betaFeedback

WorksPhysics and Astronomy

Phase transition to turbulence via moving fronts

Abstract

Directed percolation (DP), a universality class of continuous phase transitions, has recently been established as a possible route to turbulence in subcritical wall-bounded flows. In canonical straight pipe or planar flows, the transition occurs via discrete large-scale turbulent structures, known as puffs in pipe flow or bands in planar flows, which either self-replicate or laminarize. However, these processes might not be universal to all subcritical shear flows. Here, we design a numerical experiment that eliminates discrete structures in plane Couette flow and show that it follows a different, simpler transition scenario: turbulence proliferates via expanding fronts and decays via spontaneous creation of laminar zones. We map this phase transition onto a stochastic one-variable system. The level of turbulent fluctuations dictates whether moving-front transition is discontinuous, or continuous and within the DP universality class, with profound implications for other hydrodynamic systems.

Landau and Liftshitz [1] describe the subcritical transition to turbulence in shear flows such as pipe and channels as a first-order transition, driven by the competition between two states, one laminar and one turbulent. The stable state invades the metastable state via moving fronts. Below a critical Reynolds number RecRe_c, laminar flow is the stable asymptotic state. Above RecRe_c, turbulence is stable and will expand into laminar flow. Pomeau [2] recognized that the problem of subcritical transition is richer than this, since turbulence is a fluctuating state and can spontaneously decay to the laminar state, forming laminar gaps within turbulence, while laminar flow is an absorbing state that cannot spontaneously become turbulent. The subcritical transition to turbulence is thus an absorbing state transition that could be second-order and belong to a universality class of non-equilibrium statistical systems known as directed percolation (DP) [3, 4].

The apparent simplicity of this story is belied by the multiple-scale structure of transitional turbulence typical in canonical wall-bounded shear flows. At the scales of the wall separation, turbulence consists of streamwise vortices and streaks [5]. At scales an order of magnitude larger, vortices and streaks organize into discrete coherent structures, known as puffs in straight pipe flow or oblique bands in planar flows [6, 7, 8, 9]; see Figure 1(a). To date, all experimental and numerical studies confirming universal DP scaling in wall-bounded flows have shown that bands or puffs control the percolation process [10, 11, 12, 13, 14, 15]. This is at odds with the simpler scenario proposed by Pomeau based on front motion, metastability, and laminar gap formation.

Image not extracted

Figure 1. Visualizations of (a) PCF and (b) band-free PCF at Re=360Re = 360 in a domain of size (Lstrm,Lspan)=(400,200)(L_{\mathrm{strm}}, L_{\mathrm{span}}) = (400, 200). Colors show the turbulent fluctuations, visualized via the wall-normal velocity uyu_y at y=0y = 0 (blue: −0.2≤uy<0-0.2 \le u_y < 0, red: 0<uy≤0.20 < u_y \le0.2). White (uy≃0u_y \simeq0) signifies quiescent flow. Arrows show the large-scale flow (ustrm,uspanu_{\mathrm{strm}}, u_{\mathrm{span}}). The enlargements of uyu_y in (a,b) show the streakiness in the turbulent zones.

Turbulent structures are symbiotically linked to large-scale mean flows [16, 17, 18, 19, 20, 21, 22, 23] which dictate both their characteristic size and interactions [24, 25, 26, 27]. Large-scale flow energizes turbulent structures [23, 26, 28] and must be accounted for in theoretical treatments [29, 30, 31, 32, 33]. Here, we realize a numerical experiment in which the large-scale flow in plane Couette flow (PCF) is controlled to eliminate the formation of oblique turbulent bands. The resulting flow will be called band-free PCF. With this setup, we investigate the transition to turbulence in a hydrodynamic system without discrete large-scale structures and their associated mean-turbulent coupling.

We use the pseudo-spectral code Channel-flow [34] to carry out direct numerical simulations (DNS) of the three-dimensional (3D) Navier-Stokes equations governing an incompressible viscous fluid between two parallel rigid plates moving at speeds ±Uwall\pm U_{\mathrm{wall}}. Velocities are non-dimensionalized by UwallU_{\mathrm{wall}}, lengths by the half-gap hh between the plates. The Reynolds number is Re=hUwall/νRe = hU_{\mathrm{wall}}/\nu, where ν\nu is the kinematic viscosity. Figure 1(a) shows typical transitional flow containing large-scale oblique turbulent bands. The large-scale flow (arrows) is most pronounced at the interfaces separating turbulent and laminar regions [7, 18, 19].

The large-scale spanwise velocity is negligible for either fully laminar or fully turbulent flow, but appreciable along turbulent bands. By suppressing it, we seek to eliminate band formation. We introduce streamwise (strm) and spanwise (span) large-scale cut-off wavenumbers (Kstrm,Kspan)(K_{\mathrm{strm}}, K_{\mathrm{span}}). For Fourier modes ∣kstrm∣≤Kstrm|k_{\mathrm{strm}}| \leq K_{\mathrm{strm}} and ∣kspan∣≤Kspan|k_{\mathrm{span}}| \leq K_{\mathrm{span}}, we set the spanwise velocity to zero while retaining the usual momentum equations for streamwise and wall-normal velocity, including incompressibility. This yields a two-component, three-dimensional (2C-3D) hydrodynamic system at large scales. We apply the usual 3D Navier-Stokes equations at small scales, thereby preserving the mechanisms producing wall-bounded turbulence [5, 35, 36]. The numerical procedure is given in the Appendix [37].

Figure 1(b) displays the flow computed with the cut-off window (Kstrm,Kspan)=(0.24,0.47)(K_{\mathrm{strm}}, K_{\mathrm{span}}) = (0.24, 0.47), chosen so that the corresponding wavelengths (Λstrm,Λspan)≃(26,13)(\Lambda_{\mathrm{strm}}, \Lambda_{\mathrm{span}}) \simeq(26, 13) are smaller than the typical wavelengths (λstrm,λspan)≃(100,44)(\lambda_{\mathrm{strm}}, \lambda_{\mathrm{span}}) \simeq(100, 44) of the turbulent bands in PCF [7]. The large-scale flow is streamwise-oriented by construction. Our procedure has the desired effect of eliminating the turbulent bands – the laminar-turbulent interfaces do not have the well-defined angles and widths seen in Figure 1(a).

In order to study spatio-temporal dynamics, we perform simulations in a long slender domain tilted with respect to the streamwise direction by θ=24∘\theta= 24^\circ, a typical angle at which bands occur; see [7, 8]. We denote the slender direction parallel to the bands by xx and the long direction perpendicular to them by zz. This geometry has been shown to capture important features of turbulent bands while reducing their behavior to quasi-1D dynamics along zz [8, 10, 18, 38, 39]. For consistency and comparison, we use the same domain for band-free PCF, in which we suppress large-scale spanwise velocity. (See Appendix [37].)

Mean-turbulent interaction — Prior to comparing the transition scenarios in PCF and band-free PCF, we focus on the coupling between turbulence and large-scale mean flow. We decompose the velocity u=u‾+u′\mathbf{u} = \overline{\mathbf{u}} + \mathbf{u}', where averages (⋅‾)(\overline{\cdot}) are taken over xx and time periods during

which turbulent fronts are approximately stationary. The turbulent kinetic energy Eturb≡12u′⋅u′E_{\mathrm{turb}} \equiv\frac{1}{2}\mathbf{u}' \cdot\mathbf{u}' and the mean streamwise velocity u‾strm\overline{u}_{\mathrm{strm}} are plotted in Figure 2.

Image not extracted

Figure 2. Turbulent kinetic energy EturbE_{\mathrm{turb}} and mean streamwise velocity u‾strm\overline{u}_{\mathrm{strm}} in (a,b) PCF and (c,d) band-free PCF. Contour plots (a,c) show Eturb(y,z)E_{\mathrm{turb}}(y,z) and arrows show u‾strm(±0.5,z)\overline{u}_{\mathrm{strm}}(\pm0.5,z). Unlike in (a), turbulent zones in (c) do not have a selected width. Curves in (b,d) show EturbE_{\mathrm{turb}} and u‾strm\overline{u}_{\mathrm{strm}} as a function of zz at y=0.5y = 0.5. (e) Phase-plane representation of u‾strm\overline{u}_{\mathrm{strm}} and EturbE_{\mathrm{turb}} at y=0.5y = 0.5 in PCF (blue) and band-free PCF (red). In contrast to PCF, turbulence and mean-flow in band-free PCF are in phase. Abbreviations: L, laminar; T, turbulent; R, refractory; E, excited.

The significant difference between PCF and band-free PCF is in the spatial phase relation between EturbE_{\mathrm{turb}} and u‾strm\overline{u}_{\mathrm{strm}}, seen at y=0.5y = 0.5 in Figure 2(b,Figure 2d). (Reflected plots would be obtained at y=−0.5y = -0.5.) Turbulence extracts energy from the local mean shear, thereby flattening the profile and reducing u‾strm\overline{u}_{\mathrm{strm}} at y=0.5y = 0.5. For bands in PCF, advection by the large-scale flow (arrows) redistributes momentum and energy from laminar to turbulent regions [18] (see especially Fig. 6(b) in [28]). It is the competing mechanisms of mean-flow flattening by turbulence and mean flow fuelling nearby turbulence that are encrypted in the phase relation between u‾strm\overline{u}_{\mathrm{strm}} and EturbE_{\mathrm{turb}} (see Figure 2(e)). When turbulence is excited (EturbE_{\mathrm{turb}} increases, see Figure 2(b)), it is first fuelled by the large-scale flow. Hence, EturbE_{\mathrm{turb}} does not react directly to it and decreases only after a phase shift, and then vice versa in the refractory region. Excited and refractory regions are essential for sustaining localised turbulent regions and for their duplication and interactions [23, 24, 25, 30].

In band-free PCF, we find that turbulence and mean flow are not phase shifted and excited and refractory zones are absent, i.e. EturbE_{\mathrm{turb}} is a single-valued function of u‾strm\overline{u}_{\mathrm{strm}} (Figure 2(e)). This follows from the approximate zz-reflection symmetry apparent in Figure 2(c,Figure 2d). (See Supplementary Material, §I.D). This has major consequences for the way that turbulence emerges in this flow.

Transition with and without bands.—Transition scenarios in PCF and in band-free PCF are illustrated by spatio-temporal diagrams in Figure 3. Each simulation is initiated with a localized turbulent patch.

Image not extracted

Figure 3. Band and slug regimes in (a,b,c) PCF and (d,e,f) band-free PCF initiated from a turbulent patch. Colors show the spanwise velocity uspanu_{\mathrm{span}} at y=0y = 0, x=0x = 0 as a function of zz and tt (blue: −0.1≤uspan≤0-0.1 \le u_{\mathrm{span}} \le0, red: 0≤uspan≤0.10 \le u_{\mathrm{span}} \le0.1). The wide blue and red areas visible in PCF result from the large-scale flow.

The scenario for PCF is well documented in this geometry [8, 10, 27, 38]. Below Re≃350Re \simeq350, turbulent bands are localized metastable structures that survive for long times before decaying (Figure 3(a)) or proliferating by splitting (Figure 3(b)) [38]. Decay and splitting are memoryless processes, in that their associated waiting times are exponentially distributed. Above Re0≃325Re_0 \simeq325 [38], the proliferation of a single band becomes more probable than its decay, and turbulence survives, albeit in intermittent form [10, 13]. In the thermodynamic limit of large systems with many bands and long times, this intermittency gives rise to a critical point in the DP universality class at the estimated critical value Rec=328.7Re_c = 328.7 [10, 13]. (RecRe_c and Re0Re_0 are close, but distinctly different: Re0Re_0 is determined from single isolated bands while RecRe_c is influenced by interactions between bands.) For Re≳350Re \gtrsim350, e.g. Figure 3(c), turbulence proliferates via expanding fronts moving at equal and opposite speeds (slug phase [30]). The mean left-going and right-going front propagation speeds are shown in Figure 4. Below Re≃450Re \simeq450, the fronts delimit a nearly periodic interior pattern (Figure 3(c)) [38], which is absent above Re≃450Re \simeq450 (not shown).

Image not extracted

Figure 4. Front velocity as a function of ReRe in both PCF and band-free PCF. Labels (a-f) correspond to the visualizations of Figure 3(a-f). The points of zero propagation speed are Re0≃325Re_0 \simeq325 for PCF [38] and Re0≃383Re_0 \simeq383 for band-free PCF. Speeds below Re0Re_0 in band-free PCF are not shown; see Supplemental Material. The blue curve guides the eye, while the continuous red curves come from simulations of model (1).

By design, band-free PCF does not exhibit discrete turbulent structures or patterns. Turbulence contracts and expands, not via decay and splitting of discrete structures, but rather via fluctuating front motion, shown in Figure 3(d)–(Figure 3f), whose expansion speeds are also shown in Figure 4. Importantly, the decay of turbulence is not a memoryless process because the mean lifetime of a patch depends on its initial size, unlike in PCF and pipe flow [11, 38, 40], but like low-ReRe plane channel flow [41, 42].

Figure 5(a,b) shows the spatio-temporal dynamics from a fully turbulent initial state in a long domain below (Re=370Re = 370) and above (Re=385Re = 385) the critical point RecRe_c in band-free PCF. Laminar gaps nucleate within the turbulent flow and interfaces fluctuate. Below RecRe_c, turbulence predominantly contracts and asymptotically the flow is laminar. Above RecRe_c, laminar gaps are created but eventually close as a consequence of preferred turbulent expansion.

Image not extracted

Figure 5. Laminar-turbulent percolation in band-free PCF at (a) Re=370<RecRe = 370 < Re_c and (b) Re=385>RecRe = 385 > Re_c. (c) Equilibrium turbulent fraction FtF_t as a function of ϵ≡(Re−Rec)/Rec\epsilon\equiv(Re - Re_c)/Re_c in PCF (blue, data from Ref. [10]) and band-free PCF (red). The inset in (c) shows PCF data rescaled by Ft(ϵ=0.02)F_t(\epsilon= 0.02) and confirms the more abrupt transition in band-free PCF. (d, e) Space-time visualisation of model (1) with σ=2.0\sigma= 2.0 at (d) r<rc≃0.2517r < r_c \simeq0.2517 and (e) r>rcr > r_c. Time and length in the model are rescaled with T~\widetilde{T} and L~\widetilde{L}, and rr-values are chosen such that r−r0=b(Re−Re0)r - r_0 = b(Re - Re_0). (f) Same as (c) for model (1) with σ=1.0\sigma= 1.0 (first-order), 2.0 and 2.5 (both second-order cases exhibiting DP scaling ϵ0.276\epsilon^{0.276}, see dashed lines).

The order parameter for the transition is the equilibrium turbulent fraction, FtF_t, which is the mean proportion of turbulent flow at statistical equilibrium (see Supplemental Material). FtF_t is plotted as a function of ReRe in Figure 5(c) for both PCF [10] and band-free PCF. For comparison, we use reduced Reynolds numbers, ϵ≡(Re−Rec)/Rec\epsilon\equiv(Re - Re_c)/Re_c. For PCF, Rec≃328.7Re_c \simeq328.7 [10]. For band-free PCF we estimate Rec≃383.5Re_c \simeq383.5. Without bands, the transition to uniform turbulence occurs over a significantly shorter range of ReRe (Ft≃0.9F_t \simeq0.9 at ϵ≃0.01\epsilon\simeq0.01, while Ft≃0.2F_t \simeq0.2 at ϵ≃0.01\epsilon\simeq0.01 in PCF).

Measuring FtF_t near ϵ=0\epsilon= 0 is exceedingly costly because scales diverge and hence simulations become susceptible to finite-size, finite-time effects. We have not attempted to determine critical scalings associated with transition in band-free PCF.

A model for percolation via front motion.—Published models for puffs and bands contain at least two fields [29, 30, 31, 32, 33]; two are necessary to capture the interaction between mean-flow and turbulence responsible for discrete structures; see Figure 2(a,b). Band-free PCF lacks discrete structures and mean-flow and turbulence are slaved, see Figure 2(c,d). This suggests modelling band-free PCF with a single scalar field q(z,t)q(z,t) representing the local turbulent energy.

Following [2, 30, 43], we consider the stochastic model

∂tq=∂z2q−∂qV+σξ(1)\partial_t q = \partial_z^2 q - \partial_q V + \sigma\xi \tag*{(1)}

with the double-well potential

V(q)≡q22[1+(r+1)(q22−4q3)].V(q) \equiv\frac{q^2}{2}\left[1 + (r + 1)\left(\frac{q^2}{2} - \frac{4q}{3}\right)\right].

ξ\xi a space-time Gaussian white noise of unit variance and σ\sigma a parameter controlling the noise strength. Parameter rr plays the role of ReRe. VV has one local minimum at q=0q = 0 (laminar) and one at q>0q > 0 (turbulent state), the latter being the lowest minimum for r>1/8r > 1/8. Multiplicative noise σqξ\sigma q\xi represents turbulent fluctuations and vanishes at q=0q = 0, making it an absorbing state; see [44, 45, 46]. Hence model (1) describes fronts between an absorbing and a fluctuating state, just as Pomeau originally envisioned for the transition to turbulence [2, 30, 43].

We solve (1) via a finite-difference scheme under the Itô representation [47]. We first simulate expanding slugs.

Front velocities are found to closely obey the self-similar expression c≃F(r−r0(σ))c \simeq F(r-r_0(\sigma)) where F(0)=0F(0)=0 and r0(σ)≃1/8+3σ2/100r_0(\sigma)\simeq1/8+3\sigma^2/100. FF is independent of σ\sigma. Front speeds in band-free PCF, shown in Figure 4, follow cDNS≃aF(b(Re−Re0))c_{\mathrm{DNS}} \simeq aF(b(Re-Re_0)) (solid red line), where aa and bb rescale velocities and ReRe, respectively. The agreement is excellent.

The fit of the model speeds to those of band-free PCF establishes the correspondence between rr and ReRe. We determine the ratio L~≃3.0\widetilde{L}\simeq3.0 of band-free PCF to model length scales from the width of laminar-turbulent fronts. From L~\widetilde{L} and aa, we obtain a ratio of time scales between band-free PCF and model, T~=L~/a≃16\widetilde{T}=\widetilde{L}/a\simeq16 (see Supplemental Material). The final model parameter σ\sigma is chosen so as to reproduce the phenomenology of band-free PCF as we now show.

With σ=2.0\sigma= 2.0, simulations of (1) with space, time and ReRe rescaling are visualized in Figure 5(d, Figure 5e). The model closely reproduces the dynamics of fluctuating fronts and the nucleation of laminar gaps within turbulent flow as seen in band-free PCF (Figure 5(a, Figure 5b)).

The noise intensity σ\sigma indirectly controls the rate of laminar-gap nucleation, and, consequently, the order of the phase transition, as we show in Figure 5(f). With sufficiently strong noise (σ=2.0\sigma= 2.0 and 2.5), the transition is continuous and exhibits the scalings of DP (Ft∼ϵ0.276F_t \sim\epsilon^{0.276}, spatial and temporal correlations ξ⊥∼ϵ−1.097\xi_\perp\sim\epsilon^{-1.097} and ξ∥∼ϵ−1.734\xi_\parallel\sim\epsilon^{-1.734}; see Supplemental Material). Meanwhile, at low noise (σ=1.0\sigma= 1.0) the transition is discontinuous, with no intermediate 0<Ft<10 < F_t < 1 sustained. The transition becomes sharper as σ\sigma is decreased.

We find that the value σ=2.0\sigma= 2.0 captures the dependence of turbulent fraction on ReRe in band-free PCF (Figure 5(f)) with similar sharp increase in FtF_t, and specifically gives Ft(ϵ≃0.01)≃0.9F_t(\epsilon\simeq0.01)\simeq0.9. (The accuracy of this match is limited by the precision of RecRe_c.) Because this value of σ\sigma corresponds to a continuous phase transition, this strongly suggests that band-free PCF, like PCF, undergoes a continuous transition in the DP universality class.

Conclusion and discussion

In subcritical shear flows, turbulence appears, not by increasing its intensity, but by occupying an increasing proportion of space as Reynolds number is increased. In previous studies of canonical straight pipe or planar flows, transition occurs via the percolation of discrete turbulent structures that individually decay or self-replicate. Large-scale mean flow plays a crucial role in these systems by selecting these discrete structures.

Here, we investigate a plane-Couette setup without discrete structures. Transition follows a distinctly different scenario, mediated by expanding turbulent fronts and spontaneous nucleation of laminar gaps. The observed dynamics are very well captured by a simple stochastic, double-well model with a single field. The order of this phase transition is governed by the nucleation rate of laminar gaps, which depends on the level of fluctuations.

This front-moving transition should be pervasive in systems where discrete puffs or bands are absent. Numerous flows, such as bent pipes [48], body-forced pipes [49], stably-stratified flows [50], suction boundary layer [51] or constrained Couette flow [52, 53], lack discrete turbulent structures and may follow this scenario. Our Letter suggests that in such flows, the level of turbulent fluctuations plays a crucial role in dictating whether front-moving transition is first-order, as recently reported in experiments of curved and body-forced pipes [49], or second-order and within the DP universality class.

Acknowledgments

We thank Anna Frishman, Yohann Duguet, Santiago Benavides, François Pétrélis and Tobias Grafke for fruitful discussions. The calculations for this work were performed using high performance computing resources provided by the Grand Equipement National de Calcul Intensif at the Institut du Développement et des Ressources en Informatique Scientifique (IDRIS, CNRS) through Grant No. A0142A01119. This work was partly supported by a grant from the Simons Foundation (Grant No. 662985, D.B. and L.S.T.).

Supplementary Material for Phase transition to turbulence via moving fronts

Suppressing large-scale spanwise flow in PCF

Large-scale spanwise velocity is an emergent feature of bands in transitional shear flows. We have introduced a numerical procedure which eliminates band formation in plane Couette flow (PCF) by suppressing large-scale spanwise velocity. Meanwhile, we keep large-scale variations of the streamwise and wall-normal velocities, so that laminar-turbulent intermittency is preserved.

For small-scale modes, i.e. for kk such that ∣kstrm∣>Kstrm|k_{\mathrm{strm}}| > K_{\mathrm{strm}} or ∣kspan∣>Kspan|k_{\mathrm{span}}| > K_{\mathrm{span}}, we solve the usual 3D Navier-Stokes equations:

∂u^j∂t+∇^jp^=−N^j+1Re∇^2u^j for j=strm,y,span(2)\frac{\partial\hat{u}_j}{\partial t}+\hat{\nabla}_j\hat{p}=-\hat{N}_j+\frac{1}{Re}\hat{\nabla}^{2}\hat{u}_j \text{ for } j=\mathrm{strm},y,\mathrm{span} \tag*{(2)}
∇^⋅u^=0(3)\hat{\nabla}\cdot\hat{\mathbf{u}}=0 \tag*{(3)}

where N^j=∑p+q=k∂lu^l(p)u^j(q)\hat{N}_j=\sum_{p+q=k}\partial_l\hat{u}_l(p)\hat{u}_j(q) are the nonlinear advective terms.

For the large-scale modes, i.e. kk such that ∣kstrm∣≤Kstrm|k_{\mathrm{strm}}|\leq K_{\mathrm{strm}} and ∣kspan∣≤Kspan|k_{\mathrm{span}}|\leq K_{\mathrm{span}}, we solve the following system for the Fourier coefficient u^(kstrm,kspan)\hat{\mathbf{u}}(k_{\mathrm{strm}},k_{\mathrm{span}})

∂u^j∂t+∇^jp^=−N^j+1Re∇^2u^j for j=strm,y(4)\frac{\partial\hat{u}_j}{\partial t}+\hat{\nabla}_j\hat{p}=-\hat{N}_j+\frac{1}{Re}\hat{\nabla}^{2}\hat{u}_j \text{ for } j=\mathrm{strm},y \tag*{(4)}
∇^2D⋅u^=0(5)\hat{\nabla}_{2D}\cdot\hat{\mathbf{u}}=0 \tag*{(5)}
u^span=0(6)\hat{u}_{\mathrm{span}}=0 \tag*{(6)}

where ∇^2D≡(ikstrm,∂y,0)\hat{\nabla}_{2D}\equiv(ik_{\mathrm{strm}},\partial_y,0). This is a two-component, three-dimensional (2C-3D) system for these low-wavenumber modes.

Numerical setup

We solve the 3D Navier-Stokes equations using Channelflow [34], which uses a semi-implicit pseudospectral in primitive variables, decomposed in Fourier-Chebyshev modes. We briefly explain the underlying numerical principles, derived from Canuto et al. [54].

Equations (A.2) are discretized in time and in the Fourier directions as follows:

(∂y2−λ)u^strm(t)=R^strm−ikstrmp^,with u^strm(±1)=0(7)(\partial_y^2-\lambda)\hat{u}_{\mathrm{strm}}(t)=\hat{R}_{\mathrm{strm}}-ik_{\mathrm{strm}}\hat{p}, \qquad\text{with }\hat{u}_{\mathrm{strm}}(\pm1)=0 \tag*{(7)}
(∂y2−λ)u^y(t)=R^y−∂yp^,with u^y(±1)=0(8)(\partial_y^2-\lambda)\hat{u}_y(t)=\hat{R}_y-\partial_y\hat{p}, \qquad\text{with }\hat{u}_y(\pm1)=0 \tag*{(8)}
(∂y2−λ)u^span(t)=R^span−ikspanp^with u^span(±1)=0(9)(\partial_y^2-\lambda)\hat{u}_{\mathrm{span}}(t)=\hat{R}_{\mathrm{span}}-ik_{\mathrm{span}}\hat{p} \qquad\text{with }\hat{u}_{\mathrm{span}}(\pm1)=0 \tag*{(9)}
ikstrmu^strm+∂yu^y+ikspanu^span=0(10)ik_{\mathrm{strm}}\hat{u}_{\mathrm{strm}}+\partial_y\hat{u}_y+ik_{\mathrm{span}}\hat{u}_{\mathrm{span}}=0 \tag*{(10)}

The terms R^\widehat{\mathbf{R}} come from explicit integration of the non-linear term N^\widehat{\mathbf{N}} and time derivative and are given by

R^≡∑sαsΔtu^(s)−βsN^(s),(11)\widehat{\mathbf{R}} \equiv\sum_s \frac{\alpha_s}{\Delta t}\widehat{\mathbf{u}}(s)-\beta_s\widehat{\mathbf{N}}(s), \tag*{(11)}

where ss ranges over previous timesteps, Δt\Delta t is the numerical timestep, and αs\alpha_s and βs\beta_s are coefficients that depend on the time-stepping method. Equation (A.4d) is eliminated in the usual way by acting with the divergence (ikstrm,∂y,ikspan)(ik_{\mathrm{strm}},\partial_y,ik_{\mathrm{span}}) on (A.4a)–(A.4c), leading to the pressure Poisson equation, and by substituting the boundary conditions on u^strm\widehat{u}_{\mathrm{strm}}, u^span\widehat{u}_{\mathrm{span}} into (A.4d). This yields the following system for (u^y,p^)(\widehat{u}_y,\widehat{p}) for each Fourier wavenumber pair (kstrm,kspan)(k_{\mathrm{strm}},k_{\mathrm{span}}):

\begin{aligned} (\partial_y^2-k_{\mathrm{strm}}^2-k_{\mathrm{span}}^2)\widehat{p}(t)&=ik_{\mathrm{strm}}\widehat{R}_{\mathrm{strm}}+\partial_y\widehat{R}_y+ik_{\mathrm{span}}\widehat{R}_{\mathrm{span}}\\ (\partial_y^2-\lambda)\widehat{u}_y(t)&=\widehat{R}_y-\partial_y p \end{aligned} \qquad \text{with} \qquad \begin{cases} \partial_y\widehat{u}_y(\pm1)=0\\ \widehat{u}_y(\pm1)=0 \end{cases} \tag*{(12)}\end{aligned}

where λ=2/Δt+(kx2+kz2)/Re\lambda=2/\Delta t+(k_x^2+k_z^2)/Re arises from the implicit treatment of the viscous term. This system of two Helmholtz problems, coupled by their four boundary conditions on u^y\widehat{u}_y, is solved via an influence-matrix method. Equations (A.4a) and (A.4c) for the two remaining components u^strm,u^span\widehat{u}_{\mathrm{strm}},\widehat{u}_{\mathrm{span}} are uncoupled and easily solved.

We now turn to the suppression of spanwise flow in large-scale modes with ∣kstrm∣≤Kstrm|k_{\mathrm{strm}}|\le K_{\mathrm{strm}} and ∣kspan∣≤Kspan|k_{\mathrm{span}}|\le K_{\mathrm{span}}. Substituting u^span=0\widehat{u}_{\mathrm{span}}=0 for (A.4c) and carrying out the same operations as before to eliminate (A.4d) now leads to

\begin{aligned} (\partial_y^2-k_{\mathrm{strm}}^2)\widehat{p}(t)&=ik_{\mathrm{strm}}\widehat{R}_{\mathrm{strm}}+\partial_y\widehat{R}_y\\ (\partial_y^2-\lambda)\widehat{u}_y(t)&=\widehat{R}_y-\partial_y p \end{aligned} \qquad \text{with} \qquad \begin{cases} \partial_y\widehat{u}_y(\pm1)=0\\ \widehat{u}_y(\pm1)=0 \end{cases} \tag*{(13)}\end{aligned}

The influence matrix must be modified accordingly to solve system (A.7). We note that a related (though different) strategy was used by Jiménez and Pinelli [55], Jimenez [56], so as to study the cyclic mechanisms in the turbulence production in wall-bounded flows.

Simulations of band-free PCF in a tilted domain

We now define the tilted coordinate system (x,z)(x,z) via

estrm=cos⁡θ ex+sin⁡θ ez=αex+βez(14)\mathbf{e}_{\mathrm{strm}}=\cos\theta\,\mathbf{e}_x+\sin\theta\,\mathbf{e}_z=\alpha\mathbf{e}_x+\beta\mathbf{e}_z \tag*{(14)}
espan=−sin⁡θ ex+cos⁡θ ez=−βex+αez,(15)\mathbf{e}_{\mathrm{span}}=-\sin\theta\,\mathbf{e}_x+\cos\theta\,\mathbf{e}_z=-\beta\mathbf{e}_x+\alpha\mathbf{e}_z, \tag*{(15)}

illustrated in Figure 6. As with prior studies [8], the size LxL_x is chosen to Lx=10L_x=10 so that bands are homogeneous in this direction. In this geometry, we will suppress spanwise flow for modes kx=0k_x=0 and ∣kz∣<Kz|k_z|<K_z, where KzK_z denotes a cutoff wavenumber in the zz direction. Note that while the geometry is tilted with respect to the streamwise-spanwise coordinates, we still suppress spanwise flow. To set u^span=0\widehat{u}_{\mathrm{span}}=0 while preserving the evolution equations for u^strm\widehat{u}_{\mathrm{strm}} and u^y\widehat{u}_y we have

Image not extracted

Figure 6. Visualization of (a) PCF, (b) band-free PCF in a streamwise-spanwise oriented domain, and (c,d) in a quasi-1D tilted domain. (e) Spectral energy E(kz)≡12⟨u^⋅u^⟩x,tE(k_z) \equiv\frac{1}{2}\langle\widehat{\mathbf{u}}\cdot\widehat{\mathbf{u}}\rangle_{x,t}. Average is taken over (x,t)(x,t) in the tilted domain. Flow and spectra are shown at y=0y=0, Re=360Re=360. Cutoff wavenumber Kz=0.24K_z=0.24 is indicated.

(∂y2−λ)u^strm=R^strm−ikstrmp^,with u^strm(±1)=0(16)(\partial_y^2-\lambda)\widehat{u}_{\mathrm{strm}}=\widehat{R}_{\mathrm{strm}}-ik_{\mathrm{strm}}\widehat{p}, \qquad\text{with }\widehat{u}_{\mathrm{strm}}(\pm1)=0 \tag*{(16)}
(∂y2−λ)u^y=R^y−∂yp^,with u^y(±1)=0(17)(\partial_y^2-\lambda)\widehat{u}_y=\widehat{R}_y-\partial_y\widehat{p}, \qquad\text{with }\widehat{u}_y(\pm1)=0 \tag*{(17)}
u^span=0with u^span(±1)=0(18)\widehat{u}_{\mathrm{span}}=0 \qquad\text{with }\widehat{u}_{\mathrm{span}}(\pm1)=0 \tag*{(18)}
ikstrmu^strm+∂yu^y=0(19)ik_{\mathrm{strm}}\widehat{u}_{\mathrm{strm}}+\partial_y\widehat{u}_y=0 \tag*{(19)}

where kstrm=αkx+βkzk_{\mathrm{strm}}=\alpha k_x+\beta k_z and R^strm=αR^x+βR^z\widehat{R}_{\mathrm{strm}}=\alpha\widehat{R}_x+\beta\widehat{R}_z. Taking the 2D divergence (ikstrm,∂y)(ik_{\mathrm{strm}},\partial_y) of (A.9a)–(A.9b) and substituting (A.9d) leads to

\begin{aligned} (\partial_y^2-k_{\mathrm{strm}}^2)\widehat{p}&=ik_{\mathrm{strm}}\widehat{R}_{\mathrm{strm}}+\partial_y\widehat{R}_y\\ (\partial_y^2-\lambda)\widehat{u}_y&=\widehat{R}_y-\partial_y\widehat{p} \end{aligned} \qquad \text{with} \qquad \begin{cases} \partial_y\widehat{u}_y(\pm1)=0\\ \widehat{u}_y(\pm1)=0 \end{cases} \tag*{(20)}\end{aligned}

which is identical to the non-tilted case (A.7). For the remaining components, we use:

0=u^span=−βux+αuz⟹u^z=βu^xα(21)0=\widehat{u}_{\mathrm{span}}=-\beta u_x+\alpha u_z\Longrightarrow\widehat{u}_z=\frac{\beta\widehat{u}_x}{\alpha} \tag*{(21)}
u^strm=αu^x+βu^z=αu^x+β2u^xα=α2+β2αu^x=u^xα(22)\widehat{u}_{\mathrm{strm}}=\alpha\widehat{u}_x+\beta\widehat{u}_z=\alpha\widehat{u}_x+\frac{\beta^2\widehat{u}_x}{\alpha}=\frac{\alpha^2+\beta^2}{\alpha}\widehat{u}_x=\frac{\widehat{u}_x}{\alpha} \tag*{(22)}

leading to the following equations for u^x\widehat{u}_x and u^z\widehat{u}_z:

(∂y2−λ)1αu^x=(αR^x+βR^z)−i(αkx+βkz)p^,with u^x(±1)=0(23)(\partial_y^2-\lambda)\frac{1}{\alpha}\widehat{u}_x=(\alpha\widehat{R}_x+\beta\widehat{R}_z)-i(\alpha k_x+\beta k_z)\widehat{p}, \qquad\text{with }\widehat{u}_x(\pm1)=0 \tag*{(23)}
u^z=βu^xαwith u^z(±1)=0(24)\widehat{u}_z=\frac{\beta\widehat{u}_x}{\alpha} \qquad\text{with }\widehat{u}_z(\pm1)=0 \tag*{(24)}

Spectra in PCF and band-free PCF and choice of cutoff

One can view the suppression of the large-scale spanwise velocity as an effective large-scale forcing on the spanwise momentum equation, that counteracts the effects of the nonlinear term NspanN_{\mathrm{span}} and the pressure ∂spanp\partial_{\mathrm{span}}p. As a consequence, our procedure does not preserve continuity in Fourier space of the velocity components or of N\mathbf{N}.

Figure 6(e) shows the energy spectrum in a tilted domain of both PCF and band-free PCF. We compute E(kz)≡12⟨u^⋅u^⟩x,tE(k_z) \equiv\frac{1}{2}\langle\widehat{\mathbf{u}}\cdot\widehat{\mathbf{u}}\rangle_{x,t}, with u^\widehat{\mathbf{u}} the zz-Fourier transform of total velocity u\mathbf{u}, and where the average is carried out over xx and tt, at Re=360Re=360. The characteristic large-scale motions in PCF associated with the bands (kbands≃0.14k_{\mathrm{bands}}\simeq0.14) are damped to an almost constant energy spectrum in the band-free case. As expected, the energy spectrum shows a small discontinuity at the cutoff wavenumber KzK_z (here Kz=0.24K_z=0.24).

Furthermore, the energy at small scales is increased as a result of band suppression, as can be seen in an increase in the small-scale energy peak corresponding to streaks and rolls [28]. The band-suppression method that we use is indeed not designed to conserve the total dissipation in the flow, nor the dissipation at small scales; hence the energy that originally feeds bands is redistributed to the streaks in an uncontrolled way.

We chose a cutoff wavenumber KzK_z within the range 0.20<Kz<0.80.20<K_z<0.8. If KzK_z is chosen below 0.20, it is too close to kbands≃0.14k_{\mathrm{bands}}\simeq0.14 and the flow still sustains discrete bands with selected size close to 2π/Kz2\pi/K_z. For a range of KzK_z above 0.20 discrete structures are absent and we observe that transition takes place via moving fronts. However, the value of KzK_z controls the dissipation within the flow, hence the value of critical Reynolds number RecRe_c: the higher KzK_z, the more streaks and rolls are energized, hence the lower RecRe_c. We have not explored Kz>0.8K_z>0.8, as KzK_z would be close to the spacing of streaks and rolls (kz∼1k_z\sim1), and could affect the self-sustaining process. In the main text, we show results at Kz=0.24K_z=0.24. Note that in the untilted, streamwise-spanwise configuration the choice of cutoff wavenumbers (Kx,Kz)(K_x,K_z) has similar effects.

Mean flow and symmetries in PCF and band-free PCF

In the main paper we discuss the absence of excited and refractory regions for band-free PCF, and how this is related to the absence of discrete structures (bands). Here we provide further details by presenting in Figure 7 the mean flow in PCF (a,c) and band-free PCF (b,d). The mean is carried out over time and xx, for an isolated stationary band in PCF at Re=340Re = 340, and for an approximately stationary localized turbulent patch in band-free PCF at Re=360Re = 360.

Image not extracted

Figure 7. Mean flow in (a,c) PCF and (b,d) band-free PCF. (a) and (b) show u‾x−Ub,x\overline{u}_x-U_{b,x} as colors and (u‾z,u‾y)(\overline{u}_z,\overline{u}_y) streamlines in the (y,z)(y,z) plane. (c) and (d) show u‾span\overline{u}_{\mathrm{span}} as colors and profiles profiles of u‾strm(y)\overline{u}_{\mathrm{strm}}(y) in the (streamwise-yy) plane. Additionally, (d) shows streamlines associated to the two-component mean flow (u‾strm,u‾y)(\overline{u}_{\mathrm{strm}},\overline{u}_y) for band-free PCF.

Figure 7(a,Figure 7b) shows the mean flow in the tilted (z,y)(z,y) plane. Grey lines are streamlines of (u‾z,u‾y)(\overline{u}_z,\overline{u}_y) and colors depict u‾x−Ub,x\overline{u}_x-U_{b,x}, where UbU_b is the laminar flow profile. Turbulence has been centered about z=0z = 0. In Figure 7(c,Figure 7d), the mean flow is visualized in the non-tilted (strm,y)(\mathrm{strm},y) plane. Here, colors represent the spanwise component u‾span\overline{u}_{\mathrm{span}} and turbulence has been centered about strm=0\mathrm{strm} = 0. Profiles u‾strm(y)\overline{u}_{\mathrm{strm}}(y) are shown as arrows. In PCF, one cannot build a streamfunction in the (strm,y)(\mathrm{strm},y) plane since the spanwise mean flow varies as a function of the spanwise coordinate. However, in band-free PCF, u‾span=0\overline{u}_{\mathrm{span}} = 0, hence it is possible to associate the mean flow in the (strm,y)(\mathrm{strm},y) plane to a streamfunction. The streamlines for band-free PCF, which correspond to the complete mean flow, are shown in Figure 7(d).

Note that the wall-normal velocity u‾y\overline{u}_y is found to be larger in band-free PCF than in PCF. See the zz-spacing of the streamlines at the interface z≃−20z \simeq-20 in Figure 7(b), as opposed to that in Figure 7(a). This can also be seen indirectly comparing the large-scale flow for PCF and band-free PCF in Figure 1 of the main text. For band-free PCF, the streamwise flow changes rapidly, and in the absence of large-scale spanwise flow, there must be significant wall-normal flow by incompressibility.

Both PCF and band-free PCF exhibit centro-symmetry (u‾x,u‾y,u‾z)(x,y,z)→(−u‾x,−u‾y,−u‾z)(−x,−y,−z)(\overline{u}_x,\overline{u}_y,\overline{u}_z)(x,y,z) \to(-\overline{u}_x,-\overline{u}_y,-\overline{u}_z)(-x,-y,-z), as visible in Figure 7(a,Figure 7b). Centro-symmetry (u‾strm,u‾y,u‾span)(strm,y,span)→(−u‾strm,−u‾y,−u‾span)(−strm,−y,−span)(\overline{u}_{\mathrm{strm}},\overline{u}_y,\overline{u}_{\mathrm{span}})(\mathrm{strm},y,\mathrm{span}) \to(-\overline{u}_{\mathrm{strm}},-\overline{u}_y,-\overline{u}_{\mathrm{span}})(-\mathrm{strm},-y,-\mathrm{span}) is also noticeable in Figure 7(c,Figure 7d). Both centro-symmetry operations are symmetries of the problem, that is symmetries of the equations and boundary conditions. In band-free PCF, the mean flow appears to present an additional reflection symmetry (u‾x,u‾y,u‾z)(x,y,z)→(u‾x,−u‾y,u‾z)(x,y,−z)(\overline{u}_x,\overline{u}_y,\overline{u}_z)(x,y,z) \to(\overline{u}_x,-\overline{u}_y,\overline{u}_z)(x,y,-z). (Because this is not a symmetry of the PCF configuration, it holds only approximately.) In particular, the shapes of the color patches corresponding to u‾x\overline{u}_x are approximately reflection-symmetric (z→−z)(z \to-z) in Figure 7(b), in contrast to the color patches in Figure 7(a). Similarly, in usual streamwise-spanwise coordinates, the symmetry (u‾strm,u‾y,u‾span)(strm,y,span)→(u‾strm,−u‾y,u‾span)(−strm,y,span)(\overline{u}_{\mathrm{strm}},\overline{u}_y,\overline{u}_{\mathrm{span}})(\mathrm{strm},y,\mathrm{span}) \to(\overline{u}_{\mathrm{strm}},-\overline{u}_y,\overline{u}_{\mathrm{span}})(-\mathrm{strm},y,\mathrm{span}), very nearly holds for band-free PCF, as can be seen in Figure 7(d). PCF clearly does not exhibit this approximate symmetry because the mean spanwise flow (color patches in Figure 7(c)) is not even close to symmetric under streamwise reflection.

The presence of excited and refractory zones, with the associated spatial mean-turbulent feedback (see [24, 28, 31, 32]), is the main reason that shear flows sustain bands or puffs. In PCF, the presence of excited and refractory regions is consistent with the absence of reflection symmetry z→−zz \to-z, (or equivalently for pipe flow, with the absence of upstream-downstream symmetry). Downstream mean flow slowly reacts to upstream turbulence, hence the upstream interface (at z<0,y>0z < 0, y > 0 in Figure 2(a,Figure 2b) in the main text) differs from the downstream interface (at z>0,y>0z > 0, y > 0). Note that the absence of reflection symmetry z→−zz \to-z in PCF is related to the overhang regions [19, 57], in which laminar flow in one layer faces turbulent flow in the other layer.

In contrast, band-free PCF has a near-reflection symmetry z→−zz \to-z in the mean flow, so that there is no strong difference between upstream and downstream laminar-turbulent interfaces. This implies that there are no excited and refractory regions, and hence no significant overhang regions. This is why band-free PCF lacks discrete turbulent structures, in spite of its having a two-component mean flow like that of pipe flow. This is true independently of the intensity of u‾y\overline{u}_y.

Not all wall-bounded shear flows without discrete turbulent structures possess an extra reflection symmetry. For example bent pipe flow lacks puffs [48, 49], but does not have an approximate upstream-downstream symmetry. In such systems, it remains to be seen whether the absence of significant excited and refractory regions is responsible for the absence of discrete turbulent structures.

Supplement to Figure 4 of the main text: Front contraction speeds

The front velocity measured in Figure 4 in the main text lacks negative points measuring contraction speeds. The reason is illustrated on Figure 8, where a localized turbulent patch is simulated at Re=381<Re0Re = 381 < Re_0. The flow laminarizes due to two competing effects: front contraction and the nucleation of laminar gaps within the turbulent phase, as seen in Figure 8c. The model in the main text and in Eq. (A.14) below reproduces this phenomenology. Although a mean negative velocity could be extracted from the mean turbulent lifetime for a given initial width of turbulence, this would not measure single front contraction, but would average effects of contraction and gap nucleation.

Image not extracted

Figure 8. Front contraction below Re0Re_0 in band-free PCF at θ=24∘\theta= 24^\circ and Re=381Re = 381 (a,b,c), and in the stochastic model considered in the main text (see Eq. (A.14)) with σ=2.0\sigma= 2.0 (d,e,f), plotted with time and length rescaling and choice of rr as discussed in § 1.0.8. Cases shown correspond to different realizations of the same initial condition. In both DNS and model, turbulence contracts and sometimes decays in blocks, making the measurement of a contraction velocity problematic.

The asymmetry between expanding and contracting turbulent front comes about as follows. We measure the mean front speed as the change in distance between the most upstream and the most downstream turbulent-laminar fronts, divided by twice the observational time window (to account for the two front motions). For sufficiently expanding turbulence, laminar-gap formation within turbulence has no significant effect on this measurement procedure and we can measure mean front speeds precisely. For contracting turbulent patches, even at Re=381Re=381, just slightly below Re0=383Re_0=383, laminar-gap formation affects our procedure and results in values that are not representative of mean front motion, but rather a combination of front motion and laminar-gap nucleation.

Quench experiments

We expand on the quench experiments presented in the main text and show space-time visualizations of the flow at various Reynolds numbers, for both PCF and band-free PCF. The flow is initiated with uniform turbulence simulated at Re=500Re=500, and ReRe is changed to a desired value. PCF spontaneously forms its characteristic isolated bands after a few hundred time units. We visualize this process on the left column of Figure 9 in the tilted domain at angle θ=24∘\theta=24^\circ. Depending on the value of ReRe, the flow resulting from the quench will be made of either intermittent patches of turbulent bands surrounded by an otherwise turbulent flow (Re=420Re=420), regularly-spaced bands (Re=350Re=350), or sparse, isolated bands (Re=320Re=320) which populate the flow near the critical point Rec≃328.7Re_c \simeq328.7 [10].

Image not extracted

Figure 9. Quench experiments in both PCF and band-free PCF in a tilted domain with θ=24∘\theta= 24^\circ.

In band-free PCF, the flow at Re=390Re=390 is mostly uniform, although interspersed with laminar gaps of various width, which randomly nucleate from the turbulent phase. When ReRe is decreased, these gaps become wider and contaminate the flow field via fluctuating turbulent-laminar fronts. This is the equilibrium between front contraction, expansion and gap nucleation that determines the critical ReRe below which turbulence is not sustained.

Stochastic reaction-diffusion model for turbulent front propagation

In this section, we elaborate on the stochastic reaction-diffusion model and on our fitting procedure with the DNS of band-free PCF. Let us first consider a dimensional reaction-diffusion model:

T~∂tq=L~2∂zz2q−∂qV(q,r)+(L~T~)1/2σq ξ(z,t)(25)\widetilde{T}\partial_t q = \widetilde{L}^{2}\partial_{zz}^{2}q - \partial_q V(q,r) + (\widetilde{L}\widetilde{T})^{1/2}\sigma q\,\xi(z,t) \tag*{(25)}

with

V(q,r)≡q22(1+(r+1)(q22−4q3))V(q,r) \equiv\frac{q^2}{2}\left(1+(r+1)\left(\frac{q^2}{2}-\frac{4q}{3}\right)\right)

where ξ(z,t)\xi(z,t) is a space-time white Gaussian noise of unit variance, such that ⟨ξ(x,t)ξ(x′,t′)⟩=δ(t−t′)δ(x−x′)\langle\xi(x,t)\xi(x',t')\rangle= \delta(t-t')\delta(x-x'). We nondimensionalize (A.13) by the characteristic time T~\widetilde{T} and length L~\widetilde{L}, leading to:

∂tq=∂zz2q−∂qV+σq ξ(26)\partial_t q = \partial_{zz}^{2}q - \partial_q V + \sigma q\,\xi \tag*{(26)}

The space-independent and deterministic (σ=0\sigma= 0) version of the model presents a critical value r=1/8r = 1/8 above which the non-zero fixed point is the global minimum of VV. When noise is added to the local model, the critical point can be obtained by solving the associated Fokker-Planck equation. Noise both shifts the critical value above which q≠0q \ne0 is preferred, and also triggers transitions from the turbulent to the absorbing state q=0q = 0. When diffusion is added to the problem, the solution q(z,t)q(z,t) consists of fronts connecting the metastable and the stable fixed points [58]. We simulate (A.14) with periodic boundary conditions, with a finite-difference scheme under the Itô representation. We measure the expanding front speed from an initial turbulent region and denote r0r_0 the point at which this speed is zero. We find that

r0≃18+3100σ2(27)r_0 \simeq\frac{1}{8}+\frac{3}{100}\sigma^2 \tag*{(27)}

(see Figure 10a), and that the front speed behaves as (Figure 10b)

Image not extracted

Figure 10. (a) Evolution of the zero-propagation point r0r_0 as a function of σ\sigma. The left-most cross shows the deterministic critical point, r0=1/8r_0 = 1/8. Fit r0≃1/8+3/100σ2r_0 \simeq1/8 + 3/100\sigma^2 is shown as a dashed line. (b) Front speed cc as a function of r−r0(σ)r-r_0(\sigma) for the values of σ\sigma shown in (a). All curves collapse to the deterministic function FF (thick grey line).

c(r,σ)≃F(r−r0(σ))(28)c(r,\sigma) \simeq F\left(r-r_0(\sigma)\right) \tag*{(28)}

with FF a non-linear front speed, independent of σ\sigma. FF is computed in the deterministic case σ=0\sigma= 0, via nonlinear simulations (see Barkley [30], appendix A).

Note first that (A.16) is valid for r>r0r > r_0, because the negative front velocities depend on the fluctuations σ\sigma, as discussed in Figure 8. Second, FF in Eq. (A.16) may depend on σ\sigma very near r0r_0 depending on the order of the phase transition. Third, r0r_0 is close to, but a priori lower than, the critical point rcr_c: when the transition is second-order, the critical point is governed by a balance between slightly expanding fronts and laminar-gap nucleation [59].

From (A.15) and (A.16), the dimensional model (A.13), parametrized by (T~,L~,r)(\tilde{T},\tilde{L},r), has front speeds obeying the rescaled curve

cT~,L~(r)=L~T~F(r−18−3100σ2)(29)c_{\tilde{T},\tilde{L}}(r) = \frac{\tilde{L}}{\tilde{T}} F\left(r-\frac{1}{8}-\frac{3}{100}\sigma^2\right) \tag*{(29)}

We find that this curve is in good agreement with the DNS of band-free PCF, up to a rescaling of the axes r−r0r-r_0 and cc such that

cDNS(Re)=aF(b(Re−Re0))(30)c_{\mathrm{DNS}}(Re) = a F\left(b(Re-Re_0)\right) \tag*{(30)}

We first estimate the critical point in the DNS to be Re0≃383Re_0 \simeq383 from the zero front speed point. We then find aa and bb from fitting the DNS data with the form FF extracted from simulations of model (A.14). The best fit yields (a,b)=(0.189,0.0016)(a,b) = (0.189, 0.0016) with an mean-squared error of 1.9×10−41.9 \times10^{-4} (see the solid red line in Figure 4 in the main text. Figure 11a shows the mean-squared error associated to this fit.

Image not extracted

Figure 11. (a) Mean-square error associated with the fitting (A.18) with the DNS of band-free PCF. (a,b)≃(0.19,0.0016)(a,b) \simeq(0.19, 0.0016) (red cross) globally minimizes the error. (b,c,d) Visualisation of model (A.14) for different values of σ\sigma, with rr close to critical point rcr_c. (b) is initiated with a wide localised turbulent zone. As the transition is first order at this value of σ\sigma, the state near critical point is fixed by the initial condition, contrary to (c) and (d).

From parameter aa, we find a characteristic velocity scale L~/T~=a=cDNS/F\tilde{L}/\tilde{T} = a = c_{\mathrm{DNS}}/F fitting the DNS front speed. Similarly, the length scale L~\tilde{L} can be set to correspond to the front width in PCF. In the DNS, we measure the characteristic front width from Figure 2 in the main text: denoting ⟨Eturb⟩\langle E_{\mathrm{turb}}\rangle the mean turbulent energy in the turbulent zone, we define the front width lDNSl_{\mathrm{DNS}} such that EturbE_{\mathrm{turb}} goes from 0.001⟨Eturb⟩0.001\langle E_{\mathrm{turb}}\rangle to 0.999⟨Eturb⟩0.999\langle E_{\mathrm{turb}}\rangle. We carry out the same measurement in the model (with qq instead of EturbE_{\mathrm{turb}}). This yields lDNS≃39.4l_{\mathrm{DNS}} \simeq39.4 in DNS and l≃13.74l \simeq13.74 in the model (the measurement is carried out with σ=0\sigma= 0). The ratio of the two gives the rescaling length scale L~=lDNS/l≃3.0\tilde{L} = l_{\mathrm{DNS}}/l \simeq3.0. The time scale is

deduced from T~=L~/a≃16\widetilde{T}=\widetilde{L}/a\simeq16. Finally, parameter bb connects Reynolds-number scales via r−r0=b(Re−Re0)r-r_0=b(Re-Re_0) and is used to compare space-time visualizations of model and DNS at the same distance to the zero-speed point.

To find the value of σ\sigma best corresponding to the DNS, we use the measured turbulent fraction at a finite distance from critical point, RecRe_c in the DNS, and rc(σ)r_c(\sigma) in the model. Letting ϵDNS≡(Re−Rec)/Rec\epsilon_{\mathrm{DNS}}\equiv(Re-Re_c)/Re_c and ϵ≡(r−rc)/rc\epsilon\equiv(r-r_c)/r_c, we choose σ\sigma such that

Ft,DNS(ϵDNS=0.01)≃Ft,model(ϵ=0.01,σ).(31)F_{t,\mathrm{DNS}}(\epsilon_{\mathrm{DNS}}=0.01)\simeq F_{t,\mathrm{model}}(\epsilon=0.01,\sigma). \tag*{(31)}

This yields σ=2.0\sigma=2.0, and both the DNS and model turbulent fractions are approximately 0.9 at ϵ=0.01\epsilon=0.01. The precision of this determination is limited by the precision of RecRe_c, which we have only determined to ±0.5\pm0.5.

Supplement to Figure 5 of the main text: turbulent fraction and other DP exponents

In the DNS of band-free PCF, we define the local turbulent energy as

e(z,t)=12⟨uy(x,y,z,t)2+uspan(x,y,z,t)2⟩x,y(32)e(z,t)=\frac{1}{2}\left\langle u_y(x,y,z,t)^2+u_{\mathrm{span}}(x,y,z,t)^2\right\rangle_{x,y} \tag*{(32)}

and the instantaneous turbulent fraction as

Ft(t)≡∫0Lzdz Θ(e(z,t)−elam)(33)F_t(t)\equiv\int_0^{L_z} dz\ \Theta(e(z,t)-e_{\mathrm{lam}}) \tag*{(33)}

where Θ\Theta is the Heaviside function and elame_{\mathrm{lam}} is a threshold delimiting turbulent and laminar regions. We use elam=3×10−3e_{\mathrm{lam}}=3\times10^{-3}. The equilibrium turbulent fraction is measured when Ft(t)F_t(t) saturates. The measurements of FtF_t in the DNS are done in domains of size Lz=800L_z=800 for ReRe-values far from RecRe_c, and Lz=2400L_z=2400 close to RecRe_c.

In the model, turbulent fraction is defined similarly as

Ft(t)≡∫0Ldz Θ(q(z,t)−qlam)(34)F_t(t)\equiv\int_0^L dz\ \Theta(q(z,t)-q_{\mathrm{lam}}) \tag*{(34)}

where qlam=10−3q_{\mathrm{lam}}=10^{-3}. The equilibrium turbulent fraction is measured in a domain of size L=105L=10^5. We find that a different choice of elame_{\mathrm{lam}} and qlamq_{\mathrm{lam}} affects only very weakly the curves presented in Figure 5(f) in the main text.

The DP universality class is characterized by three universal scaling exponents, β\beta, ν⊥\nu_\perp, ν∥\nu_\parallel, that dictate the scaling of the mean density of active sites and their spatial and temporal correlations in the vicinity of critical point:

Ft∼ϵβ,ξ⊥∼ϵ−ν⊥,andξ∥∼ϵ−ν∥(35)F_t\sim\epsilon^\beta,\qquad\xi_\perp\sim\epsilon^{-\nu_\perp},\qquad\text{and}\qquad\xi_\parallel\sim\epsilon^{-\nu_\parallel} \tag*{(35)}

where ϵ≡(r−rc)/rc\epsilon\equiv(r-r_c)/r_c is the reduced control parameter. In one spatial dimension, β≃0.2765\beta\simeq0.2765, ν⊥≃1.097\nu_\perp\simeq1.097, and ν∥≃1.734\nu_\parallel\simeq1.734.

In Figure 5(f) of the main text, we have computed the equilibrium turbulent fraction as a function of ϵ\epsilon for σ=2.0\sigma=2.0. The data exhibits scaling consistent with Ft∼ϵβF_t\sim\epsilon^\beta. To assess correlations for this same case, we measure the distributions of empty sites (laminar gaps) in both spatial and temporal directions, giving laminar-gap data llaml_{\mathrm{lam}} and tlamt_{\mathrm{lam}}. The distributions of these data are expected to show power-law behavior N(llam)∼llamμ⊥N(l_{\mathrm{lam}})\sim l_{\mathrm{lam}}^{\mu_\perp} and N(tlam)∼tlamμ∥N(t_{\mathrm{lam}})\sim t_{\mathrm{lam}}^{\mu_\parallel}, at the critical point. Exponents μ⊥\mu_\perp and μ∥\mu_\parallel are connected to ν⊥\nu_\perp and ν∥\nu_\parallel via

μ⊥=2−βν⊥andμ∥=2−βν∥.(36)\mu_\perp=2-\frac{\beta}{\nu_\perp}\qquad\text{and}\qquad\mu_\parallel=2-\frac{\beta}{\nu_\parallel}. \tag*{(36)}

See [59, 60]. In one spatial dimension, μ⊥≃1.748\mu_\perp\simeq1.748 and μ∥≃1.840\mu_\parallel\simeq1.840.

Laminar-gap distributions from model simulations very close to the critical point (ϵ=4×10−5\epsilon=4\times10^{-5}) are shown in Figure 12. Both distributions exhibit the expected scalings. (The scaling range for the spatial gaps is not large, but a full decade of scaling, llam∈[2×102,2×103]l_{\mathrm{lam}}\in[2\times10^2,2\times10^3], is visible.) We conclude that for noise amplitude σ=2.0\sigma=2.0, model (A.14) exhibits phase transition in the DP universality class. We have not verified the critical exponents associated with correlations at other values of σ\sigma. Nevertheless, in the large-noise cases where we observe a continuous growth in the turbulent fraction with the expected exponent for DP, it is reasonable to assume that these cases are also in the DP universality class, in accordance with the Janssen-Grassberger conjecture [3, 4].

Image not extracted

Figure 12. Scalings for temporal and spatial correlations, from model simulations close to the critical point, ϵ=4×10−5\epsilon= 4 \times10^{-5}, for noise strength σ=2.0\sigma= 2.0. Distribution of laminar-gap lengths in (a) the spatial zz direction and (b) the temporal direction. Laminar gaps are obtained from simulations in a domain of size L=105L = 10^{5}, after the system reaches saturation, for time t∈[1×106,2×107]t \in[1 \times10^{6}, 2 \times10^{7}]. Critical power laws for the DP universality class are shown.

References

References

  1. [1]L. D. Landau and E. M. Lifshitz, Fluid Mechanics: Course of Theoretical Physics, Volume 6, 2nd ed. (Butterworth-Heinemann, 1987).
  2. [2]Y. Pomeau, Physica D 23, 3 (1986).
  3. [3]P. Grassberger, in Nonlinear Phenomena in Chemical Dynamics (Springer, 1981) p. 262.
  4. [4]H.-K. Janssen, Z. Phys. B 42, 151 (1981).
  5. [5]F. Waleffe, Phys. Fluids 9, 883 (1997).
  6. [6]A. Prigent, G. Grégoire, H. Chaté, O. Dauchot, and W. van Saarloos, Phys. Rev. Lett. 89, 014501 (2002).
  7. [7]A. Prigent, G. Grégoire, H. Chaté, and O. Dauchot, Physica D 174, 100 (2003).
  8. [8]D. Barkley and L. S. Tuckerman, Phys. Rev. Lett. 94, 014502 (2005).
  9. [9]Y. Duguet, P. Schlatter, and D. S. Henningson, J. Fluid Mech. 650, 119 (2010).
  10. [10]G. Lemoult, L. Shi, K. Avila, S. V. Jalikop, M. Avila, and B. Hof, Nature Physics 12, 254 (2016).
  11. [11]K. Avila, D. Moxey, A. De Lozar, M. Avila, D. Barkley, and B. Hof, Science 333, 192 (2011).
  12. [12]M. Chantry, L. S. Tuckerman, and D. Barkley, J. Fluid Mech. 824, R1 (2017).
  13. [13]L. Klotz, G. Lemoult, K. Avila, and B. Hof, Phys. Rev. Lett. 128, 014502 (2022).
  14. [14]K. Takeda, Y. Duguet, and T. Tsukahara, Entropy 22, 988 (2020).
  15. [15]K. Kohyama, M. Sano, and T. Tsukahara, Phys. Fluids 34, 084112 (2022).
  16. [16]D. Coles and C. van Atta, AIAA Journal 4, 1969 (1966).
  17. [17]I. J. Wygnanski and F. Champagne, J. Fluid Mech. 59,
  18. [18]D. Barkley and L. S. Tuckerman, J. Fluid Mech. 576, 109 (2007).
  19. [19]Y. Duguet and P. Schlatter, Phys. Rev. Lett. 110, 034502 (2013).
  20. [20]M. Couliou and R. Monchaux, Phys. Fluids 27, 034101 (2015).
  21. [21]L. Klotz, A. Pavlenko, and J. Wesfreid, J. Fluid Mech. 912 (2021).
  22. [22]E. Meriansi, G. Yalnız, and B. Hof, J. Fluid Mech. 974, A21 (2023).
  23. [23]C. W. van Doorne and J. Westerweel, Philos. Trans. R. Soc. A 367, 489 (2009).
  24. [24]B. Hof, A. De Lozar, M. Avila, X. Tu, and T. M. Schneider, Science 327, 1491 (2010).
  25. [25]D. Samanta, A. De Lozar, and B. Hof, J. Fluid Mech. 681, 193 (2011).
  26. [26]B. Song, D. Barkley, B. Hof, and M. Avila, J. Fluid Mech. 813, 1045 (2017).
  27. [27]S. Gomé, L. S. Tuckerman, and D. Barkley, J. Fluid Mech. 964, A17 (2023).
  28. [28]S. Gomé, L. S. Tuckerman, and D. Barkley, J. Fluid Mech. 964, A16 (2023).
  29. [29]D. Barkley, Phys. Rev. E 84, 016309 (2011).
  30. [30]D. Barkley, J. Fluid Mech. 803, P1 (2016).
  31. [31]S. J. Benavides and D. Barkley, arXiv:2309.12879 (2023).DOI
  32. [32]D. Barkley, J. Phys.: Conf. Ser. 318, 032001 (2011).
  33. [33]X. Wang, H.-Y. Shih, and N. Goldenfeld, Phys. Rev. Lett. 129, 034501 (2022).
  34. [34]J. Gibson, F. Reetz, S. Azimi, A. Ferraro, T. Krei- los, H. Schrobsdorff, M. Farano, A. Yesil, S. Schütz, M. Culpo, and T. Schneider, “Channelflow 2.0,” Manuscript in preparation, see channelflow.ch.
  35. [35]J. M. Hamilton, J. Kim, and F. Walefe, J. Fluid Mech. 287, 317 (1995).
  36. [36]T. Liu, B. Semin, R. Godoy-Diana, and J. E. Wesfreid, Phys. Rev. Fluids 9, 033901 (2024).
  37. [37]Appendix containing additional information on numerical methods and model (1) and Figs. 2, 4, and 5.
  38. [38]L. Shi, M. Avila, and B. Hof, Phys. Rev. Lett. 110, 204502 (2013).
  39. [39]S. Gomé, L. S. Tuckerman, and D. Barkley, Phys. Rev. Fluids 5, 083905 (2020).
  40. [40]M. Avila, A. P. Willis, and B. Hof, J. Fluid Mech. 646, 127 (2010).
  41. [41]V. Mukund, C. Paranjape, M. P. Sitte, and B. Hof, arXiv:2112.06537 (2021).DOI
  42. [42]D. Xu and B. Song, J. Fluid Mech. 950, R3 (2022).
  43. [43]Y. Pomeau, Comptes Rendus Mécanique 343, 210 (2015).
  44. [44]M. A. Munoz, Phys. Rev. E 57, 1377 (1998).
  45. [45]M. A. Munoz and R. Pastor-Satorras, Phys. Rev. Lett. 90, 204101 (2003).
  46. [46]M. A. Munoz, arXiv:cond-mat/0303650 (2003).DOI
  47. [47]C. W. Gardiner, Handbook of stochastic methods, Vol. 3 (Springer Berlin, 1985).
  48. [48]E. Rinaldi, J. Canton, and P. Schlatter, J. Fluid Mech. 866, 487 (2019).
  49. [49]H. Zhuang, B. Yang, V. Mukund, E. Marensi, and B. Hof, arXiv:2311.11474 (2023).DOI
  50. [50]C. Rorai, P. D. Mininni, and A. Pouquet, Phys. Rev. E 89, 043002 (2014).
  51. [51]T. Khapko, P. Schlatter, Y. Duguet, and D. S. Hennigson, J. Fluid Mech. 795, 356 (2016).
  52. [52]Y. Duguet, O. Le Maitre, and P. Schlatter, Phys. Rev. E 84, 066315 (2011).
  53. [53]A. Pershin, C. Beaume, and S. M. Tobias, J. Fluid Mech. 867, 414 (2019).
  54. [54]C. Canuto, M. Y. Hussaini, A. Quarteroni, and T. A. Zang, Spectral methods: fundamentals in single domains (Springer-Verlag, Berlin Heidelberg, 2007).
  55. [55]J. Jiménez and A. Pinelli, J. Fluid Mech. 389, 335 (1999).
  56. [56]J. Jimenez, arXiv:2202.09814 (2022).DOI
  57. [57]A. Lundbladh and A. V. Johansson, J. Fluid Mech. 229, 499 (1991).
  58. [58]W. Van Saarloos, Physics Reports 386, 29 (2003).
  59. [59]H. Hinrichsen, Advances in Physics 49, 815 (2000).
  60. [60]S. Lübeck, Int. J. Mod. Phys. B 18, 3977 (2004).

Paper details

vlnc:W_3n42k9ph
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