Variational Multiscale Proper Orthogonal Decomposition: Navier-Stokes Equations

1 Introduction

• For turbulent flows, simple POD-G-ROMs generally produce erroneous results because discarded modes can significantly affect dynamics despite containing little kinetic energy.The proposed VMS approach adds eddy viscosity only to small resolved scales, aiming to eliminate their oscillations without polluting large-scale components.Source: Introduction · p. 2
• The paper extends VMS-POD-ROM to Navier–Stokes equations and presents error analysis covering finite-element space and backward-Euler time discretization.The analysis considers spatial, temporal, and POD truncation error contributions.Source: Introduction · pp. 2–3
• The model is tested on a three-dimensional turbulent cylinder flow at Re = 1000, and its theoretical error estimates are checked in a two-dimensional flow simulation.The tests investigate physical accuracy and confirm the theoretical estimates, respectively.Source: Introduction · p. 3

2 Variational Multiscale Proper Orthogonal Decomposition

• The model is formulated for incompressible Navier–Stokes flow in two or three dimensions, with finite-element velocity and pressure spaces used in its analysis.The stated spaces use continuous piecewise-polynomial approximations, with the pressure degree one below the velocity degree.Source: §2 · pp. 3–4

2.1 Proper Orthogonal Decomposition

• POD constructs a low-dimensional basis that optimally approximates velocity snapshots, using an eigenproblem for the snapshot correlation matrix.The eigenvalues are ordered by descending magnitude, and the resulting modes represent descending kinetic-energy content.Source: §2.1 · p. 5
• Galerkin projection onto a few retained POD modes yields a computationally efficient ROM, but the standard POD-G-ROM is generally limited to laminar flows.The pressure term vanishes because the POD modes are solenoidal and satisfy the appropriate boundary conditions.Source: §2.1 · p. 6

2.2 Variational Multiscale Method

• VMS treats discarded POD modes as unresolved scales, the first R retained modes as resolved large scales, and the remaining retained modes as resolved small scales.The energy ordering of POD modes makes them a natural basis for this scale separation.Source: §2.2 · p. 7
• The PR-VMS-POD-ROM uses a projection operator in an eddy-viscosity term, with a constant positive coefficient, to model effects of high-index POD modes.The method introduces extra eddy viscosity only on resolved small scales.Source: §2.2 · pp. 6–7
• When R = r or α = 0, the model reduces to standard POD-G-ROM; when R = 0, it becomes the mixing-length POD-ROM.These limits correspond to introducing no eddy viscosity or applying it to all POD-ROM modes.Source: §2.2 · p. 7
• The PR-VMS-POD-ROM differs from the cited VMS-POD-ROM in its projection-based treatment and constant eddy-viscosity coefficient.The comparison distinguishes the projection operator and coefficient choice used in the models.Source: §2.2 · p. 7
• The full model discretization uses backward Euler with time step Δt and finite elements of mesh size h and polynomial degree m.The resulting PR-VMS-POD-ROM is advanced at discrete times t_k = kΔt.Source: §2.2 · p. 8

3.1 Preliminaries

• Preliminaries establish approximation and ROM stability bounds; the authors note estimates (3.16)–(3.17) are not known to be proven in a reference, though they can be assembled from [7, 8, 11] and are simulation-confirmed.The main theorem also relies on assumptions about finite-element approximation and POD projection error.Source: §3.1 · pp. 8, 9, 10, 13, 14, §3.2 · p. 14

3.2 Main Results

• Under regularity, finite-element, and POD-projection assumptions, the main error bound contains ∆t^2, h^{2m}, and h^{2m+2} terms, with the latter terms weighted by POD stiffness-matrix norms.The estimate applies to the PR-VMS-POD-ROM discretization.Source: §3 · p. 8, §3.2 · pp. 14–15
• The expression combines (1 + ∥S_r∥_2 + ∥S_R∥_2)Δt^2, h^2m, and (1 + ∥S_r∥_2 + ∥S_R∥_2)h^2m+2.The proof obtains this expression after dropping the third term on the left-hand side and applying (3.30), (3.25), and the triangle inequality.Source: §3.2 · pp. 20–21

4.1 Physical Accuracy

• POD-G-ROM's a4 amplitude is nine times the DNS projection, while ML-POD-ROM's final a4 magnitude is one eighth of DNS.The ML-POD-ROM coefficient accuracy degrades toward the end of the simulation.Source: §4.1 · p. 23
• PR-VMS-POD-ROM yields more accurate a1 and a4 time evolutions than POD-G-ROM and ML-POD-ROM over the tested interval.The coefficient trajectories are compared over [0, 300].Source: Figure 2 · p. 23, §4.1 · p. 24
• PR-VMS-POD-ROM has a more accurate energy spectrum than POD-G-ROM and ML-POD-ROM, despite high-frequency oscillations.POD-G-ROM overestimates the DNS spectrum, while ML-POD-ROM underestimates it, especially at higher frequencies.Source: Figure 3 · p. 23, §4.1 · p. 24

4.2 Numerical Accuracy

• The final-time error converges nearly linearly with the time step, as predicted by the theoretical estimate.The test uses h = 1/64, r = 99, R = 95, and α = 10^-3.Source: §4.2 · pp. 24–25, Figure 4 · p. 25
• As R increases, the error shows almost quadratic convergence against the POD-tail contribution, exceeding the theoretical linear prediction.The numerical test examines the POD contribution to the final-time L2 error.Source: §4.2 · p. 25

• Figure 5 relates the squared final-time L2 total error E2 to the EV-induced POD error contribution Pd for h = 1/64, ∆t = 10−4, r = 99, and α = 10−3.Source: Figure 5 · p. 26
• The PR-VMS-POD-ROM was physically accurate and computationally efficient in 3D cylinder flow at Re = 1000, and simulations illustrated its error estimates.Source: Conclusions · p. 26
• Its rigorous full-discretization error estimate accounts for FE spatial, backward-Euler temporal, and POD-truncation contributions.Source: Conclusions · p. 26
• The EV coefficient α is constant in this model; more accurate choices, including the Smagorinsky model, are planned.Source: Conclusions · p. 26
• Reducing computational cost through a different time discretization of the VMS term remains planned work.Source: Conclusions · p. 27