Train yourself: self-compressing reduced-order models of turbulent flows

Ian Addison-Smith \aff 1, Igor A. Maia \aff 2, Benjamin Herrmann \aff 3,4, André V. G. Cavalieri \aff 2 · 2025

Read Train yourself: self-compressing reduced-order models of turbulent flows by Ian Addison-Smith \aff 1, Igor A. Maia \aff 2, Benjamin Herrmann \aff 3,4, André V. G. Cavalieri \aff 2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-04813.txt.

Introduction

Ian Addison-Smith and three colleagues compress reduced-order models of turbulent flow using the model's own output. Galerkin models of turbulence often need too many modes to stay accurate. Getting those modes from simulation data is itself expensive.

Starting from linearized Navier-Stokes modes built without prior data, they go further. The time series of modal coefficients reveals which modes truly matter in motion. Pruning by that evidence yields a self-compressed model with far fewer degrees of freedom. First-order accuracy is maintained, they report, offering cheaper flow models for control and design loops.

Catalog record

CREATORIan Addison-Smith \aff 1, Igor A. Maia \aff 2, Benjamin Herrmann \aff 3,4, André V. G. Cavalieri \aff 2
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-04813

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Train yourself: self-compressing reduced-order models of turbulent flows Ian Addison-Smith \aff 1, Igor A. Maia \aff 2, Benjamin Herrmann \aff 3,4, André V. G. Cavalieri \aff 2 Abstract Reduced-order models (ROMs) of turbulent flows based on Galerkin projection often require many degrees of freedom to resolve the dynamics of the turbulence, or simulation data to obtain an optimal modal basis. However, obtaining simulation data is computationally expensive, and the amount of data required to obtain a converged modal basis can increase this cost. Using the linearized Navier-Stokes equations, one can achieve spatial modes through the controllability and observability Gramians, which can yield a ROM without prior simulation data. In this work, we propose a self-compression of a ROM based on controllability modes, where the time series of the modal coefficients are leveraged to reduce the dimension of the ROM. In the self-compressed ROM (SCROM), we can maintain accurate first- and second-order statistics with respect to the DNS simulation, but in a further reduced dimension. The SCROM recovers spatial structures equivalent to proper orthogonal decomposition (POD) without relying on any simulation data, recombining spatial modes from linearized equations. This method leads to a novel ROM that can represent turbulence statistics in a data-free approach in a further reduced state space. \lefttitle I. Addison-Smith, I.A. Maia, B. Herrmann and A.V.G. Cavalieri \righttitle Journal of Fluid Mechanics \corresau Ian Addison-Smith, keywords 1 Introduction In turbulent flows, obtaining reduced-order models (ROMs) from Galerkin projections may require modal bases with a large number of modes to accurately represent the relevant dynamics ( Rowley & Dawson 2017 ) . Suitable modal bases for ROMs could be obtained from data using proper orthogonal decomposition (POD) (e.g. Noack et al. 2003 ; Smith et al. 2005 ; Aubry et al. 1988 ; Chua Khoo et al. 2022 ; Sato & Schmidt 2025 ), or from the linearized Navier-Stokes equations (e.g. Cavalieri & Nogueira 2022 ; Maia & Cavalieri 2025 ; Zong et al. 2025 ). Statistical convergence of POD, as well as other data-driven modal decompositions ( Taira et al. 2017 ; Herrmann et al. 2021 ; Baddoo et al. 2022 ; Schmid 2022 ; Baddoo et al. 2023 ) , requires large sequences of full flow field data snapshots. In particular, for wall-bounded turbulence, the streamwise and spanwise directions are homogeneous, so the modal bases are obtained from different streamwise and spanwise wavenumber pairs with a fixed truncation of eigenvectors to represent the wall-normal direction. Typically, the resulting models based on the linearized equations have a larger amount of degrees of freedom compared with the POD ones, but with the advantage that the resulting ROM does not require any prior data. In this work, we present a framework to push for a further reduction in ROMs based on a modal bases from linearized equations. This self-compression leverages the time series of the modal coefficient of a ROM to solve an eigenvalue problem, where an energy optimal projection over a compressed space is obtained from the ROM. This approach, in comparison with common applications, utilizes data from the model itself, thereby maintaining the self-compression and ROM as data-free approaches. Therefore, one can significantly decrease the number of degrees of freedom of the ROM, while preserving a reasonable accuracy of first- and second-order statistics. Also, the computational cost can be significantly reduced in comparison with the ROM, which is relevant for control applications (e.g. Maia & Cavalieri 2025 ) In particular, we study a turbulent Couette flow with a Reynolds number ( \Rey=500 ), where we compare DNS simulations with a ROM built from controllability modes (same approach as Cavalieri & Nogueira 2022 ) and a self-compressed ROM (SCROM). Then we compare first- and second-order statistics of the three approaches to assess how much of the statistics are retained from the SCROM. Finally, we seek physical insight from the SCROM, obtaining spatial structures that compose the SCROM and the amount of wavenumber pairs that are selected in the model 2 Methods 2.1 Reduced-order model formulation We consider plane Couette flow, wherein two infinite planes separated by a wall-normal distance 2h move in opposite directions with the same velocity magnitude U_{w} . All quantities are normalized by wall velocity and half-wall separation. The domain is defined in Cartesian coordinates by (x,y,z) where x is the streamwise, y is the wall-normal and z is the spanwise direction with dimensions [0,2\pi)\times[-1,1]\times[0,\pi) , boundary conditions \mathbf{u}|_{y=\pm 1}=(\pm 1,0,0) in the upper and lower walls and periodic boundary conditions on streamwise and spanwise directions. The velocity field \boldsymbol{u} is decomposed into a laminar solution \boldsymbol{u}_{0}=(y,0,0) plus fluctuations, \boldsymbol{u}^{\prime} , around this state, \boldsymbol{u}(\boldsymbol{x},t)=\boldsymbol{u}_{0}(y)+\boldsymbol{u}^{\prime}(\boldsymbol{x},t), (1) The governing equations for velocity fluctuations in an incompressible fluid are given by \displaystyle\frac{\partial\boldsymbol{u}^{\prime}}{\partial t}+(\boldsymbol{u}_{0}\cdot\boldsymbol{\nabla})\boldsymbol{u}^{\prime}+(\boldsymbol{u}^{\prime}\cdot\boldsymbol{\nabla})\boldsymbol{u}_{0} \displaystyle+(\boldsymbol{u}^{\prime}\cdot\boldsymbol{\nabla})\boldsymbol{u}^{\prime}=-\boldsymbol{\nabla}p+\frac{1}{Re}\boldsymbol{\nabla}^{2}\boldsymbol{u}^{\prime}, (2) \displaystyle\boldsymbol{\nabla}\cdot\boldsymbol{u}^{\prime} \displaystyle=0. (3) where p is the pressure and Re=U_{w}h/\nu is the Reynolds number, with \nu the kinematic viscosity of the fluid. Expanding the fluctuating velocity field in terms of a basis of divergence-free orthonormal modes, \boldsymbol{\phi}_{j}(\boldsymbol{x}) , we can express the velocity as in ( Cavalieri & Nogueira 2022 ) \boldsymbol{u}^{\prime}(\boldsymbol{x},t)=\sum_{j=1}^{N}a_{j}(t)\boldsymbol{\phi}_{j}(\boldsymbol{x}), (4) where a_{j}(t) are the corresponding modal expansion coefficients. Then, we substitute ( 4 ) into ( 2 ) and take an inner product with \boldsymbol{\phi}_{i} to obtain the following equation for the dynamics of the coefficients \frac{\mathrm{d}a_{i}}{\mathrm{d}t}=\frac{1}{\Rey}\sum_{j=1}^{N}L_{ij}a_{j}+\sum_{j=1}^{N}\tilde{L}_{ij}a_{j}+\sum_{j=1}^{N}\sum_{k=1}^{N}Q_{ijk}a_{j}a_{k}, (5) where {subeqnarray} L_ij = ⟨∇^2 ϕ _j , ϕ _i ⟩, ~ L _ij = - ⟨( ϕ _j ⋅∇ u _0 + u _0 ⋅∇ ϕ _j ), ϕ _i ⟩, Q_ijk = - ⟨( ϕ _j ⋅∇ ϕ _k ), ϕ _i ⟩, and the symbols \left<\cdot\right> denote an inner product, which we define in physical space as \langle\boldsymbol{f},\boldsymbol{g}\rangle=\frac{1}{4\pi^{2}}\int_{0}^{\pi}\int_{-1}^{1}\int_{0}^{2\pi}\boldsymbol{g}(\boldsymbol{x})^{T}\boldsymbol{f}(\boldsymbol{x})\mathrm{d}x\mathrm{d}y\mathrm{d}z, (6) for real vector fields \boldsymbol{f} and \boldsymbol{g} , with (\ )^{T} denoting the transpose. The linear terms, L_{ij} and \tilde{L}_{ij} , represent the viscous term and the linear interaction with the laminar solutions, respectively, and the nonlinear term, Q_{ijk} , represents the quadratic nonlinear interactions between the different modes. The term corresponding to the projection of \phi_{i} upon pressure term vanishes due to the zero-pressure gradient condition and the divergence-free property of the modes. Here, for the choice of the modes \boldsymbol{\phi}_{j} , we follow the approach described in Cavalieri & Nogueira 2022 , who followed the method of Jovanović & Bamieh 2005 , wherein POD modes of the stochastic response of the linearized Navier-Stokes equations, also known as controllability modes, are considered. The controllability Gramian is computed for different streamwise/spanwise wavenumber combinations (k_{x},k_{z}) , through a Lyapunov equation. The resulting basis consists of the eigenvectors of the Gramian that are ranked by their corresponding eigenvalues in terms of sustained energy. For each wavenumber pair considered, n_{\text{modes}} eigenvectors are retained in the basis. 2.2 Self-compressing reduced-order model Given the timeseries of the modal coefficients a_{j}(t) obtained in a simulation of a ROM, we can separate these coefficients by wavenumber combination and create the following snapshot matrix for (k_{x}/\alpha,k_{z}/\beta)=(m_{x},m_{z}) , \mathsfbi{X}_{(m_{x},m_{z})}=\begin{bmatrix}\boldsymbol{a}_{(m_{x},m_{z})}(t_{1})&\boldsymbol{a}_{(m_{x},m_{z})}(t_{2})&\cdots&\boldsymbol{a}_{(m_{x},m_{z})}(t_{N_{s}})\end{bmatrix}, (7) where (\alpha,\beta)=(2\pi/L_{x},2\pi/L_{z}) are the fundamental streamwise and spanwise wavenumber, (k_{x}/\alpha,k_{z}/\beta) are the wavenumber pair, and N_{s} is the total number of snapshots obtained in the simulation of the ROM. This results in a matrix of dimensions (n_{\text{modes}},N_{s}) containing the total snapshots for a number of modal coefficients corresponding to

Read the complete text