betaFeedback

WorksPhysics and Astronomy

A variational multiscale rotational pressure-correction method for incompressible flows with open boundaries

Abstract

Rotational pressure correction improves pressure accuracy near open boundaries in incompressible flow simulations, but equal-order finite element discretizations can still develop spurious pressure oscillations. We develop a residual-based variational multiscale rotational pressure-correction scheme (RPCS-VMS) that stabilizes both the pressure-increment and pressure-recovery equations. By constructing stage-specific momentum residuals using a common intermediate-velocity acceleration, the scheme retains four sequential linear solves per time step. For a reduced BDF1 Oseen model with frozen stabilization coefficients, we establish augmented-energy decay and a conditional bound on the pressure gradient. Manufactured-solution tests demonstrate suppression of pressure oscillations at small time steps and reduced near-boundary pressure errors. With P1/P1P_1/P_1 and Q1/Q1Q_1/Q_1 elements, the observed spatial L2L^2 convergence rates are 2 for velocity and approximately 1.51.5 for pressure, while first- and second-order backward differentiation formulas attain their expected temporal orders. Branched-pipe and two- and three-dimensional cylinder benchmarks assess outlet flow partitioning, transient forces, and pressure differences, with results agreeing closely with reference solutions. In the three-dimensional cylinder test, RPCS-VMS achieves a 2.41-fold wall-time speedup over the tested monolithic VMS solver.

Introduction

Pressure-correction methods advance the incompressible Navier-Stokes equations through separate velocity and pressure solves, avoiding a coupled saddle-point system at each time step. Following Chorin’s projection method [1], fractional-step formulations have been developed for a range of spatial discretizations and time integration orders [2, 3, 4, 5, 6]. Their pressure accuracy depends in part on the compatibility of the auxiliary boundary conditions with the coupled problem [6, 7, 8]. When equal-order finite elements are used, control of discrete pressure modes is a further requirement. Boundary accuracy and pressure stability therefore need separate consideration.

The rotational pressure-correction scheme (RPCS), introduced by Timmermans, Minev, and Van de V esse [9], modifies the standard incremental pressure update by incorporating a correction based on the divergence of the intermediate velocity. Guermond and Shen [10] subsequently established improved velocity and pressure error estimates for this scheme under the assumptions of their analysis. At open boundaries, prescribed data couple velocity derivatives and pressure, making their treatment in the auxiliary subproblems important to splitting accuracy. For time-dependent Stokes problems, Guermond, Minev, and Shen [11] identified limitations of standard pressure correction and established error estimates for the rotational formulation. Subsequent work developed several approaches to imposing open and traction boundary data. Liu [12] proposed pressure boundary conditions within a pressure-Poisson formulation, while Poux et al. [13, 14] investigated improved boundary enforcement for pressure- and velocity-correction schemes. Bänsch [15] modified the traction condition for the intermediate velocity through a boundary correction involving the pressure increment. Bonito, Guermond, and Lee [16] instead modified the pressure-increment boundary condition to reduce the pressure boundary layer in a first-order Stokes scheme. Backflow introduces an additional difficulty because the convective boundary flux can supply kinetic energy to the computational domain. Energy estimates and open-boundary treatments addressing this issue have been developed by Bruneau and Fabrie [17, 18] and Dong and Shen [19]. Becker et al. [20] developed a general weak boundary formulation with discrete kinetic-energy control. For rotational pressure correction with open or traction boundaries, Lee and Salgado [21] established stability results for grad-div-stabilized and boundary-correction schemes. These developments provide important guidance on boundary accuracy and energy behavior. For equal-order discretizations, the resulting split formulation must also provide control of discrete pressure modes.

Equal-order velocity-pressure interpolation simplifies implementation, but standard continuous low-order pairs generally lack the inf-sup stability of pairs such as Taylor-Hood elements [22, 23]. Established remedies include residual-based stabilization [24, 25, 26], pressure-gradient and projection methods [27, 28, 29], and interior-penalty terms [30, 31]. Some fractional-step schemes also supply pressure control, but it depends on the time step and can weaken as the step decreases [32, 33]. This motivates additional spatial stabilization for pressure approximation at small time steps. Such stabilization can itself perturb the pressure near physical boundaries [28], and these perturbations must be distinguished from boundary errors introduced by splitting.

The variational multiscale (VMS) framework relates stabilization to the modeled influence of unresolved scales on the finite element solution [34, 35], with different subscale models determining the resulting stabilization terms [36, 37, 38, 39, 40]. Its use with pressure segregation is established. Codina and Soto [41] combine orthogonal-subscale stabilization with an iterative pressure-segregation procedure whose converged solution corresponds to the monolithic time discretization. Codina and Badia [42] examine fractional-step and other segregation algorithms and establish stability estimates for fractional-step schemes without an inf-sup assumption. They also identify the weakness of the resulting pressure estimates and explain that stronger control can be obtained from inf-sup stability or spatial stabilization.

In rotational pressure correction, the pressure increment and updated pressure are determined by different equations linked through the rotational update. This raises the question of how to coordinate residual stabilization across these stages while retaining sequential solves. The resulting formulation also requires an assessment of pressure control, since energy decay alone need not exclude discrete pressure oscillations.

We develop a residual-based VMS rotational pressure-correction scheme (RPCS-VMS) that coordinates stabilization in the pressure-increment and pressure-recovery equations while retaining four sequential solves. For a reduced homogeneous BDF1 Oseen model on a fixed P1/P1P_1/P_1 mesh with frozen coefficients, we establish augmented-energy decay under stated coefficient conditions and a conditional bound on the pressure gradient. Manufactured-solution tests and flow benchmarks assess convergence, pressure behavior, and computational cost.

This paper is organized as follows. Sections 2 and 3 present the baseline split and examine its boundary compatibility. Section 4 develops RPCS-VMS, and Section 5 states the reduced-model energy and pressure-gradient estimates. Sections 6 and 7 present the numerical results and conclusions, respectively.

Baseline Rotational Pressure-Correction Framework

Here, we summarize the governing equations and the classical rotational pressure-correction scheme that provide the baseline for the boundary treatment in Section 3 and the VMS formulation in Section 4.

Governing Equations and Boundary Conditions

Let Ω⊂Rd\Omega\subset\mathbb{R}^{d}, d=2d=2 or 33, be a bounded, connected Lipschitz domain. The boundary Γ=∂Ω\Gamma= \partial\Omega comprises a prescribed-velocity portion ΓD\Gamma_{D} and an open-boundary portion ΓN\Gamma_{N}, with ΓD∩ΓN=∅\Gamma_{D} \cap\Gamma_{N} = \varnothing and Γ‾D∪Γ‾N=Γ\overline{\Gamma}_{D} \cup\overline{\Gamma}_{N} = \Gamma. The outward unit normal is denoted by n\boldsymbol{n}. The incompressible Navier–Stokes problem is

∂tu+(u⋅∇)u−ν∇2u+∇p=f,in Ω×(0,T],(1)\partial_{t}\boldsymbol{u} + (\boldsymbol{u}\cdot\nabla)\boldsymbol{u} - \nu\nabla^{2}\boldsymbol{u} + \nabla p = \boldsymbol{f}, \qquad\text{in } \Omega\times(0,T], \tag*{(1)}
∇⋅u=0,in Ω×(0,T],(2)\nabla\cdot\boldsymbol{u} = 0, \qquad\text{in } \Omega\times(0,T], \tag*{(2)}
u=uD,on ΓD×(0,T],(3)\boldsymbol{u} = \boldsymbol{u}_{D}, \qquad\text{on } \Gamma_{D} \times(0,T], \tag*{(3)}
ν∂nu−pn−12(u⋅n)u=h,on ΓN×(0,T],(4)\nu\partial_{n}\boldsymbol{u} - p\boldsymbol{n} - \frac{1}{2}(\boldsymbol{u}\cdot\boldsymbol{n})\boldsymbol{u} = \boldsymbol{h}, \qquad\text{on } \Gamma_{N} \times(0,T], \tag*{(4)}
u(⋅,0)=u0,in Ω.\boldsymbol{u}(\cdot,0) = \boldsymbol{u}_{0}, \qquad\text{in } \Omega.

Here, T>0T > 0 is the final time, and u0\mathbf{u}_0 the initial velocity. The constant ν>0\nu> 0 is the kinematic viscosity, pp is the pressure divided by the constant fluid density, and f\mathbf{f} and h\mathbf{h} use the corresponding density-normalized units. We write (u⋅n)−:=min⁡{u⋅n,0}(\mathbf{u} \cdot\mathbf{n})_- := \min\{\mathbf{u} \cdot\mathbf{n}, 0\}. This backflow term is motivated by Bruneau and Fabrie [17, 18] and related energy-stable open-boundary formulations [19, 43]. It acts only where u⋅n<0\mathbf{u} \cdot\mathbf{n} < 0, controlling the adverse energy contribution of flow entering through an open boundary. The Laplacian viscous operator gives the natural boundary quantity ν∂nu−pn\nu\partial_n\mathbf{u} - p\mathbf{n}, termed the pseudo-traction. The same Laplacian/pseudo-traction pairing is used in Refs. [6, 19].

Classical Rotational Pressure-Correction Scheme

Our development builds on the classical RPCS introduced by Timmermans, Minev, and Van de Vosse [9]. Let tn=nΔtt_n = n\Delta t, with uniform time step Δt\Delta t. For backward differentiation formulas (BDF), define the history by Un=∑j=1Jαjun+1−jU^n = \sum_{j=1}^{J}\alpha_j\mathbf{u}^{n+1-j}, so that the discrete time derivative is (γ0un+1−Un)/Δt(\gamma_0\mathbf{u}^{n+1} - U^n)/\Delta t. The first- and second-order schemes, denoted BDF1 and BDF2, use the following coefficients, history, and explicit advecting velocity:

BDF1:J=1,γ0=1,α1=1,Un=un,u∗=un,BDF2:J=2,γ0=32,(α1,α2)=(2,−12),Un=2un−12un−1,u∗=2un−un−1.(5)\begin{aligned} \mathrm{BDF1}:&\quad J=1,\quad\gamma_0=1,\quad\alpha_1=1,\\ &\quad U^n=\mathbf{u}^n,\quad\mathbf{u}^*=\mathbf{u}^n,\\ \mathrm{BDF2}:&\quad J=2,\quad\gamma_0=\frac{3}{2},\quad(\alpha_1,\alpha_2)=\left(2,-\frac{1}{2}\right),\\ &\quad U^n=2\mathbf{u}^n-\frac{1}{2}\mathbf{u}^{n-1},\quad\mathbf{u}^*=2\mathbf{u}^n-\mathbf{u}^{n-1}. \tag*{(5)} \end{aligned}

BDF2 computations are initialized with one BDF1 step. The extrapolated field u∗\mathbf{u}^* is fixed during each time step. Given the required velocity history and pnp^n, the update consists of four steps.

Step 1: Intermediate velocity. Compute u~n+1\widetilde{\mathbf{u}}^{n+1} from

γ0u~n+1−UnΔt+(u∗⋅∇)u~n+1−ν∇2u~n+1+∇pn=fn+1,in Ω.(6)\frac{\gamma_0\widetilde{\mathbf{u}}^{n+1}-U^n}{\Delta t}+(\mathbf{u}^*\cdot\nabla)\widetilde{\mathbf{u}}^{n+1}-\nu\nabla^2\widetilde{\mathbf{u}}^{n+1}+\nabla p^n=\mathbf{f}^{n+1},\qquad\text{in }\Omega. \tag*{(6)}

Step 2: Pressure increment. The pressure increment ϕn+1\phi^{n+1} is obtained from

Δtγ0∇2ϕn+1=∇⋅u~n+1,in Ω.(7)\frac{\Delta t}{\gamma_0}\nabla^2\phi^{n+1}=\nabla\cdot\widetilde{\mathbf{u}}^{n+1},\qquad\text{in }\Omega. \tag*{(7)}

Step 3: Velocity correction. The corrected velocity is obtained through

un+1=u~n+1−Δtγ0∇ϕn+1,in Ω.(8)\mathbf{u}^{n+1}=\widetilde{\mathbf{u}}^{n+1}-\frac{\Delta t}{\gamma_0}\nabla\phi^{n+1},\qquad\text{in }\Omega. \tag*{(8)}

Equations (7) and (8) imply ∇⋅un+1=0\nabla\cdot\mathbf{u}^{n+1}=0 at the continuous level.

Step 4: Rotational pressure recovery. The pressure is updated as

pn+1=pn+ϕn+1−ν∇⋅u~n+1,in Ω.(9)p^{n+1}=p^n+\phi^{n+1}-\nu\nabla\cdot\widetilde{\mathbf{u}}^{n+1},\qquad\text{in }\Omega. \tag*{(9)}

Boundary Conditions and Compatibility

Section 2 specifies the interior equations of the RPCS. We now prescribe boundary conditions for the intermediate velocity and pressure increment and examine their compatibility with the boundary data for the corrected fields. The discussion is conducted at the space-continuous, time-discrete level.

Open Boundary

On ΓN\Gamma_N, we freeze the backflow coefficient β∗:=(u∗⋅n)−\beta^* := (\mathbf{u}^*\cdot\mathbf{n})_- during each step to keep the boundary term linear. The target boundary relation for the corrected velocity and updated pressure is the following semi-implicit form of Eq. (4), motivated by energy-stable pressure-correction formulations [19]:

ν∂nun+1−pn+1n−12β∗un+1=hn+1,on ΓN.(10)\nu\partial_n\mathbf{u}^{n+1}-p^{n+1}\mathbf{n}-\frac{1}{2}\beta^*\mathbf{u}^{n+1}=\mathbf{h}^{n+1},\qquad\text{on }\Gamma_N. \tag*{(10)}

At the intermediate-velocity stage, the updated pressure is not yet available. We therefore use the known pressure pnp^n in the boundary condition, keeping this solve independent of the pressure increment:

ν∂nu~n+1−pnn−12β∗u~n+1=hn+1,on ΓN.(11)\nu\partial_n \widetilde{\boldsymbol{u}}^{n+1} - p^n \boldsymbol{n} - \frac{1}{2}\beta^* \widetilde{\boldsymbol{u}}^{n+1} = \boldsymbol{h}^{n+1}, \qquad\text{on } \Gamma_N. \tag*{(11)}

We next seek the condition on ϕn+1\phi^{n+1} needed to recover the normal component of Eq. (10). Substitute the velocity correction (8) and rotational pressure update (9) into this relation and take its normal component. Subtracting the normal component of Eq. (11) then gives

ϕn+1+Δtγ0∂nnϕn+1−ν∇⋅u~n+1−Δt2γ0β∗∂nϕn+1=0,on ΓN.(12)\phi^{n+1} + \frac{\Delta t}{\gamma_0}\partial_{nn}\phi^{n+1} - \nu\nabla\cdot\widetilde{\boldsymbol{u}}^{n+1} - \frac{\Delta t}{2\gamma_0}\beta^*\partial_n\phi^{n+1}=0, \qquad\text{on } \Gamma_N. \tag*{(12)}

Here ∂nnq=nT(∇2q)n\partial_{nn}q=\boldsymbol{n}^{\mathsf{T}}(\nabla^2q)\boldsymbol{n} for a smooth scalar field qq. To obtain a relation involving only ϕn+1\phi^{n+1}, we use Eq. (7) to write ∇⋅u~n+1=(Δt/γ0)∇2ϕn+1\nabla\cdot\widetilde{\boldsymbol{u}}^{n+1}=(\Delta t/\gamma_0)\nabla^2\phi^{n+1}. We assume sufficient regularity to restrict this equation and the second derivatives of ϕn+1\phi^{n+1} to the boundary. Splitting the Laplacian into normal and surface contributions then allows the second normal derivative in Eq. (12) to cancel.

On a smooth boundary, define the surface gradient ∇Γq=(I−n⊗n)∇q\nabla_\Gamma q=(I-\boldsymbol{n}\otimes\boldsymbol{n})\nabla q, the Laplace-Beltrami operator ΔΓq=∇Γ⋅∇Γq\Delta_\Gamma q=\nabla_\Gamma\cdot\nabla_\Gamma q, and the curvature κ=∇Γ⋅n\kappa=\nabla_\Gamma\cdot\boldsymbol{n}, where II is the identity tensor. The Laplacian decomposes as

∇2q=∂nnq+κ∂nq+ΔΓq.(13)\nabla^2q=\partial_{nn}q+\kappa\partial_nq+\Delta_\Gamma q. \tag*{(13)}

Substitution into Eq. (12) yields the surface relation

ϕn+1−Δtγ0[νΔΓϕn+1+(νκ+12β∗)∂nϕn+1]=0,on ΓN.(14)\phi^{n+1}-\frac{\Delta t}{\gamma_0}\left[\nu\Delta_\Gamma\phi^{n+1}+\left(\nu\kappa+\frac{1}{2}\beta^*\right)\partial_n\phi^{n+1}\right]=0, \qquad\text{on } \Gamma_N. \tag*{(14)}

This relation expresses normal pseudo-traction compatibility, including tangential diffusion, curvature, and backflow contributions. It does not enforce tangential pseudo-traction matching.

The outlets considered here are planar, so κ=0\kappa=0. If backflow is also absent, β∗=0\beta^*=0, Eq. (14) reduces to

ϕn+1−νΔtγ0ΔΓϕn+1=0,on ΓN.(15)\phi^{n+1}-\nu\frac{\Delta t}{\gamma_0}\Delta_\Gamma\phi^{n+1}=0, \qquad\text{on } \Gamma_N. \tag*{(15)}

If the boundary term arising from surface integration by parts vanishes, for example under periodic boundary conditions or homogeneous Neumann conditions on ∂ΓN\partial\Gamma_N, testing Eq. (15) with ϕn+1\phi^{n+1} gives

∥ϕn+1∥L2(ΓN)2+νΔtγ0∥∇Γϕn+1∥L2(ΓN)2=0.(16)\left\|\phi^{n+1}\right\|_{L^2(\Gamma_N)}^2+\nu\frac{\Delta t}{\gamma_0}\left\|\nabla_\Gamma\phi^{n+1}\right\|_{L^2(\Gamma_N)}^2=0. \tag*{(16)}

Since both terms are nonnegative,

ϕn+1=0,on ΓN.(17)\phi^{n+1}=0, \qquad\text{on } \Gamma_N. \tag*{(17)}

This conditional argument recovers the conventional homogeneous pressure-increment condition at an open boundary [6], which we adopt for implementation and analysis throughout this study. The above derivation nevertheless clarifies the assumptions under which this condition is exact and identifies the contributions neglected in more general settings.

Boundary Conditions for the RPCS Subproblems

On ΓD\Gamma_D, we impose u~n+1=uDn+1\widetilde{\boldsymbol{u}}^{n+1}=\boldsymbol{u}_D^{n+1}. Requiring the velocity correction (8) to preserve the prescribed normal component gives ∂nϕn+1=0\partial_n\phi^{n+1}=0. Together with the adopted open-boundary treatment, this gives the boundary conditions used in the intermediate-velocity and pressure-increment subproblems:

u~n+1=uDn+1,on ΓD,(18)\widetilde{\boldsymbol{u}}^{n+1}=\boldsymbol{u}_D^{n+1}, \qquad\text{on } \Gamma_D, \tag*{(18)}
ν∂nu~n+1−pnn−12β∗u~n+1=hn+1,on ΓN,(19)\nu\partial_n\widetilde{\boldsymbol{u}}^{n+1}-p^n\boldsymbol{n}-\frac{1}{2}\beta^*\widetilde{\boldsymbol{u}}^{n+1}=\boldsymbol{h}^{n+1}, \qquad\text{on } \Gamma_N, \tag*{(19)}
∂nϕn+1=0,on ΓD,(20)\partial_n\phi^{n+1}=0, \qquad\text{on } \Gamma_D, \tag*{(20)}
ϕn+1=0,on ΓN.(21)\phi^{n+1}=0, \qquad\text{on } \Gamma_N. \tag*{(21)}

The homogeneous Neumann condition for the pressure increment on ΓD\Gamma_D only preserves the normal component of the prescribed velocity. The full Dirichlet velocity data are enforced in the discrete velocity-correction step in § 4.5.

RPCS-VMS Formulation

Building on the RPCS algorithm and boundary treatment developed in the previous sections, we now construct a residual-based VMS formulation for equal-order velocity-pressure interpolation. The formulation coordinates stabilization of the intermediate-velocity, pressure-increment, and rotational pressure-recovery equations while preserving the four sequential solves. Throughout this section, we omit the superscript n+1n+1 on current unknowns and write δ=Δt/γ0\delta=\Delta t/\gamma_0.

Momentum Balances for Different Stages

We construct three momentum residuals for the VMS stabilization terms in the intermediate-velocity, pressure-increment, and rotational pressure-recovery equations. To make stabilization implicit in the current unknown of each step, we base these residuals on momentum balances involving pnp^n, pn+ϕp^n+\phi, and pp, respectively.

For smooth space-continuous, time-discrete RPCS fields, define the BDF accelerations associated with the intermediate and corrected velocities, together with the provisional pressure, by

a~=γ0u~−UnΔt,a=γ0u−UnΔt,π=pn+ϕ.\tilde{a}=\frac{\gamma_0\tilde{u}-U^n}{\Delta t},\qquad a=\frac{\gamma_0u-U^n}{\Delta t},\qquad\pi=p^n+\phi.

Both accelerations use the same known velocity history UnU^n, so their difference comes entirely from the current velocity correction. The first momentum balance is Eq. (6), which contains a~\tilde{a} and the known pressure gradient ∇pn\nabla p^n.

The velocity-correction equation (8) gives a~=a+∇ϕ\tilde{a}=a+\nabla\phi. Substituting this relation into the intermediate-velocity equation yields the second momentum balance associated with the pressure-increment step,

a+(u∗⋅∇)u~−ν∇2u~+∇π=fn+1.(22)a+(u^*\cdot\nabla)\tilde{u}-\nu\nabla^2\tilde{u}+\nabla\pi=f^{n+1}. \tag*{(22)}

Eq. (9) gives π=p+ν∇⋅u~\pi=p+\nu\nabla\cdot\tilde{u}. Substituting this expression into Eq. (22) and using ∇×∇×v=∇(∇⋅v)−∇2v\nabla\times\nabla\times v=\nabla(\nabla\cdot v)-\nabla^2v gives the third momentum balance associated with the rotational pressure-recovery step,

a+(u∗⋅∇)u~+ν∇×∇×u~+∇p=fn+1.(23)a+(u^*\cdot\nabla)\tilde{u}+\nu\nabla\times\nabla\times\tilde{u}+\nabla p=f^{n+1}. \tag*{(23)}

These three momentum balances are equivalent at the space-continuous, time-discrete level. The last two balances contain the corrected-velocity acceleration aa, whose evaluation in the pressure-increment step involves the unknown current velocity correction. Section 4.3 introduces a simplified acceleration choice to address this issue.

Multiscale Decomposition and Coarse-Scale Equations

In this section, we identify the subscale contributions retained in the four RPCS equations before specifying their algebraic models. Sections 4.3 and 4.4 close the velocity- and pressure-subscale contributions, respectively.

Let Th\mathcal{T}_h be a conforming mesh, M=L2(Ω)M=L^2(\Omega) the pressure space, and Mh⊂H1(Ω)⊂MM_h\subset H^1(\Omega)\subset M a continuous Lagrange space of degree ℓ≥1\ell\ge1, using PℓP_\ell or QℓQ_\ell elements. The homogeneous velocity test space and pressure-increment space are

X={v∈[H1(Ω)]d:v∣ΓD=0},X=\{v\in[H^1(\Omega)]^d:v|_{\Gamma_D}=0\},
Xh=[Mh]d∩X,Mh,0={qh∈Mh:qh∣ΓN=0}.(24)X_h=[M_h]^d\cap X,\qquad M_{h,0}=\{q_h\in M_h:q_h|_{\Gamma_N}=0\}. \tag*{(24)}

The intermediate and corrected velocities belong to Xh,Dn+1={vh∈[Mh]d:vh∣ΓD=uDn+1}X_{h,D}^{n+1}=\{v_h\in[M_h]^d:v_h|_{\Gamma_D}=u_D^{n+1}\}. For a closed domain, take Mh,0=Mh∩L02(Ω)M_{h,0}=M_h\cap L_0^2(\Omega), fix the pressure mean, and assume zero net prescribed flux. The L2(Ω)L^2(\Omega) inner product and norm are denoted by (⋅,⋅)(\cdot,\cdot) and ∥⋅∥\|\cdot\|, respectively. For each element K∈ThK\in\mathcal{T}_h, let (⋅,⋅)K(\cdot,\cdot)_K denote the L2(K)L^2(K) inner product. We define (⋅,⋅)h=∑K∈Th(⋅,⋅)K(\cdot,\cdot)_h=\sum_{K\in\mathcal{T}_h}(\cdot,\cdot)_K and the weighted norm ∥z∥τ2=(τz,z)h\|z\|_\tau^2=(\tau z,z)_h. A subscript hh on a differential operator denotes elementwise differentiation, while a boundary subscript on an inner product indicates integration over the corresponding boundary.

Form the velocity history UhnU_h^n and extrapolated advecting velocity uh∗u_h^* from Eq. (5). Define the discrete BDF accelerations associated with the intermediate and corrected velocities and the provisional pressure by

a~h=γ0u~h−UhnΔt,ah=γ0uh−UhnΔt,πh=phn+ϕh.(25)\tilde{a}_h=\frac{\gamma_0\tilde{u}_h-U_h^n}{\Delta t},\qquad a_h=\frac{\gamma_0u_h-U_h^n}{\Delta t},\qquad\pi_h=p_h^n+\phi_h. \tag*{(25)}

With β∗=(u∗⋅n)−\beta^*=(\boldsymbol{u}^*\cdot\boldsymbol{n})_-, the forcing and open-boundary data enter through L(vh)=(fn+1,vh)+(hn+1,vh)ΓN\mathcal{L}(v_h)=(f^{n+1},v_h)+(h^{n+1},v_h)_{\Gamma_N}.

Following the VMS framework, we decompose the intermediate velocity and pressure into resolved (coarse) and unresolved (fine) scales. After lifting the prescribed velocity data, let complementary subspaces X′⊂XX'\subset X and M′⊂MM'\subset M satisfy X=Xh⊕X′X=X_h\oplus X' and M=Mh⊕M′M=M_h\oplus M'. The resulting decompositions are

u~=u~h+u~′,u~′∈X′,p=ph+p′,p′∈M′.(26)\widetilde{u}=\widetilde{u}_h+\widetilde{u}',\qquad\widetilde{u}'\in X',\qquad p=p_h+p',\qquad p'\in M'. \tag*{(26)}

Here, u~h\widetilde{u}_h and php_h represent the resolved finite element fields, while u~′\widetilde{u}' and p′p' represent the unresolved scales. We model the velocity subscale locally using element bubbles, with u~′∣K∈[H01(K)]d\left.\widetilde{u}'\right|_K\in[H_0^1(K)]^d, following the VMS construction [34, 35].

To formulate the convective coupling between the resolved and unresolved velocities, we use the skew convection operator in the weak momentum equation. This choice avoids a spurious interior kinetic-energy contribution when the discrete advecting velocity is not pointwise divergence-free [44]. For an advecting velocity ww, define

(C(w)v,z):=((w⋅∇)v,z)+12((∇⋅w)v,z).(27)(C(w)v,z):=((w\cdot\nabla)v,z)+\frac{1}{2}((\nabla\cdot w)v,z). \tag*{(27)}

Testing this form with z=vz=v leaves only the convective boundary flux. The momentum residuals used in the algebraic subscale models retain the advective convection term from § 4.1, following the distinction discussed by Codina et al. [40]. The skew and advective convection expressions coincide when ∇⋅w=0\nabla\cdot w=0.

In the intermediate-velocity momentum equation, the coarse pressure is evaluated at the known value phnp_h^n. We neglect transient and viscous velocity-subscale feedback [26, 40] and retain the convective coupling (C(uh∗)u~′,vh)h(C(u_h^* )\widetilde{u}',v_h)_h. This coupling yields SUPG-type residual stabilization with a skew convection weighting, similar to the scale-splitting treatment of Colomés et al. [45]. The pressure-subscale contribution −(p′,∇⋅vh)h-(p',\nabla\cdot v_h)_h provides divergence stabilization through an algebraic model based on the continuity residual, as developed in § 4.4 [46].

Both the pressure-increment and rotational pressure-recovery equations depend on the divergence of the intermediate velocity. Substituting u~=u~h+u~′\widetilde{u}=\widetilde{u}_h+\widetilde{u}' therefore introduces a contribution from ∇⋅u~′\nabla\cdot\widetilde{u}' into each equation. We retain and model these contributions to construct stabilization for the two scalar problems, whose unknowns are ϕh∈Mh,0\phi_h\in M_{h,0} and ph∈Mhp_h\in M_h, respectively. This treatment is inspired by VMS formulations that model velocity subscales while approximating p≈php\approx p_h [40].

For velocity correction, we project u~h−δ∇ϕh\widetilde{u}_h-\delta\nabla\phi_h onto Xh,Dn+1X_{h,D}^{n+1}. This closure uses the resolved intermediate velocity. Testing the four RPCS steps with vh∈Xhv_h\in X_h, qh∈Mh,0q_h\in M_{h,0}, and rh∈Mhr_h\in M_h gives the modeled coarse equations in the order of the algorithm:

(a~h,vh)+(C(uh∗)u~h,vh)+ν(∇u~h,∇vh)−(phn,∇⋅vh)−(p′,∇⋅vh)h(\widetilde{a}_h,v_h)+(C(u_h^*)\widetilde{u}_h,v_h)+\nu(\nabla\widetilde{u}_h,\nabla v_h)-(p_h^n,\nabla\cdot v_h)-(p',\nabla\cdot v_h)_h
−12(β∗u~h,vh)ΓN+(C(uh∗)u~′,vh)h=L(vh).(28)-\frac{1}{2}(\beta^*\widetilde{u}_h,v_h)_{\Gamma_N}+(C(u_h^*)\widetilde{u}',v_h)_h=\mathcal{L}(v_h). \tag*{(28)}
δ(∇ϕh,∇qh)=−(∇⋅u~h,qh)−(∇⋅u~′,qh)h.(29)\delta(\nabla\phi_h,\nabla q_h)=-(\nabla\cdot\widetilde{u}_h,q_h)-(\nabla\cdot\widetilde{u}',q_h)_h. \tag*{(29)}
(uh,vh)=(u~h,vh)−δ(∇ϕh,vh).(30)(u_h,v_h)=(\widetilde{u}_h,v_h)-\delta(\nabla\phi_h,v_h). \tag*{(30)}
(ph,rh)=(πh,rh)−ν(∇⋅u~h,rh)−ν(∇⋅u~′,rh)h.(31)(p_h,r_h)=(\pi_h,r_h)-\nu(\nabla\cdot\widetilde{u}_h,r_h)-\nu(\nabla\cdot\widetilde{u}',r_h)_h. \tag*{(31)}

Velocity-Subscale Models

To close the velocity-subscale contributions in Eqs. (28a), (28b), and (28d), we construct a momentum residual for each equation from the balances in § 4.1. The two pressure-stage balances (Eqs. (22), (23)) involve the corrected-velocity acceleration aha_h. To express it in terms of the intermediate velocity, let Ph:[L2(Ω)]d→XhP_h:[L^2(\Omega)]^d\to X_h be the L2L^2 projection defined by (Phz,vh)=(z,vh)(P_hz,v_h)=(z,v_h) for every vh∈Xhv_h\in X_h. Eq. (28c) gives

uh=u~h−δPh∇ϕh,ah=a~h−Ph∇ϕh.u_h=\widetilde{u}_h-\delta P_h\nabla\phi_h,\qquad a_h=\widetilde{a}_h-P_h\nabla\phi_h.

In the residual based on Eq. (22), the acceleration correction combines with ∇πh\nabla\pi_h to give ∇phn+∇ϕh−Ph∇ϕh\nabla p_h^n+\nabla\phi_h-P_h\nabla\phi_h. Retaining this expression would place a velocity mass-matrix inverse inside the pressure-increment operator. We therefore replace aha_h by a~h\widetilde{a}_h in this residual, omitting −Ph∇ϕh-P_h\nabla\phi_h, so that the increment enters through its gradient alone.

For the rotational pressure-recovery equation, we use the same residual acceleration a~h\widetilde{a}_h, although the corrected velocity is already available. This common choice allows the two pressure-stabilization contributions to be combined in the energy balance of § 5. This modeling approximation is confined to the subscale residuals and leaves the velocity correction (30) unchanged. The temporal accuracy of the resulting formulation, including this residual approximation, is assessed through convergence study in the manufactured-solution tests of § 6.1.1.

The resulting momentum residuals for Eqs. (28a), (28b), and (28d), respectively, are

Rm=a~h+(uh∗⋅∇)u~h−ν∇h2u~h+∇phn−fn+1,(32)R_m = \widetilde{a}_h + (u_h^* \cdot\nabla)\widetilde{u}_h - \nu\nabla_h^2 \widetilde{u}_h + \nabla p_h^n - f^{n+1}, \tag*{(32)}
Rϕ=a~h+(uh∗⋅∇)u~h−ν∇h2u~h+∇πh−fn+1,(33)R_\phi= \widetilde{a}_h + (u_h^* \cdot\nabla)\widetilde{u}_h - \nu\nabla_h^2 \widetilde{u}_h + \nabla\pi_h - f^{n+1}, \tag*{(33)}
Rp=a~h+(uh∗⋅∇)u~h+ν∇h×∇h×u~h+∇ph−fn+1.(34)R_p = \widetilde{a}_h + (u_h^* \cdot\nabla)\widetilde{u}_h + \nu\nabla_h \times\nabla_h \times\widetilde{u}_h + \nabla p_h - f^{n+1}. \tag*{(34)}

The residual RmR_m depends on the current intermediate velocity. After that velocity is computed, RϕR_\phi and RpR_p depend on their respective scalar unknowns through ∇πh=∇(phn+ϕh)\nabla\pi_h = \nabla(p_h^n + \phi_h) and ∇ph\nabla p_h.

We integrate the retained bubble contributions by parts within each element, transferring derivatives to the coarse test functions. The zero bubble traces eliminate the boundary terms in this integration. Following Hughes’s construction, the local fine-scale response is formally represented through an element Green’s operator [34, 35]. We approximate its contribution to the coarse-test integrals using the stage-specific algebraic models

u~m′≈−τmRm,u~ϕ′≈−τmRϕ,u~p′≈−τmRp,(35)\widetilde{u}_m' \approx-\tau_m R_m,\qquad\widetilde{u}_\phi' \approx-\tau_m R_\phi,\qquad\widetilde{u}_p' \approx-\tau_m R_p, \tag*{(35)}

where τm>0\tau_m > 0 summarizes the local response. The algebraic expressions are substituted into the coarse-test integrals, giving the following contributions on the left-hand sides of the three equations:

−(u~m′,C(uh∗)vh)h≈(τmRm,C(uh∗)vh)h,-(\widetilde{u}_m', C(u_h^*)v_h)_h \approx(\tau_m R_m, C(u_h^*)v_h)_h,
−(u~ϕ′,∇qh)h≈(τmRϕ,∇qh)h,-(\widetilde{u}_\phi', \nabla q_h)_h \approx(\tau_m R_\phi, \nabla q_h)_h,
−ν(u~p′,∇rh)h≈ν(τmRp,∇rh)h.(36)-\nu(\widetilde{u}_p', \nabla r_h)_h \approx\nu(\tau_m R_p, \nabla r_h)_h. \tag*{(36)}

The first is the convective stabilization term introduced above. Together with the continuity-residual model for p′p' in § 4.4, it stabilizes the intermediate-velocity equation and controls its divergence. The other two terms pair the appropriate momentum residual with a scalar test-function gradient, supplying stabilization to the pressure-increment and rotational pressure-recovery equations. The resulting scheme admits a conditional bound on the pressure gradient for the reduced model analyzed in § 5.

For all three models, we use the metric-based momentum scale in element KK [47]

τm,K=[(γ0Δt)2+uh,K∗⋅GKuh,K∗+CIν2GK:GK]−1/2,CI>0.(37)\tau_{m,K} = \left[\left(\frac{\gamma_0}{\Delta t}\right)^2 + u_{h,K}^* \cdot G_K u_{h,K}^* + C_I \nu^2 G_K : G_K\right]^{-1/2},\qquad C_I > 0. \tag*{(37)}

Here uh,K∗u_{h,K}^* is a representative element velocity, and (GK)ij=∑a=1d(∂ξa/∂xi)(∂ξa/∂xj)(G_K)_{ij} = \sum_{a=1}^{d}(\partial\xi_a/\partial x_i)(\partial\xi_a/\partial x_j), where ξa\xi_a are reference coordinates. GK:GK=∑i,j(GK)ij2G_K : G_K = \sum_{i,j}(G_K)_{ij}^2 scales as inverse length to the fourth power and is invariant under orthogonal changes of reference coordinates. The time-step term ensures 0<τm,K≤δ≤Δt0 < \tau_{m,K} \le\delta\le\Delta t.

Pressure-Subscale Model

The pressure-subscale term in Eq. (28a) provides divergence control through a continuity-residual model. This follows the VMS interpretation of grad-div stabilization as an algebraic pressure-subscale contribution [46, 48]. For RPCS, the relevant continuity relation includes the pressure increment. Taking the divergence of the space-continuous velocity correction gives

∇⋅u=∇⋅u~−δ∇2ϕ=0.(38)\nabla\cdot u = \nabla\cdot\widetilde{u} - \delta\nabla^2 \phi= 0. \tag*{(38)}

Evaluating this relation on the finite-element fields defines the elementwise residual

Rc=∇⋅u~h−δ∇h2ϕh.(39)R_c = \nabla\cdot\widetilde{u}_h - \delta\nabla_h^2 \phi_h. \tag*{(39)}

We represent the pressure subscale by the algebraic field p′=−τcRcp' = -\tau_c R_c, with a nonnegative elementwise coefficient τc\tau_c. Its contribution to the intermediate-velocity equation is

−(p′,∇⋅vh)h=(τcRc,∇⋅vh)h.(40)-(p', \nabla\cdot v_h)_h = (\tau_c R_c, \nabla\cdot v_h)_h. \tag*{(40)}

For affine P1P_1 elements, the elementwise pressure Laplacian vanishes, and the pressure-subscale model reduces to classical grad–div stabilization [46]. When the elementwise pressure Laplacian is nonzero, the pressure-increment contribution can be evaluated using extrapolation from previous time levels, with ϕh∗=ϕhn\phi_h^*=\phi_h^n for BDF1 and ϕh∗=2ϕhn−ϕhn−1\phi_h^*=2\phi_h^n-\phi_h^{n-1} for BDF2. This preserves the sequential structure of the four-step scheme.

To set the strength of the pressure-subscale contribution, we use the metric-based continuity scale [48, 49], with a nonnegative element factor αc,K\alpha_{c,K}:

τc,K=αc,Kτm,K(GK:GK)1/2,αc,K≥0.(41)\tau_{c,K}=\frac{\alpha_{c,K}}{\tau_{m,K}(G_K:G_K)^{1/2}},\qquad\alpha_{c,K}\ge0. \tag*{(41)}

αc,K\alpha_{c,K} is dimensionless and τc,K\tau_{c,K} has the units of kinematic viscosity. The factor αc,K\alpha_{c,K} may vary by element, and both stabilization coefficients are fixed during each step.

Fully Discrete RPCS-VMS Scheme

We now incorporate the velocity- and pressure-subscale models into the four-step RPCS framework to obtain the fully discrete RPCS-VMS scheme. At each step, form the BDF data from Eq. (5), the coefficients from Eqs. (37) and (41), and the extrapolated pressure increment from § 4.4 if needed. Keep these data and β∗\beta^* fixed during the four solves. The tests are vh∈Xhv_h\in X_h, qh∈Mh,0q_h\in M_{h,0}, and rh∈Mhr_h\in M_h.

Step 1: Intermediate velocity. Given the velocity history and phnp_h^n, solve for the intermediate velocity u~h∈Xh,Dn+1\widetilde{u}_h\in X_{h,D}^{n+1}:

(γ0u~h−UhnΔt,vh)+(C(uh∗)u~h,vh)+ν(∇u~h,∇vh)−(phn,∇⋅vh)−12(β∗u~h,vh)ΓN+(τmRm,C(uh∗)vh)h+(τc[∇⋅u~h−Δtγ0∇h2ϕh∗],∇⋅vh)h=(fn+1,vh)+(hn+1,vh)ΓN.(42)\left(\frac{\gamma_0\widetilde{u}_h-U_h^n}{\Delta t},v_h\right)+(C(u_h^*)\widetilde{u}_h,v_h)+\nu(\nabla\widetilde{u}_h,\nabla v_h)-(p_h^n,\nabla\cdot v_h)-\frac{1}{2}(\beta^*\widetilde{u}_h,v_h)_{\Gamma_N}+(\tau_m R_m,C(u_h^*)v_h)_h+\left(\tau_c\left[\nabla\cdot\widetilde{u}_h-\frac{\Delta t}{\gamma_0}\nabla_h^2\phi_h^*\right],\nabla\cdot v_h\right)_h=(f^{n+1},v_h)+(h^{n+1},v_h)_{\Gamma_N}. \tag*{(42)}

Step 2: Pressure increment. Using u~h\widetilde{u}_h from Step 1, evaluate RmR_m. With Rϕ=Rm+∇ϕhR_\phi=R_m+\nabla\phi_h, solve for ϕh∈Mh,0\phi_h\in M_{h,0}:

Δtγ0(∇ϕh,∇qh)+(τm∇ϕh,∇qh)h=−(∇⋅u~h,qh)−(τmRm,∇qh)h.(43)\frac{\Delta t}{\gamma_0}(\nabla\phi_h,\nabla q_h)+(\tau_m\nabla\phi_h,\nabla q_h)_h=-(\nabla\cdot\widetilde{u}_h,q_h)-(\tau_mR_m,\nabla q_h)_h. \tag*{(43)}

Step 3: Velocity correction. Using ϕh\phi_h from Step 2, solve for the corrected velocity uh∈Xh,Dn+1u_h\in X_{h,D}^{n+1}:

(uh,vh)=(u~h,vh)−Δtγ0(∇ϕh,vh).(44)(u_h,v_h)=(\widetilde{u}_h,v_h)-\frac{\Delta t}{\gamma_0}(\nabla\phi_h,v_h). \tag*{(44)}

Step 4: Rotational pressure recovery. Using u~h\widetilde{u}_h and ϕh\phi_h, solve for ph∈Mhp_h\in M_h:

(ph,rh)+ν(τm∇ph,∇rh)h=(phn+ϕh,rh)−ν(∇⋅u~h,rh)−ν(τm[γ0u~h−UhnΔt+(uh∗⋅∇)u~h+ν∇h×∇h×u~h−fn+1],∇rh)h.(45)(p_h,r_h)+\nu(\tau_m\nabla p_h,\nabla r_h)_h=(p_h^n+\phi_h,r_h)-\nu(\nabla\cdot\widetilde{u}_h,r_h)-\nu\left(\tau_m\left[\frac{\gamma_0\widetilde{u}_h-U_h^n}{\Delta t}+(u_h^*\cdot\nabla)\widetilde{u}_h+\nu\nabla_h\times\nabla_h\times\widetilde{u}_h-f^{n+1}\right],\nabla r_h\right)_h. \tag*{(45)}

Stability Analysis of the RPCS-VMS Formulation

To isolate the interaction between pressure stabilization and rotational recovery, we analyze a reduced BDF1 version of the four-step scheme in § 4.5. The body force and prescribed velocity and pseudo-traction data are zero. Related analyses have examined pressure segregation without convective stabilization [42] and rotational pressure correction for convection-free problems [11, 21]. Following these studies, the intermediate-velocity equation retains skew-symmetric convection and the backflow boundary term, but omits SUPG stabilization. The two pressure equations retain their stabilization terms, with convection omitted from the associated momentum residuals.

Discrete energy structure

We use continuous P1/P1P_1/P_1 fields on a fixed affine simplicial mesh, and their elementwise second derivatives vanish. The mesh is connected through shared faces and ΓD\Gamma_D has positive surface measure. The initial fields belong to Xh×MhX_h\times M_h, with a prescribed pressure mean for a closed domain. At each step the known advecting field is whn+1=uh∗w_h^{n+1}=u_h^*, with β∗=(whn+1⋅n)−\beta^*=(w_h^{n+1}\cdot n)_-. The skew convection form and backflow term together give a nonnegative outflow contribution: for vh∈Xhv_h\in X_h,

2(C(whn+1)vh,vh)−(β∗vh,vh)ΓN=∫ΓN(whn+1⋅n)+∣vh∣2 ds≥0,s+=max⁡(s,0).(46)2(C(w_h^{n+1})v_h,v_h)-(\beta^*v_h,v_h)_{\Gamma_N}=\int_{\Gamma_N}(w_h^{n+1}\cdot n)_+|v_h|^2\,\mathrm{d}s\geq0,\qquad s_+=\max(s,0). \tag*{(46)}

For the stabilization coefficients, define the extrema over the mesh by

τmmin⁡=min⁡Kτm,K,τmmax⁡=max⁡Kτm,K,τcmin⁡=min⁡Kτc,K.\tau_m^{\min}=\min_K\tau_{m,K},\qquad\tau_m^{\max}=\max_K\tau_{m,K},\qquad\tau_c^{\min}=\min_K\tau_{c,K}.

For reasons that will become clear in the energy and pressure estimates below, assume that the element coefficients are constant in time and satisfy

0<τmmin⁡≤τmmax⁡≤Δt,κc:=2τcmin⁡−(1−2d)ν≥0.(47)0<\tau_m^{\min}\leq\tau_m^{\max}\leq\Delta t,\qquad\kappa_c:=2\tau_c^{\min}-\left(1-\frac{2}{d}\right)\nu\geq0. \tag*{(47)}

The condition on τc\tau_c requires τc,K≥0\tau_{c,K}\geq0 in two dimensions and τc,K≥ν/6\tau_{c,K}\geq\nu/6 in three dimensions. Recall from § 4.2 that πhn+1=phn+ϕhn+1\pi_h^{n+1}=p_h^n+\phi_h^{n+1} is the provisional pressure. The elementwise viscous terms in Eq. (29) vanish for P1P_1 velocities. With zero body force and convection omitted from these residuals, the momentum residuals used in the pressure-increment and pressure-recovery equations reduce to Rϕn+1=a~hn+1+∇πhn+1R_\phi^{n+1}=\widetilde{a}_h^{n+1}+\nabla\pi_h^{n+1} and Rpn+1=a~hn+1+∇phn+1R_p^{n+1}=\widetilde{a}_h^{n+1}+\nabla p_h^{n+1}, where a~hn+1=(u~hn+1−uhn)/Δt\widetilde{a}_h^{n+1}=(\widetilde{u}_h^{n+1}-u_h^n)/\Delta t is the BDF1 intermediate-velocity acceleration. The four-step equations used in the analysis are listed in Appendix § 12.§ 1. To convert the pressure-increment contribution into an energy difference, introduce the accumulated increment ψhn∈Mh,0\psi_h^n\in M_{h,0}:

ψh0=0,ψhn+1=ψhn+ϕhn+1.(48)\psi_h^0=0,\qquad\psi_h^{n+1}=\psi_h^n+\phi_h^{n+1}. \tag*{(48)}

Since ϕhn+1=ψhn+1−ψhn\phi_h^{n+1}=\psi_h^{n+1}-\psi_h^n, testing the pressure increment equation with ψhn+1\psi_h^{n+1} produces a difference of squared gradient norms. To account for rotational recovery in the same balance, define ηhn=ψhn−phn∈Mh\eta_h^n=\psi_h^n-p_h^n\in M_h. Its change is exactly the difference between the provisional and final pressures:

δηhn+1:=ηhn+1−ηhn=πhn+1−phn+1.\delta\eta_h^{n+1}:=\eta_h^{n+1}-\eta_h^n=\pi_h^{n+1}-p_h^{n+1}.

The residual definitions in Section 4.3 imply Rp−Rϕ=∇h(ph−πh+ν∇⋅u~h)R_p-R_\phi=\nabla_h(p_h-\pi_h+\nu\nabla\cdot\widetilde{u}_h). On the present P1P_1 mesh, ∇h(∇⋅u~h)=0\nabla_h(\nabla\cdot\widetilde{u}_h)=0, so

Rpn+1−Rϕn+1=−∇δηhn+1.R_p^{n+1}-R_\phi^{n+1}=-\nabla\delta\eta_h^{n+1}.

This relation connects the two pressure-stabilization contributions when completing squares in the energy balance, as detailed in Appendix § 12.§ 1. The sequences ψhn\psi_h^n and ηhn\eta_h^n are introduced only for the analysis and require no additional pressure solves. The relation πhn+1=ψhn+1−ηhn\pi_h^{n+1}=\psi_h^{n+1}-\eta_h^n connects the increment and recovery tests. Testing the pressure-recovery relation with ηhn\eta_h^n completes the conversion of pressure work into differences of squared norms. These squared norms form the pressure terms in the augmented energy EnE^n. The viscous and divergence contributions are grouped in Dh\mathcal{D}_h:

En=∥uhn∥2+(Δt)2∥∇ψhn∥2+Δtν∥ηhn∥2+Δt∥∇ηhn∥τm2,E^n=\|u_h^n\|^2+(\Delta t)^2\|\nabla\psi_h^n\|^2+\frac{\Delta t}{\nu}\|\eta_h^n\|^2+\Delta t\|\nabla\eta_h^n\|_{\tau_m}^2,
Dh(vh)=2ν∥∇vh∥2+2∥∇⋅vh∥τc2−ν∥∇⋅vh∥2.(49)\mathcal{D}_h(v_h)=2\nu\|\nabla v_h\|^2+2\|\nabla\cdot v_h\|_{\tau_c}^2-\nu\|\nabla\cdot v_h\|^2. \tag*{(49)}

Under Eq. (47), Dh\mathcal{D}_h is positive definite on XhX_h, including at κc=0\kappa_c=0 (Appendix § 12.§ 1). This property also provides the divergence control needed for the pressure estimate.

Energy decay

Testing the intermediate-velocity equation and applying the projected correction gives an energy balance for the corrected velocity. The pressure-increment and pressure-recovery equations then express pressure work as changes in the auxiliary energy terms. Since τm\tau_m is constant in time, the weighted energy differences telescope when summed over successive steps.

To express the remaining terms in the energy balance, write Qh=I−PhQ_h=I-P_h, the complement of the velocity projection, where II is the identity on [L2(Ω)]d[L^2(\Omega)]^d. We collect the contributions associated with the time increment, projected correction, viscous and stabilization terms, and outflow as

Wn=∑K(1−τm,KΔt)∥uhn+1−uhn∥L2(K)2,Jn=(Δt)2∥Qh∇ϕhn+1∥2,Fn=Dh(u~hn+1)+∥∇πhn+1∥τm2,Bn=∫ΓN(whn+1⋅n)+∣u~hn+1∣2 ds.(50)\begin{aligned} W_n &= \sum_K \left(1-\frac{\tau_{m,K}}{\Delta t}\right)\left\|u_h^{n+1}-u_h^n\right\|_{L^2(K)}^2,\\ J_n &= (\Delta t)^2\left\|Q_h\nabla\phi_h^{n+1}\right\|^2,\\ F_n &= \mathcal{D}_h(\widetilde{u}_h^{n+1})+\left\|\nabla\pi_h^{n+1}\right\|_{\tau_m}^2,\\ B_n &= \int_{\Gamma_N}\left(w_h^{n+1}\cdot n\right)_+|\widetilde{u}_h^{n+1}|^2\,ds. \tag*{(50)} \end{aligned}

while pressure recovery contributes

Sn=Δt∥Rpn+1∥τm2+Δtν∥δηhn+1−ν∇⋅u~hn+1∥2.(51)S_n=\Delta t\left\|R_p^{n+1}\right\|_{\tau_m}^2+\frac{\Delta t}{\nu}\left\|\delta\eta_h^{n+1}-\nu\nabla\cdot\widetilde{u}_h^{n+1}\right\|^2. \tag*{(51)}

To complete the balance, we take rh=δηhn+1r_h=\delta\eta_h^{n+1} in the pressure-recovery relation, Eq. (A.6), and combine the residual and divergence contributions by completing squares. This produces SnS_n and the pressure-gradient term in FnF_n, together with a negative squared norm of the intermediate-velocity acceleration. That negative term combines with the squared velocity increment from the test of the intermediate-velocity equation to give WnW_n. Appendix § 12.§ 1 provides the full derivation.

Under the assumptions of Section 5.1, the solution of the reduced four-step scheme satisfies the discrete energy-dissipation identity

En+1−En+Wn+Jn+ΔtFn+Sn+ΔtBn=0.(52)E^{n+1}-E^n+W_n+J_n+\Delta t F_n+S_n+\Delta t B_n=0. \tag*{(52)}

Hence, for every N≥1N\ge1,

EN+∑n=0N−1(Wn+Jn+ΔtFn+Sn+ΔtBn)=E0,En+1≤En.(53)E^N+\sum_{n=0}^{N-1}\left(W_n+J_n+\Delta t F_n+S_n+\Delta t B_n\right)=E^0,\qquad E^{n+1}\le E^n. \tag*{(53)}

The upper bound τmmax⁡≤Δt\tau_m^{\max}\le\Delta t ensures Wn≥0W_n\ge0, and the condition κc≥0\kappa_c\ge0 ensures Dh≥0\mathcal{D}_h\ge0. The remaining contributions are squared norms or the nonnegative outflow flux in Eq. (46). Thus all dissipation terms in Eq. (52) are nonnegative, giving energy decay for both closed and open boundaries.

Pressure-gradient bounds

The energy identity controls the gradient of the provisional pressure through ∥∇πhn+1∥τm2\left\|\nabla\pi_h^{n+1}\right\|_{\tau_m}^2 in FnF_n. To obtain a bound for the gradient of the updated pressure phn+1p_h^{n+1}, we use phn+1=πhn+1−δηhn+1p_h^{n+1}=\pi_h^{n+1}-\delta\eta_h^{n+1}. The second term in SnS_n controls δηhn+1−ν∇⋅u~hn+1\delta\eta_h^{n+1}-\nu\nabla\cdot\widetilde{u}_h^{n+1}, and Dh\mathcal{D}_h controls the divergence. Together they bound the change δηhn+1\delta\eta_h^{n+1}. A finite-element inverse estimate then bounds its gradient. Summing these contributions in time and using the energy identity gives, for τmmin⁡>0\tau_m^{\min}>0, a finite constant Cp,hC_{p,h}, independent of the number of steps, such that

Δt∑n=0N−1∥∇phn+1∥2≤Cp,hτmmin⁡E0.(54)\Delta t\sum_{n=0}^{N-1}\left\|\nabla p_h^{n+1}\right\|^2\le\frac{C_{p,h}}{\tau_m^{\min}}E^0. \tag*{(54)}

The positive lower bound τmmin⁡\tau_m^{\min} converts the weighted pressure-gradient control into this unweighted estimate. The constant Cp,hC_{p,h} depends on the mesh and frozen coefficients and is specified in Appendix § 12.§ 2. On a fixed mesh, Eq. (37) gives τmmin⁡\tau_m^{\min} proportional to Δt\Delta t as Δt→0\Delta t\to0, provided the advecting velocity is bounded in L∞(Ω)L^\infty(\Omega) uniformly in the time level and Δt\Delta t. The explicit factor 1/τmmin⁡1/\tau_m^{\min} therefore grows as (Δt)−1(\Delta t)^{-1}.

For comparison, we remove the τm\tau_m-terms and prescribe τc\tau_c independently so that κc≥0\kappa_c\ge0. Energy decay still holds, but the pressure-gradient dissipation is absent. With E00=∥uh0∥2+(Δt/ν)∥ph0∥2E_0^0=\|u_h^0\|^2+(\Delta t/\nu)\|p_h^0\|^2 and T=NΔtT=N\Delta t, Appendix § 12.§ 3 gives

Δt∑n=0N−1∥∇phn+1∥2≤2T(Δt)2max⁡{1,νΔtC∇,h2}E00.(55)\Delta t\sum_{n=0}^{N-1}\left\|\nabla p_h^{n+1}\right\|^2\le\frac{2T}{(\Delta t)^2}\max\left\{1,\nu\Delta t C_{\nabla,h}^2\right\}E_0^0. \tag*{(55)}

Here C∇,hC_{\nabla,h} is a mesh-dependent constant satisfying ∥∇rh∥≤C∇,h∥rh∥\lVert\nabla r_h \rVert\le C_{\nabla,h}\lVert r_h \rVert for every rh∈Mhr_h \in M_h. At fixed mesh and final time, the factor on the right grows as (Δt)−2(\Delta t)^{-2}, compared with the explicit (Δt)−1(\Delta t)^{-1} factor in the stabilized estimate when τmmin⁡\tau_m^{\min} is proportional to Δt\Delta t. Stabilization thus supplies pressure-gradient dissipation and improves the time-step dependence of the bound.

As the time step decreases, the estimate without momentum-residual stabilization provides progressively weaker pressure-gradient control and does not rule out pressure oscillations. This behavior is consistent with the small-time-step pressure instability discussed by Codina [32]. In particular, a pressure mode orthogonal to all discrete velocity divergences can persist unchanged despite the nonincrease of energy. Appendix § 12.§ 3 gives this construction.

Numerical Results

In this section, we present the numerical results of the RPCS-VMS method. Manufactured solutions are used to examine spatial and temporal convergence, near-boundary pressure errors, and pressure stability at small time steps. The branched-pipe benchmark assesses flow partitioning between two open outlets, while two-dimensional (2D) and three-dimensional (3D) cylinder benchmarks evaluate transient forces and pressure predictions against FEATFLOW reference data [50, 51]. The 3D cylinder case also serves to assess computational efficiency relative to a monolithic VMS formulation.

The computations use FEniCSx 0.10.0 [52] and PETSc 3.24.3 [53]. The stabilized equal-order discretizations use CI=36C_I = 36 in the momentum scale of Eq. (37), following Brunelli et al. [54]. To meet the divergence-penalty threshold in Eq. (47), the divergence parameter is set elementwise to

αc,K=(1−2d)ν2τ~m,K(GK:GK)1/2.(56)\alpha_{c,K} = \left(1-\frac{2}{d}\right)\frac{\nu}{2}\widetilde{\tau}_{m,K}(G_K : G_K)^{1/2}. \tag*{(56)}

This leads to τc=0\tau_c = 0 in 2D and τc=ν/6\tau_c = \nu/6 in 3D. Except for the explicitly modified variants in § 6.1.4, all RPCS-VMS computations use the full formulation of § 4.5.

The intermediate-velocity system is solved using the generalized minimal residual method (GMRES), preconditioned with BoomerAMG, an implementation of algebraic multigrid (AMG). The pressure-increment and pressure-recovery systems are solved using the conjugate gradient method (CG), also preconditioned with BoomerAMG, while the velocity correction system is solved using CG with Jacobi preconditioning. All four solves use a relative tolerance of 10−810^{-8}, an absolute tolerance of 10−1010^{-10}, and a maximum of 1000 iterations. The GMRES restart is set to 100. Table 1 summarizes the meshes, unknown counts, and time steps for the physical cases.

CaseVertices/cellsElement pairNo. of unknownsΔt\Delta t, TT
Branched pipe8,086/15,293P1/P1P_1/P_116,172/8,0865×10−45 \times10^{-4}, 50
2D cylinder33,742/33,268Q1/Q1Q_1/Q_167,484/33,7426.25×10−46.25 \times10^{-4}, 8
3D cylinder410,800/385,920Q1/Q1Q_1/Q_11,232,400/410,8005×10−35 \times10^{-3}, 8

Table 1. Mesh and time-integration settings for the physical benchmarks. Unknown counts are listed as velocity/pressure, and TT is the final simulation time.

Manufactured-Solution Tests

The convergence and pressure tests use a common manufactured Navier–Stokes solution on the square domain Ω=[−0.5,0.5]2\Omega= [-0.5, 0.5]^2. The right edge, x=0.5x = 0.5, carries the open-boundary condition, while the exact velocity is prescribed on the other three edges. Following the spatial fields of Liu [12], set

us=(sin⁡(πy)cos⁡2(πx/2)−sin⁡(πx)cos⁡2(πy/2)),ps=cos⁡(πx/2)sin⁡(πy/2),(u,p)=A(t)(us,ps).(57)\mathbf{u}_s = \begin{pmatrix} \sin(\pi y)\cos^2(\pi x/2) \\ -\sin(\pi x)\cos^2(\pi y/2) \end{pmatrix}, \qquad p_s = \cos(\pi x/2)\sin(\pi y/2), \qquad (\mathbf{u},p) = A(t)(\mathbf{u}_s,p_s). \tag*{(57)}

The spatial velocity field is divergence-free. We use ν=1\nu= 1, with A=1A = 1 for the steady tests. The unsteady tests use A(t)=cos⁡(2ωt)A(t) = \cos(2\omega t) with ω=π\omega= \pi and T=1T = 1, and the latter tests the near-boundary pressure comparison at T=0.5T = 0.5. The body force and open-boundary data are obtained by substituting these fields into Eqs. (1) and (4), respectively.

Spatial and Temporal Convergence

We first examine spatial convergence against the steady solution, using BDF1 pseudo-time stepping with Δt=10−2\Delta t = 10^{-2} to approach steady state. Simulations are performed on P1/P1P_1/P_1 and Q1/Q1Q_1/Q_1 meshes constructed from n×nn \times n grids, with n=4,8,…,256n = 4, 8, \ldots, 256. Figure 1 shows second-order velocity convergence in L2L^2, first-order convergence in the full H1H^1 norm, and approximately order-1.5 pressure convergence for both element types.

Figure 1. Spatial convergence of RPCS-VMS for the steady manufactured solution with P1/P1P_1/P_1 and Q1/Q1Q_1/Q_1 elements. Reference slopes indicate the convergence orders.

We next examine temporal convergence against the unsteady reference solution using Q1/Q1Q_1/Q_1 elements and successively halved time steps ranging from 0.1 to 0.003125. Figure 2 shows first-order convergence in the velocity and pressure L2L^2 errors for BDF1 and second-order convergence for BDF2. As the time step decreases, spatial discretization errors become significant and limit further reduction in errors, with the velocity H1H^1 error reaching this limit earlier.

Figure 2. Temporal convergence of RPCS-VMS for the unsteady manufactured solution with Q1/Q1Q_1/Q_1 elements. BDF1 uses n=256n = 256 and BDF2 uses n=512n = 512. Errors are evaluated at T=1T = 1.

The observed spatial velocity convergence rates are consistent with standard error estimates for stabilized low-order finite element approximations of smooth solutions [55]. The pressure L2L^2 convergence rate of approximately 1.5 is consistent with that previously reported for a stabilized Q1/Q1Q_1/Q_1 monolithic discretization of Stokes flow [24].

Near-Boundary Pressure Errors

To investigate the spatial distribution of pressure errors, we compare three formulations: RPCS-VMS, the non-rotational formulation, and Monolithic-VMS. All three use BDF2 with Δt=10−3\Delta t = 10^{-3} on the same P1/P1P_1/P_1 mesh constructed from a 64×6464 \times64 uniform grid and are evaluated at T=0.5T = 0.5. The half-period endpoint reveals the boundary-pressure error before the exact pressure returns to its initial value at T=1T = 1. The non-rotational formulation retains the pressure-increment stabilization but replaces the rotational pressure-recovery equation with phn+1=phn+ϕhn+1p_h^{n+1} = p_h^n + \phi_h^{n+1}. Monolithic-VMS denotes the monolithic equal-order formulation with SUPG/PSPG stabilization and the grad-div term governed by τc\tau_c [26]. All three formulations use the same τm\tau_m model and solver tolerances.

Let the boundary element strip and its interior complement be

Ωhb=⋃{K∈Th:K has a facet on ∂Ω},Ωhi=Ω∖Ωhb.(58)\Omega_h^b = \bigcup\{K \in\mathcal{T}_h : K\text{ has a facet on }\partial\Omega\}, \qquad\Omega_h^i = \Omega\setminus\Omega_h^b. \tag*{(58)}

Table 2 reports L2L^2 errors over the full domain, boundary strip, and interior, with ∥⋅∥D\|\cdot\|_D denoting the L2(D)L^2(D) norm. As shown in Figure 3, the pressure error is concentrated near the open boundary for non-rotational formulation, because Eq. (21) imposed on the pressure increment keeps the non-rotational pressure at its initial value along this boundary. Relative to this formulation, the stabilized rotational recovery treatment reduces the boundary-strip pressure error from 3.4536×10−23.4536 \times10^{-2} to 2.6738×10−32.6738 \times10^{-3}, a factor of 12.92, and the global pressure error by a factor of 12.47. RPCS-VMS and Monolithic-VMS errors agree to within 0.4%, with nearly identical boundary-adjacent patterns in Figure 3. Localized stabilization-induced pressure perturbations remain in both formulations.

Method∣p−ph∣Ω|p-p_h|_{\Omega}∣p−ph∣Ωhb|p-p_h|_{\Omega_h^b}∣p−ph∣Ωhi|p-p_h|_{\Omega_h^i}∣u−uh∣Ω|u-u_h|_{\Omega}
Non-rotational formulation4.0855×10−24.0855 \times10^{-2}3.4536×10−23.4536 \times10^{-2}2.1826×10−22.1826 \times10^{-2}6.5958×10−46.5958 \times10^{-4}
RPCS-VMS3.2775×10−33.2775 \times10^{-3}2.6738×10−32.6738 \times10^{-3}1.8955×10−31.8955 \times10^{-3}3.0185×10−43.0185 \times10^{-4}
Monolithic-VMS3.2725×10−33.2725 \times10^{-3}2.6703×10−32.6703 \times10^{-3}1.8917×10−31.8917 \times10^{-3}3.0077×10−43.0077 \times10^{-4}

Table 2. Unsteady manufactured-solution L2L^2 errors at T=0.5T=0.5 with BDF2, Δt=10−3\Delta t=10^{-3}, and n=64n=64. All methods use P1/P1P_1/P_1 elements. Pressure errors are calculated over the full domain Ω\Omega, boundary strip Ωhb\Omega_h^b, and interior Ωhi\Omega_h^i. The velocity error is evaluated over the full domain.

Figure 3. Unsteady manufactured-solution pressure errors ph−pp_h-p at T=0.5T=0.5 for the non-rotational formulation, RPCS-VMS, and Monolithic-VMS. The common symmetric logarithmic color scale is linear for ∣ph−p∣≤0.005|p_h-p|\leq0.005.

Stability at Small Time Steps

We test small-time-step pressure stability using BDF2 and the unsteady manufactured solution on a 64×6464 \times64 grid at T=1T=1. Figure 4 compares unstabilized RPCS and RPCS-VMS with P1/P1P_1/P_1 elements against an unstabilized monolithic P2/P1P_2/P_1 reference at Δt=10−1\Delta t=10^{-1}, 10−210^{-2}, 10−310^{-3}. Decreasing Δt\Delta t produces mesh-scale pressure oscillations in RPCS, consistent with Codina’s analysis [32], while the RPCS-VMS pressure stays close to the monolithic reference. The reduced-model estimate in § 5.3 identifies the additional pressure-gradient control supplied by the τm\tau_m terms, consistent with the smoother pressure fields observed here.

Figure 4. Pressure fields at T=1T = 1 in the small-time-step manufactured-solution test. All panels use the same color scale.

Effects of Stabilization in the Two Pressure Equations

To examine the contribution of each pressure stabilization term, we compare the full RPCS-VMS formulation with two variants: RPCS-VMS (No recovery stabilization) and RPCS-VMS (No increment stabilization). In each variant, the entire residual contribution is removed from the corresponding equation, including its terms on both sides of Eq. (45) or Eq. (43). All three formulations retain the same intermediate-velocity equation and its stabilization. We repeat the BDF2 manufactured-solution test of § 6.1.3.

Figure 5 shows that removing pressure-recovery stabilization produces visible mesh-scale oscillations, whereas removing pressure-increment stabilization gives a pressure field visually similar to that of the full formulation. Table 3 reports the same L2L^2 error measures as Table 2 and includes RPCS (P1/P1P_1/P_1) and Monolithic (P2/P1P_2/P_1) from § 6.1.3. Compared with the full RPCS-VMS scheme, removing pressure-recovery stabilization increases the global pressure error by a factor of 7.60, with increases in both the boundary strip and the interior. Removing pressure-increment stabilization changes the global pressure error by less than 0.1%. Pressure-recovery stabilization therefore provides the main suppression of pressure oscillations in this test. Retaining both stabilization terms in the final algorithm is motivated by the coupled dissipation structure established for the coordinated formulation in § 5.

Figure 5. Pressure php_h at T=1T = 1 for the manufactured-solution test with BDF2, P1/P1P_1/P_1 elements, and Δt=10−3\Delta t = 10^{-3}. All three formulations retain rotational pressure recovery. All panels use the same color scale.

Method∥p−ph∥Ω\lVert p-p_h\rVert_{\Omega}∥p−ph∥Ωhb\lVert p-p_h\rVert_{\Omega_h^b}∥p−ph∥Ωhi\lVert p-p_h\rVert_{\Omega_h^i}∥u−uh∥Ω\lVert u-u_h\rVert_{\Omega}
RPCS (P1/P1P_1/P_1)4.1429×10−24.1429 \times10^{-2}1.4873×10−21.4873 \times10^{-2}3.8668×10−23.8668 \times10^{-2}2.8351×10−42.8351 \times10^{-4}
RPCS-VMS (No recovery stabilization)2.5573×10−22.5573 \times10^{-2}1.0828×10−21.0828 \times10^{-2}2.3167×10−22.3167 \times10^{-2}3.0302×10−43.0302 \times10^{-4}
RPCS-VMS (No increment stabilization)3.3626×10−33.3626 \times10^{-3}2.6642×10−32.6642 \times10^{-3}2.0516×10−32.0516 \times10^{-3}2.8223×10−42.8223 \times10^{-4}
RPCS-VMS (P1/P1P_1/P_1)3.3659×10−33.3659 \times10^{-3}2.6720×10−32.6720 \times10^{-3}2.0470×10−32.0470 \times10^{-3}3.0222×10−43.0222 \times10^{-4}
Monolithic (P2/P1P_2/P_1)3.0916×10−53.0916 \times10^{-5}1.3382×10−51.3382 \times10^{-5}2.7870×10−52.7870 \times10^{-5}8.9860×10−78.9860 \times10^{-7}

Table 3. Unsteady manufactured-solution L2L^2 errors at T=1T=1 with BDF2, Δt=10−3\Delta t=10^{-3}, and n=64n=64. The error definitions and boundary strip are the same as in Table 2. All RPCS formulations use P1/P1P_1/P_1 elements. Monolithic uses unstabilized P2/P1P_2/P_1 elements.

A possible explanation follows from the structure of the rotational update. Without pressure-recovery stabilization, Eq. (45) computes the L2L^2 projection of phn+ϕh−ν∇⋅u~hp_h^n+\phi_h-\nu\nabla\cdot\tilde{u}_h onto the pressure space. For the continuous P1P_1 velocity approximation used here, ∇⋅u~h\nabla\cdot\tilde{u}_h is elementwise constant and can jump across element interfaces. Spatial variations in this divergence correction can therefore enter the pressure even when the pressure-increment equation is stabilized. The recovery stabilization adds the positive gradient term ν(τm∇ph,∇rh)h\nu(\tau_m\nabla p_h,\nabla r_h)_h to the pressure operator, together with the associated momentum-residual forcing. This gradient term acts directly on the pressure and provides a mechanism for suppressing mesh-scale variations, consistent with the smoother fields observed when recovery stabilization is retained.

Branched Pipe Benchmark

We next simulate the branched-pipe benchmark of Liu [12], also studied by Poux et al. [13]. This case examines separated internal flow with strong turning, a thin internal obstacle, and flux partition between two open outlets. The 2D domain is

Ω=[0,8]×[−0.5,0.5]∖({[0,0.5]×[−0.5,0]}∪{[1.5,8]×[−0.1,0.2]}).(59)\Omega=[0,8]\times[-0.5,0.5]\setminus\left(\left\{[0,0.5]\times[-0.5,0]\right\}\cup\left\{[1.5,8]\times[-0.1,0.2]\right\}\right). \tag*{(59)}

The inlet is the segment x=0x=0, 0≤y≤0.50\le y\le0.5. The open outlets are the right-end segments x=8x=8, 0.2≤y≤0.50.2\le y\le0.5 and x=8x=8, −0.5≤y≤−0.1-0.5\le y\le-0.1. All other boundaries are no-slip. The simulation uses ν=1/600\nu=1/600, BDF2 with the time step in Table 1, and zero pseudo-traction at the outlets. Following the transient start-up of Liu [12], the inlet velocity is uin(y,t)=s(t)(12y(1−2y),0)u_{\mathrm{in}}(y,t)=s(t)(12y(1-2y),0), with s(t)=(1−cos⁡(πt))/2s(t)=(1-\cos(\pi t))/2 for 0≤t≤10\le t\le1 and s(t)=1s(t)=1 for t≥1t\ge1.

The mesh in Figure 6 uses local refinement near the inlet, outlets, entrance corner, and splitter tip. The streamlines show jet deflection, lower-cavity recirculation, and downstream recovery. The start-up histories in Figure 7 show how the outlet fluxes develop over 0≤t≤100\le t\le10. The analytic inflow is Qin(t)=(1−cos⁡(πt))/4Q_{\mathrm{in}}(t)=(1-\cos(\pi t))/4 during start-up and Qin=0.5Q_{\mathrm{in}}=0.5 afterwards. The summed outlet flux follows this prescribed inflow closely, indicating good global mass conservation. The individual fluxes show how the flow is partitioned between the two branches.

Figure 6. Branched-pipe mesh and flow structure. Top: computational mesh with local refinement near the inlet, outlets, entrance corner, and splitter tip. Bottom: streamlines showing jet deflection, lower-cavity recirculation, and downstream recovery.

Figure 7. Branched-pipe flux histories during start-up. The upper and lower outlet fluxes show the flow partition, while their sum is compared with the prescribed inflow to assess global mass conservation.

In addition to the start-up histories, Table 4 compares three integral diagnostics at T=50T=50 against a Monolithic P2/P1P_2/P_1 reference solution computed on the same resolution: the kinetic energy ece_c, the outlet flux QRQ_R, and the inlet-to-outlet mechanical-energy drop ΔhR\Delta h_R, referred to as the charge drop [13]. These quantities are defined as

QuantityRPCS-VMSMonolithic P2/P1P_2/P_1
Total kinetic energy1.99411.9899
Top outflux0.19340.1935
Bottom outflux0.30630.3065
Top charge drop0.63340.6238
Bottom charge drop0.56740.5576

Table 4. Branched-pipe integral diagnostics at T=50T = 50: total kinetic energy, upper and lower outlet fluxes, and inlet-to-outlet charge drops. RPCS-VMS is compared with the Monolithic P2/P1P_2/P_1 reference at the same mesh resolution.

ec=∫Ωρ02∣u∣2 dx,QR=∫Ru⋅n dl,(60)e_c=\int_{\Omega}\frac{\rho_0}{2}\lvert u\rvert^2\,dx,\qquad Q_R=\int_R u\cdot n\,dl, \tag*{(60)}
ΔhR=12(∫L(u⋅n)2 dl−∫R(u⋅n)2 dl)+(∫Lp dl−∫Rp dl).(61)\Delta h_R=\frac{1}{2}\left(\int_L(u\cdot n)^2\,dl-\int_R(u\cdot n)^2\,dl\right)+\left(\int_Lp\,dl-\int_Rp\,dl\right). \tag*{(61)}

Here, ρ0=1\rho_0=1 is the nondimensional density, and LL and RR denote the inlet and outlet boundaries, respectively. RPCS-VMS agrees closely with the Monolithic P2/P1P_2/P_1 reference in kinetic energy and branch fluxes. The charge drops differ only by about 1.5% and 1.8% at the top and bottom outlets.

Flow Past a 2D Cylinder

Flow Past a 2D Cylinder

The 2D DFG cylinder benchmark complements the branched-pipe case by testing boundary-layer resolution, vortex shedding, pressure-drop prediction, and force integration [50, 56, 57]. The domain is

Ω=[0,2.2]×[0,0.41]∖{(x−0.2)2+(y−0.2)2≤0.052}.(62)\Omega= [0, 2.2] \times[0, 0.41] \setminus\left\{(x - 0.2)^2 + (y - 0.2)^2 \le0.05^2\right\}. \tag*{(62)}

The cylinder and channel walls are no-slip, x=0x = 0 is the inflow, and x=2.2x = 2.2 is the zero-traction outlet. The prescribed inflow is

ux(y,t)=6sin⁡(πt/8)y(0.41−y)0.412,uy=0.(63)u_x(y,t) = 6\sin(\pi t/8)\frac{y(0.41-y)}{0.41^2}, \qquad u_y = 0. \tag*{(63)}

The simulation uses ν=10−3\nu= 10^{-3} and the BDF2 time step in Table 1. The pressure difference Δp\Delta p is computed between the points (0.15,0.2)(0.15, 0.2) and (0.25,0.2)(0.25, 0.2), while the force coefficients are normalized using the characteristic velocity U=1U = 1 and cylinder diameter D=0.1D = 0.1. Figure 8 shows the quadrilateral mesh and the wake at four stages of the transient. The streamline fields trace the development of alternating wake vortices as the flow develops.

Figure 8. 2D cylinder mesh and transient wake. Top: quadrilateral mesh. Below: streamlines colored by velocity magnitude.

The drag and lift coefficients are evaluated on the cylinder boundary using the pseudo-stress convention consistent with the open-boundary formulation,

Ci=2νU2Aref∫Γc−ei⋅(∇uh−phνI)n ds,i=x,y,(64)C_i = \frac{2\nu}{U^2 A_{\mathrm{ref}}} \int_{\Gamma_c} -\mathbf{e}_i \cdot\left(\nabla\mathbf{u}_h - \frac{p_h}{\nu} \mathbf{I}\right)\mathbf{n}\,\mathrm{d}s,\qquad i=x,y, \tag*{(64)}

Here Γc\Gamma_c is the cylinder boundary, ei\mathbf{e}_i the Cartesian unit vector, and I\mathbf{I} the identity tensor. The reference area is Aref=DA_{\mathrm{ref}}=D in 2D, and CD=CxC_D=C_x and CL=CyC_L=C_y denote drag and lift, respectively. Fig. 9 compares the drag, lift, and pressure-difference histories with the FEATFLOW reference data [50, 57]. RPCS-VMS captures the rise and fall of drag and pressure difference, together with the growth and subsequent decay of the lift oscillations. The close agreement in the lift history indicates that both the amplitude and timing of the vortex-shedding response are reproduced. Table 5 gives the quantitative comparison of some benchmark indices. Here t(CD,max⁡)t(C_{D,\max}) and t(CL,max⁡)t(C_{L,\max}) denote the times at which the drag and lift coefficients attain their respective maximum values over the simulated interval. All listed relative errors are within 0.18%, indicating excellent agreement.

MetricFEATFLOWRPCS-VMSRel. err. (%)
t(CD,max⁡)t(C_{D,\max})3.93623.93620
CD,max⁡C_{D,\max}2.94382.94340.01
t(CL,max⁡)t(C_{L,\max})5.69255.69500.04
CL,max⁡C_{L,\max}0.47750.47720.06
Δp(t=8)\Delta p(t=8)-0.1115-0.11170.18

Table 5. 2D cylinder benchmark indices for RPCS-VMS and the FEATFLOW reference [50], including peak force coefficients, their occurrence times, and the pressure difference at t=8t=8. Relative errors are measured against the reference values.

Figure 9. 2D cylinder histories computed with RPCS-VMS and the FEATFLOW reference data [50, 57].

Flow Past a 3D Cylinder

As shown in Figure 10, the 3D DFG cylinder benchmark extends the force comparison to a 3D wake with a smaller lift signal [51, 56, 58]. The domain is

Figure 10. 3D cylinder wake visualized by QQ-criterion isosurfaces at the common threshold Q=2.618Q = 2.618, colored by velocity magnitude.

Ω=[0,2.5]×[0,0.41]×[0,0.41]∖{(x−0.5)2+(y−0.2)2≤0.052, 0≤z≤0.41}.(65)\Omega= [0,2.5]\times[0,0.41]\times[0,0.41]\setminus\left\{(x-0.5)^2+(y-0.2)^2\le0.05^2,\ 0\le z\le0.41\right\}. \tag*{(65)}

The cylinder axis is parallel to zz. The plane x=0x=0 is the inflow, x=2.5x=2.5 is the zero-traction outlet, and the walls and cylinder surface are no-slip. The simulation uses ν=10−3\nu=10^{-3}, the BDF2 time step in Table 1, and

ux(y,z,t)=36sin⁡(πt/8)yz(0.41−y)(0.41−z)0.414,uy=uz=0.(66)u_x(y,z,t)=36\sin(\pi t/8)\frac{yz(0.41-y)(0.41-z)}{0.41^4},\qquad u_y=u_z=0. \tag*{(66)}

The drag and lift coefficients are evaluated using Eq. (64) with U=1U=1, D=0.1D=0.1, and Aref=0.041A_{\mathrm{ref}}=0.041. Figure 11 compares the force histories with the FEATFLOW reference data [51, 58]. RPCS-VMS reproduces the transient growth and decay of drag and the much smaller lift response. Table 6 quantifies the agreement at the force extrema. The maximum drag differs by 0.62%, while the maximum and minimum lift differ by 0.28% and 1.82%, respectively. Together, the histories and extrema show that both streamwise and transverse forces are accurately captured at this resolution.

MetricFEATFLOWRPCS-VMSRel. err. (%)
CD,max⁡C_{D,\max}3.29783.27750.62
CL,max⁡C_{L,\max}2.7755×10−32.7755 \times10^{-3}2.7833×10−32.7833 \times10^{-3}0.28
CL,min⁡C_{L,\min}-0.0110-0.01121.82

Table 6. 3D cylinder force extrema for RPCS-VMS and the FEATFLOW reference [51]. Relative errors are measured against the reference values.

Figure 11. 3D cylinder force histories computed with RPCS-VMS and the FEATFLOW reference data [51, 58].

We further assess computational cost by comparing RPCS-VMS with Monolithic-VMS over the first 100 time steps (T=0.5T = 0.5) using the 3D cylinder case. Both use the same mesh, ν=10−3\nu= 10^{-3}, Δt=0.005\Delta t = 0.005, CI=36C_I = 36, and τc=ν/6\tau_c = \nu/6. They share the time integration method and elementwise stabilization model. Convection is linearized using an extrapolated advecting velocity in both formulations. Each time step requires four sequential linear solves for RPCS-VMS and one coupled linear solve for Monolithic-VMS.

Each method is run three times using eight MPI ranks (one thread per rank) on an Intel Xeon Gold 6226R processor under Ubuntu 20.04. To facilitate a fair performance comparison with RPCS-VMS, the Monolithic-VMS system is solved using GMRES with a block Schur-complement preconditioner based on a lower-triangular factorization of the coupled velocity–pressure system. This preconditioner enables a sequential treatment of the velocity and pressure blocks, similar to the solution strategy of RPCS-VMS. The velocity and Schur-complement block solves are each approximated by BoomerAMG. Both formulations use the same relative and absolute solver convergence tolerances.

The total wall time includes assembly, preconditioning, and linear solves, but excludes initialization and field output. Averaged over three runs, the wall times are 5550.0 s for Monolithic-VMS and 2301.7 s for RPCS-VMS, corresponding to a 2.41-fold speedup for the reported configuration. These results highlight the computational advantage of the RPCS-VMS formulation, which replaces the coupled velocity-pressure solve with a sequence of smaller, decoupled systems, substantially reducing the overall computational cost in this case.

Conclusion

In this study, we propose a variational multiscale rotational pressure-correction method for incompressible flows with equal-order finite elements. The method incorporates residual stabilization in both pressure equations while preserving the four-solve sequential structure. For the reduced BDF1 Oseen model analyzed here, we establish augmented-energy decay and a conditional pressure-gradient bound under stated coefficient conditions.

Numerical experiments demonstrate good spatial and temporal convergence rates, mitigation of small-time-step pressure oscillations, and close agreement with reference solutions for separated flows. A comparison of the two pressure-stage stabilization terms indicates that recovery stabilization plays the principal role in suppressing oscillations in the tested case. In the 3D cylinder benchmark, RPCS-VMS achieves a 2.41-fold wall-time speedup relative to the tested Monolithic-VMS solver. These results support the use of RPCS-VMS as an efficient equal-order formulation for the incompressible flow problems considered here.

Code availability

The source code accompanying this study is available through GitHub: https://github.com/CZHU20/RPCS-VMS.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This work was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1000201), the National Natural Science Foundation of China (Grant Nos. 12293001, 12272009), and the Beijing Natural Science Foundation (Grant No. L259063).

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

The authors used Codex (OpenAI) to assist with language editing, manuscript organization, and checks of clarity and internal consistency. The authors evaluated the suggestions, revised the text, and verified the scientific content. The authors remain fully responsible for the final manuscript.

Proof of the Stability Estimates

We derive the energy identity and pressure-gradient bound stated in § 5.

Energy identity and decay

For the piecewise linear fields used in this analysis, all elementwise second derivatives vanish. Under the assumptions of § 5.1, the intermediate velocity, pressure increment, velocity correction, and pressure recovery equations are

(a~hn+1,vh)+(C(whn+1)u~hn+1,vh)+ν(∇u~hn+1,∇vh)+(τc∇⋅u~hn+1,∇⋅vh)h−(phn,∇⋅vh)−12(β∗u~hn+1,vh)ΓN=0,(67)(\widetilde{a}_h^{n+1},v_h)+(C(w_h^{n+1})\widetilde{u}_h^{n+1},v_h)+\nu(\nabla\widetilde{u}_h^{n+1},\nabla v_h)+(\tau_c\nabla\cdot\widetilde{u}_h^{n+1},\nabla\cdot v_h)_h-(p_h^n,\nabla\cdot v_h)-\frac{1}{2}(\beta^*\widetilde{u}_h^{n+1},v_h)_{\Gamma_N}=0, \tag*{(67)}
(∇⋅u~hn+1,qh)+Δt(∇ϕhn+1,∇qh)+(τmRϕn+1,∇qh)h=0,(68)(\nabla\cdot\widetilde{u}_h^{n+1},q_h)+\Delta t(\nabla\phi_h^{n+1},\nabla q_h)+(\tau_m R_\phi^{n+1},\nabla q_h)_h=0, \tag*{(68)}
uhn+1=u~hn+1−ΔtPh∇ϕhn+1,(69)u_h^{n+1}=\widetilde{u}_h^{n+1}-\Delta tP_h\nabla\phi_h^{n+1}, \tag*{(69)}
(phn+1−πhn+1+ν∇⋅u~hn+1,rh)+ν(τmRpn+1,∇rh)h=0,(70)(p_h^{n+1}-\pi_h^{n+1}+\nu\nabla\cdot\widetilde{u}_h^{n+1},r_h)+\nu(\tau_m R_p^{n+1},\nabla r_h)_h=0, \tag*{(70)}

for vh∈Xhv_h\in X_h, qh∈Mh,0q_h\in M_{h,0}, and rh∈Mhr_h\in M_h, with the momentum residuals given in § 5.1.

To combine the pressure-increment and pressure-recovery contributions, use the relations

πhn+1=ψhn+1−ηhn,phn+1=πhn+1−δηhn+1,Rpn+1=Rϕn+1−∇δηhn+1.(71)\pi_h^{n+1}=\psi_h^{n+1}-\eta_h^n,\qquad p_h^{n+1}=\pi_h^{n+1}-\delta\eta_h^{n+1},\qquad R_p^{n+1}=R_\phi^{n+1}-\nabla\delta\eta_h^{n+1}. \tag*{(71)}

Hence the pressure-recovery equation can be written as

(δηhn+1,rh)=ν(∇⋅u~hn+1,rh)+ν(τmRpn+1,∇rh)h,rh∈Mh.(72)(\delta\eta_h^{n+1},r_h)=\nu(\nabla\cdot\widetilde{u}_h^{n+1},r_h)+\nu(\tau_mR_p^{n+1},\nabla r_h)_h,\qquad r_h\in M_h. \tag*{(72)}

The energy balance for the intermediate velocity follows by testing Eq. (A.1) with 2Δtu~hn+12\Delta t\widetilde{u}_h^{n+1}, using 2(x−y,x)=∥x∥2−∥y∥2+∥x−y∥22(x-y,x)=\lVert x\rVert^2-\lVert y\rVert^2+\lVert x-y\rVert^2, and applying Eq. (46):

∥u~hn+1∥2−∥uhn∥2+Tn+2νΔt∥∇u~hn+1∥2+2Δt∥∇⋅u~hn+1∥τc2+ΔtBn−2Δt(phn,∇⋅u~hn+1)=0.(73)\lVert\widetilde{u}_h^{n+1}\rVert^2-\lVert u_h^n\rVert^2+T_n+2\nu\Delta t\lVert\nabla\widetilde{u}_h^{n+1}\rVert^2+2\Delta t\lVert\nabla\cdot\widetilde{u}_h^{n+1}\rVert_{\tau_c}^2+\Delta tB_n-2\Delta t(p_h^n,\nabla\cdot\widetilde{u}_h^{n+1})=0. \tag*{(73)}

Here Tn=∥u~hn+1−uhn∥2T_n=\lVert\widetilde{u}_h^{n+1}-u_h^n\rVert^2. To replace intermediate by corrected kinetic energy, use orthogonality of PhP_h:

(u~hn+1,Ph∇ϕhn+1)=(u~hn+1,∇ϕhn+1)(\widetilde{u}_h^{n+1},P_h\nabla\phi_h^{n+1})=(\widetilde{u}_h^{n+1},\nabla\phi_h^{n+1}).

Integration by parts produces no boundary term: u~hn+1=0\widetilde{u}_h^{n+1}=0 on ΓD\Gamma_D and ϕhn+1=0\phi_h^{n+1}=0 on ΓN\Gamma_N. Therefore

∥uhn+1∥2=∥u~hn+1∥2+2Δt(∇⋅u~hn+1,ϕhn+1)+(Δt)2∥Ph∇ϕhn+1∥2.(74)\lVert u_h^{n+1}\rVert^2=\lVert\widetilde{u}_h^{n+1}\rVert^2+2\Delta t(\nabla\cdot\widetilde{u}_h^{n+1},\phi_h^{n+1})+(\Delta t)^2\lVert P_h\nabla\phi_h^{n+1}\rVert^2. \tag*{(74)}

Substitution into Eq. (A.7) combines the pressure contributions into −2Δt(πhn+1,∇⋅u~hn+1)-2\Delta t(\pi_h^{n+1},\nabla\cdot\widetilde{u}_h^{n+1}).

To turn this pressure work into energy differences, test Eq. (A.2) with ψhn+1\psi_h^{n+1} and Eq. (A.6) with ηhn\eta_h^n. Using the pressure and residual relations in Eq. (A.5) and expanding the squared norms gives

−2Δt(πhn+1,∇⋅u~hn+1)=(Δt)2(∥∇ψhn+1∥2−∥∇ψhn∥2+∥∇ϕhn+1∥2)+Δtν(∥ηhn+1∥2−∥ηhn∥2−∥δηhn+1∥2)+Δt(∥∇ηhn+1∥τm2−∥∇ηhn∥τm2−∥∇δηhn+1∥τm2)+2Δt(τmRϕn+1,∇πhn+1)h.(75)\begin{aligned} -2\Delta t(\pi_h^{n+1},\nabla\cdot\widetilde{u}_h^{n+1}) ={}&(\Delta t)^2\left(\lVert\nabla\psi_h^{n+1}\rVert^2-\lVert\nabla\psi_h^n\rVert^2+\lVert\nabla\phi_h^{n+1}\rVert^2\right)\\ &+\frac{\Delta t}{\nu}\left(\lVert\eta_h^{n+1}\rVert^2-\lVert\eta_h^n\rVert^2-\lVert\delta\eta_h^{n+1}\rVert^2\right)\\ &+\Delta t\left(\lVert\nabla\eta_h^{n+1}\rVert_{\tau_m}^2-\lVert\nabla\eta_h^n\rVert_{\tau_m}^2-\lVert\nabla\delta\eta_h^{n+1}\rVert_{\tau_m}^2\right)\\ &+2\Delta t(\tau_mR_\phi^{n+1},\nabla\pi_h^{n+1})_h. \tag*{(75)} \end{aligned}

The weighted norm identity uses time-independence of τm\tau_m. Orthogonality also gives ∥∇ϕhn+1∥2−∥Ph∇ϕhn+1∥2=∥Qh∇ϕhn+1∥2\|\nabla\phi_h^{n+1}\|^2-\|P_h\nabla\phi_h^{n+1}\|^2=\|Q_h\nabla\phi_h^{n+1}\|^2. Combining Eqs. (A.7), (A.8), and (A.9) therefore gives

En+1−En+Tn+Jn+ΔtDh(u~hn+1)+νΔt∥∇⋅u~hn+1∥2+ΔtBn+2Δt(τmRϕn+1,∇πhn+1)h−Δtν∥δηhn+1∥2−Δt∥∇δηhn+1∥τm2=0.(76)\begin{aligned} E^{n+1}-E^n+T_n+J_n+\Delta t\mathcal{D}_h(\widetilde{u}_h^{n+1})+\nu\Delta t\|\nabla\cdot\widetilde{u}_h^{n+1}\|^2+\Delta t B_n \\ +2\Delta t(\tau_m R_\phi^{n+1},\nabla\pi_h^{n+1})_h-\frac{\Delta t}{\nu}\|\delta\eta_h^{n+1}\|^2-\Delta t\|\nabla\delta\eta_h^{n+1}\|_{\tau_m}^2=0. \tag*{(76)} \end{aligned}

It remains to combine the last three terms with the divergence contribution. Complete squares to expose their dissipation, retaining ΔtBn\Delta t B_n. Choose rh=δηhn+1r_h=\delta\eta_h^{n+1} in Eq. (A.6). Using Eq. (A.5) gives

1ν∥δηhn+1∥2+∥∇δηhn+1∥τm2=(∇⋅u~hn+1,δηhn+1)+(τmRϕn+1,∇δηhn+1)h.(77)\frac{1}{\nu}\|\delta\eta_h^{n+1}\|^2+\|\nabla\delta\eta_h^{n+1}\|_{\tau_m}^2=(\nabla\cdot\widetilde{u}_h^{n+1},\delta\eta_h^{n+1})+(\tau_m R_\phi^{n+1},\nabla\delta\eta_h^{n+1})_h. \tag*{(77)}

Expanding the two squares on the right below and applying Eq. (A.11) yields

ν∥∇⋅u~hn+1∥2+∥Rϕn+1∥τm2−1ν∥δηhn+1∥2−∥∇δηhn+1∥τm2=1ν∥δηhn+1−ν∇⋅u~hn+1∥2+∥Rpn+1∥τm2.(78)\begin{aligned} \nu\|\nabla\cdot\widetilde{u}_h^{n+1}\|^2+\|R_\phi^{n+1}\|_{\tau_m}^2-\frac{1}{\nu}\|\delta\eta_h^{n+1}\|^2-\|\nabla\delta\eta_h^{n+1}\|_{\tau_m}^2 \\ =\frac{1}{\nu}\|\delta\eta_h^{n+1}-\nu\nabla\cdot\widetilde{u}_h^{n+1}\|^2+\|R_p^{n+1}\|_{\tau_m}^2. \tag*{(78)} \end{aligned}

Finally, Rϕn+1=a~hn+1+∇πhn+1R_\phi^{n+1}=\widetilde{a}_h^{n+1}+\nabla\pi_h^{n+1} implies

2(τmRϕn+1,∇πhn+1)h−∥Rϕn+1∥τm2=∥∇πhn+1∥τm2−∥a~hn+1∥τm2.(79)2(\tau_m R_\phi^{n+1},\nabla\pi_h^{n+1})_h-\|R_\phi^{n+1}\|_{\tau_m}^2=\|\nabla\pi_h^{n+1}\|_{\tau_m}^2-\|\widetilde{a}_h^{n+1}\|_{\tau_m}^2. \tag*{(79)}

Substitution of Eqs. (A.12) and (A.13) into Eq. (A.10) leaves precisely En+1−En+(Tn−Δt∥a~hn+1∥τm2)+Jn+ΔtFn+Sn+ΔtBn=0E^{n+1}-E^n+(T_n-\Delta t\|\widetilde{a}_h^{n+1}\|_{\tau_m}^2)+J_n+\Delta t F_n+S_n+\Delta t B_n=0. The term in parentheses is WnW_n, proving Eq. (52).

To establish energy decay, we now show that each dissipation term is nonnegative. Since ∥∇⋅vh∥2≤d∥∇vh∥2\|\nabla\cdot v_h\|^2\le d\|\nabla v_h\|^2,

Dh(vh)=2ν∥∇vh∥2+2∥∇⋅vh∥τc2−ν∥∇⋅vh∥2≥[2τcmin⁡−(1−2d)ν]∥∇⋅vh∥2≥0.(80)\begin{aligned} \mathcal{D}_h(v_h)=2\nu\|\nabla v_h\|^2+2\|\nabla\cdot v_h\|_{\tau_c}^2-\nu\|\nabla\cdot v_h\|^2 \\ &\ge\left[2\tau_c^{\min}-\left(1-\frac{2}{d}\right)\nu\right]\|\nabla\cdot v_h\|^2\ge0. \tag*{(80)} \end{aligned}

The coefficient bound τm,K≤Δt\tau_{m,K}\le\Delta t gives Wn≥0W_n\ge0. Together with the other nonnegative terms, Eq. (52) gives En+1≤EnE^{n+1}\le E^n. Summing it proves Eq. (53).

For the pressure estimate, positive definiteness at κc=0\kappa_c=0 is also needed. Zero dissipation forces equality in the underlying pointwise Cauchy-Schwarz inequality, so ∇vh∣K=λKI\nabla v_h|_K=\lambda_K I, where λK\lambda_K is constant on each linear element. Thus vh∣K=λKx+bKv_h|_K=\lambda_Kx+b_K. Continuity across shared faces makes λK,bK\lambda_K,b_K common to all elements. The zero values on ΓD\Gamma_D then force vh=0v_h=0.

Guermond, Minev, and Shen [11] use a tunable coefficient χ\chi in the rotational update pn+1=pn+ϕn+1−χ∇⋅u~n+1p^{n+1}=p^n+\phi^{n+1}-\chi\nabla\cdot\widetilde{u}^{n+1}. For Laplacian viscosity with open boundaries, their sufficient coercivity condition is 0<χ<2ν/d0<\chi<2\nu/d. Here the rotational coefficient remains ν\nu. The continuity-residual stabilization in Eq. (40) supplies the additional divergence control, with the threshold in Eq. (47). Its pressure-subscale construction is described in § 4.4.

Bound for the pressure gradient

To transfer pressure control through pressure recovery, we need divergence control and an inverse estimate:

∥∇⋅vh∥2≤Cd,hDh(vh),∥∇rh∥2≤C∇,h2∥rh∥2,vh∈Xh,rh∈Mh.(81)\|\nabla\cdot v_h\|^2\le C_{d,h}\mathcal{D}_h(v_h),\qquad\|\nabla r_h\|^2\le C_{\nabla,h}^2\|r_h\|^2,\qquad v_h\in X_h,\quad r_h\in M_h. \tag*{(81)}

Both constants are finite on the chosen mesh because Dh\mathcal{D}_h is positive definite and the spaces are finite dimensional. To collect the coefficients in the pressure-gradient bound below, choose

Cp,h=max⁡{2,  4ντmmax⁡C∇,h2,  4ν2τmmax⁡C∇,h2Cd,h}.(82)C_{p,h}=\max\{2,\;4\nu\tau_m^{\max}C_{\nabla,h}^2,\;4\nu^2\tau_m^{\max}C_{\nabla,h}^2C_{d,h}\}. \tag*{(82)}

The pressure-recovery dissipation already controls zhn+1=δηhn+1−ν∇⋅u~hn+1z_h^{n+1}=\delta\eta_h^{n+1}-\nu\nabla\cdot\widetilde{u}_h^{n+1}. Applying the inverse estimate to δηhn+1∈Mh\delta\eta_h^{n+1}\in M_h gives

∥∇δηhn+1∥τm2≤τmmax⁡C∇,h2∥δηhn+1∥2≤2τmmax⁡C∇,h2(ν2∥∇⋅u~hn+1∥2+∥zhn+1∥2)≤2τmmax⁡C∇,h2(ν2Cd,hDh(u~hn+1)+∥zhn+1∥2).(83)\begin{aligned} \|\nabla\delta\eta_h^{n+1}\|_{\tau_m}^2&\le\tau_m^{\max}C_{\nabla,h}^2\|\delta\eta_h^{n+1}\|^2\\ &\le2\tau_m^{\max}C_{\nabla,h}^2\left(\nu^2\|\nabla\cdot\widetilde{u}_h^{n+1}\|^2+\|z_h^{n+1}\|^2\right)\\ &\le2\tau_m^{\max}C_{\nabla,h}^2\left(\nu^2C_{d,h}\mathcal{D}_h(\widetilde{u}_h^{n+1})+\|z_h^{n+1}\|^2\right). \tag*{(83)} \end{aligned}

Since phn+1=πhn+1−δηhn+1p_h^{n+1}=\pi_h^{n+1}-\delta\eta_h^{n+1}, the squared triangle inequality and Eq. (A.17) give

∥∇phn+1∥τm2≤Cp,h[∥∇πhn+1∥τm2+Dh(u~hn+1)+ν−1∥zhn+1∥2].(84)\left\|\nabla p_h^{n+1}\right\|_{\tau_m}^2 \le C_{p,h}\left[\left\|\nabla\pi_h^{n+1}\right\|_{\tau_m}^2+\mathcal{D}_h(\widetilde{u}_h^{n+1})+\nu^{-1}\left\|z_h^{n+1}\right\|^2\right]. \tag*{(84)}

Multiply by Δt\Delta t, sum over the steps, and apply Eq. (53). The bracketed terms are part of its nonnegative dissipation, so their Δt\Delta t-weighted sum is at most E00E_0^0. Finally, τmmin⁡∥∇phn+1∥2≤∥∇phn+1∥τm2\tau_m^{\min}\left\|\nabla p_h^{n+1}\right\|^2\le\left\|\nabla p_h^{n+1}\right\|_{\tau_m}^2 proves Eq. (54).

Comparison without momentum-residual stabilization

For this comparison, set τm=0\tau_m=0 in the four weak equations, with τc\tau_c prescribed independently at or above the divergence threshold. The energy identity remains valid with

E0n=∥uhn∥2+(Δt)2∥∇ψhn∥2+Δtν∥ηhn∥2,E0n+1≤E0n.(85)E_0^n=\left\|u_h^n\right\|^2+(\Delta t)^2\left\|\nabla\psi_h^n\right\|^2+\frac{\Delta t}{\nu}\left\|\eta_h^n\right\|^2,\qquad E_0^{n+1}\le E_0^n. \tag*{(85)}

However, its pressure-gradient dissipation is absent. Using phn+1=ψhn+1−ηhn+1p_h^{n+1}=\psi_h^{n+1}-\eta_h^{n+1} and the inverse estimate in Eq. (A.15) gives

(Δt)2∥∇phn+1∥2≤2(Δt)2∥∇ψhn+1∥2+2(Δt)2C∇,h2∥ηhn+1∥2(\Delta t)^2\left\|\nabla p_h^{n+1}\right\|^2\le2(\Delta t)^2\left\|\nabla\psi_h^{n+1}\right\|^2+2(\Delta t)^2C_{\nabla,h}^2\left\|\eta_h^{n+1}\right\|^2
≤2max⁡{1,νΔtC∇,h2}E0n+1≤2max⁡{1,νΔtC∇,h2}E00.(86)\le2\max\{1,\nu\Delta t C_{\nabla,h}^2\}E_0^{n+1}\le2\max\{1,\nu\Delta t C_{\nabla,h}^2\}E_0^0. \tag*{(86)}

Dividing by (Δt)2(\Delta t)^2, multiplying by Δt\Delta t, and summing prove Eq. (55).

To see why energy nonincrease does not ensure decay of every pressure mode, suppose there exists a nonconstant qh∈Mhq_h\in M_h, with zero mean in the closed case, such that

(qh,∇⋅vh)=0for every vh∈Xh.(q_h,\nabla\cdot v_h)=0\qquad\text{for every }v_h\in X_h.

This mode produces no pressure contribution to the intermediate-velocity equation. With τm=0\tau_m=0, the fields

u~hn+1=uhn=0,phn=qh,ϕhn+1=0\widetilde{u}_h^{n+1}=u_h^n=0,\qquad p_h^n=q_h,\qquad\phi_h^{n+1}=0

satisfy Eqs. (A.1)–(A.4) at every step. The auxiliary sequences are ψhn=0\psi_h^n=0 and ηhn=−qh\eta_h^n=-q_h, so E0n=(Δt/ν)∥qh∥2E_0^n=(\Delta t/\nu)\left\|q_h\right\|^2 is constant and satisfies E0n+1≤E0nE_0^{n+1}\le E_0^n. Whenever such a mode exists, its pressure pattern therefore persists unchanged despite the nonincrease of energy.

References

  1. [1]Alexandre Joel Chorin. Numerical solution of the Navier–Stokes equations. Mathematics of Computation, 22:745–762, 1968. doi: 10.1090/S0025-5718-1968-0242392-2.DOI
  2. [2]Katsuhiko Goda. A multistep technique with implicit difference schemes for calculating two- or three-dimensional cavity flows. Journal of Computational Physics, 30(1):76–95, 1979. doi: 10.1016/0021-9991(79)90088-3.DOI
  3. [3]J. Kim and P. Moin. Application of a fractional-step method to incompressible Navier–Stokes equations. Journal of Computational Physics, 59(2):308–323, 1985. doi: 10.1016/0021-9991(85)90148-2.DOI
  4. [4]Jos Van Kan. A second-order accurate pressure-correction scheme for viscous incompressible flow. SIAM Journal on Scientific and Statistical Computing, 7(3):870–891, 1986. doi: 10.1137/0907059.DOI
  5. [5]George Em Karniadakis, Moshe Israeli, and Steven A. Orszag. High-order splitting methods for the incompressible Navier–Stokes equations. Journal of Computational Physics, 97(2):414–443, 1991. doi: 10.1016/0021-9991(91)90007-8.DOI
  6. [6]J. L. Guermond, P. Minev, and Jie Shen. An overview of projection methods for incompressible flows. Computer Methods in Applied Mechanics and Engineering, 195(44–47):6011–6045, 2006. doi: 10.1016/j.cma.2005.10.010.DOI
  7. [7]John C. Strikwerda and Young S. Lee. The accuracy of the fractional step method. SIAM Journal on Numerical Analysis, 37(1):37–47, 1999. doi: 10.1137/S0036142997326938.DOI
  8. [8]David L. Brown, Ricardo Cortez, and Michael L. Minion. Accurate projection methods for the incompressible Navier–Stokes equations. Journal of Computational Physics, 168(2):464–499, 2001. doi: 10.1006/jcph.2001.6715.DOI
  9. [9]L. J. P. Timmermans, P. D. Minev, and F. N. Van De Vosse. An approximate projection scheme for incompressible flow using spectral elements. International Journal for Numerical Methods in Fluids, 22(7):673–688, 1996. doi: 10.1002/(SICI)1097-0363(19960415)22:7<673::AID-FLD373>3.0.CO;2-O.doi.org/10.1002/(sici)1097-0363(19960415)22:7<673::aid-fld373>3.0.co;2-o
  10. [10]Jean-Luc Guermond and Jie Shen. On the error estimates for the rotational pressure-correction projection methods. Math. Comput., 73:1719–1737, 2004. doi: 10.1090/S0025-5718-03-01621-1.DOI
  11. [11]J. L. Guermond, Peter Minev, and Jie Shen. Error analysis of pressure-correction schemes for the time-dependent Stokes equations with open boundary conditions. SIAM Journal on Numerical Analysis, 43(1):239–258, 2005. doi: 10.1137/040604418.DOI
  12. [12]Jie Liu. Open and traction boundary conditions for the incompressible Navier–Stokes equations. Journal of Computational Physics, 228(19):7250–7267, 2009. doi: 10.1016/j.jcp.2009.06.021.DOI
  13. [13]A. Poux, S. Glockner, and M. Azaïez. Improvements on open and traction boundary conditions for Navier–Stokes time-splitting methods. Journal of Computational Physics, 230(10):4011–4027, 2011. doi: 10.1016/j.jcp.2011.02.024.DOI
  14. [14]A. Poux, S. Glockner, E. Ahusborde, and M. Azaïez. Open boundary conditions for the velocity-correction scheme of the Navier–Stokes equations. Computers & Fluids, 70:29–43, 2012. doi: 10.1016/j.compfluid.2012.08.028.DOI
  15. [15]Eberhard Bänsch. A finite element pressure correction scheme for the Navier–Stokes equations with traction boundary condition. Computer Methods in Applied Mechanics and Engineering, 279:198–211, 2014. doi: 10.1016/j.cma.2014.06.030.DOI
  16. [16]Andrea Bonito, Jean-Luc Guermond, and Sanghyun Lee. Modified pressure-correction projection methods: Open boundary and variable time stepping. In Numerical Mathematics and Advanced Applications – ENUMATH 2013, pages 623–631. Springer, 2015. doi: 10.1007/978-3-319-10705-9_61.DOI
  17. [17]C.-H. Bruneau and P. Fabrie. Effective downstream boundary conditions for incompressible Navier–Stokes equations. International Journal for Numerical Methods in Fluids, 19(8):693–705, 1994. doi: 10.1002/fld.1650190805.DOI
  18. [18]C.-H. Bruneau and P. Fabrie. New efficient boundary conditions for incompressible Navier–Stokes equations: A well-posedness result. ESAIM: Mathematical Modelling and Numerical Analysis, 30(7):815–840, 1996. doi: 10.1051/m2an/1996300708151.DOI
  19. [19]S. Dong and J. Shen. A pressure correction scheme for generalized form of energy-stable open boundary conditions for incompressible flows. Journal of Computational Physics, 291:254–278, 2015. doi: 10.1016/j.jcp.2015.03.012.DOI
  20. [20]Roland Becker, Daniela Capatina, Robert Luce, and David Trujillo. Finite element formulation of general boundary conditions for incompressible flows. Computer Methods in Applied Mechanics and Engineering, 295:240–267, 2015. doi: 10.1016/j.cma.2015.07.002.DOI
  21. [21]Sanghyun Lee and Abner J. Salgado. Stability analysis of pressure correction schemes for the Navier–Stokes equations with traction boundary conditions. Computer Methods in Applied Mechanics and Engineering, 309:307–324, 2016. doi: 10.1016/j.cma.2016.05.043.DOI
  22. [22]F. Brezzi. On the existence, uniqueness and approximation of saddle-point problems arising from Lagrangian multipliers. R.A.I.R.O. Analyse Numérique, 8(R2):129–151, 1974. doi: 10.1051/m2an/197408R201291.DOI
  23. [23]C. Taylor and P. Hood. A numerical solution of the Navier–Stokes equations using the finite element technique. Computers & Fluids, 1(1):73–100, 1973. doi: 10.1016/0045-7930(73)90027-3.DOI
  24. [24]Thomas J. R. Hughes, Leopoldo P. Franca, and Marc Balestra. A new finite element formulation for computational fluid dynamics: V. Circumventing the Babuška–Brezzi condition: a stable Petrov-Galerkin formulation of the Stokes problem accommodating equal-order interpolations. Computer Methods in Applied Mechanics and Engineering, 59(1):85–99, 1986. doi: 10.1016/0045-7825(86)90025-3.DOI
  25. [25]Leopoldo P. Franca and Sérgio L. Frey. Stabilized finite element methods: II. The incompressible Navier–Stokes equations. Computer Methods in Applied Mechanics and Engineering, 99(2–3):209–233, 1992. doi: 10.1016/0045-7825(92)90041-H.DOI
  26. [26]T. E. Tezduyar, S. Mittal, S. E. Ray, and R. Shih. Incompressible flow computations with stabilized bilinear and linear equal-order interpolaton velocity-pressure elements. Computer Methods in Applied Mechanics and Engineering, 95:221–242, 1992. doi: 10.1016/0045-7825(92)90141-6.DOI
  27. [27]Ramon Codina and Jordi Blasco. A finite element formulation for the Stokes problem allowing equal velocity-pressure interpolation. Computer Methods in Applied Mechanics and Engineering, 143:373–391, 1997. doi: 10.1016/S0045-7825(96)01154-1.DOI
  28. [28]Roland Becker and Malte Braack. A finite element pressure gradient stabilization for the Stokes equations based on local projections. Calcolo, 38:173–199, 2001. doi: 10.1007/s10092-001-8180-4.DOI
  29. [29]Pavel B. Bochev, Clark R. Dohrmann, and Max D. Gunzburger. Stabilization of low-order mixed finite elements for the Stokes equations. SIAM Journal on Numerical Analysis, 44(1):82–101, 2006. doi: 10.1137/S0036142905444482.DOI
  30. [30]Erik Burman and Peter Hansbo. Edge stabilization for the generalized Stokes problem: A continuous interior penalty method. Computer Methods in Applied Mechanics and Engineering, 195(19–22):2393–2410, 2006. doi: 10.1016/j.cma.2005.05.009.DOI
  31. [31]Erik Burman and Miguel A. Fernández. Continuous interior penalty finite element method for the time-dependent Navier–Stokes equations: space discretization and convergence. Numerische Mathematik, 107(1):39–77, 2007. doi: 10.1007/s00211-007-0070-5.DOI
  32. [32]Ramon Codina. Pressure stability in fractional step finite element methods for incompressible flows. Journal of Computational Physics, 170(1):112–140, 2001. doi: 10.1006/jcph.2001.6725.DOI
  33. [33]Santiago Badia and Ramon Codina. Convergence analysis of the FEM approximation of the first order projection method for incompressible flows with and without the inf-sup condition. Numerische Mathematik, 107(4):533–557, 2007. doi: 10.1007/s00211-007-0099-5.DOI
  34. [34]Thomas J. R. Hughes. Multiscale phenomena: Green’s functions, the Dirichlet-to-Neumann formulation, subgrid scale models, bubbles and the origins of stabilized methods. Computer Methods in Applied Mechanics and Engineering, 127(1–4):387–401, 1995. doi: 10.1016/0045-7825(95)00844-9.DOI
  35. [35]Thomas J. R. Hughes, Gonzalo R. Feijóo, Luca Mazzei, and Jean-Baptiste Quincy. The variational multiscale method: a paradigm for computational mechanics. Computer Methods in Applied Mechanics and Engineering, 166(1–2):3–24, 1998. doi: 10.1016/S0045-7825(98)00079-6.DOI
  36. [36]Ramon Codina. Stabilization of incompressibility and convection through orthogonal sub-scales in finite element methods. Computer Methods in Applied Mechanics and Engineering, 190(13–14):1579–1599, 2000. doi: 10.1016/S0045-7825(00)00254-1.DOI
  37. [37]Ramon Codina. Stabilized finite element approximation of transient incompressible flows using orthogonal subscales. Computer Methods in Applied Mechanics and Engineering, 191(39–40):4295–4321, 2002. doi: 10.1016/S0045–7825(02)00337–7.DOI
  38. [38]Ramon Codina, Javier Principe, Oriol Guasch, and Santiago Badia. Time dependent subscales in the stabilized finite element approximation of incompressible flow problems. Computer Methods in Applied Mechanics and Engineering, 196(21–24):2413–2430, 2007. doi: 10.1016/j.cma.2007.01.002.DOI
  39. [39]A. Masud and R. A. Khurram. A multiscale finite element method for the incompressible Navier–Stokes equations. Computer Methods in Applied Mechanics and Engineering, 195(13–16):1750–1777, 2006. doi: 10.1016/j.cma.2005.05.048.DOI
  40. [40]Ramon Codina, Santiago Badia, Joan Baiges, and Javier Principe. Variational multiscale methods in computational fluid dynamics. In Erwin Stein, René de Borst, and Thomas J R Hughes, editors, Encyclopedia of Computational Mechanics Second Edition, pages 1–28. Wiley, 2017. doi: 10.1002/9781119176817.ecm2117.DOI
  41. [41]Ramon Codina and Orlando Soto. Approximation of the incompressible Navier–Stokes equations using orthogonal subscale stabilization and pressure segregation on anisotropic finite element meshes. Computer Methods in Applied Mechanics and Engineering, 193:1403–1419, 2004. doi: 10.1016/j.cma.2003.12.030.DOI
  42. [42]Ramon Codina and Santiago Badia. On some pressure segregation methods of fractional-step type for the finite element approximation of incompressible flow problems. Computer Methods in Applied Mechanics and Engineering, 195:2900–2918, 2006. doi: 10.1016/j.cma.2004.06.048.DOI
  43. [43]S. Dong, G. E. Karniadakis, and C. Chryssostomidis. A robust and accurate outflow boundary condition for incompressible flow simulations on severely-truncated unbounded domains. Journal of Computational Physics, 261:83–105, 2014. doi: 10.1016/j.jcp.2013.12.042.DOI
  44. [44]Sergey Charnyi, Timo Heister, Maxim A. Olshanskii, and Leo G. Rebholz. On conservation laws of Navier–Stokes Galerkin discretizations. Journal of Computational Physics, 337:289–308, 2017. doi: 10.1016/j.jcp.2017.02.039.DOI
  45. [45]Oriol Colomés, Santiago Badia, Ramon Codina, and Javier Principe. Assessment of variational multiscale models for the large eddy simulation of turbulent incompressible flows. Computer Methods in Applied Mechanics and Engineering, 285:32–63, 2015. doi: 10.1016/j.cma.2014.10.041.DOI
  46. [46]Maxim Olshanskii, Gert Lube, Timo Heister, and Johannes Löwe. Grad–div stabilization and subgrid pressure models for the incompressible Navier–Stokes equations. Computer Methods in Applied Mechanics and Engineering, 198(49–52):3975–3988, 2009. doi: 10.1016/j.cma.2009.09.005.DOI
  47. [47]Y. Bazilevs, V. M. Calo, J. A. Cottrell, T. J. R. Hughes, A. Reali, and G. Scovazzi. Variational multiscale residual-based turbulence modeling for large eddy simulation of incompressible flows. Computer Methods in Applied Mechanics and Engineering, 197(1–4):173–201, 2007. doi: 10.1016/j.cma.2007.07.016.DOI
  48. [48]Marco F. P. ten Eikelder and Ido Akkerman. Correct energy evolution of stabilized formulations: The relation between VMS, SUPG and GLS via dynamic orthogonal small-scales and isogeometric analysis. II: The incompressible Navier–Stokes equations. Computer Methods in Applied Mechanics and Engineering, 340:1135–1154, 2018. doi: 10.1016/j.cma.2018.02.030.
  49. [49]Yuri Bazilevs and Ido Akkerman. Large eddy simulation of turbulent Taylor–Couette flow using isogeometric analysis and the residual-based variational multiscale method. Journal of Computational Physics, 229(9):3402–3414, 2010. doi: 10.1016/j.jcp.2010.01.008.DOI
  50. [50]FEATFLOW. DFG benchmark 2D-3 (Re=100, fixed time interval): benchmark settings, reference and visualization results. FEATFLOW CFD Benchmarking Project, n.d.. URL https://www.mathematik.tu-dortmund.de/~featflow/en/benchmarks/cfdbenchmarking/flow/dfg_benchmark3_re100.html. Accessed: 2026-10-22.
  51. [51]FEATFLOW. DFG flow around a cylinder: 3D laminar case reference values. FEATFLOW CFD Benchmarking Project, n.d.. URL https://www.mathematik.tu-dortmund.de/~featflow/en/benchmarks/cfdbenchmarking/flow/dfg_flow3d/dfg_flow3d_reference.html. Accessed: 2026-10-22.
  52. [52]Igor A. Baratta, Joseph P. Dean, Jørgen S. Dokken, Michal Habera, Jack S. Hale, Chris N. Richardson, Marie E. Rognes, Matthew W. Scroggs, Nathan Sime, and Garth N. Wells. DOLFINx: The next generation FEniCS problem solving environment, 2025. Preprint.
  53. [53]Satish Balay, Shrirang Abhyankar, Mark F. Adams, Steven Benson, Jed Brown, Peter Brune, Kris Buschelman, Emil M. Constantinescu, Lisandro Dalcin, Alp Dener, Victor Ejkhout, Jacob Faibussowitsch, William D. Gropp, Václav Hapla, Tobin Isaac, Pierre Jolivet, Dmitry Karpeev, Dinesh Kaushik, Matthew G. Knepley, Fande Kong, Scott Kruger, Dave A. May, Lois Curfman McInnes, Richard Tran Mills, Lawrence Mitchell, Todd Munson, Jose E. Roman, Karl Rupp, Patrick Sanan, Jason Sarich, Barry F. Smith, Hansol Suh, Stefano Zampini, Hong Zhang, Hong Zhang, and Junchao Zhang. PETSc/TAO users manual. Technical Report ANL-21/39, Revision 3.25, Argonne National Laboratory, 2026.
  54. [54]Carlo Brunelli, Matija Avirovic, Bart Janssens, Benoît G. Marinus, Koen Hillewaert, and Mark Runacres. Prediction of laminar separation bubble on airfoils at low Reynolds number. Flow, Turbulence and Combustion, 116:27, 2026. doi: 10.1007/s10494-025-00727-7.DOI
  55. [55]Leopoldo P. Franca, Thomas J. R. Hughes, and Rolf Stenberg. Stabilized finite element methods. In Max D. Gunzburger and Roy A. Nicolaides, editors, Incompressible Computational Fluid Dynamics: Trends and Advances, pages 87–108. Cambridge University Press, 1993. doi: 10.1017/CBO9780511574856.005.DOI
  56. [56]M. Schäfer, S. Turek, F. Durst, E. Krause, and R. Rannacher. Benchmark computations of laminar flow around a cylinder. In Ernst Heinrich Hirschel, editor, Flow Simulation with High-Performance Computers II, volume 48 of Notes on Numerical Fluid Mechanics (NNFM), pages 547–566. Vieweg+Teubner Verlag, Wiesbaden, 1996. doi: 10.1007/978–3–322–89849–4_39.DOI
  57. [57]Volker John. Reference values for drag and lift of a two-dimensional time-dependent flow around a cylinder. International Journal for Numerical Methods in Fluids, 44(7):777–788, 2004. doi: 10.1002/fld.679.DOI
  58. [58]Evren Bayraktar, Otto Mierka, and Stefan Turek. Benchmark computations of 3D laminar flow around a cylinder with CFX, OpenFOAM and FeatFlow. International Journal of Computational Science and Engineering, 7(3):253–266, 2012. doi: 10.1504/IJCSE.2012.048245.DOI

Paper details

vlnc:W_2a2fftxd
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

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

  • Physics and Astronomy

Codex

View statement
Looking for related papers…
Contents