2.1. Continuity equation for a Poiseuille flow

We consider the laminar flow of figure 1, whose unidirectional velocity lies along the $x$ axis. This configuration implies that the channel's bed, and thus the flow, does not vary in the $x$ direction at all. Of course, such an idealized setting can only be an approximate representation of the actual flow, valid when the latter varies only weakly along the $x$ direction.

To formalize this assumption, we call $\mathcal {L}$ the cross-wise extent of the channel ($y$ axis), and ${\mathcal {L}}_s$ the characteristic length over which the bed changes downstream ($x$ axis). When the ratio ${\mathcal {L}}/{\mathcal {L}}_s$ is small enough, we may treat the flow as streamwise invariant, and represent it with its only component, $u(y,z)$. We use this approximation from now on. Under this assumption, the Stokes equation reduces to

(2.1)\begin{equation} \nu \nabla^2 u + g S = 0, \end{equation}

where $\nu$ is the kinematic viscosity of the fluid, $g$ the acceleration of gravity and $S$ the downstream slope of the channel (the sine of the angle it forms with the horizontal). The Laplacian operator, $\nabla ^2$, applies only to directions orthogonal to the flow, that is, $y$ and $z$.

A convenient consequence of our initial assumption that the flow is invariant along the $x$ axis is that its free surface is also invariant. This assumption is very limiting: it precludes any free-surface wave and, in fact, any pressure gradient other than hydrostatic. In short, it limits the present theory to two-dimensional Poiseuille flows. This, we argue, is nonetheless a rich class of systems, to which the straight laminar rivers of Abramian et al. (Reference Abramian, Devauchelle and Lajeunesse2020) belong.

We further neglect surface tension, and any other stress on the free surface. Unlike the unidirectional-flow assumption, this hypothesis could well be relaxed in the present framework but, to keep our problem simple, we here assume that the free surface is horizontal (more exactly, it belongs to the $(x,y)$ plane). We then place the origin on the free surface, which thus corresponds to $z=0$. Accordingly,

(2.2)\begin{equation} \frac{\partial u}{\partial z} = 0\quad \mbox{for}\ z = 0. \end{equation}

Finally, we treat the bed as an impervious wall, where the velocity of the fluid vanishes

(2.3)\begin{equation} u = 0 \quad \mbox{for}\ z ={-}D(y), \end{equation}

where $D(y)$ is the flow depth. This assumption is only a rough approximation of reality when the bed is made of a granular sediment, into which the velocity profile can penetrate. One could account for this penetration by allowing some slip at the porous surface (Beavers & Joseph Reference Beavers and Joseph1967). Similarly, in the context of flowing ice, a slippery bed surface would also translate into a Robin boundary condition (Schoof & Hewitt Reference Schoof and Hewitt2013). Here, for simplicity, we just keep (2.3) until Appendix C.2, where we touch upon the subject of slip.

Equation (2.1), and the associated boundary conditions (2.2) and (2.3), represent a Nusselt flow down a channel of arbitrary section (Benjamin Reference Benjamin1957). They sometimes have closed-form solutions, for instance when the channel is elliptic (§ 3.1), but in general they do not, and one needs to approximate their solution with a series expansion, or some other numerical method (§§ 3.2 and 3.4). Here, (2.1) to (2.3) serve as our reference, against which we can evaluate the new shallow-water approximation we introduce in § 2.4. In that sense, we refer to them as ‘the exact equations’ although, of course, they are but an approximation of the actual flow.

2.2. Integrated momentum balance

Integrating the exact momentum balance (2.1) over a vertical section, and invoking the free-surface boundary condition (2.2), we find

(2.4)\begin{equation} \nu \int_{{-}D}^0 \frac{\partial^2 u}{\partial y^2} \mathrm{d} z - \frac{\tau_z}{\rho} + g S D = 0, \end{equation}

where $\rho$ is the density of the fluid and $\tau _z$ is the vertical component of the viscous stress on the bed's surface. In a Newtonian fluid, the latter reads

(2.5)\begin{equation} \tau_z = \rho \nu \left. \frac{\partial u}{\partial z} \right|_{z={-}D}. \end{equation}

A more usual notation for $\tau _z$ would be $\tau _{xz}$, but in the present context the shorthand notation $\tau _z$ leaves no ambiguity.

To rewrite the integral in (2.4), we first remember that the free surface is flat and, accordingly, differentiate only the lower bound of the integral along the bed's surface

(2.6)\begin{equation} \int_{{-}D}^0 \frac{\partial^2 u}{\partial y^2} \mathrm{d} z = \frac{\partial}{\partial y } \int_{{-}D}^0 \frac{\partial u}{\partial y} \mathrm{d} z - \frac{\partial D}{\partial y} \left.\frac{\partial u}{\partial y} \right|_{z={-}D}. \end{equation}

We then use the no-slip boundary condition (2.3) which, again, we differentiate across the stream

(2.7)\begin{equation} \left. \frac{\partial u}{\partial y} \right|_{z={-}D} = \frac{\partial D}{\partial y} \left. \frac{\partial u}{\partial z} \right|_{z={-}D}. \end{equation}

The above relation tells us that the vertical component of the shear stress, $\tau _z$, its cross-wise counterpart, $\tau _y$, and the norm of the shear stress, $\tau$, are all related to each other through

(2.8)\begin{equation} \tau\equiv \sqrt{ \tau_y^2 + \tau_z^2 } = \tau_z \sqrt{ 1 + \left( \frac{\partial D}{\partial y} \right)^2 }. \end{equation}

Since the above relation is exact, and since the shape of the channel, $D(y)$, is a given of the problem, we can treat $\tau _z$ and $\tau$ as equivalent quantities. Hereafter, based on this observation, we will use $\tau _z$ or $\tau$ interchangeably, depending on which one is more convenient, or more telling.

The no-slip boundary condition (2.3) also allows us to write

(2.9)\begin{equation} \int_{{-}D}^0 \frac{\partial u}{\partial y} \mathrm{d} z = \frac{\partial}{\partial y} \int_{{-}D}^0 u \, \mathrm{d} z, \end{equation}

which, together with (2.7), we inject in the vertical integral of the Poisson equation (2.4). We finally find

(2.10)\begin{equation} \rho \nu \frac{\partial^2}{\partial y^2} (U D) - \tau_z \left(1 + \left( \frac{\partial D}{\partial y}\right)^2 \right) + \rho g S D = 0, \end{equation}

where we have defined the vertically averaged velocity $U$ as

(2.11)\begin{equation} U = \frac{1}{D} \int_{{-}D}^0 u \, \mathrm{d} z. \end{equation}

The one-dimensional momentum balance (2.10) follows from (2.1) to (2.3) without any additional assumption – in that sense, it is exact. The first term is the divergence of the cross-stream flux of momentum, the second one is the momentum lost to friction with the bottom, and the last one is the source of downstream momentum, powered by gravity.

A similar balance could be established without assuming that the free surface is flat, and the flow unidirectional; it would then explicitly couple the transverse flow to the deformations of the free surface. Assuming the flow to be unidirectional thus simplifies (2.10), but it makes the present theory unable to handle bedforms (unless they are invariant along the $x$ direction, § 3.3).

Even in the simple form of (2.10), the integrated momentum balance is incomplete: it involves two unknown functions of the cross-wise coordinate $y$ (the shear stress $\tau _z$ and the average velocity $U$). This was to be expected, since we just integrated the Poisson equation vertically, without trying to solve it. At this point, unfortunately, we can postpone no further resorting to approximation.

2.3. Classical lubrication approximation

The lubrication theory relies on the flow being shallow. To quantify this statement, we define a small parameter, $\epsilon$, as the aspect ratio of the channel's cross-section

(2.12)\begin{equation} \epsilon = \frac{\mathcal{H}}{\mathcal{L}}, \end{equation}

where $\mathcal {H}$ is the characteristic depth of the channel. Upon rescaling of the $y$ and $z$ coordinates with $\mathcal {L}$ and $\mathcal {H}$ respectively, the cross-wise derivative in the exact Poisson equation (2.10) is multiplied by $\epsilon ^2$ (Appendix B.1). Assuming $\epsilon$ is small, one can thus neglect the flux of momentum that viscosity transfers across the flow. The classical lubrication approximation is based on this idea.

Neglecting the cross-wise flux of momentum in the Poisson equation (2.10), we find the classical velocity profile of a Poiseuille flow,

(2.13)\begin{equation} u = \frac{ 3 U }{ 2 D^2 } (D^2 - z^2), \end{equation}

where the vertically averaged velocity $U$ is related to the vertical component of the shear stress on the bed, $\tau _z$, through

(2.14)\begin{equation} \tau_z = \frac{3 \rho \nu U}{D}. \end{equation}

Since the momentum that gravity delivers to the fluid does not propagate in the cross-wise direction, the surface of the bed needs to absorb it all, and thus

(2.15)\begin{equation} \tau_z = \rho g S D. \end{equation}

Equations (2.13) to (2.15) are exact when the channel is perfectly flat and infinitely wide ($\epsilon =0$). They are also the leading-order terms in the expansion of the flow as a power series in $\epsilon$ (Appendix B.1). The gist of the classical lubrication approximation is to assume that these leading-order equations provide a decent approximation of the flow when $D$ varies slowly across the channel. In the case of a unidirectional flow in a straight channel, this simply translates into (2.13) to (2.15) – only with varying coefficients $U(y)$ and $D(y)$.

The main advantage of this approach is its wonderful simplicity. The cost of this simplicity, however, is that the resulting velocity $U(y)$ violates the momentum balance (2.10) at order $\epsilon ^2$. In the next section, we improve the lubrication approximation by proposing a representation of the flow that satisfies it exactly.

2.4. Revised lubrication theory

We would like to improve the classical lubrication approximation so that it accounts for the momentum balance, while keeping the problem low-dimensional. Our objective is to predict the distribution of shear stress on the channel's bed better than the classical theory, at a small computational cost. This endeavour is in line with the work of Marche (Reference Marche2007), although here we limit ourselves to the configuration of a steady flow in a straight channel.

To refine an approximate theory based on a series expansion, it is customary to look for the next term in the expansion – here, the term of order $\epsilon ^2$. Appendix B.1 is devoted to this standard method which, as expected, yields a correction to the classical lubrication approximation. Here, however, we propose a less conventional method based on the ansatz of the classical theory, in the hope of obtaining a more flexible representation of the flow. The expansion of Appendix B.1 will be our touchstone: to qualify as progress, the revised theory should be correct at least up to order $\epsilon ^2$.

The reasoning we propose does not have the mathematical rigour of an order expansion, but it is extremely simple. It relies on the observation that it is only by enforcing (2.15) that we make (2.14) violate the momentum balance (2.10). In itself, (2.14) only links the average velocity, $U$, with the shear stress on the bed, $\tau _z$, when the vertical velocity profile is parabolic (2.13). In short, (2.14) just relates two unknowns, and thus spares one degree of freedom.

It is, we believe, in the spirit of the shallow-water approximation to make the most of this freedom, by dropping (2.15) while keeping (2.14). Specifically, we use the shape of the vertical profile (2.13) as an ansatz which relates the average velocity, $U$, to the shear stress, $\tau _z$, through (2.14). In practice, this means substituting (2.14) into the momentum balance (2.10), which then reads

(2.16)\begin{equation} \frac{1}{3} \frac{\mathrm{d}^2}{\mathrm{d}y^2} ( D^2 \tau_z ) -\left( 1 + \left(\frac{\mathrm{d} D}{\mathrm{d}y} \right)^2 \right)\tau_z + \rho g S D = 0. \end{equation}

This second-order ordinary differential equation is the substitute we propose for (2.15). Its only unknown is $\tau _z$. The flow depth, $D(y)$, although it appears at various orders of derivation, is the given boundary to which the flow adjusts.

Equation (2.14), which relates the friction on the bottom to the average velocity, is essentially a friction law. The approximate momentum balance (2.16) depends on this friction law, and it is therefore just as reliable as the latter (Appendix C).

Before we set off to investigate the validity of (2.16), we first acknowledge some of its encouraging features. First, it becomes (2.15) again over a flat bed – as it should. More interestingly, it is equivalent to the exact Poisson equation (2.1) up to order $\epsilon ^2$, which ensures that it is an improvement upon the classical (zeroth-order) theory (Appendix B.2). Finally, that it is a differential equation instead of an algebraic one is only slightly inconvenient, since (2.16) is linear in $\tau _z$, and therefore amenable to classical resolution methods. We still need to check, however, that this equation is worth the effort of solving it.

The classical lubrication theory merely equates the vertical component of the shear stress, $\tau _z$, with the momentum that gravity injects into a vertical slice of the flowing fluid, $\rho g S D$. Thus, not only does it neglect the transfer of momentum across the stream, but also approximates the flux of momentum into the bed with its vertical component, $\tau _z$. The revised theory does neither. Indeed, integrating (2.16) between two points across the stream, say $y_1$ and $y_2$, and invoking (2.7), we find

(2.17)\begin{equation} \left[\frac{1}{3} \frac{\mathrm{d}}{\mathrm{d}y} ( D^2 \tau_z ) \right]_{y_1}^{y_2} - \int_{s(y_1)}^{s(y_2)} \tau \, \mathrm{d}s + \rho g S \int_{y_1}^{y_2} D \, \mathrm{d}y = 0, \end{equation}

where $s$ denotes the arclength along the bed's surface (this change of variable is based on (2.8)). The first term in the above equation represents the cross-wise flux of momentum, the second one is the total momentum transferred to the bed and the last one is the momentum injected into the flow by gravity. The first term is an approximation, whereas the two others are exact. Regardless of the correctness of the first term, however, the three of them balance each other exactly – provided $\tau _z$ fulfils (2.16).

In other words, (2.16) is a proper continuity equation for momentum. The downstream momentum that gravity constantly supplies to the flow is either transmitted to the bed, or distributed sideways by viscosity. To close this balance, therefore, one needs to account for the momentum that is transferred across the stream. The value of the shear stress at a specific location on the bed thus depends on the shape of the entire channel. In mathematical terms, (2.16) is non-local – a property inherited form the Poisson equation (2.1).

As encouraging as it may seem, the above features do not guarantee that (2.16) is a better approximation of the flow than the expansion of Appendix B.1, which is a local, algebraic equation: both are correct up to order $\epsilon ^2$. We devote the next section to testing (2.16) in a variety of – hopefully instructive – configurations.

3.1. Exact solutions

3.1.1. Wedge flow

Let us first consider a Poiseuille flow in a wedge (figure 2a). This configuration may represent a river bank, where the water surface ($z = 0$) intersects the bed ($z = -\mu y$). Mathematically, this problem is ill posed, since it lacks a boundary condition on the open side of the wedge. We could fix such a condition, of course, but this would rule out any closed-form solution. Close enough to the corner, anyway, we expect that the flow velocity will be insensitive to the boundary condition, and will thus behave like

(3.1)\begin{equation} u =\frac{gS}{\nu} \frac{( \mu y - z )( \mu y + z )}{ 2( 1 - \mu^2 ) } =\frac{gS}{\nu} \frac{ D^2 - z^2 }{ 2( 1 - \mu^2 ) }, \end{equation}

an expression that satisfies (2.1) exactly (except for $\mu =1$, § 4). Based on this expression, the shear stress the flow exerts on the bed's surface is proportional to the water depth, $\mu y$

(3.2)\begin{equation} \tau_z = \frac{ \rho g S \mu y }{ 1 - \mu^2 }. \end{equation}

We now need to evaluate the revised lubrication approximation against this result. This is straightforward: as it turns out, (3.2) is also a solution of (2.16). This is not true for the classical lubrication approximation since, in a wedge, (2.15) reads $\tau _z = \rho g S \mu y$. The revised theory thus proves an improvement in this case, at least when the bank is not too steep ($\mu <1$; § 4 is devoted to $\mu \geq 1$).

Figure 2.

Exact solutions to both the two-dimensional Poisson equation (2.1), and the revised lubrication theory, (2.16). Aspect ratio is preserved. Blue shading show velocity contours. The width of the channel is $W$. (a) Wedge flow, (3.1) with $\mu = 1/3$. The flow extends to $y \rightarrow + \infty$ on the right-hand side. (b) Elliptic channel, with aspect ratio $R = 3.5$, (3.4).

3.1.2. Elliptic channel

We now turn our attention to a better-defined problem, and consider an elliptic channel of width $W$, and aspect ratio $R$ (figure 2b), the depth of which reads

(3.3)\begin{equation} D = \frac{1}{R} \sqrt{ W^2 - 4 y^2 }. \end{equation}

The corresponding solution of the Poisson equation (2.1) is (Boussinesq Reference Boussinesq1868, p. 388)

(3.4)\begin{equation} u = \frac{g S \left(W^2 - 4 y^2 - (Rz)^2\right)}{2 \nu \left( R^2 + 4 \right)}. \end{equation}

Differentiating the above equation, we find the expression of the vertical component of the shear stress on the bed's surface

(3.5)\begin{equation} \tau_z = \frac{\rho g S R \sqrt{W^2-4y^2}}{R^2+4}. \end{equation}

Again, this expression happens to be an exact solution to (2.16) – the approximation could not be more accurate. In fact, the velocity field associated with the revised lubrication approximation, namely

(3.6)\begin{equation} u = \frac{D}{2\rho \nu}\left( R^2 - \left( \frac{z}{D} \right)^2 \right) \tau_z, \end{equation}

is the exact solution of the original Poisson equation. This result, which seems coincidental at first, appears inevitable after careful consideration. Indeed, in an elliptic channel, the vertical velocity profile of the exact solution is parabolic, making the ansatz of the revised theory a perfect representation of the flow. Conversely, the classical lubrication theory overestimates the shear stress by a factor of $(R^2 + 4)/R^2$.

Although limited to channels of specific shapes, the perfect match between the revised theory and the reference problem is encouraging. In particular, it is remarkable that the revised theory works so well in an elliptic channel of arbitrary aspect ratio, where the hypothesis of a shallow flow does not hold at all. In the next section, we test the revised theory under even worse conditions.

3.2. Rectangular groove

Let us, for a change of context, imagine a straight microfluidic groove engraved into a solid substrate, with a rectangular cross-section (figure 3a,b). The flat bottom of the channel, of course, would be no challenge for the lubrication theory, were it not for the two sidewalls, where the no-slip boundary condition applies. There is no closed-form solution of the Poisson equation in a channel of rectangular cross-section, but there exists a solution in the form of an infinite series (e.g. White Reference White1991)

(3.7)\begin{equation} u = \frac{4 g SW^2}{\nu {\rm \pi}^3}\sum_{k=1,3,5,\ldots}^{\infty} \frac{({-}1)^{(k-1)/2}}{k^3}\left(1 - \frac{\cosh(k{\rm \pi} z/W)}{\cosh(k{\rm \pi} D/W)}\right)\cos \left(\frac{k{\rm \pi} y}{W} \right), \end{equation}

where $W$ is the width of the channel, and $D$ its (uniform) depth. In this sum, the index $k$ can be interpreted as a dimensionless wavenumber. Differentiating the above expression with respect to $z$ yields the bottom shear stress $\tau _z$ with arbitrary precision (blue line in figure 3(c,d), $\tau$ is exactly $\tau _z$ in this case). As expected, we find that viscosity conveys the influence of the walls across the flow, making the bottom shear stress reach a maximum in the middle of the channel.

Figure 3.

Laminar flow in a rectangular groove. (a,b) Streamwise velocity (blue shading), according to series expansion (3.7) truncated at $k=99$. Aspect ratio is preserved. (c,d) Intensity of flow-induced shear stress on bottom ($\tau$). Solid blue: series expansion (3.7) truncated at $k=99$, for reference. Dash-dotted grey: classical lubrication approximation (2.14). Dashed orange: present theory (3.8). Channel aspect ratio is $W/D=5$ (a,c) and $W/D=1$ (b,d).

The classical lubrication theory, in the form of (2.15), cannot account for the no-slip boundary condition at the sidewalls. In fact, according to this approximation, the shear stress just remains constant across the entire channel. As a consequence, although this constant is a reasonable estimate of the shear stress far away from the walls, the classical theory fails entirely in their neighbourhood (grey line in figure 3c,d).

Equation (2.16), conversely, requires two boundary conditions, which allows us to fix the velocity of the flow along the side walls: (2.3) translates the no-slip condition into $\tau _z=0$ in $y=\pm W/2$. The uniform depth of the channel makes this problem a textbook exercise, whose solution is

(3.8) \begin{equation} \tau_z =\rho g S D \left( 1 - \frac{\cosh ( \sqrt{3} y/ D )}{\cosh ( \sqrt{3} W/(2D) )} \right). \end{equation}

We find that this expression is a much better approximation of the actual shear stress than the constant of the classical theory (orange line in figure 3c,d). In particular, the sidewalls affect the flow over a distance comparable to the channel's depth. In a channel of aspect ratio $W/D=5$ (figure 3a,c), the largest error of the present theory is less than 10 % of the average shear stress (the average error is approximately 2 %). Combining (3.8) with the ansatz (2.13), and integrating the result across the channel, yields an estimate for the total discharge of a rectangular microfluidic groove; we find that this estimate lies within less than 3 % of the actual value, whereas the classical theory is more than 30 % off.

The revised lubrication theory relies on the friction law (2.14), which derives from the assumption that the flow profile is essentially that of a Nusselt film. Near the walls, however, this hypothesis breaks down, because the walls cause the profile to depart from its parabolic shape, thus affecting the friction law (Appendix C.1). It turns out, however, that the vicinity of a wall also lessens the contribution of bottom friction to the momentum balance. As a result, the friction term gets wrong where it does not matter much, which bolsters the validity of the revised lubrication theory near a wall.

In a narrower channel ($W/D=1$, figure 3b,d), we expect both the classical and the revised theories to fail. Indeed they do, but not to the same extent. The revised theory underestimates the discharge by approximately 12 %, whereas the classical theory overestimates it by a factor of almost 5. As visible on figure 3(d), this difference results from the ability of the revised theory to account for friction along the sidewalls (Appendix B.1).

The non-locality of (2.16) comes at a cost but, as this case illustrates, it keeps the momentum balance under control. This, we argue, is why the revised theory is more reliable than the classical theory.

3.3. Linear perturbation

We now consider a channel of infinite width, the bottom of which is perturbed with a sinusoidal corrugation of infinitesimal amplitude (figure 1). In tune with our initial assumption, the flow remains invariant in the $x$ direction; the crests and troughs of the corrugation, therefore, are aligned with the flow velocity. Mathematically,

(3.9)\begin{equation} D = D_0 + Re \left( {D}^{* }_1 \exp \left( \frac{\textrm{i}ky}{D_0} \right) \right), \end{equation}

where ${D}^{* }_1$ is the (possibly complex) amplitude of the perturbation ($| {D}^{* }_1|\ll D_0$), $D_0$ is the unperturbed depth of the channel and $k$ is the dimensionless wavenumber of the perturbation (the corresponding wavevector is aligned with the $y$ axis). When the perturbation vanishes, the bottom shear stress, $\tau _{z,0}$, is related to $D_0$ through (2.15).

Abramian et al. (Reference Abramian, Devauchelle and Lajeunesse2019a) investigated the stability of a similar perturbation when the bed over which the fluid flows is made of mobile sediment. Since it is the flow-induced shear stress that drives sediment transport in this case, one essential step of their analysis was to linearize the exact Poisson equation (our (2.1)), and to analytically solve the resulting linear problem. Their (3.7) will serve as our reference; in the present notations, it reads

(3.10)\begin{equation} \tau_z = \tau_{z,0} \left( 1 + \left( 1 - k \tanh{k} \right) Re \left( \frac{ {D}^{* }_1}{D_0} \exp \left( \frac{\textrm{i}ky}{D_0} \right) \right) \right). \end{equation}

The above equation indicates that the shear stress is in phase with the sinusoidal perturbation; its extrema lie where the perturbation's are. More remarkably, the sign of the shear-stress perturbation depends on the wavelength of the bed perturbation (blue line in figure 4). When this wavelength is large enough ($k< k_c$, where $k_c\approx 1.20$), the shear stress is at its highest in the troughs, where the flow is deeper – in tune with the classical lubrication theory. On the contrary, when the perturbation's wavelength is short enough ($k>k_c$), the crests of the corrugation find themselves more exposed to the bulk of the flow, and are therefore subjected to a stronger stress. In other words, the crests then shield the troughs, which collect a smaller share of the flow's momentum. This mechanism, the signature of the Laplacian in (2.1), is essentially the same as the shadowing that sets the shape of diffusion-limited aggregates (Witten & Sander Reference Witten and Sander1981) and causes the Saffman–Taylor instability (Saffman Reference Saffman1986). It is crucial to the stability analysis of Abramian et al. (Reference Abramian, Devauchelle and Lajeunesse2019a), as it prevents the perturbations with the shortest wavelengths to grow infinitely fast, thus selecting the wavelength most likely to appear in an experiment.

Figure 4.

Spectral response of bottom shear stress to a bed perturbation of infinitesimal amplitude. Blue line: linearized two-dimensional solution (3.10) for reference (Abramian et al. Reference Abramian, Devauchelle and Lajeunesse2019a). Blue dot: critical wavenumber $k_c\approx 1.20$. Dash-dotted grey line: classical lubrication approximation (3.11). Orange dashed line: present theory (3.13).

The classical lubrication theory cannot account for this phenomenon. Indeed, it directly relates the shear stress to the amplitude of the perturbation, regardless of its wavelength

(3.11)\begin{equation} \frac{ \tau_z - \tau_{z,0} }{ \tau_{z,0} }= \frac{ D - D_0 }{ D_0 }=Re \left( \frac{ {D}^{* }_1}{D_0} \exp \left(\frac{\textrm{i}ky}{D_0} \right) \right). \end{equation}

In the Fourier space, this translates into a flat response (grey line in figure 4). The present theory, on the other hand, is designed to account for the cross-wise flux of momentum. To compare it with (3.10), we first linearize (2.16) about a horizontal bed. We find

(3.12)\begin{equation} \frac{D_0^2}{3} \frac{\mathrm{d}^2}{\mathrm{d}y^2} \frac{\tau_{z,1}}{\tau_{z,0}} + \frac{2 D_0}{3} \frac{\mathrm{d}^2 D_1}{\mathrm{d}y^2} - \frac{\tau_{z,1}}{\tau_{z,0}} + \frac{D_1}{D_0}= 0, \end{equation}

where $D_1$ and $\tau _{z,1}$ are the linear corrections to the base state of the flow depth and the shear stress, respectively. Applying a Fourier transform to the above expression, we get

(3.13)\begin{equation} \frac{ {\tau}^{* }_{z,1}}{\tau_{z,0}} = \frac{1 - 2k^2/3}{1 + k^2/3} \frac{ {D}^{* }_1}{D_0}, \end{equation}

where the symbol $\, { }^{* }\,$ denotes the Fourier amplitude of the quantity it decorates. For long wavelengths ($k\ll 1$), this expression correctly approximates the exact linear result (3.10), but departs from it as the wavelength of the perturbation approaches the flow depth ($k \sim 1$) (orange line in figure 4). Before this happens, however, the present theory accounts for the sign change of the shear stress near $k_c$, which it estimates at $\sqrt {3/2}\approx 1.22$ – within 2 % of the true value.

The quality of the present approximation turns out to be surprisingly good. In a fashion quite typical of shallow-water theories, it works better than it should. Indeed, if we expand (3.12) and (3.13) into Taylor series, they match up to order $k^4$ only (odd orders all vanish); reasonably good, perhaps, but insufficient to account for the sign change in figure 4, since this well-controlled, but truncated, expansion fails to ever change sign, unlike the original equation (3.13). In short, it seems we can push the revised lubrication theory beyond where a well-controlled order expansion endorses it – at our own risks, of course.

Taking such risks, however, would be of little use if the revised lubrication theory were to fail beyond the linear regime, in which the two-dimensional equation it approximates can be solved relatively easily. The next section addresses this point.

3.4. Finite-amplitude corrugation

The revised lubrication theory fairly reproduces the results of the linearized, two-dimensional Poisson equation (§ 3.3). We now perturb a flat bed with a corrugation of finite amplitude, thus precluding any linearization of the problem. To test the present theory, we need a reference, which analytical calculations can no longer provide. Instead, we solve (2.1) in a corrugated channel with finite elements (FreeFem++, Hecht (Reference Hecht2012), blue shading in figure 5a,b). Upon differentiation with respect to $y$ and $z$, these numerical simulations yield an approximation of the flow-induced shear stress on the bed (blue line in figure 5c,d).

Figure 5.

Laminar flow above a corrugated bottom (a,b), and associated shear stress on the bed (c,d). Blue shading in (a,b) shows two-dimensional velocity (2.1) (finite-element simulations, arbitrary units). Flow is towards viewer. Aspect ratio is preserved. Solid blue lines in (c,d): shear stress from finite-element simulations. Dotted brown line: linearized two-dimensional flow (3.10) (Abramian et al. Reference Abramian, Devauchelle and Lajeunesse2019a). Dashed orange lines: present theory (2.16) (numerical solution): (a,c) $k=6$ and $D^*_1 = 0.03 D_0$; (b,d) $k=0.6$ and $D^*_1 = 0.7 D_0$.

We consider two configurations, in which the corrugations differ in amplitude and wavelength. In the first case, the wavelength of the corrugation is shorter than the flow depth ($k=6$), while its amplitude remains small, albeit not vanishing (${D}^{* }_1= 0.03 D_0$, figure 5a,c). We expect both the classical lubrication theory and the present theory to fail in this case, due to the short wavelength of the perturbation. In the second case, the wavelength is significantly larger than the flow depth ($k=0.6$), while the amplitude is comparable to the depth (${D}^{* }_1=0.7D_0$, figure 5b,d). It is in such a flow configuration that the revised theory could prove useful.

In the first case, the amplitude of the perturbation is small enough (with respect to depth) for the linearized theory of § 3.3 to provide a good approximation of the shear stress (dotted brown line in figure 5c). This is not true, however, for the classical lubrication theory, which wrongly locates the maxima of the shear stress in the troughs (dashed grey line). This was to be expected, since our choice of wavenumber is far beyond $k_c$, the wavenumber above which the crests shield the troughs (§ 3.3). The revised lubrication theory, in the form of a numerical solution of (2.16), does a little better, since it locates the maxima correctly, but the amplitude of its prediction is off. For very short wavelengths, therefore, the only alternative to the exact Poisson equation remains the linear, two-dimensional theory.

Conversely, in the second case, the amplitude of the perturbation is too large (with respect to depth) for the linear, two-dimensional theory (dotted brown line in figure 5d). Moreover, its wavelength is too short for the classical lubrication approximation (dash-dotted grey line). The revised lubrication theory, on the other hand, handles this case gracefully: the numerical solution of (2.16) matches the finite element simulation of the full two-dimensional problem everywhere within 2.3 % of the average stress (dashed orange line). Equation (2.16) thus represents accurately the diffusion of momentum across the flow, which makes it a convenient alternative to the Poisson equation in this case.

This encouraging result would not survive a significant shortening of the perturbation's wavelength, that is, a reduction of the aspect ratio of the flow. As it is, however, the periodic configuration of figure 5(b) evokes a series of channels of aspect ratio ${\rm \pi} D_0/( k {D_1}^{* } )\approx 7.5$ – a value comparable to that of laboratory rivers (Seizilles et al. Reference Seizilles, Devauchelle, Lajeunesse and Métivier2013). In the next section, inspired by these experiments, we consider a configuration in which the corrugation and the free surface intersect, thus forming the bank of a river.

4.1. Poisson equation

The Poisson equation (2.1) is linear and non-homogeneous. It is thus convenient to decompose its solutions into the sum of a homogeneous solution and a special solution. Dropping the source term of the Poisson equation, we are left with the Laplace equation,

(4.1)\begin{equation} \nabla^2 u = 0, \end{equation}

to which we can find solutions in the form of analytical functions of the complex variable $\omega = y + \textrm {i} z$ (the bank is located at $\omega =0$). Then, we will need to find a special solution to the original Poisson equation. Unless $\mu =1$, this will simply be (3.1), the solution we encountered in § 3.1.1.

4.1.1. Homogeneous solution

In a corner of angle $\beta$, there exists a power-law solution to the Laplace equation (4.1) that satisfies boundary conditions (2.2) and (2.3). Namely,

(4.2)\begin{equation} u^{(h)} = C \, Re [ \omega^{ {\rm \pi}/( 2 \beta)} ], \end{equation}

where $\omega = y + \textrm {i} z$ is the complex coordinate, and $C$ is a positive constant with physical dimensions (Polubarinova-Kochina Reference Polubarinova-Kochina1962). Differentiating the above expression, we get the vertical component of the shear stress, $\tau _z$

(4.3)\begin{equation} \tau_z^{(h)} = C' y^{ {\rm \pi}/(2\beta) - 1}, \end{equation}

where $C'$ is another positive constant. The exponent of the above expression is already an indication that a slope of $\mu =1$ might be a special value. Indeed, when $\beta = {\rm \pi}/4$, the exponent of the above expression is exactly 1 – the exponent of the special solution (3.2). Motivated by this observation, we now compare the homogeneous solution (4.3) with the special solution.

4.1.2. Shallow bank ($\mu <1$)

When the bank is shallow enough ($\mu <1$ or equivalently $\beta < {\rm \pi}/4$), the special solution (3.1) remains finite and positive, and thus physically acceptable. In a given channel, such as those of figure 6, the actual flow is then the sum of this special solution and a homogeneous solution (4.2), the prefactor of which, $C$, adjusts to the far field, that is, to the overall shape of the channel.

Provided the corner's angle, $\beta$, is less than ${\rm \pi} /4$, the exponent of the homogeneous solution (4.3) is larger than one, and therefore larger than that of the special solution. As a consequence, the special solution dominates the flow in the bank's neighbourhood ($\omega \rightarrow 0$), and the shear stress is well approximated by (3.2), in which there is no parameter that adjusts to the far field.

To compare this asymptotic behaviour with numerical simulations, we first devise a concrete example of a channel with a bank slope of $\mu$. We expect that, apart from the bank, the general shape of the channel will not alter the asymptotic behaviour of the flow in the corner. To fix the channel's shape, we arbitrarily define its cross-section as

(4.4)\begin{equation} D = D_m \frac{\cosh( \mu ( y/D_m - R/2 ) ) - \cosh( \mu R/2 )}{\sinh( \mu R/2 )}, \end{equation}

where $D_m$ sets the depth of the channel, and $R$ sets its aspect ratio (figure 6). In practice, we choose $R=8$, whereas $D_m$ does not matter once the problem is made dimensionless. We then run finite-element simulations in such a channel to numerically approximate the velocity field $u$, which is a solution to the Poisson equation (2.1) (blue shadings in figure 6). By differentiating this velocity, we find the shear stress on the bed, $\tau _z$. For a shallow bank ($\mu <1$), we find good agreement between the numerical solution (solid blue line in figure 7) and the special solution (3.2) (dashed blue line in figure 7), without adjusting any parameter. This shows that, near a shallow bank, the flow is indeed dominated by the special solution, the asymptotic behaviour of which is independent of the rest of the flow (the far field).

Figure 7.

Shear stress along the bed of a channel with sharp corner. Solid lines correspond to the finite-element simulations of figure 6. Dashed blue line is (3.2), without any fitted parameter. Dotted brown line is (4.3) with coefficient $C'$ fitted to finite-element simulation. Dash-dotted orange line: (4.6), with $C_b$ arbitrarily set to $D_m$.

However, as the bank's angle approaches $45^{\circ }$ ($\mu =1$), (3.1) breaks down, and the flow near the bank takes another form.

4.1.4. Intermediate bank ($\mu =1$)

When the slope of the bank is exactly one ($\beta = {\rm \pi}/4$), finding a special solution of the Poisson equation (2.1) requires more work. One method is to write the solution as the sum of a radially symmetric term that does not match the boundary conditions, with an analytical term that corrects this mismatch. A simple expression for the first is $-gS\omega \bar {\omega }/(4\nu )$ (the overbar denotes complex conjugation), which naturally satisfies the free-surface boundary condition, but not the no-slip condition on the channel's bed. We now need to find an analytical function that compensates for this shortcoming. There might be a principled method to do so, but we used trial and error to identify, among the usual suspects of complex analysis (power laws and logarithm), the one that suits our problem, namely:

(4.5)\begin{equation} u ={-} \frac{gS}{\nu} \mbox{Re} \left( \frac{ \omega \bar{\omega} }{ 4 } + \frac{\omega^2}{\rm \pi} \log \left( \frac{\omega}{ C_b } \right) \right), \end{equation}

where $C_b$ is an arbitrary positive constant. This constant relates the flow near the corner to the far field. Unlike for a steep bank (§ 4.1.3), however, this constant plays only a minor role in the neighbourhood of the bank, as $\log \omega$ overcomes $\log C_b$. This was to be expected, perhaps, in the special case that marks the transition between the shallow bank, for which the far field does not matter, and the steep bank, for which the far field dominates the flow.

Differentiating (4.5) along the $z$ direction, we finally find an expression for the shear stress near the bank

(4.6)\begin{equation} \tau_z \approx{-} \frac{ 2 \rho g S }{\rm \pi} y \log \left( \frac{y}{ C_b } \right). \end{equation}

The finite-element simulations match this peculiar scaling, even when the constant $C_b$ is arbitrarily set to $D_m$ (orange lines in figure 7).

To represent the transition from the shallow-bank asymptotic regime to steep-bank regime, we plot the exponent of the leading-order term in the shear-stress expansion, $\alpha$, as a function of the bank slope, $\mu$ (figure 8). After a plateau at $\alpha =1$ that corresponds to a shallow bank ($\mu <1$), the exponent suddenly switches to ${\rm \pi} /(2\beta )-1$ when the bank gets steeper than $45^{\circ }$ (blue line on figure 8). Right at the transition, when the bank's slope is exactly one, the leading-order term is not a power law and $\alpha$ is thus ill defined.

Figure 8.

Exponent of the leading-order term in the expansion of the shear stress, $\tau _z$, near a bank of angle $\beta$. Solid blue line: two-dimensional expansion of the exact problem for reference, (3.2) and (4.3). Dashed orange line: revised shallow-water theory, (3.2) and (4.13). Blue star: intermediate bank, for which the leading-order term is not a power law (4.6).

Now that we have established the asymptotic behaviour of the flow near a bank, we can ask how accurately the revised lubrication theory accounts for this behaviour. This is the purpose of the next section.

4.2. Revised lubrication approximation

We now consider the corner flow of § 4.1 in the framework of the revised lubrication theory. The findings of the previous section will serve as a reference, against which we will test the theory introduced in this paper. Near a bank of slope $\mu$, (2.16) becomes

(4.7)\begin{equation} \frac{1}{3}\frac{\mathrm{d}^2 }{\mathrm{d}y^2} ( y^2 \tau_z ) -\left( 1 + \frac{1}{\mu^2} \right) \tau_z + \frac{\rho g S}{\mu} y = 0. \end{equation}

Once again, to solve this linear, non-homogeneous equation, we need to find a special solution and a family of homogeneous solutions. In doing so, we hope to find the counterparts of the special and homogeneous solutions of § 4.1.

4.2.1. Homogeneous equation

We begin with the homogeneous equation associated with (4.7)

(4.8)\begin{equation} \frac{\mathrm{d}^2 }{\mathrm{d}y^2} ( y^2 \tau_z ) - A \tau_z = 0, \end{equation}

where, for simplicity, we have introduced $A=3 ( 1 + 1/\mu ^2 )$. The solutions of the above equation are the sum of two independent solutions

(4.9)\begin{equation} \tau_{z,h} = C_+ y^{\alpha_+} + C_- y^{\alpha_-}, \end{equation}

where $C_{\pm }$ denote integration constants, and the two exponents read

(4.10)\begin{equation} \alpha_{{\pm}} = \frac{-3 \pm \sqrt{ 1 + 4A}}{2}. \end{equation}

Among the above exponents, one is positive ($\alpha _+> 0$), and the other negative ($\alpha _-<0$). We discard the latter to ensure that the shear stress, $\tau _z$, remains finite at the bank. We then rewrite the homogeneous solution as

(4.11)\begin{equation} \tau_{z,h} = C_+ y^{\alpha_+}. \end{equation}

At this point, it is tempting to identify $\alpha _+$ with the exponent of (4.3). This comparison, however, makes sense only if (4.11) represents the leading-order term of the flow. To find when this is the case, we need to look for a special solution of (4.7).

4.2.3. Shallow bank ($\mu <1$)

When the bank is shallow enough ($\mu <1$), (4.10) tells us that the exponent of the homogeneous solution, $\alpha _+$, is larger than 1. As a result, near the bank ($y \rightarrow 0$), the homogeneous solution becomes negligible with respect to the special solution (3.2). Therefore,

(4.12)\begin{equation} \tau_z \sim \frac{\rho g \mu S}{1-\mu^2}y, \end{equation}

which is just (3.2). At leading order, the shear stress thus grows linearly with the distance from the bank, which translates into a plateau on figure 8 ($\alpha =1$). This plateau matches exactly the one we found in § 4.1.2 and, in retrospect, supports the rough reasoning of § 3.1.1. Encouraged by this agreement, we now consider the case of a steep bank.

4.2.4. Steep bank ($\mu >1$)

When the bank is steep ($\mu >1$), the shear stress associated with the special solution (3.2) becomes negative. This would be physically unacceptable, if the homogeneous solution did not overcome the special solution. According to (4.10), however, we find that the exponent of the special solution, $\alpha _+$, is now less than 1. This means that, near a steep bank, the homogeneous solution indeed dominates the flow. Therefore,

(4.13)\begin{equation} \tau_z \sim C'_+ y^{\alpha_+}, \end{equation}

where $C'_+$ is a constant set by the far field. This expression, together with (4.10), approximates well the exact expansion derived in § 4.1.3 when $\mu$ is close to one (figure 8). This agreement, however, becomes less accurate as the bank's slope steepens. Beyond the critical slope, the revised lubrication theory gradually becomes less reliable, and can only provide qualitative estimates of the exponent $\alpha$.

In the limit of a vertical corner ($\mu \to \infty$), the bed becomes a wall that extends down to infinity. The revised lubrication approximation then fails because the vertical integral of the momentum balance, which yields (2.4), diverges. Adding a bottom at some finite depth, however, provides a lower bound to this integral, and allows the approximation to recover (§ 3.2). In fact, near the walls of a rectangular channel, the revised lubrication approximation gives a decent estimate of the stress on the bottom (Appendix C.1). The quality of the approximation thus depends on the shape of the bottom, rather than the behaviour of the flow near the corner – another manifestation of non-locality.

4.2.5. Intermediate bank ($\mu =1$)

When the bank is at an angle of exactly $45^{\circ }$ (that is, $\mu =1$), the special solution (3.2), breaks down. Instead, we find the following special solution:

(4.14)\begin{equation} \tau_z ={-} \frac{3 \rho g S}{5} y \log \left( \frac{y}{C_b'} \right), \end{equation}

where $C_b'$ is a positive constant which, like $C_b$, is set by the far field. The exponent of the homogeneous solution is $\alpha _+=1$, and (4.14) is thus the leading-order term of the shear stress near the bank. This expression matches its two-dimensional counterpart (4.6), but for the value of the prefactor, which we find to be less than 6 % off.

Just like in § 4.1.4, the flow near a bank of slope 1 features a logarithmic correction, with an integration constant that adjusts to the far field. That, in the revised lubrication theory, the prefactor of this regime is a bit off is typical of a shallow-water approximation: it is qualitatively correct, but not exact. The error, in this case, vanishes suddenly when the bank's angle becomes less than $45^{\circ }$.