2.1 Deep flow kinematics

To describe the steady deep flow $(U_{0},W_{0})(X,Z)$ , we adopt the kinematic model of Nedderman & Tüzün (Reference Nedderman and Tüzün1979). In this model, vertical shear is assumed to cause grains to drift laterally at velocity

(2.1) $$\begin{eqnarray}U_{0}=-K\frac{\unicode[STIX]{x2202}W_{0}}{\unicode[STIX]{x2202}X},\end{eqnarray}$$

where $U_{0}(X,Z)$ , $W_{0}(X,Z)$ are respectively the horizontal and vertical components of the mean granular velocity in the deep flow region, and $K$ is a coefficient with dimensions of length called the kinematic constant by Nedderman & Tüzün (Reference Nedderman and Tüzün1979). Substitution of (2.1) into the continuity equation

(2.2) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}U_{0}}{\unicode[STIX]{x2202}X}+\frac{\unicode[STIX]{x2202}W_{0}}{\unicode[STIX]{x2202}Z}=0,\end{eqnarray}$$

then yields the diffusion equation with diffusivity $K$

(2.3) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}W_{0}}{\unicode[STIX]{x2202}Z}-K\frac{\unicode[STIX]{x2202}^{2}W_{0}}{\unicode[STIX]{x2202}X^{2}}=0,\end{eqnarray}$$

where diffusion proceeds upwards along the time-like $Z$ axis to even out horizontal variations in $W_{0}$ . The micro-mechanical basis of (2.1) and (2.3) remains to be clarified, but a simple physical explanation is as follows (Litwiniszyn Reference Litwiniszyn1963; Caram & Hong Reference Caram and Hong1991). At the discrete level, the downward motion of individual grains is associated with the upward migration of void spaces, or holes, into which individual grains fall. At each fall, the hole can move either right or left with equal probability. The resulting upward random walk of individual void spaces yields a diffusion process in the continuum limit. The opposite granular flux $W_{0}$ is then governed by the diffusion equation (2.3). Although the equation lacks a rigorous mechanical basis, it describes a physically plausible deep flow velocity field that can be used to drive surface avalanching. Upon calibrating the diffusivity $K$ , equation (2.3) has been found to yield good agreement with silo experiments (Tüzün et al. Reference Tüzün, Houlsby, Nedderman and Savage1982; Choi, Kudrolli & Bazant Reference Choi, Kudrolli and Bazant2005) and discrete element simulations (Rycroft et al. Reference Rycroft, Grest, Landry and Bazant2006; Balevičius et al. Reference Balevičius, Kačianauskas, Mróz and Sielamowicz2011).

For the silos we consider, the diffusion equation must be solved subject to the lateral boundary conditions

(2.4) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}W_{0}}{\unicode[STIX]{x2202}X}=0,\quad \text{at }X=-L\text{ and }X=0,\end{eqnarray}$$

imposing no flux across the left wall and symmetry plane (for the symmetric silo), or across the left and right walls (for the asymmetric silo). At the outlet level $Z=0$ , on the other hand, a bottom boundary condition can be written

(2.5) $$\begin{eqnarray}W_{0}=-2Q\unicode[STIX]{x1D6FF}(X),\end{eqnarray}$$

where $Q$ is either half the rate of outflow (for the symmetric case), or the full outflow (for the asymmetric case), divided by the gap width $B$ (see figure 1). Disregarding the details of the flow near the orifice, a delta function is used to represent the outflow as a point sink. Since the diffusion equation features a first derivative in $Z$ and a second derivative in $X$ (see e.g. Kevorkian Reference Kevorkian1990), no condition is required or can be applied along the granular free surface, nor can additional conditions restraining slip be applied along the side walls. The deep flow is driven by drainage from the orifice, and is not affected by avalanching at the free surface. In addition to the velocity field, we will also be interested in the streamfunction $\unicode[STIX]{x1D6F9}_{0}(X,Z)$ defined by

(2.6a,b ) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}\unicode[STIX]{x1D6F9}_{0}}{\unicode[STIX]{x2202}Z}=U_{0},\quad \frac{\unicode[STIX]{x2202}\unicode[STIX]{x1D6F9}_{0}}{\unicode[STIX]{x2202}X}=-W_{0},\end{eqnarray}$$

the contours of which give the granular streamlines. For definiteness, we associate $\unicode[STIX]{x1D6F9}_{0}=0$ with the vertical streamline through the origin $X=0$ .

The above mathematical problem can be made dimensionless by defining variables $\hat{X}=X/L$ , $\hat{Z}=ZK/L^{2}$ , ${\hat{W}}_{0}=W_{0}L/Q$ , $\hat{U} _{0}=U_{0}L^{2}/(KQ)$ and $\hat{\unicode[STIX]{x1D6F9}}_{0}=\unicode[STIX]{x1D6F9}_{0}/Q$ . As detailed in Kevorkian (Reference Kevorkian1990), the solution can be obtained using either Fourier series, advisable for $\hat{Z}$ moderate to large, or mirrored Green functions, advisable for $\hat{Z}$ small to moderate. Since the latter case better fits the conditions of our experiments, we will calculate the solution using the Green function

(2.7) $$\begin{eqnarray}G(\hat{X},\hat{Z})=\frac{1}{(\unicode[STIX]{x03C0}\hat{Z})^{1/2}}\exp \left(-\frac{\hat{X}^{2}}{4\hat{Z}}\right),\end{eqnarray}$$

which solves (2.3) subject to bottom boundary condition (2.5) on a laterally unbounded domain. To enforce lateral boundary conditions (2.4), mirrored Green functions must be added to obtain

(2.8) $$\begin{eqnarray}{\hat{W}}_{0}(\hat{X},\hat{Z})=-G(\hat{X},\hat{Z})-\mathop{\sum }_{k=1}^{n}(G(\hat{X}-2k,\hat{Z})+G(\hat{X}+2k,\hat{Z})),\end{eqnarray}$$

which is an infinite sum truncated at $n$ terms that converges to the solution for $n$ large (Kevorkian Reference Kevorkian1990). The solution for the streamfunction, likewise, can be obtained by integration of (2.8) along $X$ as

(2.9) $$\begin{eqnarray}\hat{\unicode[STIX]{x1D6F9}}_{0}(\hat{X},\hat{Z})=\unicode[STIX]{x1D6F7}(\hat{X},\hat{Z})+\mathop{\sum }_{k=1}^{n}(\unicode[STIX]{x1D6F7}(\hat{X}-2k,\hat{Z})+\unicode[STIX]{x1D6F7}(\hat{X}+2k,\hat{Z})),\end{eqnarray}$$

where

(2.10) $$\begin{eqnarray}\unicode[STIX]{x1D6F7}(\hat{X},\hat{Z})=\text{erf}\left(\frac{\hat{X}}{2\hat{Z}^{1/2}}\right).\end{eqnarray}$$

The resulting dimensionless solutions are illustrated in figure 2. Figure 2(a) shows spanwise profiles of horizontal velocity at different heights above the orifice. Figure 2(b) shows the corresponding profiles of vertical velocity. Finally figure 2(c) shows the streamlines $\hat{\unicode[STIX]{x1D6F9}}=$ constant for equally spaced values of the streamfunction. Near the orifice, flow is apparent only in the central region, but there is no sharp boundary between flowing and stationary regions. The diffusion solutions feature small non-zero motions even away from the central zone. Using the kinematic model, similar deep flow fields in silos have earlier been calculated by Nedderman & Tüzün (Reference Nedderman and Tüzün1979), using Fourier series, and by Choi et al. (Reference Choi, Kudrolli and Bazant2005) using numerical computations.

Figure 2. Solutions to the kinematic deep flow model: (a) spanwise profiles of horizontal velocity; (b) spanwise profiles of vertical velocity; (c) streamlines.

2.2 Shallow layer governing equations

To model surface avalanching (figure 3), we consider a distinct coordinate system $(x,z)$ tilted at the angle of repose $\unicode[STIX]{x1D6FC}$ , with origin at $(X,Z)=(-L/2,H(t))$ and descending with the granular free surface at speed ${\dot{H}}=-Q/L$ . The $x$ -axis of the coordinate system therefore coincides with the inclined profile (figure 3 a)

(2.11) $$\begin{eqnarray}\text{}\underline{Z}(X,t)=H(t)-(X+L/2)\tan \unicode[STIX]{x1D6FC}.\end{eqnarray}$$

Throughout this section, we assume the restricted domain $-L\leqslant X\leqslant 0$ , with mirror symmetry implied in case of a symmetric silo. In the inclined coordinate system, we define $\widetilde{z}(x,t)$ and ${\displaystyle\mathop{z}\limits_{{\sim}}}(x,t)$ to be the free surface and basal boundary of the avalanching layer. The depth of the avalanching layer is then given by $h(x,t)=\widetilde{z}-{\displaystyle\mathop{z}\limits_{{\sim}}}$ . We assume shallow flow, $h\ll L$ , and small relief $\widetilde{z}\ll L$ relative to profile $z=0$ inclined at the angle of repose. Consistent with these approximations, flow acceleration in the $z$ -direction will be neglected relative to acceleration in the downslope $x$ -direction. However both velocity components, $u(x,z,t)$ , in the $x$ -direction and $w(x,z,t)$ in the $z$ -direction, will intervene in the flow kinematics. Inspired from Liggett (Reference Liggett1994), and following Capart et al. (Reference Capart, Hung and Stark2015) and Hung et al. (Reference Hung, Stark and Capart2016), we use different lines above and below variables to denote values sampled at different locations: ${\displaystyle\mathop{\,\cdot \,}\limits_{{\sim}}}$ denotes a variable sampled along the basal boundary, $\tilde{\,\cdot \,}$ a variable sampled along the free surface and $\text{}\underline{\,\cdot \,}$ a variable sampled along the angle of repose profile (2.11). The overbar $\overline{\,\cdot \,}$ is reserved for quantities averaged over the depth $h$ .

Figure 3. Definition sketch of surface flow model: (a) overview; (b) local cutaway. The surface flow is depicted in a frame of reference descending with the free surface, in which the flow is approximately steady. In this frame of reference, the erosion flux across the basal interface exhibits a change of sign, from positive upstream to negative downstream, hence the apparent kink in (a).

Following Capart et al. (Reference Capart, Hung and Stark2015), we start from local equations and integrate them from ${\displaystyle\mathop{z}\limits_{{\sim}}}$ to $\widetilde{z}$ to obtain depth-integrated equations governing the avalanching layer. As for the deep flow, we assume incompressibility of the granular medium, hence the local continuity equation

(2.12) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}u}{\unicode[STIX]{x2202}x}+\frac{\unicode[STIX]{x2202}w}{\unicode[STIX]{x2202}z}=0.\end{eqnarray}$$

Variations of $u$ and $w$ in the transverse $y$ -direction are neglected, hence $u$ and $w$ coincide with the corresponding velocity components averaged over the gap thickness. Local balance of momentum in the $x$ -direction, on the other hand, yields the width-averaged equation of motion (Jop et al. Reference Jop, Forterre and Pouliquen2007)

(2.13) $$\begin{eqnarray}\unicode[STIX]{x1D70C}\left(\frac{\unicode[STIX]{x2202}u}{\unicode[STIX]{x2202}t}+u\frac{\unicode[STIX]{x2202}u}{\unicode[STIX]{x2202}x}+w\frac{\unicode[STIX]{x2202}u}{\unicode[STIX]{x2202}z}\right)=\unicode[STIX]{x1D70C}g_{_{\Vert }}-\frac{\unicode[STIX]{x2202}p}{\unicode[STIX]{x2202}x}+\frac{\unicode[STIX]{x2202}\unicode[STIX]{x1D70F}}{\unicode[STIX]{x2202}z}-\frac{2\unicode[STIX]{x1D70F}_{W}}{B},\end{eqnarray}$$

where $\unicode[STIX]{x1D70C}=c_{S}\unicode[STIX]{x1D70C}_{S}$ is the mean density of the bulk phase, and $p=\unicode[STIX]{x1D70C}g_{_{\bot }}(\,\widetilde{z}-z)$ is the granular pressure, assumed lithostatic. Following Jop et al. (Reference Jop, Forterre and Pouliquen2005), the normal stresses are assumed to reduce to an isotropic pressure. For shallow flow (Liggett Reference Liggett1994), granular accelerations in the $z$ direction normal to the slope can be neglected, hence a lithostatic pressure balancing the weight of the grains above. In the same equation, $g_{_{\Vert }}=g\sin \unicode[STIX]{x1D6FC}$ and $g_{\bot }=g\cos \unicode[STIX]{x1D6FC}$ are the parallel and perpendicular components of the gravitational acceleration, $\unicode[STIX]{x1D70F}_{W}=\unicode[STIX]{x1D707}_{W}p$ is the shear stress along the side walls (with $\unicode[STIX]{x1D707}_{W}$ the wall friction coefficient) and $\unicode[STIX]{x1D70F}$ is the internal shear stress. For the latter, we assume the viscoplastic, dense granular flow rheology (da Cruz et al. Reference da Cruz, Emam, Prochnow, Roux and Chevoir2005)

(2.14) $$\begin{eqnarray}\unicode[STIX]{x1D70F}=\unicode[STIX]{x1D707}_{0}p+\unicode[STIX]{x1D712}D(\unicode[STIX]{x1D70C}p)^{1/2}\dot{\unicode[STIX]{x1D6FE}},\end{eqnarray}$$

expressed as the sum of two components. The first, rate independent, is a plastic yield stress proportional to $p$ , with coefficient $\unicode[STIX]{x1D707}_{0}=\tan \unicode[STIX]{x1D6FC}$ . The second, rate dependent, is a viscous stress that varies linearly with the shear rate $\dot{\unicode[STIX]{x1D6FE}}=\unicode[STIX]{x2202}u/\unicode[STIX]{x2202}z$ , with effective viscosity $\unicode[STIX]{x1D712}D(\unicode[STIX]{x1D70C}p)^{1/2}$ dependent on grain diameter $D$ , granular pressure $p$ and dimensionless rheological coefficient $\unicode[STIX]{x1D712}$ . This is a linearized version of the $\unicode[STIX]{x1D707}(I)$ rheology proposed by Jop et al. (Reference Jop, Forterre and Pouliquen2005) and applied earlier to liquid–granular flows by Berzi & Jenkins (Reference Berzi and Jenkins2008). The nonlinear $\unicode[STIX]{x1D707}(I)$ rheology has earlier been applied to shallow granular flows over rigid beds by Gray & Edwards (Reference Gray and Edwards2014). The two rheological coefficient $\unicode[STIX]{x1D707}_{0}$ and $\unicode[STIX]{x1D712}$ will be determined in the next section by linearizing the nonlinear $\unicode[STIX]{x1D707}(I)$ rheology calibrated earlier by Jop et al. (Reference Jop, Forterre and Pouliquen2005) for the same granular material. Note that the constitutive relations for wall friction and internal shear stress are written assuming $u>u_{W}$ , and $\dot{\unicode[STIX]{x1D6FE}}>0$ . For steady uniform flow over a loose deposit of velocity ( ${\displaystyle\mathop{u}\limits_{{\sim}}}$ , ${\displaystyle\mathop{w}\limits_{{\sim}}}$ ), these equations are satisfied by the velocity profile (figure 3 b)

(2.15) $$\begin{eqnarray}u(\widehat{\unicode[STIX]{x1D702}})={\displaystyle\mathop{u}\limits_{{\sim}}}+(\overline{u}-{\displaystyle\mathop{u}\limits_{{\sim}}})({\textstyle \frac{7}{3}}-{\textstyle \frac{35}{6}}\widehat{\unicode[STIX]{x1D702}}^{3/2}+{\textstyle \frac{7}{2}}\widehat{\unicode[STIX]{x1D702}}^{5/2}),\end{eqnarray}$$

where $\widehat{\unicode[STIX]{x1D702}}=(\,\widetilde{z}-z)/h$ and $\overline{u}$ is the depth-averaged velocity (for a derivation, see Berzi & Jenkins (Reference Berzi and Jenkins2008)). As discussed in Capart et al. (Reference Capart, Hung and Stark2015), experiments exhibit some deviations from the profile (2.15) even for equilibrium flows. For unsteady, non-uniform flows, moreover, the velocity profile is expected to deviate from its equilibrium shape. As long as such deviations are small, however, integrals over depth based on the profile (2.15) should furnish a good approximation, and provide a simple way to close depth-integrated equations. Because the corresponding shear rate at the base ${\displaystyle\mathop{\{}\limits_{{\sim}}}=\unicode[STIX]{x2202}u/\unicode[STIX]{x2202}z({\displaystyle\mathop{z}\limits_{{\sim}}})$ is equal to zero, note that this profile is appropriate only for flows over erodible beds, when the basal interface can be assumed to coincide with the yield locus. For flows over rigid beds, say over rock, a finite shear rate at the base would be expected, and the base elevation prescribed, instead of the present case with zero basal shear rate and a basal elevation set by the flow itself. Likewise, the description would not apply for granular flows over deposits composed of a different type of grains. When assumptions are met, the velocity profile $u(x,z,t)$ at a given coordinate $x$ and time $t$ can be calculated using (2.15) from $\widetilde{z}(x,t)$ , $h(x,t)$ , $\overline{u}(x,t)$ and ${\displaystyle\mathop{u}\limits_{{\sim}}}(x,t)$ . The corresponding streamfunction, likewise, can be calculated from

(2.16) $$\begin{eqnarray}\unicode[STIX]{x1D713}(\hat{\unicode[STIX]{x1D702}})=\widetilde{\unicode[STIX]{x1D713}}-h{\displaystyle\mathop{u}\limits_{{\sim}}}\hat{\unicode[STIX]{x1D702}}-h(\overline{u}-{\displaystyle\mathop{u}\limits_{{\sim}}})({\textstyle \frac{7}{3}}\hat{\unicode[STIX]{x1D702}}-{\textstyle \frac{7}{3}}\hat{\unicode[STIX]{x1D702}}^{5/2}+\hat{\unicode[STIX]{x1D702}}^{7/2}),\end{eqnarray}$$

where $\unicode[STIX]{x1D713}$ satisfies $\unicode[STIX]{x2202}\unicode[STIX]{x1D713}/\unicode[STIX]{x2202}z=u$ , $\unicode[STIX]{x2202}\unicode[STIX]{x1D713}/\unicode[STIX]{x2202}x=-w$ , and such that

(2.17) $$\begin{eqnarray}\widetilde{\unicode[STIX]{x1D713}}-{\displaystyle\mathop{\unicode[STIX]{x1D713}}\limits_{{\sim}}}=h\overline{u}.\end{eqnarray}$$

To obtain unsteady, non-uniform governing equations for the avalanching layer, we first integrate (2.12) and apply the Leibniz integral rule to get

(2.18) $$\begin{eqnarray}\int _{{\displaystyle\mathop{z}\limits_{{\sim}}}}^{\widetilde{z}}\left(\frac{\unicode[STIX]{x2202}u}{\unicode[STIX]{x2202}x}+\frac{\unicode[STIX]{x2202}w}{\unicode[STIX]{x2202}z}\right)\text{d}z=\frac{\unicode[STIX]{x2202}(h\overline{u})}{\unicode[STIX]{x2202}x}-\widetilde{u}\frac{\unicode[STIX]{x2202}\widetilde{z}}{\unicode[STIX]{x2202}x}+{\displaystyle\mathop{u}\limits_{{\sim}}}\frac{\unicode[STIX]{x2202}{\displaystyle\mathop{z}\limits_{{\sim}}}}{\unicode[STIX]{x2202}x}+\widetilde{w}-{\displaystyle\mathop{w}\limits_{{\sim}}}=0,\end{eqnarray}$$

where $q=h\overline{u}=\int _{{\displaystyle\mathop{z}\limits_{{\sim}}}}^{\widetilde{z}}u\,\text{d}z$ . This can be further worked out using the kinematic boundary condition along the free surface,

(2.19) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}\widetilde{z}}{\unicode[STIX]{x2202}t}+\widetilde{u}\frac{\unicode[STIX]{x2202}\widetilde{z}}{\unicode[STIX]{x2202}x}=\widetilde{w},\end{eqnarray}$$

and the definition $h=\widetilde{z}-{\displaystyle\mathop{z}\limits_{{\sim}}}$ to obtain the depth-integrated continuity equation

(2.20) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}h}{\unicode[STIX]{x2202}t}+\frac{\unicode[STIX]{x2202}(h\overline{u})}{\unicode[STIX]{x2202}x}=e={\displaystyle\mathop{w}\limits_{{\sim}}}-{\displaystyle\mathop{u}\limits_{{\sim}}}\frac{\unicode[STIX]{x2202}{\displaystyle\mathop{z}\limits_{{\sim}}}}{\unicode[STIX]{x2202}x}-\frac{\unicode[STIX]{x2202}{\displaystyle\mathop{z}\limits_{{\sim}}}}{\unicode[STIX]{x2202}t},\end{eqnarray}$$

where $e(x,t)$ is the erosion or entrainment rate, or rate of granular volume transfer across the basal interface ${\displaystyle\mathop{z}\limits_{{\sim}}}(x,t)$ and into the avalanching layer. This rate is defined as the deviation from volume conservation of the layer, expressed by the left-hand side. Likewise, the local equation of motion (2.13) can be integrated over depth to obtain the depth-integrated momentum equation

(2.21) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}t}(h\overline{u})+\frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}x}(h\,\overline{u^{2}})=e{\displaystyle\mathop{u}\limits_{{\sim}}}+g_{_{\Vert }}h-g_{_{\bot }}h\frac{\unicode[STIX]{x2202}\widetilde{z}}{\unicode[STIX]{x2202}x}-\frac{1}{\unicode[STIX]{x1D70C}}{\displaystyle\mathop{\unicode[STIX]{x1D70F}}\limits_{{\sim}}}-\frac{2}{\unicode[STIX]{x1D70C}B}h\overline{\unicode[STIX]{x1D70F}}_{W}.\end{eqnarray}$$

Assuming ${\displaystyle\mathop{u}\limits_{{\sim}}}(x,t)$ , ${\displaystyle\mathop{w}\limits_{{\sim}}}(x,t)$ to be otherwise given, we must solve for three evolving profiles $h(x,t)$ , $\widetilde{z}(x,t)$ and $\overline{u}(x,t)$ . We therefore supplement (2.20) and (2.21) with an additional equation expressing the balance of depth-integrated kinetic energy (Capart et al. Reference Capart, Hung and Stark2015; Hung et al. Reference Hung, Stark and Capart2016). This is obtained by integrating over depth the product of (2.13) with velocity $u(x,z,t)$ , yielding

(2.22) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}t}\left(\frac{h\,\overline{u^{2}}}{2}\right)+\frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}x}\left(\frac{h\,\overline{u^{3}}}{2}\right)=\frac{e{\displaystyle\mathop{u}\limits_{{\sim}}}^{2}}{2}+g_{_{\Vert }}h\overline{u}-g_{_{\bot }}h\overline{u}\frac{\unicode[STIX]{x2202}\widetilde{z}}{\unicode[STIX]{x2202}x}-{\displaystyle\mathop{\unicode[STIX]{x1D70F}}\limits_{{\sim}}}{\displaystyle\mathop{u}\limits_{{\sim}}}-\frac{1}{\unicode[STIX]{x1D70C}}h\,\overline{\unicode[STIX]{x1D70F}\dot{\unicode[STIX]{x1D6FE}}}-\frac{2}{\unicode[STIX]{x1D70C}B}h\,\overline{\unicode[STIX]{x1D70F}_{W}u}.\end{eqnarray}$$

In this equation, the next-to-last and last terms represent dissipation by granular contacts inside the avalanching layer and dissipation by friction along the walls, respectively. Equations (2.20)–(2.22) represent a generalization of the equations derived in Capart et al. (Reference Capart, Hung and Stark2015) for the case of stationary deposits ( ${\displaystyle\mathop{u}\limits_{{\sim}}}={\displaystyle\mathop{w}\limits_{{\sim}}}=0$ ). We now specialize them to the problem of avalanching flows over loose deposits undergoing uneven subsidence due to silo discharge. Ideally, we would wish to match the velocity at the base of the avalanching layer with the deep flow velocity along the same curve, i.e. let $({\displaystyle\mathop{u}\limits_{{\sim}}},{\displaystyle\mathop{w}\limits_{{\sim}}})=({\displaystyle\mathop{u}\limits_{{\sim}}}_{0},{\displaystyle\mathop{w}\limits_{{\sim}}}_{0})$ . The precise shape of the basal interface ${\displaystyle\mathop{z}\limits_{{\sim}}}(x,z,t)$ , however, is not known in advance. As in linearized groundwater seepage problems (see e.g. Ni & Capart Reference Ni and Capart2006), therefore, we choose instead to apply kinematic matching to the approximate basal interface given by $\text{}\underline{Z}(X,t)$ . This corresponds to the linear profile (2.11), inclined at the angle of repose. We therefore approximate

(2.23) $$\begin{eqnarray}\displaystyle & \displaystyle {\displaystyle\mathop{u}\limits_{{\sim}}}\approx \text{}\underline{u}_{0}=\text{}\underline{U\!}_{\,0}\cos \unicode[STIX]{x1D6FC}-(\text{}\underline{W\!}_{\,0}-{\dot{H}})\sin \unicode[STIX]{x1D6FC}, & \displaystyle\end{eqnarray}$$

(2.24) $$\begin{eqnarray}\displaystyle & \displaystyle {\displaystyle\mathop{w}\limits_{{\sim}}}\approx \text{}\underline{w}_{0}=\text{}\underline{U\!}_{\,0}\sin \unicode[STIX]{x1D6FC}+(\text{}\underline{W\!}_{\,0}-{\dot{H}})\cos \unicode[STIX]{x1D6FC}, & \displaystyle\end{eqnarray}$$

where velocities in the upright, stationary silo frame of reference are converted to the tilted frame of reference descending at constant speed ${\dot{H}}$ . Likewise we approximate

(2.25) $$\begin{eqnarray}{\displaystyle\mathop{w}\limits_{{\sim}}}+{\displaystyle\mathop{u}\limits_{{\sim}}}\frac{\unicode[STIX]{x2202}{\displaystyle\mathop{z}\limits_{{\sim}}}}{\unicode[STIX]{x2202}x}\approx \text{}\underline{w}_{0}+\text{}\underline{u}_{0}\frac{\unicode[STIX]{x2202}\text{}\underline{z}}{\unicode[STIX]{x2202}x}=\text{}\underline{w}_{0}.\end{eqnarray}$$

In terms of the streamfunction, equivalently, we let

(2.26) $$\begin{eqnarray}{\displaystyle\mathop{\unicode[STIX]{x1D713}}\limits_{{\sim}}}\approx \text{}\underline{\unicode[STIX]{x1D713}\!}_{\,0}=\text{}\underline{\unicode[STIX]{x1D6F9}\!}_{\,0}+{\dot{H}}\text{}\underline{X},\end{eqnarray}$$

where $\text{}\underline{\unicode[STIX]{x1D6F9}\!}_{\,0}$ is the streamfunction (2.9) sampled along $(\text{}\underline{X},\text{}\underline{Z})(x,t)=(-L/2+x\cos \unicode[STIX]{x1D6FC},H(t)-x\sin \unicode[STIX]{x1D6FC})$ . The solutions presented below are obtained by applying the matching condition at this known, but approximate interface position $\text{}\underline{Z}(X,t)$ . We could instead proceed by iterations, applying the matching condition again at each step to the interface location calculated from the previous step. Because the deep flow varies gradually with height above orifice, however, the error of the approximate solution with respect to the converged iterative solution is small, hence we prefer the simpler approach. With this approximation, the depth-integrated continuity equation becomes

(2.27) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}\widetilde{z}}{\unicode[STIX]{x2202}t}+\frac{\unicode[STIX]{x2202}(h\overline{u})}{\unicode[STIX]{x2202}x}=\text{}\underline{w}_{0}.\end{eqnarray}$$

For quasi-steady flow, the dominant balance is between the second and third terms, hence $\text{}\underline{w}_{0}\sim h\overline{u}/L$ . For shallow flow the ratio $h/L=\unicode[STIX]{x1D716}$ is small, thus the deep flow velocity components $\text{}\underline{u_{0}}\sim \text{}\underline{w_{0}}\sim \unicode[STIX]{x1D716}\overline{u}$ are one order of magnitude smaller than the depth-averaged avalanche velocity $\overline{u}$ . Terms associated with the basal velocity ${\displaystyle\mathop{u}\limits_{{\sim}}}\approx \text{}\underline{u}_{0}$ can therefore be neglected in the momentum and kinetic energy equations (2.21) and (2.22). Likewise we can set ${\displaystyle\mathop{u}\limits_{{\sim}}}=0$ in (2.15). With these simplifications, equations (2.21)–(2.22) governing the dynamics of the avalanching layer become

(2.28) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}t}(h\overline{u})+\frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}x}\left(\frac{77}{48}h\overline{u}^{2}\right)=-g_{_{\bot }}h\frac{\unicode[STIX]{x2202}\widetilde{z}}{\unicode[STIX]{x2202}x}-\frac{\unicode[STIX]{x1D707}_{W}}{B}g_{_{\bot }}h^{2}, & \displaystyle\end{eqnarray}$$

(2.29) $$\begin{eqnarray}\displaystyle & \displaystyle \frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}t}\left(\frac{77}{96}h\overline{u}^{2}\right)+\frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}x}(\unicode[STIX]{x1D705}h\overline{u}^{3})=-g_{_{\bot }}h\overline{u}\frac{\unicode[STIX]{x2202}\widetilde{z}}{\unicode[STIX]{x2202}x}-\frac{35}{9}\unicode[STIX]{x1D712}Dg_{\bot }^{1/2}\frac{\overline{u}^{2}}{h^{1/2}}-\frac{5}{9}\frac{\unicode[STIX]{x1D707}_{W}}{B}g_{_{\bot }}h^{2}\overline{u}, & \displaystyle\end{eqnarray}$$

where $\unicode[STIX]{x1D705}=342853/233376\approx 1.469$ , and where the depth-integrated terms $\overline{u^{2}}$ and $\overline{u^{3}}$ were integrated using profile (2.15) with ${\displaystyle\mathop{u}\limits_{{\sim}}}=0$ . The last two terms of (2.29), representing energy dissipation, are obtained by integration of the corresponding terms $\overline{\unicode[STIX]{x1D70F}\dot{\unicode[STIX]{x1D6FE}}}$ and $\overline{\unicode[STIX]{x1D70F}_{W}u}$ in (2.22), using profile (2.15) and the rheology (2.14). The next-to-last term of (2.29) is thus controlled by the rheology coefficient $\unicode[STIX]{x1D712}$ that determines the effective viscosity. After simplification, equations (2.28) and (2.29) are identical to the momentum and kinetic energy equations derived earlier in Capart et al. (Reference Capart, Hung and Stark2015). Together with (2.27) they form a closed set of governing equations for profiles $\widetilde{z}(x,t)$ , $h(x,t)$ and $\overline{u}(x,t)$ , driven by the deep flow velocities $\text{}\underline{U\!}_{\,0}(x,t)$ and $\text{}\underline{W\!}_{\,0}(x,t)$ . These in turn can be estimated using the kinematic model described in the previous section. While these unsteady equations could in principle be solved numerically to describe the transient response, here we consider only the quasi-steady response obtained by neglecting the time derivatives in (2.27)–(2.29). The semi-analytical approach used to construct solutions for this case is described in the next sub-section.

2.3 Quasi-steady avalanching solutions

For quasi-steady flow, the continuity equation (2.27) reduces to

(2.30) $$\begin{eqnarray}\frac{\unicode[STIX]{x2202}q}{\unicode[STIX]{x2202}x}=\text{}\underline{w}_{0}=-\frac{\unicode[STIX]{x2202}\text{}\underline{\unicode[STIX]{x1D713}\!}_{\,0}}{\unicode[STIX]{x2202}x},\end{eqnarray}$$

where $q=h\overline{u}$ is the depth-integrated granular discharge of the avalanching layer. Integration with boundary conditions $q=0$ , $\text{}\underline{\unicode[STIX]{x1D713}\!}_{\,0}=0$ at $x=L/(2\cos \unicode[STIX]{x1D6FC})$ then yields

(2.31) $$\begin{eqnarray}q(x,t)=-\text{}\underline{\unicode[STIX]{x1D713}\!}_{\,0}(x,t),\end{eqnarray}$$

such that $q=0$ is also automatically satisfied at $x=-L/(2\cos \unicode[STIX]{x1D6FC})$ . As expected, it follows from (2.17) and (2.31) that $\widetilde{\unicode[STIX]{x1D713}}=0$ , i.e. the quasi-steady free surface coincides with the zero streamline. For quasi-steady flow, the downslope distribution of the avalanching discharge $q(x,t)$ is therefore fully known, including its maximum value $q_{max}(t)$ at any given time $t$ during the descent. An example is illustrated in figure 4(a). Hereafter, we drop the explicit time dependence, but note that the evolving height $H(t)$ of the avalanching layer during its descent will affect its discharge profile $q(x)$ .

Figure 4. Dynamic model solution for the avalanching layer: (a) imposed spanwise profile of normal to slope velocity ${\hat{w}}_{0}=\text{d}\hat{q}/\text{d}\hat{x}$ (dashed line) and the resulting avalanching discharge $\hat{q}(\hat{x})$ (continuous line); (b) dimensionless phase diagram $(\hat{x},{\hat{h}})$ with the solution curve ${\hat{h}}(\hat{x})$ in black; (c) basal interface (dashed line) and streamlines (continuous lines) in the frame of reference descending with the free surface.

To solve the momentum and kinetic energy equations (2.28) and (2.29) subject to the known discharge profile $q(x)$ , we substitute $\overline{u}=q/h$ and introduce the dimensionless variables $\hat{x}=x/(L/\cos \unicode[STIX]{x1D6FC})$ , $\hat{q}=q/Q$ , ${\hat{h}}=h/h_{a}$ , ${\hat{S}}=(-\unicode[STIX]{x2202}\widetilde{z}/\unicode[STIX]{x2202}x)/S_{a}$ , where

(2.32a,b ) $$\begin{eqnarray}h_{a}=\left(\frac{\unicode[STIX]{x1D712}DQ}{g_{\bot }^{1/2}\unicode[STIX]{x1D707}_{W}/B}\right)^{2/7}\quad \text{and}\quad S_{a}=\frac{\unicode[STIX]{x1D707}_{W}}{B}h_{a}\end{eqnarray}$$

are the characteristic depth and excess surface inclination associated with equilibrium avalanching flow (Hung et al. Reference Hung, Stark and Capart2016). At equilibrium, the flow is paced (via the rheological coefficient $\unicode[STIX]{x1D712}$ ) by the rate-dependent term of the dense granular flow rheology (2.14). Translated to dimensionless form, (2.28) and (2.29) become

(2.33) $$\begin{eqnarray}\displaystyle & \displaystyle \hat{Q}^{8/7}\frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}\hat{x}}\left(\frac{77}{48}\frac{\hat{q}^{2}}{{\hat{h}}}\right)={\hat{h}}{\hat{S}}-{\hat{h}}^{2}, & \displaystyle\end{eqnarray}$$

(2.34) $$\begin{eqnarray}\displaystyle & \displaystyle \hat{Q}^{8/7}\frac{\unicode[STIX]{x2202}}{\unicode[STIX]{x2202}\hat{x}}\left(\frac{\unicode[STIX]{x1D705}\hat{q}^{3}}{{\hat{h}}^{2}}\right)=\hat{q}{\hat{S}}-\frac{35}{9}\frac{\hat{q}^{2}}{{\hat{h}}^{5/2}}-\frac{5}{9}{\hat{h}}\hat{q}, & \displaystyle\end{eqnarray}$$

where $\hat{Q}$ is the dimensionless silo discharge

(2.35) $$\begin{eqnarray}\hat{Q}=\frac{Q}{(\unicode[STIX]{x1D707}_{W}/B)^{1/8}g_{\bot }^{1/2}(\unicode[STIX]{x1D712}D)^{3/4}(L/\cos \unicode[STIX]{x1D6FC})^{7/8}}.\end{eqnarray}$$

An equivalent dimensionless number, found to control avalanching behaviour in rotating drums, was given the name entrainment number by Hung et al. (Reference Hung, Stark and Capart2016). It can be seen from (2.33) and (2.34) that this number controls the strength of the left-hand side flux divergence terms associated with convective inertia. For low $\hat{Q}$ , the avalanching layer is locally in equilibrium and its behaviour is dominated by the dense granular flow rheology. In the inertialess limit $\hat{Q}\rightarrow 0$ , in particular, the flux divergence terms disappear, and the solution is simply

(2.36a,b ) $$\begin{eqnarray}{\hat{h}}(\hat{x})=({\textstyle \frac{35}{4}}\hat{q}(\hat{x}))^{2/7},\quad {\hat{S}}(\hat{x})={\hat{h}}(\hat{x}).\end{eqnarray}$$

For high $\hat{Q}$ , in contrast, inertia effects dominate and take the flow far from local equilibrium (Hung et al. Reference Hung, Stark and Capart2016). For intermediate $\hat{Q}$ , both inertia and rheology intervene. The equations in that case can be solved by eliminating ${\hat{S}}$ between (2.33) and (2.34), yielding for ${\hat{h}}$ the ordinary differential equation of the first order

(2.37) $$\begin{eqnarray}\frac{\text{d}{\hat{h}}}{\text{d}\hat{x}}=\frac{{\textstyle \frac{35}{9}}{\hat{h}}^{1/2}\hat{q}-{\textstyle \frac{4}{9}}{\hat{h}}^{4}+\hat{Q}^{8/7}(3\unicode[STIX]{x1D705}-{\textstyle \frac{77}{24}}){\hat{h}}\hat{q}\,\text{d}\hat{q}/\text{d}\hat{x}}{\hat{Q}^{8/7}(2\unicode[STIX]{x1D705}-{\textstyle \frac{77}{48}})\hat{q}^{2}}.\end{eqnarray}$$

Starting from ${\hat{h}}(-1/2)=0$ at the upstream boundary, this nonlinear equation can be integrated numerically in the phase plane $(\hat{x},{\hat{h}})$ to obtain the solution curve ${\hat{h}}(\hat{x})$ (figure 4 b). Because the solution has vertical tangents at end points $\hat{x}=\pm 1/2$ , we integrate using discrete steps of constant distance $\text{d}\hat{\unicode[STIX]{x1D709}}=(\text{d}\hat{x}^{2}+\text{d}{\hat{h}}^{2})^{1/2}$ rather than constant intervals $\text{d}\hat{x}$ . Next, we deduce the excess inclination profile ${\hat{S}}$ by substituting the solution to (2.37) back into (2.33). Finally, the free surface profile $\hat{\widetilde{z}}(\hat{x})=\widetilde{z}/(S_{a}L/\cos \unicode[STIX]{x1D6FC})$ is determined by integrating $\text{d}\hat{\widetilde{z}}/\text{d}\hat{x}=-{\hat{S}}$ numerically subject to the integral condition $\int _{-1/2}^{1/2}\hat{\widetilde{z}}\text{d}\hat{x}=0$ . Results are plotted in figure 4(c), together with the streamlines calculated from (2.16). For the experiments below, the dimensionless silo discharge $\hat{Q}$ takes relatively low values, hence the inertialess solution (2.36) provides a good approximation to the full solution of (2.37). Because the basal forcing term and discharge distribution $\hat{q}(\hat{x})$ gradually evolve as the avalanching layer descends down the silo, note that solutions must be calculated anew for each time at which a snapshot of the flow is desired.

4.1 Deep flow comparison

Prior to comparing theory with experiments for the deep flow region, the kinematic constant or diffusivity $K$ must be calibrated. This is done for each experiment using the following approach. Consider the rate of material transfer (per unit interface area) across an imaginary horizontal interface $Z(X,t)=H(t)$ descending at speed ${\dot{H}}=-Q/L$ down the silo, and given by

(4.1) $$\begin{eqnarray}W_{0}^{\prime }(X,Z)=W_{0}(X,Z)-{\dot{H}},\end{eqnarray}$$

where only the deep flow is considered. Assuming that the flow above this interface is steady in the descending frame of reference, this influx must be balanced by a horizontal granular current $J(X,Z)$ (per unit gap width) given by

(4.2) $$\begin{eqnarray}J(X,Z)=-\int _{0}^{X}W_{0}^{\prime }(X^{\prime },Z)\,\text{d}X^{\prime }.\end{eqnarray}$$

The control volumes associated with this balance are illustrated in figure 6(a). Let $J_{max}(Z)$ denote the maximum value of the granular current $J(X,Z)$ . Taken along horizontal interface $Z=H$ instead of the tilted interface $\text{}\underline{Z}(X)$ , it provides an approximation to the maximum avalanching discharge $q_{max}$ . Here we use measurements of $J_{max}$ for different elevations $Z$ to calibrate the diffusivity $K$ for each experiment. Using the deep flow theory of sub-section 2.1, the current $J$ is given by

(4.3) $$\begin{eqnarray}J(X,Z)=\unicode[STIX]{x1D6F9}_{0}(X,Z)+{\dot{H}}X,\end{eqnarray}$$

or in dimensionless form

(4.4) $$\begin{eqnarray}\hat{J}=\frac{J}{Q}=\hat{\unicode[STIX]{x1D6F9}}_{0}(\hat{X},\hat{Z})-\hat{X},\end{eqnarray}$$

where, like before, $\hat{X}=X/L$ and $\hat{Z}=ZK/L^{2}$ . Let $\hat{J}_{max}=F(\hat{Z})$ denote the maximum dimensionless current for a given dimensionless elevation $\hat{Z}$ . This relation can be inverted (numerically) to obtain for $Z$ as a function of $J_{max}$ the relation

(4.5) $$\begin{eqnarray}\frac{Z}{L}=\frac{L}{K}F^{-1}\left(\frac{J_{max}}{Q}\right),\end{eqnarray}$$

where we have reverted to dimensional variables. This provides a simple linear relation in coefficient $L/K$ that can be fitted to measurements of $J_{max}/Q$ for different relative elevations $Z/L$ .

Results obtained via this calibration procedure are presented in figure 8 for one of the experimental runs. In figure 8(a), relation (4.5) is compared with the measured data for the $K$ value yielding the best fit. In figure 8(b), we then compare the theoretical profiles $W_{0}(X,Z)$ obtained with this $K$ value with the measured vertical velocity profiles. Finally, in figure 8(c), we compare the theoretical profiles $J(X,Z)$ with the corresponding experimental profiles. In addition to matching well the values of maximum current $J_{max}(Z)$ , on which the fitting was performed (figure 8 a), the results show good agreement for other features of the deep flow, including the detailed velocity profiles (figure 8 b) and the locus $X_{max}(Z)$ where the maximum current $J_{max}$ is attained (figure 8 c).

Figure 8. Comparison of the kinematic theory (blue curves) with experimental measurements (black symbols) for the deep flow region (run 2): (a) dependence of maximum horizontal current $J_{max}$ on elevation above orifice $Z$ , used to calibrate the diffusivity $K$ ; (b) vertical velocity profiles $W(X,Z)$ ; (c) horizontal current profiles $J(X,Z)$ , with locus of maximum values indicated by dashed line (theory) and circles (experiment). The error bars indicate $\pm$ one standard deviation, estimated from measurements of three repeat experiments for the same conditions, and multiplied for clarity by factors 20, 40 and 3 for panels (a), (b) and (c) respectively.

Considering that the diffusivity $K$ is the only parameter requiring calibration, agreement is good overall. Nevertheless, there are some significant differences between the predicted and measured velocity profiles. Most conspicuously (figure 8 b), the experiments feature a zone of low velocity near the side wall above the orifice, the thickness of which grows with height. This departure from our assumed condition of perfect slip may be partly due to the use of wood strips for the sides. The same behaviour, however, was also observed by Maiti et al. (Reference Maiti, Das and Das2016) in their experiments conducted in an eccentric silo made entirely from acrylic (polymethyl methacrylate) plates. Regardless of the side wall material, it is likely that grain mobility near the sides is reduced due to corner effects (grains in contact with both the front and side walls). This vertical boundary layer is not modelled by the theory, which instead assumes perfect slip along the side walls. Fortunately (figure 8 c), this local discrepancy exerts a limited influence on the integrated granular current profiles $J(X,Z)$ , hence its effect on the dynamics of the avalanching layer is expected to be small.

Figure 9. Variation of the dimensionless diffusivity with the dimensionless gap thickness for all experimental runs.

Using the larger set of experiments, in which we varied the outflow discharge, silo gap thickness and grain diameter (see table 1), we can further investigate the influence of these parameters on the calibrated diffusivity $K$ . The results, plotted in figure 9, are unexpectedly simple. For the range of conditions examined, $K$ is simply proportional to the gap thickness $B$ , satisfying to a close approximation the linear relation $K=\unicode[STIX]{x1D706}B$ , where $\unicode[STIX]{x1D706}=0.26$ . When the gap thickness is held constant, $K$ is independent of the discharge $Q$ , consistent with the rate-independent character of the kinematic model. Likewise, $K$ is found to be independent of the grain diameter $D$ .

Figure 10. Comparison of the deep flow theory (blue line) with all experimental measurements (black dots): (a) maximum current $J_{max}$ ; (b) location $X_{max}$ where the maximum current occurs.

As a consequence of this simple dependence, the $\hat{J}_{max}$ and $\hat{X}_{max}$ data for all the experimental runs should collapse together when plotted against dimensionless number ${\hat{H}}=HB/L^{2}$ . This is checked in figure 10. Upon substituting $\hat{Z}=\unicode[STIX]{x1D706}{\hat{H}}$ in (4.4), the data collapse and agree closely with the theoretical curve $\hat{J}_{max}({\hat{H}})$ (figure 10 a). Agreement for the locus $\hat{X}_{max}({\hat{H}})$ is also good (figure 10 b). The two-dimensional deep flow pattern above the orifice, therefore, is effectively the same for all cases when plotted in the dimensionless variables $(\hat{X},{\hat{H}})$ . To test whether this holds more generally, one would need to vary parameters over a broader range. For instance, it is unclear if the same reduction would hold for grains and walls of materials different from glass, which would affect their frictional properties. For the present experimental conditions, however, this reduction to a single dimensionless flow pattern accounts well for the deep flow observations.

4.2 Surface flow comparison

Once the basal forcing is known from the kinematic model, the dynamic model can be used to predict the surface avalanching flow. For this purpose, we estimate the basal normal velocity $\text{}\underline{w}_{0}$ or equivalently the basal streamfunction $\text{}\underline{\unicode[STIX]{x1D713}\!}_{\,0}$ from the analytical solutions (2.8) and (2.9). For the diffusivity, we adopt the values $K=\unicode[STIX]{x1D706}B$ estimated from the gap thickness $B$ using the linear relation calibrated in the previous section. For the rheological coefficients $\unicode[STIX]{x1D707}_{0}$ and $\unicode[STIX]{x1D712}$ , we adopt the values determined in § 3 from previous experiments by Jop et al. (Reference Jop, Forterre and Pouliquen2005) using the same granular material. Detailed comparisons will be presented for two runs, corresponding respectively to gap thicknesses $B=3.5$ and 5 mm. In figure 11, we first compare predicted and measured streamfunctions for the entire flow field, at three successive times. The streamlines shown represent equally spaced streamfunction contours, hence the distance between contours is inversely proportional to the discharge through the corresponding streamtube. For the theory, the streamfunction is obtained from

(4.6) $$\begin{eqnarray}\unicode[STIX]{x1D6F9}=\unicode[STIX]{x1D6F9}_{0}+\unicode[STIX]{x1D713}-{\displaystyle\mathop{\unicode[STIX]{x1D713}}\limits_{{\sim}}},\end{eqnarray}$$

where the first term represents the deep flow, the second term surface avalanching and the third their common part that must be subtracted in order not to be counted twice. This approach is used in boundary layer problems to match inner and outer solutions (Kevorkian Reference Kevorkian1990). By analogy, the same approach can be used here to bridge between the avalanching and deep flow models. In our case, however, a more formal boundary layer analysis is not feasible because we do not know the more general equations that would reduce to the avalanching and deep flow models in the appropriate limits. For the experiments, the streamfunction is obtained from the measured velocity field $(U,W)(X,Z)$ by integrating $U$ column-wise from the bottom and $W$ row-wise from the far side wall, then averaging the two results. Side by side comparisons between the theoretical and experimental streamlines obtained in this fashion are shown in figure 11. Going from top to bottom, the streamlines are first evenly spaced across the silo width, due the quasi-steady drawdown of the free surface. They then get focused by the avalanching layer, where the streamlines become very closely spaced, to connect with the uneven deep flow streamlines at their base. Thereafter, the streamlines gradually converge toward the orifice where they again become closely spaced. Theory and experiment are seen to be in good agreement regarding this overall pattern.

Figure 11. Comparison of predicted and measured streamlines for two different gap thicknesses, at three successive times: (a) prediction, and (b) measurements, for gap thickness $B=3.5~\text{mm}$ (run 3); (c) prediction, and (d) measurements for gap thickness $B=5~\text{mm}$ (run 5). Streamlines are plotted at equal streamfunction increments, each corresponding to $1/8$ of the total outflow discharge.

In figure 12, we compare predictions and measurements for the granular velocity fields. At five equally spaced times during the descent, we compare calculated results for the downslope velocity $u(x,z,t)$ (figure 12 a,c) with measured results for the magnitude of the velocity norm $((U-U_{0})^{2}+(W-W_{0})^{2})^{1/2}$ , from which the deep flow was subtracted (figure 12 b,d). On all panels, the last, lowermost snapshot shows the magnitude of the total velocity when the avalanching layer approaches the orifice level. In both theory and experiment, the avalanching layers exhibit lenticular depth profiles, with vanishing depths at both side walls and maximum depths $h_{max}$ that occur towards the orifice side of the silo. This asymmetry and the steepness of the avalanching layer become stronger as the surface descends closer to the orifice, becoming more strongly influenced by its localized pull. Surface velocities likewise accelerate as the avalanching layer drops.

Figure 12. Comparison of predicted and measured granular velocity fields for two different gap thicknesses, at six successive times: (a) prediction, and (b) measurements, for gap thickness $B=3.5~\text{mm}$ (run 3); (c) prediction, and (d) measurements for gap thickness $B=5~\text{mm}$ (run 5). Colours from blue to red indicate increasing velocity magnitude.

Comparison of the left and right panels of figure 12 further shows the influence of gap thickness $B$ on the avalanching flows. The maps of figure 12(a,b) correspond to gap thickness $B=3.5~\text{mm}$ , whereas those of figure 12(c,d) correspond to gap thickness $B=5~\text{mm}$ . For the larger gap thickness (figure 12 c,d), kinematic diffusion is stronger, hence a less localized surface drawdown by the deep flow. The maximum discharge $q_{max}$ that the avalanching layer must transfer to even out the drawdown is therefore reduced, hence weaker avalanching flows. The larger gap thickness, moreover, reduces the relative magnitude of wall friction, leading to deeper and slower avalanching layers for the same discharge. Note that this does not imply any change in the magnitude of the wall friction coefficient, assumed constant in the theory. For increasing gap width, the magnitude of the wall friction force relative to the other terms decreases because, in the width-averaged equation (2.13) and the ensuing depth-integrated equations, the wall friction force per unit width $-2\unicode[STIX]{x1D70F}_{W}/B$ is the only term that gets divided by the width $B$ . For the smaller gap thickness (figure 12 a,b), kinematic diffusion is weaker hence the deep flow drawdown is more concentrated above the orifice. The induced maximum avalanching discharge $q_{max}$ is therefore greater. Due to the smaller gap thickness, wall friction causes the avalanching layers to reduce their depth and steepen their inclination. As a result of these different effects, surface velocities are significantly increased compared to figure 12(c,d), even though the discharge $Q$ through the orifice is nearly the same.

Although the free surface is nearly linear for the wider gap (figure 12 c,d), for the thinner gap the surface curvature is more pronounced, transitioning from concave downward to concave upward going from upstream to downstream (figure 12 a,b). As it nears the orifice wall, the avalanching layer cusps upwards as it thins to a sharp tip. Similar behaviour was earlier described for symmetric silos, with avalanching layers that cusp upwards as they reach the silo axis of symmetry (Samadani et al. Reference Samadani, Mahadevan and Kudrolli2002). Although this behaviour was interpreted as resulting from shock formation, our shallow flow theory captures this phenomenon without a shock. Similar behaviour can be observed for flows in rotating drums, where the granular free surface goes from straight to curved as the rotation rate increases (Rajchenbach Reference Rajchenbach1990).

To make comparisons more precise, figure 13 plots together theoretical and experimental profiles for two depth-integrated quantities associated with the avalanching layers: the depth-integrated discharge $q=\int \,u\,\text{d}z$ , and the depth-integrated kinetic energy, $k=\int (u^{2}\,\text{d}z)/2$ , both varying with height above orifice $H$ and with the downslope coordinate $x$ . Since the discharge profile is controlled by the deep flow, the comparisons for $q$ represent a test of the kinematic model (figure 13 a,c). For a given discharge profile, on the other hand, the kinetic energy profile is controlled by the avalanche dynamics, hence comparisons for $k$ represent a test of the dynamic model (figure 13 b,d). In both the theory and experiments, peaks in $q$ and $k$ occur further downstream as the granular surface drops and approaches the orifice. Also, profiles for the smaller gap thickness (figure 13 a,b) are more asymmetric than profiles for the larger gap thickness (figure 13 c,d). Finally, the profiles for $k$ are more strongly peaked than the profiles for $q$ . For all these features, fairly good agreement is registered between theory and experiment. Although the measured signal for the kinetic energy is noisier, the level of agreement obtained for $q$ and $k$ indicates that both the kinematic and dynamic models perform reasonably well.

Figure 13. Comparison of theory (blue lines) and experiment (black lines) for profiles of discharge and kinetic energy, at six successive times during the descent of the avalanching layers, for two different gap thicknesses: (a) discharge profiles $q(x)$ , and (b) kinetic energy profiles $k(x)$ , for gap thickness $B=3.5~\text{mm}$ (run 3); (c) discharge profiles $q(x)$ , and (d) kinetic energy profiles $k(x)$ , for gap thickness $B=5~\text{mm}$ (run 5). Circles indicate peaks in the predicted profiles. The error bars in (a,b) indicate $\pm$ one standard deviation, estimated from three repeat experiments for the same conditions.