betaFeedback

WorksPhysics and Astronomy

Wavelet-based resolvent analysis for statistically-stationary and temporally-evolving flows

Abstract

This work introduces a formulation of resolvent analysis that uses wavelet transforms rather than Fourier transforms in time. This allows resolvent analysis to be extended to turbulent flows with non-stationary means in addition to statistically-stationary flows. The optimal resolvent modes for this formulation correspond to the potentially time-transient structures that are most amplified by the linearized Navier-Stokes operator. We validate this methodology for turbulent channel flow and show that the wavelet-based and Fourier-based resolvent analyses are equivalent for statistically-stationary flows. We then apply the wavelet-based resolvent analysis to study the transient growth mechanism in the buffer layer of a turbulent channel flow by windowing the resolvent operator in time and frequency. The method is also applied to temporally-evolving parallel shear flows such as an oscillating boundary layer and three-dimensional channel flow, in which a lateral pressure gradient perturbs a fully-developed turbulent flow in a channel.

Nomenclature

x1x_1 = streamwise direction

x2x_2 = wall-normal direction

x3x_3 = spanwise direction

uˉi\bar{u}_i = total velocity in the xix_i direction

UiU_i = average velocity over ensembles and homogeneous directions in the xix_i direction

uiu_i = fluctuating velocity in the xix_i direction

pˉ\bar{p} = total pressure

PP = average pressure over ensembles and homogeneous directions

pp = fluctuating pressure

ν\nu = kinematic viscosity

ρ\rho = density

⟨⋅⟩\langle\cdot\rangle = average over ensembles and homogeneous directions

FF, (⋅^)(\hat{\cdot}) = Fourier transform in homogeneous directions and time

WW, (⋅~)(\tilde{\cdot}) = wavelet transform in time and Fourier transform in homogeneous directions

LL = discretized Laplacian

DtD_t = discretized temporal derivative

DiD_i = discretized spatial derivative in the xix_i direction

TT = time horizon

NxiN_{x_i} = number of spatial points in the xix_i direction

Δxi\Delta x_i = grid size in the xix_i direction

NtN_t = number of discrete temporal points

kik_i = wavenumber in the xix_i direction

ω\omega = temporal frequency

φ\varphi = wavelet function

α\alpha = wavelet scale parameter

β\beta = wavelet shift parameter

H^,H~\hat{\mathcal{H}}, \widetilde{\mathcal{H}} = resolvent operator (Fourier- and wavelet-based resolvent, respectively)

ϕ^,ϕ~\hat{\phi}, \widetilde{\phi} = principal resolvent forcing mode (for the Fourier- and wavelet-based resolvent, respectively)

ψ^,ψ~\hat{\psi}, \widetilde{\psi} = principal resolvent response mode (for the Fourier- and wavelet-based resolvent, respectively)

ψi\psi_i = ii-th component of the principal resolvent response mode

σ\sigma = principal resolvent singular value

B\mathcal{B} = windowing matrix that restricts the forcing to a subset of the full space

C\mathcal{C} = windowing matrix that restricts the response to a subset of the full space

(⋅)†(\cdot)^\dagger = Moore-Penrose pseudo-inverse

δ\delta = channel half-height

ReτRe_\tau = friction Reynolds number

uτu_\tau = friction velocity

(⋅)+(\cdot)^+ = viscous units

Ω\Omega = channel wall oscillation frequency for the Stokes boundary layer

ReΩRe_\Omega = Reynolds number for the Stokes boundary layer

δΩ\delta_\Omega = Stokes laminar boundary layer thickness

Umax⁡U_{\max} = maximum streamwise wall velocity for the Stokes boundary layer

Ui,rmsU_{i,rms} = root-mean-square velocities profiles in the xix_i direction

Π\Pi = ratio of the mean spanwise and streamwise pressure gradient

uτ,0u_{\tau,0} = friction velocity at time t=0t = 0

uτ,Tu_{\tau,T} = friction velocity at time t=Tt = T

τi\tau_i = wall-shear stress in the xix_i direction

γ\gamma = wall-shear angle tan⁡−1(τ3/τ1)\tan^{-1}(\tau_3/\tau_1)

Introduction

Fundamental studies of unsteady aerodynamics and turbulent flow have mostly focused on statistically stationary configurations, where all statistics are invariant under a shift in time. However, in real-world applications of external aerodynamics, truly unsteady transient effects become important. Such events include separation leading to stall, a sudden change in yaw angle, gust encounters, and shockwave formation. The additional complexity of these transient problems makes it harder to perform controlled experiments, both numerical and in the laboratory. Our goal is to develop reduced-order models that can be used to understand, predict, and control highly-unsteady transient turbulent flow in various engineering systems.

Resolvent analysis has been a popular reduced-order model for understanding a wide variety of turbulent flows. Resolvent analysis refers to the inspection of the resolvent operator, a linear operator that consists of the linearized Navier-Stokes and maps forcing inputs (which can be due to the nonlinear advection terms, or other exogenous inputs) onto the flow states. This operator governs how inputs are amplified by the linear dynamics of the system. Its singular value decomposition (SVD) identifies the inputs to which the linearized equations of motion are most receptive, their gains, and the most amplified outputs [1, 2]. Often, the truncated SVD is enough to capture most of the effect of the operator due to the fast decay of the singular values. The resulting low-rank approximation of the forcing-response dynamics of the full system is extremely valuable for modeling, controlling, and understanding the physics of fluid flows [3, 4, 5, 6, 7, 8, 9], and has been extended to compressible and stratified flows [10, 11, 12]. However, a Fourier transform in time is traditionally used to formulate the resolvent operator, which restricts its analysis and application to statistically-steady and quasi-periodic flows [13]. Indeed, the resulting SVD modes will be Fourier modes in time, and cannot represent temporally local effects.

To construct the resolvent operator, we instead propose using a wavelet transform [14] in time. The most effective inputs and the most amplified outputs given by the SVD of the newly-formulated resolvent operator will be spatiotemporally localized. There have been numerous research efforts applying wavelet methods in fluid mechanics, with early work predating the application of resolvent analysis to turbulent flows and focusing on spatial transformations [15, 16].

More recent work has focused on data-driven wavelet transforms [17, 18]. The proposed research effort will provide an important bridge connecting such data-driven analyses to the underlying physics, analogous to recent work connecting and comparing resolvent analysis to data-driven spectral proper orthogonal decomposition [19, 20, 21].

In the present work, we develop and validate wavelet-based resolvent analysis for a variety of systems, ranging from quasi-parallel wall-bounded turbulent flows to spatio-temporally evolving systems. In §III, we formulate the Fourier- and wavelet-based resolvents; we highlight the properties of the wavelet transform and justify the choice of wavelet basis. In §IV we first establish the equivalence of Fourier- and wavelet-based resolvent analyses for the statistically stationary turbulent channel flow and showcase the additional capacity of the wavelet-based resolvent to capture the linear transient growth of streaks under transient forcing. We also apply wavelet-based resolvent analysis to statistically non-stationary flows in §V, notably the Stokes boundary layer flow as well as a turbulent channel flow subjected to a sudden lateral pressure gradient. Finally, the summary of the work and a discussion of the results are given in §VI.

Mathematical formulation

Fourier-based resolvent analysis

The incompressible Navier-Stokes equations in Einstein notation are given by

∂uˉi∂t+uˉj∂uˉi∂xj=−1ρ∂p~∂xi+ν∂2uˉi∂xj∂xj,∂uˉi∂xi=0,(1)\frac{\partial\bar{u}_i}{\partial t}+\bar{u}_j\frac{\partial\bar{u}_i}{\partial x_j}=-\frac{1}{\rho}\frac{\partial\tilde{p}}{\partial x_i}+\nu\frac{\partial^2\bar{u}_i}{\partial x_j\partial x_j},\qquad\frac{\partial\bar{u}_i}{\partial x_i}=0, \tag*{(1)}

where uˉi\bar{u}_i is the total velocity (including the mean and the fluctuating component), p~\tilde{p} is the total pressure, ρ\rho is the density, and ν\nu is the kinematic viscosity. The total velocity can be divided into uˉi=Ui+ui\bar{u}_i=U_i+u_i, where UiU_i is the average over ensembles and homogeneous directions, and uiu_i is the fluctuating component. Similarly, pressure can be decomposed as p~=P+p\tilde{p}=P+p.

Equation (1) can then be divided into the mean and the fluctuating equations

∂Ui∂t+⟨uˉj∂uˉi∂xj⟩=−1ρ∂P∂xi+ν∂2Ui∂xj∂xj,∂Ui∂xi=0,(2)\frac{\partial U_i}{\partial t}+\left\langle\bar{u}_j\frac{\partial\bar{u}_i}{\partial x_j}\right\rangle=-\frac{1}{\rho}\frac{\partial P}{\partial x_i}+\nu\frac{\partial^2U_i}{\partial x_j\partial x_j},\qquad\frac{\partial U_i}{\partial x_i}=0, \tag*{(2)}
∂ui∂t+Uj∂ui∂xj+uj∂Ui∂xj=−1ρ∂p∂xi+ν∂2ui∂xj∂xj+fi,∂ui∂xi=0,(3)\frac{\partial u_i}{\partial t}+U_j\frac{\partial u_i}{\partial x_j}+u_j\frac{\partial U_i}{\partial x_j}=-\frac{1}{\rho}\frac{\partial p}{\partial x_i}+\nu\frac{\partial^2u_i}{\partial x_j\partial x_j}+f_i,\qquad\frac{\partial u_i}{\partial x_i}=0, \tag*{(3)}

where fif_i is the remaining nonlinear terms in the fluctuating equations and ⟨⋅⟩\langle\cdot\rangle is the average over ensembles and in homogeneous directions. Note that some of the terms in the fluctuating equations may be zero depending on the flow configuration. The equations above do not have an analytic solution unless in very particular situations and are most commonly solved numerically. Discretizing the fluctuating equations, we get

Dtui+UjDjui+ujdUi,j=−1ρDip+νLui+fi,Diui=0,(4)D_tu_i+U_jD_ju_i+u_jdU_{i,j}=-\frac{1}{\rho}D_ip+\nu Lu_i+f_i,\qquad D_iu_i=0, \tag*{(4)}

where DtD_t is the discrete derivative in time, DiD_i is the discrete derivative in the xix_i direction, LL is the discrete Laplacian, and dUi,jdU_{i,j} denotes the matrix form of ∂Ui/∂xj\partial U_i/\partial x_j. Each discretized equation is an Nt×Nx1×Nx2×Nx3N_t\times N_{x_1}\times N_{x_2}\times N_{x_3}-dimensional system, where NtN_t is the temporal resolution, and NxiN_{x_i} are the spatial resolutions in the xix_i, i=1,2,3i=1,2,3 directions respectively. The discretized velocity and velocity gradient, UjU_j and dUi,jdU_{i,j}, are (Nt×Nx1×Nx2×Nx3)2(N_t\times N_{x_1}\times N_{x_2}\times N_{x_3})^2 diagonal matrices. In the traditional resolvent analysis, we apply the Fourier transform operator in homogeneous directions and time, FF, to the left in Eq. (4), leading to the equation

(FDtF−1)(Fui)+(FUjDjF−1)(Fui)+(Fuj)(FdUi,jF−1)=−1ρ(FDiF−1)(Fp)+ν(FLF−1)(Fui)+Ffi,(FDiF−1)(Fui)=0,(5)(FD_tF^{-1})(Fu_i)+(FU_jD_jF^{-1})(Fu_i)+(Fu_j)(FdU_{i,j}F^{-1})=-\frac{1}{\rho}(FD_iF^{-1})(Fp)+\nu(FLF^{-1})(Fu_i)+Ff_i,\qquad(FD_iF^{-1})(Fu_i)=0, \tag*{(5)}

where F−1F^{-1} is the inverse transformation, or equivalently

D^tu^i+U^jD^ju^i+dU^i,ju^j=−1ρD^ip^+νL^u^i+f^i,D^iu^i=0.(6)\widehat{D}_t\widehat{u}_i+\widehat{U}_j\widehat{D}_j\widehat{u}_i+\widehat{dU}_{i,j}\widehat{u}_j=-\frac{1}{\rho}\widehat{D}_i\widehat{p}+\nu\widehat{L}\widehat{u}_i+\widehat{f}_i,\qquad\widehat{D}_i\widehat{u}_i=0. \tag*{(6)}

Note that for an arbitrary matrix MM and vector bb, M^:=FMF−1\hat{M}:=FMF^{-1} and b^:=Fb\hat{b}:=Fb. For temporally stationary systems, this equation can typically be decoupled for each wavenumber and frequency combination. For example, in the case of channel flow where the flow is homogeneous in the streamwise (x1x_1) and spanwise (x3x_3) directions, the linear operator can be cast as

[u^1(k1,k3,ω)u^2(k1,k3,ω)u^3(k1,k3,ω)p^(k1,k3,ω)]=H^(k1,k3,ω)[f^1(k1,k3,ω)f^2(k1,k3,ω)f^3(k1,k3,ω)0](7)\begin{bmatrix} \hat{u}_1(k_1,k_3,\omega)\\ \hat{u}_2(k_1,k_3,\omega)\\ \hat{u}_3(k_1,k_3,\omega)\\ \hat{p}(k_1,k_3,\omega) \end{bmatrix} = \hat{\mathcal{H}}(k_1,k_3,\omega) \begin{bmatrix} \hat{f}_1(k_1,k_3,\omega)\\ \hat{f}_2(k_1,k_3,\omega)\\ \hat{f}_3(k_1,k_3,\omega)\\ 0 \tag*{(7)} \end{bmatrix}

for a given (k1,k3,ω)(k_1,k_3,\omega) triplet. Typically, the singular value decomposition of the linear operator H^∈C4Nx2×C4Nx2\hat{\mathcal{H}}\in\mathbb{C}^{4N_{x_2}\times\mathbb{C}^{4N_{x_2}}} is taken to study the left and right singular vectors as response and forcing modes, and the singular values as amplification factors or gains.

Wavelet-based resolvent analysis

Formulation

To account for transient behavior in the mean flow or the fluctuations, we introduce the wavelet-based resolvent analysis. The benefit of the wavelet transform in time is that it preserves both time and frequency information. The wavelet transform projects a function onto a wavelet basis composed of scaled and shifted versions of a mother function φ(t)\varphi(t). The transformed function depends on the scale (α\alpha) and shift (β\beta) parameters respectively linked to frequency and time information, whereas the Fourier transform is a function of only frequency.

We propose using a wavelet transform in time while keeping the Fourier transform in homogeneous directions. We denote the total transformation operator (wavelet in time and Fourier in homogeneous directions) as WW and the left inverse operator as W−1W^{-1}, which is also the right inverse for unitary transforms. The inverse operator is well-defined and unique for orthogonal wavelet bases. We can then apply WW on the left in (4), which gives

(WDtW−1)(Wui)+(WUjDjW−1)(Wui)+(WdUi,jW−1)(Wuj)=−1ρ(WDiW−1)(Wp)+ν(WLW−1)(Wui)+Wfi,(WDiW−1)(Wui)=0,(8)(WD_tW^{-1})(Wu_i)+(WU_jD_jW^{-1})(Wu_i)+(WdU_{i,j}W^{-1})(Wu_j) = -\frac{1}{\rho}(WD_iW^{-1})(Wp)+\nu(WLW^{-1})(Wu_i)+Wf_i,\qquad(W D_i W^{-1})(W u_i)=0, \tag*{(8)}

or

D~tu~i+U~jD~ju~i+dU~i,ju~j=−1ρD~ip~+νL~u~i+f~i,D~iu~i=0.(9)\widetilde{D}_t\tilde{u}_i+\widetilde{U}_j\widetilde{D}_j\tilde{u}_i+\widetilde{dU}_{i,j}\tilde{u}_j=-\frac{1}{\rho}\widetilde{D}_i\tilde{p}+\nu\widetilde{L}\tilde{u}_i+\tilde{f}_i,\qquad\widetilde{D}_i\tilde{u}_i=0. \tag*{(9)}

Note that for an arbitrary matrix MM and vector bb, M~:=WMW−1\widetilde{M}:=WMW^{-1} and b~:=Wb\widetilde{b}:=Wb. These equations can be separated for each spatial wavenumber in the homogeneous direction, and thus the dimension of each linear equation is smaller than the full Navier-Stokes equations. Note that if the transformation in time is given by the Fourier transform rather than a wavelet transform, this would recover the traditional Fourier-based resolvent analysis [2] (if the flow is temporally stationary) or the harmonic resolvent [13] analysis (if the flow is periodic in time). Similar to the Fourier-based resolvent analysis, for flows that are homogeneous in the x1x_1 and x3x_3 directions, this can be written in matrix form as

\begin{bmatrix} \tilde{u}_1(k_1,k_3)\\ \tilde{u}_2(k_1,k_3)\\ \tilde{u}_3(k_1,k_3)\\ \tilde{p}(k_1,k_3) \end{bmatrix} = \widetilde{\mathcal{H}}(k_1,k_3) \begin{bmatrix} \tilde{f}_1(k_1,k_3)\\ \tilde{f}_2(k_1,k_3)\\ \tilde{f}_3(k_1,k_3)\\ 0 \end{bmatrix}, \tag*{(10)}\end{bmatrix}

where the wavelet-based resolvent operator H~∈C4NtNx2×C4NtNx2\widetilde{\mathcal{H}}\in\mathbb{C}^{4N_tN_{x_2}\times\mathbb{C}^{4N_tN_{x_2}}} is defined as

H~=[(D~t−νL~+U~iD~i)(1000010000100000)+(dU~1,1dU~1,2dU~1,31ρD~1dU~2,1dU~2,2dU~2,31ρD~2dU~3,1dU~3,2dU~3,31ρD~3D~1D~2D~30)]−1.(11)\widetilde{\mathcal{H}}= \left[ \left(\widetilde{D}_t-\nu\widetilde{L}+\widetilde{U}_i\widetilde{D}_i\right) \begin{pmatrix} 1&0&0&0\\ 0&1&0&0\\ 0&0&1&0\\ 0&0&0&0 \end{pmatrix} + \begin{pmatrix} \widetilde{dU}_{1,1}&\widetilde{dU}_{1,2}&\widetilde{dU}_{1,3}&\frac{1}{\rho}\widetilde{D}_1\\ \widetilde{dU}_{2,1}&\widetilde{dU}_{2,2}&\widetilde{dU}_{2,3}&\frac{1}{\rho}\widetilde{D}_2\\ \widetilde{dU}_{3,1}&\widetilde{dU}_{3,2}&\widetilde{dU}_{3,3}&\frac{1}{\rho}\widetilde{D}_3\\ \widetilde{D}_1&\widetilde{D}_2&\widetilde{D}_3&0 \end{pmatrix} \right]^{-1}. \tag*{(11)}

This formulation allows us to study transient flows using resolvent analysis.

Wavelet-based resolvent analysis with windowing

We can reformulate a resolvent map between forcing and response at specific time shifts and scales by defining a windowed resolvent operator

\begin{bmatrix} \tilde{u}_1(k_1,k_3) \\ \tilde{u}_2(k_1,k_3) \\ \tilde{u}_3(k_1,k_3) \\ \tilde{p}(k_1,k_3) \end{bmatrix} = \mathcal{C}\widetilde{\mathcal{H}}(k_1,k_3)\mathcal{B} \begin{bmatrix} \tilde{f}_1(k_1,k_3) \\ \tilde{f}_2(k_1,k_3) \\ \tilde{f}_3(k_1,k_3) \\ 0 \end{bmatrix}, \tag*{(12)}\end{bmatrix}

where B\mathcal{B} and C\mathcal{C} are windowing matrices on the forcing and response modes, respectively [22, 23]. The windowing matrices select a subset of the full forcing and response states. For example, to select a particular scale and shift parameter (αs,βs)(\alpha_s,\beta_s) for the forcing mode, we set

B=diag⁡(1(α=αs)1(β=βs)),(13)\mathcal{B}=\operatorname{diag}\left(\mathbb{1}(\alpha=\alpha_s)\mathbb{1}(\beta=\beta_s)\right), \tag*{(13)}

where 1(⋅)\mathbb{1}(\cdot) is an indicator function. The SVD of the windowed resolvent operator, CH~(k1,k3)B\mathcal{C}\widetilde{\mathcal{H}}(k_1,k_3)\mathcal{B}, allows us to identify forcing and response modes restricted to a limited frequency and time interval.

The windowed resolvent operator, however, is equivalent to taking the Moore-Penrose pseudo-inverse of B†H~(k1,k3)−1C†\mathcal{B}^{\dagger}\widetilde{\mathcal{H}}(k_1,k_3)^{-1}\mathcal{C}^{\dagger}, where the superscript †\dagger indicates the pseudo-inverse. In this case, B†=B\mathcal{B}^{\dagger}=\mathcal{B} and C†=C\mathcal{C}^{\dagger}=\mathcal{C}. By applying the pseudo-inverse to the linearized Navier-Stokes operator prior to inverting it to compute H~(k1,k3)\widetilde{\mathcal{H}}(k_1,k_3), we can reduce the computational cost. The matrix inversion and SVD will apply to a matrix with its size defined by the nonzero block of BC\mathcal{B}\mathcal{C} rather than the full system.

Choice of wavelet basis

Wavelet transforms are not unique and are defined by the choice of the mother wavelet φ(t)\varphi(t). Consider an arbitrary function ff in L2(R)L_2(\mathbb{R}). Its Fourier and wavelet transforms are

f^(ω)=∫−∞+∞f(t)e−iωt dt,(14)\hat{f}(\omega)=\int_{-\infty}^{+\infty}f(t)e^{-\mathrm{i}\omega t}\,\mathrm{d}t, \tag*{(14)}
f~(α,β)=∫−∞+∞f(t)φα,β∗(t) dt,φα,β(t)≡1αφ(t−βα).(15)\tilde{f}(\alpha,\beta)=\int_{-\infty}^{+\infty}f(t)\varphi_{\alpha,\beta}^{*}(t)\,\mathrm{d}t,\qquad\varphi_{\alpha,\beta}(t)\equiv\frac{1}{\sqrt{\alpha}}\varphi\left(\frac{t-\beta}{\alpha}\right). \tag*{(15)}

Note that there is a trade-off between precision in frequency and precision in time, i.e., one cannot find a function φ(t)\varphi(t) that is well localized in both time and frequency [24]. As two extreme examples, consider the Dirac delta centered at t=1t=1, which is perfectly localized in space but with an infinite spread in frequency space, and the Fourier mode eite^{\mathrm{i}t}, i=−1i=\sqrt{-1}, which is perfectly localized in frequency space at ω=1\omega=1 but has infinite spread in time.

For this study, we choose the Shannon wavelet for its frequency properties. Each Shannon wavelet acts as a perfect band-pass filter and covers a band 2−i/T([−2π,−π]∪[π,2π])2^{-i}/T\left([-2\pi,-\pi]\cup[\pi,2\pi]\right), i∈Zi\in\mathbb{Z} (Figure 1), where TT denotes the time horizon considered. Though the Shannon wavelet does not have perfect frequency localization provided by the Fourier transform, it allows the separation of the frequency content into distinct non-overlapping bands for different scales. In addition, the discrete transform matrix WW [24, 25] for the Shannon wavelet is an orthonormal basis for L2(R)L_2(\mathbb{R}), making WW unitary [25]. However, we note that the results in the current study do not change significantly with different choices of the wavelet basis.

Shannon wavelet in time and frequency domains

Figure 1. Shannon wavelet in (a) time and (b) frequency domains.

Computational cost

The construction of H~\widetilde{\mathcal{H}} requires the inversion of a 4NyNt×4NyNt4N_yN_t\times4N_yN_t matrix, a computation that costs O(64Ny3Nt3)O(64N_y^3N_t^3) operations. The full SVD of H~\widetilde{\mathcal{H}} would also require O(64Ny3Nt3)O(64N_y^3N_t^3) operations. With a direct solve, the wavelet-based resolvent analysis would cost O(Nt2)O(N_t^2)-times more than performing NtN_t separate Fourier-based resolvent for each temporal scale, though the latter would fail to capture the interactions between the different time scales. This penalty of O(Nt2)O(N_t^2) is the nominal cost of constructing time-localized resolvent modes.

One method for reducing the memory and computational cost of wavelet-based resolvent analysis is to use sparse finite difference operators and wavelet transforms when constructing H~−1\widetilde{\mathcal{H}}^{-1}. The resulting sparse matrix can be factored with specialized packages like PARDISO [26, 27, 28], which we utilize in the current study, in order to efficiently solve linear equations of the form H~−1v=w\widetilde{\mathcal{H}}^{-1}v=w, where vv and ww are arbitrary vectors. To further take advantage of the sparsity of H~\widetilde{\mathcal{H}}, we opt for an iterative method to perform the SVD. In this work, we use a one-sided Lanczos bidiagonalization [29], an iterative algorithm that allows us to compute a truncated SVD and accurately estimate a number q<4NyNtq<4N_yN_t of the most significant singular input and output modes. Other efficient SVD algorithms rely on randomized approaches, in particular by sub-sampling the high-dimensional matrix and performing the SVD on the lower-dimensional approximation [30, 31, 32]. A randomized SVD of a high-dimensional discrete resolvent operator was used in [33, 34].

In some cases such as §III.B.2, where only some wavelet scales or time shifts are relevant, the windowing matrices CC and B\mathcal{B} can be chosen to select the significant regions of the time-frequency domain. We then neglect the zero rows and columns in CH~BC\widetilde{\mathcal{H}}\mathcal{B} associated with non-significant wavelet scales and shifts and perform the SVD and pseudo-inversion on the lower-dimensional system.

Application to statistically-stationary flow

We first validate the wavelet-based resolvent analysis on a statistically-stationary turbulent channel flow. We have shown that for unitary wavelet transforms the Fourier-based and wavelet-based resolvent modes are equivalent. Thus, for the channel flow case, we expect the two methods to produce identical resolvent modes. We then utilize the temporally-local property of wavelet-based resolvent analysis to study transient growth in a turbulent channel flow.

Turbulent channel flow

We first compare the results of both the traditional Fourier-based resolvent analysis and wavelet-based resolvent analysis. For this, we consider a turbulent channel flow. The mean profile of a turbulent channel flow at friction Reynolds number Reτ=186Re_\tau=186 is obtained by a direct numerical simulation (DNS) using a second-order staggered finite-difference [35] and a fractional-step method [36] with a third-order Runge–Kutta time-advancing scheme [37]. Periodic boundary conditions are imposed in the streamwise and spanwise directions and the no-slip and no-penetration boundary conditions are used at the top and bottom walls. The code has been validated in previous studies in turbulent channel flows [38, 39, 40] and flat-plate boundary layers [41]. The numerical domain is 8πδ×2δ×3πδ8\pi\delta\times2\delta\times3\pi\delta and is discretized using 768, 130, and 288 grid points in the streamwise, wall-normal, and spanwise directions, respectively. We use uniform grid spacing in the streamwise and spanwise directions of Δx1+=6\Delta x_1^+=6 and Δx3+=6\Delta x_3^+=6, and a wall-normal grid stretched away from the wall using a hyperbolic tangent with min⁡(Δx2+)=0.16\min(\Delta x_2^+)=0.16 and max⁡(Δx2+)=7.2\max(\Delta x_2^+)=7.2. Here the superscript ++ denotes wall units defined in terms of ν\nu and uτu_\tau, where uτu_\tau is the friction velocity. The simulations were run for 100 eddy turnover times (defined as δ/uτ\delta/u_\tau) after transients to compute the mean quantities.

For the resolvent analysis, the wall-normal discretization is discretized using a Chebyshev collocation method using Nx2=128N_{x_2}=128, and the mean streamwise velocity profile and its wall-normal derivative from the DNS are interpolated to the Chebyshev collocation points. We use a periodic boundary condition for the temporal domain, TT, with a temporal resolution of Nt=128N_t=128. We choose DtD_t to be a Fourier differentiation matrix, and Dx2D_{x_2} a Chebyshev differentiation matrix with a no-slip and no-penetration boundary condition at the wall. Note that these matrices are not sparse; for this case, sparse differentiation matrices were not needed, though finite difference matrices may be used in higher-dimensional problems to improve efficiency. We choose the spanwise and streamwise wavelengths of λ1+≈1000\lambda_1^+ \approx1000 and λ3+≈100\lambda_3^+ \approx100, in line with the most energetic structures close to the wall.

Since the mean profiles are statistically steady, we have U‾jDj=U^jDj\overline{U}_jD_j = \widehat{U}_jD_j, dU‾i,j=dU^i,jd\overline{U}_{i,j} = d\widehat{U}_{i,j}, L~=L^\widetilde{L} = \widehat{L}. The wavelet- and Fourier-based cases thus only differ by the time differentiation matrix such that W−1D~tW=F−1D^tF=DtW^{-1}\widetilde{D}_tW = F^{-1}\widehat{D}_tF = D_t. Note that WW is unitary for the choice of wavelet. Since the singular value decomposition is unique up to multiplication by a unitary matrix, we expect the singular values of H~\widetilde{H} to be the same as that of

H^(k1,k3)=(H^(k1,k3,ω1)H^(k1,k3,ω2)H^(k1,k3,ω3)⋱)(16)\widehat{H}(k_1,k_3) = \begin{pmatrix} \widehat{H}(k_1,k_3,\omega_1) & & \\ & \widehat{H}(k_1,k_3,\omega_2) & \\ & & \widehat{H}(k_1,k_3,\omega_3) \\ & & \ddots \end{pmatrix} \tag*{(16)}

where ωi=(2πi)/T\omega_i = (2\pi i)/T for i=−Nt/2,⋯ ,Nt/2−1i = -N_t/2,\cdots,N_t/2-1. Here, we set T=5.5δ/uτT = 5.5\delta/u_\tau. Moreover, we expect the response and forcing modes of both systems to be related by the unitary transform given by the Fourier and inverse-wavelet transform in time, FW−1FW^{-1}. In Figure 2(a), we show the singular values of the leading Fourier- and wavelet-based resolvent response modes, ψ^\widehat{\psi} and ψ~\widetilde{\psi} respectively. The Fourier-based resolvent modes were computed by applying the resolvent analysis at each ωi\omega_i and taking the principal singular value. The wavelet-based resolvent modes were computed by solving the full space-time system at once. We also plot the most amplified streamwise resolvent response mode for the two methods in Figure 2(b). As expected, the singular values and the corresponding modes are equivalent. The small deviations in the singular values for the two cases are due to the truncation errors of the Shannon wavelet basis.

Singular values and streamwise resolvent response modes

Figure 2. (a) First ten singular values for the wavelet-based resolvent (red) and the largest singular value for the Fourier-based resolvent (black) computed for each ωi\omega_i. The vertical gray lines indicate the temporal resolution in frequency space, and the vertical black lines delimit the frequency band covered by each of the chosen wavelet scales. (b) Magnitude of the Fourier-transformed wavelet-based (red) and Fourier-based (black) streamwise resolvent response mode for ω=17.14uτ/δ\omega= 17.14u_\tau/\delta.

Transient growth mechanism of turbulent channel flow

The added advantage of the wavelet-based method lies in its ability to preserve temporal localization, which lets us formulate a time-scale sensitive resolvent for transient flows even when the mean profiles are statistically stationary. For this, we use the windowed wavelet-based resolvent analysis framework from §III.B, but setting the windowing of the response modes, CC, to the identity matrix. We then calculate the resolvent modes of H~B\widetilde{H}B. The principal forcing mode will lie in the span of a subset of the complete wavelet basis, as selected by B\mathcal{B}. By picking a particular wavelet

scale and shift via B\mathcal{B}, we can restrict the forcing to a chosen frequency range, mostly localized in a particular time range.

To demonstrate this capability, we study the effect of a time-localized forcing on the buffer layer (x2+∈[15,40]x_2^+ \in[15,40]). We choose B\mathcal{B} such that the forcing modes are limited to a single wavelet with the frequency range containing the critical layer corresponding to the buffer layer, i.e., [U(x2+=15),U(x2+=40)]=[ωmin⁡,ωmax⁡]/k1[U(x_2^+ = 15), U(x_2^+ = 40)] = [\omega_{\min},\omega_{\max}]/k_1 (Figure 3). For the current case, we have ωmin⁡=12.52uτ/δ\omega_{\min} = 12.52u_\tau/\delta and ωmax⁡=17.17uτ/δ\omega_{\max} = 17.17u_\tau/\delta. We pick a temporal resolution of Nt=64N_t = 64, and a time horizon of T=5.5δ/uτT = 5.5\delta/u_\tau. We project the flow variables onto Shannon wavelets that cover frequency intervals (2−i−1Nt/T)([−2π,−π]∪[π,2π])(2^{-i-1}N_t/T)([-2\pi,-\pi]\cup[\pi,2\pi]), for i=0,1,⋯5i = 0,1,\cdots5. The windowing matrix B\mathcal{B} is chosen to restrict the forcing to the unshifted wavelet (τ=0\tau= 0, i.e., centered at t=0t = 0) that has been scaled to cover [9.14,18.28]uτ/δ[9.14,18.28]u_\tau/\delta (i.e. i=1i = 1).

Mean streamwise velocity profile for channel flow

Figure 3. Mean streamwise velocity profile for channel flow at Reτ=186Re_{\tau}=186. The buffer layer is delimited by the dashed lines; the black horizontal lines correspond to frequency bands covered by the chosen wavelets, mapped to streamwise velocities using U1=ω/k1U_1=\omega/k_1.

The resulting principal response mode is confined to the frequency band determined by the forcing, as shown in Figure 4(a), which is expected since the time scales are decoupled for statistically stationary flows. The response mode additionally varies in time; the magnitude of the mode growth as a function of time is shown in Figure 4(b). The principal response mode peaks at a time tuτ/δ=0.79tu_\tau/\delta= 0.79 before decaying. This transient growth can be explained by the non-normality of the linearized system [42]. The response modes at three different time shifts are shown in Figure 5. The modes are predominantly in the streamwise direction, forming alternating low- and high-speed streamwise streaks. Although not shown, the principal forcing mode only has a small contribution from the streamwise direction and is in the form of streamwise rolls, in line with the linear theory [43, 44, 45] and the self-sustaining process of wall turbulence [10, 46, 47, 48, 49, 50, 51].

Magnitude of the principal response mode and inverse wavelet-transformed principal response mode

Figure 4. (a) Magnitude of the principal response mode of H~\widetilde{\mathcal{H}} for channel flow at Reτ=186Re_{\tau}=186 under transient forcing, in the frequency-time plane. The forcing window is highlighted in red. (b) Magnitude of the inverse wavelet-transformed principal response mode as a function of time.

Principal response mode of the wavelet-based resolvent for the channel flow

Figure 5. Principal response mode of the wavelet-based resolvent for the channel flow at Reτ=186Re_{\tau}=186 and under transient forcing, shown at (a) tuτ/δ=0.26t u_{\tau}/\delta=0.26, (b) tuτ/δ=0.79t u_{\tau}/\delta=0.79, and (c) tuτ/δ=1.92t u_{\tau}/\delta=1.92. The arrows in the right panels are colored according to the sign of ∣ψ~2∣2+∣ψ~3∣2sgn⁡(ψ~1)\sqrt{|\widetilde{\psi}_{2}|^{2}+|\widetilde{\psi}_{3}|^{2}}\operatorname{sgn}(\widetilde{\psi}_{1}) to show the intensity of the streamwise rolls relative to the streamwise streaks.

Application to non-stationary flow

We now apply wavelet-based resolvent analysis to problems with a time-varying mean flow. In particular, we study the turbulent Stokes boundary layer and a turbulent channel flow with a sudden lateral pressure gradient. The Stokes boundary layer is a purely oscillatory flow in time, and thus, Fourier-based resolvent analysis still may be used [13]. However, in the case of the temporally-changing channel flow, the flow is truly unsteady, and a Fourier transform in time will not be applicable.

Turbulent Stokes boundary layer

The Stokes boundary layer is simulated through a channel flow with the lower and upper walls oscillating in tandem at a velocity of Uw(t)=Umax⁡cos⁡(Ωt)U_w(t) = U_{\max}\cos(\Omega t) with no imposed pressure gradient. The relevant nondimensional number is the Reynolds number ReΩ=Umax⁡δΩ/νRe_\Omega= U_{\max}\delta_\Omega/\nu, where δΩ=2ν/Ω\delta_\Omega= \sqrt{2\nu/\Omega} denotes the Stokes boundary layer thickness. For the current case, we consider ReΩ=1500Re_\Omega= 1500, which lies within the intermittently turbulent regime [52, 53, 54, 55, 56]. This problem has been well-studied numerically and experimentally in the literature [52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63]. The same numerical solver used to simulate the turbulent channel flow is used to generate the statistics, with modifications in the boundary condition to accommodate the oscillating walls. The domain size of the channel for the DNS is given by 6πδΩ×80δΩ×3πδΩ6\pi\delta_\Omega\times80\delta_\Omega\times3\pi\delta_\Omega and discretized using 64, 385 and 64 points in each direction. We compute the mean velocity profiles by averaging in homogeneous directions and phase. Figure 6 shows the mean and the root-mean-square (rms). velocity profiles at three different temporal locations. Note that U1(tΩ+π)=−U(tΩ)U_1(t\Omega+ \pi) = -U(t\Omega) and Ui,rms(tΩ+π)=Ui,rms(tΩ)U_{i,\mathrm{rms}}(t\Omega+ \pi) = U_{i,\mathrm{rms}}(t\Omega). We observe that the turbulent energy peaks near the wall at tΩ=2.65t\Omega= 2.65 and propagates away from the wall in time.

Mean streamwise velocity profile and streamwise r.m.s. velocity

Figure 6. (a) Mean streamwise velocity profile and (b) streamwise r.m.s. velocity at from tΩ=0t\Omega= 0 (blue) to tΩ=πt\Omega= \pi (red). The profiles shown are at tΩ=0t\Omega= 0, π/8\pi/8, π/4\pi/4, 3π/83\pi/8, π/2\pi/2, 5π/85\pi/8, 3π/43\pi/4, 7π/87\pi/8, π\pi.

To construct the resolvent operator, we first choose the spatial scales for the homogeneous directions. Using the DNS data, we calculate the streamwise energy spectrum at x2/δΩ=1.43x_2/\delta_\Omega= 1.43 and tΩ=2.51t\Omega= 2.51, which correspond to the peak spatio-temporal location of the U1,rmsU_{1,\mathrm{rms}}. The most energetic spanwise and spanwise scales at that location are given by k1δΩ=0.67k_1\delta_\Omega= 0.67 and k3δΩ=2.67k_3\delta_\Omega= 2.67, which we choose as the streamwise and spanwise scales for the resolvent operator. We use Nx2=140N_{x_2} = 140 and Nt=780N_t = 780 to solve the discrete system. We construct DtD_t and Dx2D_{x_2} as sparse finite difference matrices. We additionally choose the time-derivative operator DtD_t to be circulant to enforce periodicity in time. We compute the modes for the half-channel and modify Dx2D_{x_2} to enforce a no-slip and no-penetration boundary condition at the wall, and a free-slip and no-penetration boundary condition at the centerline.

The principal input and output modes corresponding to the chosen spatial scales and boundary conditions in time are shown in Figure 7(a). We observe that the principal input and output modes are synchronized with the peaks in U1,rmsU_{1,\mathrm{rms}}. The energy propagation towards the center occurs at a similar rate for the resolvent modes as for the DNS results. This suggests that the energy amplification in the Stokes boundary layer can partially be explained by the optimal linear mechanism, as for turbulent channel flow [45]. Moreover, we observe that the principal input modes precede the principal output modes with time delay, and thus a transient growth mechanism can be explained through the wavelet-based resolvent modes. This is in line with a physical interpretation of the modes in which the input modes

Streamwise velocity component of the principal resolvent input and output modes and magnitude of the principal response mode

Figure 7. (a) Streamwise velocity component of the principal resolvent input (top) and output (bottom) modes of the Stokes boundary layer. The black contour lines are U1,rmsU_{1,\mathrm{rms}} with the levels indicating 30%, 50%, 75%, 95% of its maximum value. The vertical dashed lines show the location of the amplitude peak for the input mode (tΩ=2.54t\Omega= 2.54) and output mode (tΩ=2.67t\Omega= 2.67). (b) Magnitude of the principal response mode of H~\tilde{H} for the Stokes boundary layer in the frequency-time plane.

‘cause’ the output modes. In Figure 7(b), we show the magnitude of the response modes in the frequency-time plane, which highlights that the bulk of the response is limited to a single wavelet scale and a few time shifts. Using the windowed resolvent operator to restrict our forcing and our response to the first wavelet scale, we indeed obtain nearly identical modes as in 7 while reducing the effective dimension of our resolvent operator by a factor of four.

Figure 8 shows that the peak amplification occurs at the location of zero mean streamwise velocity. This is in line with the fact that Fourier-based resolvent modes are often centered around a critical layer [2], where the critical layer x2=x2cx_2 = x_2^c is defined as U1(x2c)=cU_1(x_2^c) = c and kxc=ωk_x c = \omega. For the Stokes boundary layer with the chosen length scales, the critical layer seems to occur at U1(x2c)=0U_1(x_2^c) = 0, which corresponds to ω=0\omega= 0. For this time-periodic problem, we note that wavelet resolvent analysis should be equivalent to a harmonic resolvent analysis [13] that includes the interactions between all the resolved time scales. The wavelet resolvent modes should indeed map to the harmonic resolvent modes via an inverse wavelet transform and a Fourier transform in time. We expect that harmonic resolvent analysis will also reveal a peak amplification for waves at ω=0\omega= 0; we will conduct a harmonic resolvent analysis on this system and compare it with the wavelet-based resolvent modes in future works.

Mean streamwise velocity profiles

Figure 8. Mean streamwise velocity profiles normalized by uτ(tΩ)u_\tau(t\Omega) at (a) tΩ=2.35t\Omega= 2.35 and (b) tΩ=2.67t\Omega= 2.67, which correspond to 10% and 100% of the maximum intensity of the principal response mode. The vertical red lines show the wall-normal location of the instantaneous peak of the streamwise component of the output mode, the dashed black lines show the viscous and logarithmic layers, and the horizontal blue line marks where the velocity is zero.

Channel flow with sudden lateral pressure gradient

Finally, we study a fully-developed turbulent channel flow at Reτ=186Re_\tau= 186 that is subjected to a sudden lateral pressure gradient dP/dx3=ΠdP/dx1dP/dx_3 = \Pi dP/dx_1 at t=0t = 0 with Π=30\Pi= 30 [64, 65]. This flow, commonly referred to as a three-dimensional (3D) channel flow, has an initial transient period dominated by 3D non-equilibrium effects. Eventually, the flow will reach a new statistically steady state with the mean flow in the (dP/dx1,dP/dx3)(dP/dx_1, dP/dx_3) direction parallel to the wall. In the transient period, the tangential Reynolds stress initially decreases before increasing linearly, with depletion and increase rate that scales as Πx2/δ\Pi x_2/\delta [64].

The mean flow profiles are obtained from [64] and have nonzero streamwise and spanwise components U1U_1 and U3U_3 (Figure 9) as well as nonzero wall-normal gradients of streamwise and spanwise components dU1,2dU_{1,2} and dU3,2dU_{3,2}. The time domain of the simulation is Tuτ,0/δ=2.34T u_{\tau,0}/\delta= 2.34, where uτ,0u_{\tau,0} is the initial friction velocity. To construct the discrete resolvent operator, we use a spatial resolution of Nx2=65N_{x_2} = 65 for the half-channel, and a finite difference matrix Dx2D_{x_2} enforcing a no-slip and no-penetration boundary condition at the wall and a free-slip and no-penetration condition at the centerline. The boundary condition for the temporal finite difference operator DtD_t is chosen to enforce a Neumann-type condition, ∂t(⋅)∣t=0=∂t(⋅)∣t=T=0\partial_t(\cdot)|_{t=0} = \partial_t(\cdot)|_{t=T} = 0. To reduce the impact of the boundary condition on the modes at t=0t = 0 we extend U1U_1 and U3U_3 to the time interval tuτ,0/δ∈[−0.58,2.34]t u_{\tau,0}/\delta\in[-0.58, 2.34] and assume Ui(t≤0,y)=Ui(t=0,y)U_i(t \le0, y) = U_i(t = 0, y) and dP/dx3(t<0)=0dP/dx_3(t < 0) = 0. When the modes are plotted, we only show the original time domain tuτ,0/δ∈[0,2.34]t u_{\tau,0}/\delta\in[0, 2.34] and exclude the contribution from negative times. We use a temporal resolution of Nt=1000N_t = 1000 for the extended time frame.

Mean streamwise and spanwise velocity profiles

Figure 9. Mean (a) streamwise and (b) spanwise velocity profile from tuτ,0/δ=0t u_{\tau,0}/\delta= 0 (blue) to tuτ,0/δ=2.34t u_{\tau,0}/\delta= 2.34 (red). The times shown are tuτ,0/δ=0t u_{\tau,0}/\delta= 0, 0.58, 1.17, 1.76, 2.34. Data taken from [64].

Regarding the spatial scales for the homogeneous directions, we choose (λ1+,λ3+)=(1000,100)(\lambda_1^+, \lambda_3^+) = (1000, 100) as well as (λ1+,λ3+)=(189,1890)(\lambda_1^+, \lambda_3^+) = (189, 1890) and (λ1+,λ3+)=(297,2970)(\lambda_1^+, \lambda_3^+) = (297, 2970). Here, (⋅)+(\cdot)^+ indicates the wall scaling with respect to uτ,0u_{\tau,0}, before the lateral pressure gradient is applied. The first combination of λ1+\lambda_1^+ and λ3+\lambda_3^+ is the same as in §IV.A and corresponds to the spatial scales preferred by the near-wall streaks at Reτ=186Re_\tau= 186 prior to the lateral pressure gradient. The resolvent modes for these scales are shown in Figure 11(a,Figure 11b). The magnitude of the modes in frequency-time space is also plotted in Figure 12(a). The resolvent modes are temporally centered around t=0t = 0 and exhibit a predominant streamwise component. The modes are located in a region x2/δ<0.25x_2/\delta< 0.25, which corresponds to x2+<45x_2^+ < 45, i.e., the buffer region. Thus, at t=0t = 0, the modes capture the highly energetic near-wall streaks. The subsequent temporal decay of these modes can be explained by the changing flow conditions, notably the growth of the spanwise wall-shear stress τ3\tau_3, and consequently uτu_\tau (see Figure 10). Under these conditions, the spatial scales preferred by the near-wall streaks stretch as uτu_\tau increases and the wall-shear stress tensor rotates toward the x3x_3 direction.

Friction velocity and wall-shear stress angle as functions of time

Figure 10. (a) Friction velocity uτu_\tau and (b) wall-shear stress angle γ=tan⁡−1(τ3/τ1)\gamma= \tan^{-1}(\tau_3/\tau_1) as a function of time. Data taken from [64]. The vertical dashed lines are at tΩ=0,1.38,1.94t\Omega= 0, 1.38, 1.94, and correspond to the choice of λ1+\lambda_1^+ and λ3+\lambda_3^+ for the modes plotted in Figure 11.

Streamwise and spanwise velocity components of the resolvent output modes

Figure 11. Streamwise (left) and spanwise (right) velocity components of the resolvent output modes, for spatial scales of (a, b) λ1+=1000\lambda_1^+ = 1000, λ3+=100\lambda_3^+ = 100, (c, d) λ1+=189\lambda_1^+ = 189, λ3+=1890\lambda_3^+ = 1890, and (e, f) λ1+=297\lambda_1^+ = 297, λ3+=2970\lambda_3^+ = 2970. The vertical dashed lines mark (a, b) tΩ=0t\Omega= 0, (c, d) tΩ=1.38t\Omega= 1.38, and (e, f) tΩ=1.94t\Omega= 1.94

Image not extracted

Figure 12. Magnitude of the principal response mode in the frequency-time plane for (a) λ1+=1000\lambda_1^+ = 1000, λ3+=100\lambda_3^+ = 100, (b) λ1+=189\lambda_1^+ = 189, λ3+=1890\lambda_3^+ = 1890, and (c) λ1+=297\lambda_1^+ = 297, λ3+=2970\lambda_3^+ = 2970.

The second and third pairs of spatial scales are chosen so that the resolvent modes can capture the near-wall streaks under the new shear condition at times tΩ=1.3t\Omega= 1.3 and tΩ=1.94t\Omega= 1.94 respectively. To take into account the stronger mean shear in the spanwise direction, which increases proportionally to dP/dx3dP/dx_3 [64], the quantities in wall units must be scaled by a factor of uτ(t)/uτ,0u_\tau(t)/u_{\tau,0}. We also take into account the new orientation of the streaks by applying a rotation by the wall-shear stress angle γ(t)=tan⁡−1(τ3/τ1)\gamma(t) = \tan^{-1}(\tau_3/\tau_1), where τi\tau_i is the instantaneous wall-shear stress in the xix_i direction (see Fig.

Summary

This work expands the resolvent analysis framework to non-stationary flow problems. The resolvent operator is traditionally constructed for flow quantities that are Fourier-transformed in the homogeneous spatial directions and in time. Such a resolvent operator cannot be used to study time-localized nonlinear forcing or a time-varying mean flow. Instead, we construct a wavelet-based resolvent operator, applying a wavelet transform in time while keeping the Fourier transform for the homogeneous spatial directions. This resolvent operator, provided we use an orthonormal wavelet basis, is equivalent to the Fourier-based resolvent analysis for statistically stationary flows. Even in stationary cases, wavelet-based resolvent analysis can be modified to explore the effects of transient forcing localized to time scales of interest, such as those characterizing the buffer or logarithmic layers.

In the channel flow case, the wavelet-based resolvent modes are able to capture the transient growth expected for non-normal systems. We observe that the input modes precede the output modes, opening the possibility to study causality in turbulent flows using resolvent analysis. The wavelet-based resolvent analysis is notable in its ability to reflect the effects of a non-stationary mean flow. In the case of the Stokes oscillatory flow, the resolvent modes show increased sensitivity to forcing and perturbation amplification near the peaks of the streamwise root-mean-square velocity. The wavelet-based resolvent modes allow us to track the spatial and temporal location of the peak amplification alongside the varying mean flow. Finally, for the 3D channel flow, the resolvent modes are able to identify the effect of the varying flow conditions, mainly the increasing shear velocity and rotating wall shear stress, on the principal resolvent modes. We compute the resolvent modes using the length scales preferred by near-wall streaks for flow conditions at three different times. The resulting resolvent response modes peak around the chosen times, with the exception of the time close to the end of the temporal domain. The predominant velocity component for the resolvent modes also shifts from the streamwise component to the spanwise one, mirroring the reorientation of the mean flow. Wavelet resolvent modes reflect time-varying mean flow conditions and help locate energetic near-wall streaks in space and time, and identify their preferred spatial scales. This can shed light on the flow conditions that amplify these coherent structures. The cases considered in this work thus show the value of wavelet-based resolvent analysis as a new tool to study non-stationary turbulent flows.

Acknowledgements

The authors acknowledge support from the Air Force Office of Scientific Research under grant number FA9550-22-1-0109.

References

  1. [1]Jovanović, M. R., and Bamieh, B., “Componentwise energy amplification in channel flows,” J. Fluid Mech., Vol. 534, 2005, pp. 145–183.DOI
  2. [2]McKeon, B. J., and Sharma, A. S., “A critical-layer framework for turbulent pipe flow,” J. Fluid Mech., Vol. 658, 2010, pp. 336–382.DOI
  3. [3]Sharma, A. S., and McKeon, B. J., “On coherent structure in wall turbulence,” J. Fluid Mech., Vol. 728, 2013, pp. 196–238.DOI
  4. [4]Moarref, R., Sharma, A. S., McKeon, B. J., and Tropp, J. A., “Model-based scaling and prediction of the streamwise energy intensity in high-Reynolds number turbulent channels,” J. Fluid Mech., Vol. 734, 2013, pp. 275–316.
  5. [5]Luhar, M., Sharma, A. S., and McKeon, B. J., “Opposition control within the resolvent analysis framework,” J. Fluid Mech., Vol. 749, 2014, pp. 597–626.DOI
  6. [6]Martini, E., Cavalieri, A. V. G., Jordan, P., Towne, A., and Lesshafft, L., “Resolvent-based optimal estimation of transitional and turbulent flows,” J. Fluid Mech., Vol. 900, 2020, p. A2.DOI
  7. [7]Towne, A., Lozano-Durán, A., and Yang, X. I. A., “Resolvent-based estimation of space–time flow statistics,” J. Fluid Mech., Vol. 883, 2020, p. A17.DOI
  8. [8]Bae, H. J., Lozano-Durán, A., and McKeon, B. J., “Nonlinear mechanism of the self-sustaining process in the buffer and logarithmic layers of wall-bounded flows,” J. Fluid Mech., Vol. 914, 2021, p. A3.DOI
  9. [9]Yeh, C.-A., and Taira, K., “Resolvent-analysis-based design of airfoil separation control,” J. Fluid Mech., Vol. 867, 2019, pp. 572–610.DOI
  10. [10]Bae, H. J., Dawson, S. T. M., and McKeon, B. J., “Resolvent-based study of compressibility effects on supersonic turbulent boundary layers,” J. Fluid Mech., Vol. 883, 2020, p. A29.DOI
  11. [11]Bae, H. J., Dawson, S. T. M., and McKeon, B. J., “Studying the effect of wall cooling in supersonic boundary layer flow using resolvent analysis,” AIAA Scitech 2020 Forum, 2020, p. 0575.DOI
  12. [12]Ahmed, M. A., Bae, H. J., Thompson, A. F., and McKeon, B. J., “Resolvent analysis of stratification effects on wall-bounded shear flows,” Phys. Rev. Fluids, Vol. 6, 2021, p. 084804.DOI
  13. [13]Padovan, A., Otto, S. E., and Rowley, C. W., “Analysis of amplification mechanisms and cross-frequency interactions in nonlinear flows via the harmonic resolvent,” J. Fluid Mech., Vol. 900, 2020, p. A14.DOI
  14. [14]Meyer, Y., Wavelets and Operators: Volume 1, Cambridge University Press, 1992.
  15. [15]Meneveau, C., “Analysis of turbulence in the orthonormal wavelet representation,” J. Fluid Mech., Vol. 232, 1991, pp. 469–520.DOI
  16. [16]Lewalle, J., “Wavelet transforms of the Navier-Stokes equations and the generalized dimensions of turbulence,” Appl. Sci. Res., Vol. 51, No. 1-2, 1993, pp. 109–113.
  17. [17]Ren, J., Mao, X., and Fu, S., “Image-based flow decomposition using empirical wavelet transform,” J. Fluid Mech., Vol. 906, 2021, p. A22.DOI
  18. [18]Floryan, D., and Graham, M. D., “Discovering multiscale and self-similar structure with data-driven wavelets,” Proc. Natl. Acad. Sci. USA, Vol. 118, No. 1, 2021, p. e2021299118.DOI
  19. [19]Towne, A., Schmidt, O. T., and Colonius, T., “Spectral proper orthogonal decomposition and its relationship to dynamic mode decomposition and resolvent analysis,” J. Fluid Mech., Vol. 847, 2018, pp. 821–867.DOI
  20. [20]Abreu, L. I., Cavalieri, A. V. G., Schlatter, P., Vinuesa, R., and Henningson, D. S., “Spectral proper orthogonal decomposition and resolvent analysis of near-wall coherent structures in turbulent pipe flows,” J. Fluid Mech., Vol. 900, 2020, p. A11.DOI
  21. [21]Tissot, G., Cavalieri, A. V. G., and Mémin, E., “Stochastic linear modes in a turbulent channel flow,” J. Fluid Mech., Vol. 912, 2021, p. A51.DOI
  22. [22]Jeun, J., Nichols, J. W., and Jovanović, M. R., “Input-output analysis of high-speed axisymmetric isothermal jet noise,” Phys. Fluids, Vol. 28.4, 2016, p. 047101.DOI
  23. [23]Kojima, Y., Yeh, C., Taira, K., and Kameda, M., “Resolvent analysis on the origin of two-dimensional transonic buffet,” J. Fluid Mech., Vol. 885, 2020, p. R1.DOI
  24. [24]Mallat, S., A Wavelet Tour of Signal Processing, Academic Press, 2001.
  25. [25]Najmi, A.-H., Wavelets: A Concise Guide, The Johns Hopkins University Press, 2012.
  26. [26]Alappat, C., Basermann, A., Bishop, A. R., Fehske, H., Hager, G., Schenk, O., Thies, J., and Wellein, G., “A recursive algebraic coloring technique for hardware-efficient symmetric sparse matrix-vector multiplication,” ACM Trans. Parallel Comput., Vol. 7, No. 3, 2020, pp. 1–37.DOI
  27. [27]Bollhöfer, M., Schenk, O., Janalik, R., Hamm, S., and Gullapalli, K., “State-of-the-art sparse direct solvers,” Parallel Algorithms in Computational Science and Engineering, Springer, 2020, pp. 3–33.arxiv.org/abs/1907.05309
  28. [28]Bollhöfer, M., A., E., Scheidegger, S., and Schenk, O., “Large-Scale Sparse Inverse Covariance Matrix Estimation,” SIAM J. Sci. Comput., Vol. 41, No. 1, 2019, pp. A380–A401.DOI
  29. [29]Simon, H. D., and Zha, H., “Low-rank matrix approximation using the Lanczos bidiagonalization process with applications,” SIAM Journal on Scientific Computing, Vol. 21, No. 6, 2000, pp. 2257–2274.DOI
  30. [30]Halko, N., Martinsson, P.-G., and Tropp, J. A., “Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions,” SIAM review, Vol. 53, No. 2, 2011, pp. 217–288.arxiv.org/abs/0909.4061
  31. [31]Drineas, P., and Mahoney, M. W., “RandNLA: randomized numerical linear algebra,” Communications of the ACM, Vol. 59, No. 6, 2016, pp. 80–90.DOI
  32. [32]Tropp, J. A., Yurtsever, A., Udell, M., and Cevher, V., “Practical sketching algorithms for low-rank matrix approximation,” SIAM Journal on Matrix Analysis and Applications, Vol. 38, No. 4, 2017, pp. 1454–1485.DOI
  33. [33]Ribeiro, J. H. M., Yeh, C.-A., and Taira, K., “Randomized resolvent analysis,” Phys. Rev. Fluids, Vol. 5, No. 3, 2020, p. 033902.DOI
  34. [34]Yeh, C.-A., Benton, S. I., Taira, K., and Garmann, D. J., “Resolvent analysis of an airfoil laminar separation bubble at Re= 500 000,” Physical Review Fluids, Vol. 5, No. 8, 2020, p. 083906.arxiv.org/abs/2007.16197
  35. [35]Orlandi, P., Fluid Flow Phenomena: A Numerical Toolkit, Fluid Flow Phenomena: A Numerical Toolkit, Springer, 2000.DOI
  36. [36]Kim, J., and Moin, P., “Application of a fractional-step method to incompressible Navier-Stokes equations,” J. Comp. Phys., Vol. 59, 1985, pp. 308–323.DOI
  37. [37]Wray, A. A., “Minimal-storage time advancement schemes for spectral methods,” Tech. rep., NASA Ames Research Center, 1990.
  38. [38]Bae, H. J., Lozano-Durán, A., Bose, S. T., and Moin, P., “Turbulence intensities in large-eddy simulation of wall-bounded flows,” Phys. Rev. Fluids, Vol. 3, 2018, p. 014610.DOI
  39. [39]Bae, H. J., Lozano-Durán, A., Bose, S. T., and Moin, P., “Dynamic slip wall model for large-eddy simulation,” J. Fluid Mech., Vol. 859, 2019, pp. 400–432.DOI
  40. [40]Lozano-Durán, A., and Bae, H. J., “Characteristic scales of Townsend’s wall-attached eddies.” J. Fluid Mech., Vol. 868, 2019, p. 698.DOI
  41. [41]Lozano-Durán, A., Hack, M. J. P., and Moin, P., “Modeling boundary-layer transition in direct and large-eddy simulations using parabolized stability equations,” Phys. Rev. Fluids, Vol. 3, 2018, p. 023901.DOI
  42. [42]Schmid, P. J., and Henningson, D. S., Stability and Transition in Shear Flows, Vol. 142, Springer Science & Business Media, 2000.DOI
  43. [43]Orr, W., “The stability or instability of the steady motions of a perfect liquid and of a viscous liquid. Part I: A perfect liquid,” Proc. R. Ir. Acad., Vol. A 27, 1907, pp. 9–68.
  44. [44]Landahl, M. T., “Wave breakdown and turbulence,” SIAM J. Appl. Math, Vol. 28, No. 4, 1975, pp. 735–756.
  45. [45]Jiménez, J., “How linear is wall-bounded turbulence?” J. Fluid Mech., Vol. 25, 2013, p. 110814.DOI
  46. [46]Jiménez, J., and Moin, P., “The minimal flow unit in near-wall turbulence,” J. Fluid Mech., Vol. 225, 1991, pp. 213–240.
  47. [47]Hamilton, J. M., Kim, J., and Waleffe, F., “Regeneration mechanisms of near-wall turbulence structures,” J. Fluid Mech., Vol. 287, 1995, pp. 317–348.DOI
  48. [48]Jiménez, J., and Pinelli, A., “The autonomous cycle of near-wall turbulence,” J. Fluid Mech., Vol. 389, 1999, pp. 335–359.DOI
  49. [49]Waleffe, F., “On a self-sustaining process in shear flows,” Phys. Fluids, Vol. 9, No. 4, 1997, pp. 883–900.DOI
  50. [50]Schoppa, W., and Hussain, F., “Coherent structure generation in near-wall turbulence,” J. Fluid Mech., Vol. 453, 2002, pp. 57–108.DOI
  51. [51]Farrell, B. F., Gayme, D. F., and Ioannou, P. J., “A statistical state dynamics approach to wall turbulence,” Phil. Trans. R. Soc. Lond. A, Vol. 375, No. 2089, 2017, p. 20160081.DOI
  52. [52]Hino, M., Sawamoto, M., and Takasu, S., “Experiments on transition to turbulence in an oscillatory pipe flow,” J. Fluid Mech., Vol. 75, No. 2, 1976, pp. 193–207.DOI
  53. [53]Akhavan, R., Kamm, R. D., and Shapiro, A. H., “An investigation of transition to turbulence in bounded oscillatory Stokes flows Part 1. Experiments,” J. Fluid Mech., Vol. 225, 1991, pp. 395–422.DOI
  54. [54]Verzicco, R., and Vittori, G., “Direct simulation of transition in Stokes boundary layers,” Phys. Fluids, Vol. 8, No. 6, 1996, pp. 1341–1343.
  55. [55]Vittori, G., and Verzicco, R., “Direct simulation of transition in an oscillatory boundary layer,” J. Fluid Mech., Vol. 371, 1998, pp. 207–232.DOI
  56. [56]Costamagna, P., Vittori, G., and Blondeaux, P., “Coherent structures in oscillatory boundary layers,” J. Fluid Mech., Vol. 474, 2003, pp. 1–33.
  57. [57]Spalart, P. R., and Baldwin, B. S., “Direct simulation of a turbulent oscillating boundary layer,” Turbulent shear flows 6, Springer, 1989, pp. 417–440.
  58. [58]Jensen, B. L., Sumer, B. M., and Fredsøe, J., “Turbulent oscillatory boundary layers at high Reynolds numbers,” J. Fluid Mech., Vol. 206, 1989, pp. 265–297.DOI
  59. [59]Von Kerczek, C., and Davis, S. H., “Linear stability theory of oscillatory Stokes layers,” Journal of Fluid Mechanics, Vol. 62, No. 4, 1974, pp. 753–773.DOI
  60. [60]Sarpkaya, T., “Coherent structures in oscillatory boundary layers,” Journal of Fluid Mechanics, Vol. 253, 1993, pp. 105–140.
  61. [61]Blondeaux, P., and Vittori, G., “Wall imperfections as a triggering mechanism for Stokes-layer transition,” Journal of Fluid Mechanics, Vol. 264, 1994, pp. 107–135.DOI
  62. [62]Carstensen, S., Sumer, B. M., and Fredsøe, J., “Coherent structures in wave boundary layers. Part 1. Oscillatory motion,” Journal of Fluid Mechanics, Vol. 646, 2010, pp. 169–206.DOI
  63. [63]Ozdemir, C. E., Hsu, T.-J., and Balachandar, S., “Direct numerical simulations of transition and turbulence in smooth-walled Stokes boundary layer,” Physics of Fluids, Vol. 26, No. 4, 2014, p. 045108.DOI
  64. [64]Lozano-Durán, A., Giometto, M. G., Park, G. I., and Moin, P., “Non-equilibrium three-dimensional boundary layers at moderate Reynolds numbers,” J. Fluid Mech., Vol. 883, 2021, p. A20.
  65. [65]Moin, P., Shih, T.-H., Driver, D. M., and Mansour, N. N., “Direct numerical simulation of a three-dimensional turbulent boundary layer,” Phys. Fluids, Vol. 2, No. 10, 1990, pp. 1846–1853.

Paper details

vlnc:W_22qvvaqu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents