}^n,\quad\dot{\widehat{\eta}}^n,\quad\widehat{\omega}^n,\quad\dot{\widehat{\omega}}^n,\quad\widehat{t}_f^n,\quad f_s^{\mathrm{ref},n} \tag*{(47)} u n , p n , η n , η ˙ n , ω n , ω ˙ n , t f n , f s ref , n (47)
are available, together with the fluid map Φ ^ f n \widehat{\Phi}_f^n Φ f n and mesh displacement η ^ f n \widehat{\eta}_f^n η f n . Here, the two stored loads have separate roles. t ^ f n \widehat{t}_f^n t f n is the reference fluid traction, including the surface Jacobian, and f s r e f , n f_s^{\mathrm{ref},n} f s ref , n is the load applied to the shell. For the continuous solution, action and reaction give f s r e f = − t ^ f f_s^{\mathrm{ref}} = -\widehat{t}_f f s ref = − t f . The discrete update permits a force imbalance as the two subproblems advance independently.
Robin conditions at the interface
The exact interface conditions on the reference parameter domain are
u ∘ Φ s = η ˙ , t ^ f + f s r e f = 0. (48) u \circ\Phi_s = \dot{\eta},\qquad\widehat{t}_f + f_s^{\mathrm{ref}} = 0. \tag*{(48)} u ∘ Φ s = η ˙ , t f + f s ref = 0. (48)
For our splitting scheme, we choose the following two Robin conditions at the interface:
t ^ f n + 1 + γ u ^ n + 1 ∘ φ = γ 2 ( u ^ n ∘ φ + η ^ ˙ n ) + 1 2 ( t ^ f n − f s r e f , n ) , (49) \widehat{t}_f^{n+1} + \gamma\widehat{u}^{n+1} \circ\varphi= \frac{\gamma}{2}\left(\widehat{u}^n \circ\varphi+ \dot{\widehat{\eta}}^n\right) + \frac{1}{2}\left(\widehat{t}_f^n - f_s^{\mathrm{ref},n}\right), \tag*{(49)} t f n + 1 + γ u n + 1 ∘ φ = 2 γ ( u n ∘ φ + η ˙ n ) + 2 1 ( t f n − f s ref , n ) , (49)
f s r e f , n + 1 + γ η ^ ˙ n + 1 = γ 2 ( u ^ n ∘ φ + η ^ ˙ n ) − 1 2 ( t ^ f n − f s r e f , n ) . (50) f_s^{\mathrm{ref},n+1} + \gamma\dot{\widehat{\eta}}^{n+1} = \frac{\gamma}{2}\left(\widehat{u}^n \circ\varphi+ \dot{\widehat{\eta}}^n\right) - \frac{1}{2}\left(\widehat{t}_f^n - f_s^{\mathrm{ref},n}\right). \tag*{(50)} f s ref , n + 1 + γ η ˙ n + 1 = 2 γ ( u n ∘ φ + η ˙ n ) − 2 1 ( t f n − f s ref , n ) . (50)
Here, we choose the same constant γ > 0 \gamma> 0 γ > 0 for both Robin conditions, and γ \gamma γ is an independent coupling parameter. Its units in k g m − 2 s − 1 \mathrm{kg\,m^{-2}\,s^{-1}} kg m − 2 s − 1 , given that multiplying a velocity by γ \gamma γ gives a reference force density.
Both right-hand sides of (49) and (50) use the same average of the previous fluid and shell velocities, while the previous loads enter with opposite signs, following the action-reaction convention. Since all these quantities are already known, neither subproblem needs the new solution of the other. To check the relation with the continuous coupling conditions, set all time levels in (49) –(50) to the same values. The two equations then reduce to
t ^ f + f s r e f + γ ( u ^ ∘ φ − η ˙ ) = 0 , \widehat{t}_f + f_s^{\mathrm{ref}} + \gamma(\widehat{u} \circ\varphi- \dot{\eta}) = 0, t f + f s ref + γ ( u ∘ φ − η ˙ ) = 0 ,
t ^ f + f s r e f − γ ( u ^ ∘ φ − η ˙ ) = 0. (51) \widehat{t}_f + f_s^{\mathrm{ref}} - \gamma(\widehat{u} \circ\varphi- \dot{\eta}) = 0. \tag*{(51)} t f + f s ref − γ ( u ∘ φ − η ˙ ) = 0. (51)
Adding these equations gives force balance; subtracting them gives velocity continuity.
Fluid domain update
In our splitting scheme, the fluid branch needs a domain update before the new shell solution is available. We predict its boundary from the old shell displacement and velocity:
Φ s n + 1 , ∗ ( x , y a r c ) = φ ( x , y a r c ) + η n ( x , y a r c ) + Δ t η ˙ n ( x , y a r c ) . (52) \Phi_s^{n+1,*}(x,y_{\mathrm{arc}})=\varphi(x,y_{\mathrm{arc}})+\eta^n(x,y_{\mathrm{arc}})+\Delta t\,\dot{\eta}^n(x,y_{\mathrm{arc}}). \tag*{(52)} Φ s n + 1 , ∗ ( x , y arc ) = φ ( x , y arc ) + η n ( x , y arc ) + Δ t η ˙ n ( x , y arc ) . (52)
This equation will provide the predicted fluid interface. Next, we extend the predicted boundary displacement into the reference fluid domain by solving the harmonic equation
{ − Δ ^ η ^ f n + 1 = 0 in Ω ^ f , η ^ f n + 1 ∘ φ = η n + Δ t η ˙ n on ω , η ^ f n + 1 = 0 on ∂ Ω ^ f ∖ Γ ^ s . (53) \begin{cases}
-\widehat{\Delta}\widehat{\eta}_f^{n+1}=0 & \text{in }\widehat{\Omega}_f,\\
\widehat{\eta}_f^{n+1}\circ\varphi=\eta^n+\Delta t\,\dot{\eta}^n & \text{on }\omega,\\
\widehat{\eta}_f^{n+1}=0 & \text{on }\widehat{\partial\Omega}_f\setminus\widehat{\Gamma}_s. \tag*{(53)}
\end{cases} ⎩ ⎨ ⎧ − Δ η f n + 1 = 0 η f n + 1 ∘ φ = η n + Δ t η ˙ n η f n + 1 = 0 in Ω f , on ω , on ∂ Ω f ∖ Γ s . (53)
This extension defines the new fluid map, its gradient and domain velocity in reference domain:
φ ^ f n + 1 ( x ^ ) = x ^ + η ^ f n + 1 ( x ^ ) , Ω f n + 1 = φ ^ f n + 1 ( Ω ^ f ) , Γ s n + 1 , ∗ = Φ s n + 1 , ∗ ( ω ) , \widehat{\varphi}_f^{n+1}(\widehat{x})=\widehat{x}+\widehat{\eta}_f^{n+1}(\widehat{x}),\qquad
\Omega_f^{n+1}=\widehat{\varphi}_f^{n+1}(\widehat{\Omega}_f),\qquad
\Gamma_s^{n+1,*}=\Phi_s^{n+1,*}(\omega), φ f n + 1 ( x ) = x + η f n + 1 ( x ) , Ω f n + 1 = φ f n + 1 ( Ω f ) , Γ s n + 1 , ∗ = Φ s n + 1 , ∗ ( ω ) ,
F ^ f n + 1 = ∇ ^ φ ^ f n + 1 , J ^ f n + 1 = det F ^ f n + 1 , w ^ f n + 1 = η ^ f n + 1 − η ^ f n Δ t . (54) \widehat{\mathbf{F}}_f^{n+1}=\widehat{\nabla}\widehat{\varphi}_f^{n+1},\qquad
\widehat{J}_f^{n+1}=\det\widehat{\mathbf{F}}_f^{n+1},\qquad
\widehat{\mathbf{w}}_f^{n+1}=\frac{\widehat{\eta}_f^{n+1}-\widehat{\eta}_f^n}{\Delta t}. \tag*{(54)} F f n + 1 = ∇ φ f n + 1 , J f n + 1 = det F f n + 1 , w f n + 1 = Δ t η f n + 1 − η f n . (54)
Assuming that the resulting maps are invertible and preserve orientation. Their displacement increment gives the mesh velocity
w f n + 1 = w ^ f n + 1 ∘ ( φ ^ f n + 1 ) − 1 . (55) \mathbf{w}_f^{n+1}=\widehat{\mathbf{w}}_f^{n+1}\circ\left(\widehat{\varphi}_f^{n+1}\right)^{-1}. \tag*{(55)} w f n + 1 = w f n + 1 ∘ ( φ f n + 1 ) − 1 . (55)
The fluid interface integrals use the Jacobian of the predicted surface,
J s n + 1 , ∗ = ∣ ∂ 1 Φ s n + 1 , ∗ × ∂ 2 Φ s n + 1 , ∗ ∣ . (56) J_s^{n+1,*}=\left\lvert\partial_1\Phi_s^{n+1,*}\times\partial_2\Phi_s^{n+1,*}\right\rvert. \tag*{(56)} J s n + 1 , ∗ = ∂ 1 Φ s n + 1 , ∗ × ∂ 2 Φ s n + 1 , ∗ . (56)
Noticing that the shell branch will compute its own new surface, which may differ from this prediction. Including the old velocity gives a first-order predictor and a consistent initial mesh velocity when the wall starts with nonzero velocity. The Jacobian in (56) belongs to the fluid prediction, while the new shell surface is φ + η n + 1 \varphi+\eta^{n+1} φ + η n + 1 . From the second step onward, the old fluid surface is also a prediction. Using the predictor and shell kinematics, we therefore obtain, for n ≥ 1 n\geq1 n ≥ 1 ,
w ^ f n + 1 ∘ φ = η n + Δ t η ˙ n − η n − 1 − Δ t η ˙ n − 1 Δ t = 2 η ˙ n − η ˙ n − 1 . (57) \widehat{\mathbf{w}}_f^{n+1}\circ\varphi
=\frac{\eta^n+\Delta t\dot{\eta}^n-\eta^{n-1}-\Delta t\dot{\eta}^{n-1}}{\Delta t}
=2\dot{\eta}^n-\dot{\eta}^{n-1}. \tag*{(57)} w f n + 1 ∘ φ = Δ t η n + Δ t η ˙ n − η n − 1 − Δ t η ˙ n − 1 = 2 η ˙ n − η ˙ n − 1 . (57)
The interface mesh velocity thus involves two shell time levels and generally differs from either the old or new shell velocity. At the first step, the fluid map starts from η 0 \eta^0 η 0 without prediction, so w ^ f 1 ∘ φ = η ˙ 0 \widehat{\mathbf{w}}_f^1\circ\varphi=\dot{\eta}^0 w f 1 ∘ φ = η ˙ 0 .
Remark 3.1 (Alternative geometry prediction). One may instead use the lagged surface Φ s n + 1 , ∗ = φ + η n \Phi_s^{n+1,*}=\varphi+\eta^n Φ s n + 1 , ∗ = φ + η n , which also allows independent fluid and shell solves. We use (52) to include the prescribed initial wall velocity in the first mesh update. Both choices give the same scheme when the fluid domain is fixed, so the fixed-domain proofs below apply to either choice. Their moving-domain behavior requires a separate analysis.
Fluid subproblem on the moving domain
The fluid fields in the physical domain can be related to their reference representations by
u ^ n + 1 = u n + 1 ∘ φ ^ f n + 1 , p ^ n + 1 = p n + 1 ∘ φ ^ f n + 1 . (58) \widehat{u}^{n+1}=u^{n+1}\circ\widehat{\varphi}_f^{n+1},\qquad\widehat{p}^{n+1}=p^{n+1}\circ\widehat{\varphi}_f^{n+1}. \tag*{(58)} u n + 1 = u n + 1 ∘ φ f n + 1 , p n + 1 = p n + 1 ∘ φ f n + 1 . (58)
Both u ^ n \widehat{u}^{n} u n and u ^ n + 1 \widehat{u}^{n+1} u n + 1 are defined on the reference domain. Their backward difference therefore compares velocities at matching reference points. In physical coordinates, the old velocity must first be transported to the new domain. The fully implicit fluid subproblem is
ρ f [ u n + 1 − u ^ n ∘ ( φ ^ f n + 1 ) − 1 Δ t + ( ( u n + 1 − w f n + 1 ) ⋅ ∇ ) u n + 1 ] − ∇ ⋅ σ f ( u n + 1 , p n + 1 ) = F f n + 1 , \rho_f\left[\frac{u^{n+1}-\widehat{u}^{n}\circ(\widehat{\varphi}_f^{n+1})^{-1}}{\Delta t}+\left((u^{n+1}-w_f^{n+1})\cdot\nabla\right)u^{n+1}\right]-\nabla\cdot\boldsymbol{\sigma}_f(u^{n+1},p^{n+1})=F_f^{n+1}, ρ f [ Δ t u n + 1 − u n ∘ ( φ f n + 1 ) − 1 + ( ( u n + 1 − w f n + 1 ) ⋅ ∇ ) u n + 1 ] − ∇ ⋅ σ f ( u n + 1 , p n + 1 ) = F f n + 1 ,
∇ ⋅ u n + 1 = 0 in Ω f n + 1 . (59) \nabla\cdot u^{n+1}=0\qquad\text{in }\Omega_f^{n+1}. \tag*{(59)} ∇ ⋅ u n + 1 = 0 in Ω f n + 1 . (59)
The reference form of the divergence constraint is tr ( ∇ ^ u ^ n + 1 ( F ^ f n + 1 ) − 1 ) = 0 \operatorname{tr}(\widehat{\nabla} \widehat{u}^{n+1}(\widehat{F}_f^{n+1})^{-1})=0 tr ( ∇ u n + 1 ( F f n + 1 ) − 1 ) = 0 . On the predicted interface impose (49) , with the reference Cauchy traction
t ^ f n + 1 = J s n + 1 , ∗ [ σ f ( u n + 1 , p n + 1 ) n f n + 1 ] ∘ Φ s n + 1 , ∗ . (60) \widehat{t}_f^{n+1}=J_s^{n+1,*}\left[\boldsymbol{\sigma}_f(u^{n+1},p^{n+1})n_f^{n+1}\right]\circ\Phi_s^{n+1,*}. \tag*{(60)} t f n + 1 = J s n + 1 , ∗ [ σ f ( u n + 1 , p n + 1 ) n f n + 1 ] ∘ Φ s n + 1 , ∗ . (60)
The other external conditions are
u n + 1 = 0 on Σ f D , n + 1 , u^{n+1}=0\qquad\text{on }\Sigma_f^{D,n+1}, u n + 1 = 0 on Σ f D , n + 1 ,
σ f ( u n + 1 , p n + 1 ) n f n + 1 = − p i n n + 1 n f n + 1 on Γ f i n , n + 1 , \boldsymbol{\sigma}_f(u^{n+1},p^{n+1})n_f^{n+1}=-p_{\mathrm{in}}^{n+1}n_f^{n+1}\qquad\text{on }\Gamma_f^{\mathrm{in},n+1}, σ f ( u n + 1 , p n + 1 ) n f n + 1 = − p in n + 1 n f n + 1 on Γ f in , n + 1 ,
σ f ( u n + 1 , p n + 1 ) n f n + 1 = − p o u t n + 1 n f n + 1 on Γ f o u t , n + 1 . (61) \boldsymbol{\sigma}_f(u^{n+1},p^{n+1})n_f^{n+1}=-p_{\mathrm{out}}^{n+1}n_f^{n+1}\qquad\text{on }\Gamma_f^{\mathrm{out},n+1}. \tag*{(61)} σ f ( u n + 1 , p n + 1 ) n f n + 1 = − p out n + 1 n f n + 1 on Γ f out , n + 1 . (61)
To write the weak problem on the reference domain, we use the stress
σ ^ f n + 1 = − p ^ n + 1 I + μ f [ ∇ ^ u ^ n + 1 ( F ^ f n + 1 ) − 1 + ( F ^ f n + 1 ) − T ( ∇ ^ u ^ n + 1 ) T ] . (62) \widehat{\boldsymbol{\sigma}}_f^{n+1}=-\widehat{p}^{n+1}I+\mu_f\left[\widehat{\nabla}\widehat{u}^{n+1}(\widehat{F}_f^{n+1})^{-1}+(\widehat{F}_f^{n+1})^{-T}(\widehat{\nabla}\widehat{u}^{n+1})^T\right]. \tag*{(62)} σ f n + 1 = − p n + 1 I + μ f [ ∇ u n + 1 ( F f n + 1 ) − 1 + ( F f n + 1 ) − T ( ∇ u n + 1 ) T ] . (62)
Then, the weak fluid problem is:
For all reference test functions v ~ ^ \widehat{\widetilde{v}} v vanishing on the reference Dirichlet boundary and all admissible q ^ \widehat{q} q , require
ρ f ∫ Ω ^ f J ^ f n + 1 u ^ n + 1 − u ^ n Δ t ⋅ v ~ ^ d x ^ + ρ f ∫ Ω ^ f J ^ f n + 1 [ ∇ ^ u ^ n + 1 ( F ^ f n + 1 ) − 1 ( u n + 1 − w f n + 1 ) ] ⋅ v ~ ^ d x ^ + ∫ Ω ^ f J ^ f n + 1 σ ^ f n + 1 ( F ^ f n + 1 ) − T : ∇ ^ v ~ ^ d x ^ + ∫ Ω ^ f J ^ f n + 1 q ^ tr ( ∇ ^ u ^ n + 1 ( F ^ f n + 1 ) − 1 ) d x ^ + γ ∫ ω ( a ^ n + 1 ∘ φ ) ⋅ ( v ~ ^ ∘ φ ) d x d y a r c = ∫ Ω f n + 1 F f n + 1 ⋅ v ~ d x − ∫ Γ f i n , n + 1 p i n n + 1 n f n + 1 ⋅ v ~ d S − ∫ Γ f o u t , n + 1 p o u t n + 1 n f n + 1 ⋅ v ~ d S + ∫ ω [ γ 2 ( u ^ n ∘ φ + η ˙ n ) + 1 2 ( t ^ f n − f s r e f , n ) ] ⋅ ( v ~ ∘ φ ) d x d y a r c . (63) \begin{aligned}
\rho_f\int_{\widehat{\Omega}_f}\widehat{J}_f^{n+1}\frac{\widehat{u}^{n+1}-\widehat{u}^{n}}{\Delta t}\cdot\widehat{\widetilde{v}}\,d\widehat{x}
+\rho_f\int_{\widehat{\Omega}_f}\widehat{J}_f^{n+1}\left[\widehat{\nabla}\widehat{u}^{n+1}(\widehat{F}_f^{n+1})^{-1}(u^{n+1}-w_f^{n+1})\right]\cdot\widehat{\widetilde{v}}\,d\widehat{x} \\
&+\int_{\widehat{\Omega}_f}\widehat{J}_f^{n+1}\widehat{\boldsymbol{\sigma}}_f^{n+1}(\widehat{F}_f^{n+1})^{-T}:\widehat{\nabla}\widehat{\widetilde{v}}\,d\widehat{x}
+\int_{\widehat{\Omega}_f}\widehat{J}_f^{n+1}\widehat{q}\operatorname{tr}\left(\widehat{\nabla}\widehat{u}^{n+1}(\widehat{F}_f^{n+1})^{-1}\right)\,d\widehat{x} \\
&+\gamma\int_{\omega}(\widehat{a}^{n+1}\circ\varphi)\cdot(\widehat{\widetilde{v}}\circ\varphi)\,dx\,dy_{\mathrm{arc}} \\
&=\int_{\Omega_f^{n+1}}F_f^{n+1}\cdot\widetilde{v}\,dx
-\int_{\Gamma_f^{\mathrm{in},n+1}}p_{\mathrm{in}}^{n+1}n_f^{n+1}\cdot\widetilde{v}\,dS
-\int_{\Gamma_f^{\mathrm{out},n+1}}p_{\mathrm{out}}^{n+1}n_f^{n+1}\cdot\widetilde{v}\,dS \\
&+\int_{\omega}\left[\frac{\gamma}{2}(\widehat{u}^{n}\circ\varphi+\dot{\eta}^{n})+\frac{1}{2}\left(\widehat{t}_f^{n}-f_s^{\mathrm{ref},n}\right)\right]\cdot(\widetilde{v}\circ\varphi)\,dx\,dy_{\mathrm{arc}}. \tag*{(63)}
\end{aligned} ρ f ∫ Ω f J f n + 1 Δ t u n + 1 − u n ⋅ v d x + ρ f ∫ Ω f J f n + 1 [ ∇ u n + 1 ( F f n + 1 ) − 1 ( u n + 1 − w f n + 1 ) ] ⋅ v d x + ∫ Ω f J f n + 1 σ f n + 1 ( F f n + 1 ) − T : ∇ v d x + ∫ Ω f J f n + 1 q tr ( ∇ u n + 1 ( F f n + 1 ) − 1 ) d x + γ ∫ ω ( a n + 1 ∘ φ ) ⋅ ( v ∘ φ ) d x d y arc = ∫ Ω f n + 1 F f n + 1 ⋅ v d x − ∫ Γ f in , n + 1 p in n + 1 n f n + 1 ⋅ v d S − ∫ Γ f out , n + 1 p out n + 1 n f n + 1 ⋅ v d S + ∫ ω [ 2 γ ( u n ∘ φ + η ˙ n ) + 2 1 ( t f n − f s ref , n ) ] ⋅ ( v ∘ φ ) d x d y arc . (63)
Here v ~ = v ~ ^ ∘ ( φ ^ f n + 1 ) − 1 \widetilde{v}=\widehat{\widetilde{v}}\circ(\widehat{\varphi}_f^{n+1})^{-1} v = v ∘ ( φ f n + 1 ) − 1 . The stress already contains pressure, and the last volume term imposes the transformed divergence constraint. The interface integrals are over reference parameter area because t ^ f \widehat{t}_f t f already contains the surface Jacobian.
After the fluid solve, recover t ^ f n + 1 \widehat{\mathbf{t}}_f^{n+1} t f n + 1 from (49)
t ^ f n + 1 = γ 2 ( a ^ n ∘ φ + η ˙ n ) + 1 2 ( t ^ f n − f r e f , n ) − γ u f n + 1 ∘ φ . (64) \widehat{\mathbf{t}}_f^{n+1}=\frac{\gamma}{2}\left(\widehat{\mathbf{a}}^n\circ\varphi+\dot{\boldsymbol{\eta}}^n\right)+\frac{1}{2}\left(\widehat{\mathbf{t}}_f^n-\mathbf{f}^{\mathrm{ref},n}\right)-\gamma\mathbf{u}_f^{n+1}\circ\varphi. \tag*{(64)} t f n + 1 = 2 γ ( a n ∘ φ + η ˙ n ) + 2 1 ( t f n − f ref , n ) − γ u f n + 1 ∘ φ . (64)
For a sufficiently regular solution, the Robin condition makes (64) agree with the Cauchy traction in (60) . In a finite element calculation, we use the Robin relation to define the exchanged load in the chosen interface space.
Lagged skew-symmetric convection formulation. In our numerical implementation, we linearize the fluid convection term by using a known velocity from the previous time step. Specifically, we define the convection velocity on the updated fluid domain as
a n + 1 = u ^ n ∘ ( φ ^ f n + 1 ) − 1 − w f n + 1 , (65) \mathbf{a}^{n+1}=\widehat{\mathbf{u}}^n\circ\left(\widehat{\varphi}_f^{n+1}\right)^{-1}-\mathbf{w}_f^{n+1}, \tag*{(65)} a n + 1 = u n ∘ ( φ f n + 1 ) − 1 − w f n + 1 , (65)
where the previous fluid velocity is mapped to the current fluid domain and the mesh velocity is subtracted to account for the moving domain.
Using this known convection velocity, we replace the fully implicit convection term with the following skew-symmetric bilinear form:
c n + 1 ( z , v ) = ρ f 2 ∫ Ω f n + 1 [ ( ( a n + 1 ⋅ ∇ ) z ) ⋅ v − ( ( a n + 1 ⋅ ∇ ) v ) ⋅ z ] d x . (66) c^{n+1}(\mathbf{z},\mathbf{v})=\frac{\rho_f}{2}\int_{\Omega_f^{n+1}}\left[((\mathbf{a}^{n+1}\cdot\nabla)\mathbf{z})\cdot\mathbf{v}-((\mathbf{a}^{n+1}\cdot\nabla)\mathbf{v})\cdot\mathbf{z}\right]\,dx. \tag*{(66)} c n + 1 ( z , v ) = 2 ρ f ∫ Ω f n + 1 [ (( a n + 1 ⋅ ∇ ) z ) ⋅ v − (( a n + 1 ⋅ ∇ ) v ) ⋅ z ] d x . (66)
The main reasoning is that this formulation has the important property of c n + 1 ( z , z ) = 0 c^{n+1}(\mathbf{z},\mathbf{z})=0 c n + 1 ( z , z ) = 0 , which holds regardless of whether a n + 1 \mathbf{a}^{n+1} a n + 1 is divergence-free or satisfies any particular boundary condition. Thus, the convection term does not contribute to the energy when the test and trial functions are identical. When the convection term is omitted, this distinction disappears, and both formulations reduce to the Stokes problem considered in the analysis below.
Pressure stabilization. In our numerical example, we use continuous piecewise-linear finite elements for both the fluid velocity and pressure (P 1 P_1 P 1 -P 1 P_1 P 1 ) on the tetrahedral mesh. Since this equal-order finite element pair does not satisfy the inf-sup stability condition, we introduce the following pressure stabilization term:
s h n + 1 ( p h n + 1 , q h ) = δ p ∑ K ∈ T h n + 1 h K 2 ∫ K ∇ p h n + 1 ⋅ ∇ q h d x , δ p > 0 , (67) s_h^{n+1}(p_h^{n+1},q_h)=\delta_p\sum_{K\in\mathcal{T}_h^{n+1}}h_K^2\int_K\nabla p_h^{n+1}\cdot\nabla q_h\,dx,\qquad\delta_p>0, \tag*{(67)} s h n + 1 ( p h n + 1 , q h ) = δ p K ∈ T h n + 1 ∑ h K 2 ∫ K ∇ p h n + 1 ⋅ ∇ q h d x , δ p > 0 , (67)
where h K h_K h K denotes the diameter of the current mesh element K K K , and δ p \delta_p δ p is a positive stabilization parameter. With this stabilization, the discrete continuity equation becomes
( q h , ∇ ⋅ u h n + 1 ) + s h n + 1 ( p h n + 1 , q h ) = 0. (q_h,\nabla\cdot\mathbf{u}_h^{n+1})+s_h^{n+1}(p_h^{n+1},q_h)=0. ( q h , ∇ ⋅ u h n + 1 ) + s h n + 1 ( p h n + 1 , q h ) = 0.
When testing the fluid equations with the new velocity and pressure, the mixed pressure terms cancel. In addition, the skew-symmetric convection term vanishes, while the pressure stabilization contributes the nonnegative term
Δ t s h n + 1 ( p h n + 1 , p h n + 1 ) ≥ 0 \Delta t\,s_h^{n+1}(p_h^{n+1},p_h^{n+1})\ge0 Δ t s h n + 1 ( p h n + 1 , p h n + 1 ) ≥ 0
to the discrete energy balance. Since the fluid domain geometry and convection velocity are known at each time step, the resulting fluid problem is linear. We emphasize that the analysis presented below focuses on the time discretization of the Stokes problem on a fixed domain. Therefore, the effects of spatial stabilization and errors arising from the transfer of interface quantities are not included in the analysis.
Naghdi shell subproblem on the reference parameter domain
The Naghdi shell problem is solved on the fixed reference domain ω \omega ω . We seek the displacement η n + 1 \eta^{n+1} η n + 1 , rotation ω n + 1 \omega^{n+1} ω n + 1 , and their corresponding velocities η ˙ n + 1 \dot{\eta}^{n+1} η ˙ n + 1 and ω ˙ n + 1 \dot{\omega}^{n+1} ω ˙ n + 1 satisfying
η n + 1 − η n Δ t = η ˙ n + 1 , ω n + 1 − ω n Δ t = ω ˙ n + 1 , (68) \frac{\eta^{n+1}-\eta^n}{\Delta t}=\dot{\eta}^{n+1}, \qquad\frac{\omega^{n+1}-\omega^n}{\Delta t}=\dot{\omega}^{n+1}, \tag*{(68)} Δ t η n + 1 − η n = η ˙ n + 1 , Δ t ω n + 1 − ω n = ω ˙ n + 1 , (68)
and, for every ( v , w ) ∈ V N ( ω ) (v,w)\in V_N(\omega) ( v , w ) ∈ V N ( ω ) ,
ε ρ s ∫ ω η ˙ n + 1 − η ˙ n Δ t ⋅ v d x d y a r c + ε 3 ρ s 12 ∫ ω ω ˙ n + 1 − ω ˙ n Δ t ⋅ w d x d y a r c + B m ( ( η n + 1 , ω n + 1 ) , ( v , w ) ) + B b ( ( η n + 1 , ω n + 1 ) , ( v , w ) ) + γ ∫ ω η ˙ n + 1 ⋅ v d x d y a r c = ∫ ω [ γ 2 ( u ^ n ∘ φ + η ˙ n ) − 1 2 ( t ^ f n − f s r e f , n ) ] ⋅ v d x d y a r c . (69) \begin{aligned}
\varepsilon\rho_s\int_\omega\frac{\dot{\eta}^{n+1}-\dot{\eta}^n}{\Delta t}\cdot v\,dx\,dy_{\mathrm{arc}}+\frac{\varepsilon^3\rho_s}{12}\int_\omega\frac{\dot{\omega}^{n+1}-\dot{\omega}^n}{\Delta t}\cdot w\,dx\,dy_{\mathrm{arc}}+B_m((\eta^{n+1},\omega^{n+1}),(v,w))+B_b((\eta^{n+1},\omega^{n+1}),(v,w))+\gamma\int_\omega\dot{\eta}^{n+1}\cdot v\,dx\,dy_{\mathrm{arc}} \\
&=\int_\omega\left[\frac{\gamma}{2}(\hat{u}^n\circ\varphi+\dot{\eta}^n)-\frac{1}{2}(\hat{t}_f^n-f_s^{\mathrm{ref},n})\right]\cdot v\,dx\,dy_{\mathrm{arc}}. \tag*{(69)}
\end{aligned} ε ρ s ∫ ω Δ t η ˙ n + 1 − η ˙ n ⋅ v d x d y arc + 12 ε 3 ρ s ∫ ω Δ t ω ˙ n + 1 − ω ˙ n ⋅ w d x d y arc + B m (( η n + 1 , ω n + 1 ) , ( v , w )) + B b (( η n + 1 , ω n + 1 ) , ( v , w )) + γ ∫ ω η ˙ n + 1 ⋅ v d x d y arc = ∫ ω [ 2 γ ( u ^ n ∘ φ + η ˙ n ) − 2 1 ( t ^ f n − f s ref , n ) ] ⋅ v d x d y arc . (69)
Once the shell problem is solved, the new shell interface load is recovered from the Robin condition:
f s r e f , n + 1 = γ 2 ( u ^ n ∘ φ + η ˙ n ) − 1 2 ( t ^ f n − f s r e f , n ) − γ η ˙ n + 1 . (70) f_s^{\mathrm{ref},n+1}=\frac{\gamma}{2}(\hat{u}^n\circ\varphi+\dot{\eta}^n)-\frac{1}{2}(\hat{t}_f^n-f_s^{\mathrm{ref},n})-\gamma\dot{\eta}^{n+1}. \tag*{(70)} f s ref , n + 1 = 2 γ ( u ^ n ∘ φ + η ˙ n ) − 2 1 ( t ^ f n − f s ref , n ) − γ η ˙ n + 1 . (70)
Equation (70) recovers the load that balances the computed shell inertia and elastic forces. As it only uses the new shell velocity and old interface data, we store this load for the next step.
Finally, the updated shell midsurface is
Φ s n + 1 = φ + η n + 1 . (71) \Phi_s^{n+1}=\varphi+\eta^{n+1}. \tag*{(71)} Φ s n + 1 = φ + η n + 1 . (71)
Since the fluid domain is constructed using the predicted interface Φ s n + 1 , ∗ \Phi_s^{n+1,*} Φ s n + 1 , ∗ , the fluid and shell geometries may differ at the end of the time step. From (68) , this difference is
Φ s n + 1 − Φ s n + 1 , ∗ = Δ t ( η ˙ n + 1 − η ˙ n ) . (72) \Phi_s^{n+1}-\Phi_s^{n+1,*}=\Delta t(\dot{\eta}^{n+1}-\dot{\eta}^n). \tag*{(72)} Φ s n + 1 − Φ s n + 1 , ∗ = Δ t ( η ˙ n + 1 − η ˙ n ) . (72)
Equation (72) gives the geometric difference between the two at the completed step. For a smooth exact shell motion, the corresponding local geometry prediction error is O ( Δ t 2 ) O(\Delta t^2) O ( Δ t 2 ) .
Initialization and time-stepping algorithm
The initial ALE map is constructed by extending the prescribed shell displacement η 0 \eta^0 η 0 into the reference fluid domain, such that
φ ^ f 0 ∘ φ = φ + η 0 . \hat{\varphi}_f^0\circ\varphi=\varphi+\eta^0. φ ^ f 0 ∘ φ = φ + η 0 .
The initial reference fluid velocity is u ^ 0 = u 0 ∘ φ ^ f 0 \hat{u}^0=u^0\circ\hat{\varphi}_f^0 u ^ 0 = u 0 ∘ φ ^ f 0 , with the compatibility conditions given in § 2 . If a compatible initial pressure p 0 p^0 p 0 is available, the initial fluid traction and shell load are
t ^ f 0 = J s 0 [ σ f ( u 0 , p 0 ) n f 0 ] ∘ ( φ + η 0 ) , J s 0 = ∣ ∂ 1 ( φ + η 0 ) × ∂ 2 ( φ + η 0 ) ∣ , f s r e f , 0 = − t ^ f 0 . (73) \hat{t}_f^0=J_s^0\left[\sigma_f(u^0,p^0)n_f^0\right]\circ(\varphi+\eta^0), \qquad J_s^0=\left|\partial_1(\varphi+\eta^0)\times\partial_2(\varphi+\eta^0)\right|,\qquad f_s^{\mathrm{ref},0}=-\hat{t}_f^0. \tag*{(73)} t ^ f 0 = J s 0 [ σ f ( u 0 , p 0 ) n f 0 ] ∘ ( φ + η 0 ) , J s 0 = ∂ 1 ( φ + η 0 ) × ∂ 2 ( φ + η 0 ) , f s ref , 0 = − t ^ f 0 . (73)
If the initial traction is unavailable, it may be obtained from a compatible initialization solve or approximated with a controlled error. This error enters the initial augmented error energy. Both initial loads may be set to zero when the physical initial interface traction vanishes.
The complete procedure is given in Algorithm 1.
Algorithm 1 Parallel Robin-Robin fluid-Naghdi shell scheme
Require: Initial fluid and shell states u 0 \mathbf{u}^{0} u 0 , η 0 \eta^{0} η 0 , η ˙ 0 \dot{\eta}^{0} η ˙ 0 , ω 0 \omega^{0} ω 0 , ω ˙ 0 \dot{\omega}^{0} ω ˙ 0 ; initial interface loads t ^ f 0 \widehat{\mathbf{t}}_{f}^{0} t f 0 , f s r e f , 0 \mathbf{f}_{s}^{\mathrm{ref},0} f s ref , 0 ; Δ t > 0 \Delta t > 0 Δ t > 0 , γ > 0 \gamma> 0 γ > 0 .
Initialization
Construct the initial ALE map φ ^ f 0 \widehat{\varphi}_{f}^{0} φ f 0 .
Set u ^ 0 = u 0 ∘ φ ^ f 0 \widehat{\mathbf{u}}^{0} = \mathbf{u}^{0} \circ\widehat{\varphi}_{f}^{0} u 0 = u 0 ∘ φ f 0 .
for n = 0 , … , N − 1 n = 0,\ldots,N-1 n = 0 , … , N − 1 do
Step 1: Construct the Robin data.
Form the fluid Robin right-hand side: γ 2 ( u ^ n ∘ φ + η ˙ n ) + 1 2 ( t ^ f n − f s r e f , n ) \frac{\gamma}{2}(\widehat{\mathbf{u}}^{n} \circ\varphi+ \dot{\eta}^{n}) + \frac{1}{2}(\widehat{\mathbf{t}}_{f}^{n} - \mathbf{f}_{s}^{\mathrm{ref},n}) 2 γ ( u n ∘ φ + η ˙ n ) + 2 1 ( t f n − f s ref , n ) .
Form the shell Robin right-hand side: γ 2 ( u ^ n ∘ φ + η ˙ n ) − 1 2 ( t ^ f n − f s r e f , n ) \frac{\gamma}{2}(\widehat{\mathbf{u}}^{n} \circ\varphi+ \dot{\eta}^{n}) - \frac{1}{2}(\widehat{\mathbf{t}}_{f}^{n} - \mathbf{f}_{s}^{\mathrm{ref},n}) 2 γ ( u n ∘ φ + η ˙ n ) − 2 1 ( t f n − f s ref , n ) .
Step 2: Solve the fluid and shell subproblems concurrently.
_Fluid subproblem:_
Predict the interface Φ s n + 1 , ∗ \Phi_{s}^{n+1,*} Φ s n + 1 , ∗ using (52) .
Extend the predicted displacement using (53) .
Construct the ALE map φ ^ f n + 1 \widehat{\varphi}_{f}^{n+1} φ f n + 1 and mesh velocity w f n + 1 \mathbf{w}_{f}^{n+1} w f n + 1 .
Solve (63) for ( u n + 1 , p n + 1 ) (\mathbf{u}^{n+1},p^{n+1}) ( u n + 1 , p n + 1 ) .
Compute the reference fluid velocity u ^ n + 1 = u n + 1 ∘ φ ^ f n + 1 \widehat{\mathbf{u}}^{n+1} = \mathbf{u}^{n+1} \circ\widehat{\varphi}_{f}^{n+1} u n + 1 = u n + 1 ∘ φ f n + 1 .
Recover t ^ f n + 1 \widehat{\mathbf{t}}_{f}^{n+1} t f n + 1 using (64) .
_Shell subproblem:_
Solve (68) and (69) for η n + 1 \eta^{n+1} η n + 1 , ω n + 1 \omega^{n+1} ω n + 1 , η ˙ n + 1 \dot{\eta}^{n+1} η ˙ n + 1 , and ω ˙ n + 1 \dot{\omega}^{n+1} ω ˙ n + 1 .
Recover f s r e f , n + 1 \mathbf{f}_{s}^{\mathrm{ref},n+1} f s ref , n + 1 using (70) .
Update the shell surface Φ s n + 1 = φ + η n + 1 \Phi_{s}^{n+1} = \varphi+ \eta^{n+1} Φ s n + 1 = φ + η n + 1 .
Step 3: Exchange the new interface velocities and loads after both subproblems finish. Store u ^ n + 1 \widehat{\mathbf{u}}^{n+1} u n + 1 , η ˙ n + 1 \dot{\eta}^{n+1} η ˙ n + 1 , t ^ f n + 1 \widehat{\mathbf{t}}_{f}^{n+1} t f n + 1 , and f s r e f , n + 1 \mathbf{f}_{s}^{\mathrm{ref},n+1} f s ref , n + 1 for the next time step.
end for
Remark 3.2 (Discrete coupling defects and scope of the analysis). Adding the two Robin conditions gives the force imbalance
t ^ f n + 1 + f s r e f , n + 1 = − γ [ u ^ n + 1 ∘ φ − u ^ n ∘ φ + η ˙ n + 1 − η ˙ n ] . (74) \widehat{\mathbf{t}}_{f}^{n+1}+\mathbf{f}_{s}^{\mathrm{ref},n+1}
=-\gamma\left[\widehat{\mathbf{u}}^{n+1}\circ\varphi-\widehat{\mathbf{u}}^{n}\circ\varphi+\dot{\eta}^{n+1}-\dot{\eta}^{n}\right]. \tag*{(74)} t f n + 1 + f s ref , n + 1 = − γ [ u n + 1 ∘ φ − u n ∘ φ + η ˙ n + 1 − η ˙ n ] . (74)
Subtracting the two conditions gives the velocity mismatch
γ ( u ^ n + 1 ∘ φ − η ˙ n + 1 ) = − ( t ^ f n + 1 − t ^ f n ) + ( f s r e f , n + 1 − f s r e f , n ) . (75) \gamma\left(\widehat{\mathbf{u}}^{n+1}\circ\varphi-\dot{\eta}^{n+1}\right)
=-\left(\widehat{\mathbf{t}}_{f}^{n+1}-\widehat{\mathbf{t}}_{f}^{n}\right)
+\left(\mathbf{f}_{s}^{\mathrm{ref},n+1}-\mathbf{f}_{s}^{\mathrm{ref},n}\right). \tag*{(75)} γ ( u n + 1 ∘ φ − η ˙ n + 1 ) = − ( t f n + 1 − t f n ) + ( f s ref , n + 1 − f s ref , n ) . (75)
Thus, the interface forces and velocities need not match exactly at each time step. Their differences depend on the corresponding time increments and are controlled through the Robin work identity in the fixed-domain analysis.
For the moving-domain scheme, additional effects arise from the difference between the predicted fluid interface and the computed shell surface, as well as the possible mismatch between fluid and mesh velocities. To isolate the coupling mechanism, the analysis in § 4 considers a fixed fluid domain without convection, while retaining the full Naghdi shell dynamics. Stability analysis on a moving domain would also require accounting for changes in the fluid volume, discrete geometric conservation, and boundary energy fluxes.
Fixed-domain Stokes-Naghdi Shell formulation
We consider the fixed-domain Stokes-Naghdi shell problem for the stability and error analysis. In this setting, shell deformation affects the interface forces but does not change the fluid geometry. With convection omitted, the fully implicit fluid formulation (59) and the lagged skew-symmetric formulation (66) reduce to the same Stokes problem. The analysis considers only time discretization and excludes the pressure stabilization (67) used in our numerical implementation.
For every t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] , take
Ω ^ f ( t ) = Ω ^ f , Γ ^ s ( t ) = Γ ^ s , φ ^ f = Id , w f = 0 , F ^ f = I 3 , J ^ f = 1. (76) \widehat{\Omega}_f(t)=\widehat{\Omega}_f,\qquad\widehat{\Gamma}_s(t)=\widehat{\Gamma}_s,\qquad\widehat{\varphi}_f=\operatorname{Id},\qquad\boldsymbol{w}_f=0,\qquad\widehat{\boldsymbol{F}}_f=\boldsymbol{I}_3,\qquad\widehat{J}_f=1. \tag*{(76)} Ω f ( t ) = Ω f , Γ s ( t ) = Γ s , φ f = Id , w f = 0 , F f = I 3 , J f = 1. (76)
The interface is parameterized by the reference cylindrical map in (2) . The reference surface measure is d S 0 = d x d y a r c dS_0=dx\,dy_{\mathrm{arc}} d S 0 = d x d y arc , and fluid traces are evaluated through φ \varphi φ . The fluid normal and all integration measures stay fixed as the shell moves. Accordingly, a unit interface Jacobian describes the frozen coupling surface, rather than the geometry of the displaced shell.
Since the fluid reference and current coordinates now coincide, we omit hats on velocity and pressure. We keep the hat on t ^ f \widehat{\boldsymbol{t}}_f t f to identify a load on the shell parameter domain:
t ^ f ( x , y a r c , t ) = [ σ f ( u , p ) n f ] ( φ ( x , y a r c ) , t ) . (77) \widehat{\boldsymbol{t}}_f(x,y_{\mathrm{arc}},t)=\left[\boldsymbol{\sigma}_f(\boldsymbol{u},p)\boldsymbol{n}_f\right]\left(\varphi(x,y_{\mathrm{arc}}),t\right). \tag*{(77)} t f ( x , y arc , t ) = [ σ f ( u , p ) n f ] ( φ ( x , y arc ) , t ) . (77)
Here n f \boldsymbol{n}_f n f is the fixed outward fluid normal. The shell frame normal a 3 \boldsymbol{a}_3 a 3 keeps its definition from Section 2.
Function spaces and analytical assumptions
Assume that Ω ^ f \widehat{\Omega}_f Ω f is a bounded Lipschitz domain with the boundary decomposition specified in Section 2. Define
V f = { v ~ ∈ H 1 ( Ω ^ f ; R 3 ) : v ~ = 0 on Σ ^ f D } . (78) V_f=\left\{\widetilde{\boldsymbol{v}}\in H^1(\widehat{\Omega}_f;\mathbb{R}^3):\widetilde{\boldsymbol{v}}=0\text{ on }\widehat{\Sigma}_f^D\right\}. \tag*{(78)} V f = { v ∈ H 1 ( Ω f ; R 3 ) : v = 0 on Σ f D } . (78)
For the mixed traction formulation with pressure level fixed by the boundary conditions, take
Q f = L 2 ( Ω ^ f ) . (79) Q_f=L^2(\widehat{\Omega}_f). \tag*{(79)} Q f = L 2 ( Ω f ) . (79)
Assume the appropriate divergence inf-sup condition: there exists β f > 0 \beta_f>0 β f > 0 , such that
inf q ∈ Q f q ≠ 0 sup v ~ ∈ V f v ~ ≠ 0 ∫ Ω ^ f q ∇ ⋅ v ~ d x ∥ q ∥ L 2 ( Ω ^ f ) ∥ v ~ ∥ H 1 ( Ω ^ f ) ≥ β f . (80) \inf_{\substack{q\in Q_f\\q\ne0}}\sup_{\substack{\widetilde{\boldsymbol{v}}\in V_f\\\widetilde{\boldsymbol{v}}\ne0}}
\frac{\int_{\widehat{\Omega}_f}q\nabla\cdot\widetilde{\boldsymbol{v}}\,dx}
{\|q\|_{L^2(\widehat{\Omega}_f)}\|\widetilde{\boldsymbol{v}}\|_{H^1(\widehat{\Omega}_f)}}\ge\beta_f. \tag*{(80)} q ∈ Q f q = 0 inf v ∈ V f v = 0 sup ∥ q ∥ L 2 ( Ω f ) ∥ v ∥ H 1 ( Ω f ) ∫ Ω f q ∇ ⋅ v d x ≥ β f . (80)
The prescribed traction fixes the pressure level in this setting. If a different boundary condition leaves a pressure nullspace, the pressure space must instead include the corresponding normalization.
We use the shell trial and test space V N ( ω ) V_N(\omega) V N ( ω ) defined in (35) . We retain B m B_m B m and B b B_b B b from Section 2 and assume that they are continuous, symmetric, and nonnegative on V N ( ω ) V_N(\omega) V N ( ω ) . These properties are the basis of the energy calculation.
To prove solvability in the stated H 1 H^1 H 1 shell space, we also use the following shifted coercivity bound, with c N > 0 c_N>0 c N > 0 and C N ≥ 0 C_N\ge0 C N ≥ 0 :
B m ( ( v , w ) , ( v , w ) ) + B b ( ( v , w ) , ( v , w ) ) + C N ( ∥ v ∥ L 2 ( ω ) 2 + ∥ w ∥ L 2 ( ω ) 2 ) ≥ c N ( ∥ v ∥ H 1 ( ω ) 2 + ∥ w ∥ H 1 ( ω ) 2 ) for every ( v , w ) ∈ V N ( ω ) . (81) B_m((\boldsymbol{v},\boldsymbol{w}),(\boldsymbol{v},\boldsymbol{w}))+B_b((\boldsymbol{v},\boldsymbol{w}),(\boldsymbol{v},\boldsymbol{w}))+C_N\left(\|\boldsymbol{v}\|_{L^2(\omega)}^2+\|\boldsymbol{w}\|_{L^2(\omega)}^2\right)
\ge c_N\left(\|\boldsymbol{v}\|_{H^1(\omega)}^2+\|\boldsymbol{w}\|_{H^1(\omega)}^2\right)
\quad\text{for every }(\boldsymbol{v},\boldsymbol{w})\in V_N(\omega). \tag*{(81)} B m (( v , w ) , ( v , w )) + B b (( v , w ) , ( v , w )) + C N ( ∥ v ∥ L 2 ( ω ) 2 + ∥ w ∥ L 2 ( ω ) 2 ) ≥ c N ( ∥ v ∥ H 1 ( ω ) 2 + ∥ w ∥ H 1 ( ω ) 2 ) for every ( v , w ) ∈ V N ( ω ) . (81)
For the cylindrical material law, this bound follows from the full membrane-shear and bending strains. We give the argument because the same bounds will later control the shell error. Set
c s = E s ν s 1 − ν s 2 , k s = min { μ , 2 ( μ + c s ) } > 0 , (82) c_s = \frac{E_s\nu_s}{1-\nu_s^2}, \qquad k_s = \min\{\mu,2(\mu+c_s)\} > 0, \tag*{(82)} c s = 1 − ν s 2 E s ν s , k s = min { μ , 2 ( μ + c s )} > 0 , (82)
where μ + c s = E s / [ 2 ( 1 − ν s ) ] > 0 \mu+c_s=E_s/[2(1-\nu_s)]>0 μ + c s = E s / [ 2 ( 1 − ν s )] > 0 . Decomposition of a 2 × 2 2 \times2 2 × 2 matrix into its trace and trace-free parts gives
c s ( tr S ) 2 + 2 μ ∣ S ∣ 2 ≥ min { 2 μ , 2 ( μ + c s ) } ∣ S ∣ 2 . (83) c_s(\operatorname{tr}S)^2 + 2\mu\lvert S\rvert^2 \ge\min\{2\mu,2(\mu+c_s)\}\lvert S\rvert^2. \tag*{(83)} c s ( tr S ) 2 + 2 μ ∣ S ∣ 2 ≥ min { 2 μ , 2 ( μ + c s )} ∣ S ∣ 2 . (83)
Since the reference frame is orthonormal and J \mathbf{J} J is orthogonal,
B m ( ( v , w ) , ( v , w ) ) ≥ ε k s ∑ α = 1 2 ∥ ∂ α v + a α × w ∥ L 2 ( ω ) 2 , B b ( ( v , w ) , ( v , w ) ) ≥ ε 3 k s 12 ∥ ∇ w ∥ L 2 ( ω ) 2 . (84) \begin{aligned}
B_m((\mathbf{v},\mathbf{w}),(\mathbf{v},\mathbf{w})) &\ge\varepsilon k_s \sum_{\alpha=1}^{2}\lVert\partial_\alpha\mathbf{v}+\mathbf{a}_\alpha\times\mathbf{w}\rVert_{L^2(\omega)}^2,\\
B_b((\mathbf{v},\mathbf{w}),(\mathbf{v},\mathbf{w})) &\ge\frac{\varepsilon^3 k_s}{12}\lVert\nabla\mathbf{w}\rVert_{L^2(\omega)}^2. \tag*{(84)}
\end{aligned} B m (( v , w ) , ( v , w )) B b (( v , w ) , ( v , w )) ≥ ε k s α = 1 ∑ 2 ∥ ∂ α v + a α × w ∥ L 2 ( ω ) 2 , ≥ 12 ε 3 k s ∥ ∇ w ∥ L 2 ( ω ) 2 . (84)
Moreover,
∥ ∇ v ∥ L 2 ( ω ) 2 ≤ 2 ∑ α = 1 2 ∥ ∂ α v + a α × w ∥ L 2 ( ω ) 2 + 4 ∥ w ∥ L 2 ( ω ) 2 . (85) \lVert\nabla\mathbf{v}\rVert_{L^2(\omega)}^2 \le2\sum_{\alpha=1}^{2}\lVert\partial_\alpha\mathbf{v}+\mathbf{a}_\alpha\times\mathbf{w}\rVert_{L^2(\omega)}^2+4\lVert\mathbf{w}\rVert_{L^2(\omega)}^2. \tag*{(85)} ∥ ∇ v ∥ L 2 ( ω ) 2 ≤ 2 α = 1 ∑ 2 ∥ ∂ α v + a α × w ∥ L 2 ( ω ) 2 + 4 ∥ w ∥ L 2 ( ω ) 2 . (85)
The preceding estimates control the shell gradients after adding sufficiently large positive L 2 L^2 L 2 terms. This proves (81) for the cylindrical law. For other constitutive tensors, we retain that inequality as an assumption.
Clamping both ends gives a stronger result for the present cylinder. Poincaré’s inequality in the axial direction yields C P > 0 C_P>0 C P > 0 such that
∥ v ∥ L 2 ( ω ) ≤ C P ∥ ∇ v ∥ L 2 ( ω ) , ∥ w ∥ L 2 ( ω ) ≤ C P ∥ ∇ w ∥ L 2 ( ω ) (86) \lVert\mathbf{v}\rVert_{L^2(\omega)}\le C_P\lVert\nabla\mathbf{v}\rVert_{L^2(\omega)},\qquad\lVert\mathbf{w}\rVert_{L^2(\omega)}\le C_P\lVert\nabla\mathbf{w}\rVert_{L^2(\omega)} \tag*{(86)} ∥ v ∥ L 2 ( ω ) ≤ C P ∥ ∇ v ∥ L 2 ( ω ) , ∥ w ∥ L 2 ( ω ) ≤ C P ∥ ∇ w ∥ L 2 ( ω ) (86)
for every ( v , w ) ∈ V N ( ω ) (\mathbf{v},\mathbf{w})\in V_N(\omega) ( v , w ) ∈ V N ( ω ) . The bending bound first controls w \mathbf{w} w in H 1 H^1 H 1 . The membrane bound then controls ∇ v \nabla\mathbf{v} ∇ v , since the rotation term is already bounded. Applying Poincaré’s inequality to v \mathbf{v} v gives
∥ v ∥ H 1 ( ω ) 2 + ∥ w ∥ H 1 ( ω ) 2 ≤ C e l [ B m ( ( v , w ) , ( v , w ) ) + B b ( ( v , w ) , ( v , w ) ) ] . (87) \lVert\mathbf{v}\rVert_{H^1(\omega)}^2+\lVert\mathbf{w}\rVert_{H^1(\omega)}^2\le C_{\mathrm{el}}\left[B_m((\mathbf{v},\mathbf{w}),(\mathbf{v},\mathbf{w}))+B_b((\mathbf{v},\mathbf{w}),(\mathbf{v},\mathbf{w}))\right]. \tag*{(87)} ∥ v ∥ H 1 ( ω ) 2 + ∥ w ∥ H 1 ( ω ) 2 ≤ C el [ B m (( v , w ) , ( v , w )) + B b (( v , w ) , ( v , w )) ] . (87)
Here C e l C_{\mathrm{el}} C el depends on geometry, thickness, and material parameters, but not on Δ t \Delta t Δ t . Thus the elastic energy is equivalent to the full shell H 1 H^1 H 1 norm for the clamped cylinder. For a more general law with only shifted coercivity, the elastic form may define a seminorm; the arguments below also cover that case.
Throughout the analysis, ρ f \rho_f ρ f , μ f \mu_f μ f , ρ s \rho_s ρ s , and ε \varepsilon ε are positive. To write external work as L 2 L^2 L 2 integrals, we take F f ( t ) ∈ L 2 ( Ω ^ f ; R 3 ) F_f(t)\in L^2(\widehat{\Omega}_f;\mathbb{R}^3) F f ( t ) ∈ L 2 ( Ω f ; R 3 ) and p i n / o u t ( t ) ∈ L 2 ( Γ ^ f i n / o u t ) p_{\mathrm{in/out}}(t)\in L^2(\widehat{\Gamma}_f^{\mathrm{in/out}}) p in/out ( t ) ∈ L 2 ( Γ f in/out ) . The same work terms can be understood as continuous functionals for more general data. We state the time regularity needed for consistency in Section 6.
Continuous equations in the fixed domain
The fluid equations are
{ ρ f ∂ t u − ∇ ⋅ σ f ( u , p ) = F f in Ω ^ f , ∇ ⋅ u = 0 in Ω ^ f , σ f ( u , p ) = − p I 3 + 2 μ f D ( u ) , D ( u ) = 1 2 ( ∇ u + ( ∇ u ) T ) . (88) \left\{
\begin{aligned}
\rho_f\partial_t\mathbf{u}-\nabla\cdot\boldsymbol{\sigma}_f(\mathbf{u},p)&=F_f &&\text{in }\widehat{\Omega}_f,\\
\nabla\cdot\mathbf{u}&=0 &&\text{in }\widehat{\Omega}_f,\\
\boldsymbol{\sigma}_f(\mathbf{u},p)&=-pI_3+2\mu_f\mathbb{D}(\mathbf{u}),\\
\mathbb{D}(\mathbf{u})&=\frac{1}{2}\left(\nabla\mathbf{u}+(\nabla\mathbf{u})^{\mathsf{T}}\right).
\end{aligned}
\right. \tag*{(88)} ⎩ ⎨ ⎧ ρ f ∂ t u − ∇ ⋅ σ f ( u , p ) ∇ ⋅ u σ f ( u , p ) D ( u ) = F f = 0 = − p I 3 + 2 μ f D ( u ) , = 2 1 ( ∇ u + ( ∇ u ) T ) . in Ω f , in Ω f , (88)
The external boundary conditions are
u = 0 on Σ ^ f D , σ f ( u , p ) n f = − p i n n f on Γ ^ f i n , σ f ( u , p ) n f = − p o u t n f on Γ ^ f o u t . (89) \begin{aligned}
u &= 0 && \text{on } \widehat{\Sigma}^{D}_{f},\\
\boldsymbol{\sigma}_{f}(u,p)\boldsymbol{n}_{f} &= -p_{\mathrm{in}}\boldsymbol{n}_{f} && \text{on } \widehat{\Gamma}^{\mathrm{in}}_{f},\\
\boldsymbol{\sigma}_{f}(u,p)\boldsymbol{n}_{f} &= -p_{\mathrm{out}}\boldsymbol{n}_{f} && \text{on } \widehat{\Gamma}^{\mathrm{out}}_{f}. \tag*{(89)}
\end{aligned} u σ f ( u , p ) n f σ f ( u , p ) n f = 0 = − p in n f = − p out n f on Σ f D , on Γ f in , on Γ f out . (89)
The shell equation is:
ε ρ s ∫ ω ∂ t η ˙ ⋅ v d x d y a r c + ε 3 ρ s 12 ∫ ω ∂ t ω ˙ ⋅ w d x d y a r c + B m ( ( η , ω ) , ( v , w ) ) + B b ( ( η , ω ) , ( v , w ) ) = ∫ ω f s r e f ⋅ v d x d y a r c , ( v , w ) ∈ V N ( ω ) . (90) \varepsilon\rho_{s}\int_{\omega}\partial_{t}\dot{\boldsymbol{\eta}}\cdot\boldsymbol{v}\,dx\,dy_{\mathrm{arc}}+\frac{\varepsilon^{3}\rho_{s}}{12}\int_{\omega}\partial_{t}\dot{\boldsymbol{\omega}}\cdot\boldsymbol{w}\,dx\,dy_{\mathrm{arc}}+B_{m}((\boldsymbol{\eta},\boldsymbol{\omega}),(\boldsymbol{v},\boldsymbol{w}))+B_{b}((\boldsymbol{\eta},\boldsymbol{\omega}),(\boldsymbol{v},\boldsymbol{w}))
=
\int_{\omega}f_{s}^{\mathrm{ref}}\cdot\boldsymbol{v}\,dx\,dy_{\mathrm{arc}},\qquad(\boldsymbol{v},\boldsymbol{w})\in V_{N}(\omega). \tag*{(90)} ε ρ s ∫ ω ∂ t η ˙ ⋅ v d x d y arc + 12 ε 3 ρ s ∫ ω ∂ t ω ˙ ⋅ w d x d y arc + B m (( η , ω ) , ( v , w )) + B b (( η , ω ) , ( v , w )) = ∫ ω f s ref ⋅ v d x d y arc , ( v , w ) ∈ V N ( ω ) . (90)
The coupling conditions are imposed on the fixed reference surface:
u ∘ φ = η ˙ , t ^ f + f s r e f = 0 . (91) \boldsymbol{u}\circ\varphi=\dot{\boldsymbol{\eta}},\qquad\widehat{\boldsymbol{t}}_{f}+\boldsymbol{f}_{s}^{\mathrm{ref}}=\boldsymbol{0}. \tag*{(91)} u ∘ φ = η ˙ , t f + f s ref = 0 . (91)
The shell satisfies the clamped and periodic conditions from § 2 , together with compatible initial data.
Using the interface conditions with compatible test functions v ~ ∘ φ = v \widetilde{\boldsymbol{v}}\circ\varphi=\boldsymbol{v} v ∘ φ = v , we obtain the coupled weak formulation
ρ f ∫ Ω ^ f ∂ t u ⋅ v ~ d x + 2 μ f ∫ Ω ^ f D ( u ) : D ( v ~ ) d x − ∫ Ω ^ f p ∇ ⋅ v ~ d x + ∫ Ω ^ f q ∇ ⋅ u d x + ε ρ s ∫ ω ∂ t η ˙ ⋅ v d x d y a r c + ε 3 ρ s 12 ∫ ω ∂ t ω ˙ ⋅ w d x d y a r c + B m ( ( η , ω ) , ( v , w ) ) + B b ( ( η , ω ) , ( v , w ) ) = ∫ Ω ^ f F f ⋅ v ~ d x − ∫ Γ ^ f i n p i n n f ⋅ v ~ d S 0 − ∫ Γ ^ f o u t p o u t n f ⋅ v ~ d S 0 . (92) \begin{aligned}
\rho_{f}\int_{\widehat{\Omega}_{f}}\partial_{t}\boldsymbol{u}\cdot\widetilde{\boldsymbol{v}}\,dx
+2\mu_{f}\int_{\widehat{\Omega}_{f}}\mathbb{D}(\boldsymbol{u}):\mathbb{D}(\widetilde{\boldsymbol{v}})\,dx
-\int_{\widehat{\Omega}_{f}}p\nabla\cdot\widetilde{\boldsymbol{v}}\,dx
+\int_{\widehat{\Omega}_{f}}q\nabla\cdot\boldsymbol{u}\,dx\\
+\varepsilon\rho_{s}\int_{\omega}\partial_{t}\dot{\boldsymbol{\eta}}\cdot\boldsymbol{v}\,dx\,dy_{\mathrm{arc}}
+\frac{\varepsilon^{3}\rho_{s}}{12}\int_{\omega}\partial_{t}\dot{\boldsymbol{\omega}}\cdot\boldsymbol{w}\,dx\,dy_{\mathrm{arc}}
+B_{m}((\boldsymbol{\eta},\boldsymbol{\omega}),(\boldsymbol{v},\boldsymbol{w}))
+B_{b}((\boldsymbol{\eta},\boldsymbol{\omega}),(\boldsymbol{v},\boldsymbol{w}))\\
=\int_{\widehat{\Omega}_{f}}\boldsymbol{F}_{f}\cdot\widetilde{\boldsymbol{v}}\,dx
-\int_{\widehat{\Gamma}^{\mathrm{in}}_{f}}p_{\mathrm{in}}\boldsymbol{n}_{f}\cdot\widetilde{\boldsymbol{v}}\,dS_{0}
-\int_{\widehat{\Gamma}^{\mathrm{out}}_{f}}p_{\mathrm{out}}\boldsymbol{n}_{f}\cdot\widetilde{\boldsymbol{v}}\,dS_{0}. \tag*{(92)}
\end{aligned} ρ f ∫ Ω f ∂ t u ⋅ v d x + 2 μ f ∫ Ω f D ( u ) : D ( v ) d x − ∫ Ω f p ∇ ⋅ v d x + ∫ Ω f q ∇ ⋅ u d x + ε ρ s ∫ ω ∂ t η ˙ ⋅ v d x d y arc + 12 ε 3 ρ s ∫ ω ∂ t ω ˙ ⋅ w d x d y arc + B m (( η , ω ) , ( v , w )) + B b (( η , ω ) , ( v , w )) = ∫ Ω f F f ⋅ v d x − ∫ Γ f in p in n f ⋅ v d S 0 − ∫ Γ f out p out n f ⋅ v d S 0 . (92)
Interface work cancels for these compatible tests. The prescribed inlet and outlet tractions still contribute external work.
Time-semidiscrete Robin-Robin splitting scheme in the fixed domain
Let Δ t = T / N \Delta t=T/N Δ t = T / N , t n = n Δ t t^{n}=n\Delta t t n = n Δ t , and fix γ > 0 \gamma>0 γ > 0 . The fields remain continuous in space. To start the recursion, take u 0 ∈ V f \boldsymbol{u}^{0}\in V_{f} u 0 ∈ V f , ( η 0 , ω 0 ) ∈ V N ( ω ) (\boldsymbol{\eta}^{0},\boldsymbol{\omega}^{0})\in V_{N}(\omega) ( η 0 , ω 0 ) ∈ V N ( ω ) , η ˙ 0 , ω ˙ 0 ∈ L 2 ( ω ; R 3 ) \dot{\boldsymbol{\eta}}^{0},\dot{\boldsymbol{\omega}}^{0}\in L^{2}(\omega;\mathbb{R}^{3}) η ˙ 0 , ω ˙ 0 ∈ L 2 ( ω ; R 3 ) , and t ^ f 0 , f s r e f , 0 ∈ L 2 ( ω ; R 3 ) \widehat{\boldsymbol{t}}_{f}^{0},\boldsymbol{f}_{s}^{\mathrm{ref},0}\in L^{2}(\omega;\mathbb{R}^{3}) t f 0 , f s ref , 0 ∈ L 2 ( ω ; R 3 ) . Exact initialization also satisfies the continuous compatibility conditions. At each step we solve the following two problems with independent test functions, without imposing equality of their traces.
Fluid subproblem. Find ( u n + 1 , p n + 1 ) ∈ V f × Q f (\boldsymbol{u}^{n+1},p^{n+1})\in V_{f}\times Q_{f} ( u n + 1 , p n + 1 ) ∈ V f × Q f such that for every ( v ~ , q ) ∈ V f × Q f (\widetilde{\boldsymbol{v}},q)\in V_{f}\times Q_{f} ( v , q ) ∈ V f × Q f ,
ρ f ∫ Ω ^ f u n + 1 − u n Δ t ⋅ v ~ d x + 2 μ f ∫ Ω ^ f D ( u n + 1 ) : D ( v ~ ) d x − ∫ Ω ^ f p n + 1 ∇ ⋅ v ~ d x + ∫ Ω ^ f q ∇ ⋅ u n + 1 d x + γ ∫ ω ( u n + 1 ∘ φ ) ⋅ ( v ~ ∘ φ ) d x d y a r c = ∫ Ω ^ f F f n + 1 ⋅ v ~ d x − ∫ Γ ^ f i n p i n n + 1 n f ⋅ v ~ d S 0 − ∫ Γ ^ f o u t p o u t n + 1 n f ⋅ v ~ d S 0 + ∫ ω [ γ 2 ( u n ∘ φ + η ˙ n ) + 1 2 ( t ^ f n − f s r e f , n ) ] ⋅ ( v ~ ∘ φ ) d x d y a r c . (93) \begin{aligned}
\rho_{f}\int_{\widehat{\Omega}_{f}}\frac{\boldsymbol{u}^{n+1}-\boldsymbol{u}^{n}}{\Delta t}\cdot\widetilde{\boldsymbol{v}}\,dx
+2\mu_{f}\int_{\widehat{\Omega}_{f}}\mathbb{D}(\boldsymbol{u}^{n+1}):\mathbb{D}(\widetilde{\boldsymbol{v}})\,dx
-\int_{\widehat{\Omega}_{f}}p^{n+1}\nabla\cdot\widetilde{\boldsymbol{v}}\,dx
+\int_{\widehat{\Omega}_{f}}q\nabla\cdot\boldsymbol{u}^{n+1}\,dx\\
+\gamma\int_{\omega}(\boldsymbol{u}^{n+1}\circ\varphi)\cdot(\widetilde{\boldsymbol{v}}\circ\varphi)\,dx\,dy_{\mathrm{arc}}\\
=\int_{\widehat{\Omega}_{f}}\boldsymbol{F}_{f}^{n+1}\cdot\widetilde{\boldsymbol{v}}\,dx
-\int_{\widehat{\Gamma}^{\mathrm{in}}_{f}}p_{\mathrm{in}}^{n+1}\boldsymbol{n}_{f}\cdot\widetilde{\boldsymbol{v}}\,dS_{0}
-\int_{\widehat{\Gamma}^{\mathrm{out}}_{f}}p_{\mathrm{out}}^{n+1}\boldsymbol{n}_{f}\cdot\widetilde{\boldsymbol{v}}\,dS_{0}\\
+\int_{\omega}\left[\frac{\gamma}{2}\left(\boldsymbol{u}^{n}\circ\varphi+\dot{\boldsymbol{\eta}}^{n}\right)+\frac{1}{2}\left(\widehat{\boldsymbol{t}}_{f}^{n}-\boldsymbol{f}_{s}^{\mathrm{ref},n}\right)\right]\cdot(\widetilde{\boldsymbol{v}}\circ\varphi)\,dx\,dy_{\mathrm{arc}}. \tag*{(93)}
\end{aligned} ρ f ∫ Ω f Δ t u n + 1 − u n ⋅ v d x + 2 μ f ∫ Ω f D ( u n + 1 ) : D ( v ) d x − ∫ Ω f p n + 1 ∇ ⋅ v d x + ∫ Ω f q ∇ ⋅ u n + 1 d x + γ ∫ ω ( u n + 1 ∘ φ ) ⋅ ( v ∘ φ ) d x d y arc = ∫ Ω f F f n + 1 ⋅ v d x − ∫ Γ f in p in n + 1 n f ⋅ v d S 0 − ∫ Γ f out p out n + 1 n f ⋅ v d S 0 + ∫ ω [ 2 γ ( u n ∘ φ + η ˙ n ) + 2 1 ( t f n − f s ref , n ) ] ⋅ ( v ∘ φ ) d x d y arc . (93)
The fluid load is updated by
t ^ f n + 1 = γ 2 ( u n ∘ φ + η ˙ n ) + 1 2 ( t ^ f n − f s r e f , n ) − γ u n + 1 ∘ φ . (94) \widehat{\mathbf{t}}_{f}^{n+1}=\frac{\gamma}{2}\left(\mathbf{u}^{n}\circ\varphi+\dot{\boldsymbol{\eta}}^{n}\right)+\frac{1}{2}\left(\widehat{\mathbf{t}}_{f}^{n}-\mathbf{f}_{s}^{\mathrm{ref},n}\right)-\gamma\mathbf{u}^{n+1}\circ\varphi. \tag*{(94)} t f n + 1 = 2 γ ( u n ∘ φ + η ˙ n ) + 2 1 ( t f n − f s ref , n ) − γ u n + 1 ∘ φ . (94)
Shell subproblem. Find ( η n + 1 , ω n + 1 ) (\boldsymbol{\eta}^{n+1},\boldsymbol{\omega}^{n+1}) ( η n + 1 , ω n + 1 ) and ( η ˙ n + 1 , ω ˙ n + 1 ) (\dot{\boldsymbol{\eta}}^{n+1},\dot{\boldsymbol{\omega}}^{n+1}) ( η ˙ n + 1 , ω ˙ n + 1 ) in V N ( ω ) V_{N}(\omega) V N ( ω ) satisfying
η n + 1 − η n Δ t = η ˙ n + 1 , ω n + 1 − ω n Δ t = ω ˙ n + 1 , (95) \frac{\boldsymbol{\eta}^{n+1}-\boldsymbol{\eta}^{n}}{\Delta t}=\dot{\boldsymbol{\eta}}^{n+1},\qquad\frac{\boldsymbol{\omega}^{n+1}-\boldsymbol{\omega}^{n}}{\Delta t}=\dot{\boldsymbol{\omega}}^{n+1}, \tag*{(95)} Δ t η n + 1 − η n = η ˙ n + 1 , Δ t ω n + 1 − ω n = ω ˙ n + 1 , (95)
and, for all ( v , w ) ∈ V N ( ω ) (\mathbf{v},\mathbf{w})\in V_{N}(\omega) ( v , w ) ∈ V N ( ω ) ,
ε ρ s Δ t ∫ ω η ˙ n + 1 − η ˙ n Δ t ⋅ v d x d y a r c + ε 3 ρ s 12 ∫ ω ω ˙ n + 1 − ω ˙ n Δ t ⋅ w d x d y a r c + B m ( ( η n + 1 , ω n + 1 ) , ( v , w ) ) + B b ( ( η n + 1 , ω n + 1 ) , ( v , w ) ) + γ ∫ ω η ˙ n + 1 ⋅ v d x d y a r c = ∫ ω [ γ 2 ( u n ∘ φ + η ˙ n ) − 1 2 ( t ^ f n − f s r e f , n ) ] ⋅ v d x d y a r c . (96) \begin{aligned}
\frac{\varepsilon\rho_{s}}{\Delta t}\int_{\omega}\frac{\dot{\boldsymbol{\eta}}^{n+1}-\dot{\boldsymbol{\eta}}^{n}}{\Delta t}\cdot\mathbf{v}\,dx\,dy_{\mathrm{arc}}+\frac{\varepsilon^{3}\rho_{s}}{12}\int_{\omega}\frac{\dot{\boldsymbol{\omega}}^{n+1}-\dot{\boldsymbol{\omega}}^{n}}{\Delta t}\cdot\mathbf{w}\,dx\,dy_{\mathrm{arc}}+B_{m}((\boldsymbol{\eta}^{n+1},\boldsymbol{\omega}^{n+1}),(\mathbf{v},\mathbf{w}))+B_{b}((\boldsymbol{\eta}^{n+1},\boldsymbol{\omega}^{n+1}),(\mathbf{v},\mathbf{w}))+\gamma\int_{\omega}\dot{\boldsymbol{\eta}}^{n+1}\cdot\mathbf{v}\,dx\,dy_{\mathrm{arc}} \\
&=\int_{\omega}\left[\frac{\gamma}{2}(\mathbf{u}^{n}\circ\varphi+\dot{\boldsymbol{\eta}}^{n})-\frac{1}{2}(\widehat{\mathbf{t}}_{f}^{n}-\mathbf{f}_{s}^{\mathrm{ref},n})\right]\cdot\mathbf{v}\,dx\,dy_{\mathrm{arc}}. \tag*{(96)}
\end{aligned} Δ t ε ρ s ∫ ω Δ t η ˙ n + 1 − η ˙ n ⋅ v d x d y arc + 12 ε 3 ρ s ∫ ω Δ t ω ˙ n + 1 − ω ˙ n ⋅ w d x d y arc + B m (( η n + 1 , ω n + 1 ) , ( v , w )) + B b (( η n + 1 , ω n + 1 ) , ( v , w )) + γ ∫ ω η ˙ n + 1 ⋅ v d x d y arc = ∫ ω [ 2 γ ( u n ∘ φ + η ˙ n ) − 2 1 ( t f n − f s ref , n ) ] ⋅ v d x d y arc . (96)
The shell load is updated by
f s r e f , n + 1 = γ 2 ( u n ∘ φ + η ˙ n ) − 1 2 ( t ^ f n − f s r e f , n ) − γ η ˙ n + 1 . (97) \mathbf{f}_{s}^{\mathrm{ref},n+1}=\frac{\gamma}{2}(\mathbf{u}^{n}\circ\varphi+\dot{\boldsymbol{\eta}}^{n})-\frac{1}{2}(\widehat{\mathbf{t}}_{f}^{n}-\mathbf{f}_{s}^{\mathrm{ref},n})-\gamma\dot{\boldsymbol{\eta}}^{n+1}. \tag*{(97)} f s ref , n + 1 = 2 γ ( u n ∘ φ + η ˙ n ) − 2 1 ( t f n − f s ref , n ) − γ η ˙ n + 1 . (97)
Both right-hand sides of (93) and (96) use only old interface data, so the solves can run concurrently. Each branch then recovers its own load for the next step. The trace theorem and the recovery relations ensure that the loads remain in L 2 ( ω ; R 3 ) L^{2}(\omega;\mathbb{R}^{3}) L 2 ( ω ; R 3 ) . For smooth solutions, the recovered fluid load agrees with (77) . For weaker solutions, it is defined through the Robin condition.
Solvability of the subproblems
For fixed old data, the fluid subproblem has the symmetric velocity form
ρ f Δ t ∥ v ~ ∥ L 2 ( Ω ^ f ) 2 + 2 μ f ∥ D ( v ~ ) ∥ L 2 ( Ω ^ f ) 2 + γ ∥ v ~ ∘ φ ∥ L 2 ( ω ) 2 . (98) \frac{\rho_{f}}{\Delta t}\|\widetilde{\mathbf{v}}\|_{L^{2}(\widehat{\Omega}_{f})}^{2}+2\mu_{f}\|\mathbb{D}(\widetilde{\mathbf{v}})\|_{L^{2}(\widehat{\Omega}_{f})}^{2}+\gamma\|\widetilde{\mathbf{v}}\circ\varphi\|_{L^{2}(\omega)}^{2}. \tag*{(98)} Δ t ρ f ∥ v ∥ L 2 ( Ω f ) 2 + 2 μ f ∥ D ( v ) ∥ L 2 ( Ω f ) 2 + γ ∥ v ∘ φ ∥ L 2 ( ω ) 2 . (98)
Korn’s inequality with an L 2 L^{2} L 2 term gives coercivity in H 1 ( Ω ^ f ) H^{1}(\widehat{\Omega}_{f}) H 1 ( Ω f ) for every fixed Δ t > 0 \Delta t>0 Δ t > 0 . Together with (80) and continuity of the data, this gives a unique velocity and a pressure with the stated normalization.
For the shell, substitute (95) into (96) . The bilinear part for the new velocities is
ε ρ s Δ t ∥ η ˙ n + 1 ∥ L 2 ( ω ) 2 + ε 3 ρ s 12 Δ t ∥ ω ˙ n + 1 ∥ L 2 ( ω ) 2 + γ ∥ η ˙ n + 1 ∥ L 2 ( ω ) 2 + Δ t B m ( ( η ˙ n + 1 , ω ˙ n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 ) ) + Δ t B b ( ( η ˙ n + 1 , ω ˙ n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 ) ) . (99) \begin{aligned}
\frac{\varepsilon\rho_{s}}{\Delta t}\|\dot{\boldsymbol{\eta}}^{n+1}\|_{L^{2}(\omega)}^{2}+\frac{\varepsilon^{3}\rho_{s}}{12\Delta t}\|\dot{\boldsymbol{\omega}}^{n+1}\|_{L^{2}(\omega)}^{2}+\gamma\|\dot{\boldsymbol{\eta}}^{n+1}\|_{L^{2}(\omega)}^{2} \\
&+\Delta t B_{m}((\dot{\boldsymbol{\eta}}^{n+1},\dot{\boldsymbol{\omega}}^{n+1}),(\dot{\boldsymbol{\eta}}^{n+1},\dot{\boldsymbol{\omega}}^{n+1}))+\Delta t B_{b}((\dot{\boldsymbol{\eta}}^{n+1},\dot{\boldsymbol{\omega}}^{n+1}),(\dot{\boldsymbol{\eta}}^{n+1},\dot{\boldsymbol{\omega}}^{n+1})). \tag*{(99)}
\end{aligned} Δ t ε ρ s ∥ η ˙ n + 1 ∥ L 2 ( ω ) 2 + 12Δ t ε 3 ρ s ∥ ω ˙ n + 1 ∥ L 2 ( ω ) 2 + γ ∥ η ˙ n + 1 ∥ L 2 ( ω ) 2 + Δ t B m (( η ˙ n + 1 , ω ˙ n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 )) + Δ t B b (( η ˙ n + 1 , ω ˙ n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 )) . (99)
The known displacement and rotation contribute B m ( ( η n , ω n ) , ( v , w ) ) B_{m}((\boldsymbol{\eta}^{n},\boldsymbol{\omega}^{n}),(\mathbf{v},\mathbf{w})) B m (( η n , ω n ) , ( v , w )) and B b ( ( η n , ω n ) , ( v , w ) ) B_{b}((\boldsymbol{\eta}^{n},\boldsymbol{\omega}^{n}),(\mathbf{v},\mathbf{w})) B b (( η n , ω n ) , ( v , w )) to the right-hand side, with a minus sign after eliminating the kinematics. The new-velocity form combines positive translational and rotational mass terms with nonnegative elasticity. To see its coercivity, let m Δ t = min { ε ρ s / Δ t , ε 3 ρ s / ( 12 Δ t ) } > 0 m_{\Delta t}=\min\{\varepsilon\rho_{s}/\Delta t,\varepsilon^{3}\rho_{s}/(12\Delta t)\}>0 m Δ t = min { ε ρ s /Δ t , ε 3 ρ s / ( 12Δ t )} > 0 . If C N > 0 C_{N}>0 C N > 0 , the form in (99) controls the left-hand side of (81) with coefficient min { Δ t , m Δ t / C N } \min\{\Delta t,m_{\Delta t}/C_{N}\} min { Δ t , m Δ t / C N } . If C N = 0 C_{N}=0 C N = 0 , the elastic terms alone provide the same bound with coefficient Δ t \Delta t Δ t . Thus, for every fixed Δ t > 0 \Delta t>0 Δ t > 0 , the shell velocities are uniquely determined. The displacement and rotation follow from (95) . The coercivity constant may depend on Δ t \Delta t Δ t , which is sufficient for solvability at each time step.
Discrete energy stability estimate
We now prove stability of the fixed-domain scheme. In the continuous problem, interface work cancels because the velocities agree and the loads balance. The Robin-Robin splitting scheme does not enforce these conditions at the new time level. Instead, its Robin relations express the interface work through changes in the numerical interface energy and two nonnegative defect terms. Combining this relation with the fluid and shell energy balances yields the discrete energy identity.
The mechanical energy at time t n t^n t n is
E n = ρ f 2 ∥ w n ∥ L 2 ( Ω f ) 2 + ϵ ρ s 2 ∥ η ˙ n ∥ L 2 ( ω ) 2 + ϵ 3 ρ s 24 ∥ ω ˙ n ∥ L 2 ( ω ) 2 + 1 2 B m ( ( η n , ω n ) , ( η n , ω n ) ) + 1 2 B b ( ( η n , ω n ) , ( η n , ω n ) ) . (100) \mathcal{E}^{n} = \frac{\rho_f}{2}\lVert w^n\rVert_{L^2(\Omega_f)}^2 + \frac{\epsilon\rho_s}{2}\lVert\dot{\eta}^n\rVert_{L^2(\omega)}^2 + \frac{\epsilon^3\rho_s}{24}\lVert\dot{\omega}^n\rVert_{L^2(\omega)}^2
+ \frac{1}{2}B_m\bigl((\eta^n,\omega^n),(\eta^n,\omega^n)\bigr) + \frac{1}{2}B_b\bigl((\eta^n,\omega^n),(\eta^n,\omega^n)\bigr). \tag*{(100)} E n = 2 ρ f ∥ w n ∥ L 2 ( Ω f ) 2 + 2 ϵ ρ s ∥ η ˙ n ∥ L 2 ( ω ) 2 + 24 ϵ 3 ρ s ∥ ω ˙ n ∥ L 2 ( ω ) 2 + 2 1 B m ( ( η n , ω n ) , ( η n , ω n ) ) + 2 1 B b ( ( η n , ω n ) , ( η n , ω n ) ) . (100)
The shell terms include membrane-shear and bending energy, with the displacement-rotation coupling retained. The rotational kinetic coefficient ϵ 3 ρ s / 24 \epsilon^3\rho_s/24 ϵ 3 ρ s /24 is half the corresponding mass coefficient in the shell equation.
The Robin conditions introduce additional interface terms in the augmented energy
E a u g n = E n + Δ t γ 2 ( ∥ u n ∘ φ ∥ L 2 ( ω ) 2 + ∥ η ˙ n ∥ L 2 ( ω ) 2 ) + Δ t 2 γ ( ∥ t ^ f n ∥ L 2 ( ω ) 2 + ∥ f s r e f , n ∥ L 2 ( ω ) 2 ) . (101) \mathcal{E}_{\mathrm{aug}}^n = \mathcal{E}^n + \frac{\Delta t\,\gamma}{2}\left(\lVert u^n\circ\varphi\rVert_{L^2(\omega)}^2+\lVert\dot{\eta}^n\rVert_{L^2(\omega)}^2\right) + \frac{\Delta t}{2\gamma}\left(\lVert\widehat{t}_f^n\rVert_{L^2(\omega)}^2+\lVert f_s^{\mathrm{ref},n}\rVert_{L^2(\omega)}^2\right). \tag*{(101)} E aug n = E n + 2 Δ t γ ( ∥ u n ∘ φ ∥ L 2 ( ω ) 2 + ∥ η ˙ n ∥ L 2 ( ω ) 2 ) + 2 γ Δ t ( ∥ t f n ∥ L 2 ( ω ) 2 + ∥ f s ref , n ∥ L 2 ( ω ) 2 ) . (101)
These terms account for the Robin interface exchange and are distinct from the mechanical energy.
Discrete energy identity
The load recovery relations (94) and (97) give
t ^ f n + 1 + γ u n + 1 ∘ φ = γ 2 ( u n ∘ φ + η ˙ n ) + 1 2 ( t ^ f n − f s r e f , n ) , \widehat{t}_f^{n+1}+\gamma u^{n+1}\circ\varphi=\frac{\gamma}{2}(u^n\circ\varphi+\dot{\eta}^n)+\frac{1}{2}(\widehat{t}_f^n-f_s^{\mathrm{ref},n}), t f n + 1 + γ u n + 1 ∘ φ = 2 γ ( u n ∘ φ + η ˙ n ) + 2 1 ( t f n − f s ref , n ) ,
f s r e f , n + 1 + γ η ˙ n + 1 = γ 2 ( u n ∘ φ + η ˙ n ) − 1 2 ( t ^ f n − f s r e f , n ) . (102) f_s^{\mathrm{ref},n+1}+\gamma\dot{\eta}^{n+1}=\frac{\gamma}{2}(u^n\circ\varphi+\dot{\eta}^n)-\frac{1}{2}(\widehat{t}_f^n-f_s^{\mathrm{ref},n}). \tag*{(102)} f s ref , n + 1 + γ η ˙ n + 1 = 2 γ ( u n ∘ φ + η ˙ n ) − 2 1 ( t f n − f s ref , n ) . (102)
Squaring and adding these relations cancels the cross terms on the right and gives, almost everywhere on ω \omega ω ,
∣ t ^ f n + 1 + γ u n + 1 ∘ φ ∣ 2 + ∣ f s r e f , n + 1 + γ η ˙ n + 1 ∣ 2 = γ 2 2 ∣ u n ∘ φ + η ˙ n ∣ 2 + 1 2 ∣ t ^ f n − f s r e f , n ∣ 2 = γ 2 ( ∣ u n ∘ φ ∣ 2 + ∣ η ˙ n ∣ 2 ) + ∣ t ^ f n ∣ 2 + ∣ f s r e f , n ∣ 2 − γ 2 2 ∣ u n ∘ φ − η ˙ n ∣ 2 − 1 2 ∣ t ^ f n + f s r e f , n ∣ 2 . (103) \begin{aligned}
\left|\widehat{t}_f^{n+1}+\gamma u^{n+1}\circ\varphi\right|^2+\left|f_s^{\mathrm{ref},n+1}+\gamma\dot{\eta}^{n+1}\right|^2
=\frac{\gamma^2}{2}\left|u^n\circ\varphi+\dot{\eta}^n\right|^2+\frac{1}{2}\left|\widehat{t}_f^n-f_s^{\mathrm{ref},n}\right|^2 \\
&=\gamma^2\left(\lvert u^n\circ\varphi\rvert^2+\lvert\dot{\eta}^n\rvert^2\right)+\lvert\widehat{t}_f^n\rvert^2+\lvert f_s^{\mathrm{ref},n}\rvert^2-\frac{\gamma^2}{2}\left|u^n\circ\varphi-\dot{\eta}^n\right|^2-\frac{1}{2}\left|\widehat{t}_f^n+f_s^{\mathrm{ref},n}\right|^2. \tag*{(103)}
\end{aligned} t f n + 1 + γ u n + 1 ∘ φ 2 + f s ref , n + 1 + γ η ˙ n + 1 2 = 2 γ 2 ∣ u n ∘ φ + η ˙ n ∣ 2 + 2 1 t f n − f s ref , n 2 = γ 2 ( ∣ u n ∘ φ ∣ 2 + ∣ η ˙ n ∣ 2 ) + ∣ t f n ∣ 2 + ∣ f s ref , n ∣ 2 − 2 γ 2 ∣ u n ∘ φ − η ˙ n ∣ 2 − 2 1 t f n + f s ref , n 2 . (103)
Expanding the left-hand side, dividing by 2 γ 2\gamma 2 γ , and integrating over the interface yields
∫ ω [ t ^ f n + 1 ⋅ ( u n + 1 ∘ φ ) + f s r e f , n + 1 ⋅ η ˙ n + 1 ] d x d y a r c = γ 2 [ ∥ u n ∘ φ ∥ L 2 ( ω ) 2 + ∥ η ˙ n ∥ L 2 ( ω ) 2 − ∥ u n + 1 ∘ φ ∥ L 2 ( ω ) 2 − ∥ η ˙ n + 1 ∥ L 2 ( ω ) 2 ] + 1 2 γ [ ∥ t ^ f n ∥ L 2 ( ω ) 2 + ∥ f s r e f , n ∥ L 2 ( ω ) 2 − ∥ t ^ f n + 1 ∥ L 2 ( ω ) 2 − ∥ f s r e f , n + 1 ∥ L 2 ( ω ) 2 ] \begin{aligned}
\int_{\omega}\left[\widehat{t}_f^{n+1}\cdot(u^{n+1}\circ\varphi)+f_s^{\mathrm{ref},n+1}\cdot\dot{\eta}^{n+1}\right]\,dx\,dy_{\mathrm{arc}} \\
&=\frac{\gamma}{2}\left[\lVert u^n\circ\varphi\rVert_{L^2(\omega)}^2+\lVert\dot{\eta}^n\rVert_{L^2(\omega)}^2-\lVert u^{n+1}\circ\varphi\rVert_{L^2(\omega)}^2-\lVert\dot{\eta}^{n+1}\rVert_{L^2(\omega)}^2\right] \\
&+\frac{1}{2\gamma}\left[\lVert\widehat{t}_f^n\rVert_{L^2(\omega)}^2+\lVert f_s^{\mathrm{ref},n}\rVert_{L^2(\omega)}^2-\lVert\widehat{t}_f^{n+1}\rVert_{L^2(\omega)}^2-\lVert f_s^{\mathrm{ref},n+1}\rVert_{L^2(\omega)}^2\right]
\end{aligned} ∫ ω [ t f n + 1 ⋅ ( u n + 1 ∘ φ ) + f s ref , n + 1 ⋅ η ˙ n + 1 ] d x d y arc = 2 γ [ ∥ u n ∘ φ ∥ L 2 ( ω ) 2 + ∥ η ˙ n ∥ L 2 ( ω ) 2 − ∥ u n + 1 ∘ φ ∥ L 2 ( ω ) 2 − ∥ η ˙ n + 1 ∥ L 2 ( ω ) 2 ] + 2 γ 1 [ ∥ t f n ∥ L 2 ( ω ) 2 + ∥ f s ref , n ∥ L 2 ( ω ) 2 − ∥ t f n + 1 ∥ L 2 ( ω ) 2 − ∥ f s ref , n + 1 ∥ L 2 ( ω ) 2 ]
− γ 4 ∥ u n ∘ φ − η ˙ n ∥ L 2 ( ω ) 2 − 1 4 γ ∥ t ^ f n + f s r e f , n ∥ L 2 ( ω ) 2 . (104) -\frac{\gamma}{4}\lVert u^n\circ\varphi-\dot{\eta}^n\rVert_{L^2(\omega)}^2-\frac{1}{4\gamma}\lVert\widehat{t}_f^n+f_s^{\mathrm{ref},n}\rVert_{L^2(\omega)}^2. \tag*{(104)} − 4 γ ∥ u n ∘ φ − η ˙ n ∥ L 2 ( ω ) 2 − 4 γ 1 ∥ t f n + f s ref , n ∥ L 2 ( ω ) 2 . (104)
The last two terms are nonpositive and measure the velocity mismatch and force imbalance at t n t^n t n , the time level used in the Robin data.
Theorem 5.1 (Exact discrete energy identity). Under the assumptions of Section 4 , suppose that the fixed-domain scheme admits a sufficiently regular solution such that the computed velocities can be used as test functions in the corresponding variational formulations. Then, for any γ > 0 \gamma> 0 γ > 0 and Δ t > 0 \Delta t > 0 Δ t > 0 , the following discrete energy identity holds at each time step:
E a u g n + 1 − E a u g n + ρ f 2 ∥ u n + 1 − u n ∥ L 2 ( Ω ^ f ) 2 + 2 μ f Δ t ∥ D ( u n + 1 ) ∥ L 2 ( Ω ^ f ) 2 + ε ρ s 2 ∥ η ˙ n + 1 − η ˙ n ∥ L 2 ( ω ) 2 + ε 3 ρ s 24 ∥ ω ˙ n + 1 − ω ˙ n ∥ L 2 ( ω ) 2 + 1 2 B m ( ( η n + 1 − η n , ω n + 1 − ω n ) , ( η n + 1 − η n , ω n + 1 − ω n ) ) + 1 2 B b ( ( η n + 1 − η n , ω n + 1 − ω n ) , ( η n + 1 − η n , ω n + 1 − ω n ) ) + Δ t γ 4 ∥ u n ∘ φ − η ˙ n ∥ L 2 ( ω ) 2 + Δ t 4 γ ∥ t ^ f n + f s r e f , n ∥ L 2 ( ω ) 2 = Δ t ∫ Ω ^ f F f n + 1 ⋅ u n + 1 d x − Δ t ∫ Γ ^ i n p i n n + 1 n f ⋅ u n + 1 d S 0 − Δ t ∫ Γ ^ o u t p o u t n + 1 n f ⋅ u n + 1 d S 0 . (105) \begin{aligned}
\mathcal{E}_{\mathrm{aug}}^{n+1}-\mathcal{E}_{\mathrm{aug}}^n
&+\frac{\rho_f}{2}\lVert u^{n+1}-u^n\rVert_{L^2(\widehat{\Omega}_f)}^2+2\mu_f\Delta t\lVert D(u^{n+1})\rVert_{L^2(\widehat{\Omega}_f)}^2\\
&+\frac{\varepsilon\rho_s}{2}\lVert\dot{\eta}^{n+1}-\dot{\eta}^n\rVert_{L^2(\omega)}^2+\frac{\varepsilon^3\rho_s}{24}\lVert\dot{\omega}^{n+1}-\dot{\omega}^n\rVert_{L^2(\omega)}^2\\
&+\frac{1}{2}B_m((\eta^{n+1}-\eta^n,\omega^{n+1}-\omega^n),(\eta^{n+1}-\eta^n,\omega^{n+1}-\omega^n))\\
&+\frac{1}{2}B_b((\eta^{n+1}-\eta^n,\omega^{n+1}-\omega^n),(\eta^{n+1}-\eta^n,\omega^{n+1}-\omega^n))\\
&+\frac{\Delta t\,\gamma}{4}\lVert u^n\circ\varphi-\dot{\eta}^n\rVert_{L^2(\omega)}^2+\frac{\Delta t}{4\gamma}\lVert\widehat{t}_f^n+f_s^{\mathrm{ref},n}\rVert_{L^2(\omega)}^2\\
&=\Delta t\int_{\widehat{\Omega}_f}F_f^{n+1}\cdot u^{n+1}\,dx-\Delta t\int_{\widehat{\Gamma}_{\mathrm{in}}}p_{\mathrm{in}}^{n+1}n_f\cdot u^{n+1}\,dS_0-\Delta t\int_{\widehat{\Gamma}_{\mathrm{out}}}p_{\mathrm{out}}^{n+1}n_f\cdot u^{n+1}\,dS_0. \tag*{(105)}
\end{aligned} E aug n + 1 − E aug n + 2 ρ f ∥ u n + 1 − u n ∥ L 2 ( Ω f ) 2 + 2 μ f Δ t ∥ D ( u n + 1 ) ∥ L 2 ( Ω f ) 2 + 2 ε ρ s ∥ η ˙ n + 1 − η ˙ n ∥ L 2 ( ω ) 2 + 24 ε 3 ρ s ∥ ω ˙ n + 1 − ω ˙ n ∥ L 2 ( ω ) 2 + 2 1 B m (( η n + 1 − η n , ω n + 1 − ω n ) , ( η n + 1 − η n , ω n + 1 − ω n )) + 2 1 B b (( η n + 1 − η n , ω n + 1 − ω n ) , ( η n + 1 − η n , ω n + 1 − ω n )) + 4 Δ t γ ∥ u n ∘ φ − η ˙ n ∥ L 2 ( ω ) 2 + 4 γ Δ t ∥ t f n + f s ref , n ∥ L 2 ( ω ) 2 = Δ t ∫ Ω f F f n + 1 ⋅ u n + 1 d x − Δ t ∫ Γ in p in n + 1 n f ⋅ u n + 1 d S 0 − Δ t ∫ Γ out p out n + 1 n f ⋅ u n + 1 d S 0 . (105)
Proof. Testing the fluid equation with ( v ~ , q ) = ( u n + 1 , p n + 1 ) (\widetilde{v},q)=(u^{n+1},p^{n+1}) ( v , q ) = ( u n + 1 , p n + 1 ) cancels the pressure terms. Using the recovered fluid load to express the interface work gives
ρ f ∫ Ω ^ f u n + 1 − u n Δ t ⋅ u n + 1 d x + 2 μ f ∥ D ( u n + 1 ) ∥ L 2 ( Ω ^ f ) 2 = ∫ ω t ^ f n + 1 ⋅ ( u n + 1 ∘ φ ) d x d y a r c + ∫ Ω ^ f F f n + 1 ⋅ u n + 1 d x − ∫ Γ ^ i n p i n n + 1 n f ⋅ u n + 1 d S 0 − ∫ Γ ^ o u t p o u t n + 1 n f ⋅ u n + 1 d S 0 . (106) \begin{aligned}
\rho_f\int_{\widehat{\Omega}_f}\frac{u^{n+1}-u^n}{\Delta t}\cdot u^{n+1}\,dx+2\mu_f\lVert D(u^{n+1})\rVert_{L^2(\widehat{\Omega}_f)}^2
&=\int_{\omega}\widehat{t}_f^{n+1}\cdot(u^{n+1}\circ\varphi)\,dx\,dy_{\mathrm{arc}}+\int_{\widehat{\Omega}_f}F_f^{n+1}\cdot u^{n+1}\,dx\\
&\quad-\int_{\widehat{\Gamma}_{\mathrm{in}}}p_{\mathrm{in}}^{n+1}n_f\cdot u^{n+1}\,dS_0-\int_{\widehat{\Gamma}_{\mathrm{out}}}p_{\mathrm{out}}^{n+1}n_f\cdot u^{n+1}\,dS_0. \tag*{(106)}
\end{aligned} ρ f ∫ Ω f Δ t u n + 1 − u n ⋅ u n + 1 d x + 2 μ f ∥ D ( u n + 1 ) ∥ L 2 ( Ω f ) 2 = ∫ ω t f n + 1 ⋅ ( u n + 1 ∘ φ ) d x d y arc + ∫ Ω f F f n + 1 ⋅ u n + 1 d x − ∫ Γ in p in n + 1 n f ⋅ u n + 1 d S 0 − ∫ Γ out p out n + 1 n f ⋅ u n + 1 d S 0 . (106)
By the polarization identity
2 ( a − b ) ⋅ a = ∣ a ∣ 2 − ∣ b ∣ 2 + ∣ a − b ∣ 2 , (107) 2(a-b)\cdot a=\lvert a\rvert^2-\lvert b\rvert^2+\lvert a-b\rvert^2, \tag*{(107)} 2 ( a − b ) ⋅ a = ∣ a ∣ 2 − ∣ b ∣ 2 + ∣ a − b ∣ 2 , (107)
the first term becomes the fluid kinetic-energy difference divided by Δ t \Delta t Δ t , plus the nonnegative increment ρ f ∥ u n + 1 − u n ∥ L 2 ( Ω ^ f ) 2 / ( 2 Δ t ) \rho_f\lVert u^{n+1}-u^n\rVert_{L^2(\widehat{\Omega}_f)}^2/(2\Delta t) ρ f ∥ u n + 1 − u n ∥ L 2 ( Ω f ) 2 / ( 2Δ t ) .
For the shell, choose ( v , w ) = ( η ˙ n + 1 , ω ˙ n + 1 ) (v,w)=(\dot{\eta}^{n+1},\dot{\omega}^{n+1}) ( v , w ) = ( η ˙ n + 1 , ω ˙ n + 1 ) and use the recovered shell load. Then
ε ρ s ∫ ω η ˙ n + 1 − η ˙ n Δ t ⋅ η ˙ n + 1 d x d y a r c + ε 3 ρ s 12 ∫ ω ω ˙ n + 1 − ω ˙ n Δ t ⋅ ω ˙ n + 1 d x d y a r c + B m ( ( η n + 1 , ω n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 ) ) + B b ( ( η n + 1 , ω n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 ) ) = ∫ ω f s r e f , n + 1 ⋅ η ˙ n + 1 d x d y a r c . (108) \begin{aligned}
\varepsilon\rho_s\int_{\omega}\frac{\dot{\eta}^{n+1}-\dot{\eta}^n}{\Delta t}\cdot\dot{\eta}^{n+1}\,dx\,dy_{\mathrm{arc}}
+\frac{\varepsilon^3\rho_s}{12}\int_{\omega}\frac{\dot{\omega}^{n+1}-\dot{\omega}^n}{\Delta t}\cdot\dot{\omega}^{n+1}\,dx\,dy_{\mathrm{arc}}\\
+B_m((\eta^{n+1},\omega^{n+1}),(\dot{\eta}^{n+1},\dot{\omega}^{n+1}))
+B_b((\eta^{n+1},\omega^{n+1}),(\dot{\eta}^{n+1},\dot{\omega}^{n+1}))\\
=\int_{\omega}f_s^{\mathrm{ref},n+1}\cdot\dot{\eta}^{n+1}\,dx\,dy_{\mathrm{arc}}. \tag*{(108)}
\end{aligned} ε ρ s ∫ ω Δ t η ˙ n + 1 − η ˙ n ⋅ η ˙ n + 1 d x d y arc + 12 ε 3 ρ s ∫ ω Δ t ω ˙ n + 1 − ω ˙ n ⋅ ω ˙ n + 1 d x d y arc + B m (( η n + 1 , ω n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 )) + B b (( η n + 1 , ω n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 )) = ∫ ω f s ref , n + 1 ⋅ η ˙ n + 1 d x d y arc . (108)
The same identity applies to both inertial terms, with kinetic-energy coefficients ε ρ s / 2 \varepsilon\rho_s/2 ε ρ s /2 and ε 3 ρ s / 24 \varepsilon^3\rho_s/24 ε 3 ρ s /24 . For the elastic terms, the discrete kinematic relations express the new velocities through displacement and rotation differences.
Using symmetry of B m B_m B m , we obtain
2 Δ t B m ( ( η n + 1 , ω n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 ) ) = B m ( ( η n + 1 , ω n + 1 ) , ( η n + 1 , ω n + 1 ) ) − B m ( ( η n , ω n ) , ( η n , ω n ) ) + B m ( ( η n + 1 − η n , ω n + 1 − ω n ) , ( η n + 1 − η n , ω n + 1 − ω n ) ) , (109) \begin{aligned}
2\Delta t B_m\left((\eta^{n+1},\omega^{n+1}),(\dot{\eta}^{n+1},\dot{\omega}^{n+1})\right)
={}&B_m\left((\eta^{n+1},\omega^{n+1}),(\eta^{n+1},\omega^{n+1})\right)\\
&-B_m\left((\eta^n,\omega^n),(\eta^n,\omega^n)\right)\\
&+B_m\left((\eta^{n+1}-\eta^n,\omega^{n+1}-\omega^n),(\eta^{n+1}-\eta^n,\omega^{n+1}-\omega^n)\right), \tag*{(109)}
\end{aligned} 2Δ t B m ( ( η n + 1 , ω n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 ) ) = B m ( ( η n + 1 , ω n + 1 ) , ( η n + 1 , ω n + 1 ) ) − B m ( ( η n , ω n ) , ( η n , ω n ) ) + B m ( ( η n + 1 − η n , ω n + 1 − ω n ) , ( η n + 1 − η n , ω n + 1 − ω n ) ) , (109)
The same calculation for B b B_b B b gives
2 Δ t B b ( ( η n + 1 , ω n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 ) ) = B b ( ( η n + 1 , ω n + 1 ) , ( η n + 1 , ω n + 1 ) ) − B b ( ( η n , ω n ) , ( η n , ω n ) ) + B b ( ( η n + 1 − η n , ω n + 1 − ω n ) , ( η n + 1 − η n , ω n + 1 − ω n ) ) . (110) \begin{aligned}
2\Delta t B_b\left((\eta^{n+1},\omega^{n+1}),(\dot{\eta}^{n+1},\dot{\omega}^{n+1})\right)
={}&B_b\left((\eta^{n+1},\omega^{n+1}),(\eta^{n+1},\omega^{n+1})\right)\\
&-B_b\left((\eta^n,\omega^n),(\eta^n,\omega^n)\right)\\
&+B_b\left((\eta^{n+1}-\eta^n,\omega^{n+1}-\omega^n),(\eta^{n+1}-\eta^n,\omega^{n+1}-\omega^n)\right). \tag*{(110)}
\end{aligned} 2Δ t B b ( ( η n + 1 , ω n + 1 ) , ( η ˙ n + 1 , ω ˙ n + 1 ) ) = B b ( ( η n + 1 , ω n + 1 ) , ( η n + 1 , ω n + 1 ) ) − B b ( ( η n , ω n ) , ( η n , ω n ) ) + B b ( ( η n + 1 − η n , ω n + 1 − ω n ) , ( η n + 1 − η n , ω n + 1 − ω n ) ) . (110)
Both identities act on the displacement-rotation pair and retain all elastic cross terms.
Multiply the tested fluid and shell equations by Δ t \Delta t Δ t and add them. Their energy differences give E n + 1 − E n \mathcal{E}^{n+1}-\mathcal{E}^n E n + 1 − E n , while the remaining interface work is Δ t \Delta t Δ t times the left-hand side of (104) . Substituting this identity gives the augmented-energy difference and the two nonnegative defect terms, proving (105) .
Corollary 5.1 (Unconditional stability under homogeneous data). Suppose that F f n + 1 = 0 F_f^{n+1}=0 F f n + 1 = 0 , p i n n + 1 = 0 p_{\mathrm{in}}^{n+1}=0 p in n + 1 = 0 , and p o u t n + 1 = 0 p_{\mathrm{out}}^{n+1}=0 p out n + 1 = 0 for all time steps. Then the augmented energy satisfies
E a u g n + 1 ≤ E a u g n ≤ E a u g 0 . (111) \mathcal{E}_{\mathrm{aug}}^{n+1}\leq\mathcal{E}_{\mathrm{aug}}^n\leq\mathcal{E}_{\mathrm{aug}}^0. \tag*{(111)} E aug n + 1 ≤ E aug n ≤ E aug 0 . (111) Moreover, summing the discrete energy inequality over n = 0 , … , N − 1 n=0,\ldots,N-1 n = 0 , … , N − 1 and discarding the nonnegative kinetic and elastic increment terms yields
E a u g N + 2 μ f Δ t ∑ n = 0 N − 1 ∥ D ( u n + 1 ) ∥ L 2 ( Ω f ) 2 + Δ t γ 4 ∑ n = 0 N − 1 ∥ u n ∘ φ − η ˙ n ∥ L 2 ( ω ) 2 + Δ t 4 γ ∑ n = 0 N − 1 ∥ f ^ f n + f s r e f , n ∥ L 2 ( ω ) 2 ≤ E a u g 0 . (112) \begin{aligned}
\mathcal{E}_{\mathrm{aug}}^N
&+2\mu_f\Delta t\sum_{n=0}^{N-1}\|\mathbb{D}(u^{n+1})\|_{L^2(\Omega_f)}^2\\
&+\frac{\Delta t\,\gamma}{4}\sum_{n=0}^{N-1}\|u^{n}\circ\varphi-\dot{\eta}^{n}\|_{L^2(\omega)}^2
+\frac{\Delta t}{4\gamma}\sum_{n=0}^{N-1}\|\widehat{f}_f^n+f_s^{\mathrm{ref},n}\|_{L^2(\omega)}^2
\leq\mathcal{E}_{\mathrm{aug}}^0. \tag*{(112)}
\end{aligned} E aug N + 2 μ f Δ t n = 0 ∑ N − 1 ∥ D ( u n + 1 ) ∥ L 2 ( Ω f ) 2 + 4 Δ t γ n = 0 ∑ N − 1 ∥ u n ∘ φ − η ˙ n ∥ L 2 ( ω ) 2 + 4 γ Δ t n = 0 ∑ N − 1 ∥ f f n + f s ref , n ∥ L 2 ( ω ) 2 ≤ E aug 0 . (112) Consequently, for any Δ t > 0 \Delta t>0 Δ t > 0 and fixed γ > 0 \gamma>0 γ > 0 , the scheme is stable without a time-step restriction from the added-mass effect or the fluid-to-structure density ratio. Since E n ≤ E a u g n \mathcal{E}^n\leq\mathcal{E}_{\mathrm{aug}}^n E n ≤ E aug n , the mechanical energy is bounded but need not decrease at every step. The cumulative estimate also controls the interface defects. Their behavior under refinement is addressed in § 6 and § 7 .
External work and driven stability For nonzero external data, we estimate the work on the right of (105) using F f n + 1 = F f ( t n + 1 ) F_f^{n+1}=F_f(t^{n+1}) F f n + 1 = F f ( t n + 1 ) and p i n / o u t n + 1 = p i n / o u t ( t n + 1 ) p_{\mathrm{in/out}}^{n+1}=p_{\mathrm{in/out}}(t^{n+1}) p in/out n + 1 = p in/out ( t n + 1 ) . These samples must be well defined, with bounded discrete sums of squares as Δ t \Delta t Δ t decreases. Continuity in time with values in the stated spatial L 2 L^2 L 2 spaces is sufficient. An L 2 L^2 L 2 assumption in time alone does not justify point evaluation. To estimate the external work, we use the Korn inequality appropriate to the fluid test space. We consider separately the cases with and without an L 2 L^2 L 2 velocity term.
Korn inequality without an L 2 L^2 L 2 term. Suppose first that the fluid test space satisfies
∥ v ~ ∥ H 1 ( Ω ^ f ) ≤ C K ∥ D ( v ~ ) ∥ L 2 ( Ω ^ f ) . (113) \|\widetilde{v}\|_{H^1(\widehat{\Omega}_f)} \le C_K \|D(\widetilde{v})\|_{L^2(\widehat{\Omega}_f)}. \tag*{(113)} ∥ v ∥ H 1 ( Ω f ) ≤ C K ∥ D ( v ) ∥ L 2 ( Ω f ) . (113)
This holds, for example, when the homogeneous Dirichlet boundary has positive surface measure. Let C i n C_{\mathrm{in}} C in and C o u t C_{\mathrm{out}} C out be trace constants such that
∥ v ~ ∥ L 2 ( Γ ^ i n ) ≤ C i n ∥ v ~ ∥ H 1 ( Ω ^ f ) , \|\widetilde{v}\|_{L^2(\widehat{\Gamma}_{\mathrm{in}})} \le C_{\mathrm{in}}\|\widetilde{v}\|_{H^1(\widehat{\Omega}_f)}, ∥ v ∥ L 2 ( Γ in ) ≤ C in ∥ v ∥ H 1 ( Ω f ) ,
∥ v ~ ∥ L 2 ( Γ ^ o u t ) ≤ C o u t ∥ v ~ ∥ H 1 ( Ω ^ f ) . (114) \|\widetilde{v}\|_{L^2(\widehat{\Gamma}_{\mathrm{out}})} \le C_{\mathrm{out}}\|\widetilde{v}\|_{H^1(\widehat{\Omega}_f)}. \tag*{(114)} ∥ v ∥ L 2 ( Γ out ) ≤ C out ∥ v ∥ H 1 ( Ω f ) . (114)
The Cauchy-Schwarz and trace inequalities, followed by Young’s inequality, give
∣ ∫ Ω ^ f F f n + 1 ⋅ u n + 1 d x − ∫ Γ ^ i n p i n n + 1 n f ⋅ u n + 1 d S 0 − ∫ Γ ^ o u t p o u t n + 1 n f ⋅ u n + 1 d S 0 ∣ ≤ μ f ∥ D ( u n + 1 ) ∥ L 2 ( Ω ^ f ) 2 + C K 2 4 μ f ( ∥ F f n + 1 ∥ L 2 ( Ω ^ f ) + C i n ∥ p i n n + 1 ∥ L 2 ( Γ ^ i n ) + C o u t ∥ p o u t n + 1 ∥ L 2 ( Γ ^ o u t ) ) 2 . (115) \left|\int_{\widehat{\Omega}_f}\mathbf{F}_f^{n+1}\cdot\mathbf{u}^{n+1}\,dx-\int_{\widehat{\Gamma}_{\mathrm{in}}}p_{\mathrm{in}}^{n+1}\mathbf{n}_f\cdot\mathbf{u}^{n+1}\,dS_0-\int_{\widehat{\Gamma}_{\mathrm{out}}}p_{\mathrm{out}}^{n+1}\mathbf{n}_f\cdot\mathbf{u}^{n+1}\,dS_0\right|
\le\mu_f\|D(\mathbf{u}^{n+1})\|_{L^2(\widehat{\Omega}_f)}^2+\frac{C_K^2}{4\mu_f}\left(\|\mathbf{F}_f^{n+1}\|_{L^2(\widehat{\Omega}_f)}+C_{\mathrm{in}}\|p_{\mathrm{in}}^{n+1}\|_{L^2(\widehat{\Gamma}_{\mathrm{in}})}+C_{\mathrm{out}}\|p_{\mathrm{out}}^{n+1}\|_{L^2(\widehat{\Gamma}_{\mathrm{out}})}\right)^2. \tag*{(115)} ∫ Ω f F f n + 1 ⋅ u n + 1 d x − ∫ Γ in p in n + 1 n f ⋅ u n + 1 d S 0 − ∫ Γ out p out n + 1 n f ⋅ u n + 1 d S 0 ≤ μ f ∥ D ( u n + 1 ) ∥ L 2 ( Ω f ) 2 + 4 μ f C K 2 ( ∥ F f n + 1 ∥ L 2 ( Ω f ) + C in ∥ p in n + 1 ∥ L 2 ( Γ in ) + C out ∥ p out n + 1 ∥ L 2 ( Γ out ) ) 2 . (115)
The first term is absorbed into the viscous dissipation. Summing the resulting inequality and discarding the remaining nonnegative terms gives
E a u g N + μ f Δ t ∑ n = 0 N − 1 ∥ D ( u n + 1 ) ∥ L 2 ( Ω ^ f ) 2 ≤ E a u g 0 + C K 2 Δ t 4 μ f ∑ n = 0 N − 1 ( ∥ F f n + 1 ∥ L 2 ( Ω ^ f ) + C i n ∥ p i n n + 1 ∥ L 2 ( Γ ^ i n ) + C o u t ∥ p o u t n + 1 ∥ L 2 ( Γ ^ o u t ) ) 2 . (116) \mathcal{E}_{\mathrm{aug}}^N+\mu_f\Delta t\sum_{n=0}^{N-1}\|D(\mathbf{u}^{n+1})\|_{L^2(\widehat{\Omega}_f)}^2
\le\mathcal{E}_{\mathrm{aug}}^0+\frac{C_K^2\Delta t}{4\mu_f}\sum_{n=0}^{N-1}\left(\|\mathbf{F}_f^{n+1}\|_{L^2(\widehat{\Omega}_f)}+C_{\mathrm{in}}\|p_{\mathrm{in}}^{n+1}\|_{L^2(\widehat{\Gamma}_{\mathrm{in}})}+C_{\mathrm{out}}\|p_{\mathrm{out}}^{n+1}\|_{L^2(\widehat{\Gamma}_{\mathrm{out}})}\right)^2. \tag*{(116)} E aug N + μ f Δ t n = 0 ∑ N − 1 ∥ D ( u n + 1 ) ∥ L 2 ( Ω f ) 2 ≤ E aug 0 + 4 μ f C K 2 Δ t n = 0 ∑ N − 1 ( ∥ F f n + 1 ∥ L 2 ( Ω f ) + C in ∥ p in n + 1 ∥ L 2 ( Γ in ) + C out ∥ p out n + 1 ∥ L 2 ( Γ out ) ) 2 . (116)
This estimate holds for every positive time step.
Korn inequality with an L 2 L^2 L 2 term. The tube benchmark has an elastic lateral wall and traction conditions at both ends. Its fluid test space has no Dirichlet portion of positive surface measure, since clamping the shell at the end rings does not impose a fluid Dirichlet condition. We therefore use
∥ v ~ ∥ H 1 ( Ω ^ f ) ≤ C K ( ∥ D ( v ~ ) ∥ L 2 ( Ω ^ f ) + ∥ v ~ ∥ L 2 ( Ω ^ f ) ) . (117) \|\widetilde{v}\|_{H^1(\widehat{\Omega}_f)}\le C_K\left(\|D(\widetilde{v})\|_{L^2(\widehat{\Omega}_f)}+\|\widetilde{v}\|_{L^2(\widehat{\Omega}_f)}\right). \tag*{(117)} ∥ v ∥ H 1 ( Ω f ) ≤ C K ( ∥ D ( v ) ∥ L 2 ( Ω f ) + ∥ v ∥ L 2 ( Ω f ) ) . (117)
Applying Young’s inequality to the symmetric-gradient and velocity terms separately gives the bound
μ f ∥ D ( u n + 1 ) ∥ L 2 ( Ω ^ f ) 2 + ρ f 2 T ∥ u n + 1 ∥ L 2 ( Ω ^ f ) 2 + C K 2 ( 1 4 μ f + T 2 ρ f ) ( ∥ F f n + 1 ∥ L 2 ( Ω ^ f ) + C i n ∥ p i n n + 1 ∥ L 2 ( Γ ^ i n ) + C o u t ∥ p o u t n + 1 ∥ L 2 ( Γ ^ o u t ) ) 2 . (118) \mu_f\|D(\mathbf{u}^{n+1})\|_{L^2(\widehat{\Omega}_f)}^2+\frac{\rho_f}{2T}\|\mathbf{u}^{n+1}\|_{L^2(\widehat{\Omega}_f)}^2+C_K^2\left(\frac{1}{4\mu_f}+\frac{T}{2\rho_f}\right)\left(\|\mathbf{F}_f^{n+1}\|_{L^2(\widehat{\Omega}_f)}+C_{\mathrm{in}}\|p_{\mathrm{in}}^{n+1}\|_{L^2(\widehat{\Gamma}_{\mathrm{in}})}+C_{\mathrm{out}}\|p_{\mathrm{out}}^{n+1}\|_{L^2(\widehat{\Gamma}_{\mathrm{out}})}\right)^2. \tag*{(118)} μ f ∥ D ( u n + 1 ) ∥ L 2 ( Ω f ) 2 + 2 T ρ f ∥ u n + 1 ∥ L 2 ( Ω f ) 2 + C K 2 ( 4 μ f 1 + 2 ρ f T ) ( ∥ F f n + 1 ∥ L 2 ( Ω f ) + C in ∥ p in n + 1 ∥ L 2 ( Γ in ) + C out ∥ p out n + 1 ∥ L 2 ( Γ out ) ) 2 . (118)
Since ρ f ∥ u n + 1 ∥ 2 / ( 2 T ) ≤ E a u g n + 1 / T \rho_f\|\mathbf{u}^{n+1}\|^2/(2T)\le\mathcal{E}_{\mathrm{aug}}^{n+1}/T ρ f ∥ u n + 1 ∥ 2 / ( 2 T ) ≤ E aug n + 1 / T , the energy identity implies
( 1 − Δ t T ) E a u g n + 1 + μ f Δ t ∥ D ( u n + 1 ) ∥ L 2 ( Ω ^ f ) 2 ≤ E a u g n + C K 2 Δ t ( 1 4 μ f + T 2 ρ f ) ( ∥ F f n + 1 ∥ L 2 ( Ω ^ f ) + C i n ∥ p i n n + 1 ∥ L 2 ( Γ ^ i n ) + C o u t ∥ p o u t n + 1 ∥ L 2 ( Γ ^ o u t ) ) 2 . (119) \left(1-\frac{\Delta t}{T}\right)\mathcal{E}_{\mathrm{aug}}^{n+1}+\mu_f\Delta t\|D(\mathbf{u}^{n+1})\|_{L^2(\widehat{\Omega}_f)}^2
\le\mathcal{E}_{\mathrm{aug}}^n+C_K^2\Delta t\left(\frac{1}{4\mu_f}+\frac{T}{2\rho_f}\right)\left(\|\mathbf{F}_f^{n+1}\|_{L^2(\widehat{\Omega}_f)}+C_{\mathrm{in}}\|p_{\mathrm{in}}^{n+1}\|_{L^2(\widehat{\Gamma}_{\mathrm{in}})}+C_{\mathrm{out}}\|p_{\mathrm{out}}^{n+1}\|_{L^2(\widehat{\Gamma}_{\mathrm{out}})}\right)^2. \tag*{(119)} ( 1 − T Δ t ) E aug n + 1 + μ f Δ t ∥ D ( u n + 1 ) ∥ L 2 ( Ω f ) 2 ≤ E aug n + C K 2 Δ t ( 4 μ f 1 + 2 ρ f T ) ( ∥ F f n + 1 ∥ L 2 ( Ω f ) + C in ∥ p in n + 1 ∥ L 2 ( Γ in ) + C out ∥ p out n + 1 ∥ L 2 ( Γ out ) ) 2 . (119)
For Δ t ≤ T / 2 \Delta t \le T/2 Δ t ≤ T /2 , the inequality − log ( 1 − Δ t / T ) ≤ 2 Δ t / T -\log(1-\Delta t/T) \le2\Delta t/T − log ( 1 − Δ t / T ) ≤ 2Δ t / T bounds the product of the recurrence factors by exp ( 2 ) \exp(2) exp ( 2 ) over N N N steps. Iteration bounds the maximum energy, and summation then gives
max 0 ≤ n ≤ N E a u g n + μ f Δ t ∑ n = 0 N − 1 ∥ D ( u n + 1 ) ∥ L 2 ( Ω ^ f ) 2 ≤ C [ E a u g 0 + C K 2 Δ t ( 1 4 μ f + T 2 ρ f ) ∑ n = 0 N − 1 ( ∥ F f n + 1 ∥ L 2 ( Ω ^ f ) + C i n ∥ p i n n + 1 ∥ L 2 ( Γ ^ f i n ) + C o u t ∥ p o u t n + 1 ∥ L 2 ( Γ ^ f o u t ) ) 2 ] . (120) \max_{0\le n\le N}\mathcal{E}_{\mathrm{aug}}^{n}+\mu_f\Delta t\sum_{n=0}^{N-1}\left\|\mathsf{D}(u^{n+1})\right\|_{L^2(\widehat{\Omega}_f)}^2
\le C\left[\mathcal{E}_{\mathrm{aug}}^0+C_K^2\Delta t\left(\frac{1}{4\mu_f}+\frac{T}{2\rho_f}\right)\sum_{n=0}^{N-1}\left(\left\|\boldsymbol{F}_f^{n+1}\right\|_{L^2(\widehat{\Omega}_f)}+C_{\mathrm{in}}\left\|p_{\mathrm{in}}^{n+1}\right\|_{L^2(\widehat{\Gamma}_f^{\mathrm{in}})}+C_{\mathrm{out}}\left\|p_{\mathrm{out}}^{n+1}\right\|_{L^2(\widehat{\Gamma}_f^{\mathrm{out}})}\right)^2\right]. \tag*{(120)} 0 ≤ n ≤ N max E aug n + μ f Δ t n = 0 ∑ N − 1 D ( u n + 1 ) L 2 ( Ω f ) 2 ≤ C [ E aug 0 + C K 2 Δ t ( 4 μ f 1 + 2 ρ f T ) n = 0 ∑ N − 1 ( F f n + 1 L 2 ( Ω f ) + C in p in n + 1 L 2 ( Γ f in ) + C out p out n + 1 L 2 ( Γ f out ) ) 2 ] . (120)
Here C C C is independent of Δ t \Delta t Δ t . One may take C = 1 + 2 exp ( 2 ) C=1+2\exp(2) C = 1 + 2 exp ( 2 ) , since the maximum energy is bounded by exp ( 2 ) \exp(2) exp ( 2 ) times the bracketed quantity and the viscous sum by that quantity plus the maximum energy. The condition Δ t ≤ T / 2 \Delta t\le T/2 Δ t ≤ T /2 applies only to this driven-data estimate and is independent of mesh size, density ratio, and shell stiffness. The homogeneous energy identity has no such restriction.
Consistency analysis
To estimate consistency, we insert the exact solution into the discrete equations. Backward differences introduce residuals in the fluid and shell equations and in both shell kinematic relations. The Robin conditions based on previous data also introduce two interface residuals. In the following section, we will estimate each of these terms before deriving the error equations.
We keep γ > 0 \gamma>0 γ > 0 fixed, and evaluate the external data at t n + 1 t^{n+1} t n + 1 : F f n + 1 = F f ( t n + 1 ) \boldsymbol{F}_f^{n+1}=\boldsymbol{F}_f(t^{n+1}) F f n + 1 = F f ( t n + 1 ) , p i n n + 1 = p i n ( t n + 1 ) p_{\mathrm{in}}^{n+1}=p_{\mathrm{in}}(t^{n+1}) p in n + 1 = p in ( t n + 1 ) , and p o u t n + 1 = p o u t ( t n + 1 ) p_{\mathrm{out}}^{n+1}=p_{\mathrm{out}}(t^{n+1}) p out n + 1 = p out ( t n + 1 ) . Here, the analysis is time-semidiscrete, so no spatial projection or interpolation is involved. We denote the exact solution at time n n n as u ( t n ) u(t^n) u ( t n ) and numerical values as u n u^n u n . The exact solution satisfies the equations in § 4 , including
u ( t ) ∘ φ = η ˙ ( t ) , f s r e f ( t ) = − t ^ f ( t ) . (121) u(t)\circ\varphi=\dot{\eta}(t),\qquad\boldsymbol{f}^{\mathrm{ref}}_s(t)=-\widehat{\boldsymbol{t}}_f(t). \tag*{(121)} u ( t ) ∘ φ = η ˙ ( t ) , f s ref ( t ) = − t f ( t ) . (121)
For the bounds below, assume
u ∈ H 2 ( 0 , T ; L 2 ( Ω ^ f ; R 3 ) ) ∩ C ( [ 0 , T ] ; V f ) , ( η , ω ) ∈ H 2 ( 0 , T ; V N ( ω ) ) , u\in H^2(0,T;L^2(\widehat{\Omega}_f;\mathbb{R}^3))\cap C([0,T];V_f),\qquad(\eta,\omega)\in H^2(0,T;V_N(\omega)), u ∈ H 2 ( 0 , T ; L 2 ( Ω f ; R 3 )) ∩ C ([ 0 , T ] ; V f ) , ( η , ω ) ∈ H 2 ( 0 , T ; V N ( ω )) ,
η ˙ , ω ˙ ∈ H 2 ( 0 , T ; L 2 ( ω ; R 3 ) ) , t ^ f ∈ H 1 ( 0 , T ; L 2 ( ω ; R 3 ) ) . (122) \dot{\eta},\dot{\omega}\in H^2(0,T;L^2(\omega;\mathbb{R}^3)),\qquad\widehat{\boldsymbol{t}}_f\in H^1(0,T;L^2(\omega;\mathbb{R}^3)). \tag*{(122)} η ˙ , ω ˙ ∈ H 2 ( 0 , T ; L 2 ( ω ; R 3 )) , t f ∈ H 1 ( 0 , T ; L 2 ( ω ; R 3 )) . (122)
Here η ˙ = ∂ t η \dot{\eta}=\partial_t\eta η ˙ = ∂ t η and ω ˙ = ∂ t ω \dot{\omega}=\partial_t\omega ω ˙ = ∂ t ω . Thus the regularity in η ˙ , ω ˙ \dot{\eta},\dot{\omega} η ˙ , ω ˙ include third order time derivatives of displacement and rotation in the L 2 L^2 L 2 spaces. We also assume the following continuity in time so that the pressure and external data can be evaluated at the discrete time levels:
p ∈ C ( [ 0 , T ] ; Q f ) , F f ∈ C ( [ 0 , T ] ; L 2 ( Ω ^ f ; R 3 ) ) , p\in C([0,T];Q_f),\qquad\boldsymbol{F}_f\in C([0,T];L^2(\widehat{\Omega}_f;\mathbb{R}^3)), p ∈ C ([ 0 , T ] ; Q f ) , F f ∈ C ([ 0 , T ] ; L 2 ( Ω f ; R 3 )) ,
p i n ∈ C ( [ 0 , T ] ; L 2 ( Γ ^ f i n ) ) , p o u t ∈ C ( [ 0 , T ] ; L 2 ( Γ ^ f o u t ) ) . (123) p_{\mathrm{in}}\in C([0,T];L^2(\widehat{\Gamma}_f^{\mathrm{in}})),\qquad p_{\mathrm{out}}\in C([0,T];L^2(\widehat{\Gamma}_f^{\mathrm{out}})). \tag*{(123)} p in ∈ C ([ 0 , T ] ; L 2 ( Γ f in )) , p out ∈ C ([ 0 , T ] ; L 2 ( Γ f out )) . (123)
Together with (6.2), these assumptions ensure that the weak equations, initially satisfied almost everywhere in time, can be evaluated at each t n t^n t n . The interface traction requires a separate regularity assumption, since the continuity of pressure in Q f Q_f Q f does not guarantee an L 2 L^2 L 2 boundary traction. We estimate the consistency residuals in discrete L 2 L^2 L 2 time norms, for which the assumed square-integrability of the corresponding time derivatives is sufficient.
Interface residuals We define each residual as the discrete left-hand side minus the right-hand side, evaluated at the exact solution. For the fluid Robin condition, this gives
R f n + 1 : = t ^ f ( t n + 1 ) + γ u ( t n + 1 ) ∘ φ − γ 2 ( u ( t n ) ∘ φ + η ˙ ( t n ) ) − 1 2 ( t ^ f ( t n ) − f s r e f ( t n ) ) . (124) \mathcal{R}_{f}^{n+1} := \hat{t}_{f}(t^{n+1}) + \gamma u(t^{n+1}) \circ\varphi- \frac{\gamma}{2}\left(u(t^{n}) \circ\varphi+ \dot{\eta}(t^{n})\right) - \frac{1}{2}\left(\hat{t}_{f}(t^{n}) - f_{s}^{\mathrm{ref}}(t^{n})\right). \tag*{(124)} R f n + 1 := t ^ f ( t n + 1 ) + γ u ( t n + 1 ) ∘ φ − 2 γ ( u ( t n ) ∘ φ + η ˙ ( t n ) ) − 2 1 ( t ^ f ( t n ) − f s ref ( t n ) ) . (124)
Using exact velocity continuity and force balance from (121) , we simplify this expression to
R f n + 1 = γ ( η ˙ ( t n + 1 ) − η ˙ ( t n ) ) + t ^ f ( t n + 1 ) − t ^ f ( t n ) = ∫ t n t n + 1 ( γ ∂ t η ˙ ( t ) + ∂ t t ^ f ( t ) ) d t . (125) \mathcal{R}_{f}^{n+1} = \gamma\left(\dot{\eta}(t^{n+1})-\dot{\eta}(t^{n})\right)+\hat{t}_{f}(t^{n+1})-\hat{t}_{f}(t^{n})=\int_{t^{n}}^{t^{n+1}}\left(\gamma\partial_{t}\dot{\eta}(t)+\partial_{t}\hat{t}_{f}(t)\right)\,\mathrm{d}t. \tag*{(125)} R f n + 1 = γ ( η ˙ ( t n + 1 ) − η ˙ ( t n ) ) + t ^ f ( t n + 1 ) − t ^ f ( t n ) = ∫ t n t n + 1 ( γ ∂ t η ˙ ( t ) + ∂ t t ^ f ( t ) ) d t . (125)
For the shell Robin condition, the same sign convention gives
R s n + 1 : = f s r e f ( t n + 1 ) + γ η ˙ ( t n + 1 ) − γ 2 ( u ( t n ) ∘ φ + η ˙ ( t n ) ) + 1 2 ( t ^ f ( t n ) − f s r e f ( t n ) ) . (126) \mathcal{R}_{s}^{n+1} := f_{s}^{\mathrm{ref}}(t^{n+1})+\gamma\dot{\eta}(t^{n+1})-\frac{\gamma}{2}\left(u(t^{n})\circ\varphi+\dot{\eta}(t^{n})\right)+\frac{1}{2}\left(\hat{t}_{f}(t^{n})-f_{s}^{\mathrm{ref}}(t^{n})\right). \tag*{(126)} R s n + 1 := f s ref ( t n + 1 ) + γ η ˙ ( t n + 1 ) − 2 γ ( u ( t n ) ∘ φ + η ˙ ( t n ) ) + 2 1 ( t ^ f ( t n ) − f s ref ( t n ) ) . (126)
It follows that
R s n + 1 = γ ( η ˙ ( t n + 1 ) − η ˙ ( t n ) ) − t ^ f ( t n + 1 ) + t ^ f ( t n ) = ∫ t n t n + 1 ( γ ∂ t η ˙ ( t ) − ∂ t t ^ f ( t ) ) d t . (127) \mathcal{R}_{s}^{n+1}=\gamma\left(\dot{\eta}(t^{n+1})-\dot{\eta}(t^{n})\right)-\hat{t}_{f}(t^{n+1})+\hat{t}_{f}(t^{n})=\int_{t^{n}}^{t^{n+1}}\left(\gamma\partial_{t}\dot{\eta}(t)-\partial_{t}\hat{t}_{f}(t)\right)\,\mathrm{d}t. \tag*{(127)} R s n + 1 = γ ( η ˙ ( t n + 1 ) − η ˙ ( t n ) ) − t ^ f ( t n + 1 ) + t ^ f ( t n ) = ∫ t n t n + 1 ( γ ∂ t η ˙ ( t ) − ∂ t t ^ f ( t ) ) d t . (127)
The two residuals contain the same acceleration term and opposite traction increments. Squaring and adding cancels the cross terms. Applying the Cauchy-Schwarz in time gives
∥ R f n + 1 ∥ L 2 ( ω ) 2 + ∥ R s n + 1 ∥ L 2 ( ω ) 2 = 2 γ 2 ∥ η ˙ ( t n + 1 ) − η ˙ ( t n ) ∥ L 2 ( ω ) 2 + 2 ∥ t ^ f ( t n + 1 ) − t ^ f ( t n ) ∥ L 2 ( ω ) 2 ≤ 2 Δ t ∫ t n t n + 1 [ γ 2 ∥ ∂ t η ˙ ( t ) ∥ L 2 ( ω ) 2 + ∥ ∂ t t ^ f ( t ) ∥ L 2 ( ω ) 2 ] d t . (128) \left\|\mathcal{R}_{f}^{n+1}\right\|_{L^{2}(\omega)}^{2}+\left\|\mathcal{R}_{s}^{n+1}\right\|_{L^{2}(\omega)}^{2}=2\gamma^{2}\left\|\dot{\eta}(t^{n+1})-\dot{\eta}(t^{n})\right\|_{L^{2}(\omega)}^{2}+2\left\|\hat{t}_{f}(t^{n+1})-\hat{t}_{f}(t^{n})\right\|_{L^{2}(\omega)}^{2}
\leq2\Delta t\int_{t^{n}}^{t^{n+1}}\left[\gamma^{2}\left\|\partial_{t}\dot{\eta}(t)\right\|_{L^{2}(\omega)}^{2}+\left\|\partial_{t}\hat{t}_{f}(t)\right\|_{L^{2}(\omega)}^{2}\right]\,\mathrm{d}t. \tag*{(128)} R f n + 1 L 2 ( ω ) 2 + R s n + 1 L 2 ( ω ) 2 = 2 γ 2 η ˙ ( t n + 1 ) − η ˙ ( t n ) L 2 ( ω ) 2 + 2 t ^ f ( t n + 1 ) − t ^ f ( t n ) L 2 ( ω ) 2 ≤ 2Δ t ∫ t n t n + 1 [ γ 2 ∥ ∂ t η ˙ ( t ) ∥ L 2 ( ω ) 2 + ∂ t t ^ f ( t ) L 2 ( ω ) 2 ] d t . (128)
Multiplication by Δ t \Delta t Δ t and summation over all intervals yield
Δ t ∑ n = 0 N − 1 ( ∥ R f n + 1 ∥ L 2 ( ω ) 2 + ∥ R s n + 1 ∥ L 2 ( ω ) 2 ) ≤ 2 Δ t 2 ∫ 0 T [ γ 2 ∥ ∂ t η ˙ ( t ) ∥ L 2 ( ω ) 2 + ∥ ∂ t t ^ f ( t ) ∥ L 2 ( ω ) 2 ] d t . (129) \Delta t\sum_{n=0}^{N-1}\left(\left\|\mathcal{R}_{f}^{n+1}\right\|_{L^{2}(\omega)}^{2}+\left\|\mathcal{R}_{s}^{n+1}\right\|_{L^{2}(\omega)}^{2}\right)\leq2\Delta t^{2}\int_{0}^{T}\left[\gamma^{2}\left\|\partial_{t}\dot{\eta}(t)\right\|_{L^{2}(\omega)}^{2}+\left\|\partial_{t}\hat{t}_{f}(t)\right\|_{L^{2}(\omega)}^{2}\right]\,\mathrm{d}t. \tag*{(129)} Δ t n = 0 ∑ N − 1 ( R f n + 1 L 2 ( ω ) 2 + R s n + 1 L 2 ( ω ) 2 ) ≤ 2Δ t 2 ∫ 0 T [ γ 2 ∥ ∂ t η ˙ ( t ) ∥ L 2 ( ω ) 2 + ∂ t t ^ f ( t ) L 2 ( ω ) 2 ] d t . (129)
Momentum residuals The backward-Euler residuals from the momentum equations satisfying that
τ u n + 1 = u ( t n + 1 ) − u ( t n ) Δ t − ∂ t u ( t n + 1 ) = − 1 Δ t ∫ t n t n + 1 ( t − t n ) ∂ t t u ( t ) d t , \tau_{u}^{n+1}=\frac{u(t^{n+1})-u(t^{n})}{\Delta t}-\partial_{t}u(t^{n+1})=-\frac{1}{\Delta t}\int_{t^{n}}^{t^{n+1}}(t-t^{n})\partial_{tt}u(t)\,\mathrm{d}t, τ u n + 1 = Δ t u ( t n + 1 ) − u ( t n ) − ∂ t u ( t n + 1 ) = − Δ t 1 ∫ t n t n + 1 ( t − t n ) ∂ tt u ( t ) d t ,
τ η n + 1 = η ˙ ( t n + 1 ) − η ˙ ( t n ) Δ t − ∂ t η ˙ ( t n + 1 ) = − 1 Δ t ∫ t n t n + 1 ( t − t n ) ∂ t t η ˙ ( t ) d t , \tau_{\eta}^{n+1}=\frac{\dot{\eta}(t^{n+1})-\dot{\eta}(t^{n})}{\Delta t}-\partial_{t}\dot{\eta}(t^{n+1})=-\frac{1}{\Delta t}\int_{t^{n}}^{t^{n+1}}(t-t^{n})\partial_{tt}\dot{\eta}(t)\,\mathrm{d}t, τ η n + 1 = Δ t η ˙ ( t n + 1 ) − η ˙ ( t n ) − ∂ t η ˙ ( t n + 1 ) = − Δ t 1 ∫ t n t n + 1 ( t − t n ) ∂ tt η ˙ ( t ) d t ,
τ ω n + 1 = ω ˙ ( t n + 1 ) − ω ˙ ( t n ) Δ t − ∂ t ω ˙ ( t n + 1 ) = − 1 Δ t ∫ t n t n + 1 ( t − t n ) ∂ t t ω ˙ ( t ) d t . (130) \tau_{\omega}^{n+1}=\frac{\dot{\omega}(t^{n+1})-\dot{\omega}(t^{n})}{\Delta t}-\partial_{t}\dot{\omega}(t^{n+1})=-\frac{1}{\Delta t}\int_{t^{n}}^{t^{n+1}}(t-t^{n})\partial_{tt}\dot{\omega}(t)\,\mathrm{d}t. \tag*{(130)} τ ω n + 1 = Δ t ω ˙ ( t n + 1 ) − ω ˙ ( t n ) − ∂ t ω ˙ ( t n + 1 ) = − Δ t 1 ∫ t n t n + 1 ( t − t n ) ∂ tt ω ˙ ( t ) d t . (130)
Since ∫ t n t n + 1 ( t − t n ) 2 d t = Δ t 3 / 3 \int_{t^{n}}^{t^{n+1}}(t-t^{n})^{2}\,\mathrm{d}t=\Delta t^{3}/3 ∫ t n t n + 1 ( t − t n ) 2 d t = Δ t 3 /3 , we have:
∥ τ u n + 1 ∥ L 2 ( Ω ^ f ) 2 ≤ Δ t 3 ∫ t n t n + 1 ∥ ∂ t t u ( t ) ∥ L 2 ( Ω ^ f ) 2 d t , \left\|\tau_{u}^{n+1}\right\|_{L^{2}(\widehat{\Omega}_{f})}^{2}\leq\frac{\Delta t}{3}\int_{t^{n}}^{t^{n+1}}\left\|\partial_{tt}u(t)\right\|_{L^{2}(\widehat{\Omega}_{f})}^{2}\,\mathrm{d}t, τ u n + 1 L 2 ( Ω f ) 2 ≤ 3 Δ t ∫ t n t n + 1 ∥ ∂ tt u ( t ) ∥ L 2 ( Ω f ) 2 d t ,
∥ τ η n + 1 ∥ L 2 ( ω ) 2 ≤ Δ t 3 ∫ t n t n + 1 ∥ ∂ t t η ˙ ( t ) ∥ L 2 ( ω ) 2 d t , \left\|\tau_{\eta}^{n+1}\right\|_{L^{2}(\omega)}^{2}\leq\frac{\Delta t}{3}\int_{t^{n}}^{t^{n+1}}\left\|\partial_{tt}\dot{\eta}(t)\right\|_{L^{2}(\omega)}^{2}\,\mathrm{d}t, τ η n + 1 L 2 ( ω ) 2 ≤ 3 Δ t ∫ t n t n + 1 ∥ ∂ tt η ˙ ( t ) ∥ L 2 ( ω ) 2 d t ,
∥ τ ω n + 1 ∥ L 2 ( ω ) 2 ≤ Δ t 3 ∫ t n t n + 1 ∥ ∂ t t ω ˙ ( t ) ∥ L 2 ( ω ) 2 d t . (131) \left\|\tau_{\omega}^{n+1}\right\|_{L^{2}(\omega)}^{2}\leq\frac{\Delta t}{3}\int_{t^{n}}^{t^{n+1}}\left\|\partial_{tt}\dot{\omega}(t)\right\|_{L^{2}(\omega)}^{2}\,\mathrm{d}t. \tag*{(131)} τ ω n + 1 L 2 ( ω ) 2 ≤ 3 Δ t ∫ t n t n + 1 ∥ ∂ tt ω ˙ ( t ) ∥ L 2 ( ω ) 2 d t . (131)
Multiplying by Δ t \Delta t Δ t and summing gives
Δ t ∑ n = 0 N − 1 ∥ τ u n + 1 ∥ L 2 ( Ω ^ f ) 2 ≤ Δ t 2 3 ∫ 0 T ∥ ∂ t t u ( t ) ∥ L 2 ( Ω ^ f ) 2 d t , Δ t ∑ n = 0 N − 1 ∥ τ η n + 1 ∥ L 2 ( ω ) 2 ≤ Δ t 2 3 ∫ 0 T ∥ ∂ t t η ˙ ( t ) ∥ L 2 ( ω ) 2 d t , Δ t ∑ n = 0 N − 1 ∥ τ ω n + 1 ∥ L 2 ( ω ) 2 ≤ Δ t 2 3 ∫ 0 T ∥ ∂ t t ω ˙ ( t ) ∥ L 2 ( ω ) 2 d t . (132) \begin{aligned}
\Delta t \sum_{n=0}^{N-1} \left\|\tau_u^{n+1}\right\|_{L^2(\widehat{\Omega}_f)}^2 &\le\frac{\Delta t^2}{3}\int_0^T \left\|\partial_{tt}u(t)\right\|_{L^2(\widehat{\Omega}_f)}^2\,dt,\\
\Delta t \sum_{n=0}^{N-1} \left\|\tau_\eta^{n+1}\right\|_{L^2(\omega)}^2 &\le\frac{\Delta t^2}{3}\int_0^T \left\|\partial_{tt}\dot{\eta}(t)\right\|_{L^2(\omega)}^2\,dt,\\
\Delta t \sum_{n=0}^{N-1} \left\|\tau_\omega^{n+1}\right\|_{L^2(\omega)}^2 &\le\frac{\Delta t^2}{3}\int_0^T \left\|\partial_{tt}\dot{\omega}(t)\right\|_{L^2(\omega)}^2\,dt. \tag*{(132)}
\end{aligned} Δ t n = 0 ∑ N − 1 τ u n + 1 L 2 ( Ω f ) 2 Δ t n = 0 ∑ N − 1 τ η n + 1 L 2 ( ω ) 2 Δ t n = 0 ∑ N − 1 τ ω n + 1 L 2 ( ω ) 2 ≤ 3 Δ t 2 ∫ 0 T ∥ ∂ tt u ( t ) ∥ L 2 ( Ω f ) 2 d t , ≤ 3 Δ t 2 ∫ 0 T ∥ ∂ tt η ˙ ( t ) ∥ L 2 ( ω ) 2 d t , ≤ 3 Δ t 2 ∫ 0 T ∥ ∂ tt ω ˙ ( t ) ∥ L 2 ( ω ) 2 d t . (132)
Kinematic residuals in the shell energy. The displacement and rotation updates have residuals as well. Define
κ η n + 1 = η ( t n + 1 ) − η ( t n ) Δ t − η ˙ ( t n + 1 ) , κ ω n + 1 = ω ( t n + 1 ) − ω ( t n ) Δ t − ω ˙ ( t n + 1 ) . (133) \kappa_\eta^{n+1}=\frac{\eta(t^{n+1})-\eta(t^n)}{\Delta t}-\dot{\eta}(t^{n+1}),\qquad
\kappa_\omega^{n+1}=\frac{\omega(t^{n+1})-\omega(t^n)}{\Delta t}-\dot{\omega}(t^{n+1}). \tag*{(133)} κ η n + 1 = Δ t η ( t n + 1 ) − η ( t n ) − η ˙ ( t n + 1 ) , κ ω n + 1 = Δ t ω ( t n + 1 ) − ω ( t n ) − ω ˙ ( t n + 1 ) . (133)
Their exact representations are
κ η n + 1 = − 1 Δ t ∫ t n t n + 1 ( t − t n ) ∂ t t η ( t ) d t , κ ω n + 1 = − 1 Δ t ∫ t n t n + 1 ( t − t n ) ∂ t t ω ( t ) d t . (134) \kappa_\eta^{n+1}=-\frac{1}{\Delta t}\int_{t^n}^{t^{n+1}}(t-t^n)\partial_{tt}\eta(t)\,dt,\qquad
\kappa_\omega^{n+1}=-\frac{1}{\Delta t}\int_{t^n}^{t^{n+1}}(t-t^n)\partial_{tt}\omega(t)\,dt. \tag*{(134)} κ η n + 1 = − Δ t 1 ∫ t n t n + 1 ( t − t n ) ∂ tt η ( t ) d t , κ ω n + 1 = − Δ t 1 ∫ t n t n + 1 ( t − t n ) ∂ tt ω ( t ) d t . (134)
In particular,
η ( t n + 1 ) − η ( t n ) Δ t = η ˙ ( t n + 1 ) + κ η n + 1 , ω ( t n + 1 ) − ω ( t n ) Δ t = ω ˙ ( t n + 1 ) + κ ω n + 1 . (135) \frac{\eta(t^{n+1})-\eta(t^n)}{\Delta t}=\dot{\eta}(t^{n+1})+\kappa_\eta^{n+1},\qquad
\frac{\omega(t^{n+1})-\omega(t^n)}{\Delta t}=\dot{\omega}(t^{n+1})+\kappa_\omega^{n+1}. \tag*{(135)} Δ t η ( t n + 1 ) − η ( t n ) = η ˙ ( t n + 1 ) + κ η n + 1 , Δ t ω ( t n + 1 ) − ω ( t n ) = ω ˙ ( t n + 1 ) + κ ω n + 1 . (135)
The exact kinematics relate velocities to time derivatives, whereas the scheme uses backward differences. This is why the two residuals in (135) remain even for the exact solution.
Cauchy-Schwarz holds for each symmetric nonnegative elastic form, including when it defines only a seminorm. Applying it to (134) , followed by Cauchy–Schwarz in time, gives
B m ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ≤ Δ t 3 ∫ t n t n + 1 B m ( ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) , ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) ) d t . (136) B_m\left((\kappa_\eta^{n+1},\kappa_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right)
\le\frac{\Delta t}{3}\int_{t^n}^{t^{n+1}}B_m\left((\partial_{tt}\eta(t),\partial_{tt}\omega(t)),(\partial_{tt}\eta(t),\partial_{tt}\omega(t))\right)\,dt. \tag*{(136)} B m ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ≤ 3 Δ t ∫ t n t n + 1 B m ( ( ∂ tt η ( t ) , ∂ tt ω ( t )) , ( ∂ tt η ( t ) , ∂ tt ω ( t )) ) d t . (136)
B b ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ≤ Δ t 3 ∫ t n t n + 1 B b ( ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) , ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) ) d t . (137) B_b\left((\kappa_\eta^{n+1},\kappa_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right)
\le\frac{\Delta t}{3}\int_{t^n}^{t^{n+1}}B_b\left((\partial_{tt}\eta(t),\partial_{tt}\omega(t)),(\partial_{tt}\eta(t),\partial_{tt}\omega(t))\right)\,dt. \tag*{(137)} B b ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ≤ 3 Δ t ∫ t n t n + 1 B b ( ( ∂ tt η ( t ) , ∂ tt ω ( t )) , ( ∂ tt η ( t ) , ∂ tt ω ( t )) ) d t . (137)
Consequently,
Δ t ∑ n = 0 N − 1 B m ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ≤ Δ t 2 3 ∫ 0 T B m ( ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) , ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) ) d t . (138) \Delta t\sum_{n=0}^{N-1}B_m\left((\kappa_\eta^{n+1},\kappa_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right)
\le\frac{\Delta t^2}{3}\int_0^T B_m\left((\partial_{tt}\eta(t),\partial_{tt}\omega(t)),(\partial_{tt}\eta(t),\partial_{tt}\omega(t))\right)\,dt. \tag*{(138)} Δ t n = 0 ∑ N − 1 B m ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ≤ 3 Δ t 2 ∫ 0 T B m ( ( ∂ tt η ( t ) , ∂ tt ω ( t )) , ( ∂ tt η ( t ) , ∂ tt ω ( t )) ) d t . (138)
and
Δ t ∑ n = 0 N − 1 B b ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ≤ Δ t 2 3 ∫ 0 T B b ( ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) , ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) ) d t . (139) \Delta t\sum_{n=0}^{N-1}B_b\left((\kappa_\eta^{n+1},\kappa_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right)
\le\frac{\Delta t^2}{3}\int_0^T B_b\left((\partial_{tt}\eta(t),\partial_{tt}\omega(t)),(\partial_{tt}\eta(t),\partial_{tt}\omega(t))\right)\,dt. \tag*{(139)} Δ t n = 0 ∑ N − 1 B b ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ≤ 3 Δ t 2 ∫ 0 T B b ( ( ∂ tt η ( t ) , ∂ tt ω ( t )) , ( ∂ tt η ( t ) , ∂ tt ω ( t )) ) d t . (139)
Residual equations and first-order consistency We can now write the exact solution in the same form as the numerical scheme, with the residuals added. For every ( v ~ , q ) ∈ V f × Q f (\tilde{v},q)\in V_f\times Q_f ( v ~ , q ) ∈ V f × Q f , we obtain
ρ f ∫ Ω ^ f u ( t n + 1 ) − u ( t n ) Δ t ⋅ v ~ d x + 2 μ f ∫ Ω ^ f D ( u ( t n + 1 ) ) : D ( v ~ ) d x \rho_f\int_{\widehat{\Omega}_f}\frac{u(t^{n+1})-u(t^n)}{\Delta t}\cdot\tilde{v}\,dx+2\mu_f\int_{\widehat{\Omega}_f}\mathbb{D}(u(t^{n+1})):\mathbb{D}(\tilde{v})\,dx ρ f ∫ Ω f Δ t u ( t n + 1 ) − u ( t n ) ⋅ v ~ d x + 2 μ f ∫ Ω f D ( u ( t n + 1 )) : D ( v ~ ) d x
− ∫ Ω ^ f p ( t n + 1 ) ∇ ⋅ v ~ d x + ∫ Ω ^ f q ∇ ⋅ u ( t n + 1 ) d x + γ ∫ ω ( u ( t n + 1 ) ∘ φ ) ⋅ ( v ~ ∘ φ ) d x d y a r c = ∫ Ω ^ f F f ( t n + 1 ) ⋅ v ~ d x − ∫ Γ ^ i n f p i n ( t n + 1 ) n f ⋅ v ~ d S 0 − ∫ Γ ^ o u t f p o u t ( t n + 1 ) n f ⋅ v ~ d S 0 + ∫ ω [ γ 2 ( u ( t n ) ∘ φ + η ˙ ( t n ) ) + 1 2 ( t ^ f ( t n ) − f s r e f ( t n ) ) ] ⋅ ( v ~ ∘ φ ) d x d y a r c + ρ f ∫ Ω ^ f τ u n + 1 ⋅ v ~ d x + ∫ ω R f n + 1 ⋅ ( v ~ ∘ φ ) d x d y a r c . (140) \begin{aligned}
-\int_{\widehat{\Omega}_f}p(t^{n+1})\nabla\cdot\tilde{v}\,dx+\int_{\widehat{\Omega}_f}q\nabla\cdot u(t^{n+1})\,dx+\gamma\int_{\omega}(u(t^{n+1})\circ\varphi)\cdot(\tilde{v}\circ\varphi)\,dx\,dy_{\mathrm{arc}} \\
&=\int_{\widehat{\Omega}_f}F_f(t^{n+1})\cdot\tilde{v}\,dx-\int_{\widehat{\Gamma}_{\mathrm{in}}^f}p_{\mathrm{in}}(t^{n+1})n_f\cdot\tilde{v}\,dS_0-\int_{\widehat{\Gamma}_{\mathrm{out}}^f}p_{\mathrm{out}}(t^{n+1})n_f\cdot\tilde{v}\,dS_0 \\
&+\int_{\omega}\left[\frac{\gamma}{2}\left(u(t^n)\circ\varphi+\dot{\eta}(t^n)\right)+\frac{1}{2}\left(\hat{t}_f(t^n)-f_s^{\mathrm{ref}}(t^n)\right)\right]\cdot(\tilde{v}\circ\varphi)\,dx\,dy_{\mathrm{arc}} \\
&+\rho_f\int_{\widehat{\Omega}_f}\tau_u^{n+1}\cdot\tilde{v}\,dx+\int_{\omega}\mathcal{R}_f^{n+1}\cdot(\tilde{v}\circ\varphi)\,dx\,dy_{\mathrm{arc}}. \tag*{(140)}
\end{aligned} − ∫ Ω f p ( t n + 1 ) ∇ ⋅ v ~ d x + ∫ Ω f q ∇ ⋅ u ( t n + 1 ) d x + γ ∫ ω ( u ( t n + 1 ) ∘ φ ) ⋅ ( v ~ ∘ φ ) d x d y arc = ∫ Ω f F f ( t n + 1 ) ⋅ v ~ d x − ∫ Γ in f p in ( t n + 1 ) n f ⋅ v ~ d S 0 − ∫ Γ out f p out ( t n + 1 ) n f ⋅ v ~ d S 0 + ∫ ω [ 2 γ ( u ( t n ) ∘ φ + η ˙ ( t n ) ) + 2 1 ( t ^ f ( t n ) − f s ref ( t n ) ) ] ⋅ ( v ~ ∘ φ ) d x d y arc + ρ f ∫ Ω f τ u n + 1 ⋅ v ~ d x + ∫ ω R f n + 1 ⋅ ( v ~ ∘ φ ) d x d y arc . (140)
The divergence equation has no residual because the exact velocity is divergence-free at t n + 1 t^{n+1} t n + 1 .
For all ( v , w ) ∈ V N ( ω ) (v,w)\in V_N(\omega) ( v , w ) ∈ V N ( ω ) , the exact shell solution satisfies
ε ρ s ∫ ω η ˙ ( t n + 1 ) − η ˙ ( t n ) Δ t ⋅ v d x d y a r c + ε 3 ρ s 12 ∫ ω ω ˙ ( t n + 1 ) − ω ˙ ( t n ) Δ t ⋅ w d x d y a r c + B m ( ( η ( t n + 1 ) , ω ( t n + 1 ) ) , ( v , w ) ) + B b ( ( η ( t n + 1 ) , ω ( t n + 1 ) ) , ( v , w ) ) + γ ∫ ω η ˙ ( t n + 1 ) ⋅ v d x d y a r c = ∫ ω [ γ 2 ( u ( t n ) ∘ φ + η ˙ ( t n ) ) − 1 2 ( t ^ f ( t n ) − f s r e f ( t n ) ) ] ⋅ v d x d y a r c + ε ρ s ∫ ω τ η n + 1 ⋅ v d x d y a r c + ε 3 ρ s 12 ∫ ω τ ω n + 1 ⋅ w d x d y a r c + ∫ ω R s n + 1 ⋅ v d x d y a r c . (141) \begin{aligned}
\varepsilon\rho_s\int_{\omega}\frac{\dot{\eta}(t^{n+1})-\dot{\eta}(t^n)}{\Delta t}\cdot v\,dx\,dy_{\mathrm{arc}}+\frac{\varepsilon^3\rho_s}{12}\int_{\omega}\frac{\dot{\omega}(t^{n+1})-\dot{\omega}(t^n)}{\Delta t}\cdot w\,dx\,dy_{\mathrm{arc}} \\
&+B_m((\eta(t^{n+1}),\omega(t^{n+1})),(v,w))+B_b((\eta(t^{n+1}),\omega(t^{n+1})),(v,w))+\gamma\int_{\omega}\dot{\eta}(t^{n+1})\cdot v\,dx\,dy_{\mathrm{arc}} \\
&=\int_{\omega}\left[\frac{\gamma}{2}(u(t^n)\circ\varphi+\dot{\eta}(t^n))-\frac{1}{2}(\hat{t}_f(t^n)-f_s^{\mathrm{ref}}(t^n))\right]\cdot v\,dx\,dy_{\mathrm{arc}} \\
&+\varepsilon\rho_s\int_{\omega}\tau_\eta^{n+1}\cdot v\,dx\,dy_{\mathrm{arc}}+\frac{\varepsilon^3\rho_s}{12}\int_{\omega}\tau_\omega^{n+1}\cdot w\,dx\,dy_{\mathrm{arc}}+\int_{\omega}\mathcal{R}_s^{n+1}\cdot v\,dx\,dy_{\mathrm{arc}}. \tag*{(141)}
\end{aligned} ε ρ s ∫ ω Δ t η ˙ ( t n + 1 ) − η ˙ ( t n ) ⋅ v d x d y arc + 12 ε 3 ρ s ∫ ω Δ t ω ˙ ( t n + 1 ) − ω ˙ ( t n ) ⋅ w d x d y arc + B m (( η ( t n + 1 ) , ω ( t n + 1 )) , ( v , w )) + B b (( η ( t n + 1 ) , ω ( t n + 1 )) , ( v , w )) + γ ∫ ω η ˙ ( t n + 1 ) ⋅ v d x d y arc = ∫ ω [ 2 γ ( u ( t n ) ∘ φ + η ˙ ( t n )) − 2 1 ( t ^ f ( t n ) − f s ref ( t n )) ] ⋅ v d x d y arc + ε ρ s ∫ ω τ η n + 1 ⋅ v d x d y arc + 12 ε 3 ρ s ∫ ω τ ω n + 1 ⋅ w d x d y arc + ∫ ω R s n + 1 ⋅ v d x d y arc . (141)
We note that the exact shell fields also satisfy (135) . The kinematic residuals will enter the energy estimate when we express velocity errors through displacement and rotation differences. They are not additional forces in the momentum equation above. Moreover, the elastic forms are time-independent and evaluated at the exact new-time fields, so they introduce no further temporal residual.
Proposition (Temporal consistency on the fixed domain). Assume (122) and (123) , fix γ > 0 \gamma>0 γ > 0 , and sample the external data exactly. Then the exact Stokes–Naghdi solution satisfies the time-semidiscrete equations with residuals (124) , (126) , (130) , and (133) . The interface and momentum residuals are O ( Δ t ) O(\Delta t) O ( Δ t ) in the discrete L 2 L^2 L 2 time norm ( Δ t ∑ n = 0 N − 1 ∥ r X n + 1 ∥ 2 ) 1 / 2 (\Delta t\sum_{n=0}^{N-1}\|r_X^{n+1}\|^2)^{1/2} ( Δ t ∑ n = 0 N − 1 ∥ r X n + 1 ∥ 2 ) 1/2 . The paired kinematic residual has the same order in each elastic seminorm, with the squared norm replaced by the corresponding bilinear form. These statements concern the weighted time norm; the stated regularity does not give a uniform pointwise residual bound.
Proof. Equations (140) , (141) , and (135) give the required residual equations. Taking square roots in (129) , (132) , (138) , and (139) proves the stated orders. Their regularity integrals are finite by assumption, and the constants are independent of Δ t \Delta t Δ t for fixed γ \gamma γ .
With stronger regularity, Taylor expansion also gives the leading interface terms
R f n + 1 = Δ t [ γ ∂ t η ˙ ( t n + 1 ) + ∂ t t ^ f ( t n + 1 ) ] + O ( Δ t 2 ) , \mathcal{R}_f^{n+1}=\Delta t\left[\gamma\partial_t\dot{\eta}(t^{n+1})+\partial_t\hat{t}_f(t^{n+1})\right]+O(\Delta t^2), R f n + 1 = Δ t [ γ ∂ t η ˙ ( t n + 1 ) + ∂ t t ^ f ( t n + 1 ) ] + O ( Δ t 2 ) ,
R s n + 1 = Δ t [ γ ∂ t η ˙ ( t n + 1 ) − ∂ t t ^ f ( t n + 1 ) ] + O ( Δ t 2 ) . (142) \mathcal{R}_s^{n+1}=\Delta t\left[\gamma\partial_t\dot{\eta}(t^{n+1})-\partial_t\hat{t}_f(t^{n+1})\right]+O(\Delta t^2). \tag*{(142)} R s n + 1 = Δ t [ γ ∂ t η ˙ ( t n + 1 ) − ∂ t t ^ f ( t n + 1 ) ] + O ( Δ t 2 ) . (142)
The O ( Δ t 2 ) O(\Delta t^2) O ( Δ t 2 ) remainders here require bounded second time derivatives of η ˙ \dot{\eta} η ˙ and t ^ f \hat{t}_f t ^ f in the interface norm. These expansions explain the leading terms; the proposition above uses the weaker integral estimates.
A priori temporal error estimate
We define the error variables from subtracting numerical solution from the exact solution:
e u n = u ( t n ) − u n , e p n = p ( t n ) − p n , e η n = η ( t n ) − η n , e η ˙ n = η ˙ ( t n ) − η ˙ n , e ω n = ω ( t n ) − ω n , e ω ˙ n = ω ˙ ( t n ) − ω ˙ n , e t ^ f n = t ^ f ( t n ) − t ^ f n , e f s n = f s r e f ( t n ) − f s r e f , n . (143) \begin{aligned}
e_u^n &= u(t^n)-u^n, \qquad e_p^n = p(t^n)-p^n, \qquad e_\eta^n = \eta(t^n)-\eta^n, \qquad e_{\dot{\eta}}^n = \dot{\eta}(t^n)-\dot{\eta}^n,\\
e_\omega^n &= \omega(t^n)-\omega^n, \qquad e_{\dot{\omega}}^n = \dot{\omega}(t^n)-\dot{\omega}^n, \qquad e_{\hat{t}_f}^n = \hat{t}_f(t^n)-\hat{t}_f^n, \qquad e_{f_s}^n = f_s^{\mathrm{ref}}(t^n)-f_s^{\mathrm{ref},n}. \tag*{(143)}
\end{aligned} e u n e ω n = u ( t n ) − u n , e p n = p ( t n ) − p n , e η n = η ( t n ) − η n , e η ˙ n = η ˙ ( t n ) − η ˙ n , = ω ( t n ) − ω n , e ω ˙ n = ω ˙ ( t n ) − ω ˙ n , e t ^ f n = t ^ f ( t n ) − t ^ f n , e f s n = f s ref ( t n ) − f s ref , n . (143)
Here, we also include separate errors e t ^ f n e_{\hat{t}_f}^n e t ^ f n and e f s n e_{f_s}^n e f s n for the two recovered interface loads ((94) and (97) ), as our numerical scheme does not impose exact force balance.
The mechanical error energy is
E e r r n = ρ f 2 ∥ e u n ∥ L 2 ( Ω ^ f ) 2 + ε ρ s 2 ∥ e η ˙ n ∥ L 2 ( ω ) 2 + ε 3 ρ s 24 ∥ e ω ˙ n ∥ L 2 ( ω ) 2 + 1 2 B m ( ( e η n , e ω n ) , ( e η n , e ω n ) ) + 1 2 B b ( ( e η n , e ω n ) , ( e η n , e ω n ) ) . (144) \mathcal{E}_{\mathrm{err}}^n =
\frac{\rho_f}{2}\lVert e_u^n\rVert_{L^2(\widehat{\Omega}_f)}^2
+\frac{\varepsilon\rho_s}{2}\lVert e_{\dot{\eta}}^n\rVert_{L^2(\omega)}^2
+\frac{\varepsilon^3\rho_s}{24}\lVert e_{\dot{\omega}}^n\rVert_{L^2(\omega)}^2
+\frac{1}{2}B_m\bigl((e_\eta^n,e_\omega^n),(e_\eta^n,e_\omega^n)\bigr)
+\frac{1}{2}B_b\bigl((e_\eta^n,e_\omega^n),(e_\eta^n,e_\omega^n)\bigr). \tag*{(144)} E err n = 2 ρ f ∥ e u n ∥ L 2 ( Ω f ) 2 + 2 ε ρ s ∥ e η ˙ n ∥ L 2 ( ω ) 2 + 24 ε 3 ρ s ∥ e ω ˙ n ∥ L 2 ( ω ) 2 + 2 1 B m ( ( e η n , e ω n ) , ( e η n , e ω n ) ) + 2 1 B b ( ( e η n , e ω n ) , ( e η n , e ω n ) ) . (144)
The augmented error energy is
E e r r , a u g n = E e r r n + Δ t γ 2 ( ∥ e u n ∘ φ ∥ L 2 ( ω ) 2 + ∥ e η ˙ n ∥ L 2 ( ω ) 2 ) + Δ t 2 γ ( ∥ e t ^ f n ∥ L 2 ( ω ) 2 + ∥ e f s n ∥ L 2 ( ω ) 2 ) . (145) \mathcal{E}_{\mathrm{err,aug}}^n
=
\mathcal{E}_{\mathrm{err}}^n
+\frac{\Delta t\,\gamma}{2}
\left(\lVert e_u^n\circ\varphi\rVert_{L^2(\omega)}^2+\lVert e_{\dot{\eta}}^n\rVert_{L^2(\omega)}^2\right)
+\frac{\Delta t}{2\gamma}
\left(\lVert e_{\hat{t}_f}^n\rVert_{L^2(\omega)}^2+\lVert e_{f_s}^n\rVert_{L^2(\omega)}^2\right). \tag*{(145)} E err , aug n = E err n + 2 Δ t γ ( ∥ e u n ∘ φ ∥ L 2 ( ω ) 2 + ∥ e η ˙ n ∥ L 2 ( ω ) 2 ) + 2 γ Δ t ( ∥ e t ^ f n ∥ L 2 ( ω ) 2 + ∥ e f s n ∥ L 2 ( ω ) 2 ) . (145)
For exact initialization, we take
u 0 = u ( 0 ) , η 0 = η ( 0 ) , η ˙ 0 = η ˙ ( 0 ) , ω 0 = ω ( 0 ) , ω ˙ 0 = ω ˙ ( 0 ) , u^0=u(0), \qquad\eta^0=\eta(0), \qquad\dot{\eta}^0=\dot{\eta}(0), \qquad\omega^0=\omega(0), \qquad\dot{\omega}^0=\dot{\omega}(0), u 0 = u ( 0 ) , η 0 = η ( 0 ) , η ˙ 0 = η ˙ ( 0 ) , ω 0 = ω ( 0 ) , ω ˙ 0 = ω ˙ ( 0 ) ,
t ^ f 0 = [ σ f ( u ( 0 ) , p ( 0 ) ) n f ] ∘ φ , f s r e f , 0 = − t ^ f 0 . (146) \hat{t}_f^0=[\sigma_f(u(0),p(0))n_f]\circ\varphi, \qquad f_s^{\mathrm{ref},0}=-\hat{t}_f^0. \tag*{(146)} t ^ f 0 = [ σ f ( u ( 0 ) , p ( 0 )) n f ] ∘ φ , f s ref , 0 = − t ^ f 0 . (146)
This requires a compatible initial pressure and sufficiently regular traction. The initial augmented error then vanishes. If the initial fields or loads are approximated, their errors will enter the initial augmented error energy.
Error equations Subtract ((93) ) and ((96) ) from ((140) ) and ((141) ), respectively. Using the corresponding Robin load relations (((94) ), ((97) ), ((124) ) and ((126) )), we express the interface terms through the new-time load errors. We can derive the error equation for the fluid subproblem:
ρ f ∫ Ω ^ f e u n + 1 − e u n Δ t ⋅ v ~ d x + 2 μ f ∫ Ω ^ f D ( e u n + 1 ) : D ( v ~ ) d x − ∫ Ω ^ f e p n + 1 ∇ ⋅ v ~ d x + ∫ Ω ^ f q ∇ ⋅ e u n + 1 d x = ∫ ω e t ^ f n + 1 ⋅ ( v ~ ∘ φ ) d x d y a r c + ρ f ∫ Ω ^ f τ u n + 1 ⋅ v ~ d x , ( v ~ , q ) ∈ V f × Q f . (147) \begin{aligned}
\rho_f\int_{\widehat{\Omega}_f}\frac{e_u^{n+1}-e_u^n}{\Delta t}\cdot\tilde{v}\,dx
+2\mu_f\int_{\widehat{\Omega}_f}D(e_u^{n+1}):D(\tilde{v})\,dx
-\int_{\widehat{\Omega}_f}e_p^{n+1}\nabla\cdot\tilde{v}\,dx
+\int_{\widehat{\Omega}_f}q\nabla\cdot e_u^{n+1}\,dx\\
=\int_{\omega}e_{\hat{t}_f}^{n+1}\cdot(\tilde{v}\circ\varphi)\,dx\,dy_{\mathrm{arc}}
+\rho_f\int_{\widehat{\Omega}_f}\tau_u^{n+1}\cdot\tilde{v}\,dx,
\qquad(\tilde{v},q)\in V_f\times Q_f. \tag*{(147)}
\end{aligned} ρ f ∫ Ω f Δ t e u n + 1 − e u n ⋅ v ~ d x + 2 μ f ∫ Ω f D ( e u n + 1 ) : D ( v ~ ) d x − ∫ Ω f e p n + 1 ∇ ⋅ v ~ d x + ∫ Ω f q ∇ ⋅ e u n + 1 d x = ∫ ω e t ^ f n + 1 ⋅ ( v ~ ∘ φ ) d x d y arc + ρ f ∫ Ω f τ u n + 1 ⋅ v ~ d x , ( v ~ , q ) ∈ V f × Q f . (147)
The shell error equation is
ε ρ s ∫ ω e η ˙ n + 1 − e η ˙ n Δ t ⋅ v d x d y a r c + ε 3 ρ s 12 ∫ ω e ω ˙ n + 1 − e ω ˙ n Δ t ⋅ w d x d y a r c + B m ( ( e η n + 1 , e ω n + 1 ) , ( v , w ) ) + B b ( ( e η n + 1 , e ω n + 1 ) , ( v , w ) ) = ∫ ω e f s n + 1 ⋅ v d x d y a r c + ε ρ s ∫ ω τ η n + 1 ⋅ v d x d y a r c + ε 3 ρ s 12 ∫ ω τ ω n + 1 ⋅ w d x d y a r c , ( v , w ) ∈ V N ( ω ) . (148) \begin{aligned}
\varepsilon\rho_s\int_{\omega}\frac{e_{\dot{\eta}}^{n+1}-e_{\dot{\eta}}^n}{\Delta t}\cdot v\,dx\,dy_{\mathrm{arc}}
+\frac{\varepsilon^3\rho_s}{12}\int_{\omega}\frac{e_{\dot{\omega}}^{n+1}-e_{\dot{\omega}}^n}{\Delta t}\cdot w\,dx\,dy_{\mathrm{arc}}\\
+B_m\bigl((e_\eta^{n+1},e_\omega^{n+1}),(v,w)\bigr)
+B_b\bigl((e_\eta^{n+1},e_\omega^{n+1}),(v,w)\bigr)\\
=\int_{\omega}e_{f_s}^{n+1}\cdot v\,dx\,dy_{\mathrm{arc}}
+\varepsilon\rho_s\int_{\omega}\tau_\eta^{n+1}\cdot v\,dx\,dy_{\mathrm{arc}}
+\frac{\varepsilon^3\rho_s}{12}\int_{\omega}\tau_\omega^{n+1}\cdot w\,dx\,dy_{\mathrm{arc}},
\qquad(v,w)\in V_N(\omega). \tag*{(148)}
\end{aligned} ε ρ s ∫ ω Δ t e η ˙ n + 1 − e η ˙ n ⋅ v d x d y arc + 12 ε 3 ρ s ∫ ω Δ t e ω ˙ n + 1 − e ω ˙ n ⋅ w d x d y arc + B m ( ( e η n + 1 , e ω n + 1 ) , ( v , w ) ) + B b ( ( e η n + 1 , e ω n + 1 ) , ( v , w ) ) = ∫ ω e f s n + 1 ⋅ v d x d y arc + ε ρ s ∫ ω τ η n + 1 ⋅ v d x d y arc + 12 ε 3 ρ s ∫ ω τ ω n + 1 ⋅ w d x d y arc , ( v , w ) ∈ V N ( ω ) . (148)
And the two interface error relations:
e ^ t f n + 1 + γ e u n + 1 ∘ φ = γ 2 ( e u n ∘ φ + e η n ) + 1 2 ( e ^ t f n − e f s n ) + R f n + 1 , (149) \widehat{e}_{tf}^{n+1}+\gamma e_u^{n+1}\circ\varphi=\frac{\gamma}{2}\left(e_u^n\circ\varphi+e_\eta^n\right)+\frac{1}{2}\left(\widehat{e}_{tf}^n-e_{fs}^n\right)+\mathcal{R}_f^{n+1}, \tag*{(149)} e t f n + 1 + γ e u n + 1 ∘ φ = 2 γ ( e u n ∘ φ + e η n ) + 2 1 ( e t f n − e f s n ) + R f n + 1 , (149)
e f s n + 1 + γ e η n + 1 = γ 2 ( e u n ∘ φ + e η n ) − 1 2 ( e ^ t f n − e f s n ) + R s n + 1 . (150) e_{fs}^{n+1}+\gamma e_\eta^{n+1}=\frac{\gamma}{2}\left(e_u^n\circ\varphi+e_\eta^n\right)-\frac{1}{2}\left(\widehat{e}_{tf}^n-e_{fs}^n\right)+\mathcal{R}_s^{n+1}. \tag*{(150)} e f s n + 1 + γ e η n + 1 = 2 γ ( e u n ∘ φ + e η n ) − 2 1 ( e t f n − e f s n ) + R s n + 1 . (150)
And the kinematic error relations:
e η n + 1 − e η n Δ t = e η ˙ n + 1 + κ η n + 1 , e ω n + 1 − e ω n Δ t = e ω ˙ n + 1 + κ ω n + 1 . (151) \frac{e_\eta^{n+1}-e_\eta^n}{\Delta t}=e_{\dot{\eta}}^{n+1}+\kappa_\eta^{n+1},\qquad
\frac{e_\omega^{n+1}-e_\omega^n}{\Delta t}=e_{\dot{\omega}}^{n+1}+\kappa_\omega^{n+1}. \tag*{(151)} Δ t e η n + 1 − e η n = e η ˙ n + 1 + κ η n + 1 , Δ t e ω n + 1 − e ω n = e ω ˙ n + 1 + κ ω n + 1 . (151)
We note that the interface residuals already enter through (149) and (150) , so they did not appear in the momentum equations.
Mechanical error energy balance Test the fluid error equation with ( v ~ , q ) = ( e u n + 1 , e p n + 1 ) (\widetilde{v},q)=(e_u^{n+1},e_p^{n+1}) ( v , q ) = ( e u n + 1 , e p n + 1 ) and the shell equation with ( v , w ) = ( e η n + 1 , e ω n + 1 ) (v,w)=(e_\eta^{n+1},e_\omega^{n+1}) ( v , w ) = ( e η n + 1 , e ω n + 1 ) . Pressure cancels, and the three inertial terms give kinetic-energy differences and nonnegative increments, as in the stability proof.
For the elastic terms B m B_m B m and B b B_b B b , since they are symmetric and bilinear, for example, we have:
B m ( a , a − b ) = 1 2 [ B m ( a , a ) − B m ( b , b ) + B m ( a − b , a − b ) ] . B_m(a,a-b)=\frac{1}{2}\left[B_m(a,a)-B_m(b,b)+B_m(a-b,a-b)\right]. B m ( a , a − b ) = 2 1 [ B m ( a , a ) − B m ( b , b ) + B m ( a − b , a − b ) ] .
Using (151) gives
Δ t B m ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) = 1 2 B m ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) − 1 2 B m ( ( e η n , e ω n ) , ( e η n , e ω n ) ) + 1 2 B m ( ( e η n + 1 − e η n , e ω n + 1 − e ω n ) , ( e η n + 1 − e η n , e ω n + 1 − e ω n ) ) − Δ t B m ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) . (152) \begin{aligned}
\Delta t B_m\left((e_\eta^{n+1},e_\omega^{n+1}),(e_\eta^{n+1},e_\omega^{n+1})\right)
={}&\frac{1}{2}B_m\left((e_\eta^{n+1},e_\omega^{n+1}),(e_\eta^{n+1},e_\omega^{n+1})\right)\\
&-\frac{1}{2}B_m\left((e_\eta^n,e_\omega^n),(e_\eta^n,e_\omega^n)\right)\\
&+\frac{1}{2}B_m\left((e_\eta^{n+1}-e_\eta^n,e_\omega^{n+1}-e_\omega^n),(e_\eta^{n+1}-e_\eta^n,e_\omega^{n+1}-e_\omega^n)\right)\\
&-\Delta t B_m\left((e_\eta^{n+1},e_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right). \tag*{(152)}
\end{aligned} Δ t B m ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) = 2 1 B m ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) − 2 1 B m ( ( e η n , e ω n ) , ( e η n , e ω n ) ) + 2 1 B m ( ( e η n + 1 − e η n , e ω n + 1 − e ω n ) , ( e η n + 1 − e η n , e ω n + 1 − e ω n ) ) − Δ t B m ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) . (152)
Similarly, the corresponding bending identity is
Δ t B b ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) = 1 2 B b ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) − 1 2 B b ( ( e η n , e ω n ) , ( e η n , e ω n ) ) + 1 2 B b ( ( e η n + 1 − e η n , e ω n + 1 − e ω n ) , ( e η n + 1 − e η n , e ω n + 1 − e ω n ) ) − Δ t B b ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) . (153) \begin{aligned}
\Delta t B_b\left((e_\eta^{n+1},e_\omega^{n+1}),(e_\eta^{n+1},e_\omega^{n+1})\right)
={}&\frac{1}{2}B_b\left((e_\eta^{n+1},e_\omega^{n+1}),(e_\eta^{n+1},e_\omega^{n+1})\right)\\
&-\frac{1}{2}B_b\left((e_\eta^n,e_\omega^n),(e_\eta^n,e_\omega^n)\right)\\
&+\frac{1}{2}B_b\left((e_\eta^{n+1}-e_\eta^n,e_\omega^{n+1}-e_\omega^n),(e_\eta^{n+1}-e_\eta^n,e_\omega^{n+1}-e_\omega^n)\right)\\
&-\Delta t B_b\left((e_\eta^{n+1},e_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right). \tag*{(153)}
\end{aligned} Δ t B b ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) = 2 1 B b ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) − 2 1 B b ( ( e η n , e ω n ) , ( e η n , e ω n ) ) + 2 1 B b ( ( e η n + 1 − e η n , e ω n + 1 − e ω n ) , ( e η n + 1 − e η n , e ω n + 1 − e ω n ) ) − Δ t B b ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) . (153)
Multiply the tested equations (147) and (148) by Δ t \Delta t Δ t and add them. Apply the kinetic-energy polarization identity and (152) –(153) . After dropping the nonnegative kinetic and elastic increments, we obtain:
E e r r n + 1 − E e r r n + 2 μ f Δ t ∥ D ( e u n + 1 ) ∥ L 2 ( Ω ^ f ) 2 ≤ Δ t ∫ ω [ e ^ t f n + 1 ⋅ ( e u n + 1 ∘ φ ) + e f s n + 1 ⋅ e η n + 1 ] d x d y a r c + ρ f Δ t ∫ Ω ^ f τ u n + 1 ⋅ e u n + 1 d x + ϵ ρ s Δ t ∫ ω τ η n + 1 ⋅ e η n + 1 d x d y a r c + ϵ 3 ρ s Δ t 12 ∫ ω τ ω n + 1 ⋅ e ω n + 1 d x d y a r c + Δ t B m ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) + Δ t B b ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) . (154) \begin{aligned}
\mathcal{E}_{\mathrm{err}}^{n+1}-\mathcal{E}_{\mathrm{err}}^n+2\mu_f\Delta t\left\lVert D(e_u^{n+1})\right\rVert_{L^2(\widehat{\Omega}_f)}^2
\le{}&\Delta t\int_\omega\left[\widehat{e}_{tf}^{n+1}\cdot(e_u^{n+1}\circ\varphi)+e_{fs}^{n+1}\cdot e_\eta^{n+1}\right]\,dx\,dy_{\mathrm{arc}}\\
&+\rho_f\Delta t\int_{\widehat{\Omega}_f}\tau_u^{n+1}\cdot e_u^{n+1}\,dx\\
&+\epsilon\rho_s\Delta t\int_\omega\tau_\eta^{n+1}\cdot e_\eta^{n+1}\,dx\,dy_{\mathrm{arc}}\\
&+\frac{\epsilon^3\rho_s\Delta t}{12}\int_\omega\tau_\omega^{n+1}\cdot e_\omega^{n+1}\,dx\,dy_{\mathrm{arc}}\\
&+\Delta t B_m\left((e_\eta^{n+1},e_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right)\\
&+\Delta t B_b\left((e_\eta^{n+1},e_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right). \tag*{(154)}
\end{aligned} E err n + 1 − E err n + 2 μ f Δ t D ( e u n + 1 ) L 2 ( Ω f ) 2 ≤ Δ t ∫ ω [ e t f n + 1 ⋅ ( e u n + 1 ∘ φ ) + e f s n + 1 ⋅ e η n + 1 ] d x d y arc + ρ f Δ t ∫ Ω f τ u n + 1 ⋅ e u n + 1 d x + ϵ ρ s Δ t ∫ ω τ η n + 1 ⋅ e η n + 1 d x d y arc + 12 ϵ 3 ρ s Δ t ∫ ω τ ω n + 1 ⋅ e ω n + 1 d x d y arc + Δ t B m ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) + Δ t B b ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) . (154)
The last two terms come from the exact solution’s kinematic residuals. They account for the difference between its time derivatives and the backward differences used by the scheme.
Momentum and kinematic residual work We bound the residual work in (154) using Young’s inequality, with the final time T > 0 T > 0 T > 0 as a fixed scale. For the momentum terms, we have:
ρ f ∣ ∫ Ω ^ f τ u n + 1 ⋅ e u n + 1 d x ∣ ≤ ρ f 2 T ∥ e u n + 1 ∥ L 2 ( Ω ^ f ) 2 + ρ f T 2 ∥ τ u n + 1 ∥ L 2 ( Ω ^ f ) 2 , \rho_f \left|\int_{\widehat{\Omega}_f} \boldsymbol{\tau}_u^{n+1}\cdot\boldsymbol{e}_u^{n+1}\,dx\right|
\leq\frac{\rho_f}{2T}\|\boldsymbol{e}_u^{n+1}\|_{L^2(\widehat{\Omega}_f)}^2+\frac{\rho_f T}{2}\|\boldsymbol{\tau}_u^{n+1}\|_{L^2(\widehat{\Omega}_f)}^2, ρ f ∫ Ω f τ u n + 1 ⋅ e u n + 1 d x ≤ 2 T ρ f ∥ e u n + 1 ∥ L 2 ( Ω f ) 2 + 2 ρ f T ∥ τ u n + 1 ∥ L 2 ( Ω f ) 2 ,
ε ρ s ∣ ∫ ω τ η n + 1 ⋅ e η n + 1 d x d y a r c ∣ ≤ ε ρ s 2 T ∥ e η n + 1 ∥ L 2 ( ω ) 2 + ε ρ s T 2 ∥ τ η n + 1 ∥ L 2 ( ω ) 2 , \varepsilon\rho_s\left|\int_{\omega}\boldsymbol{\tau}_{\eta}^{n+1}\cdot\boldsymbol{e}_{\eta}^{n+1}\,dx\,dy_{\mathrm{arc}}\right|
\leq\frac{\varepsilon\rho_s}{2T}\|\boldsymbol{e}_{\eta}^{n+1}\|_{L^2(\omega)}^2+\frac{\varepsilon\rho_sT}{2}\|\boldsymbol{\tau}_{\eta}^{n+1}\|_{L^2(\omega)}^2, ε ρ s ∫ ω τ η n + 1 ⋅ e η n + 1 d x d y arc ≤ 2 T ε ρ s ∥ e η n + 1 ∥ L 2 ( ω ) 2 + 2 ε ρ s T ∥ τ η n + 1 ∥ L 2 ( ω ) 2 ,
ε 3 ρ s 12 ∣ ∫ ω τ ω n + 1 ⋅ e ω n + 1 d x d y a r c ∣ ≤ ε 3 ρ s 24 T ∥ e ω n + 1 ∥ L 2 ( ω ) 2 + ε 3 ρ s T 24 ∥ τ ω n + 1 ∥ L 2 ( ω ) 2 . (155) \frac{\varepsilon^3\rho_s}{12}\left|\int_{\omega}\boldsymbol{\tau}_{\omega}^{n+1}\cdot\boldsymbol{e}_{\omega}^{n+1}\,dx\,dy_{\mathrm{arc}}\right|
\leq\frac{\varepsilon^3\rho_s}{24T}\|\boldsymbol{e}_{\omega}^{n+1}\|_{L^2(\omega)}^2+\frac{\varepsilon^3\rho_sT}{24}\|\boldsymbol{\tau}_{\omega}^{n+1}\|_{L^2(\omega)}^2. \tag*{(155)} 12 ε 3 ρ s ∫ ω τ ω n + 1 ⋅ e ω n + 1 d x d y arc ≤ 24 T ε 3 ρ s ∥ e ω n + 1 ∥ L 2 ( ω ) 2 + 24 ε 3 ρ s T ∥ τ ω n + 1 ∥ L 2 ( ω ) 2 . (155)
Regarding the elastic terms, Cauchy-Schwarz in the shell seminorms gives
∣ B m ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ∣ ≤ 1 2 T B m ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) + T 2 B m ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) , \left|B_m\left((e_\eta^{n+1},e_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right)\right|
\leq\frac{1}{2T}B_m\left((e_\eta^{n+1},e_\omega^{n+1}),(e_\eta^{n+1},e_\omega^{n+1})\right)+\frac{T}{2}B_m\left((\kappa_\eta^{n+1},\kappa_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right), B m ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ≤ 2 T 1 B m ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) + 2 T B m ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ,
∣ B b ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ∣ ≤ 1 2 T B b ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) + T 2 B b ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) . (156) \left|B_b\left((e_\eta^{n+1},e_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right)\right|
\leq\frac{1}{2T}B_b\left((e_\eta^{n+1},e_\omega^{n+1}),(e_\eta^{n+1},e_\omega^{n+1})\right)+\frac{T}{2}B_b\left((\kappa_\eta^{n+1},\kappa_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right). \tag*{(156)} B b ( ( e η n + 1 , e ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ≤ 2 T 1 B b ( ( e η n + 1 , e ω n + 1 ) , ( e η n + 1 , e ω n + 1 ) ) + 2 T B b ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) . (156)
Together, the error-dependent terms on the right are E e r r n + 1 / T \mathcal{E}_{\mathrm{err}}^{n+1}/T E err n + 1 / T . We bound this by E e r r , a u g n + 1 / T \mathcal{E}_{\mathrm{err,aug}}^{n+1}/T E err , aug n + 1 / T when combining the estimates.
Interface work with residuals The Robin square identity (103) must also account for the two interface residuals. We use
∣ a + b ∣ 2 ≤ ( 1 + Δ t T ) ∣ a ∣ 2 + ( 1 + T Δ t ) ∣ b ∣ 2 (157) |a+b|^2\leq\left(1+\frac{\Delta t}{T}\right)|a|^2+\left(1+\frac{T}{\Delta t}\right)|b|^2 \tag*{(157)} ∣ a + b ∣ 2 ≤ ( 1 + T Δ t ) ∣ a ∣ 2 + ( 1 + Δ t T ) ∣ b ∣ 2 (157)
on the right of each Robin error relation ((7.7 and (7.8), taking b b b as its interface residual. Applying the same square-expansion and averaging calculation as in (103) yields
∥ e ^ t f n + 1 + γ e u n + 1 ∘ φ ∥ L 2 ( ω ) 2 + ∥ e f s n + 1 + γ e η n + 1 ∥ L 2 ( ω ) 2 ≤ ( 1 + Δ t T ) [ γ 2 ( ∥ e u n ∘ φ ∥ L 2 ( ω ) 2 + ∥ e η n ∥ L 2 ( ω ) 2 ) + ∥ e ^ t f n ∥ L 2 ( ω ) 2 + ∥ e f s n ∥ L 2 ( ω ) 2 − γ 2 2 ∥ e u n ∘ φ − e η n ∥ L 2 ( ω ) 2 − 1 2 ∥ e ^ t f n + e f s n ∥ L 2 ( ω ) 2 ] + ( 1 + T Δ t ) ( ∥ R f n + 1 ∥ L 2 ( ω ) 2 + ∥ R s n + 1 ∥ L 2 ( ω ) 2 ) . (158) \begin{aligned}
\|\widehat{e}_{t_f}^{n+1}+\gamma e_u^{n+1}\circ\varphi\|_{L^2(\omega)}^2+\|e_{f_s}^{n+1}+\gamma e_\eta^{n+1}\|_{L^2(\omega)}^2
&\leq\left(1+\frac{\Delta t}{T}\right)\bigg[\gamma^2\left(\|e_u^n\circ\varphi\|_{L^2(\omega)}^2+\|e_\eta^n\|_{L^2(\omega)}^2\right)+\|\widehat{e}_{t_f}^n\|_{L^2(\omega)}^2+\|e_{f_s}^n\|_{L^2(\omega)}^2\\
&\qquad-\frac{\gamma^2}{2}\|e_u^n\circ\varphi-e_\eta^n\|_{L^2(\omega)}^2-\frac{1}{2}\|\widehat{e}_{t_f}^n+e_{f_s}^n\|_{L^2(\omega)}^2\bigg]\\
&\quad+\left(1+\frac{T}{\Delta t}\right)\left(\|\mathcal{R}_f^{n+1}\|_{L^2(\omega)}^2+\|\mathcal{R}_s^{n+1}\|_{L^2(\omega)}^2\right). \tag*{(158)}
\end{aligned} ∥ e t f n + 1 + γ e u n + 1 ∘ φ ∥ L 2 ( ω ) 2 + ∥ e f s n + 1 + γ e η n + 1 ∥ L 2 ( ω ) 2 ≤ ( 1 + T Δ t ) [ γ 2 ( ∥ e u n ∘ φ ∥ L 2 ( ω ) 2 + ∥ e η n ∥ L 2 ( ω ) 2 ) + ∥ e t f n ∥ L 2 ( ω ) 2 + ∥ e f s n ∥ L 2 ( ω ) 2 − 2 γ 2 ∥ e u n ∘ φ − e η n ∥ L 2 ( ω ) 2 − 2 1 ∥ e t f n + e f s n ∥ L 2 ( ω ) 2 ] + ( 1 + Δ t T ) ( ∥ R f n + 1 ∥ L 2 ( ω ) 2 + ∥ R s n + 1 ∥ L 2 ( ω ) 2 ) . (158)
The square bracket is the sum of squares of the two old-data Robin right-hand sides before the residuals are added. It is therefore nonnegative, allowing us to keep the displayed defect terms. Expand the new-time squares and multiply by Δ t / ( 2 γ ) \Delta t/(2\gamma) Δ t / ( 2 γ ) . The cross terms give the interface work in (154) , and the remaining terms give the new interface error energy. The old interface energy is multiplied by 1 + Δ t / T 1+\Delta t/T 1 + Δ t / T . The coefficient of the residual squares is
T + Δ t 2 γ ( ∥ R f n + 1 ∥ L 2 ( ω ) 2 + ∥ R s n + 1 ∥ L 2 ( ω ) 2 ) . (159) \frac{T+\Delta t}{2\gamma}\left(\|\mathcal{R}_f^{n+1}\|_{L^2(\omega)}^2+\|\mathcal{R}_s^{n+1}\|_{L^2(\omega)}^2\right). \tag*{(159)} 2 γ T + Δ t ( ∥ R f n + 1 ∥ L 2 ( ω ) 2 + ∥ R s n + 1 ∥ L 2 ( ω ) 2 ) . (159)
This coefficient has no extra factor of Δ t \Delta t Δ t multiplying the residual squares. As a result, the interface residual contributes one lower power of Δ t \Delta t Δ t than the momentum and kinematic terms in the final squared-error bound.
One-step augmented error inequality. Combining (154) , (155) , (156) , and (158) gives
( 1 − Δ t T ) E e r r , a u g n + 1 + 2 μ f Δ t ∥ D ( e u n + 1 ) ∥ L 2 ( Ω ^ f ) 2 + ( 1 + Δ t T ) Δ t γ 4 ∥ e u n ∘ φ − e η n ∥ L 2 ( ω ) 2 + ( 1 + Δ t T ) Δ t 4 γ ∥ e ^ t f n + e f s n ∥ L 2 ( ω ) 2 ≤ ( 1 + Δ t T ) E e r r , a u g n + T Δ t 2 [ ρ f ∥ τ u n + 1 ∥ L 2 ( Ω ^ f ) 2 + ϵ ρ s ∥ τ η n + 1 ∥ L 2 ( ω ) 2 + ϵ 3 ρ s 12 ∥ τ ω n + 1 ∥ L 2 ( ω ) 2 + B m ( ( κ η n + 1 , κ η n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) + B b ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ] + T + Δ t 2 γ ( ∥ R f n + 1 ∥ L 2 ( ω ) 2 + ∥ R s n + 1 ∥ L 2 ( ω ) 2 ) . (160) \begin{aligned}
\left(1-\frac{\Delta t}{T}\right)\mathcal{E}_{\mathrm{err,aug}}^{n+1}
&+2\mu_f\Delta t\left\|D\left(e_u^{n+1}\right)\right\|_{L^2(\widehat{\Omega}_f)}^2 \\
&+\left(1+\frac{\Delta t}{T}\right)\frac{\Delta t\gamma}{4}\left\|e_u^n\circ\varphi-e_\eta^n\right\|_{L^2(\omega)}^2
+\left(1+\frac{\Delta t}{T}\right)\frac{\Delta t}{4\gamma}\left\|\widehat e_{tf}^n+e_{fs}^n\right\|_{L^2(\omega)}^2 \\
&\leq\left(1+\frac{\Delta t}{T}\right)\mathcal{E}_{\mathrm{err,aug}}^n \\
&\quad+\frac{T\Delta t}{2}\left[\rho_f\left\|\tau_u^{n+1}\right\|_{L^2(\widehat{\Omega}_f)}^2+\epsilon\rho_s\left\|\tau_\eta^{n+1}\right\|_{L^2(\omega)}^2+\frac{\epsilon^3\rho_s}{12}\left\|\tau_\omega^{n+1}\right\|_{L^2(\omega)}^2\right.\\
&\qquad\left.+B_m\left((\kappa_\eta^{n+1},\kappa_\eta^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right)+B_b\left((\kappa_\eta^{n+1},\kappa_\omega^{n+1}),(\kappa_\eta^{n+1},\kappa_\omega^{n+1})\right)\right]\\
&\quad+\frac{T+\Delta t}{2\gamma}\left(\left\|R_f^{n+1}\right\|_{L^2(\omega)}^2+\left\|R_s^{n+1}\right\|_{L^2(\omega)}^2\right). \tag*{(160)}
\end{aligned} ( 1 − T Δ t ) E err , aug n + 1 + 2 μ f Δ t D ( e u n + 1 ) L 2 ( Ω f ) 2 + ( 1 + T Δ t ) 4 Δ t γ e u n ∘ φ − e η n L 2 ( ω ) 2 + ( 1 + T Δ t ) 4 γ Δ t e t f n + e f s n L 2 ( ω ) 2 ≤ ( 1 + T Δ t ) E err , aug n + 2 T Δ t [ ρ f τ u n + 1 L 2 ( Ω f ) 2 + ϵ ρ s τ η n + 1 L 2 ( ω ) 2 + 12 ϵ 3 ρ s τ ω n + 1 L 2 ( ω ) 2 + B m ( ( κ η n + 1 , κ η n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) + B b ( ( κ η n + 1 , κ ω n + 1 ) , ( κ η n + 1 , κ ω n + 1 ) ) ] + 2 γ T + Δ t ( R f n + 1 L 2 ( ω ) 2 + R s n + 1 L 2 ( ω ) 2 ) . (160)
To obtain the above inequality in terms of augmented energy, we increased the old mechanical energy by the factor 1 + Δ t / T 1+\Delta t/T 1 + Δ t / T and replaced E e r r n + 1 / T \mathcal{E}_{\mathrm{err}}^{n+1}/T E err n + 1 / T with E e r r , a u g n + 1 / T \mathcal{E}_{\mathrm{err,aug}}^{n+1}/T E err , aug n + 1 / T . Specifically, the momentum and kinematic residual estimates contribute ( Δ t / T ) E e r r n + 1 (\Delta t/T)\mathcal{E}_{\mathrm{err}}^{n+1} ( Δ t / T ) E err n + 1 to the right-hand side. Moving this term to the left and using (145) , we obtain
( 1 − Δ t T ) E e r r , a u g n + 1 ≤ E e r r , a u g n + 1 − Δ t T E e r r n + 1 , \left(1-\frac{\Delta t}{T}\right)\mathcal{E}_{\mathrm{err,aug}}^{n+1}
\leq\mathcal{E}_{\mathrm{err,aug}}^{n+1}-\frac{\Delta t}{T}\mathcal{E}_{\mathrm{err}}^{n+1}, ( 1 − T Δ t ) E err , aug n + 1 ≤ E err , aug n + 1 − T Δ t E err n + 1 ,
because E e r r n + 1 ≤ E e r r , a u g n + 1 \mathcal{E}_{\mathrm{err}}^{n+1}\leq\mathcal{E}_{\mathrm{err,aug}}^{n+1} E err n + 1 ≤ E err , aug n + 1 . At the old time level, (158) multiplies the interface terms by 1 + Δ t / T 1+\Delta t/T 1 + Δ t / T . Together with the old mechanical energy, these terms equal
E e r r n + ( 1 + Δ t T ) ( E e r r , a u g n − E e r r n ) = ( 1 + Δ t T ) E e r r , a u g n − Δ t T E e r r n ≤ ( 1 + Δ t T ) E e r r , a u g n . \mathcal{E}_{\mathrm{err}}^n+\left(1+\frac{\Delta t}{T}\right)\left(\mathcal{E}_{\mathrm{err,aug}}^n-\mathcal{E}_{\mathrm{err}}^n\right)
=\left(1+\frac{\Delta t}{T}\right)\mathcal{E}_{\mathrm{err,aug}}^n-\frac{\Delta t}{T}\mathcal{E}_{\mathrm{err}}^n
\leq\left(1+\frac{\Delta t}{T}\right)\mathcal{E}_{\mathrm{err,aug}}^n. E err n + ( 1 + T Δ t ) ( E err , aug n − E err n ) = ( 1 + T Δ t ) E err , aug n − T Δ t E err n ≤ ( 1 + T Δ t ) E err , aug n .
Using these lower and upper bounds gives (160) .
Theorem 7.1 (Temporal error estimate). Assume the fixed-domain hypotheses of § 4 and the exact-solution regularity in (122) and (123) . Let Δ t = T / N \Delta t=T/N Δ t = T / N with N ≥ 2 N\geq2 N ≥ 2 , and let γ > 0 \gamma>0 γ > 0 be fixed independently of Δ t \Delta t Δ t . Suppose that the numerical and exact external data agree at the time levels. Then there exists a constant C > 0 C>0 C > 0 , independent of Δ t \Delta t Δ t , such that
max 0 ≤ n ≤ N E e r r , a u g n + 2 μ f Δ t ∑ n = 0 N − 1 ∥ D ( e u n + 1 ) ∥ L 2 ( Ω ^ f ) 2 + Δ t γ 4 ∑ n = 0 N − 1 ∥ e u n ∘ φ − e η n ∥ L 2 ( ω ) 2 + Δ t 4 γ ∑ n = 0 N − 1 ∥ e ^ t f n + e f s n ∥ L 2 ( ω ) 2 ≤ C { E e r r , a u g 0 + T Δ t 2 6 ∫ 0 T [ ρ f ∥ ∂ t t u ( t ) ∥ L 2 ( Ω ^ f ) 2 + ϵ ρ s ∥ ∂ t t η ˙ ( t ) ∥ L 2 ( ω ) 2 + ϵ 3 ρ s 12 ∥ ∂ t t ω ˙ ( t ) ∥ L 2 ( ω ) 2 + B m ( ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) , ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) ) + B b ( ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) , ( ∂ t t η ( t ) , ∂ t t ω ( t ) ) ) ] d t + ( T + Δ t ) Δ t γ ∫ 0 T [ γ 2 ∥ ∂ t η ˙ ( t ) ∥ L 2 ( ω ) 2 + ∥ ∂ t e ^ t f ( t ) ∥ L 2 ( ω ) 2 ] d t } . (161) \begin{aligned}
\max_{0\leq n\leq N}\mathcal{E}_{\mathrm{err,aug}}^n
&+2\mu_f\Delta t\sum_{n=0}^{N-1}\left\|D\left(e_u^{n+1}\right)\right\|_{L^2(\widehat{\Omega}_f)}^2
+\frac{\Delta t\gamma}{4}\sum_{n=0}^{N-1}\left\|e_u^n\circ\varphi-e_\eta^n\right\|_{L^2(\omega)}^2\\
&\quad+\frac{\Delta t}{4\gamma}\sum_{n=0}^{N-1}\left\|\widehat e_{tf}^n+e_{fs}^n\right\|_{L^2(\omega)}^2\\
&\leq C\left\{\mathcal{E}_{\mathrm{err,aug}}^0+\frac{T\Delta t^2}{6}\int_0^T\left[\rho_f\left\|\partial_{tt}u(t)\right\|_{L^2(\widehat{\Omega}_f)}^2+\epsilon\rho_s\left\|\partial_{tt}\dot\eta(t)\right\|_{L^2(\omega)}^2+\frac{\epsilon^3\rho_s}{12}\left\|\partial_{tt}\dot\omega(t)\right\|_{L^2(\omega)}^2\right.\right.\\
&\qquad\left.\left.+B_m\left((\partial_{tt}\eta(t),\partial_{tt}\omega(t)),(\partial_{tt}\eta(t),\partial_{tt}\omega(t))\right)+B_b\left((\partial_{tt}\eta(t),\partial_{tt}\omega(t)),(\partial_{tt}\eta(t),\partial_{tt}\omega(t))\right)\right]\,dt\right.\\
&\qquad\left.+\frac{(T+\Delta t)\Delta t}{\gamma}\int_0^T\left[\gamma^2\left\|\partial_t\dot\eta(t)\right\|_{L^2(\omega)}^2+\left\|\partial_t\widehat e_{tf}(t)\right\|_{L^2(\omega)}^2\right]\,dt\right\}. \tag*{(161)}
\end{aligned} 0 ≤ n ≤ N max E err , aug n + 2 μ f Δ t n = 0 ∑ N − 1 D ( e u n + 1 ) L 2 ( Ω f ) 2 + 4 Δ t γ n = 0 ∑ N − 1 e u n ∘ φ − e η n L 2 ( ω ) 2 + 4 γ Δ t n = 0 ∑ N − 1 e t f n + e f s n L 2 ( ω ) 2 ≤ C { E err , aug 0 + 6 T Δ t 2 ∫ 0 T [ ρ f ∥ ∂ tt u ( t ) ∥ L 2 ( Ω f ) 2 + ϵ ρ s ∥ ∂ tt η ˙ ( t ) ∥ L 2 ( ω ) 2 + 12 ϵ 3 ρ s ∥ ∂ tt ω ˙ ( t ) ∥ L 2 ( ω ) 2 + B m ( ( ∂ tt η ( t ) , ∂ tt ω ( t )) , ( ∂ tt η ( t ) , ∂ tt ω ( t )) ) + B b ( ( ∂ tt η ( t ) , ∂ tt ω ( t )) , ( ∂ tt η ( t ) , ∂ tt ω ( t )) ) ] d t + γ ( T + Δ t ) Δ t ∫ 0 T [ γ 2 ∥ ∂ t η ˙ ( t ) ∥ L 2 ( ω ) 2 + ∥ ∂ t e t f ( t ) ∥ L 2 ( ω ) 2 ] d t } . (161) The physical parameters and γ \gamma γ are fixed. This estimate does not imply a uniform unweighted bound as the shell thickness or density tends to zero.
Proof. The tested error equations and residual-work estimates give (160) . For brevity, let R n + 1 \mathcal{R}^{n+1} R n + 1 denote the sum of the momentum, kinematic, and interface residual terms on the right-hand side of that inequality, excluding ( 1 + Δ t / T ) E e r r , a u g n (1+\Delta t/T)\mathcal{E}^{n}_{\mathrm{err,aug}} ( 1 + Δ t / T ) E err , aug n . All terms in R n + 1 \mathcal{R}^{n+1} R n + 1 are nonnegative.
We first bound the maximum augmented error energy. Since N ≥ 2 N \ge2 N ≥ 2 , we have Δ t / T ≤ 1 / 2 \Delta t/T \le1/2 Δ t / T ≤ 1/2 . Dropping the nonnegative viscous and interface terms in (160) and dividing by 1 − Δ t / T 1-\Delta t/T 1 − Δ t / T gives
E e r r , a u g n + 1 ≤ 1 + Δ t / T 1 − Δ t / T E e r r , a u g n + R n + 1 1 − Δ t / T . \mathcal{E}^{n+1}_{\mathrm{err,aug}} \le\frac{1+\Delta t/T}{1-\Delta t/T}\mathcal{E}^{n}_{\mathrm{err,aug}}+\frac{\mathcal{R}^{n+1}}{1-\Delta t/T}. E err , aug n + 1 ≤ 1 − Δ t / T 1 + Δ t / T E err , aug n + 1 − Δ t / T R n + 1 .
For 0 ≤ Δ t / T ≤ 1 / 2 0 \le\Delta t/T \le1/2 0 ≤ Δ t / T ≤ 1/2 ,
log ( 1 + Δ t / T 1 − Δ t / T ) ≤ 4 Δ t T , 1 1 − Δ t / T ≤ 2. \log\left(\frac{1+\Delta t/T}{1-\Delta t/T}\right) \le\frac{4\Delta t}{T}, \qquad\frac{1}{1-\Delta t/T}\le2. log ( 1 − Δ t / T 1 + Δ t / T ) ≤ T 4Δ t , 1 − Δ t / T 1 ≤ 2.
Since N Δ t = T N\Delta t=T N Δ t = T , iteration over at most N N N steps yields
max 0 ≤ n ≤ N E e r r , a u g n ≤ 2 exp ( 4 ) ( E e r r , a u g 0 + ∑ n = 0 N − 1 R n + 1 ) . (162) \max_{0\le n\le N}\mathcal{E}^{n}_{\mathrm{err,aug}} \le2\exp(4)\left(\mathcal{E}^{0}_{\mathrm{err,aug}}+\sum_{n=0}^{N-1}\mathcal{R}^{n+1}\right). \tag*{(162)} 0 ≤ n ≤ N max E err , aug n ≤ 2 exp ( 4 ) ( E err , aug 0 + n = 0 ∑ N − 1 R n + 1 ) . (162)
We next bound the accumulated viscous dissipation and interface defects. Summing (160) over n = 0 , … , N − 1 n=0,\ldots,N-1 n = 0 , … , N − 1 and replacing 1 + Δ t / T 1+\Delta t/T 1 + Δ t / T by 1 1 1 in front of the nonnegative interface defects gives
2 μ f Δ t ∑ n = 0 N − 1 ∥ D ( e u n + 1 ) ∥ L 2 ( Ω f ) 2 + Δ t γ 4 ∑ n = 0 N − 1 ∥ e u n ∘ φ − e η n ∥ L 2 ( ω ) 2 + Δ t 4 γ ∑ n = 0 N − 1 ∥ e ^ t f n + e f s n ∥ L 2 ( ω ) 2 (163) 2\mu_f\Delta t\sum_{n=0}^{N-1}\lVert\mathsf{D}(e_u^{n+1})\rVert_{L^2(\Omega_f)}^2+\frac{\Delta t\,\gamma}{4}\sum_{n=0}^{N-1}\lVert e_u^{n}\circ\varphi-e_\eta^{n}\rVert_{L^2(\omega)}^2+\frac{\Delta t}{4\gamma}\sum_{n=0}^{N-1}\lVert\widehat e_{tf}^{n}+e_{fs}^{n}\rVert_{L^2(\omega)}^2 \tag*{(163)} 2 μ f Δ t n = 0 ∑ N − 1 ∥ D ( e u n + 1 ) ∥ L 2 ( Ω f ) 2 + 4 Δ t γ n = 0 ∑ N − 1 ∥ e u n ∘ φ − e η n ∥ L 2 ( ω ) 2 + 4 γ Δ t n = 0 ∑ N − 1 ∥ e t f n + e f s n ∥ L 2 ( ω ) 2 (163)
≤ E e r r , a u g 0 − E e r r , a u g N + Δ t T ∑ n = 0 N − 1 ( E e r r , a u g n + 1 + E e r r , a u g n ) + ∑ n = 0 N − 1 R n + 1 (164) \le\mathcal{E}^{0}_{\mathrm{err,aug}}-\mathcal{E}^{N}_{\mathrm{err,aug}}+\frac{\Delta t}{T}\sum_{n=0}^{N-1}\left(\mathcal{E}^{n+1}_{\mathrm{err,aug}}+\mathcal{E}^{n}_{\mathrm{err,aug}}\right)+\sum_{n=0}^{N-1}\mathcal{R}^{n+1} \tag*{(164)} ≤ E err , aug 0 − E err , aug N + T Δ t n = 0 ∑ N − 1 ( E err , aug n + 1 + E err , aug n ) + n = 0 ∑ N − 1 R n + 1 (164)
≤ E e r r , a u g 0 + 2 max 0 ≤ n ≤ N E e r r , a u g n + ∑ n = 0 N − 1 R n + 1 . (165) \le\mathcal{E}^{0}_{\mathrm{err,aug}}+2\max_{0\le n\le N}\mathcal{E}^{n}_{\mathrm{err,aug}}+\sum_{n=0}^{N-1}\mathcal{R}^{n+1}. \tag*{(165)} ≤ E err , aug 0 + 2 0 ≤ n ≤ N max E err , aug n + n = 0 ∑ N − 1 R n + 1 . (165)
The viscous and interface sums are then bounded by the same residual quantity, with a constant independent of Δ t \Delta t Δ t . Here we use nonnegativity of the final energy and N Δ t = T N\Delta t=T N Δ t = T . Combining this inequality with (162) bounds the maximum energy and all accumulated terms by
( 1 + 6 exp ( 4 ) ) ( E e r r , a u g 0 + ∑ n = 0 N − 1 R n + 1 ) . (1+6\exp(4))\left(\mathcal{E}^{0}_{\mathrm{err,aug}}+\sum_{n=0}^{N-1}\mathcal{R}^{n+1}\right). ( 1 + 6 exp ( 4 )) ( E err , aug 0 + n = 0 ∑ N − 1 R n + 1 ) .
It remains to estimate the summed residual terms. For the momentum and kinematic residuals, apply (132) , (138) , and (139) . Their coefficient T Δ t / 2 T\Delta t/2 T Δ t /2 in (160) gives the factor T Δ t 2 / 6 T\Delta t^2/6 T Δ t 2 /6 in (161) .
For the interface residuals, divide (129) by Δ t \Delta t Δ t to obtain
∑ n = 0 N − 1 ( ∥ R f n + 1 ∥ L 2 ( ω ) 2 + ∥ R s n + 1 ∥ L 2 ( ω ) 2 ) ≤ 2 Δ t ∫ 0 T [ γ 2 ∥ ∂ t η ˙ ( t ) ∥ L 2 ( ω ) 2 + ∥ ∂ t f ^ ( t ) ∥ L 2 ( ω ) 2 ] d t . \sum_{n=0}^{N-1}\left(\lVert\mathcal{R}^{n+1}_{f}\rVert_{L^2(\omega)}^2+\lVert\mathcal{R}^{n+1}_{s}\rVert_{L^2(\omega)}^2\right)\le2\Delta t\int_0^T\left[\gamma^2\lVert\partial_t\dot{\eta}(t)\rVert_{L^2(\omega)}^2+\lVert\partial_t\widehat f(t)\rVert_{L^2(\omega)}^2\right]\,\mathrm{d}t. n = 0 ∑ N − 1 ( ∥ R f n + 1 ∥ L 2 ( ω ) 2 + ∥ R s n + 1 ∥ L 2 ( ω ) 2 ) ≤ 2Δ t ∫ 0 T [ γ 2 ∥ ∂ t η ˙ ( t ) ∥ L 2 ( ω ) 2 + ∥ ∂ t f ( t ) ∥ L 2 ( ω ) 2 ] d t .
Multiplying by ( T + Δ t ) / ( 2 γ ) (T+\Delta t)/(2\gamma) ( T + Δ t ) / ( 2 γ ) gives the final residual term in (161) . Absorbing the Gronwall factors into C C C completes the proof.
Convergence rate and initialization
With exact initialization of the fluid and shell variables and both interface loads, E e r r , a u g 0 = 0 \mathcal{E}_{\mathrm{err,aug}}^{0}=0 E err , aug 0 = 0 . For fixed T T T , γ \gamma γ , and an exact solution with the stated regularity, the momentum and kinematic terms in (161) are O ( Δ t 2 ) O(\Delta t^{2}) O ( Δ t 2 ) , while the interface term is O ( Δ t ) O(\Delta t) O ( Δ t ) . Hence
max 0 ≤ n ≤ N E e r r , a u g n = O ( Δ t ) , max 0 ≤ n ≤ N E e r r n = O ( Δ t ) . (166) \max_{0\le n\le N}\mathcal{E}_{\mathrm{err,aug}}^{n}=O(\Delta t), \qquad\max_{0\le n\le N}\mathcal{E}_{\mathrm{err}}^{n}=O(\Delta t). \tag*{(166)} 0 ≤ n ≤ N max E err , aug n = O ( Δ t ) , 0 ≤ n ≤ N max E err n = O ( Δ t ) . (166)
Taking square roots gives the mechanical energy-norm bound O ( Δ t ) O(\sqrt{\Delta t}) O ( Δ t ) . Its component estimates are
max 0 ≤ n ≤ N ρ f ∥ e u n ∥ L 2 ( Ω ^ f ) = O ( Δ t ) , \max_{0\le n\le N}\sqrt{\rho_f}\lVert e_u^n\rVert_{L^2(\widehat{\Omega}_f)}=O(\sqrt{\Delta t}), 0 ≤ n ≤ N max ρ f ∥ e u n ∥ L 2 ( Ω f ) = O ( Δ t ) ,
max 0 ≤ n ≤ N ε ρ s ∥ e η ˙ n ∥ L 2 ( ω ) = O ( Δ t ) , \max_{0\le n\le N}\sqrt{\varepsilon\rho_s}\lVert e_{\dot{\eta}}^n\rVert_{L^2(\omega)}=O(\sqrt{\Delta t}), 0 ≤ n ≤ N max ε ρ s ∥ e η ˙ n ∥ L 2 ( ω ) = O ( Δ t ) ,
max 0 ≤ n ≤ N ε 3 ρ s / 12 ∥ e ω ˙ n ∥ L 2 ( ω ) = O ( Δ t ) , \max_{0\le n\le N}\sqrt{\varepsilon^3\rho_s/12}\lVert e_{\dot{\omega}}^n\rVert_{L^2(\omega)}=O(\sqrt{\Delta t}), 0 ≤ n ≤ N max ε 3 ρ s /12 ∥ e ω ˙ n ∥ L 2 ( ω ) = O ( Δ t ) ,
max 0 ≤ n ≤ N [ B m ( ( e η n , e ω n ) , ( e η n , e ω n ) ) + B b ( ( e η n , e ω n ) , ( e η n , e ω n ) ) ] 1 / 2 = O ( Δ t ) . (167) \max_{0\le n\le N}\left[B_m\left((e_\eta^n,e_\omega^n),(e_\eta^n,e_\omega^n)\right)+B_b\left((e_\eta^n,e_\omega^n),(e_\eta^n,e_\omega^n)\right)\right]^{1/2}=O(\sqrt{\Delta t}). \tag*{(167)} 0 ≤ n ≤ N max [ B m ( ( e η n , e ω n ) , ( e η n , e ω n ) ) + B b ( ( e η n , e ω n ) , ( e η n , e ω n ) ) ] 1/2 = O ( Δ t ) . (167)
The square root of the accumulated symmetric-gradient error has the same order in its discrete time norm.
The last line controls the shell elastic seminorm. For the clamped cylinder, (87) directly gives the full H 1 H^1 H 1 error bound. Under only shifted coercivity, we also need to control the shell L 2 L^2 L 2 error. Summing the kinematic error equations and applying Cauchy–Schwarz in time gives
max 0 ≤ n ≤ N ( ∥ e η n ∥ L 2 ( ω ) 2 + ∥ e ω n ∥ L 2 ( ω ) 2 ) ≤ 3 ( ∥ e η 0 ∥ L 2 ( ω ) 2 + ∥ e ω 0 ∥ L 2 ( ω ) 2 ) + 3 T 2 max 0 ≤ n ≤ N ( ∥ e η ˙ n ∥ L 2 ( ω ) 2 + ∥ e ω ˙ n ∥ L 2 ( ω ) 2 ) + 3 T Δ t ∑ n = 0 N − 1 ( ∥ κ η n + 1 ∥ L 2 ( ω ) 2 + ∥ κ ω n + 1 ∥ L 2 ( ω ) 2 ) . (168) \max_{0\le n\le N}\left(\lVert e_\eta^n\rVert_{L^2(\omega)}^2+\lVert e_\omega^n\rVert_{L^2(\omega)}^2\right)
\le3\left(\lVert e_\eta^0\rVert_{L^2(\omega)}^2+\lVert e_\omega^0\rVert_{L^2(\omega)}^2\right)+3T^2\max_{0\le n\le N}\left(\lVert e_{\dot{\eta}}^n\rVert_{L^2(\omega)}^2+\lVert e_{\dot{\omega}}^n\rVert_{L^2(\omega)}^2\right)
+3T\Delta t\sum_{n=0}^{N-1}\left(\lVert\kappa_\eta^{n+1}\rVert_{L^2(\omega)}^2+\lVert\kappa_\omega^{n+1}\rVert_{L^2(\omega)}^2\right). \tag*{(168)} 0 ≤ n ≤ N max ( ∥ e η n ∥ L 2 ( ω ) 2 + ∥ e ω n ∥ L 2 ( ω ) 2 ) ≤ 3 ( ∥ e η 0 ∥ L 2 ( ω ) 2 + ∥ e ω 0 ∥ L 2 ( ω ) 2 ) + 3 T 2 0 ≤ n ≤ N max ( ∥ e η ˙ n ∥ L 2 ( ω ) 2 + ∥ e ω ˙ n ∥ L 2 ( ω ) 2 ) + 3 T Δ t n = 0 ∑ N − 1 ( ∥ κ η n + 1 ∥ L 2 ( ω ) 2 + ∥ κ ω n + 1 ∥ L 2 ( ω ) 2 ) . (168)
The final sum, with its Δ t \Delta t Δ t weight, is O ( Δ t 2 ) O(\Delta t^2) O ( Δ t 2 ) by (134) and the regularity assumptions. With exact initialization and fixed physical parameters, the squared shell L 2 L^2 L 2 error is therefore O ( Δ t ) O(\Delta t) O ( Δ t ) . Combining this with shifted coercivity and the elastic bound gives
max 0 ≤ n ≤ N ( ∥ e η n ∥ H 1 ( ω ) 2 + ∥ e ω n ∥ H 1 ( ω ) 2 ) 1 / 2 = O ( Δ t ) . (169) \max_{0\le n\le N}\left(\lVert e_\eta^n\rVert_{H^1(\omega)}^2+\lVert e_\omega^n\rVert_{H^1(\omega)}^2\right)^{1/2}=O(\sqrt{\Delta t}). \tag*{(169)} 0 ≤ n ≤ N max ( ∥ e η n ∥ H 1 ( ω ) 2 + ∥ e ω n ∥ H 1 ( ω ) 2 ) 1/2 = O ( Δ t ) . (169)
For approximate initialization under only shifted coercivity, this argument additionally requires ∥ e η 0 ∥ L 2 ( ω ) 2 + ∥ e ω 0 ∥ L 2 ( ω ) 2 = O ( Δ t ) \lVert e_\eta^0\rVert_{L^2(\omega)}^2+\lVert e_\omega^0\rVert_{L^2(\omega)}^2=O(\Delta t) ∥ e η 0 ∥ L 2 ( ω ) 2 + ∥ e ω 0 ∥ L 2 ( ω ) 2 = O ( Δ t ) . For the clamped cylinder, unshifted coercivity provides this control through the initial augmented error bound, so no separate initial L 2 L^2 L 2 condition is needed.
The mechanical energy-norm rate also holds for approximate initialization provided E e r r , a u g 0 = O ( Δ t ) \mathcal{E}_{\mathrm{err,aug}}^{0}=O(\Delta t) E err , aug 0 = O ( Δ t ) . This condition includes the initial mechanical error, the fluid and shell velocity trace errors, and both interface load errors. For fixed γ \gamma γ , uniformly bounded initial errors in the interface traces and loads contribute O ( Δ t ) O(\Delta t) O ( Δ t ) to the augmented initial error energy. To obtain the stated convergence rate, the initial mechanical error energy must also be O ( Δ t ) O(\Delta t) O ( Δ t ) . Larger initial errors remain in the final estimate. The full shell H 1 H^1 H 1 rate then follows directly for the clamped cylinder, or with the additional initial L 2 L^2 L 2 condition under shifted coercivity.
Because the exact solution satisfies velocity continuity and force balance,
e u n ∘ φ − e η ˙ n = − ( u n ∘ φ − η ˙ n ) , e ^ t f n + e f s n = − ( t ^ f n + f s r e f , n ) . (170) e_u^n\circ\varphi-e_{\dot{\eta}}^n=-\left(u^n\circ\varphi-\dot{\eta}^n\right),\qquad\widehat{e}_{t_f}^n+e_{f_s}^n=-\left(\widehat{t}_f^n+f_s^{\mathrm{ref},n}\right). \tag*{(170)} e u n ∘ φ − e η ˙ n = − ( u n ∘ φ − η ˙ n ) , e t f n + e f s n = − ( t f n + f s ref , n ) . (170)
Equation (161) therefore also controls the numerical velocity mismatch and force imbalance. For fixed γ \gamma γ and either exact initialization or E e r r , a u g 0 = O ( Δ t ) \mathcal{E}^{0}_{\mathrm{err,aug}} = O(\Delta t) E err , aug 0 = O ( Δ t ) , their weighted discrete L 2 L^2 L 2 time norms are O ( Δ t ) O(\sqrt{\Delta t}) O ( Δ t ) . These estimates use the old time levels n = 0 , … , N − 1 n = 0,\ldots,N-1 n = 0 , … , N − 1 at which the interface data are exchanged. They follow from the error equations and consistency bounds in addition to stability.
Numerical examples
Pressure-wave propagation in a straight tube
We evaluate the proposed scheme using a moving-domain benchmark problem involving a pressure pulse in a straight compliant tube. We compare its results with those of the β \beta β -scheme [2 ] and a monolithic solver. We first examine the effect of varying the Robin parameter γ \gamma γ at a fixed time step, then investigate temporal refinement with γ \gamma γ held fixed. The comparisons focus on fluid pressure and radial wall displacement.
The reference tube has length L = 0.06 L = 0.06 L = 0.06 m and radius R 0 = 0.005 R_0 = 0.005 R 0 = 0.005 m. Its wall motion is governed by the Naghdi shell model described in § 2 . All three solvers use the same membrane, shear, bending, and inertia operators and the parameters as listed in Table 1 . The pressure-pulse setting is based on straight-tube β \beta β -scheme benchmarks [4 ] . Initially, the fluid and shell are at rest, with
Table 1. Physical parameters of the pressure-pulse in a straight tube study.
zero displacement and interface loads. The shell is clamped at both ends and periodic in its circumferential parameter. The inlet traction is − p i n ( t ) n x -p_{\mathrm{in}}(t)\boldsymbol{n}_x − p in ( t ) n x , with
p i n ( t ) = { p max 2 ( 1 − cos ( 2 π t t p ) ) , 0 ≤ t ≤ t p , 0 , t > t p , p max = 2000 P a , t p = 5 m s . (171) p_{\mathrm{in}}(t) =
\begin{cases}
\dfrac{p_{\max}}{2}\left(1-\cos\left(\dfrac{2\pi t}{t_p}\right)\right), & 0 \le t \le t_p, \\
0, & t > t_p,
\end{cases}
\qquad p_{\max}=2000\ \mathrm{Pa},\qquad t_p=5\ \mathrm{ms}. \tag*{(171)} p in ( t ) = ⎩ ⎨ ⎧ 2 p m a x ( 1 − cos ( t p 2 π t ) ) , 0 , 0 ≤ t ≤ t p , t > t p , p m a x = 2000 Pa , t p = 5 ms . (171)
The outlet traction is zero. Each simulation runs from 0 to 12 ms. We exam the pressure and displacement profiles at t k ∈ { 2 , 4 , 6 , 8 , 10 , 12 } t_k \in\{2,4,6,8,10,12\} t k ∈ { 2 , 4 , 6 , 8 , 10 , 12 } ms for our comparisons.
Figure 1 shows the meshes used in our test. The fluid mesh contains 5978 vertices and 29160 tetrahedra. The shell parameter mesh has 3840 triangles and 2013 vertices. Fluid velocity and pressure use continuous P 1 P_1 P 1 elements, with Brezzi-Pitkäranta stabilization term 0.01 ∑ K h K 2 ( ∇ p h , ∇ q h ) K 0.01\sum_K h_K^2(\nabla p_h,\nabla q_h)_K 0.01 ∑ K h K 2 ( ∇ p h , ∇ q h ) K . All six shell components use continuous P 1 P_1 P 1 elements. The two partitioned solvers use the same
Figure 1. Reference meshes: (a) fluid exterior, (b) shell mesh in the parametric domain.
lagged skew-convection form and pressure stabilization, but advance geometry and loads differently. The β \beta β scheme uses the new shell-stage displacement for mesh motion and projects Cauchy traction. Our explicit scheme predicts geometry from the old state and recovers the load from its Robin condition. The monolithic solver instead iterates the current geometry and convecting velocity. For our comparison, β \beta β scheme uses β = 1 \beta=1 β = 1 and Δ t = 10 − 4 \Delta t=10^{-4} Δ t = 1 0 − 4 s. The monolithic solver uses the same time step, and within each step, Picard iterations update the ALE geometry and skew convection until both velocity increments meet a relative tolerance of 10 − 7 10^{-7} 1 0 − 7 . A step is rejected if it has not converged within 30 iterations.
For result comparison, we divide the reference axial interval into 61 equal bins. Fluid cells are assigned by their reference centroids, and we use B i ( t ) B_i(t) B i ( t ) to denote the current cells in bin i i i , and S i S_i S i to denote the corresponding shell parameter cells. We analyze the following data profiles:
p ‾ i ( t ) = 1 ∣ B i ( t ) ∣ ∫ B i ( t ) p h ( t ) d x , η ‾ r , i ( t ) = 1 ∣ S i ∣ ∫ S i η h ( t ) ⋅ ( 0 , cos ( y / R 0 ) , sin ( y / R 0 ) ) d y d x . (172) \overline{p}_i(t)=\frac{1}{|B_i(t)|}\int_{B_i(t)}p_h(t)\,dx,\qquad\overline{\eta}_{r,i}(t)=\frac{1}{|S_i|}\int_{S_i}\eta_h(t)\cdot(0,\cos(y/R_0),\sin(y/R_0))\,dy\,dx. \tag*{(172)} p i ( t ) = ∣ B i ( t ) ∣ 1 ∫ B i ( t ) p h ( t ) d x , η r , i ( t ) = ∣ S i ∣ 1 ∫ S i η h ( t ) ⋅ ( 0 , cos ( y / R 0 ) , sin ( y / R 0 )) d y d x . (172)
The shell integral uses axial and arc-length coordinates x x x and y y y . Both quantities average over a slab of cells, rather than sample an exact cross-section. The radial displacement is also obtained from the stored mean diameter through η ‾ r , i = ( D ‾ i − 2 R 0 ) / 2 \overline{\eta}_{r,i}=(\overline{D}_i-2R_0)/2 η r , i = ( D i − 2 R 0 ) /2 .
We take the β \beta β solution as the numerical comparison standard. For either q = p ‾ q=\overline{p} q = p or q = η ‾ r q=\overline{\eta}_r q = η r , define
D q = 100 ( ∑ k = 1 6 ∑ i = 1 61 ∣ q i ( t k ) − q i β ( t k ) ∣ 2 ∑ k = 1 6 ∑ i = 1 61 ∣ q i β ( t k ) ∣ 2 ) 1 / 2 % . (173) D_q=100\left(\frac{\sum_{k=1}^{6}\sum_{i=1}^{61}\left|q_i(t_k)-q_i^\beta(t_k)\right|^2}{\sum_{k=1}^{6}\sum_{i=1}^{61}\left|q_i^\beta(t_k)\right|^2}\right)^{1/2}\%. \tag*{(173)} D q = 100 ∑ k = 1 6 ∑ i = 1 61 q i β ( t k ) 2 ∑ k = 1 6 ∑ i = 1 61 q i ( t k ) − q i β ( t k ) 2 1/2 %. (173)
Here, k k k indexes the time snapshots and i i i indexes the equally spaced bins. This measure compares six profiles at matching physical times, without phase shifts or amplitude rescaling, and is distinct from a full-field error or the analytical energy norm. We also compute discrepancies relative to the monolithic solution by replacing q β q^\beta q β with the corresponding monolithic values. The β \beta β and monolithic solutions differ by 4.09% in pressure and 4.97% in radial displacement, providing a numerical cross-check, although neither is established as an exact or mesh-converged reference.
Sensitivity study of the Robin parameter We hold Δ t = 10 − 4 s \Delta t = 10^{-4}\,\mathrm{s} Δ t = 1 0 − 4 s and take γ ∈ { 500 , 1000 , 1500 , 3000 , 5000 } k g m − 2 s − 1 \gamma\in\{500, 1000, 1500, 3000, 5000\}\,\mathrm{kg\,m^{-2}\,s^{-1}} γ ∈ { 500 , 1000 , 1500 , 3000 , 5000 } kg m − 2 s − 1 . Figure 2 and 3 show the pressure and wall motion, and Table 2 summarizes their discrepancies. For smaller γ \gamma γ , the results obtained from the explicit scheme differ substantially from both solutions obtained from the β \beta β scheme and monolithic solver. However, increasing γ \gamma γ reduces the discrepancies in both quantities, and the improvement becomes marginal as γ \gamma γ increases from 3000 to 5000. Thus, γ = 3000 \gamma= 3000 γ = 3000 is a reasonable choice for the current mesh and time step considered.
Table 2. Robin parameter sensitivity study with Δ t = 10 − 4 s \Delta t = 10^{-4}\,\mathrm{s} Δ t = 1 0 − 4 s . Discrepancies relative to the β \beta β scheme solutions are defined by (173) and evaluated using six snapshots at 2, 4, 6, 8, 10, and 12 ms.
Figure 2. Robin parameter sensitivity study: slab-averaged pressure profiles at t = 2 , 4 , 6 , 8 , 10 , 12 t = 2, 4, 6, 8, 10, 12 t = 2 , 4 , 6 , 8 , 10 , 12 ms. All curves use Δ t = 10 − 4 s \Delta t = 10^{-4}\,\mathrm{s} Δ t = 1 0 − 4 s , and the β \beta β scheme and monolithic solution are included for comparison.
Time refinement at fixed Robin parameter We study γ = 1500 \gamma= 1500 γ = 1500 and 3000 separately, using explicit time steps 10 − 4 10^{-4} 1 0 − 4 , 5 × 10 − 5 5 \times10^{-5} 5 × 1 0 − 5 , and 2.5 × 10 − 5 s 2.5 \times10^{-5}\,\mathrm{s} 2.5 × 1 0 − 5 s . All other settings are unchanged. The β \beta β and monolithic solutions stay at Δ t = 10 − 4 s \Delta t = 10^{-4}\,\mathrm{s} Δ t = 1 0 − 4 s . Figure 4 –7 show the six profile times, and Table 3
Table 3. Explicit time refinement at γ = 1500 \gamma= 1500 γ = 1500 and 3000. The β \beta β -scheme comparison uses Δ t = 10 − 4 \Delta t = 10^{-4} Δ t = 1 0 − 4 s. Pressure and radial-displacement discrepancies use the six snapshots at 2–12 ms and 61 axial bins.
Figure 4. Time refinement at γ = 1500 \gamma= 1500 γ = 1500 : pressure profiles at the six snapshot times. Only the explicit splitting scheme refine time step from t = ( 10 − 4 , 5 × 10 − 5 , 2.5 × 10 − 5 ) t = (10^{-4}, 5 \times10^{-5}, 2.5 \times10^{-5}) t = ( 1 0 − 4 , 5 × 1 0 − 5 , 2.5 × 1 0 − 5 ) . Both β \beta β scheme and monolithic comparison solvers use Δ t = 10 − 4 \Delta t = 10^{-4} Δ t = 1 0 − 4 s.
Figure 3. Robin parameter sensitivity study: mean radial wall displacement at six different time moments. All curves use Δ t = 10 − 4 s \Delta t = 10^{-4}\,\mathrm{s} Δ t = 1 0 − 4 s , and the β \beta β scheme and monolithic solution are included for comparison. Displacement is reported in micrometres.
reports the corresponding discrepancies against the solution from β \beta β -scheme. The table uses only the displayed 2 to 12 ms snapshots.
Both pressure and displacement discrepancies decrease at each refinement in Table 3 . The same trend holds against the monolithic solution. For γ = 1500 \gamma= 1500 γ = 1500 , the pressure discrepancy changes from 27.65% to 8.69%, and the displacement discrepancy from 34.37% to 12.08%. For γ = 3000 \gamma= 3000 γ = 3000 , the pressure discrepancy changes from 18.28% to 6.28%, and the displacement discrepancy from 23.98% to 9.64%. Both measures decrease at each refinement for both Robin parameters. At each tested time step, γ = 3000 \gamma= 3000 γ = 3000 gives smaller discrepancies than γ = 1500 \gamma= 1500 γ = 1500 . Together with the fixed-step study, this shows that the useful choice of γ \gamma γ depends on the time step and the quantity being compared.
Figure 6. Time refinement study with γ = 3000 \gamma= 3000 γ = 3000 : pressure profiles at six selected times. The time steps used for the explicit scheme and comparison solvers are the same as those in Figure 4 .
Figure 7. Time refinement study with γ = 3000 \gamma= 3000 γ = 3000 : mean radial displacement, comparing with Figure 5 .
Figure 5. Time refinement study with γ = 1500 \gamma= 1500 γ = 1500 : mean radial displacement at six selected times. Only the explicit splitting scheme is refined in time, using Δ t = 10 − 4 \Delta t = 10^{-4} Δ t = 1 0 − 4 , 5 × 10 − 5 5 \times10^{-5} 5 × 1 0 − 5 , and 2.5 × 10 − 5 2.5 \times10^{-5} 2.5 × 1 0 − 5 , s. Both the β \beta β scheme and the monolithic solver use Δ t = 10 − 4 \Delta t = 10^{-4} Δ t = 1 0 − 4 , s.
Noticing that the computations include interface mismatch, load imbalance, and geometry staging, so they do not separate a single mechanism responsible for the remaining differences.
Pressure-wave propagation in a stented artery
We next apply the explicit splitting scheme to an artery supported with a stent. The reference artery has a length of 33 mm and a radius of 1.5 mm. Let s ∈ [ 0 , 33 ] s \in[0, 33] s ∈ [ 0 , 33 ] mm denote distance from the inlet and θ \theta θ the circumferential angle. The connected strut network lies in the central region, approximately 6 ≤ s ≤ 27 6 \le s \le27 6 ≤ s ≤ 27 mm. It has eight axial cells and eight circumferential repeats, with strut width 0.13 mm.
We reported all the parameters used in this example in Table 4 . The shell stiffness is assembled
Table 4. Layer parameters used in the stented-artery study.
by integrating through both layers, rather than replacing the wall by steel. The total shell thickness is 0.4 mm at the two ends and 0.55 mm at the stented region. To specify this extension of the homogeneous model, let z z z measure distance outward from the artery midsurface, let h a h_a h a and h s h_s h s be the artery and stent thicknesses, and let χ s \chi_s χ s be the indicator of the strut bands. The elastic moments used in the implementation are
A j ( s , θ ) = E a ∫ − h a / 2 h a / 2 z j d z + χ s ( s , θ ) E s ∫ h a / 2 h a / 2 + h s z j d z , m j ( s , θ ) = ρ a ∫ − h a / 2 h a / 2 z j d z + χ s ( s , θ ) ρ s ∫ h a / 2 h a / 2 + h s z j d z , j = 0 , 1 , 2. (174) \begin{aligned}
A_j(s,\theta) &= E_a \int_{-h_a/2}^{h_a/2} z^j\,\mathrm{d}z + \chi_s(s,\theta)E_s \int_{h_a/2}^{h_a/2+h_s} z^j\,\mathrm{d}z,\\
m_j(s,\theta) &= \rho_a \int_{-h_a/2}^{h_a/2} z^j\,\mathrm{d}z + \chi_s(s,\theta)\rho_s \int_{h_a/2}^{h_a/2+h_s} z^j\,\mathrm{d}z,\qquad j=0,1,2. \tag*{(174)}
\end{aligned} A j ( s , θ ) m j ( s , θ ) = E a ∫ − h a /2 h a /2 z j d z + χ s ( s , θ ) E s ∫ h a /2 h a /2 + h s z j d z , = ρ a ∫ − h a /2 h a /2 z j d z + χ s ( s , θ ) ρ s ∫ h a /2 h a /2 + h s z j d z , j = 0 , 1 , 2. (174)
With the common Poisson ratio, A 0 A_0 A 0 , A 1 A_1 A 1 , and A 2 A_2 A 2 supply the membrane, membrane-bending coupling, and bending coefficients in the Naghdi equation. The corresponding inertial terms use m 0 m_0 m 0 , m 1 m_1 m 1 , and m 2 m_2 m 2 . The first moments are retained because the added layer is outside the artery midsurface. This is a layered shell approximation.
Figure 8 shows the fitted material mesh. The shell has 21,032 triangles and 10,705 parameter-mesh vertices. The fluid mesh has 156,320 tetrahedra and 30,620 vertices. Fluid velocity and pressure use continuous P 1 P_1 P 1 elements with pressure stabilization coefficient 0.01. The six shell components also use continuous P 1 P_1 P 1 elements, with periodicity around the circumference and clamped ends. Harmonic extension supplies the ALE displacement. The fluid properties, inlet pressure pulse, and outlet boundary conditions are the same as those in our first numerical example in § 8.1 . The initial state is at rest. The explicit calculation uses γ = 2000 \gamma= 2000 γ = 2000 , Δ t = 10 − 4 s \Delta t = 10^{-4}\,\mathrm{s} Δ t = 1 0 − 4 s , and the implemented lagged skew-convection form.
Figure 8. Stented-artery mesh: (Top) fluid mesh; (Bottom) shell mesh in the parametric domain.
The Figure 9 shows pressure traveling along the center axis at six different times 2, 4, 6, 8, 10, and 12 ms. Figure 10 shows the corresponding stented artery displacements. Figure 11 represents both sides of the longitudinal plane θ = 0 , π \theta= 0,\pi θ = 0 , π . In the central region, local peaks and troughs occur along the interface, and thus the response is not spatially uniform. This is also why both axial and circumferential resolution are needed to describe this heterogeneous shell response.
Figure 9. Stented artery pressure profiles along the center axis at six different time moments, obtained from using the explicit splitting scheme with γ = 2000 \gamma= 2000 γ = 2000 and Δ t = 10 − 4 s \Delta t = 10^{-4}\,\mathrm{s} Δ t = 1 0 − 4 s .
Figure 10. Displacement of the stented artery at six time snapshots, t = 2 , 4 , 6 , 8 , 10 , t = 2, 4, 6, 8, 10, t = 2 , 4 , 6 , 8 , 10 , and 12 12 12 ms, obtained from using the explicit Robin-Robin splitting scheme. The displacement is magnified by 10 10 10 times for better visualization of the arterial wall deformation.
Figure 11. Local radial displacement at the interface along the longitudinal slicing θ = 0 , π \theta= 0, \pi θ = 0 , π . The horizontal coordinate is distance from the inlet. The two curves are evaluated on opposite sides of the artery.
Conclusions
We have developed an explicit Robin-Robin splitting scheme for fluid interaction with a Naghdi shell. The coupling exchanges translational velocity and interface forces, while the shell formulation retains independent rotations and rotational inertia. Both subproblems use previous-time interface data, allowing them to be solved independently and in parallel without coupling subiterations.
For the fixed-domain Stokes-Naghdi problem, we derive a discrete energy identity that accounts for the mechanical energy, additional interface terms, and nonnegative contributions from velocity mismatch and force imbalance. We establish unconditional stability for homogeneous external data with any fixed positive common Robin coefficient, along with stability estimates for nonzero external data. We further prove first-order temporal consistency and an O ( Δ t ) O(\sqrt{\Delta t}) O ( Δ t ) error estimate in the mechanical energy norm under suitable regularity and initialization assumptions.
The straight-tube numerical experiments investigate the effects of the Robin parameter and time step on the computed pressure and displacement. At a fixed time step, increasing the Robin parameter reduces the discrepancies relative to the reference solutions, although the improvement becomes marginal for larger values of γ \gamma γ . Time refinement with γ = 1500 \gamma= 1500 γ = 1500 and 3000 further reduces the pressure and displacement discrepancies relative to the saved β \beta β solution. These results demonstrate that the accuracy of the proposed scheme depends on both the Robin parameter and the time step.
The stented-artery example extends the numerical application to a bonded layered shell with a large stiffness contrast between the steel struts and the surrounding arterial wall. The pressure profiles show the pulse propagation, while the shell displacement exhibits spatial variations in both the axial and circumferential directions. These variations motivate the use of a Naghdi shell formulation that accounts for membrane, bending, and transverse-shear deformation. The independent shell rotations provide additional flexibility in representing localized bending near stiffness transitions. More importantly, the low-regularity Naghdi formulation avoids requiring second derivatives of the reference midsurface in its geometric coefficients, as discussed in [2 ] .
More generally, the proposed Robin-Robin splitting scheme provides a framework for fluid interaction with thin structures represented by 2D surfaces embedded in a 3D fluid domain, rather than 3D elastic bodies. Since the coupling involves only the structural translational velocity and the corresponding interface force, the same splitting strategy can be extended to other reduced-dimensional structural models, including membrane, plate, and shell models. This allows the fluid and structural subproblems to be advanced independently without introducing a 3D structural domain. The stability and convergence analysis for such extensions would depend on the properties of the corresponding structural model and remains a subject for future investigation.
Data Availability
The data that support the findings of this study are available from the corresponding author upon reasonable request.