betaFeedback

WorksPhysics and Astronomy

Viscosity ratio across interfaces controls the stability and self-assembly of microrollers

Abstract

We investigate the individual and collective dynamics of torque-driven particles, called microrollers, near fluid-fluid interfaces. We find that the viscosity ratio across the interface controls the speed and direction of the particles, their relative motion, the growth of a fingering instability, and the self-assembled motile structures that emerge from it. By combining theory and large scale numerical simulations, we show how the viscosity ratio across the interface governs the long-range hydrodynamic interactions between particles and thus their collective behavior.

The motion of small driven particles near interfaces is observed in numerous biological and artificial systems [3] [1, 2]. Examples include droplets filled with self-propelled particles, where hydrodynamic couplings between the particles and the surrounding fluid-fluid interface induce collective motion and translation of the droplet [4, 5, 6, 7, 8, 9]. Navigation near interfaces also happens in natural environments with air-water interfaces, such as marine foams [10] or films [11–13], and in biomedical environments where microswimmers move near soft vessel walls, visco-elastic media (biofilms, gels, tissues) or mucus layers [14–16].

Among them, spinning particles, such as torque-driven colloids (also called microrollers), use rotation-translation couplings near surfaces to self-propel. Synthetic microrollers are actuated by an external magnetic field rotating about an axis parallel to the interface. The orientation of the magnetic field can be varied over time to guide these particles in a variety of environments. Thanks to their steerability and to the strong long-range flows they generate, they offer promising perspectives for particle micromanipulation, fluid pumping and drug delivery in microfluidic and biological systems [17–24]. Above a rigid wall, hydrodynamic interactions between microrollers induce a variety of collective motions such as periodic leapfrog orbits [25, 26], the formation of fast moving layers [27–29], the emergence of dense fronts [30] and the growth of a fingering instability that releases stable motile clusters [31, 32]. However, despite the rich and well documented dynamics observed above rigid boundaries, little is known about their collective behavior near interfaces.

In this Letter, we combine theory and simulations to investigate the dynamics of microrollers near fluid-fluid interfaces. We show how the viscosity ratio across the interface, denoted ξ\xi, modifies the flow around the spinning particles, with amazing consequences on their individual and collective motion. At the individual level, we find that ξ\xi controls the direction and speed of a single microroller. At the pair level, ξ\xi acts as an order parameter on their relative motion that determines the existence of periodic orbits. At the collective level, ξ\xi controls both the growth rate of the fingering instability initially observed above a no-slip surface, and the self-assembly process that leads to the emergence of motile clusters. As discussed in the conclusions, the ability to control the macroscopic response of these active suspensions opens promising perspectives for microfluidics applications.

We consider the motion of small torque-driven spherical particles with radius aa suspended in a fluid with dynamic viscosity ηin\eta_{\mathrm{in}} above a fluid-fluid interface (Figure 1a). The outer fluid on the other side of the interface has dynamic viscosity ηout\eta_{\mathrm{out}} which can vary from zero (e.g. air) to +∞+\infty (a rigid wall). We denote ξ=ηout/ηin\xi= \eta_{\mathrm{out}}/\eta_{\mathrm{in}} the viscosity ratio between the two fluids across the interface. In this work we consider particles with large contact angles that are not adsorbed to the surface (such as paramagnetic beads above air/water interfaces [27]). In typical experiments the particles are micron-sized (a=O(1) μma = O(1)\,\mu\mathrm{m}) and suspended in water (ηin=10−3 Pa s\eta_{\mathrm{in}} = 10^{-3}\,\mathrm{Pa\,s}). They rotate in synchrony with an external magnetic field with frequency f=O(10) Hzf = O(10)\,\mathrm{Hz} thanks to a magnetic torque aligned with the yy-axis denoted τ\tau. The corresponding capillary number is Ca=ηinu/γ≈10−6−10−5\mathrm{Ca} = \eta_{\mathrm{in}}u/\gamma\approx10^{-6} - 10^{-5}, where u=ωa=2πfau = \omega a = 2\pi f a is the maximal fluid velocity, reached on the particle surface due to the no-slip condition, and γ=O(10−2) N/m\gamma= O(10^{-2})\,\mathrm{N/m} is the typical surface tension between the two phases. Owing to the very small value of Ca\mathrm{Ca} and to the small size of the particles, the interface can be approximated as flat and nondeformable for all values of ξ\xi [33]. In this limit, one can compute the flow field around the particles using the Green’s function GG of the Stokes equations which, by linearity, is given by [34]:

Schematic of a torque-driven particle above a fluid-fluid interface and self-induced velocity as a function of viscosity ratio and normalized height

Figure 1. a) Schematic of a torque-driven particle above a fluid-fluid interface. b) Self-induced velocity of the particle v0v_0 in response to a constant torque τ\tau as a function of ξ=ηout/ηin\xi= \eta_{\mathrm{out}}/\eta_{\mathrm{in}} and normalized height h/ah/a.

G=1(ξ+1)GFS+ξ(ξ+1)GW(1)G = \frac{1}{\left(\xi+ 1\right)}G^{FS} + \frac{\xi}{\left(\xi+ 1\right)}G^{W} \tag*{(1)}

where GFSG^{FS} is the Green’s function of a domain bounded by a flat free-surface and GWG^{W} a domain bounded by a flat no-slip wall, both of which have known analytical expressions based on image systems [33, 35][36]. Changing ξ\xi from 0 to +∞+\infty in (1) transitions smoothly from a free-surface to a no-slip wall.

Single particle motion.– Using the Faxen formulae for the motion of a spherical particle in a non-uniform flow [37], the far-field approximation of the self-induced velocity of a single microroller above the interface, is given by [38]

v0(ξ,h/a)=14(ah)2[−1ξ+1+ξξ+112(ah)2](2)v_0(\xi,h/a)=\frac{1}{4}\left(\frac{a}{h}\right)^2\left[-\frac{1}{\xi+1}+\frac{\xi}{\xi+1}\frac{1}{2}\left(\frac{a}{h}\right)^2\right] \tag*{(2)}

where the velocity has been normalized by τ/8πηa2\tau/8\pi\eta a^2. Interestingly, the sign of the velocity in (2) depends both on the viscosity ratio and the particle height (Figure 1b). Below a critical value ξc=2\xi_c=2 the velocity is always negative, even though the applied torque is clockwise, and decays as (a/h)2(a/h)^2 for ξ=0\xi=0. This surprising backward motion has already been observed experimentally [27] and is due to the fact there is no velocity gradient on the free-surface underneath. The particle therefore experiences more viscous stress on its upper side, where velocity gradients are larger, than on the lower side. As a result, the particle “rolls” on the liquid above, which resists more against rotational motion, and thus moves backwards. When ξ≥ξc\xi\geq\xi_c, there is a critical height hc=ξ/2ah_c=\sqrt{\xi/2}a at which viscous stresses balance between both sides such that the particle rotates in place, i.e. v0(ξ>ξc,hc)=0v_0(\xi>\xi_c,h_c)=0. v0v_0 is positive below hch_c and negative above. In the limit of a no-slip wall, ξ→+∞\xi\to+\infty, v0v_0 is always positive and decays as (a/h)4(a/h)^4. These results show that both the direction and speed of a rotating particle can be controlled with the viscosity ratio.

Pair dynamics.– When adding another particle in the system, the trajectories exhibit a richer dynamics than the individual motion described above: particles can follow a variety of periodic orbits, change direction or end up translating at a steady speed (Figure 2a-Figure 2c). We consider two torque-driven particles, lying in the (x,z)(x,z)-plane vertical to the floor and separated by a distance r12=x122+z122r_{12}=\sqrt{x_{12}^2+z_{12}^2}, where x12=x1−x2x_{12}=x_1-x_2 and z12=z1−z2z_{12}=z_1-z_2. For simplicity, we focus on the limit a≪min⁡(r12,z1,z2)a\ll\min(r_{12},z_1,z_2), where the particles are considered as point-torques (rotlets) so that the equations of motion can be written compactly as a simple dynamical system with three degrees of freedom [38]:

Phase diagrams and streamlines of rotlet trajectories

Figure 2. a) Effect of ξ\xi on the periodicity of trajectories of two rotlets (blue and black disks) with identical initial separation: x120=4x_{12}^{0}=4. Red disk and dashed line: position and trajectory of the center of mass (xC,zC)(x_C,z_C). Black arrow: direction of motion of xCx_C. b) Phase diagram of rotlet orbits near the critical points as a function of ξ\xi and x12x_{12}. Black line: critical distance x12⋆x_{12}^{\star}. c) Effect of ξ\xi on the direction of motion of a pair with x120=2x_{12}^{0}=2. d) Streamlines and flow field around a rotlet in the (x,z)(x,z)-plane for various ξ\xi. Cyan solid line: iso-value uz=0u_z=0. Cyan dotted line: z=1z=1. Cyan circles: position of the critical horizontal separation x12⋆x_{12}^{\star}, reported below each panel.

[x˙12z˙12z˙C]=[2z12(1r123−ξξ+12x122−4zC2R125)+δv02x12(1R123−1r123+ξξ+112zC2R125)ξξ+16x12z12zCR125](3)\begin{bmatrix} \dot{x}_{12}\\ \dot{z}_{12}\\ \dot{z}_{C} \end{bmatrix} = \begin{bmatrix} 2z_{12}\left(\frac{1}{r_{12}^{3}}-\frac{\xi}{\xi+1}\frac{2x_{12}^{2}-4z_{C}^{2}}{R_{12}^{5}}\right)+\delta v_0\\ 2x_{12}\left(\frac{1}{R_{12}^{3}}-\frac{1}{r_{12}^{3}}+\frac{\xi}{\xi+1}\frac{12z_{C}^{2}}{R_{12}^{5}}\right)\\ \frac{\xi}{\xi+1}\frac{6x_{12}z_{12}z_{C}}{R_{12}^{5}} \tag*{(3)} \end{bmatrix}

where zC=(z1+z2)/2z_C=(z_1+z_2)/2, and R12=x122+4zC2R_{12}=\sqrt{x_{12}^2+4z_C^2}. δv0=−141ξ+1((zC+z12/2)−2−(zC−z12/2)−2)\delta v_0=-\frac{1}{4}\frac{1}{\xi+1}\left((z_C+z_{12}/2)^{-2}-(z_C-z_{12}/2)^{-2}\right) is the difference between the self-induced velocities in the point-particle limit. Here lengths have been rescaled by the initial height of the system lC=zC0=zC(t=0)l_C=z_C^0=z_C(t=0) and time by tC=8πηlC3/τt_C=8\pi\eta l_C^3/\tau. The dynamical system (3) has two critical saddle points with coordinates x12±=(x12=±x12∗,z12=0,zC=1)x_{12}^{\pm}=(x_{12}= \pm x_{12}^{*},z_{12}=0,z_C=1), where x12∗≠0x_{12}^{*}\neq0 is a zero of ξξ+12x122−4zC2R125\frac{\xi}{\xi+1}\frac{2x_{12}^{2}-4z_C^2}{R_{12}^{5}} [38]. These critical points separate periodic (leapfrog) and non-periodic orbits, as shown on Figure 2b. Interestingly, their position, x12∗x_{12}^{*}, is controlled by the viscosity ratio ξ\xi. To visualize this dependency, we examine the changes in the topology of the flow field around a spinning particle as a function of ξ\xi (Figure 2d). When ξ=0\xi=0 all the streamlines in the domain are closed and circular. As ξ\xi increases, the region of closed streamlines contracts in the vicinity of the particle, while open streamlines, which push the fluid forward, occupy more space. To understand the effect of the flow on the periodicity of the orbits we place the particles at the critical point x12=x12∗x_{12}=x_{12}^{*} where z1=z2=1z_1=z_2=1 and we assume x1>x2x_1>x_2. At this position their relative, and thus absolute, vertical velocity vanishes z˙12=0⇒z˙1=z˙2=z˙C=0\dot{z}_{12}=0 \Rightarrow\dot{z}_1=\dot{z}_2=\dot{z}_C=0. x12⋆x_{12}^{\star} therefore corresponds to the intersection between the contour of zero vertical fluid velocity induced by a particle, uz=0u_z=0, and the line z=1z=1 as shown on Figure 2d. If x12>x12⋆x_{12}>x_{12}^{\star} then z˙12>0\dot{z}_{12}>0 and the particles cannot perform the leapfrog motion, as can be intuited by superimposing the flow of two rotlets side by side (Figure 2a and d). Below a critical value ξc=1\xi_c=1, z˙12\dot{z}_{12} is always negative at z=1z=1, which implies that there is no critical point x12⋆x_{12}^{\star} (uz=0u_z=0 and z=1z=1 never meet) and thus that all the trajectories are periodic regardless of the initial separation (Figure 2b-Figure 2d). The viscosity ratio therefore controls the existence of periodic orbits. The viscosity ratio also sets the speed and direction of these periodic motions. In the limit ξ=0\xi=0 the backward motion is fastest and the height of the system remains constant (z˙C=0\dot{z}_C=0) (Figure 2c, left panel). Otherwise the direction of motion depends on the relative position of the particles so that, given an initial configuration, there is always a threshold value for which the pair orbits in place (e.g. ξ=3.66\xi=3.66 for x120=4x_{12}^{0}=4 in Figure 2c, middle panel). In the limit ξ=+∞\xi=+\infty, the pair always moves forward: x˙C=6zCx122/R15>0\dot{x}_C=6z_Cx_{12}^{2}/R_{1}^{5}>0 (Figure 2c, right panel).

Collective dynamics.— We further explore the dynamics at the collective level where large collections of particles interact. In the rigid-wall limit (ξ→∞\xi\to\infty) the suspension exhibits a cascade of events: an initially uniform strip of particles forms a dense front which is subject to a transverse fingering instability, from which the fingertips then detach to generate dense motile structures. As shown in previous work, the wavelength at the onset of the instability is set by the mean height of the front [31, 32]. We investigate the effect of the viscosity ratio on this rich collective dynamics using 3D Stokesian dynamics simulations that include both hydrodynamic and steric interactions between tens of thousands of microrollers above the interface [38]. Figure 3a shows the time evolution of 10,000 microrollers initially lying on a monolayer with area fraction ϕ=0.65\phi=0.65, at a given height zC0/a=1.2z_C^0/a=1.2, for various viscosity ratios ξ∈{0,0.5,1.5,5,+∞}\xi\in\{0,0.5,1.5,5,+\infty\} (see SI Movies 1-2). In the case of a no-slip wall, ξ=+∞\xi=+\infty, the strips evolves as described above. The rollers inside the fingertips perform a treadmill motion that follows the clockwise rotation of the torque. A few particles are occasionally shed from the front and lifted up by flow of the rotating structures. When ξ=5\xi=5 the suspension behaves similarly with a slower forward motion. At a critical value ξ≈1.5\xi\approx1.5, the suspension treadmills without translating, but the transverse instability still occurs in place, leading to a lateral separation of the particles.

Time evolution of microrollers and growth rates

Figure 3. a) Time-evolution of 10,000 microrollers initially uniformly distributed in a monolayer (delimited by the black dotted lines) above the interface. Each color corresponds to a different value of ξ∈{0,0.5,1.5,5,+∞}\xi\in\{0, 0.5, 1.5, 5, +\infty\} simulated independently. Bottom panel: side view at t=87.9t = 87.9. Arrows: direction of motion. b) Growth rate of the two line model (4)–(5) for various ξ\xi. Dashed line: fastest growing mode kmax⁡k_{\max}. c) Maximum growth rate (solid cyan line) as a function of ξ\xi. Dashed black line: prefactor ∼ξ/(ξ+1)\sim\xi/(\xi+ 1) due to the transverse flows generated above a no-slip boundary (inset).

Below that threshold, the suspension self-assembles into a long roll treadmilling clockwise but moving backwards. When ξ=0.5\xi= 0.5 the roll is wavy due to the transverse instability, but evolves at a significantly slower rate, and sporadically sheds particles from the rear. In the free-surface limit (ξ=0\xi= 0), the transverse instability is suppressed and the roll remains straight and stable. It does not shed any particle because, as shown in the previous section, the streamlines of the flow induced by the microrollers are closed regardless of the separation distance (Figure 2b, bottom panel). Finally, we note that the mean height of the particles in the fingertips slightly increases with ξ\xi (Figure 3a, bottom panel), which in turn increases the wavelength of the transverse instability. This is due to the upward advective flows ahead of the particles that get stronger with ξ\xi (see (Figure 2b) and allow the particles behind to lift the leading front where the instability occurs.

In order to better understand the effect of ξ\xi on the transverse instability, we use a simple model consisting of two continuous lines of rotlets separated by a distance dd in a plane parallel to the interface at a constant height hh. The lines are aligned along the yy-axis with rotlet density ρi(y,t)\rho_i(y,t) and position xi(y,t)x_i(y,t), i=1,2i = 1, 2. The model is governed by the equations of motion of each line together with the conservation of rotlets along their length:

∂xi(y,t)∂t=ux(xi(y,t),y)∂ρi(y,t)∂t=−∂[ρi(y,t)uy(xi(y,t),y)]∂y,i=1,2(4)\begin{aligned} \frac{\partial x_i(y,t)}{\partial t} &= u_x(x_i(y,t),y) \\ \frac{\partial\rho_i(y,t)}{\partial t} &= -\frac{\partial[\rho_i(y,t)u_y(x_i(y,t),y)]}{\partial y}, \qquad i = 1,2 \tag*{(4)} \end{aligned}

where length and time have been rescaled with lc=hl_c = h and tct_c defined above. The velocity (ux,uy)(u_x,u_y) in the RHS arises from the long-ranged hydrodynamic interactions between the microrollers of each line, e.g.

ux(xi,y)=∑j=12∫−∞+∞μxuτ(xi−xj,y−y′)ρj(y′) dy′(5)u_x(x_i,y) = \sum_{j=1}^{2}\int_{-\infty}^{+\infty}\mu_x^{u\tau}(x_i-x_j,y-y')\rho_j(y')\,\mathrm{d}y' \tag*{(5)}

where μyFS(x−x′)\mu_y^{\mathrm{FS}}(x-x') is the xx-component of the operator μyFS(x−x′)=12(∇x′×G(x−x′))⋅ey\mu_y^{\mathrm{FS}}(x-x')=\frac{1}{2}\left(\nabla_{x'}\times G(x-x')\right)\cdot e_y that computes the fluid velocity at position xx induced by a rotlet, directed along eye_y, at position x′x' [38]. As shown in previous work, this model contains all the essential ingredients to faithfully reproduce the fingering instability above a no-slip wall [31] and naturally extends to fluid-fluid interfaces by using the linear combination of Green’s functions in (1). The base state of this system corresponds to two straight lines with uniform rotlet densities (ρi(y,t)=ρ0)(\rho_i(y,t)=\rho_0) translating at a steady speed, so that their position are constant in the moving frame (x1(y,t)=0,x2(y,t)=d)(x_1(y,t)=0,x_2(y,t)=d). After carrying a linear stability analysis of the system around the base state, we obtain an analytical expression for the growth rate σ(k,ξ)\sigma(k,\xi) [38], plotted on Figure 3b. First, we notice that an identical fastest growing mode is selected for all values of ξ\xi, as seen by the clear bump at λmax⁡=2π/kmax⁡=23\lambda_{\max}=2\pi/k_{\max}=23, except for ξ=0\xi=0 for which there is no clear selection. The existence of a plateau at large kk is specific to the two-lines model, and its dependence on ξ\xi is explained in the SI [38]. Second the corresponding growth rate, σmax⁡\sigma_{\max}, increases with ξ\xi. The increase of σmax⁡\sigma_{\max} is quantified on Figure 3c, and shows a plateau at large ξ\xi. A detailed analysis of the two-line model shows that the growth rate of the transverse instability is proportional to the transverse flow uyu_y in (4). Since the microrollers do not induce transverse flows above a free surface (uyFS=0(u_y^{\mathrm{FS}}=0 [38]), σmax⁡\sigma_{\max} scales as ξ/(ξ+1)\xi/(\xi+1), which is the prefactor of the no-slip wall contribution GWG^W in (1) (see inset of Figure 3c). Overall the two-line model highlights the crucial role of these transverse flows and exhibits a good agreement with the simulations in the sense that, for a given particle height, ξ\xi only changes the growth rate of the instability without affecting much the dominant wavelength. However since the height of the particles is kept constant in the model, it cannot capture the slight increase of the wavelength with ξ\xi that is observed on Figure 3a. The nature of the interface does not only affect the direction of motion of the suspension and the growth of the transverse instability, but also the shape and structure of the emerging clusters. We study these changes from a macroscopic point of view, with a mean field description of the density of microrollers ρ(x,t)\rho(x,t) in the (x,z)(x,z)-plane perpendicular to the interface, denoted P\mathcal{P}. ρ(x,t)\rho(\boldsymbol{x},t) obeys a conservation equation

∂ρ∂t+∇⋅(uρ)=0.(6)\frac{\partial\rho}{\partial t}+\nabla\cdot(\boldsymbol{u}\rho)=0. \tag*{(6)}

The velocities in the flux term arise from the long-ranged hydrodynamic interactions between microrollers in the plane P\mathcal{P}, e.g. ux(x,t)=∫Pμxu(x−x′)ρ(x′,t)dx′u_x(\boldsymbol{x},t)=\int_{\mathcal{P}}\mu_x^u(\boldsymbol{x}-\boldsymbol{x}')\rho(\boldsymbol{x}',t)\mathrm{d}\boldsymbol{x}'. The nonlocal equation (6) is solved numerically [38] and the rollers are initially uniformly distributed (ρ=ρ0\rho=\rho_0) over a thin strip of aspect ratio Γ=L/H=9.4\Gamma=L/H=9.4 near the interface. After some time, the system reaches a quasi-steady state where a main cluster emerges and translates at a steady speed (see SI Movies 3-7). Figure 4 shows a snapshot of the microroller density at the same dimensionless time in the quasi-permanent regime for various values of ξ\xi together with the density ρ(xC,z)\rho(x_C,z) and velocity profiles ux(xC,z)u_x(x_C,z) at the central cross-section of the cluster. The cluster is delimited by the iso-value ρ=0.4ρ0\rho=0.4\rho_0 (red line). As in the discrete particle simulations, the velocity of the cluster is fastest and backwards for ξ=0\xi=0, vanishes around ξ≈1.1\xi\approx1.1, where the cluster treadmills in place, and increases forward beyond. The shape of the cluster evolves from a near-hemisphere of aspect ratio Γ=3.1\Gamma=3.1 at ξ=0\xi=0, to a “pancake” shape (Γ=8.4\Gamma=8.4) at ξ=1.1\xi=1.1, and becomes circular when ξ\xi increases further (Γ=2\Gamma=2 and Γ=1.4\Gamma=1.4 for ξ=5\xi=5 and ξ=+∞\xi=+\infty). In addition to the shape, ξ\xi controls the particle distribution inside the cluster. The distribution is uniform at ξ=0\xi=0 (ρ≈ρ0\rho\approx\rho_0) and becomes more peaked as ξ\xi increases, with a maximum ρmax⁡≈3.8ρ0\rho_{\max}\approx3.8\rho_0 reached at ξ=+∞\xi=+\infty. These changes in concentration, together with the boundary condition at the interface underneath, determine the velocity profile within the cluster. At ξ=0\xi=0, the treadmilling motion is fastest near the bottom interface, where the slip condition allows for large velocities. As ξ\xi increases, the particles concentrate at the center and the no-slip condition becomes stronger, which shifts the maximum velocity upwards and generates a rigid-body motion near the core.

Particle density distributions and velocity profiles for various $\xi$

Figure 4. Particle density distribution ρ(x,t)\rho(x,t) obtained from (6) at t=1418t=1418 (quasi-steady state) in the (x,z)(x,z)-plane for various ξ∈{0,0.5,1.5,5,+∞}\xi\in\{0,0.5,1.5,5,+\infty\}. Black arrows: flow field. Solid red line: ρ/ρ0=0.4\rho/\rho_0=0.4. Red cross: position of the COM (xC,zC)(x_C,z_C). Side panels: density (dotted blue line) and velocity (solid black line) profiles at x=xCx=x_C.

Discussion.– In this Letter, we have shown that the viscosity ratio across interfaces controls both the microscopic and macroscopic response of an active suspension. These fundamental findings are also of technological importance. Our results suggest that the strategies for particle transport and fluid pumping initially developed with microrollers above solid walls can readily be extended to a variety of biological and microfluidic environments with different type of surfaces. Our results could even generalize to solid boundaries with a finite slip length [39], which are found in many experimental systems, such as hydrophobic surfaces. In this work we considered particles lying above the interface, i.e. with large contact angles. In some cases the particles wet the outer fluid and get adsorbed at the interface. Adsorbed active particle layers can be used as active surfactants to modulate interfacial properties in emulsions or films, or to pump and mix flows in the surrounding fluids [3] [40]. Up to now, the hydrodynamics of torque-driven particles straddling fluid-fluid interfaces is still not well-understood [41] and should deserve more attention given their exciting applications.

I thank M. Driscoll, A. Donev and S. Michelin for their critical reading of the manuscript as well as for stimulating discussions. I acknowledge support from the French National Research Agency (ANR), under award ANR-20-CE30-0006. I also thank the NVIDIA Academic Partnership program for providing GPU hardware for performing some of the simulations reported here.

Far-field approximation of the self-induced velocity of a single microroller above a fluid-fluid interface

The flow induced by a particle located at x′\mathbf{x}' and driven by a torque τ\boldsymbol{\tau} is given by

u(x)=12[∇x′×G(x−x′)]⋅τ(7)\mathbf{u}(\mathbf{x}) = \frac{1}{2} \left[\nabla_{\mathbf{x}'} \times\mathbf{G}(\mathbf{x} - \mathbf{x}')\right] \cdot\boldsymbol{\tau} \tag*{(7)}

where G(x−x′)=1/(ξ+1)GFS(x−x′)+ξ/(ξ+1)GW(x−x′)\mathbf{G}(\mathbf{x} - \mathbf{x}') = 1/(\xi+ 1)\mathbf{G}^{FS}(\mathbf{x} - \mathbf{x}') + \xi/(\xi+ 1)\mathbf{G}^{W}(\mathbf{x} - \mathbf{x}') is the Green’s function of the domain (Eq. (1) in the main text). The self-induced velocity v0\mathbf{v}_0 of the particle due to its own flow is given by the Faxen formula

v0=(1+a26∇2)u(x)∣x=x′(8)\mathbf{v}_0 = \left(1 + \frac{a^2}{6}\nabla^2\right)\mathbf{u}(\mathbf{x})\bigg|_{\mathbf{x}=\mathbf{x}'} \tag*{(8)}

Using the analytical expressions of GFS\mathbf{G}^{FS} and GW\mathbf{G}^{W} provided in [1] and [2] respectively, the self-induced response to an external torque aligned along the yy-axis, τ=τey\boldsymbol{\tau} = \tau\mathbf{e}_y, is given by v0=v0ex\mathbf{v}_0 = v_0\mathbf{e}_x where

v0=τ32πηa2(ah)2[−1ξ+1+ξξ12(ah)2](9)v_0 = \frac{\tau}{32\pi\eta a^2}\left(\frac{a}{h}\right)^2\left[-\frac{1}{\xi+1}+\frac{\xi}{\xi}\frac{1}{2}\left(\frac{a}{h}\right)^2\right] \tag*{(9)}

which leads to Eq. (2) in the main text after normalizing by τ/8πηa2\tau/8\pi\eta a^2. Note that the self-induced velocity is always negative in the point-particle limit (a/h→0)(a/h \to0)

lim⁡a/h→0v0=−τ32πη1ξ+11h2≤0(10)\lim_{a/h\to0} v_0 = -\frac{\tau}{32\pi\eta}\frac{1}{\xi+1}\frac{1}{h^2} \le0 \tag*{(10)}

Dynamical system approach for the two point-torques (rotlets)

Equations of motion

Consider two point particles with positions x=(xi,zi)\mathbf{x}=(x_i,z_i), i=1,2i=1,2 driven by a torque τ\boldsymbol{\tau} in the (x,z)(x,z)-plane above a fluid-fluid interface. Since the particles have no size, the velocity of particle 1 due to the torque on particle 2 is equal to the fluid velocity at the particle position x1\mathbf{x}_1:

v1=u2(x=x1)=12[∇x′×G(x1−x′)]x′=x2⋅τ(11)\mathbf{v}_1=\mathbf{u}_2(\mathbf{x}=\mathbf{x}_1)=\frac{1}{2}\left[\nabla_{\mathbf{x}'}\times\mathbf{G}(\mathbf{x}_1-\mathbf{x}')\right]_{\mathbf{x}'=\mathbf{x}_2}\cdot\boldsymbol{\tau} \tag*{(11)}

From this formula, and by adding the self-induced contribution (S4), one obtains simple analytical expressions for the particle velocities vi=(x˙i,z˙i)\mathbf{v}_i=(\dot{x}_i,\dot{z}_i), i=1,2i=1,2. Using the translational invariance of the system along the xx-direction, one can reduce the system to three degrees of freedom, (x12=x1−x2,z12=z1−z2,zC=(z1+z2)/2)(x_{12}=x_1-x_2,z_{12}=z_1-z_2,z_C=(z_1+z_2)/2), which leads to Eq. (3) in the main text. Note that the mean horizontal position xC=(x1+x2)/2x_C=(x_1+x_2)/2 is omitted in Eq. (3). Due to the translational invariance of the system along the xx-axis, xCx_C decouples from the other degrees of freedom in the sense that it does not affect their evolution in Eq. (3) (the contrary is not true, x˙C\dot{x}_C does depend on x12x_{12}, z12z_{12} and zCz_C). As a result xCx_C does not affect the existence, and the values, of the critical points that separate periodic and non-periodic orbits (see section below).

Critical saddle points

The critical points of the dynamical system are found by solving

\begin{bmatrix} \dot{x}_{12}\\ \dot{z}_{12}\\ \dot{z}_C \end{bmatrix} = \begin{bmatrix} 2z_{12}\left(\frac{1}{r_{12}^{3}}-\frac{\xi}{\xi+1}\frac{2x_{12}^{2}-4z_{12}^{2}}{R_{12}^{5}}\right)-\frac{1}{4}\frac{1}{\xi+1}\left(\frac{1}{(z_C+z_{12}/2)^2}-\frac{1}{(z_C-z_{12}/2)^2}\right)\\ 2x_{12}\left(\frac{1}{R_{12}^{3}}-\frac{1}{r_{12}^{3}}+\frac{\xi}{\xi+1}\frac{12z_C^{2}}{R_{12}^{5}}\right)\\ \frac{\xi}{\xi+1}\frac{6x_{12}z_{12}z_C}{R_{12}^{5}} \end{bmatrix} =\mathcal{G}(\mathbf{x})=\mathbf{0} \tag*{(12)}\end{bmatrix}

where G(x)\mathcal{G}(\mathbf{x}) is the vector field that describes the body dynamics in phase space. Their coordinate is given by x±∗=(x12=±x12∗,z12=0,zC=1)\mathbf{x}^{*}_{\pm}=(x_{12}=\pm x^{*}_{12},z_{12}=0,z_C=1), where x12∗≠0x^{*}_{12}\ne0 is a zero of z˙12(x12,0,1,ξ)\dot{z}_{12}(x_{12},0,1,\xi), which is found by solving

x123(x122+4+12ξξ+1)−(x122+4)5/2=0.(13)x_{12}^{3}\left(x_{12}^{2}+4+12\frac{\xi}{\xi+1}\right)-(x_{12}^{2}+4)^{5/2}=0. \tag*{(13)}

In order to characterize these fixed points, we compute the eigenvalues of the Jacobian matrix of the system evaluated at x⋆x^\star (∇xG(x)∣x±⋆\nabla_x\mathcal{G}(x)|_{x_\pm^\star}). Figure S1 shows these eigenvalues as a function of ξ∈[1,100]\xi\in[1,100] (remember that the critical points do not exist for ξ<1\xi< 1, see main text). Two eigenvalues are real with opposite signs and the third one is zero, which clearly shows that these critical points are saddle points and that the dynamics lies on a two-dimensional manifold in the vicinity of x±⋆x_\pm^\star in phase space. These two saddle points with symmetric positions delimit a separatrix between periodic and non-periodic orbits. Since x12⋆x_{12}^\star decreases when ξ\xi increases, the size of the domain enclosed by the separatrix decreases as well. Note that the magnitude of the eigenvalues increases with ξ\xi, which indicates an acceleration of the relative motion between particles with ξ\xi in the vicinity of the fixed points.

Figure S1. Eigenvalues of ∇xG(x)∣x±⋆\nabla_x\mathcal{G}(x)|_{x_\pm^\star} as a function of the viscosity ratio ξ∈[1,100]\xi\in[1,100]

Details on the large scale 3D Stokesian dynamics simulations

Consider a collection of NN microrollers with radius aa and positions X={x1,…,xN}\mathbf{X} = \{x_1,\ldots,x_N\} subject external forces F={f1,…,fN}\mathbf{F} = \{f_1,\ldots,f_N\} and torques T={τ1,…,τN}\mathbf{T} = \{\tau_1,\ldots,\tau_N\}, where τ1=⋯=τN=τ=τey\tau_1 = \cdots= \tau_N = \tau= \tau e_y. The equation of motion of the particles follow the mobility problem:

X˙=V=MVF⋅F+MVT⋅T(14)\dot{\mathbf{X}} = \mathbf{V} = \mathbf{M}^{VF} \cdot\mathbf{F} + \mathbf{M}^{VT} \cdot\mathbf{T} \tag*{(14)}

where MVFM^{VF} and MVTM^{VT} are the mobility matrices that relate the particle velocities to their external forces and torques respectively. Similarly to Section I, they are obtained by applying the Faxen laws to the flow induced by the particles above the interface, see [3, 4] for more details on their calculation. This modelling includes only leading-order corrections for the finite size of the particle to limit the computational cost required to simulate large numbers (O(104)O(10^{4})) of particles. Even though this low-resolution model overestimates the particle mobility, our previous work shows that it reproduces qualitatively well the experimental results above a no-slip boundary [4, 5, 6].

The external force on each microroller fi\mathbf{f}_i is the resultant of a short-ranged repulsion between particles fP\mathbf{f}_P, a short range repulsion with the interface fI\mathbf{f}_I to avoid overlaps, and gravity fG\mathbf{f}_G, so that fi=fP+fI+fG\mathbf{f}_i = \mathbf{f}_P + \mathbf{f}_I + \mathbf{f}_G. The short range repulsive forces fP\mathbf{f}_P and fI\mathbf{f}_I derive from an exponentially decaying repulsive potential of the form [5]

U(r)={U0(1+d−rb)if r<d,U0exp⁡(d−rb)if r≥d.(15)U(r) = \begin{cases} U_0\left(1+\frac{d-r}{b}\right) & \text{if } r<d,\\ U_0\exp\left(\frac{d-r}{b}\right) & \text{if } r\ge d. \end{cases} \tag*{(15)}

For particle-particle interactions, rr is the center-to-center distance and d=2ad=2a. For particle-interface interactions, rr is the height of the particle center and d=ad=a. The energy scale U0U_0 and interaction range bb control the strength and decay of the potential respectively. We found that taking b=0.1ab=0.1a, U0=τ/2U_0=\tau/2 for particle-interface interactions and b=0.1ab=0.1a, U0=τ/26U_0=\tau/26 for particle-particle interactions prevents overlaps while keeping close contact.

In the large scale particle simulations, a small amount of gravity is added: fG=−mgez\mathbf{f}_G=-mg\mathbf{e}_z, with mg=0.0015τ/amg=0.0015\tau/a. This is done because, in the absence of gravity, the front particles are lifted up quite high by the rear particles when ξ=+∞\xi=+\infty. As a result, the transverse instability would take a long time to develop and, because it is proportional to the particle height [4, 5, 7], the wavelength of the instability would be large and becomes similar to the domain size. The addition of a small gravitational force slightly reduces the height of the front particles and therefore increases the growth rate of the instability and allows to have several wavelengths along the domain. However, the effects of gravity remain negligible compared to the torque induced flows (τ≫mga\tau\gg mga) and we have carefully checked that it did not change the qualitative behaviour of the suspension for all the values of ξ\xi simulated.

Particle trajectories (S8) are integrated with the two-step Adams–Bashforth scheme. Mobility-vector products and steric interactions are computed with PyCUDA on an Nvidia Titan V GPU. The typical simulation time is 30 min for 4,000 time iterations with 10,000 particles.

Two-lines model

This section describes the two-line model and the details of the linear stability analysis. This model has already been used to describe the transverse instability of micro-rollers above a no-slip wall [7]. The derivation and analysis below are a direct extension of this work, where only the Green’s function has been changed. For the sake of clarity we recall the governing equations and the main steps of the stability analysis.

Governing equations

Consider two infinite lines of rotlets, labelled L1\mathcal{L}_1, at the rear, and L2\mathcal{L}_2, at the front, rotating around the yy axis at a fixed height hh in a plane parallel to the fluid-fluid interface (see Figure S2). We denote x1(y,t)x_1(y,t) and x2(y,t)x_2(y,t) their position and ρ1(y,t)≡ρ(x1(y,t),y,t)\rho_1(y,t) \equiv\rho(x_1(y,t),y,t) and ρ2(y,t)≡ρ(x2(y,t),y,t)\rho_2(y,t) \equiv\rho(x_2(y,t),y,t) their rotlet densities. Here, we neglect out of plane motion in the zz-direction and only consider the velocities in the xyxy-plane. Although out of plane motion is seen in 3D simulations, our previous work [4, 7] shows that the fingering instability still occurs when considering particles confined to a plane above the surface.

Sketch of the two infinite lines

Figure S2. Sketch of the two infinite lines. The rotlet densities are rotating with a constant torque, τ\tau, about the yy-axis and interact hydrodynamically in the xyxy-plane at a height hh above the floor.

The model is governed by the equations of motion of each line together with the conservation of rotlets along their length:

∂xi(y,t)∂t=ux(xi(y,t),y)\frac{\partial x_i(y,t)}{\partial t}=u_x(x_i(y,t),y)
∂ρi(y,t)∂t=−∂[ρi(y,t)uy(xi(y,t),y)]∂y,i=1,2,(16)\frac{\partial\rho_i(y,t)}{\partial t}=-\frac{\partial[\rho_i(y,t)u_y(x_i(y,t),y)]}{\partial y},\quad i=1,2, \tag*{(16)}

where length and time have been rescaled with lc=hl_c=h and tc=8πηl3/τt_c=8\pi\eta l^3/\tau respectively. These four PDEs are nonlinear and nonlocal. The velocity at a given point (xi,y)(x_i,y) along a line LiL_i is given by the sum of the flows induced by each line,

ux(xi,y)=∑j=12∫−∞+∞μxuτ(xi−xj,y−y′)ρj(y′)dy′,(17)u_x(x_i,y)=\sum_{j=1}^{2}\int_{-\infty}^{+\infty}\mu_x^{u\tau}(x_i-x_j,y-y')\rho_j(y')\mathrm{d}y', \tag*{(17)}
uy(xi,y)=∑j=12∫−∞+∞μyuτ(xi−xj,y−y′)ρj(y′)dy′,(18)u_y(x_i,y)=\sum_{j=1}^{2}\int_{-\infty}^{+\infty}\mu_y^{u\tau}(x_i-x_j,y-y')\rho_j(y')\mathrm{d}y', \tag*{(18)}

where μuτ(x−x′)\boldsymbol{\mu}^{u\tau}(\boldsymbol{x}-\boldsymbol{x}') is the operator which provides the fluid velocity in the plane at position x=(x,y,h)\boldsymbol{x}=(x,y,h) induced by a rotlet at position x′=(x′,y′,h)\boldsymbol{x}'=(x',y',h). It is given by

μxuτ(x−x′,y−y′)=12[(∇x′×G(x−x′,y−y′;h))⋅ey]⋅ex(19)\mu_x^{u\tau}(x-x',y-y')=\frac{1}{2}\left[(\nabla_{x'}\times\boldsymbol{G}(x-x',y-y';h))\cdot\boldsymbol{e}_y\right]\cdot\boldsymbol{e}_x \tag*{(19)}
=ξξ+16h3(x−x′)2R5−1ξ+12h3R3(20)=\frac{\xi}{\xi+1}\frac{6h^3(x-x')^2}{R^5}-\frac{1}{\xi+1}\frac{2h^3}{R^3} \tag*{(20)}
=ξξ+1uxW+1ξ+1uxFS(21)=\frac{\xi}{\xi+1}u_x^W+\frac{1}{\xi+1}u_x^{FS} \tag*{(21)}
μyuτ(x−x′,y−y′)=12[(∇x′×G(x−x′,y−y′;h))⋅ey]⋅ey(22)\mu_y^{u\tau}(x-x',y-y')=\frac{1}{2}\left[(\nabla_{x'}\times\boldsymbol{G}(x-x',y-y';h))\cdot\boldsymbol{e}_y\right]\cdot\boldsymbol{e}_y \tag*{(22)}
=ξξ+16h3(x−x′)(y−y′)R5+1ξ+1×0(23)=\frac{\xi}{\xi+1}\frac{6h^3(x-x')(y-y')}{R^5}+\frac{1}{\xi+1}\times0 \tag*{(23)}
=ξξ+1uyW+1ξ+1uyFS(24)=\frac{\xi}{\xi+1}u_y^W+\frac{1}{\xi+1}u_y^{FS} \tag*{(24)}

where R=[(x−x′)2+(y−y′)2+4h2]1/2R=[(x-x')^2+(y-y')^2+4h^2]^{1/2}. As shown on Figure S3 the free-surface contribution has no transverse flow: uyFS=0u_y^{FS}=0.

Flow induced by a microroller in the plane parallel to the surface

Figure S3. Flow induced by a microroller in the plane parallel to the surface. a) No-slip wall (W), b) Free-surface (FS).

Linear stability analysis

The base state of this system corresponds to two straight lines with uniform rotlet densities (ρi(y,t)=ρ0)(\rho_i(y,t)=\rho_0) translating at a steady speed, so that their position are constant in the moving frame (x1(y,t)=0, x2(y,t)=d)(x_1(y,t)=0,\ x_2(y,t)=d), where dd is the initial distance between the two lines.

We perturb the system about the base state

x1(y,t)=0+δx1(y,t)(25)x_1(y,t) = 0 + \delta x_1(y,t) \tag*{(25)}
x2(y,t)=d+δx2(y,t),(26)x_2(y,t) = d + \delta x_2(y,t), \tag*{(26)}
ρ1(y,t)=ρ0+δρ1(y,t)(27)\rho_1(y,t) = \rho_0 + \delta\rho_1(y,t) \tag*{(27)}
ρ2(y,t)=ρ0+δρ2(y,t),(28)\rho_2(y,t) = \rho_0 + \delta\rho_2(y,t), \tag*{(28)}

where δx1,δx2≪min⁡(d,h)\delta x_1, \delta x_2 \ll\min(d,h) and δρ1,δρ2≪ρ0\delta\rho_1, \delta\rho_2 \ll\rho_0.

After Taylor expanding the functionals in Eqs. (16)–(18) we obtain the linearized governing equations

∂δx1∂t=ρ0G0δx1+ρ0G1∗δx2+μxuτ(0−d,⋅)∗δρ2(29)\frac{\partial\delta x_1}{\partial t} = \rho_0 \mathcal{G}_0 \delta x_1 + \rho_0 \mathcal{G}_1 * \delta x_2 + \mu_x^{u\tau}(0-d,\cdot) * \delta\rho_2 \tag*{(29)}
∂δx2∂t=−ρ0G0δx2−ρ0G1∗δx1+μxuτ(0−d,⋅)∗δρ1(30)\frac{\partial\delta x_2}{\partial t} = -\rho_0 \mathcal{G}_0 \delta x_2 - \rho_0 \mathcal{G}_1 * \delta x_1 + \mu_x^{u\tau}(0-d,\cdot) * \delta\rho_1 \tag*{(30)}
∂δρ1∂t=−ρ0∂[(ρ0G2∗δx1+ρ0G3∗δx2+μyuτ(0−d,⋅)∗δρ2)]∂y(31)\frac{\partial\delta\rho_1}{\partial t} = -\rho_0 \frac{\partial\left[(\rho_0 \mathcal{G}_2 * \delta x_1 + \rho_0 \mathcal{G}_3 * \delta x_2 + \mu_y^{u\tau}(0-d,\cdot) * \delta\rho_2)\right]}{\partial y} \tag*{(31)}
∂δρ2∂t=−ρ0∂[(ρ0G2∗δx2+ρ0G3∗δx1−μyuτ(0−d,⋅)∗δρ1)]∂y,(32)\frac{\partial\delta\rho_2}{\partial t} = -\rho_0 \frac{\partial\left[(\rho_0 \mathcal{G}_2 * \delta x_2 + \rho_0 \mathcal{G}_3 * \delta x_1 - \mu_y^{u\tau}(0-d,\cdot) * \delta\rho_1)\right]}{\partial y}, \tag*{(32)}

where we have used the symmetries of μxuτ\mu_x^{u\tau} and μyuτ\mu_y^{u\tau}. The star “∗*” denotes the one dimensional convolution product. The interaction kernels are given by

G0=∂∂x1[∫−∞+∞μxuτ(x1−d,y−y′) dy′]∣x1=0(33)\mathcal{G}_0 = \left.\frac{\partial}{\partial x_1}\left[\int_{-\infty}^{+\infty}\mu_x^{u\tau}(x_1-d,y-y')\,\mathrm{d}y'\right]\right|_{x_1=0} \tag*{(33)}
G1(y−y′)=∂μxuτ(0−x2,y−y′)∂x2∣x2=d,(34)\mathcal{G}_1(y-y') = \left.\frac{\partial\mu_x^{u\tau}(0-x_2,y-y')}{\partial x_2}\right|_{x_2=d}, \tag*{(34)}
G2(y−y′)=∂μyuτ(0−x1,y−y′)∂x1∣x1=0,(35)\mathcal{G}_2(y-y') = \left.\frac{\partial\mu_y^{u\tau}(0-x_1,y-y')}{\partial x_1}\right|_{x_1=0}, \tag*{(35)}
G3(y−y′)=∂μyuτ(0−x2,y−y′)∂x2∣x2=d.(36)\mathcal{G}_3(y-y') = \left.\frac{\partial\mu_y^{u\tau}(0-x_2,y-y')}{\partial x_2}\right|_{x_2=d}. \tag*{(36)}

Note that G0\mathcal{G}_0 is a constant which depends only on the geometric parameters of the problem dd, hh.

Looking for periodic solutions of Eqs. (S23)-(S26) of the form

u=(δx1,δx2,δρ1,δρ2)=u~ eiky+σt,\mathbf{u}=(\delta x_1,\delta x_2,\delta\rho_1,\delta\rho_2)=\tilde{\mathbf{u}}\,e^{\mathrm{i}ky+\sigma t},

where the wavenumber k=2π/λk=2\pi/\lambda and

u~=(δx~1,δx~2,δρ~1,δρ~2),\tilde{\mathbf{u}}=(\delta\tilde{x}_1,\delta\tilde{x}_2,\delta\tilde{\rho}_1,\delta\tilde{\rho}_2),

we obtain the following eigenvalue problem

Au~=σu~(37)A\tilde{\mathbf{u}}=\sigma\tilde{\mathbf{u}} \tag*{(37)}

where

A=(ρ0G0ρ0G~10μ~xuτ(−d,k)−ρ0G~1−ρ0G0μ~xuτ(−d,k)0−ikρ02G~2−ikρ02G~30−ikρ0μ~yuτ(−d,k)−ikρ02G~3−ikρ02G~2ikρ0μ~yuτ(−d,k)0),(38)A= \begin{pmatrix} \rho_0\mathcal{G}_0 & \rho_0\tilde{\mathcal{G}}_1 & 0 & \tilde{\mu}_x^{u\tau}(-d,k) \\ -\rho_0\tilde{\mathcal{G}}_1 & -\rho_0\mathcal{G}_0 & \tilde{\mu}_x^{u\tau}(-d,k) & 0 \\ -\mathrm{i}k\rho_0^2\tilde{\mathcal{G}}_2 & -\mathrm{i}k\rho_0^2\tilde{\mathcal{G}}_3 & 0 & -\mathrm{i}k\rho_0\tilde{\mu}_y^{u\tau}(-d,k) \\ -\mathrm{i}k\rho_0^2\tilde{\mathcal{G}}_3 & -\mathrm{i}k\rho_0^2\tilde{\mathcal{G}}_2 & \mathrm{i}k\rho_0\tilde{\mu}_y^{u\tau}(-d,k) & 0 \end{pmatrix}, \tag*{(38)}

and the tilde denotes the Fourier transform with respect to yy, e.g.

μ~xuτ(−d,k)=∫−∞∞μxuτ(−d,y)exp⁡(−iky) dy.(39)\tilde{\mu}_x^{u\tau}(-d,k)=\int_{-\infty}^{\infty}\mu_x^{u\tau}(-d,y)\exp(-\mathrm{i}ky)\,\mathrm{d}y. \tag*{(39)}

For the sake of conciseness, the exact expressions of the entries of AA are not reported here. They correspond to linear combinations of the expressions given in the main text and appendix A of [7]. Due to the particular structure of AA, Eq. (37) can be solved analytically. The four solutions σ1(k,ξ),…,σ4(k,ξ)\sigma_1(k,\xi),\ldots,\sigma_4(k,\xi) are given by

σ1,2,3,4(k,ξ)=±12[2A112−2A122+4A32A14−2A342±2(A114−2A112A122+4A112A14A32+2A112A342−8A11A14A31A34+A124−4A122A14A32−2A122A342+8A12A14A32A34+4A142A312−4A14A32A342+A344)1/2]1/2.(40)\sigma_{1,2,3,4}(k,\xi)=\pm\frac{1}{2}\left[2A_{11}^{2}-2A_{12}^{2}+4A_{32}A_{14}-2A_{34}^{2}\pm2\left(A_{11}^{4}-2A_{11}^{2}A_{12}^{2}+4A_{11}^{2}A_{14}A_{32}+2A_{11}^{2}A_{34}^{2}-8A_{11}A_{14}A_{31}A_{34}+A_{12}^{4}-4A_{12}^{2}A_{14}A_{32}-2A_{12}^{2}A_{34}^{2}+8A_{12}A_{14}A_{32}A_{34}+4A_{14}^{2}A_{31}^{2}-4A_{14}A_{32}A_{34}^{2}+A_{34}^{4}\right)^{1/2}\right]^{1/2}. \tag*{(40)}

The first two eigenvalues are real and of opposite sign σ1=−σ2\sigma_1=-\sigma_2. The two other are imaginary and conjugate σ3=σˉ4\sigma_3=\bar{\sigma}_4. Figure 3b in the main text shows the positive real eigenvalue σ1\sigma_1, i.e. the unstable one, for d=10hd=10h.

Effect of the viscosity ratio on the growth rate at large wavenumbers

One specificity of the two-lines model, that is not seen in 3d numerical simulations, is the existence of a plateau at large kk (see Figure 3b in the main text). This plateau was already identified in the case of a no-slip wall [6]. This plateau is specific to the two-lines model, which, even though far from the actual system, is the simplest representation we could find to carry out a linear stability analysis analytically. Its value is set by the term G0\mathcal{G}_0 in eqs. (S23)-(S24) and (S27): lim⁡k→∞σ=ρ0∣G0∣\lim_{k\to\infty}\sigma=\rho_0|\mathcal{G}_0|. This term represents the uniform, unidirectional, advective flow induced by a straight line with a uniform density of rotlets on the other line at a distance dd. Because it depends on the local distance between the lines, this flow amplifies shape perturbations of the other line at all scales (see Figure 3d in [6] for an example above a no-slip wall). In the case of a fluid-fluid interface, it has two contributions, one from the wall term and one from free-surface term:

G0=ξξ+1G0W+1ξ+1G0FS(41)\mathcal{G}_0=\frac{\xi}{\xi+1}\mathcal{G}_0^W+\frac{1}{\xi+1}\mathcal{G}_0^{FS} \tag*{(41)}
=ξξ+183d(d2−4h2)(d2+4h2)3−1ξ+143d(d2+4h2)2(42)=\frac{\xi}{\xi+1}\frac{8}{3}\frac{d(d^2-4h^2)}{(d^2+4h^2)^3}-\frac{1}{\xi+1}\frac{4}{3}\frac{d}{(d^2+4h^2)^2} \tag*{(42)}

where dd is the separation distance between the two lines and hh their height above the interface. Note that G0\mathcal{G}_0 does not depend on kk and therefore, unlike the other terms in the stability analysis, does not vanish in the large kk limit. In the case of a no-slip wall (ξ=+∞\xi=+\infty), G0\mathcal{G}_0 is positive because the lines move forward. Above a free-surface (ξ=0\xi=0), G0\mathcal{G}_0 is negative because the lines move backward. There is therefore a value of the viscosity ratio ξ∗\xi_{*} for which these two contributions cancel:

ξ∗=12d2+4h2d2−4h2.(43)\xi_{*} = \frac{1}{2}\frac{d^{2}+4h^{2}}{d^{2}-4h^{2}}. \tag*{(43)}

When ξ=ξ∗\xi= \xi_{*}, the plateau at large kk vanishes (G0=0G_{0}=0). In the case of Figure 3d, I had chosen d=10hd=10h, which gives ξ∗=13/24=0.541667\xi_{*}=13/24=0.541667.

Figure S4 below shows the growth rate in the range close to ξ∗\xi_{*}. We can clearly see that the value of the plateau decreases, vanishes at ξ∗=13/24\xi_{*}=13/24 and increases again to reach ρ0∣G0∣=1.03×10−4\rho_{0}|G_{0}|=1.03 \times10^{-4} (in dimensionless units) at ξ=0\xi=0.

Growth rate of the two-line model for viscosity ratios from 0 to 1

Figure S4. Growth rate of the two-line model in the range ξ∈[0;1]\xi\in[0;1]

However, even though this G0G_{0} term arises naturally in the two-lines model, it does not

affect the lengthscale selection of the transverse instability: if one removes it from the linear stability analysis (see Figure S5), the fastest growing mode is still preserved, the plateau disappears and the system because fully stable in the free-surface case (as in the 3d Stokesian Dynamics simulation shown on Figure 3a).

Growth rate of the two-line model

Figure S5. Growth rate of the two-line model in the range ξ∈[0;106]\xi\in[0;10^{6}] without the term G0\mathcal{G}_{0}.

Details on the mean-field simulations

The nonlocal mean-field equation Eq. (6) in the main text is solved numerically with a third-order second-time finite volume method optimally designed for hyperbolic scalar conservation equations [8, 9]. The rectangular computational domain is discretized with a cartesian grid with cell sizes Δx\Delta x and Δz\Delta z respectively. The domain dimensions LxL_x and LzL_z are chosen to ensure that the roller density remains small at the boundaries (ρ/ρ0<ϵ=10−4\rho/\rho_{0}<\epsilon=10^{-4}) during the simulation time. The time step Δt\Delta t is automatically chosen at each time iteration to ensure that the CFL condition is always satisfied: Δt=0.9min⁡(Δx/ux,Δz/uz)\Delta t=0.9\min(\Delta x/u_x,\Delta z/u_z). The convolution products due to hydrodynamic interactions,

ux(x,t)=∫PμxT(x−x′)ρ(x′,t) dx′(44)u_x(\mathbf{x},t)=\int_{\mathcal{P}}\mu_x^{\mathrm{T}}(\mathbf{x}-\mathbf{x}')\rho(\mathbf{x}',t)\,\mathrm{d}\mathbf{x}' \tag*{(44)}
uz(x,t)=∫PμzT(x−x′)ρ(x′,t) dx′,(45)u_z(\mathbf{x},t)=\int_{\mathcal{P}}\mu_z^{\mathrm{T}}(\mathbf{x}-\mathbf{x}')\rho(\mathbf{x}',t)\,\mathrm{d}\mathbf{x}', \tag*{(45)}

are evaluated with a Riemann sum using PyCUDA on an Nvidia Titan V GPU for efficiency. The typical simulation time is 50 min for 9,000 time iterations with N=Nx×Nz=1500×200=3⋅105N=N_x\times N_z=1500\times200=3\cdot10^{5} grid cells.

Note that corrections for the finite size of the particles could be included in the hydrodynamic interactions in Eq. (7) (as done in the particle simulations) with no additional complexity, but these effects are negligible at the macroscopic level where collective effects dominate at large scales.

Movie Captions

  • SI Movie 1, “Particle_simulations_compare_xi_top_view.avi”: Top view of the time-evolution of 10,000 microrollers initially uniformly distributed in a monolayer above the interface. Each color, from dark blue to yellow, corresponds to a different value of ξ∈{0,0.5,1.5,5,+∞}\xi\in\{0, 0.5, 1.5, 5, +\infty\} simulated independently with the 3D Stokesian Dynamics approach.

  • SI Movie 2, “Particle_simulations_compare_xi_side_view.avi”: Side view of the numerical simulations shown in SI Movie 1.

  • SI Movie 3, “Continuum_model_xi_0.mp4”: numerical simulation of the continuum model above an interface with viscosity ratio ξ=0\xi= 0 (free-slip). The rollers are initially uniformly distributed (ρ/ρ0=1\rho/\rho_0 = 1) over a thin strip of aspect ratio γ=L/H=9.4\gamma= L/H = 9.4 near the interface. Colorbar: normalized particle density ρ/ρ0\rho/\rho_0. Black arrow: fluid velocity field. The boundary of the cluster is defined by the green isocontour ρ/ρ0=0.4\rho/\rho_0 = 0.4. The motion along the xx-axis is expressed relative to the initial position of the center of mass, xC0=xC(t=0)x_C^0 = x_C(t = 0), in order to show how far the cluster has travelled at the end of the simulation.

  • SI Movie 4, “Continuum_model_xi_0_5.mp4”: same as SI Movie 3 but with ξ=0.5\xi= 0.5.

  • SI Movie 5, “Continuum_model_xi_1_1.mp4”: same as SI Movie 3 but with ξ=1.1\xi= 1.1.

  • SI Movie 6, “Continuum_model_xi_5_0.mp4”: same as SI Movie 3 but with ξ=5\xi= 5.

  • SI Movie 7, “Continuum_model_xi_infty.mp4”: same as SI Movie 3 but with ξ=+∞\xi= +\infty (no-slip wall).

References

  1. [3]J. W. Swan and J. F. Brady, Simulation of hydrodynamically interacting particles near a no-slip boundary, Physics of Fluids (1994-present) 19, 113306 (2007).DOI
  2. [4]M. Driscoll, B. Delmotte, M. Youssef, S. Sacanna, A. Donev, and P. Chaikin, Unstable fronts and motile structures formed by microrollers, Nature Physics (2016).DOI
  3. [5]F. B. Usabiaga, B. Delmotte, and A. Donev, Brownian dynamics of confined suspensions of active microrollers, J. Chem. Phys. 146, 134104 (2017), software available at https://github.com/ stochastichydroTools/RigidMultiblobsWall.DOI
  4. [6]B. Delmotte, M. Driscoll, P. Chaikin, and A. Donev, Hydrodynamic shocks in microroller suspensions, Physical Review Fluids 2, 092301(R) (2017).arxiv.org/abs/1702.03350
  5. [7]B. Delmotte, A. Donev, M. Driscoll, and P. Chaikin, Minimal model for a hydrodynamic fingering instability in microroller suspensions, Physical Review Fluids 2, 114301 (2017).
  6. [8]J. B. Bell, C. N. Dawson, and G. R. Shubin, An unsplit, higher order godunov method for scalar conservation laws in multiple dimensions, Journal of Computational Physics 74, 1 (1988).DOI
  7. [9]S. May, A. Nonaka, A. Almgren, and J. Bell, An unsplit, higher-order godunov method using quadratic reconstruction for advection in two dimensions, Communications in Applied Mathematics and Computational Science 6, 27 (2011).DOI

Paper details

vlnc:W_x9wkuwtf
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents