Hostname: page-component-7b7567c8c-fvjsw Total loading time: 0 Render date: 2026-11-15T16:48:28.812Z Has data issue: false hasContentIssue false

Sparsifying the resolvent forcing mode via gradient-based optimisation

Published online by Cambridge University Press:  06 July 2022

Calum S. Skene*
Affiliation:
Department of Mechanical and Aerospace Engineering, University of California, Los Angeles, CA 90095, USA
Chi-An Yeh
Affiliation:
Department of Mechanical and Aerospace Engineering, University of California, Los Angeles, CA 90095, USA
Peter J. Schmid
Affiliation:
Department of Mathematics, Imperial College London, London SW7 2AZ, UK
Kunihiko Taira
Affiliation:
Department of Mechanical and Aerospace Engineering, University of California, Los Angeles, CA 90095, USA
*
†Email address for correspondence: c.s.skene@leeds.ac.uk

Abstract

We consider the use of sparsity-promoting norms in obtaining localised forcing structures from resolvent analysis. By formulating the optimal forcing problem as a Riemannian optimisation, we are able to maximise cost functionals whilst maintaining a unit-energy forcing. Taking the cost functional to be the energy norm of the driven response results in a traditional resolvent analysis and is solvable by a singular value decomposition (SVD). By modifying this cost functional with the $L_1$-norm, we target spatially localised structures that provide an efficient amplification in the energy of the response. We showcase this optimisation procedure on two flows: plane Poiseuille flow at Reynolds number $Re=4000$, and turbulent flow past a NACA 0012 aerofoil at $Re=23\,000$. In both cases, the optimisation yields sparse forcing modes that maintain important features of the structures arising from an SVD in order to provide a gain in energy. These results showcase the benefits of utilising a sparsity-promoting resolvent formulation to uncover sparse forcings, specifically with a view to using them as actuation locations for flow control.

Information

Type
JFM Papers
Copyright
© The Author(s), 2022. Published by Cambridge University Press
Figure 0

Figure 1.A sketch illustrating the concept of Riemannian optimisation. First, the Euclidean gradient $\boldsymbol {\nabla }_{\boldsymbol {f}^{n}_\boldsymbol{\mathsf{M}}} \sigma ^{(L_2)}$ is found from the vector $\boldsymbol {f}^{n}_\boldsymbol{\mathsf{M}}$ that is situated on the hypersphere $\mathcal {S}$. This vector is then mapped to the tangent space $T_{\boldsymbol {f}^{n}_\boldsymbol{\mathsf{M}}}\mathcal {S}$ via the projection $\text {Proj}_{\boldsymbol {f}^{n}_\boldsymbol{\mathsf{M}}}$. Next, the Riemannian gradient is extended along the tangent space by the step size $\alpha$. Finally, we map this gradient back to the manifold via the retraction $R_{\boldsymbol {f}^{n}_\boldsymbol{\mathsf{M}}}$, yielding the updated forcing vector $\boldsymbol {f}^{n+1}_\boldsymbol{\mathsf{M}}$. For varying values of $\alpha$, the retraction traces out a smooth curve over the manifold, displayed as a dotted line. The link between $\alpha$ and $\theta$ is also shown.

Figure 1

Figure 2.Sketches of the $L_2$- and $L_1$-norms. The effect of a coordinate rotation on the norms is also demonstrated. (a) The unit $L_2$-norm (circle) and $L_1$-norm (diamond) with respect to the coordinate system $(x,y)$. (b) The $L_2$-norm (green dashed circle) and $L_1$-norm (green dashed square) with respect to the coordinate system $(\xi,\eta )$.

Figure 2

Figure 3.The eigenvalue-based wavemakers and resolvent wavemakers for cylinder flow at $Re=100$ shown for illustration. Computations are performed with the immersed boundary projection method ibmos (Fosas de Pando 2020). (a) Wavemaker for the real part of the $u$-component. (b) Wavemaker for the real part of the $v$-component. (c) Resolvent wavemaker for the $u$-component at $St=0.162$. (d) Resolvent wavemaker for the $v$-component at ${St}=0.162$.

Figure 3

Figure 4.The optimal gains over frequency for Poiseuille flow. Our analysis focuses on the optimal gain that occurs at $\omega =0.278$, and on the second peak at $\omega =1.14$ (both shown with a black $+$).

Figure 4

Figure 5.The real parts of the forcing and response modes obtained by SVD of the resolvent at $\omega =0.278$. The forcing is unit-norm, whereas the response mode has norm equal to the gain. (a) Real part of the full forcing mode, $u$-velocity component. (b) Real part of the full forcing mode, $v$-velocity component. (c) Real part of the full response mode, $u$-velocity component. (d) Real part of the full response mode, $v$-velocity component.

Figure 5

Figure 6.The real parts of the forcing and response modes obtained by SVD of the resolvent at $\omega =1.14$. The forcing is unit-norm, whereas the response mode has norm equal to the gain. (a) Real part of the full forcing mode, $u$-velocity component. (b) Real part of the full forcing mode, $v$-velocity component. (c) Real part of the full response mode, $u$-velocity component. (d) Real part of the full response mode, $v$-velocity component.

Figure 6

Figure 7.The real parts of the forcing and response modes obtained by sparsification at $\omega =0.278$. The forcing is unit-norm, whereas the response mode has norm equal to the gain. (a) Real part of the sparse forcing mode, $u$-velocity component. (b) Real part of the sparse forcing mode, $v$-velocity component. (c) Real part of the sparse response mode, $u$-velocity component. (d) Real part of the sparse response mode, $v$-velocity component.

Figure 7

Figure 8.The real parts of the forcing and response modes obtained using sparsification at $\omega =1.14$. The forcing is unit-norm, whereas the response mode has norm equal to the gain. The location of the sparse forcing mode is shown with $\times$. (a) Real part of the sparse forcing mode, $u$-velocity component. (b) Real part of the sparse forcing mode, $v$-velocity component. (c) Real part of the sparse response mode, $u$-velocity component. (d) Real part of the sparse response mode, $v$-velocity component.

Figure 8

Figure 9.The streamwise velocity component of the base-flow.

Figure 9

Figure 10.The optimal gains against the Strouhal number for flow past an aerofoil. Our analysis will focus on the optimal gain, which occurs at (a) ${St}\approx 5.22$ for $\beta =0$, and (b) ${St}\approx 5.90$ for $\beta =20{\rm \pi}$ (both highlighted with a black $+$).

Figure 10

Figure 11.The full resolvent modes for aerofoil flow at ${St}=5.22$ for $\beta =0$, and ${St}=5.90$ for $\beta =20{\rm \pi}$. The linearised component of the streamwise momentum is shown. (a) Full forcing mode for $\beta =0$. (b) Full response mode for $\beta =0$. (c) Full forcing mode for $\beta =20{\rm \pi}$. (d) Full response mode for $\beta =20{\rm \pi}$.

Figure 11

Figure 12.The sparse resolvent modes for aerofoil flow at ${St}=5.22$ for $\beta =0$, and ${St}=5.90$ for $\beta =20{\rm \pi}$. The linearised component of the streamwise momentum is shown. (a) Sparse forcing mode for $\beta =0$. (b) Sparse response mode for $\beta =0$. (c) Sparse forcing mode for $\beta =20{\rm \pi}$. (d) Sparse response mode for $\beta =20{\rm \pi}$.

Figure 12

Figure 13.The resolvent wavemakers for aerofoil flow at ${St}=5.22$ for $\beta =0$, and ${St}=5.90$ for $\beta =20{\rm \pi}$. The linearised component of the streamwise momentum is shown. (a) Wavemaker for $\beta =0$. (b) Wavemaker for $\beta =20{\rm \pi}$.

Figure 13

Figure 14.The surface forcing on the suction side of the aerofoil at ${St}=5.22$ for $\beta =0$, and ${St}=5.90$ for $\beta =20{\rm \pi}$. The sparse components are shown with a star. (a) Surface forcing on the suction side of the aerofoil for $\beta =0$. (b) Surface forcing on the suction side of the aerofoil for $\beta =20{\rm \pi}$.

Figure 14

Figure 15.Dependence of the results on $\delta$ for the plane Poiseuille flow examples. (a) $L_1$-norm and gains with $\delta $ for $\omega =0.278$. (b) Convergence of the pseudo-Huber-norm to the true $L_1$-norm for $\omega =0.278$. (c) $L_1$-norm and gains with $\delta $ for $\omega =1.14$. (d) Convergence of the pseudo-Huber-norm to the true $L_1$-norm for $\omega =1.14$.

Figure 15

Figure 16.The optimisation convergence behaviour for the plane Poiseuille flow examples. Convergence behaviour of the conjugate gradient algorithm for plane Poiseuille flow at (a) $\omega =0.278$ and $\delta =5^{-7}$, (b) $\omega =1.14$ and $\delta =5^{-7}$.