Verifiability and Limit Consistency of Eddy Viscosity Large Eddy Simulation Reduced Order Models

TL;DR

This paper introduces the Ladyzhenskaya ROM (L-ROM), a generalized eddy viscosity reduced order model, and proves its verifiability and limit consistency, supported by numerical tests on convection-dominated flows.

Contribution

The paper develops the L-ROM, extending classical EV-ROMs, and provides rigorous proofs of its verifiability and limit consistency, enhancing ROM reliability for turbulent flows.

Findings

L-ROM is verifiable with bounded error by closure error.

L-ROM and S-ROM are limit consistent as ROM dimension increases.

Numerical tests confirm verifiability and limit consistency in convection-dominated problems.

Abstract

Large eddy simulation reduced order models (LES-ROMs) are ROMs that leverage LES ideas (e.g., filtering and closure modeling) to construct accurate and efficient ROMs for convection-dominated (e.g., turbulent) flows. Eddy viscosity (EV) ROMs (e.g., Smagorinsky ROM (S-ROM)) are LES-ROMs whose closure model consists of a diffusion-like operator in which the viscosity depends on the ROM velocity. We propose the Ladyzhenskaya ROM (L-ROM), which is a generalization of the S-ROM. Furthermore, we prove two fundamental numerical analysis results for the new L-ROM and the classical S-ROM: (i) We prove the verifiability of the L-ROM and S-ROM, i.e, that the ROM error is bounded (up to a constant) by the ROM closure error. (ii) We introduce the concept of ROM limit consistency (in a discrete sense), and prove that the L-ROM and S-ROM are limit consistent, i.e., that as the ROM dimension approaches the rank of the snapshot matrix, $d$, and the ROM lengthscale goes to zero, the ROM solution converges to the \emph{``true solution"}, i.e., the solution of the $d$-dimensional ROM. Finally, we illustrate numerically the verifiability and limit consistency of the new L-ROM and S-ROM in two under-resolved convection-dominated problems that display sharp gradients: (i) the 1D Burgers equation with a small diffusion coefficient; and (ii) the 2D lid-driven cavity flow at Reynolds number $Re=15,000$.