2 Mathematical models

We consider the two configurations in figures 1(a) and 1(b), where both phases consist of Newtonian fluids with constant density $\unicode[STIX]{x1D70C}_{i}$ and viscosity $\unicode[STIX]{x1D707}_{i}$ (the subscript $i=1,2$ differentiates between the two phases), and where $g$ designates gravitational acceleration. The surface tension $\unicode[STIX]{x1D70E}$ will be assumed constant except in § 9, where we will study the additional effect of thermal Marangoni stresses. We assume that the (dimensionless) film thickness $h$ is small compared to the (dimensionless) wavelength $\unicode[STIX]{x1D6EC}$ and use the weighted residual integral boundary layer (WRIBL) model of Dietze & Ruyer-Quil (Reference Dietze and Ruyer-Quil2013), which accounts for inertia, longitudinal viscous diffusion and inter-phase coupling. In dimensionless form, this reads:

where $i$ and $j$ are to be permuted through the phase indices 1 and 2 using Einstein summation. In (2.1), $h$ designates the film thickness, $q_{i}$ the phase-specific flow rate per unit width and $\unicode[STIX]{x1D705}=\unicode[STIX]{x2202}_{xx}h$ the interfacial curvature (at second order in the long-wave expansion). Following Yiantsios & Higgins (Reference Yiantsios and Higgins1989), we have used for non-dimensionalization the length scale ${\mathcal{L}}=h_{0}$ , corresponding to the average film thickness, the velocity scale ${\mathcal{U}}=|\unicode[STIX]{x0394}\unicode[STIX]{x1D70C}|gh_{0}^{2}/\unicode[STIX]{x1D707}_{1}$ with $\unicode[STIX]{x0394}\unicode[STIX]{x1D70C}=\unicode[STIX]{x1D70C}_{1}-\unicode[STIX]{x1D70C}_{2}$ , obtained by balancing viscous drag and gravity, and the time scale ${\mathcal{T}}={\mathcal{L}}/{\mathcal{U}}=\unicode[STIX]{x1D707}_{1}/|\unicode[STIX]{x0394}\unicode[STIX]{x1D70C}|/g/h_{0}$ . This choice yields the Reynolds number $\mathit{Re}={\mathcal{U}}h_{0}|\unicode[STIX]{x0394}\unicode[STIX]{x1D70C}|/\unicode[STIX]{x1D707}_{1}$ and the Bond number $\mathit{Bo}=|\unicode[STIX]{x0394}\unicode[STIX]{x1D70C}|gh_{0}^{2}/\unicode[STIX]{x1D70E}$ , which are completed by the density and viscosity ratios $\unicode[STIX]{x1D6F1}_{\unicode[STIX]{x1D70C}}=\unicode[STIX]{x1D70C}_{2}/\unicode[STIX]{x1D70C}_{1}$ and $\unicode[STIX]{x1D6F1}_{\unicode[STIX]{x1D707}}=\unicode[STIX]{x1D707}_{2}/\unicode[STIX]{x1D707}_{1}$ . At places, we will also relate the dimensionless horizontal coordinate $x$ to the dimensionless wavelength  $\unicode[STIX]{x1D6EC}$ .

In (2.1c ), the sign of the gravity term (first term on the right-hand side) is positive for the suspended film (figure 1 a) and negative for the gas film (figure 1 b). The coefficients $F_{ij}$ , $G_{ij}$ , $C_{ij}$ , $S_{j}$ , $J_{j}$ , $K_{j}$ and $M_{j}$ are known functions of $h$ and the domain height  $D$ . Our coefficients are slightly different than in Dietze & Ruyer-Quil (Reference Dietze and Ruyer-Quil2013), as we impose a slip boundary at $y=D$ ( $\unicode[STIX]{x2202}_{y}u|_{D}=v|_{D}=0$ ) instead of a wall (the coefficient definitions have been provided in a Mathematica^® file in the supplementary material). The slip boundary is sufficiently far to prevent influencing the large humps produced by the primary instability, i.e.  $D\gg 1$ , and sufficiently close to satisfy the long-wave approximation in both layers, i.e.  $D\ll \unicode[STIX]{x1D6EC}$ . We have verified for both the suspended film (figure 1 a, $D=4=0.16\unicode[STIX]{x1D6EC}$ ) and the gas film (figure 1 b, $D=10=0.04\unicode[STIX]{x1D6EC}$ ) that the quasi-steady state reached prior to sliding is virtually insensitive to  $D$ . In this sense, our simulations mimic an unconfined outer phase.

We solve (2.1) numerically, using second-order central differences for spatial and the Crank–Nicolson method for time discretization, and linearizing nonlinear terms around the old time step. In terms of boundary conditions, we distinguish two cases: (i) periodic simulations on a domain of length  $\unicode[STIX]{x1D6EC}$ , where $\unicode[STIX]{x2202}_{x^{i}}h|_{x=0}=\unicode[STIX]{x2202}_{x^{i}}h|_{x=\unicode[STIX]{x1D6EC}}$ , $\unicode[STIX]{x2202}_{x^{i}}q|_{x=0}=\unicode[STIX]{x2202}_{x^{i}}q|_{x=\unicode[STIX]{x1D6EC}}$ and the film is free to slide, and (ii) symmetric simulations on a domain of length $\unicode[STIX]{x1D6EC}/2$ , where $\unicode[STIX]{x2202}_{x}h=\unicode[STIX]{x2202}_{xxx}h=0$ (implying $q=0$ ) at $x=0$ and $\unicode[STIX]{x1D6EC}/2$ , in order to capture the non-sliding quasi-steady solution. The dimensionless wavelength $\unicode[STIX]{x1D6EC}$ is set to the most-amplified wavelength of the Rayleigh–Taylor instability for a passive outer phase $\unicode[STIX]{x1D6EC}=\sqrt{2}\unicode[STIX]{x1D6EC}_{c}$ , where $\unicode[STIX]{x1D6EC}_{c}=2\unicode[STIX]{x03C0}/\sqrt{\mathit{Bo}}$ is the corresponding cutoff wavelength. This quantity is convenient because it is known in closed form and, for all our simulations, it differs by less than 0.5 per cent from the actual most-amplified wavelength (i.e. for an active outer phase). We will loosely refer to $\sqrt{2}\unicode[STIX]{x1D6EC}_{c}$ as the most-amplified wavelength of the Rayleigh–Taylor instability.

For the suspended water film, which we mainly focus on, we have used the full model (2.1) as a reference to identify those ingredients that are sufficient for the sliding instability, i.e. gravity, surface tension and cross-wise ( $y$ -direction) viscous diffusion. Retaining only these ingredients in (2.1), we obtain the following simplified model:

where the outer phase is neglected ( $\unicode[STIX]{x1D6F1}_{\unicode[STIX]{x1D70C}}=\unicode[STIX]{x1D6F1}_{\unicode[STIX]{x1D707}}=0$ ) and thus the phase index has been dropped. The Bond number reduces to $\mathit{Bo}=\unicode[STIX]{x1D70C}_{1}gh_{0}^{2}/\unicode[STIX]{x1D70E}$ and remains the sole dimensionless group. We will use (2.2) for our stability analysis and most of the discussions in §§ 3–7. We point out that it is the same as the lubrication equation in Lister et al. (Reference Lister, Rallison, King, Cummings and Jensen2006b ).

In § 9, we will study the effect of additional thermal Marangoni stresses due to heating the suspended film from the bounding wall, assuming $\unicode[STIX]{x2202}_{T}\unicode[STIX]{x1D70E}<0$ . To account for this, equation (2.2b ) needs to be extended:

where $\mathit{Ma}=\unicode[STIX]{x2202}_{T}\unicode[STIX]{x1D70E}(T_{w}-T_{\infty })/\unicode[STIX]{x1D70E}$ designates a modified Marangoni number, $\unicode[STIX]{x1D703}|_{h}=(T|_{h}-T_{w})/(T_{w}-T_{\infty })=-\mathit{Bi}\,h/(1+\mathit{Bi}\,h)$ the dimensionless film surface temperature, $\mathit{Bi}=\mathscr{H}h_{0}/k_{1}$ the Biot number and $T_{w}$ and $T_{\infty }$ the wall and ambient temperature. The Biot number contains the interfacial heat transfer coefficient $\mathscr{H}$ and the thermal conductivity  $k_{1}$ . We point out that (2.3) was previously used in Alexeev & Oron (Reference Alexeev and Oron2007), where the film was cooled from the wall, and thus Marangoni stresses were stabilizing ( $\mathit{Ma}>0$ ) in terms of the primary instability, as opposed to our case (we will set $\mathit{Ma}=-0.2$ ).

In § 8, we will show that a very thin gas film underneath a (much more viscous) liquid layer is also prone to the sliding instability. For this configuration, we will use the full model (2.1) in order to account for viscous coupling with the outer phase.

All our simulations were started from a symmetric initial condition:

with a very small relative perturbation amplitude $\unicode[STIX]{x1D700}=0.0009$ . When using the full model (2.1), the initial flow rate $q|_{t=0}$ was computed from the inertialess limit (2.2b ) using (2.4). Our initial condition ensures that sliding, if it occurs, does so spontaneously.

3 Kinematics of the sliding instability

For the time being, we focus on the configuration of a suspended water film of average thickness $h_{0}=1~\text{mm}$ and $\mathit{Bo}=0.134$ which is surrounded by air, as illustrated in figure 1(a) (see caption for other properties). We have simulated the evolution of this film with the full (2.1) and simplified (2.2) models, starting from the fully symmetrical initial condition (2.4) (perturbation amplitude $\unicode[STIX]{x1D700}=0.0009$ ), using periodic boundary conditions on a domain spanning the wavelength $\unicode[STIX]{x1D6EC}=2\sqrt{2}\unicode[STIX]{x03C0}/\sqrt{\mathit{Bo}}=24.2$ and discretized with 1001 grid points. Figure 3 shows how the film evolves from the symmetrical initial state to an asymmetrical sliding state through four characteristic stages, which are also discernible in figure 2. In contrast to figure 1(a), gravity points upward in figures 2 and 3.

Figure 3.

Kinematics of the sliding sequence for the suspended water film (see figure 1 a): $h_{0}=1~\text{mm}$ , $\mathit{Bo}=0.134$ , $\unicode[STIX]{x1D6EC}=24.2$ . In (a,b), dashed lines correspond to the full model (2.1), solid lines to (2.2), and the red solid line in (b) to a simulation of the full Navier–Stokes equations (discussed at the end of § 3). Symbols refer to characteristic stages in (c–f), where profiles evolve from dashed to dot–dashed lines. The horizontal coordinate $x$ has been related to the (dimensionless) domain length  $\unicode[STIX]{x1D6EC}$ . (a) Time trace of the trough position (left trough after buckling); (b) film thickness at trough position corresponding to (a); (c) surface profiles during first stage: progressive growth; (d) flattening and buckling of the film surface; (e) quasi-steady two-trough shape (see also supplementary movie 1); (f) loss of symmetry and sliding (see also supplementary movie 2).

Figures 3(a) and 3(b) represent time traces of the position $x_{min}$ and thickness $h_{min}$ of the film surface minimum. Different symbols refer to different evolution stages, which are illustrated through surface profiles in figures 3(c)–3(f). Data were obtained with the inertialess model (2.2), except for the dashed lines in figures 3(a) and 3(b), which correspond to the full model (2.1), and the red line in figure 3(b), which was obtained from a simulation of the full Navier–Stokes equations (detailed at the end of § 3). For convenience, we have normalized $x$ with the domain length  $\unicode[STIX]{x1D6EC}$ . The large values of $t$ in figures 3(a) and 3(b) occur because sliding sets in very late in terms of the typical time scales of viscous capillary–gravity flows (Yiantsios & Higgins Reference Yiantsios and Higgins1989; Lister et al. Reference Lister, Rallison, King, Cummings and Jensen2006b ; Glasner Reference Glasner2007).

The first three evolution stages in figure 3 have been discussed in detail by Yiantsios & Higgins (Reference Yiantsios and Higgins1989) and so we recap them only briefly. In the first stage (crosses in figures 3 a and 3 b), growth of the surface perturbation is progressive and the corresponding spatial profiles (figure 3 c) exhibit a single trough that increasingly thins while remaining in the middle of the domain. During the second stage (filled circles in figures 3 a and 3 b), the film surface around the trough flattens and then buckles upon further approaching the wall, forming two secondary troughs enclosing a secondary hump in the middle (figure 3 d, where the range of the abscissa has been reduced). In figures 3(a) and 3(b), it is the left secondary trough that is tracked from the buckling event onwards. This secondary trough (and its twin on the other side) moves outward and increasingly thins. At the same time, the secondary hump in the middle grows more pronounced. This evolution continues for some time but increasingly slows down, until the film reaches a quasi-steady state (diamonds in figures 3 a and 3 b), constituting the third evolution stage. Corresponding surface profiles in figure 3(e) change only very slightly over a considerable time interval. In particular, the locations of the secondary troughs remain virtually fixed. The supplementary movie 1 shows the first three evolution stages in action (the ordinate has been scaled logarithmically to highlight the secondary troughs).

In the fourth evolution stage (open circles in figures 3 a and 3 b), the quasi-steady buckled-film surface spontaneously loses its symmetry, causing the entire film to slide to the left (figure 3 f). The supplementary movie 2 shows these events in action (the ordinate has again been scaled logarithmically). The speed of the sliding motion, based on the displacement of the right trough in figure 3(f), is roughly $c=1.2\times 10^{-4}$ (corresponding to a dimensional value of $1.2~\text{mm}~\text{s}^{-1}$ ).

We now focus on the loss of symmetry with the help of figure 4 by comparing our periodic simulation (dashed and dot–dot–dashed lines in figures 4 a–4 c) with a symmetric simulation on a domain spanning $\unicode[STIX]{x1D6EC}/2$ (solid lines in figures 4 a–4 c). Although the symmetric simulation represents only one of the secondary troughs, we have produced the other by mirroring the simulation data to the other side. Comparing the two solutions in figure 4(a), we conclude that symmetry is lost at $t\approx 7\times 10^{4}$ , when the periodic simulation departs from the symmetric one in that both the left and right secondary troughs move to the left. At the same time, the film thickness at the left secondary trough starts to decrease, while it increases at the right trough (figure 4 b). During the leftward migration of the secondary troughs, the film is peeled off on their right and deposited on their left, in accordance with the motion described by Lister et al. (Reference Lister, Rallison, King, Cummings and Jensen2006b ). This is comparable to a track vehicle putting down its chains while moving forward.

At the left trough, deposition occurs faster than peeling and thus the trough becomes increasingly flat, whereas the opposite occurs at the right trough, which becomes increasingly curved. Quantitative evidence of this is shown in figure 4(c), which plots time traces of the surface curvature $\unicode[STIX]{x2202}_{xx}h$ at the two troughs. Comparing the periodic with the symmetric data after the film surface has buckled (unshaded region), shows that, at the onset of sliding ( $t\approx 7\times 10^{4}$ ), the curvature at the left secondary trough (dot–dot–dashed line) suddenly decreases, i.e. the trough flattens, while it increases at the right secondary trough (dashed line). By contrast, $\unicode[STIX]{x2202}_{xx}h$ in the symmetric simulation (solid line) never ceases to increase, as the film converges to its final equilibrium state shown in figure 4(e).

Figure 4.

Symmetry loss of the quasi-steady state in figure 3(e). Panels (a-c) compare the periodic simulation (discontinuous lines) to a symmetric simulation on a domain spanning $\unicode[STIX]{x1D6EC}/2$ (solid lines). Dashed lines correspond to the right secondary trough and dot–dot–dashed lines to the left secondary trough. Open circles mark time points highlighted in figure 3(a). (a) Trough positions; (b) film thickness at the troughs; (c) surface curvature $\unicode[STIX]{x2202}_{xx}h$ at the troughs; (d) film surface in the two-trough region immediately after symmetry loss ( $t=6.8\times 10^{4}$ until $t=10.5\times 10^{4}$ ); (e) symmetric simulation showing evolution to final equilibrium state (4.1), represented as a dashed line. The supplementary movie 2 shows the loss of symmetry and sliding motion in action.

To verify that the simplified model (2.2) does not preclude any dominant physical effects, we return to figures 3(a) and 3(b), where we have also included results obtained with the full model (2.1), represented with dashed lines. We see that both calculations evolve exactly to the same quasi-steady state (diamonds on solid line). After that, sliding sets in slightly later for the full-model calculation, but the ensuing evolution is the same. However, before reaching the quasi-steady regime, the full model produces a number of oscillations that consist in the secondary troughs periodically moving toward and away from each other (see figure 5). These oscillations result from inertia, but they do not cause any loss of symmetry before the quasi-steady state has been reached.

Figure 5.

Inertia-driven oscillations of the buckled film in figure 3. Solid lines represent data obtained from the periodic simulation of the full model (2.1) and diamonds represent a corresponding DNS of the Navier–Stokes equations using the code Gerris (Popinet Reference Popinet2009). (a) Time traces of the secondary trough positions (DNS data are only shown at two time points); (b) surface profiles at the two characteristic time points marked by diamonds in figure 5(a).

We have validated our full model (2.1) with a direct numerical simulation (DNS) based on the full Navier–Stokes equations (diamonds in figure 5 and red line in figure 3 b). The DNS was performed with the finite-volume code Gerris (Popinet Reference Popinet2009), using periodic boundary conditions and adaptive grid refinement. Grid refinement was limited to a minimum cell size of $\unicode[STIX]{x0394}x=\unicode[STIX]{x0394}y=0.004$ . As a result, the DNS data in figure 3(b) can be trusted as long as $h_{min}\geqslant 0.016$ , when the thickness of the secondary troughs is resolved by at least 4 grid points. We have continued our DNS past this point and, although the accuracy of the ensuing data is open to discussion, they do exhibit the same sliding behaviour as the full model (dashed line in figure 3 b), albeit earlier.

In figures 2–4, sliding occurs in leftward direction, but the film is equally likely to slide to the right. The direction in a given computational run is decided by uncontrollable numerical noise, which perturbs the unstable film and sets off the sliding motion. We will see later, from our linear stability analysis, just what sort of perturbation in this noise is needed for the sliding to occur and how sensitive the sliding onset is with respect to the noise level.

4 Draining mechanisms

In the absence of noisy perturbations, the buckled film in figure 3(e) would evolve in a perfectly symmetrical manner until attaining its final equilibrium state. In our case, where $\unicode[STIX]{x1D6EC}<2\unicode[STIX]{x1D6EC}_{c}$ , this final state consists of two sinusoidal drop halves spanning the cutoff wavelength $\unicode[STIX]{x1D6EC}_{c}=2\unicode[STIX]{x03C0}/\sqrt{\mathit{Bo}}$ of the Rayleigh–Taylor instability and separated by a zero-thickness film segment (Hammond Reference Hammond1983). It is obtained by setting (2.2b ) to zero and the left half of this symmetric solution is given by:

The final state is represented with a blue line in figures 6(a), 6(b), 6(e) and 6(f). We use figure 6 to discuss the draining mechanisms driving evolution toward this state. It represents surface plots and profiles of the driving pressure gradient at four characteristic time points, as obtained from the symmetric simulation of (2.2) for the conditions in figure 3. In the surface plots 6(a), 6(b), 6(e) and 6(f), the red line represents the initial condition (2.4). In the corresponding pressure gradient plots (figure 6 c,d,g,h), solid lines represent the full pressure gradient $\unicode[STIX]{x2202}_{x}p$ , dot–dot–dashed lines the contribution of gravity $\unicode[STIX]{x2202}_{x}p|_{g}$ , and dashed lines the capillary contribution $\unicode[STIX]{x2202}_{x}p|_{\unicode[STIX]{x1D70E}}$ according to:

The driving pressure gradient $\unicode[STIX]{x2202}_{x}p$ is always counteracted by viscous drag, which moderates the action of $\unicode[STIX]{x2202}_{x}p$ on the flow rate through the term  $h^{3}/3$ :

Figure 6.

Symmetric simulation of the suspended water film (figure 1 a): $h_{0}=1~\text{mm}$ , $\mathit{Bo}=0.134$ . Capillary and gravity-induced drainage driving the film from the initial condition to the final equilibrium state. (a,b,e,f) Surface profiles at four characteristic time points ( $t=427$ , 641, 1068, $6.4\times 10^{4}$ ). Solid lines: solution of (2.2) using symmetry conditions on a domain of length $\unicode[STIX]{x1D6EC}/2$ (data were mirrored onto the full-wavelength domain); red and blue lines: initial condition (2.4) and final equilibrium state (4.1); (c,d,g,h) profiles of the pressure gradient. Solid line: full pressure gradient (4.2); dot–dot–dashed: gravity-induced contribution $\unicode[STIX]{x2202}_{x}p|_{g}$ ; dashed: capillary contribution $\unicode[STIX]{x2202}_{x}p|_{\unicode[STIX]{x1D70E}}$ . Open circles in (e–h) highlight loci of the secondary troughs.

In order for the weakly deformed film in figure 6(a) to reach its final equilibrium state, the liquid contained underneath the trough region needs to be fully drained to the main hump. During the first evolution stage (figure 6 a), the pressure gradient for this is provided by gravity, which symmetrically drives liquid outward from underneath the initial single trough (dot–dot–dashed line in figure 6 c), while capillarity counteracts this drainage (dashed line in figure 6 c).

When the trough becomes sufficiently thin (figure 6 b), viscous drag causes the film surface there to flatten (Yiantsios & Higgins Reference Yiantsios and Higgins1989), and this attenuates the gravity-induced flow rate contribution $q|_{g}=(h^{3}/3)\unicode[STIX]{x2202}_{x}h$ . Gravity alone can no longer drain sufficient liquid from underneath the trough to accommodate the growth of the main hump, where viscous drag is much weaker and the initial balance of power is maintained (figure 6 d). At the same time, flattening of the trough region increases $|\unicode[STIX]{x2202}_{xxx}h|$ such that the capillary flow rate contribution $q|_{\unicode[STIX]{x1D70E}}=(h^{3}/3/\mathit{Bo})\unicode[STIX]{x2202}_{xxx}h$ now helps and even dominates drainage there.

As the trough becomes even thinner (figure 6 e), capillary drainage needs to further increase, in order to continue draining sufficient liquid to the main hump, and this eventually requires the film surface to buckle (Yiantsios & Higgins Reference Yiantsios and Higgins1989). Drainage in the region between the newly formed secondary troughs is now entirely driven by capillarity, the sign of $\unicode[STIX]{x2202}_{x}p|_{g}=-\unicode[STIX]{x2202}_{x}h$ having changed due to the inversion of surface slope (figure 6 g).

In figure 6(f), showing the quasi-steady state, the film has almost attained its final equilibrium state (blue line). In particular, the width of the main hump has reached the cutoff wavelength of the Rayleigh–Taylor instability $\unicode[STIX]{x1D6EC}_{c}=2\unicode[STIX]{x03C0}/\sqrt{\mathit{Bo}}$ and thus the position of the secondary troughs is fixed from now on. To fully reach the final state, all liquid remaining in the secondary hump needs to be drained into the main hump through the very thin secondary troughs. This drainage is entirely driven by capillarity, as $\unicode[STIX]{x2202}_{x}p$ and $\unicode[STIX]{x2202}_{x}p|_{\unicode[STIX]{x1D70E}}$ are virtually identical around the troughs (dashed and solid lines in figure 6 h).

Moreover, the pressure gradient is considerable only around the secondary troughs, where it exhibits large-magnitude pulses (figure 6 h). By contrast, $\unicode[STIX]{x2202}_{x}p$ is almost exactly zero within the main hump and thus the latter is virtually in static equilibrium. This results from the slowness of the drag-limited draining process in the trough region, allowing the main hump to always relax toward equilibrium (Hammond Reference Hammond1983). In fact, the main hump closely follows the sinusoidal profile given by (4.1a ) (blue line in figure 6 f), which is the neutral mode of the Rayleigh–Taylor instability at the cutoff wavelength  $\unicode[STIX]{x1D6EC}_{c}$ . It is thus neutrally stable toward a pure translation and stable toward any other volume-preserving perturbation and can be displaced with minimal energy input.

Within the secondary hump, $\unicode[STIX]{x2202}_{x}p$ is also very small but its magnitude increases noticeably toward the secondary troughs (figure 6 h). According to Hammond (Reference Hammond1983), the secondary hump continually adjusts to the short sinusoidal equilibrium shape known as a lobe (Lister et al. Reference Lister, Morrison and Rallison2006a ). However, such a lobe exhibits a finite slope at its extremities and thus cannot connect smoothly to the secondary troughs, as opposed to the equilibrium solution of the main hump (4.1a ), the slope of which decreases to zero at the troughs. Also, the pressure within the lobe is higher than that within the main hump (Lister et al. Reference Lister, Rallison, King, Cummings and Jensen2006b ). Therefore, lobes eventually drain out completely and the final state of the film cannot include lobes (Yiantsios & Higgins Reference Yiantsios and Higgins1989).

Further change of the quasi-steady state in figure 6(f) is driven by capillary pressure gradients $\unicode[STIX]{x2202}_{x}p|_{\unicode[STIX]{x1D70E}}=-(1/\mathit{Bo})\unicode[STIX]{x2202}_{xxx}h$ , which are governed by surface curvature variations. When symmetry is imposed, they drive the film toward its final equilibrium state by symmetrically draining the remaining liquid from underneath the secondary hump, otherwise they drive the sliding motion of the film (Lister et al. Reference Lister, Morrison and Rallison2006a ).

5 Frozen-time linear stability analysis

We have shown in figures 4(a)–4(c) that the quasi-steady buckled-film surface obtained from our periodic simulation loses symmetry roughly at $t=7\times 10^{4}$ , when the left secondary trough starts to thin and flatten and the right trough starts to thicken and curve with respect to a fully symmetric simulation. Figure 7(a) represents surface profiles corresponding to this time, as obtained from the periodic (crosses) and symmetric (solid line) simulation, respectively (symmetric data were mirrored onto the full length of the periodic domain). We have checked that both simulations have fully converged in terms of grid resolution (1001 grid points per wavelength  $\unicode[STIX]{x1D6EC}$ ).

We now perform a linear stability analysis upon the perfectly symmetric surface profile in figure 7(a) (solid line). We designate this profile as base profile and denote it $H(x)$ , neglecting its temporal evolution, which is extremely slow. This amounts to a so-called frozen-time approach. Next, we perturb the base profile infinitesimally, introducing the linear film thickness perturbation $h^{\ast }(x,t)$ , which is assumed to grow exponentially:

Figure 7.

Frozen-time linear stability analysis of the suspended water film ( $h_{0}=1~\text{mm}$ , $\mathit{Bo}=0.134$ ) at the time of symmetry loss in figure 4(a): $t=7\times 10^{4}$ . Open circles mark loci of secondary troughs. (a) Solid line: base state $H(x)$ obtained from symmetric simulation on domain of length $\unicode[STIX]{x1D6EC}/2$ (501 grid points) and mirrored onto full-wavelength domain; crosses: profile from periodic simulation on domain of length $\unicode[STIX]{x1D6EC}$ (1001 grid points) after loss of symmetry; (b) linear stability results. Solid lines: most-unstable asymmetric ( $A_{j}=0$ ) and symmetric ( $B_{j}=0$ ) eigenfunctions ${\hat{h}}(x)$ (5.3) obtained from linear stability analysis of the perfectly symmetric profile in (a) (solid line there); asterisks: actual perturbation associated with symmetry loss, i.e. difference between periodic and symmetric profiles in (a); red-dashed line: perturbation resulting from pure translation of base profile $H(x)$ with speed $c$ , i.e.  $\unicode[STIX]{x2202}_{t}h=-c\unicode[STIX]{x2202}_{x}H$ ; (c) perturbation profiles from (b) normalized with local film thickness $H(x)$ ; (d) time derivative of surface curvature $\unicode[STIX]{x2202}_{t}(\unicode[STIX]{x2202}_{xx}h)=\unicode[STIX]{x1D702}\unicode[STIX]{x2202}_{xx}{\hat{h}}$ associated with most-unstable eigenfunctions in (b).

Upon inserting (5.1) into (2.2) and linearizing in terms of  ${\hat{h}}$ , an eigenvalue problem with eigenvalue $\unicode[STIX]{x1D702}$ and eigenfunction ${\hat{h}}$ is obtained:

We choose a Fourier series ansatz with $N=100$ for the eigenfunction  ${\hat{h}}$ :

and solve the eigenvalue problem with the Galerkin approach (Boyd Reference Boyd1989). We then identify the most-unstable (greatest  $\unicode[STIX]{x1D702}$ ) eigenfunctions for two perturbation types: (i) asymmetric perturbations, when $A_{j}=0$ ; and (ii) symmetric perturbations, when $B_{j}=0$ .

Figure 7(b) represents the thus obtained eigenfunctions (solid black lines). The asymmetric eigenfunction is associated with a positive eigenvalue $\unicode[STIX]{x1D702}=1.8\times 10^{-4}$ . This proves that the film is subject to a symmetry-breaking secondary instability, secondary in the sense that it occurs after the primary Rayleigh–Taylor instability has developed. We call this instability sliding instability. It is associated with a very particular eigenfunction. In fact, the unsymmetric eigenfunction in figure 7(b) is the sole unstable unsymmetric eigenmode. Moreover, it is the sole unstable eigenmode altogether, as all symmetric eigenfunctions are stable, the greatest symmetric eigenvalue being negative $\unicode[STIX]{x1D702}=-1.9\times 10^{-5}$ (figure 7 b).

Crosses in figure 7(b) represent the actual perturbation associated with the loss of symmetry of the periodic simulation. This is easily obtained by taking the difference between the periodic and symmetric surface profiles in figure 7(a). Good agreement in figure 7(b) between the actual perturbation (crosses) and the asymmetric eigenfunction (solid black line) validates both our frozen-time decomposition (5.1) and our Fourier series ansatz (5.3). Validity of the frozen-time approach is further confirmed by the fact that our greatest eigenvalue $\unicode[STIX]{x1D702}=1.8\times 10^{-4}$ is an order of magnitude greater than the actual growth rate of the base state at the secondary troughs. We also point out that our stability results have been checked for convergence in terms of $N$ in (5.3). Moreover, the same results were obtained independently by the first two authors and were additionally cross-checked by the second author with a pseudo-spectral solver.

To better understand how the asymmetric eigenfunction ${\hat{h}}(x)$ in figure 7(b) affects the base profile $H(x)$ in figure 7(a), we have included an additional curve in figure 7(b). The red-dashed line there represents a pure translation of $H(x)$ at constant speed  $c$ , in which case the perturbed film thickness would satisfy $\unicode[STIX]{x2202}_{t}h=-c\unicode[STIX]{x2202}_{x}H$ . It turns out that ${\hat{h}}(x)$ corresponds exactly to such a translation within the main hump, as the solid black and red-dashed lines in figure 7(b) collapse there. Differences between the two curves are more apparent when normalizing with the base profile $H(x)$ , and this is represented in figure 7(c). We see that the red-dashed and black solid lines almost perfectly coincide within the main hump. This is because the main hump has virtually attained its sinusoidal equilibrium shape of width  $\unicode[STIX]{x1D6EC}_{c}$ , corresponding to the neutral mode of the Rayleigh–Taylor instability (see § 4). This mode is neutrally stable toward a pure translation and stable toward all other volume-preserving perturbations. Pure translation is thus the only symmetry-breaking option for the main hump and it requires a minimal energy input. It also means that the main hump does not actively contribute to the sliding instability mechanism.

By contrast, at the secondary troughs, ${\hat{h}}/H$ , which through (5.1) sets the linear growth rate $\unicode[STIX]{x2202}_{t}h/H=({\hat{h}}/H)\unicode[STIX]{x1D702}\exp (\unicode[STIX]{x1D702}\,t)$ , cannot be represented by a pure translation. Such a translation would impose $\unicode[STIX]{x2202}_{t}h=-c\unicode[STIX]{x2202}_{x}H=0$ at the troughs, but our eigenfunction ${\hat{h}}/H$ is clearly non-zero there (open circles in figure 7 c). At the left trough, ${\hat{h}}/H<0$ implying $\unicode[STIX]{x2202}_{t}h/H<0$ , thus the trough is pushed down and further thins, whereas, at the right trough, ${\hat{h}}/H>0$ implying $\unicode[STIX]{x2202}_{t}h/H>0$ , thus the trough is pulled up and further thickens.

Closer investigation of the ${\hat{h}}/H$ profile in figure 7(c) shows that the film at the trough locus itself is less affected than the immediate surroundings. Indeed, the film is more strongly pushed down to the left of the left trough and more strongly pulled up to the right of the right trough. This tends to move the trough loci leftward, constituting a sliding motion. It also produces a localized surface curvature decrease/increase at the left/right secondary trough. In figure 7(d), we have plotted $\unicode[STIX]{x1D702}\unicode[STIX]{x2202}_{xx}{\hat{h}}$ , which according to (5.1) is proportional to the time derivative of the perturbed surface curvature $\unicode[STIX]{x2202}_{t}(\unicode[STIX]{x2202}_{xx}h)=\unicode[STIX]{x1D702}\unicode[STIX]{x2202}_{xx}{\hat{h}}\exp (\unicode[STIX]{x1D702}t)$ . This quantity displays highly localized pulses at the secondary troughs, evidencing a flattening of the left and a curving of the right trough.

6 Mechanism of the sliding instability

We seek to identify the positive feedback mechanism causing amplification of the unstable perturbation ${\hat{h}}/H$ in figure 7(c) and we focus on the secondary troughs, as the main hump does not actively participate in the instability mechanism (see § 5). Additional evidence for the dynamical importance of the secondary troughs is provided by the transient stability analysis in § 7.

Figure 8.

Feedback mechanism amplifying the action of the unstable eigenfunction ${\hat{h}}$ in figure 7(c). (a) Zoomed view of the base profile $H(x)$ (figure 7 a). Open circles mark loci of secondary troughs, asterisks loci of growth rate extrema (e), and the thick red line represents the final equilibrium state (4.1); (b) surface curvature profiles $\unicode[STIX]{x2202}_{xx}H$ corresponding to (a). Dashed lines indicate evolution toward equilibrium state; (c) pressure gradient perturbation amplitude according to (6.1a,b ). Solid line: total amplitude $\unicode[STIX]{x2202}_{x}\hat{p}$ ; dots: capillary contribution $\unicode[STIX]{x2202}_{x}\hat{p}|_{\unicode[STIX]{x1D70E}}$ ; dot–dashed: gravity-induced contribution $\unicode[STIX]{x2202}_{x}\hat{p}|_{g}$ ; (d) flow rate perturbation amplitude. Solid line: total amplitude  $\hat{q}$ ; dashes: pressure gradient contribution $\hat{q}|_{p}=-(1/3)\unicode[STIX]{x2202}_{x}\hat{p}H^{3}$ ; dot–dot–dashed: viscous drag contribution $\hat{q}|_{h}=-\unicode[STIX]{x2202}_{x}pH^{2}{\hat{h}}$ ; (e) initial growth rate of the perturbation $\unicode[STIX]{x1D702}{\hat{h}}/H=-\unicode[STIX]{x2202}_{x}\hat{q}$ .

Figure 8(a) represents an enlarged view of the buckled portion of the surface profile $H(x)$ used for the stability analysis (black line). The thick red line corresponds to the final equilibrium state (4.1) toward which the film evolves by draining the remaining liquid from the secondary hump through the troughs. Drainage is governed by a balance between the driving capillary pressure gradient $\unicode[STIX]{x2202}_{x}P|_{\unicode[STIX]{x1D70E}}=-(1/\mathit{Bo})\unicode[STIX]{x2202}_{xxx}H$ , generated through variations in surface curvature $\unicode[STIX]{x2202}_{xx}H$ , and viscous drag, which scales with $1/H^{3}$ . Because the secondary troughs are so thin, and viscous drag there is so strong, a very steep $\unicode[STIX]{x2202}_{xx}H$ profile is established to drive liquid through them. This is plotted with a solid black line in figure 8(b), whereas the thick red line corresponds to the final equilibrium state, with a $\unicode[STIX]{x2202}_{xx}H$ discontinuity at the juncture of the sinusoidal and zero-thickness film segments (Yiantsios & Higgins Reference Yiantsios and Higgins1989). Dashed lines indicate the evolution toward this state.

Due to the steepness of the base state curvature profile (solid black line in figure 8 b), the localized $\unicode[STIX]{x2202}_{xx}{\hat{h}}$ pulses caused by the unstable eigenfunction (figure 7 d) produce large opposite-sign perturbations of the third derivative $\unicode[STIX]{x2202}_{xxx}{\hat{h}}$ either side of the secondary troughs. These translate into perturbation extrema of the capillary pressure gradient $\unicode[STIX]{x2202}_{x}\hat{p}|_{\unicode[STIX]{x1D70E}}\propto \unicode[STIX]{x2202}_{xxx}{\hat{h}}$ that destabilize the film, as shown in figure 8(c).

In this panel, we have represented the linear response of the driving pressure gradient (4.2) to the perturbed film thickness $h$ (5.1):

where, as in (5.1), $\unicode[STIX]{x2202}_{x}p$ corresponds to the perturbed quantity, $\unicode[STIX]{x2202}_{x}\hat{p}$ to the perturbation amplitude, and $\unicode[STIX]{x2202}_{x}P$ to the quasi-steady base state. The total perturbation amplitude $\unicode[STIX]{x2202}_{x}\hat{p}$ is plotted with a solid black line in figure 8(c), whereas the dot–dashed line represents the gravity-induced constituent $\unicode[STIX]{x2202}_{x}\hat{p}|_{g}$ , and filled circles the capillary constituent $\unicode[STIX]{x2202}_{x}\hat{p}|_{\unicode[STIX]{x1D70E}}$ . We see that the pressure perturbation is dominated by its capillary contribution $\unicode[STIX]{x2202}_{x}\hat{p}|_{\unicode[STIX]{x1D70E}}$ .

The sign difference of the $\unicode[STIX]{x2202}_{x}\hat{p}$ extrema either side of the secondary troughs acts to produce a disparity of flow rate disturbances there. We show this in figure 8(d), which represents the linear response of the flow rate $q=-(h^{3}/3)\unicode[STIX]{x2202}_{x}p$ (4.3) to the perturbed film thickness $h$ (5.1). The solid line corresponds to the total flow rate perturbation amplitude  $\hat{q}$ , the dashed line to the contribution of the pressure perturbation $\hat{q}|_{p}=-(1/3)\unicode[STIX]{x2202}_{x}\hat{p}H^{3}$ and the dot–dot–dashed line to the contribution of the viscous drag perturbation $\hat{q}|_{h}=-\unicode[STIX]{x2202}_{x}pH^{2}{\hat{h}}$ . The effect of the pressure perturbation alone ( $\hat{q}|_{p}$ , dashed line) is to reduce the flow toward the left trough and to increase the flow away from it (note that the base flow rate around the left trough is negative, i.e. to the left), which tends to drive the left trough to thin (and vice versa for the right trough), amplifying the action of the unstable eigenfunction ${\hat{h}}/H$ . At the same time, ${\hat{h}}/H$ alters the viscous drag around the secondary troughs and this has the opposite effect on the flow rate disturbances ( $\hat{q}|_{h}$ , dot–dashed line). Indeed, around the left trough, viscous drag is increased to a greater extent on its left than on its right, and thus the flow away from the trough is reduced more than the flow toward it (the opposite holds at the right trough). However, the net result of the two opposing effects on the total flow rate perturbation $\hat{q}$ (solid line) is to increase the flow rate difference across the left trough and to reduce it across the right one.

The spatial variation of $\hat{q}$ (solid line in figure 8 d) governs growth and decay of the film thickness through the continuity equation (2.2a ), yielding the growth rate $\unicode[STIX]{x1D702}{\hat{h}}/H=-\unicode[STIX]{x2202}_{x}\hat{q}$ , which we have plotted in figure 8(e). Concentrating on the left secondary trough (opposite arguments apply to the right trough), we see that the growth rate minimum is not situated at the trough locus itself but slightly to the left, while there is a smaller local maximum slightly to the right. This has two consequences. First, it causes the trough to move even further to the left, amplifying the action of the ${\hat{h}}/H$ perturbation in figure 7(c). Second, the film is deposited more rapidly on the left of the trough than it is peeled off on the right and this further reduces the local surface curvature, amplifying the action of the $\unicode[STIX]{x2202}_{xx}{\hat{h}}$ perturbation in figure 7(d).

These two positive feedbacks are caused by the way in which the flow rate perturbations in figure 8(d) behave around the trough. The pressure-related perturbation $\hat{q}|_{p}$ alone (dashed line) tends to produce a growth rate minimum at the trough locus itself, where $\unicode[STIX]{x2202}_{x}\hat{q}|_{p}$ is strongest. However, the drag-related flow rate perturbation $\hat{q}|_{h}$ (dot–dot–dashed line) strongly counteracts this effect in the immediate vicinity of the trough, as its slope is opposed. Outside of this region, the relevance of the drag-related perturbation rapidly decays with increasing film thickness (and decreasing magnitude of ${\hat{h}}/H$ ). Indeed, while the magnitude of $\hat{q}|_{p}$ keeps increasing to the left of the left trough, that of $\hat{q}|_{h}$ slowly decreases after having reached a maximum. This shifts the total growth rate minimum to the left of the trough locus (left-most star in figure 8 e).

The growth rate maximum to the right of the left trough (also marked by a star in figure 8 e) occurs for the same reasons. However, it is smaller in magnitude because of the asymmetric shape of the trough region, which connects the steep front of the main hump to the weakly sloped flank of the secondary hump. The thickness of the base state increases more rapidly to the left of the trough than to the right and so the double-pulsed pressure perturbation $\unicode[STIX]{x2202}_{x}\hat{p}$ (figure 8 c) unfolds its effect on the flow rate differently on either side (see $\hat{q}|_{p}$ marked by dashed line in figure 8 d).

Although the engine of the sliding instability is the capillary-induced migration of the secondary troughs, its most-visible consequence is the translation of the main hump, which contains most of the liquid. At first sight, it is surprising that the inconspicuous secondary troughs drive the main hump and not vice versa. However, the main hump has virtually attained a static equilibrium shape (i.e. the cutoff mode of the primary instability) that is neutrally stable toward a pure translation and stable toward all other volume-preserving perturbations. On the one hand, this means that only a minimal driving force is required to move the main hump and thus the latter follows the motion dictated by the secondary troughs. The actual driving force moving the hump results from a capillary pressure difference built up around it by the curvature perturbations in figure 7(d). These induce a greater pressure in the trailing edge of the main hump (left flank in figure 7 a) and a lower one in the leading edge (right flank in figure 7 a). On the other hand, by resisting all other volume-preserving perturbations, the main hump selects the possible sliding instability modes. In particular, the hump’s width is fixed and thus it resists compression/expansion. The two secondary troughs, which in our periodic setting enclose the main hump, are thus required to move in concert in the same direction. They are coupled in that the left one must be perturbed in the exact opposite manner than the right one. This requires the corresponding eigenfunction to be point symmetric about the main hump, a condition satisfied by the unstable mode uncovered in figure 7(c).

7 Transient stability analysis and the onset of sliding

The frozen-time analysis has demonstrated that the film is susceptible to a secondary sliding instability and shed light on the mechanisms involved. However, this method cannot be applied at earlier times when the film evolves more rapidly. Therefore, to investigate the onset of sliding, we relax the assumption of a frozen base state and instead linearize (2.2) around the time-evolving base state $H(x,t)$ :

where $h^{\ast }$ and $q^{\ast }$ denote the linear perturbations of film thickness and flow rate. We begin by computing the linear noise response of the perfectly symmetrical base state, starting from three representative time points $t_{i}=1922$ , 20 000 and 70 000 (see figure 3 a to situate $t_{i}$ in the evolution of the sliding film), for which the profiles $H(x,t_{i})$ are plotted in figure 9(a) (a logarithmic ordinate is chosen for better distinction).

Figure 9.

Transient instability and the onset of sliding. (a) Surface profiles $H(x,t_{i})$ of the perfectly symmetrical base state at three representative time points (see figure 3 a): $t_{i}=1922$ (red), 20 000 (green) and 70 000 (blue). A logarithmic ordinate is used for better distinction; (b) long-term linear responses $h^{\ast }(x,t_{i}+T)$ to a noisy perturbation (black line) of the $H(x,t_{i})$ profiles over time horizons $T=2000$ (red), 6000 (green) and 10 000 (blue). Solutions to (7.1) while advancing the base state $H(x,t)$ from $t_{i}$ to $t_{i}+T$ ; (c) most-unstable perturbations (solid lines) from transient stability analysis (Balestra et al. Reference Balestra, Brun and Gallaire2016) applied to the $H(x,t_{i})$ profiles over relatively short time horizons $T=200$ (red) and 1000 (green/blue). Dashed lines represents corresponding responses $h^{\ast }(x,t_{i}+T)$ , rescaled for a clear comparison; (d) nonlinear response of the periodic simulation in figure 3(a) to an injection of noise $h_{noise}$ (7.6) at $t=683$ . Time traces of the centre of mass $x_{C}$ are represented for different noise levels $\unicode[STIX]{x1D700}=\max (h_{noise})-\min (h_{noise})=0$ (solid), $1.3\times 10^{-4}$ (dashed), $1.3\times 10^{-3}$ (dotted), $1.3\times 10^{-2}$ (dot–dashed) and 0.04 (dot–dot–dashed). These correspond to the typical surface roughness of different materials, ranging from glass to steel.

We solve (7.1) for $h^{\ast }(x,t)$ , starting from a noisy initial condition $h^{\ast }(x,t_{i})=h_{noise}$ (defined in (7.6) and represented by a black line in figure 9 b), while advancing $H(x,t)$ from $t=t_{i}$ to $t=t_{i}+T$ over a relatively long time horizon $T$ (see caption of figure 11). Coloured lines in figure 9(b) represent the obtained linear responses $h^{\ast }(x,t_{i}+T)$ . For all three cases, the noisy initial perturbation evolves into a sliding mode similar to the eigenfunction obtained with the frozen-time approach (see figure 7 b). The growth rate is largest for the earliest (red line, $t_{i}=1922$ ) and lowest for the latest (blue line, $t_{i}=70\,000$ ) base state profile $H(x,t_{i})$ .

Next, we follow the transient stability analysis outlined in Balestra et al. (Reference Balestra, Brun and Gallaire2016) (see also Schmid Reference Schmid2007) to identify the most-unstable perturbations associated with the base state profiles $H(x,t_{i})$ . We repeatedly solve the direct problem (7.1) from an iteratively improved initial condition:

obtained by solving the adjoint problem:

starting from the end condition:

and stepping backwards in time from $t_{i}+T$ to $t_{i}$ . The procedure converges to a maximal value of the gain  $G(T)$ :

which quantifies growth over the time horizon $T$ . Figure 9(c) represents the thus-obtained most-unstable perturbations $h^{\ast }(x,t_{i})$ for the three base state profiles $H(x,t_{i})$ in figure 9(a) over relatively short time horizons $T$ (see caption).

All $h^{\ast }(x,t_{i})$ profiles display localized pulses that respectively thin and thicken the two secondary troughs. Over the time horizon  $T$ , they all evolve toward the sliding eigenmode obtained with our frozen-time analysis (5.1), as shown by the dashed lines in figure 11(c), which represent the linear responses $h^{\ast }(x,t_{i}+T)$ . Thus, the most-unstable scenario identified through our transient analysis exhibits the same long-time asymptotic behaviour, i.e. a concerted sliding motion of the entire film. This sliding eigenmode, which grows in a shape-preserving form, is most-effectively triggered by localized asymmetric disturbances at the secondary troughs, highlighting the importance of these regions for the onset of sliding.

Applying the transient analysis at earlier times, we have observed a qualitative change in the behaviour of the film near the time of buckling. In the pre-buckling regime, the most-unstable disturbance corresponds to a pure translation that merely produces a phase shift in the evolving film. It is only after buckling has occurred that the most-unstable disturbance mode takes on the non-trivial, localized form shown in figure 9(c). Thereafter, it remains virtually unchanged in form, with a gain $G(T)$ that is always greater than unity. In fact, we find that $G(T)$ is lower for later  $t_{i}$ , which is probably due to the increase in viscous stresses. Thus, there is no intrinsic/inherent fixed time for the onset of sliding. Rather, the onset of a macroscopically visible sliding motion is controlled by the level of ambient noise, e.g. due to surface roughness or pressure fluctuations. To demonstrate this, we have solved our nonlinear model (2.2), subject to periodic boundary conditions and starting from the initial condition (2.4), while injecting synthetic noise at a specific time $t_{noise}=683$ , i.e. just after the film surface has buckled (see figure 3 a). This is done through the random film thickness perturbation $h_{noise}$ (Chang, Demekhin & Kalaidin Reference Chang, Demekhin and Kalaidin1996):

which consists of sinusoidal modes of random phase shift $\unicode[STIX]{x1D711}_{rand}$ that cover 100 times the unstable range of the primary instability. By changing the coefficient  $\unicode[STIX]{x1D716}$ , the noise level $\unicode[STIX]{x1D700}=\max (h_{noise})-\min (h_{noise})$ was varied in five simulation runs: $\unicode[STIX]{x1D700}=0$ , $1.3\times 10^{-4}$ , $1.3\times 10^{-3}$ , $1.3\times 10^{-2}$ and 0.04. These values correspond to the typical surface roughness of different materials, ranging from glass to steel. Figure 9(d) represents time traces of the position $x_{C}$ of the film’s centre of mass (initially in the middle of the domain, i.e.  $x_{C}/L=0.5$ ), as obtained from the five runs. The onset of sliding is considerably precipitated with increasing noise level but, in the range studied, always occurs in the quasi-steady regime of the base state ( $t\geqslant 2\times 10^{-4}$ , see figure 3 a). Thus, although linear theory suggests that the film is susceptible to sliding at any time after buckling, nonlinearly, we find that the sliding eigenmode is only able to emerge after the film has slowed down to a quasi-steady state with quasi-equilibrium humps and sharp secondary troughs.

8 Gas film underneath a liquid layer

The ingredients of the sliding instability, identified for a suspended liquid film in § 6, are quite general and can be found in other systems as well, albeit in the presence of additional effects that may call for more complex models. One such scenario, which involves two active fluid phases, is related to the spontaneous motion of Leidenfrost drops on a heated surface. Liquid in contact with the hot surface evaporates to form a thin vapour film that supports the drop. Burton et al. (Reference Burton, Sharpe, van der Veen, Franco and Nagel2012) measured the vapour film underneath a Leidenfrost drop made of water and showed that the vapour–liquid interface is buckled, similar to what we have observed in figure 3(d) for the suspended water film. In a follow-up paper, Ma et al. (Reference Ma, Liétor-Santos and Burton2015) mention that the smooth aluminium substrate heating their Leidenfrost drop was curved ‘in order to keep drops stationary and suppress the buoyancy-driven Rayleigh–Taylor instability in the vapor layer’. They also contend that the dynamical traits of Leidenfrost drops, such as self-propulsion (Linke et al. Reference Linke, Alemán, Melling, Taormina, Francis, Dow-Hygelund, Narayanan, Taylor and Stout2006; Quéré Reference Quéré2013), depend ‘on a sensitive coupling between deformations of the liquid/vapor interface and lubrication flow in the thin ( ${\approx}100~\unicode[STIX]{x03BC}\text{m}$ ) vapor layer’. Most recently, experiments of Ma, Liétor-Santos & Burton (Reference Ma, Liétor-Santos and Burton2017) have shown that the oscillatory dynamics of Leidenfrost drops, which is linked to the drainage of vapour below the drop, depends only on the capillary length of the liquid, ‘indicating a purely hydrodynamic (non-thermal) origin for the oscillations’.

Figure 10.

Spontaneous sliding of a very thin gas film underneath a liquid layer (see sketch in figure 1 b), as simulated with the full model (2.1). Fluid properties (see caption of figure 1 b for values) correspond to a water vapour film underneath a water drop, according to the experiments of Burton et al. (Reference Burton, Sharpe, van der Veen, Franco and Nagel2012). The mean film thickness is $h_{0}=100~\unicode[STIX]{x03BC}\text{m}$ and the Bond number $\mathit{Bo}=0.0016$ . The domain length corresponds to the most-amplified wavelength of the Rayleigh–Taylor instability $\unicode[STIX]{x1D6EC}=2\sqrt{2}\unicode[STIX]{x03C0}/\sqrt{\mathit{Bo}}$ . (a) Logarithmic profiles of the film surface. Solid lines: profiles just before and after the onset of sliding; circles: suspended water film from figure 3; (b) film thickness time traces at the left and right secondary troughs. Solid: full model (2.1), dashed: full model in the limit $\unicode[STIX]{x1D6F1}_{\unicode[STIX]{x1D707}}\rightarrow 0$ , dot–dot–dashed: simplified model (2.2).

These experimental observations have prompted us to check whether a vapour film underneath a liquid layer, such as illustrated in figure 1(b), is prone to the sliding instability in a purely hydrodynamical sense, i.e. without accounting for evaporation. We consider the parameters quantified in the caption of figure 1(b), which are based on the experiments of Burton et al. (Reference Burton, Sharpe, van der Veen, Franco and Nagel2012), i.e. the gas layer consists of water vapour and the liquid layer of liquid water. The physical properties of the liquid ( $\unicode[STIX]{x1D707}_{2}$ , $\unicode[STIX]{x1D70C}_{2}$ ) and the surface tension $\unicode[STIX]{x1D70E}$ were evaluated at the experimental saturation temperature $T=100\,^{\circ }\text{C}$ , whereas the vapour properties ( $\unicode[STIX]{x1D707}_{1}$ , $\unicode[STIX]{x1D70C}_{1}$ ) were evaluated at $T=235\,^{\circ }\text{C}$ , corresponding to the average between the experimental wall and saturation temperatures. The mean vapour thickness was set to $h_{0}=100~\unicode[STIX]{x03BC}\text{m}$ , yielding a Bond number of $\mathit{Bo}=0.0016$ comparable to the experiments.

We have performed a periodic simulation of this configuration with our full model (2.1), which accounts for coupling between the thin gas film and the much more viscous liquid phase, the viscosity ratio being $\unicode[STIX]{x1D6F1}_{\unicode[STIX]{x1D707}}=2\times 10^{3}$ . The domain length was set to $\unicode[STIX]{x1D6EC}=2\sqrt{2}\unicode[STIX]{x03C0}/\sqrt{\mathit{Bo}}$ (the most-amplified wavelength of the Rayleigh–Taylor instability) and its height $D=10$ was chosen so large that the liquid phase is quasi-unconfined.

Our simulation has shown that the vapour film indeed slides spontaneously, displaying all the characteristic features of the sliding instability identified in § 6. Figure 10(a) represents the vapour film surface at two characteristic time points, just before and somewhat after the onset of sliding. We have additionally included a profile of the suspended water film from § 3 just before it slides (symbols in figure 10 a). We have chosen logarithmic scaling on the ordinate to accentuate any differences between the vapour film and the suspended water film. At the sliding onset, the surface profiles for the two cases, which have been scaled horizontally with the domain length $\unicode[STIX]{x1D6EC}=2\sqrt{2}\unicode[STIX]{x03C0}/\sqrt{\mathit{Bo}}$ , are virtually identical. This follows from capillary pressure gradients dominating the long-time evolution of the film, in which case (2.2) reduces to $\unicode[STIX]{x2202}_{t}h\propto \unicode[STIX]{x2202}_{x}(h^{3}\unicode[STIX]{x2202}_{xxx}h)$ after rescaling the horizontal coordinate $x$ with $\unicode[STIX]{x1D6EC}$ and adjusting the time scale accordingly.

The sliding onset for the vapour film is discernible in figure 10(b), showing time traces of the film thickness at the two secondary troughs. The role of viscous coupling is evidenced by comparing the full-model prediction (2.1) (solid line) with the limit $\unicode[STIX]{x1D6F1}_{\unicode[STIX]{x1D707}}\rightarrow 0$ (dashed line). Viscous stresses in the outer phase, through their action at the gas–liquid interface, modify the velocity profile and the associated viscous drag within the gas film. As a result, the onset of sliding is significantly delayed. These stresses also suppress the inertia-induced oscillation of the secondary troughs which had been observed for the suspended water film in figure 3(b) and figure 5. However, they do not qualitatively alter the loss of symmetry and sliding. In figure 10(b), we have also included the prediction of the simplified model (2.2) for completeness (see dot–dot–dashed line).

Our simplified analysis is far from proving that the sliding instability is linked to the spontaneous motion of Leidenfrost drops. Nonetheless, it has identified a possible mechanism. To verify whether this mechanism holds underneath a Leidenfrost drop, further analysis needs to include evaporation. The effect of evaporation may be stabilizing in that it tends to thicken a trough that has been thinned by a perturbation. On the other hand, it creates additional fluid within the secondary hump that needs to be drained through the troughs and this should be destabilizing. Evaporation maintains the interface at uniform temperature. This precludes the development of Marangoni stresses, which normally play a key role in the evolution of heated films. This is the subject of the next section, in which we show that such stresses can suppress the sliding instability.

9 Suppression of sliding by Marangoni stresses

Let us revisit the suspended liquid film in figure 1(a), but with the wall at a higher temperature than the ambient passive gas. Then modulations of the film thickness will result in temperature variations along the interface, which will in turn produce thermal Marangoni stresses. Since they act along the interface, the effect of these stresses on fluid drainage will be very different from that of capillary pressure gradients, which act within the bulk of the fluid. In fact, as we will show, these stresses can completely suppress sliding provided the wall is sufficiently hot.

In this analysis, we assume that surface tension decreases with temperature, $\unicode[STIX]{x2202}_{T}\unicode[STIX]{x1D70E}<0$ , in which case the problem is governed by (2.3), where the modified Marangoni number $\mathit{Ma}=\unicode[STIX]{x2202}_{T}\unicode[STIX]{x1D70E}(T_{w}-T_{\infty })/\unicode[STIX]{x1D70E}$ is negative. The Marangoni effect is thus destabilizing in terms of the primary instability. We have performed periodic simulations of (2.3) using the same parameter values as in § 3 (see caption of figure 1 a), additionally setting the Biot number to $\mathit{Bi}=1$ and increasing the magnitude of $\mathit{Ma}<0$ step by step from zero.

Figure 11.

Suppression of sliding by thermal Marangoni stresses. Heating the suspended water film from the wall suppresses the sliding instability and instead causes a cascade of buckling events (see also supplementary movie 3) similar to the traditional Marangoni problem (Boos & Thess Reference Boos and Thess1999; Oron Reference Oron2000). Numerical simulation of (2.3) using $\mathit{Ma}=-0.2$ , $\mathit{Bi}=1.0$ and $\mathit{Bo}=0.134$ (see caption of figure 1 a for other quantities). (a) Surface profile after three buckling events ( $t=2\times 10^{4}$ ). Inset shows enlarged view of boxed region around left tertiary hump; (b) formation of left tertiary hump from second buckling event; (c) formation of left-most quaternary hump (boxed region in inset of (a)) from third buckling event; (d) flow rate contributions (9.1) (normalized with  $h^{3}$ ) corresponding to the thick red profile in (b). Open circles mark loci of tertiary troughs. Thick red solid line: total flow rate  $q$ ; dashed: thermocapillary contribution $q|_{\mathit{Ma}}$ ; dot–dot–dashed: capillary contribution $q|_{\unicode[STIX]{x1D70E}}$ . Thick red profiles all correspond to same time.

Our simulations have shown that sliding is suppressed above a certain threshold value for $|\mathit{Ma}|$ . Instead of sliding, the film undergoes a cascade of buckling events that constantly produce new generations of humps and associated troughs, as has been observed in the traditional Marangoni problem (Boos & Thess Reference Boos and Thess1999; Oron Reference Oron2000). We focus our remaining discussion on a representative simulation for $\mathit{Ma}=-0.2$ , results of which are plotted in figure 11. We have also provided a supplementary movie, movie 3, which shows the buckling cascade in action (therein, the ordinate has been scaled logarithmically to highlight the evolution of the troughs).

Figure 11(a) represents the film surface after three buckling events. The inset shows an enlarged view of the boxed region surrounding the left tertiary hump, which results from the second buckling event. Figure 11(b) displays the time evolution of this second buckling event, from the dot–dashed unbuckled profile to the thick red profile, where the tertiary hump and associated troughs have already formed. Subsequently, the tertiary troughs (marked by open circles in figure 11 b) undergo a third buckling event, forming quaternary humps and troughs. This is shown in figure 11(c) for the left-most tertiary trough.

We now focus on the thick red profile in figure 11(b), which results from the second buckling event, and evaluate the different flow rate contributions:

where $q|_{\mathit{Ma}}$ denotes the contribution due to Marangoni stresses and $q|_{g}$ and $q|_{\unicode[STIX]{x1D70E}}$ are defined according to (4.3). These contributions are plotted in figure 11(d), where we have normalized $q$ with  $h^{3}$ . The thick red line corresponds to the total flow rate  $q$ , the dot–dot–dashed line to the capillary contribution $q|_{\unicode[STIX]{x1D70E}}$ , and the dashed line to the thermocapillary contribution $q|_{\mathit{Ma}}$ , while $q|_{g}$ is negligible and thus not plotted.

Considering the region around the left tertiary trough (left-most open circles in figures 11 b and 11 d), we see that the thermocapillary contribution $q|_{\mathit{Ma}}$ (dashed line in figure 11 d) is significant compared to the capillary one (dot–dot–dashed line). In contrast to the isothermal case in figure 6(h), drainage is thus not dominated by variations in surface curvature $\unicode[STIX]{x2202}_{xx}h$ . Instead, the surface slope $\unicode[STIX]{x2202}_{x}h$ , which determines $q|_{\mathit{Ma}}$ in (9.1), also plays an important role. We have verified that this holds for subsequent buckling events.

The reason $q|_{\mathit{Ma}}$ remains relevant even when the film is very thin, in contrast to $q|_{g}$ (4.3) which is also proportional to $\unicode[STIX]{x2202}_{x}h$ but subsides after producing the first buckling event (see figure 6), is that it scales with $h^{2}$ instead of  $h^{3}$ . The Marangoni effect, which acts at the film surface, is less hindered by viscous drag.

Thermocapillary drainage $q|_{\mathit{Ma}}\propto \unicode[STIX]{x2202}_{x}h$ is fundamentally different from capillary drainage in that it is symmetric about the troughs (where $\unicode[STIX]{x2202}_{x}h=0$ ). That is, liquid underneath a trough is driven away to both sides, as evidenced by the profile of $q|_{\mathit{Ma}}$ (dashed line) around the left-most tertiary trough in figure 11(d). Marangoni stresses thus help capillary drainage on one side of a trough and counteract it on the other and this is responsible for both the buckling cascade and the suppression of sliding.

Closer investigation of figures 11(b) and 11(c) shows that the buckling events from the second one onward differ from the first buckling event (figure 3 d) in that the buckling trough does not split into two new identical daughter troughs on either side. Instead, a new trough forms always on the inside of the original trough, which itself moves outward. This follows from the competition between thermocapillary $q|_{\mathit{Ma}}$ and capillary $q|_{\unicode[STIX]{x1D70E}}$ drainage, which produces a divergence point ( $q=0$ ) to the right of the left-most tertiary trough in figure 11(d) (see thick red line). From this divergence point, liquid is drained to either side, ultimately producing a quaternary trough there, when the slope of the secondary hump $\unicode[STIX]{x2202}_{x}h$ , which scales $q|_{\mathit{Ma}}$ in (9.1), has sufficiently grown.

Marangoni stresses, which in the present case are sufficiently strong to compete with capillary drainage, are also directly responsible for suppressing the sliding instability. First, they prevent the film from attaining a quasi-steady state. In fact, equation (2.3) possesses no steady solution for $\mathit{Ma}<0$ , in contrast to the final equilibrium state for the isothermal case (4.1). Instead, the film repeatedly buckles, producing ever thinner troughs, which would eventually disjoin due to long-range van der Waals forces between the wall and the film surface. Thereby, the fact that the width of the main hump is no longer constrained by an equilibrium state allows it to be increasingly compressed by the two adjacent troughs, which increasingly approach one another following each buckling event.

Second, Marangoni stresses counteract the way in which the film surface around a secondary trough would be modified by a sliding motion. Such a motion would peel off the film on the inside of an outward sliding secondary trough, whereas thermocapillary buckling would pull it down in the process of forming a new tertiary trough (see figure 11 b). Third, the growth rate contribution of Marangoni stresses (in the small $h$ limit):

reduces to a single term proportional to the surface curvature $\unicode[STIX]{x2202}_{xx}h$ (second term above) when evaluated at a trough (where $\unicode[STIX]{x2202}_{x}h=0$ ). For $\mathit{Ma}<0$ , this term tends to increase the thickness of a flattened secondary trough and reduce it at a curved trough, thereby opposing the positive feedback mechanism of the sliding instability discussed in § 6.