3.1. Amplification of vertical shear

We begin by studying the transient growth of the initial state defined in (2.9) for the simulations F05, F1 and F2, where the results from simulation F07 follow a similar pattern and are omitted from this section for clarity. The initial perturbation $(\tilde {u}(z), \tilde {v}(z),0)$ causes the magnitude of the streamwise velocity $u$ to increase rapidly before saturating at around $O(1)$, which we investigate in detail below. First, figure 1 shows visualisations of the streamwise velocity field $u$ at time $t\approx 48$ after saturation of the initial growth. Panel $(a)$ shows the vertical profile of the centreline velocity $u(0,0,z)$ for simulations F05 (red solid line), F1 (blue dashed line) and F2 (orange dot-dashed line). The black dotted line shows the shape of the initial perturbation $\tilde {v}(z)$ of the spanwise velocity normalised to have similar amplitude for ease of visualisation. A distinct vertical mode structure in $u$ is present, in all cases taking a similar (reflected) shape to the initial profile, which strongly suggests the growth in $u$ is locally determined by $\tilde {v}$ in a manner consistent with (2.8).

Figure 1.

(a) Vertical profiles of streamwise velocity $u(0,0,z)$ taken at time $t\approx 48$ for simulations F05 (red solid line), F1 (blue dashed line) and F2 (orange dot-dashed line). The black dotted line shows the shape of the initial spanwise velocity profile $\hat {\tilde {v}}(z)$ normalised to a similar amplitude for comparison. (b,c) Contour plots showing vertical slices from simulation F1 in the $y$ and $x$ planes of the streamwise velocity field (b) $u(x,0,z)$ and (c) $u(0,y,z)$. The directory including the data and Jupyter notebook for producing the figure can be found at https://multimediafluids.uam.elogim.com/S0022112024001216/JFM-Notebooks/files/fig1.

There are small but noticeable differences between the simulations with different initial Froude number $Fr_0$: in general, a lower Froude number results in more pronounced higher wavenumber modes. Recalling that $\tilde {v}$ is constructed of a sum of randomly phased Fourier modes ${\rm e}^{{\rm i}mz}$ for the lowest seven permissible wavenumbers in the vertical, this means that the higher wavenumber modes exhibit enhanced growth for smaller $Fr_0$. This is entirely consistent with the results of Deloncle et al. (Reference Deloncle, Chomaz and Billant2007) who show that, for the same base flow without the additional vertical perturbations, three-dimensional infinitesimal normal mode perturbations with vertical wavenumber $m$ less than $1/Fr_0$ grow almost equally as fast as the most unstable two-dimensional mode with $m=0$ corresponding to that of classical unstratified shear instability, with a larger growth rate for smaller $Fr_0$. It is important to stress here that, though there exists a clear scaling similarity, the finite-amplitude modes we are considering here are entirely different and grow over a much shorter time scale than the infinitesimal perturbations considered by Deloncle et al. (Reference Deloncle, Chomaz and Billant2007), the latter of which can be seen emerging in the form of a vertical column defect length scale in the fully nonlinear simulations of BS06.

The quasi-steady state reached by the algebraic growth consists of horizontal layers in the streamwise velocity field that are localised at the velocity interface in the $y$ direction and extend across the whole domain in the $x$ direction, as can be seen by looking at panels (b,c) of figure 1. We note that $v$, $w$ and $\rho$ are small in comparison to $u$ and only grow substantially once horizontal shear instability occurs later during the flow evolution. Panel (c) shows that the interface in the $y$ direction is distorted by the layering, though the magnitude of the maximum velocity gradient $\partial u/\partial y$ remains similar and, hence, as we will see, this distortion does not prevent the rolling up of the sheet of vorticity into Kelvin–Helmholtz billows in the $x$–$y$ plane caused by inflectional shear instability.

We now investigate the dynamics of the transient growth in detail. In figure 2 the centreline streamwise velocity magnitude $\langle |u(y=0)|\rangle$ averaged over the $x$ and $z$ directions can be seen to increase linearly with time initially, matching the algebraic growth predicted by the lift-up mechanism of Ellingsen & Palm (Reference Ellingsen and Palm1975). Moreover, we can plot the corresponding $x$- and $z$-averaged terms from (2.8) to find a very good match with the predicted growth rate $\langle |v\partial \bar {u}/\partial y(y=0)|\rangle$. The streamwise velocity magnitude continues to grow until it starts to saturate at a time proportional to $Fr_0$, so that the saturation time for F1 is approximately double that of F05 and so on. Noting that the dimensionless buoyancy frequency is given by $1/Fr_0$, this indicates that the stratification is responsible for equilibrating the flow to a quasi-steady vertical layered state approximately over two buoyancy periods $4{\rm \pi} Fr_0$, which is then sustained until the growth of the horizontal shear instability becomes significant. Perhaps somewhat surprisingly, the agreement with (2.8) is maintained throughout the saturation of the layered state. This is because, as suggested by figure 1(b), $u$ and $p$ continue to remain independent of the streamwise coordinate $x$ as the perturbations grow. Meanwhile, $v$ decays and $w$ remains small, with the result being that the nonlinearities and pressure gradient term in the streamwise momentum equation are negligible (along with the viscous term) meaning it still reduces to the form (2.8).

Figure 2.

The evolution of streamwise velocity magnitude evaluated at $y=0$ and integrated over the $x$ and $z$ directions (right axis), as well as the corresponding terms in the linearised evolution equation (2.8) (left axis). Results are shown for simulations (a) F05, (c) F1 and (c) F2. The directory including the data and Jupyter notebook for producing the figure can be found at https://multimediafluids.uam.elogim.com/S0022112024001216/JFM-Notebooks/files/fig2.

It is worth briefly comparing the transient growth observed here to that seen in existing studies of the hyperbolic tangent shear layer profile. For the case of the unstratified shear layer (here obtained by taking $Fr_0\to \infty$), Arratia et al. (Reference Arratia, Caulfield and Chomaz2013) show that infinitesimal perturbations of the form $\boldsymbol {u}_0(y) \exp ({{\rm i}k_xx+{\rm i}k_zz})$ optimizing the (linear) energy growth over sufficiently short time horizons are inherently three dimensional, i.e. both horizontal and vertical wavenumbers $k_x$ and $k_z$ are non-zero, in contrast to the classical most unstable two-dimensional shear instability mode with $k_z=0$ obtained from linear stability analysis. Kaminski et al. (Reference Kaminski, Caulfield and Taylor2014) find similar results for the stratified shear layer in the case where the shear is parallel to gravity (equivalent to gravity acting in the $y$ direction in the coordinate system we use here). In both studies, optimal perturbations over short time horizons consist of oblique waves localised at the centre of the shear layer and tilted against the shear, growing through a combination of the lift-up and Orr mechanisms (Orr Reference Orr1907; Farrell & Ioannou Reference Farrell and Ioannou1993), though the behaviour may be markedly different in the presence of stratification due to the excitation of internal gravity waves at some distance from the central interface. The inviscid linear evolution of non-normal perturbations to the initial condition (2.1a,b) used in this work where shear is orthogonal to gravity is studied by Arratia (Reference Arratia2011). This system is also discussed by Bakas & Farrell (Reference Bakas and Farrell2009a,Reference Bakas and Farrellb) as an extension of a more general investigation on the behaviour of internal wave perturbations on a flow with constant linear background shear. For the study here in which we focus primarily on the nonlinear flow evolution, it suffices to say that in both cases, optimal perturbations in the limit of $k_x \to 0$ exhibit linear streamwise velocity growth with time purely via the lift-up mechanism, precisely as we have observed here.

3.2. Turbulent transition

The initial conditions of BS06 are equivalent to ours in the absence of the vertical perturbations leading to horizontal layer formation. They observe the development of horizontal shear instability in the form of columnar vortices that interact with adjacent vortices over a characteristic vertical length scale dependent on $Fr_0$. However, this length scale only becomes substantially smaller than the length scale of the horizontal shear mode for $Fr_0\lesssim 1$. Even when this is the case, the induced three-dimensional behaviour does not lead to a transition to small-scale turbulence. In our DNS, the lift-up mechanism enables the transient growth of layers with a significantly smaller vertical scale than the vertical length scale observed in BS06 associated with the fastest growing three-dimensional normal mode of the background flow. As discussed at the end of § 2.3, the magnitudes of vertical perturbations we impose on top of the set-up of BS06 are chosen so that the transient growth and saturation of the vertically sheared layers does not by itself destabilize the flow. Here we show that the subsequent onset of horizontal shear instability facilitates a transition to small-scale turbulence by interacting with the existing horizontal layers and enhancing the vertical shear between them, as well as creating local statically unstable overturns, or regions where the total density gradient $-1+ \partial \rho /\partial z >0$.

In figure 3 we use sequential violin plots (see the caption for a detailed description) to illustrate the evolution of the shape of the probability density function (p.d.f.) of the local gradient Richardson number $Ri_g(x,0,z,t)$ defined in (2.10), evaluated in the plane $y=0$ at the centre of the horizontal shear layer. For clarity, only those values $-0.5< Ri_g<2$ are considered. At time $t=55$, the horizontal layers in $u$ have saturated in magnitude as is evident from figure 2. As per our design, the vertical shear at this point between the layers alone is not strong enough to trigger vertical shear instability and growing Kelvin–Helmholtz billows. This is reflected in the distributions of $Ri_g$, which do not exhibit values less than $1/4$ corresponding to the classical marginal stability value from the Miles–Howard criterion (though this is strictly only valid for normal mode perturbations to parallel inviscid stratified shear flows (Howard Reference Howard1961; Miles Reference Miles1961)) for any value of $Fr_0$.

Figure 3.

Violin plots showing the evolution of the shape of the p.d.f. of the gradient Richardson number $-0.5< Ri_g<2$ in the central plane $y=0$. Each plot is centred at its corresponding time on the $x$ axis. The breadth of the shaded region corresponds to the magnitude of the p.d.f. evaluated at the corresponding value of $Ri_g$ shown on the $y$ axis. Results are shown for simulations (a) F07, (b) F1 and (c) F2. The black dot-dashed lines correspond to the notional marginal value of $Ri_g=1/4$. Red lines (right axis) show the evolution of the spanwise perturbation velocity $\langle v'^2\rangle$ as an indicator for the onset of horizontal shear instability.

To trigger a transition to turbulence, the onset of horizontal shear instability is required. A proxy for the growth of this instability is the root-mean-square magnitude of the spanwise velocity $\langle v^2\rangle$, here averaged over $y=0$, whose evolution is shown by the solid red lines in figure 3. Because the fastest growing mode of instability is purely horizontal (Deloncle et al. Reference Deloncle, Chomaz and Billant2007), the onset and development is not expected to be influenced strongly by $Fr_0$, as is consistent with the figure. However, the growing horizontal perturbation clearly interacts with and enhances the vertical shear between the existing layers, distinctly modifying both the shape and centre of the distribution of $Ri_g$. The effect is more pronounced for weaker stratification, or larger $Fr_0$. The growth of the horizontal mode results in the existence of local values of $Ri_g<1/4$ for all simulations, with an increasing relative density for smaller $Fr_0$. This indicates the potential for vertical shear instability (though we emphasise again that $Ri_g<1/4$ is only a notional, indicative proxy), which has often been assumed to be the canonical route to turbulence in LAST (e.g. Brethouwer et al. Reference Brethouwer, Billant, Lindborg and Chomaz2007). Moreover, values of $Ri_g<0$ corresponding to regions of static instability are observed, suggesting the advection of denser fluid over lighter fluid. One way this may be achieved is through the strong vertical shearing of horizontal gradients in density that arise due to the fact that the density perturbation field is strongly coupled to the vertical vorticity field, a feature that is discussed in detail in BS06.

In general, figure 3 illustrates how the growth of the horizontal shear mode renders the flow increasingly locally unstable and, hence, facilitates turbulent transition at small scales, provided $Fr_0$ is not too small. As the trend in the distribution shape with $Fr_0$ suggests, for simulation F05, it was found that $Ri_g$ remains larger than $1/4$ everywhere even after the growth of the horizontal mode and, correspondingly, turbulence did not develop. This simulation is therefore excluded from the remainder of the discussion, where we focus on F07, F1 and F2. We stress that, by design, the existence of the horizontal layers formed by transient growth and the subsequent development of horizontal shear instability are both necessary for turbulent transition. In the absence of either, insufficient regions of vertical shear and static instability are generated for the development of small-scale instabilities. Though they are not considered here, it seems likely that initial perturbations with increasingly large amplitudes would eventually lead to local vertical shear instability dominating proceedings as discussed in § 2.3.

To illustrate the highly three-dimensional transition to turbulence fully, figures 4 and 5 show horizontal and vertical slices of the vertical vorticity field $\zeta _z = \partial v/\partial x -\partial u/\partial y$ for simulations F07, F1 and F2 at various times throughout the flow evolution, following the initial development of horizontal layers. Despite the significant early growth of the horizontal layers described above, the subsequent flow still develops vertically coherent columnar billow structures due to horizontal shear instability in the manner described in BS06. This can clearly be seen in the left-hand panels of the two figures. Note that, in order to resolve small-scale turbulence, our simulations have a significantly smaller domain size in both the $x$ and $z$ directions than those of BS06 and, as a result, we do not see the billow interlocking behaviour they observe that occurs on vertical length scales close to the horizontal extent of the primary instability for $Fr_0=1$, though we note there is clear large-scale vertical decorrelation of billow pairing visible in figures 5(b) and 5(c) for the simulation F07 with the lowest $Fr_0$.

Figure 4.

Horizontal planes $z=0$ showing contours of the vertical vorticity field for simulations (a–c) F07, (d–f) F1, (g–i) F2. Plots (a,d,g) are taken at time $t=88$, (b,e,h) at $t=118$ and (c,f,i) at time $t=148$. Blue and red colours denote negative and positive values. An animated movie for simulation F1 is included in the supplementary materials available at https://doi.org/10.1017/jfm.2024.121. The directory including the data and Jupyter notebook for producing the figure can be found at https://multimediafluids.uam.elogim.com/S0022112024001216/JFM-Notebooks/files/fig4.

Figure 5.

Similar to figure 4 but this time vertical planes $y=0$ showing contours of the vertical vorticity field for simulations (a–c) F07, (d–f) F1, (g–i) F2. Plots (a,d,g) are taken at time $t=88$, (b,e,h) at $t=118$ and (c,f,i) at time $t=148$. Blue and red colours denote negative and positive values. An animated movie for simulation F1 is included in the supplementary materials. The directory including the data and Jupyter notebook for producing the figure can be found at https://multimediafluids.uam.elogim.com/S0022112024001216/JFM-Notebooks/files/fig5.

We observe distinctly different behaviour to that reported by BS06 occurring once columnar billows have formed, with the horizontal layers in the streamwise velocity field arising due to the lift-up mechanism strongly distorting the billow in the vertical at scales much smaller than its horizontal extent. This has the greatest effect on the outer regions, most notably the ‘braid’ structure that joins the two billows as can be seen in the left panels of figure 5. A breakdown to three-dimensional turbulence characterised by small-scale motions in the vorticity field follows the development of positive vertical vorticity structures in the braid seen in simulations F1 and F2, shown in figures 5(d) and 5(g). The braid vortex structures are not formed in flow F07, presumably due to the stronger influence of the stratification, though fully three-dimensional turbulence still eventually develops, as shown in panels (b,c).

At the same time, the enhancement of local vertical shear by the growing horizontal mode of instability means that the distributions of $Ri_g$ shown in figure 3 for each simulation have an increased density of points with $Ri_g<1/4$. Moreover, the existence of horizontal gradients in density aligned with the structures in the vertical vorticity field in the presence of this vertical shear generates statically unstable regions with $Ri_g<0$ as discussed above. Thus, conditions are created consistent with both shear and convective instability leading to turbulence production. Based on the results of Parker et al. (Reference Parker, Howland, Caulfield and Kerswell2021) who study optimal perturbations on a breaking internal gravity wave through its interaction with a vertical shear layer, we think it reasonable to suppose here that the precise sequence of mechanisms leading to turbulence will depend on the energy available to growing perturbations from the time-evolving background flow, whose evolution itself is governed (at least) by the initial conditions and parameters $Re_0$, $Fr_0$ and $Pr$. This same complexity is also apparent during the breakdown of Kelvin–Helmholtz billows in a stratified shear layer, which develop a ‘zoo’ of secondary instabilities facilitating transition to small-scale turbulence that vary strongly with the strength of the background stratification, as well as $Re_0$ and $Pr$ (Mashayek & Peltier Reference Mashayek and Peltier2012a,Reference Mashayek and Peltierb; Salehipour, Peltier & Mashayek Reference Salehipour, Peltier and Mashayek2015).

The spatial locality of regions where turbulence is produced leads to obvious large-scale intermittency that is present in all simulations. Horizontal intermittency is mainly due to the coherent billow core structure, with both figures demonstrating the columnar vortices eventually pairing that is noticeably decorrelated at different heights for flows F07 and F1, as can be seen most clearly in figure 5(b). At these smaller values of $Fr_0$, an increasing fraction of the vertical extent of the domain remains mostly laminar, likely due to the fact that the breakdown of the braid is inhibited by early pairing in these regions. Despite the differences in intermittency however, the centre and right-hand panels of figure 5 show that in general, where they are present, turbulent regions exhibit vertical anisotropy for all Froude numbers investigated, where emerging layered structures have a horizontal to vertical aspect ratio that increases with decreasing $Fr_0$. This is in qualitative agreement with the LAST regime.

3.3. Computation of turbulent statistics

Before exploring the dynamics of the three-dimensional flow regime quantitatively, we briefly outline how relevant statistics are calculated. Due to the presence of the initial background horizontal shear in the problem, the perturbation velocity field $\boldsymbol {u}'(\boldsymbol {x},t)$ is defined in the same manner as in BS06 as the perturbation of the total velocity field $\boldsymbol {u}$ away from the ‘background profile’ obtained by averaging in the $x$ and $z$ directions:

(3.1)\begin{equation} \boldsymbol{u'}(\boldsymbol{x}, t) = \boldsymbol{u}(\boldsymbol{x},t) - \frac{1}{L_xL_z}\int_{x=0}^{L_x}\int_{z=0}^{L_z} \boldsymbol{u}\, {\rm d}\kern0.06em x\,{\rm d}z. \end{equation}

As seen above, the flows we are considering are transient and, especially at points early during turbulence development, highly spatially intermittent in both the vertical and horizontal directions. This makes computing representative bulk turbulent statistics challenging, since any box average will contain an a priori unknown volume fraction of quiescent regions. Perhaps the most obvious approach is to average statistics over a time-evolving three-dimensional turbulent region identified according to some threshold criterion on a field variable such as the turbulent kinetic energy, however the disk-space requirements for saving fully three-dimensional flow fields are too demanding to achieve a satisfactory resolution in time by this method. Since turbulence is generally centred on the location of the initial strip of vertical vorticity at $y=0$, we suggest that a reasonable compromise for computing statistics is to consider vertical snapshots in the plane $y=0$ as illustrated in figure 5. Bulk statistics for the field variable $f(x,y,z,t)$ are then calculated by a plane average over the active turbulent region denoted by angle brackets $\langle {\cdot } \rangle _T$ with a subscript $T$,

(3.2)\begin{equation} \langle \,f(x,y,z) \rangle_T(t) = \frac{1}{L_x(L_z-z_{min})}\int_{x=0}^{L_x}\int_{z=z_{min}}^{L_z} f(x,0,z) \, {\rm d}\kern0.06em x\,{\rm d}z, \end{equation}

where $z_{min}$ is chosen to neglect the largely quiescent lower regions in simulations F07 and F1. Specifically, we take $z_{min}=L_z/2 = 7.14$ for simulation F07, $z_{min}=4.0$ for simulation F1 and $z_{min} = 0$ for simulation F2. This approach removes the majority of the vertical intermittency and, on average, minimises the horizontal intermittency, though fluctuations in bulk values are expected as localised regions of turbulence are swept around the central part of the domain by the spinning and merging columnar billow structures seen in figure 4. The trade-off for a high time resolution is therefore that computed statistics are at best only broadly representative of flow behaviour. Nonetheless, we find that they prove to be sufficiently effective for characterising the key features of the flows we consider, as we show below. It is also worth pointing out the clear practical similarities between this problem and its observational analogue when dealing with experimental and field data: particularly relevant for the motivation of this study are oceanographic microstructure observations, for which the present most complete spatio-temporal datasets come from a chain of moored sensors (see, e.g. Ivey, Bluteau & Jones Reference Ivey, Bluteau and Jones2018).

3.4. Stratified turbulence characteristics

At least qualitatively, turbulence in all simulations exhibits behaviour that is consistent with the LAST regime. It is natural to ask whether this behaviour, despite arising in a highly spatially localised horizontal shear flow whose fully turbulent behaviour has not previously been considered, can be characterised by an appropriately defined set of turbulence parameters as discussed in the introduction. As argued by Zhou & Diamessis (Reference Zhou and Diamessis2019) and Portwood et al. (Reference Portwood, de Bruyn Kops, Taylor, Salehipour and Caulfield2016), since the flow we are considering is highly transient, it is perhaps most appropriate to define the horizontal Reynolds and Froude numbers $Re_h$ and $Fr_h$ directly in terms of diagnosed energy-containing integral length scales $L$ and $L_v$ in the horizontal and vertical directions rather than indirectly appealing to the inertial scaling $\varepsilon \sim U^3/L$ often used in forced statistically stationary DNS (Brethouwer et al. Reference Brethouwer, Billant, Lindborg and Chomaz2007). To determine $L$ and $L_v$, we integrate the spectra of horizontal velocities in the plane $y=0$ using the method described in the appendix of Zhou & Diamessis (Reference Zhou and Diamessis2019). The results are shown in figure 6 for times $t>80$ following the onset of billow roll up. Figure 6(a) demonstrates that, despite the onset of turbulence, $L$ remains largely determined by the horizontal extent of the columnar billow structures throughout turbulent flow evolution. Billow pairing is clearly indicated by the rapid increase in $L$ (though the timing depends non-monotonically on $Fr_0$ due to the vertical decorrelation seen in figure 5). Consistent with the qualitative horizontal evolution shown in figure 4, this indicates that the horizontal kinetic energy and momentum balances are dominated by the emerging and interacting vortex columns. Looking at figure 6(b), the integral vertical scale $L_v$ is set by the layer growth that occurs prior to billow formation and notionally represents a dimensionless layer length scale. Despite some reasonable fluctuations (likely due to the manner in which it is calculated, as discussed above), $L_v$ maintains a similar magnitude over time, indicating that these layers are maintained throughout turbulence evolution. This behaviour of $L$ increasing whilst $L_v$ remains a similar magnitude is fairly typical of freely evolving flows in the LAST regime. In particular, we note the similarities with the freely evolving DNS of Riley & de Bruyn Kops (Reference Riley and de Bruyn Kops2003) initiated from a laminar Taylor–Green vortex configuration, which quickly develop strong horizontal vertically sheared layers that are maintained in structure and magnitude throughout the turbulent flow evolution despite the horizontal length scale increasing significantly. We define $Re_h$ and $Fr_h$ according to (1.1a,b), where the velocity scale $U=\mathcal {K}_h^{1/2}$ for $\mathcal {K}_h = \langle u'^2+v'^2\rangle _T/2$ and the average is taken as discussed above using (3.2). In dimensionless form we have

(3.3a,b)\begin{equation} Re_h=Re_0 \mathcal{K}_h^{1/2} L, \quad Fr_h = \frac{Fr_0 \mathcal{K}_h^{1/2}}{L}. \end{equation}

The evolution of each turbulent flow in $Re_h$-$Fr_h$ parameter space is shown in figure 7(a) by plotting $Re_h$ against $1/Fr_h$ as suggested by Brethouwer et al. (Reference Brethouwer, Billant, Lindborg and Chomaz2007), where their delineation of the ‘strongly stratified’ LAST regime according to $Fr_h<0.02$ and $Re_h Fr_h^{2}>1$ is shaded. According to our estimation of $Re_h$ and $Fr_h$, flows F1 and F07 can be seen in the figure to both narrowly fall within this regime during at least some of their evolution, whilst flow F2 remains only weakly affected by the stratification throughout. It is interesting to compare our results with Zhou & Diamessis (Reference Zhou and Diamessis2019), who, for a stratified turbulent wake behind a bluff body, derive a predicted turbulent Reynolds number $Re_h\sim Re_0Fr_0^{-2/3}$ based on the initial Reynolds and Froude numbers associated with the size of the body, using this to estimate empirically that strongly stratified turbulence is accessed when

(3.4)\begin{equation} Re_0 Fr_0^{-2/3} \gtrsim 5\times 10^3. \end{equation}

Figure 6.

Evolution of the horizontal and vertical integral length scales (a) $L$ and (b) $L_v$ for simulations F07 (blue lines), F1 (red lines) and F2 (orange lines). The directory including the data and Jupyter notebook for producing the figure can be found at https://multimediafluids.uam.elogim.com/S0022112024001216/JFM-Notebooks/files/fig6.

Figure 7.

(a) Trajectories of $Re_h$ vs $1/Fr_h$ for simulations F07, F1 and F2. The direction of time is from left to right, indicated by the grey arrow. Markers indicate the points at which the ‘fully turbulent’ snapshots considered in figures 8 and 9 are taken. The light blue shaded region denotes the ‘strongly stratified’ region of parameter space delineated by Brethouwer et al. (Reference Brethouwer, Billant, Lindborg and Chomaz2007) according to $Re_hFr_h^2>1$ and $Fr_h<0.02$. Panels (b,c) show the evolution of the buoyancy Reynolds number $Re_b$ and vertical Froude number $Fr_v$ for each simulation, where the dashed lines correspond to the time instant at which the markers are located in (a). The directory including the data and Jupyter notebook for producing the figure can be found at https://multimediafluids.uam.elogim.com/S0022112024001216/JFM-Notebooks/files/fig7.

Calculating a Reynolds and Froude number associated with the cylindrical billow structures that form with a diameter approximately $L_x/2=14.28$ in our simulations as $Re_0^{{{{{\dagger}}}}} = 14.28 Re_0$ and $Fr_0^{{{{{\dagger}}}}} = Fr_0/14.28$, we find that $Re_0^{{{{{\dagger}}}}} Fr_0^{{{{{{\dagger}}}}} -2/3}\approx 6100$ for simulation F07R1, $Re_0^{{{{{\dagger}}}}} Fr_0^{{{{{{\dagger}}}}} -2/3}\approx 4900$ for F1R1 and $Re_0^{{{{{\dagger}}}}} Fr_0^{{{{{{\dagger}}}}} -2/3}\approx 3100$ for F2. This provides a perhaps fortuitously good match with the prediction of Zhou & Diamessis (Reference Zhou and Diamessis2019) despite the different flows under consideration, suggesting at least a non-trivial level of underlying similarity in the dynamics. Indeed, the derivation of their criterion essentially relies on the assumptions that turbulence parameters are primarily dependent on the large-scale properties of the flow on top of which turbulence develops, a feature that we also observe here.

We also define the buoyancy Reynolds number $Re_b$ and the vertical Froude number $Fr_v$ (using the cyclic buoyancy period $2{\rm \pi} Fr_0$ as, for example, in the freely evolving flows of de Bruyn Kops & Riley Reference de Bruyn Kops and Riley2019; Zhou & Diamessis Reference Zhou and Diamessis2019) by

(3.5a,b)\begin{equation} Re_b = Re_0 Fr_0^2 \varepsilon, \quad Fr_v = \frac{ 2{\rm \pi} Fr_0 \mathcal{K}_h^{1/2}}{L_v}, \end{equation}

where $\varepsilon = \langle \partial _i u'_j\partial _i u'_j\rangle _T/Re_0$ is the dimensionless bulk turbulent kinetic energy dissipation rate. The time evolution of these quantities is shown in figure 7(b,c). Looking first at figure 7(b), both the maximum value of $Re_b$ and the time taken to reach this maximum depend strongly on the dimensionless buoyancy period $Fr_0$. In particular, just as with the prior development of the horizontal layers, the time taken for $Re_b$ to reach its first local maximum scales approximately linearly with $Fr_0$, demonstrating that turbulent transition is affected at leading order by the stratification. Once billow pairing has taken place, $Re_b$ starts to increase again for flows F1 and F2, likely due to the same mechanisms that caused the initial burst (though as discussed above, the simulations are stopped at this point as the dynamics start to become constrained by the periodicity). Peak values of $Re_b\approx 10$ and $Re_b\approx 5$ for the strongly stratified simulations F1 and F07 are high enough for the development of small-scale turbulent motions, though nonetheless imply at best a modest range of scales between the Ozmidov scale and the Kolmogorov length scale. Similar to the vertical length scale, the vertical Froude number plotted in figure 7(c) remains similar to its initial value determined by the shear layers that form prior to billow development and turbulence, and importantly, is $O(1)$ throughout turbulence evolution that is a characteristic signal of the stratified turbulence regime.

For each turbulent simulation, we look at a snapshot of the dissipation field $\varepsilon$ when $Re_b$ is close to its peak value whilst $Fr_h$ is either less than $0.02$ (for flows F07 and F1) or as small as possible (for flow F2). The corresponding locations in parameter space are shown by the markers in figure 7(a) with times indicated by the dashed lines in panels (b,c). Vertical plane snapshots of the dissipation field are plotted in figure 8. Looking first at snapshots in the plane $x=0$ shown in the left panels, we can see that a range of small scales are present in all simulations, though remain highly localised to the central strip of vorticity associated with the initial horizontal shear flow that is distorted in the vertical by the vertically sheared layers that form. Horizontal layering of dissipation is obvious in simulations F07 and F1, while for the more weakly stratified flow F2, the distribution is more sparse. This is also clear from looking at the snapshots in the plane $y=0$ in the right panels where horizontally localised patches of turbulence are apparent in flow F2. We suggest that the reason for the differences between the spatial patterns in $\varepsilon$ is due to the timing of the development of turbulence, which depends strongly on $Fr_0$, relative to the evolution of the background horizontal flow set by the vortex columns, which is similar across all simulations. The results are consistent with a picture of turbulence onset being highly localised to unstable regions that are advected and extended by the background horizontal flow, matching the behaviour in the vertical vorticity field shown in figure 4. When $Fr_0$ is larger, turbulence develops rapidly and $Re_b$ is largest whilst the unstable regions are local to the outer regions of the billows causing the patchiness seen in figure 8(f). For smaller $Fr_0$, the time scale of turbulence development is large enough relative to that of the horizontal flow such that it is distributed throughout the domain by horizontal motions before decaying.

Figure 8.

Vertical slices showing the dissipation field $\varepsilon$, with (a,c,e) taken in the plane $x=0$ and (b,d,f) taken in the plane $y=0$. Plots (a,b) are from simulation F07, (c,d) from F1 and (e,f) from F2. For simulations F07 and F1 that are vertically intermittent, only the fully turbulent section of the vertical domain is shown. The snapshots are taken at times corresponding to the location of the markers in parameter space in figure 7(a), indicated explicitly by the dashed lines in 7(b,c). Note the logarithmic colour scale. The directory including the data and Jupyter notebook for producing the figure can be found at https://multimediafluids.uam.elogim.com/S0022112024001216/JFM-Notebooks/files/fig8.

Finally, we look at the one-dimensional horizontal and vertical spectra of horizontal kinetic energy for each simulation at the times corresponding to the markers in figure 7(a). Our pseudo-spectral DNS resolve discrete streamwise Fourier modes $\exp ({\rm i}k_x)$ with wavenumbers $k_x=2{\rm \pi} n_x/L_x$ for integers $n_x$ between $0$ and $384$ (corresponding to the maximum wavenumber after dealiasing) and analagous vertical modes $\exp ({\rm i}k_z)$ with wavenumbers $k_z=2{\rm \pi} n_z/L_z$ for $n_z$ between $0$ and $170$. Spectra are calculated using the fully three-dimensional flow fields in the central turbulent region $-5< y<5$. Precisely,

(3.6)$$\begin{gather} E(k_x) = \frac{1}{2}\int_{y=-5}^{5}\left(\sum_{k_z}\hat{u}\hat{u}^* + \hat{v}\hat{v}^* \right) {\rm d}y, \end{gather}$$

(3.7)$$\begin{gather}E(k_z) = \frac{1}{2}\int_{y=-5}^{5}\left(\sum_{k_x}\hat{u}\hat{u}^* + \hat{v}\hat{v}^* \right) {\rm d}y, \end{gather}$$

where a hat denotes a Fourier transform in the $x$ and $z$ directions and $*$ denotes complex conjugation.

Figure 9(a) demonstrates that, whilst $Re_b$ is at, or close to, its maximum value, the horizontal spectrum for each simulation displays a plateau with a $k_x^{-5/3}$ slope indicating an incipient inertial range in the horizontal scales that is consistent with other vortically forced simulations at similar $Re_b$ by, e.g. Howland et al. (Reference Howland, Taylor and Caulfield2020), Augier et al. (Reference Augier, Billant and Chomaz2015) and Lindborg (Reference Lindborg2006). For sufficiently small $Fr_h$, the buoyancy wavenumber $k_b = (Fr_0 \mathcal {K}^{1/2})^{-1}$ becomes a dynamically relevant parameter (Billant & Chomaz Reference Billant and Chomaz2001) that, under the fundamental assumption that the vertical Froude number $Fr_v$ defined in (3.5a,b) is $O(1)$, scales with the wavenumber $2{\rm \pi} /L_v$ corresponding to the integral vertical scale $L_v$ set by the initial layers that form. The latter is easily recognisable as the location of the distinctive sharp peaks visible in figure 9(b). To draw attention to dynamics at scales influenced strongly by buoyancy, it is revealing to plot the vertical spectra as a function of the vertical wavenumber normalised by the Ozmidov wavenumber

(3.8)\begin{equation} k_O \equiv (Fr_0^{-3}/\varepsilon)^{1/2}. \end{equation}

In the LAST regime a range of anisotropic scales is expected to emerge between $k_b$ and $m_O$ where $E(k_z)\sim k_z^{-3}$, followed at wavenumbers larger than $k_O$ by an increase in gradient towards the isotropic slope $k_z^{-5/3}$ provided that $Re_b$ is sufficiently large, that is, there is a range of turbulent scales between $k_O$ and the Kolmogorov wavenumber $k_K = (Re_0^{-3}/\varepsilon )^{-1/4}$ that do not feel the influence of the stratification. In all of our simulations there are at most a few decades of scales within the entire range $k_b< k_O< k_K$ so it is not surprising that these subranges are not distinct in figure 9(b). However, clearly the range of wavenumbers between $k_O$ and the wavenumber of the vertical layers that form corresponding to the sharp peaks increases with decreasing $Fr_0$. For the lowest $Fr_0=0.71$, a hint of an $k_z^{-3}$ plateau emerges, which is consistent with being at least marginally inside the strongly stratified turbulent regime. As suggested by Almalkie & de Bruyn Kops (Reference Almalkie and de Bruyn Kops2012) and explored in detail by Maffioli (Reference Maffioli2017) and Howland et al. (Reference Howland, Taylor and Caulfield2020), due to the natural spatial variation in the vertical length scales associated with a given horizontal scale, the use of one-dimensional spectra results in the aliasing of some of the energy contained in small-scale horizontal motions onto relatively low vertical wavenumbers that acts to obscure the narrow $k_z^{-3}$ plateau at low to moderate values of $Re_b$.

Figure 9.

(a) Compensated horizontal (streamwise) spectra of kinetic energy $k_x^{5/3} E(k_x)$; (b) compensated vertical spectra of kinetic energy $k_z^3 E(k_z)$; (c) vertical spectra of kinetic energy as in (b), where this time the sum is taken over streamwise modes $k_x>0$. Lines are coloured consistent with the previous figures. Spectra are evaluated using the full three-dimensional flow fields. The directory including the data and Jupyter notebook for producing the figure can be found at https://multimediafluids.uam.elogim.com/S0022112024001216/JFM-Notebooks/files/fig9.

We can remove the low wavenumber peaks in the vertical spectra by noting that the shear layers that cause them extend horizontally across the entire domain. Therefore, by computing the power spectral density whilst taking the sum in (3.7) over $k_x\neq 0$, we can isolate the dynamics in the absence of the these largest-scale horizontal motions. The results are shown in figure 9(c), which somewhat more convincingly demonstrates the tendency of the vertical spectra towards an $k_z^{-3}$ plateau as $Fr_h$ decreases, suggesting the emergence of the associated buoyancy-inertial range. Additionally, as $Re_b$ increases from simulation F07 to simulation F2, the compensated spectra exhibit persistently steeper slopes towards vertical wavenumbers larger than $k_O$, displaying a plausible tendency towards the existence of an inertial range whose slope is indicated by the dashed line. In general, comparing the three spectra demonstrates a clear transition from the behaviour of stratified turbulence containing wavenumbers mostly smaller than $k_O$ (simulation F07) to that containing wavenumbers mostly at or larger than $k_O$ (simulation F2), whilst the overlap in the buoyancy and inertial subranges at wavenumbers close to $k_O$ highlights the difficulty in performing DNS with a large enough dynamic range to resolve both simultaneously. As of yet, DNS have not been able to reveal distinctly both the proposed isotropic and buoyancy-inertial vertical wavenumber range in stratified turbulence due to present computational limitations. Nonetheless, despite the time-dependent nature of the flow we are considering, the results here are at least suggestive that our DNS can transiently access a regime whose corresponding spectra behave in a manner consistent with arguably the most closely matching vortically forced flows of Maffioli (Reference Maffioli2017).

3.5. Mixing

We assume the irreversible mixing rate to be equivalent to the rate of destruction of buoyancy variance $\chi$ defined for density perturbations $\rho$ from a uniform background density gradient as

(3.9)\begin{equation} \chi = \frac{1}{Re_0 Pr Fr_0^2} \langle \partial_i \rho\partial_i\rho\rangle_T . \end{equation}

Technically, a precise definition describing the evolution of the potential energy associated with an appropriate background density profile as proposed by Winters & D'Asaro (Reference Winters and D'Asaro1996) should be invoked, however, the practical difficulties in implementing this method and the feasibility of direct measurements of $\chi$ in the ocean have led to (3.9) being regularly taken to be the definition of the mixing rate. Howland et al. (Reference Howland, Taylor and Caulfield2021) show that $\chi$ as defined above remains within 10 % of the ‘true’ mixing rate in a variety of forced stratified turbulent flows, with similarly good agreements being found for freely evolving vertically sheared flows (Lewin & Caulfield Reference Lewin and Caulfield2021).

An associated mixing efficiency $\eta$ and flux coefficient $\varGamma$ can be defined instantaneously or cumulatively (denoted with subscripts ‘i’ and ‘c’, respectively) as

(3.10a,b)$$\begin{gather} \eta_i(t) = \frac{\chi}{\chi + \varepsilon}, \quad \varGamma_i(t) = \frac{\chi}{\varepsilon}; \end{gather}$$

(3.11a,b)$$\begin{gather}\eta_c(t) = \frac{\displaystyle\int\nolimits_0^t \chi \,{\rm d}t}{\displaystyle\int\nolimits_0^t (\chi+\varepsilon)\, {\rm d}t}, \quad \varGamma_c(t) = \frac{\displaystyle\int\nolimits_0^t \chi \,{\rm d}t}{ \displaystyle\int\nolimits_0^t \varepsilon\, {\rm d}t}. \end{gather}$$

Note that $\eta = \varGamma /(1+\varGamma )$ in both the instantaneous and cumulative sense. Plots of $\chi$ and the associated dissipation $\varepsilon$ are shown in figure 10(a). The behaviour of $\varepsilon$ was previously discussed in the form of $Re_b= Re_0 Fr_0^{2}\varepsilon$, but is included as the dashed lines to facilitate a direct comparison with $\chi$ in the context of mixing. There is a notable similarity in the early behaviour of both $\chi$ and $\varepsilon$ in simulations F1 and F2, with both quantities increasing almost immediately following the formation of columnar billows. The onset of growth for simulation F07 is substantially more delayed. This may be attributed to the vortex structures that form in the braid region early during flow evolution for flows F1 and F2 but not for the more strongly stratified F07, as discussed in § 3.2. Corresponding behaviour is also marked in the plots of $\varGamma _i$ and $\varGamma _c$ shown in figure 10(b), where values early during turbulence development are similar for flows F1 and F2 and significantly larger than those of F07, representing larger mixing efficiencies during this stage. Indeed, the difference in the late-stage values of cumulative $\varGamma _c$ between flows F07 and F1 is likely due to this early behaviour. The pattern of larger values of the mixing efficiency being linked to coherent small-scale structures formed prior to turbulent transition has an analogue for the stratified vertical shear layer problem demonstrated by Mashayek, Caulfield & Peltier (Reference Mashayek, Caulfield and Peltier2013) and Lewin & Caulfield (Reference Lewin and Caulfield2021) in the context of coherent secondary shear instabilities that develop on the periphery of the (vertical) Kelvin–Helmholtz billow structures that form.

Figure 10.

Time evolution of (a) $\chi$ (solid lines) with $\varepsilon$ (dashed lines) superposed; (b) instantaneous and cumulative flux coefficients $\varGamma _i$ (solid lines) and $\varGamma _c$ (dotted lines). Each simulation is indicated by the colours consistent with previous figures. The directory including the data and Jupyter notebook for producing the figure can be found at https://multimediafluids.uam.elogim.com/S0022112024001216/JFM-Notebooks/files/fig10.

The behaviour of $\chi$ later during flow evolution closely follows that of $\varepsilon$, peaking at around the same time for each flow. However, the associated ratio $\varGamma _i$ adjusts once turbulence develops so that flows F07 and F1 match very closely with values of $\varGamma _i$ between $0.3$ and $0.4$, whilst flow F2 displays larger values of $\varGamma _i$ between $0.4$ and $0.5$. This is consistent with the results of Maffioli et al. (Reference Maffioli, Brethouwer and Lindborg2016), who find that, for stratified turbulent flows maintained by vortical forcing, an appropriately defined $\varGamma$ appears to exhibit a maximum value of around $0.5$ at moderate values of $Fr_h\approx 0.3$ before decreasing and eventually tending towards a constant value of $\varGamma \approx 0.35$ as $Fr_h$ decreases towards around $O(10^{-2})$. Based on the results from the previous section, the point at which $\varGamma _i$ becomes roughly equal to this asymptotic value (or rather, for our transient flows, the point at which $\varGamma _i$ settles on the same trajectory for flows F07 and F1) appears to be determined by the entrance of the flow into the LAST regime, controlled by $Fr_h$. The physical reason for this has been argued through the consideration of the dominant scaling balances in this regime, for example, by Maffioli et al. (Reference Maffioli, Brethouwer and Lindborg2016) and Garanaik & Venayagamoorthy (Reference Garanaik and Venayagamoorthy2019). Our values of $\varGamma _i$ and their dependence on $Fr_h$ are consistent with simulations forced with vertically uniform large-scale vortical modes reported by both Howland et al. (Reference Howland, Taylor and Caulfield2020) and Maffioli et al. (Reference Maffioli, Brethouwer and Lindborg2016). Values of (an appropriately defined) $\varGamma \approx 0.4$ were also reported in stratified turbulence resulting from the freely evolving Taylor–Green vortex configuration of Riley & de Bruyn Kops (Reference Riley and de Bruyn Kops2003), as well as in the simulations of Jacobitz & Sarkar (Reference Jacobitz and Sarkar2000) forced with uniform horizontal shear.

3.6. Length scales

Motivated by the recent developments in analysing the relationship between important dynamical length scales in the flow, we define the Ozmidov scale $L_O=2{\rm \pi} /k_O$, where $k_O$ is given by (3.8), as well as the Thorpe scale $L_T$ as the root-mean-square displacement of fluid parcels in each individual vertical column from their height in the corresponding profile that is sorted so that density monotonically decreases with depth. Precisely, $L_T=\sqrt {\langle \delta _T^2\rangle _T }$, where $\delta _T(\boldsymbol {x},t) = z(\rho ^*(\boldsymbol {x},t)) - z(\rho (\boldsymbol {x},t))$ is the displacement of a parcel at position $z(\rho ^*(\boldsymbol {x},t))$ in the sorted density field $\rho ^*(\boldsymbol {x},t)$ (obtained by rearranging the values of $\rho$ in each vertical column to be monotonically increasing with depth) from its position in the observed density field $z(\rho (\boldsymbol {x},t))$. The ratio of Ozmidov to Thorpe scales $R_{OT} = L_O/L_T$ has been proposed by Dillon (Reference Dillon1982) as a measure of the age of a turbulent event in the context of vertical shear-driven mixing as discussed in Smyth, Moum & Caldwell (Reference Smyth, Moum and Caldwell2001) and more recently, for example, in Lewin & Caulfield (Reference Lewin and Caulfield2021), with values of $R_{OT}\ll 1$ associated with early stage turbulence with larger corresponding $\varGamma$ and values $R_{OT}\gg 1$ with late-stage decay where $\varGamma \to 0$. For strongly stratified turbulence, Garanaik & Venayagamoorthy (Reference Garanaik and Venayagamoorthy2019) argue (using the Ellison scale discussed in Lewin & Caulfield (Reference Lewin and Caulfield2021) as a substitute for the Thorpe scale) that $R_{OT}$ may be a reasonable proxy for the Froude number $Fr_h$, though without considering developing turbulence do not cover values of $R_{OT}\ll 1$. They find $\varGamma$ roughly tends to a constant of $O(1)$ for $R_{OT}\sim 1$. Since our simulations contain both a period of young turbulence, and later of strongly stratified turbulence, we can investigate whether or not a corresponding signal exists in $R_{OT}$ for each of these stages.

The solid lines in figure 11(a) show the time evolution of $R_{OT}$ for each simulation. In general, for our simulations, $R_{OT}$ remains $O(1)$ throughout the mixing event that is broadly consistent with Garanaik & Venayagamoorthy (Reference Garanaik and Venayagamoorthy2019) who predict that $\varGamma$ should remain roughly constant at around $O(1)$ in this regime. Actually $R_{OT}\sim 1$ also corresponds to the optimal ‘Goldilocks mixing’ regime of Mashayek et al. (Reference Mashayek, Caulfield and Alford2021) within which $\varGamma \sim O(1)$. Though this is derived within the setting of mixing by Kelvin–Helmholtz instability, which is in principle entirely different to the LAST regime, we point towards the possibility of the Kelvin–Helmholtz instability paradigm existing locally at scales below the buoyancy scale set by the vertically sheared layers within LAST. All flows studied here follow a pattern of $R_{OT}$ being smaller during transition and then increasing once turbulence is fully developed, which is at least consistent with the ‘young’ turbulence picture of Mashayek et al. (Reference Mashayek, Caulfield and Peltier2017, Reference Mashayek, Caulfield and Alford2021), though we do not observe $R_{OT}\ll 1$ as would be needed to fully verify this picture. This is perhaps unsurprising as such a regime would normally be associated with values of $\varGamma \geq 1$ not seen here. Whether or not there are mechanisms by which strongly stratified flows can produce local overturns with much larger scale than the global Ozmidov scale $L_O$ and, therefore, access a regime with $R_{OT}\ll 1$ and $\varGamma \gtrsim 1$ remains an open question.

Figure 11.

(a) Time evolution of the ratio $R_{OT}=L_O/L_T$ of Ozmidov to Thorpe scales (solid lines, left axis), where the corresponding shaded regions (right axis) represent the fraction of the domain that has a Thorpe displacement $\delta _T>0$. Lines are coloured as in previous figures. (b) Plots of the p.d.f. of $\log _{10}\varepsilon$ over the turbulent domain at times corresponding to the markers in figure 7, also indicated by the dashed lines in panel (a). The directory including the data and Jupyter notebook for producing the figure can be found at https://multimediafluids.uam.elogim.com/S0022112024001216/JFM-Notebooks/files/fig11.

An alternative, though closely related, view comes considering the fraction of the domain that is statically unstable, or equivalently, has $|\delta _T|>0$. This is represented by the background shaded regions colour matched to the lines in figure 11(a). The behaviour is as expected, with the fraction of overturning increasing with increasing $Fr_0$. Flow F2 peaks at around $50\,\%$ overturning fraction, F1 around $35\,\%$ and F07 at just under $25\,\%$. In their classification of dynamically distinct regions in forced stratified turbulence, Portwood et al. (Reference Portwood, de Bruyn Kops, Taylor, Salehipour and Caulfield2016) determine turbulent patches as having an overturning fraction greater than $35\,\%$, layers as having a fraction between $20\,\%$ and $30\,\%$ and quiescent regions as having less than $20\,\%$. This roughly classifies F2 as a turbulent patch and F1 and F07 as layers, which is consistent with the entering of the latter two into the LAST regime accompanied by a change in the mixing properties. We have not investigated whether or not the turbulence we are considering in each simulation can be further subdivided into smaller patches with dynamically distinct signatures, which would require a careful filtering procedure. However, figure 11(b) offers some evidence that, at least when turbulence is fully developed, this is not necessary for our purposes. Plotted are the p.d.f.s (equivalent to normalised frequency distributions) of $\log _{10}(\varepsilon )$ over the domain at the time indicated by the vertical dashed lines in panel (a). We note that they are close to log-normal as expected for stratified turbulent flows with a sufficiently broad inertial range (de Bruyn Kops Reference de Bruyn Kops2015), and in particular, lack the right ‘shoulder’ feature that was a characteristic signature of high $Re_b$ patches in the flows studied by Portwood et al. (Reference Portwood, de Bruyn Kops, Taylor, Salehipour and Caulfield2016) (though flows F07 and F2 exhibit slight shoulders on the left of the distribution indicating the possible presence of small quiescent regions).