betaFeedback

WorksPhysics and Astronomy

Simulating the rheology of dense suspensions using pairwise formulation of contact, lubrication and Brownian forces

Abstract

Dense suspensions of solid particles in viscous liquid are ubiquitous in both industry and nature, and there is a clear need for efficient numerical routines to simulate their rheology and microstructure. Particles of micron size present a particular challenge: at low shear rates colloidal interactions control their dynamics while at high rates granular-like contacts dominate. While there are established particle-based simulation schemes for large-scale non-Brownian suspensions using only pairwise lubrication and contact forces, common schemes for colloidal suspensions generally are more computationally costly and thus restricted to relatively small system sizes. Here we present a minimal particle-based numerical model for dense colloidal suspensions which incorporates Brownian forces in pairwise form alongside contact and lubrication forces. We show that this scheme reproduces key features of dense suspension rheology near the collodial-to-granular transition, including both shear-thinning due to entropic forces at low rates and shear thickening at high rate due to contact formation. This scheme is implemented in LAMMPS, a widely-used open source code for parallelized particle-based simulations, with a runtime that scales linearly with the number of particles making it amenable for large-scale simulations.

Introduction

Dense suspensions of Brownian and non-Brownian solid particles in viscous liquid present intriguing flow properties, and understanding their rheology is a subject of both fundamental and technological relevance [1, 2]. Of particular interest are suspensions comprising particles with radius a≈1 μma \approx1\,\mathrm{\mu m}, or more broadly in the range 0.10.1–10 μm10\,\mathrm{\mu m}. These are present in numerous applications, in all areas of food science [3] and consumer products, as well as across manufacturing and construction sectors [4] and indeed in many geophysical contexts [5]. Often their physics are challenging because their Brownian diffusion time may be comparable to the processing or macroscopic timescales involved in their use, so that they sit at the boundary of colloidal and granular systems [6].

Particle-based simulation offers a promising route to better understand the physics of these materials, providing highly-resolved information complementary to what can be obtained by experiment. With simultaneous access to particle trajectories and bulk rheology, one might devise new micromechanical constitutive equations [7] or develop microstructural insight that could guide the future analysis of experimental data. Numerical models might also be useful for exploring parameter space and systematically linking aspects of particle-level physics (friction [8], adhesion [9], roughness [10]) to bulk flow behaviour. As a result, one might aim to optimise industrial processes such as mixing and extrusion or indeed to optimise the design of materials themselves through additives, using insight gained through particle-based simulation.

Stokesian Dynamics (SD) [11, 12] is a computational method used to simulate the rheological behavior of colloidal and granular particles suspended in a viscous fluid, addressing the special case of inertia-free flow where the Stokes number is zero [13]. The method involves balancing all of the forces on each particle by evaluating their velocities via a grand mobility matrix containing information on the relative positions of every particle in the system. Despite accurately capturing the long and short range hydrodynamic interactions between particles, SD has not been adopted widely as a predictive tool in applied and industrial settings in the same way as other particle-based simulation methods have, due to the complexity of its implementation and its computational expense (notwithstanding recent developments that significantly speed it up [14, 15]).

The discrete element method (DEM) [16], on the other hand, is a particle-based computational method (a variant of molecular dynamics) that is used widely to simulate the behavior of granular materials including powders, particles and grains taking into account their pairwise interactions. In contrast to SD, DEM does not balance forces on each particle. Instead inertia is present, and one simply sums the forces and the resultant leads to an acceleration that can be realised through a conventional time-stepping algorithm such as Velocity-Verlet. This approach has proven to be useful for studying overdamped suspensions under shear flow nonetheless [17, 18], where one introduces short ranged lubrication forces and sets the Stokes number to be O(10−2)O(10^{-2}) or smaller. This approach is pragmatic in the sense that the physics associated with flowing dense suspensions can be implemented in existing, widely-used codes with large user bases, so that they have a clear path to adoption in engineering and other applied contexts. To date there is not, to our knowledge, a discrete element method simulation that includes the relevant physics of dense suspensions at the colloidal-to-granular interface, accounting for short-ranged hydrodynamics, Brownian forces, and (frictional) particle-particle contacts.

Here we present a minimal particle-based simulation model for predicting the rheology of dense Brownian and non-Brownian suspensions. Our model comprises hydrodynamic lubrication, particle-particle contacts and Brownian forces. After first describing the model in detail, we present some aspects of the effective interactions and diffusion that arise, before giving a detailed account of the rheological predictions of the model. The model reproduces well the main features of the experimentally observed rheology of dense suspensions, namely a low shear rate plateau that gives way to shear thinning and later shear thickening as the shear rate is increased, with the relative viscosity of the suspension increasing sharply with solids volume fraction and particle-particle friction coefficient.

Methodology

We consider a model system of nearly-monodisperse solid spheres, dispersed at high solids volume fraction ϕ\phi in a density-matched Newtonian liquid. The microscopic physics included in our model represent a minimal set of ingredients necessary to make useful predictions of the rheology of suspensions comprising particles with radius in the range 10−710^{-7} to 10−4 m10^{-4}\,\mathrm{m}. The trajectories of individual particles with translational and rotational motion are governed by Langevin equations that comprise three force (F\mathbf{F}) and torque (T\mathbf{T}) contributions: direct particle contacts (FC\mathbf{F}^{\mathrm{C}}, TC\mathbf{T}^{\mathrm{C}}), hydrodynamics (FH\mathbf{F}^{\mathrm{H}}, TH\mathbf{T}^{\mathrm{H}}), and Brownian noise (FB\mathbf{F}^{\mathrm{B}}, TB\mathbf{T}^{\mathrm{B}}). The equations of motion for translation and rotation of the particles are written, respectively, as

mid2xidt2=∑j(Fi,jC+FiH,D)+∑jFi,jH,L+FiB,D+∑jFi,jB,L,(1)m_i \frac{d^2 x_i}{dt^2} = \sum_j \left(F^{\mathrm{C}}_{i,j} + F^{\mathrm{H,D}}_i\right) + \sum_j F^{\mathrm{H,L}}_{i,j} + F^{\mathrm{B,D}}_i + \sum_j F^{\mathrm{B,L}}_{i,j}, \tag*{(1)}
25miai2dΩidt=∑jTi,jC+TiH,D+∑jTi,jH,L+TiB,D+∑jTi,jB,L,(2)\frac{2}{5}m_i a_i^2 \frac{d\Omega_i}{dt} = \sum_j T^{\mathrm{C}}_{i,j} + T^{\mathrm{H,D}}_i + \sum_j T^{\mathrm{H,L}}_{i,j} + T^{\mathrm{B,D}}_i + \sum_j T^{\mathrm{B,L}}_{i,j}, \tag*{(2)}

where xix_i represents the position of particle ii, Ωi\Omega_i represents its rotational velocity, and aia_i and mim_i are its radius and mass respectively. The subscript ii represents single-body forces and torques acting on particle ii, while the subscript i,ji,j represents pairwise forces and torques acting between particles labelled ii and jj. The superscripts C, H, B, D, L refer to the force and torque components arising due to contacts (C), hydrodynamics (H) and Brownian (B) effects, with the latter two acting both through drag (D) and lubrication (L). Each of these force and torque terms is described in detail below. These equations of motion can be understood as Langevin equations in which the ⟨⋅⟩C\langle\cdot\rangle^{\mathrm{C}} terms represent particle-particle interactions; the ⟨⋅⟩H\langle\cdot\rangle^{\mathrm{H}} terms represent configuration-dependent viscous friction (i.e. dissipative forces linear in the particle velocities); and the ⟨⋅⟩B\langle\cdot\rangle_{B} terms represent configuration-dependent (multiplicative) noise. Although particle inertia is present in the model we omit fluid inertia [19], arguing that for the regimes of interest the principle contributions to the overall bulk rheology will come from particle-particle contact and hydrodynamic lubrication interactions. Particles are subjected to a liquid flow field given by U∞\mathbf{U}^{\infty} (acting through the body force FiH,D\mathbf{F}_{i}^{H,D} as described below), leading to a rate of strain tensor E=12(∇U∞+(∇U∞)T)\mathbf{E} = \frac{1}{2}\left(\nabla\mathbf{U}^{\infty} + \left(\nabla\mathbf{U}^{\infty}\right)^{\mathrm{T}}\right). Pairwise forces and torques are summed over the neighbours jj of each particle ii, and the positions, velocities and acceleration are updated in a stepwise manner following the Velocity-Verlet algorithm [20]. Below we describe each of the force and torque contributions in detail; shown in Figure 1 are illustrative schematics of each of the forces.

Figure 1. Snapshot of simulation box, and schematics of the leading pairwise force terms present in the model. In all cases the leading component of the pairwise force acts along the (positive or negative) direction of the unit vector ni,j\mathbf{n}_{i,j} pointing from the centre of particle ii to the centre of particle jj. (a) Snapshot of the simulation box showing particles of radius aa (red) and 1.4a1.4a (blue). Particles are mixed in approximately equal volume. (b) Contact force, with force acting along ni,j\mathbf{n}_{i,j}; shown from left to right are [i] sketch of contacting particles ii and jj, with the overlap δi,j\delta_{i,j} shown in green; [ii] the unit vector ni,j\mathbf{n}_{i,j} pointing from the centre of particle ii to jj; [iii] repulsive contact FC\mathbf{F}^{\mathrm{C}} forces acting along the positive and negative directions of ni,j\mathbf{n}_{i,j}. (c) Pairwise hydrodynamic lubrication force, with force set by the component of the relative particle velocity acting along ni,j\mathbf{n}_{i,j}; Shown from left to right are [i] sketch showing the velocity vectors U\mathbf{U} of neighbouring particles ii and jj; [ii] the relative velocity Uj−Ui\mathbf{U}_{j}-\mathbf{U}_{i} breaks down into tangential (yellow) and normal (red) components, with the latter pointing along ni,j\mathbf{n}_{i,j}; [iii] lubrication forces act along the positive and negative directions of ni,j\mathbf{n}_{i,j}, proportional to the normal part of the relative velocity. (d) Pairwise Brownian lubrication force with a random pairwise vector θi,j\boldsymbol{\theta}_{i,j} projected onto ni,j\mathbf{n}_{i,j}. Shown from left to right are [i] neighbouring particles with centre-to-centre unit vector ni,j\mathbf{n}_{i,j}; [ii] random vectors θi,j\boldsymbol{\theta}_{i,j} drawn from a Gaussian distribution (green) are projected onto ni,j\mathbf{n}_{i,j} by the tangential (yellow) and normal (red) operators; [iii] Brownian forces act along the positive and negative directions of ni,j\mathbf{n}_{i,j}, proportional to the normal part of the random vector.

Contact forces and torques

The particle-particle contact force Fi,jC\mathbf{F}_{i,j}^{C} follows a conventional granular-type interaction [16], and is activated for any two particles ii and jj for which the centre-to-centre distance ∣ri,j∣\lvert\mathbf{r}_{i,j}\rvert is smaller than the sum of the radii ai+aja_i+a_j. Contact forces include a repulsive part acting normal to the pairwise centre-to-centre vector ri,j\mathbf{r}_{i,j} (we define a unit vector ni,j=ri,j/∣ri,j∣\mathbf{n}_{i,j}=\mathbf{r}_{i,j}/\lvert\mathbf{r}_{i,j}\rvert), and a tangential part. For simplicity we model contacts as linear springs, so that particle pairs experience repulsive contact forces proportional to their scalar overlap, defined once in contact as δi,j=(ai+aj)−∣ri,j∣\delta_{i,j}=(a_i+a_j)-\lvert\mathbf{r}_{i,j}\rvert. The implementation of our model within LAMMPS [21] nonetheless allows straightforward implementation of more complex δi,j\delta_{i,j} dependence. Tangential forces are linear in ξi,j\boldsymbol{\xi}_{i,j}, a vector describing the accumulated displacement of the particle pair perpendicular to ni,j\mathbf{n}_{i,j} since the initiation of the contact. Contact force and torque magnitudes are controlled by normal and tangential stiffness constants knk_n and ktk_t that set the hardness of the particles. The force and torque are given respectively by:

Fi,jC=knδi,jni,j−ktξi,j,(3)\mathbf{F}_{i,j}^{C}=k_n\delta_{i,j}\mathbf{n}_{i,j}-k_t\boldsymbol{\xi}_{i,j}, \tag*{(3)}
Ti,jC=ai(ni,j×ktξi,j).(4)\mathbf{T}_{i,j}^{C}=a_i\left(\mathbf{n}_{i,j}\times k_t\boldsymbol{\xi}_{i,j}\right). \tag*{(4)}

We additionally introduce a static friction coefficient μ\mu that constrains the tangential force to ∣ktξi,j∣≤μknδi,j\lvert k_t\boldsymbol{\xi}_{i,j}\rvert\leq\mu k_n\delta_{i,j}. For larger values of ktξi,jk_t\boldsymbol{\xi}_{i,j}, the tangential part of the force and the torque are truncated, and particle contacts transition from a rolling to a sliding regime. We present data for μ=0\mu=0 throughout, except in Figure 6(b)-(Figure 6c) where we explore the role of contact friction. Each pairwise contact between particles ii and jj contributes to the overall contact stress of the system with a tensorial stresslet given by the outer product −Fi,jC⊗ri,j-\mathbf{F}_{i,j}^{C}\otimes\mathbf{r}_{i,j}. The contact stress Σi,jC\boldsymbol{\Sigma}_{i,j}^{C} is obtained by summing this quantity over all contacting particle pairs and dividing by the system volume and dimension.

The viscosity variation with volume fraction and particle-particle contact friction

Figure 6. The viscosity variation with volume fraction and particle-particle contact friction. (a) Variation of ηr\eta_{\mathrm{r}} with volume fraction ϕ\phi at three PePe, showing fits to ηr=(1−ϕ/ϕJ)−λ\eta_{\mathrm{r}} = (1-\phi/\phi_J)^{-\lambda}; (b) ηr\eta_{\mathrm{r}} as a function of PePe for several particle-particle friction coefficients μ\mu and repulsive force magnitudes A/kbTA/k_{\mathrm{b}}T at a volume fraction of ϕ=0.5\phi= 0.5; (c) ηr\eta_{\mathrm{r}} as a function of PePe for several ϕ\phi and A/kbTA/k_{\mathrm{b}}T, with μ=0.5\mu= 0.5. The colour legend in (b) refers also to (c).

Contact forces of this kind have successfully been deployed in numerical models for rate-independent granular suspension rheology [17, 22, 23] and for models of shear thickening suspensions [8] (in the latter case rate dependence arises from a ‘critical load’ that the contact force must exceed before static friction is activated).

Hydrodynamic forces and torques

In general, hydrodynamic interactions in suspensions appear as single particle drag forces FH,D\mathbf{F}^{H,D}, pairwise near-contact lubrication forces FH,L\mathbf{F}^{H,L}, and many-body long range forces. In high volume fraction dense suspensions, however, it is argued by many authors that the hydrodynamic interactions are dominated by near-contact lubrication interactions [24] (which diverge on close approach) and that long range interactions are effectively screened by intervening particles [8, 25]. We follow this reasoning and therefore omit long range hydrodynamics from our model. Below we describe in detail the drag and lubrication forces deployed in the model. Single particle drag forces and torques are given by

FiH,D=6πηai(U∞(xi)−Ui),(5)\mathbf{F}_{i}^{H,D}=6\pi\eta a_i\left(\mathbf{U}^{\infty}(\mathbf{x}_i)-\mathbf{U}_i\right), \tag*{(5)}
TiH,D=8πηai3(Ω∞−Ωi),(6)T_i^{\mathrm{H,D}} = 8\pi\eta a_i^3\left(\Omega^\infty-\Omega_i\right), \tag*{(6)}

where we use the isolated-particle Stokes terms and, for simplicity, do not introduce volume fraction dependent hindrance functions. Here η\eta is the liquid viscosity, U∞(xi)\mathbf{U}^\infty(\mathbf{x}_i) is the value of the liquid streaming velocity at the position of the centre of mass of particle ii, and Ω∞=12(∇×U∞)\boldsymbol{\Omega}^\infty= \frac{1}{2}\left(\nabla\times\mathbf{U}^\infty\right) (spatially uniform assuming U∞\mathbf{U}^\infty is uniform in space). The drag forces lead to a per particle stress given by ΣiH,D=203πηai3E\boldsymbol{\Sigma}_i^{\mathrm{H,D}} = \frac{20}{3}\pi\eta a_i^3\mathbf{E}.

For pairwise lubrication forces and torques acting between interacting particles ii and jj we start from the conventional representation given by Kim and Karrila [26] as

\begin{pmatrix} \mathbf{F}_{i,j}^{\mathrm{H,L}} \\ \mathbf{F}_{j,i}^{\mathrm{H,L}} \\ \mathbf{T}_{i,j}^{\mathrm{H,L}} \\ \mathbf{T}_{j,i}^{\mathrm{H,L}} \\ \boldsymbol{\Sigma}_{i}^{\mathrm{H,L}} \\ \boldsymbol{\Sigma}_{j,i}^{\mathrm{H,L}} \end{pmatrix} = \eta\mathbf{R} \begin{pmatrix} \mathbf{U}^\infty(\mathbf{x}_i)-\mathbf{U}_i \\ \mathbf{U}^\infty(\mathbf{x}_j)-\mathbf{U}_j \\ \boldsymbol{\Omega}^\infty-\boldsymbol{\Omega}_i \\ \boldsymbol{\Omega}^\infty-\boldsymbol{\Omega}_j \\ \mathbf{E} \\ \mathbf{E} \end{pmatrix}, \tag*{(7)}\end{pmatrix}

where R\mathbf{R} is the resistance matrix containing tensorial operations that linearly couple particle forces (torques) to velocities (rotational velocities), taking into account relative particle positions. After some algebra and omitting terms that vanish with the size of the interparticle gap (see Radhakrishnan [27] for details) one can obtain the forces in a simplified pairwise form as

Fi,jH,L=−Fj,iH,L=(X11ANi,j+Y11ATi,j)(Uj−Ui)+Y11B(Ωi×ni,j)+Y21B(Ωj×ni,j),(8)F^{\mathrm{H,L}}_{i,j}=-F^{\mathrm{H,L}}_{j,i}=\left(X^{A}_{11}\mathbf{N}_{i,j}+Y^{A}_{11}\mathbf{T}_{i,j}\right)\left(U_j-U_i\right) +Y^{B}_{11}\left(\Omega_i\times n_{i,j}\right) +Y^{B}_{21}\left(\Omega_j\times n_{i,j}\right), \tag*{(8)}

where Fi,jH,LF^{\mathrm{H,L}}_{i,j} is the force acting on particle ii by particle jj; N=ni,j⊗ni,j\mathbf{N}=n_{i,j}\otimes n_{i,j} is a tensorial normal operator; T=I−ni,j⊗ni,j\mathbf{T}=\mathbf{I}-n_{i,j}\otimes n_{i,j} is a tensorial projection operator; ni,jn_{i,j} is the unit vector pointing from particle ii to particle jj; UiU_i is the velocity of particle ii; Ωi\Omega_i is the rotational velocity of particle ii; and I\mathbf{I} is the identity tensor in three dimensions. The scalar prefactors XX and YY encode the geometry of the interacting pair, namely the size of the interparticle gap and the size ratio of the interacting particles. Their superscripts AA, BB and subscripts 11, 22 are more appropriate to the labelling convention used by Kim and Karrila [26] but nonetheless we retain them here for ease of referencing to that work. The particle size ratio is written as β=aj/ai\beta=a_j/a_i and the dimensionless interparticle gap is ξ=2(∣ri,j∣−(ai+aj))/(ai+aj)\xi=2\left(\lvert r_{i,j}\rvert-(a_i+a_j)\right)/(a_i+a_j). The scalar prefactors are given by

X11A=6πηai(2β2(1+β)31ξ+β(1+7β+β2)5(1+β)3ln⁡(1ξ)),(9)X^{A}_{11}=6\pi\eta a_i\left(\frac{2\beta^2}{(1+\beta)^3}\frac{1}{\xi}+\frac{\beta(1+7\beta+\beta^2)}{5(1+\beta)^3}\ln\left(\frac{1}{\xi}\right)\right), \tag*{(9)}
Y11A=6πηai(4β(2+β+2β2)15(1+β)3ln⁡(1ξ)),(10)Y^{A}_{11}=6\pi\eta a_i\left(\frac{4\beta(2+\beta+2\beta^2)}{15(1+\beta)^3}\ln\left(\frac{1}{\xi}\right)\right), \tag*{(10)}
Y11B=−4πηai2(β(4+β)5(1+β)2ln⁡(1ξ)),(11)Y^{B}_{11}=-4\pi\eta a_i^2\left(\frac{\beta(4+\beta)}{5(1+\beta)^2}\ln\left(\frac{1}{\xi}\right)\right), \tag*{(11)}
Y21B=−4πηaj2(β−1(4+β−1)5(1+β−1)2ln⁡(1ξ)).(12)Y^{B}_{21}=-4\pi\eta a_j^2\left(\frac{\beta^{-1}(4+\beta^{-1})}{5(1+\beta^{-1})^2}\ln\left(\frac{1}{\xi}\right)\right). \tag*{(12)}

Meanwhile the torques on particle ii and jj as a result of their interaction with particles jj and ii respectively are written as

Ti,jH,L=Y11B(Uj−Ui)×ni,j−Ti,j(Y11CΩi+Y12CΩj),(13)T^{\mathrm{H,L}}_{i,j}=Y^{B}_{11}(U_j-U_i)\times n_{i,j}-\mathbf{T}_{i,j}\left(Y^{C}_{11}\Omega_i+Y^{C}_{12}\Omega_j\right), \tag*{(13)}
Tj,iH,L=Y21B(Uj−Ui)×ni,j−Ti,j(Y21CΩi+Y22CΩj),(14)T^{\mathrm{H,L}}_{j,i}=Y^{B}_{21}(U_j-U_i)\times n_{i,j}-\mathbf{T}_{i,j}\left(Y^{C}_{21}\Omega_i+Y^{C}_{22}\Omega_j\right), \tag*{(14)}

with scalar prefactors given by

Y11C=8πηai3(2β5(1+β)ln⁡(1ξ)),(15)Y^{C}_{11}=8\pi\eta a_i^3\left(\frac{2\beta}{5(1+\beta)}\ln\left(\frac{1}{\xi}\right)\right), \tag*{(15)}
Y12C=8πηai3(β210(1+β)ln⁡(1ξ)),(16)Y^{C}_{12}=8\pi\eta a_i^3\left(\frac{\beta^2}{10(1+\beta)}\ln\left(\frac{1}{\xi}\right)\right), \tag*{(16)}
Y22C=8πηaj3(2β−15(1+β−1)ln⁡(1ξ)),(17)Y^{C}_{22}=8\pi\eta a_j^3\left(\frac{2\beta^{-1}}{5(1+\beta^{-1})}\ln\left(\frac{1}{\xi}\right)\right), \tag*{(17)}
Y21C=8πηaj3(β−210(1+β−1)ln⁡(1ξ)).(18)Y^{C}_{21}=8\pi\eta a_j^3\left(\frac{\beta^{-2}}{10(1+\beta^{-1})}\ln\left(\frac{1}{\xi}\right)\right). \tag*{(18)}

Similar expressions may be obtained for the elements of the hydrodynamic lubrication stress tensor, though these can be shown to be equivalent (up to an order ξ\xi term in the normal stresses) to the form used for the contact forces. The contribution to the hydrodynamic stress coming from each pairwise interaction is thus given by Σi,jH,L=−Fi,jH,L⊗ri,j\mathbf{\Sigma}^{\mathrm{H,L}}_{i,j}=-F^{\mathrm{H,L}}_{i,j}\otimes r_{i,j}. To mitigate against divergence in the scalar prefactors at particle contacts (that is, where ξ→0\xi\to0) we use ξeff=10−3\xi_{\mathrm{eff}}=10^{-3} in the calculation whenever ξ<10−3\xi<10^{-3}. We do not calculate pairwise lubrication forces when particles are separated by gaps ξ>0.05\xi>0.05, having verified that this choice does not affect our conclusions.

Brownian forces and torques

To satisfy fluctuation-dissipation theorem, we must produce Brownian forces that follow

⟨FB⊗FB⟩=2kbTΔtR,(19)\langle F_B\otimes F_B\rangle=\frac{2k_bT}{\Delta t}\mathcal{R}, \tag*{(19)}

where FBF_B is a list of the Brownian forces and torques, R\mathcal{R} is the overall resistance operator for the system (taking into account both one body and pairwise hydrodynamic dissipation that we describe separately below), kBTk_\mathrm{B}T is the thermal energy and Δt\Delta t is the computational timestep (discussed in more detail below).

For one-body Brownian forces we need 6 random numbers (i.e. two vectors in three-dimensional space ψi\boldsymbol{\psi}_i and φi\boldsymbol{\varphi}_i) to satisfy the translational and rotational degrees of freedom of each particle ii. The elements of the random vectors ψi\boldsymbol{\psi}_i, φi\boldsymbol{\varphi}_i are drawn from a Gaussian distribution and satisfy ⟨φαφβ⟩=⟨ψαψβ⟩=δαβ\langle\varphi_\alpha\varphi_\beta\rangle=\langle\psi_\alpha\psi_\beta\rangle=\delta_{\alpha\beta} and they are uncorrelated with each other so that ⟨φαψβ⟩=0\langle\varphi_\alpha\psi_\beta\rangle=0. The following forces and torques satisfy fluctuation-dissipation theorem (we label them as Brownian drag ‘B,D’ to align with the hydrodynamic drag forces and torques defined above). The one-body Brownian force and torque on particle ii are given by:

FiB,D=2kBTΔt6πηaiψi.(20)\mathbf{F}_i^{B,D}=\sqrt{\frac{2k_\mathrm{B}T}{\Delta t}}\sqrt{6\pi\eta a_i}\boldsymbol{\psi}_i. \tag*{(20)}
TiB,D=2kBTΔt8πηai3φi.(21)\mathbf{T}_i^{B,D}=\sqrt{\frac{2k_\mathrm{B}T}{\Delta t}}\sqrt{8\pi\eta a_i^3}\boldsymbol{\varphi}_i. \tag*{(21)}

Averaging ⟨FiB,D⊗FiB,D⟩\langle\mathbf{F}_i^{B,D}\otimes\mathbf{F}_i^{B,D}\rangle and ⟨TiB,D⊗TiB,D⟩\langle\mathbf{T}_i^{B,D}\otimes\mathbf{T}_i^{B,D}\rangle over many realisations of the vectors ψi\boldsymbol{\psi}_i and φi\boldsymbol{\varphi}_i leads, respectively, to 2kBTΔt6πηaiI\frac{2k_\mathrm{B}T}{\Delta t}6\pi\eta a_i\mathbf{I} and 2kBTΔt8πηai3I\frac{2k_\mathrm{B}T}{\Delta t}8\pi\eta a_i^3\mathbf{I} as required (with I\mathbf{I} the identity matrix in three dimensions).

Pairwise Brownian forces and torques similarly require two random vectors θi,j\boldsymbol{\theta}_{i,j} and χi,j\boldsymbol{\chi}_{i,j} (independent of ψi\boldsymbol{\psi}_i and φi\boldsymbol{\varphi}_i but with the same properties) to satisfy the relative translational and rotational motion of two interacting particles. The pairwise forces and torques must be constructed in such a way that, for particles ii and jj, averaging ⟨Fi,jB,L⊗Fi,jB,L⟩\langle\mathbf{F}_{i,j}^{B,L}\otimes\mathbf{F}_{i,j}^{B,L}\rangle and ⟨Ti,jB,L⊗Ti,jB,L⟩\langle\mathbf{T}_{i,j}^{B,L}\otimes\mathbf{T}_{i,j}^{B,L}\rangle over many realisations of θi,j\boldsymbol{\theta}_{i,j} and χi,j\boldsymbol{\chi}_{i,j} recovers the form of the pairwise hydrodynamic lubrication forces and torques described above. Doing so, which involves exploiting that the normal and projection operators present in the definition of the lubrication forces and torques are idempotent (i.e. ⟨(Ni,jθi,j)⊗(Ni,jθi,j)⟩=Ni,j\langle(\mathbf{N}_{i,j}\boldsymbol{\theta}_{i,j})\otimes(\mathbf{N}_{i,j}\boldsymbol{\theta}_{i,j})\rangle=\mathbf{N}_{i,j}) and orthogonal (i.e. ⟨(Ni,jθi,j)⊗(Ti,jθi,j)⟩=0\langle(\mathbf{N}_{i,j}\boldsymbol{\theta}_{i,j})\otimes(\mathbf{T}_{i,j}\boldsymbol{\theta}_{i,j})\rangle=0), one obtains the following expressions for the pairwise Brownian force and torque

Fi,jB,L=−Fj,iB,L=2kBTΔt(X11ANi,j+Y11ATi,j)θi,j,(22)\mathbf{F}_{i,j}^{B,L}=-\mathbf{F}_{j,i}^{B,L}=\sqrt{\frac{2k_\mathrm{B}T}{\Delta t}}\left(\sqrt{X_{11}^{A}}\mathbf{N}_{i,j}+\sqrt{Y_{11}^{A}}\mathbf{T}_{i,j}\right)\boldsymbol{\theta}_{i,j}, \tag*{(22)}
Ti,jB,L=2kBTΔt(YB11YA11θi,j×ni,j+YC11−(YB11)2YA11Tχi,j),(23)\mathbf{T}_{i,j}^{B,L}=\sqrt{\frac{2k_\mathrm{B}T}{\Delta t}}\left(\frac{Y_{B}^{11}}{\sqrt{Y_{A}^{11}}}\boldsymbol{\theta}_{i,j}\times\mathbf{n}_{i,j}+\sqrt{Y_{C}^{11}-\frac{(Y_{B}^{11})^2}{Y_{A}^{11}}}\mathbf{T}\boldsymbol{\chi}_{i,j}\right), \tag*{(23)}
Tj,iB,L=2kBTΔt(YB21YA11θi,j×ni,j+YC22−(YB21)2YA11Tχi,j).(24)\mathbf{T}_{j,i}^{B,L}=\sqrt{\frac{2k_\mathrm{B}T}{\Delta t}}\left(\frac{Y_{B}^{21}}{\sqrt{Y_{A}^{11}}}\boldsymbol{\theta}_{i,j}\times\mathbf{n}_{i,j}+\sqrt{Y_{C}^{22}-\frac{(Y_{B}^{21})^2}{Y_{A}^{11}}}\mathbf{T}\boldsymbol{\chi}_{i,j}\right). \tag*{(24)}

Our model thus involves computing Equations (1) and (2) to evaluate the trajectory of each particle, subject to imposed forces given by Equations (3), (5), (8), (20), and (22), and torques given by Equations (4), (6), (13), (14), (21), (23) and (24).

Brownian stress calculation

One can similarly obtain from fluctuation-dissipation theorem an expression for the Brownian stress resulting from the pairwise interaction between particles ii and jj that averages over many realisations so that ⟨Fi,jB,L⊗Fi,jB,L⟩\langle\mathbf{F}_{i,j}^{B,L}\otimes\mathbf{F}_{i,j}^{B,L}\rangle recovers the form of the hydrodynamic lubrication stress, but as described above this can similarly be shown to be equivalent to Σi,jB,L=−Fi,jB,L⊗ri,j\boldsymbol{\Sigma}_{i,j}^{B,L}=-\mathbf{F}_{i,j}^{B,L}\otimes\mathbf{r}_{i,j}. Since the pairwise Brownian force term contains the normal operator Ni,j\mathbf{N}_{i,j} acting on the random vector θi,j\boldsymbol{\theta}_{i,j}, one obtains a prefactor in the stress containing the dot product ni,j⋅θi,j\mathbf{n}_{i,j}\cdot\boldsymbol{\theta}_{i,j}. This quantity will always approach zero when averaged over many realisations of θi,j\theta_{i,j}, so that the Brownian stress computed in this way averages to zero. Nonetheless, particle pairs do experience non-zero Brownian forces acting at all timesteps that will influence their trajectories so that the resulting contact and lubrication stresses will be altered by the presence of the Brownian forces. Below we describe a method that allows us to estimate the contribution of Brownian motion to the overall stress.

It is important to note here that our method, in which particle inertia is accounted for, is fundamentally different to other computational approaches, notably Stokesian Dynamics (SD) [11, 13, 28] in which the trajectories are evolved with a timestep longer than the inertial one. In the latter methods (see in particular Banchio and Brady [12]) the Brownian stress for the overall system is obtained as ΣB=kbT∇⋅(RSU⋅RFU−1)\boldsymbol{\Sigma}^{B}=k_bT\nabla\cdot\left(\mathcal{R}_{SU}\cdot\mathcal{R}_{FU}^{-1}\right), in practice using a midpoint scheme in which the positions and velocities of every particle are sampled at some increment of the overall timestep. Here RSU\mathcal{R}_{SU} and RFU\mathcal{R}_{FU} represent parts of the overall resistance matrix that couple, respectively, stresses to velocities and forces to velocities. Our method described above is based on the Langevin equation so that particle inertia is small but present, and force balance is not strictly achieved at each timestep. In order to obtain an estimate of the Brownian contribution to the stress, we deploy a structural method that exploits the anisotropy of the radial distribution function, using the approach described by Brady [29]. The Brownian stress attributable to the pair i,ji,j can be written as

Σi,jB=−nkbTa∫S2(ni,j⊗ni,j)p1/1(xj∣xi) dS2,(25)\boldsymbol{\Sigma}^{B}_{i,j}=-nk_bTa\int_{S_2}\left(\mathbf{n}_{i,j}\otimes\mathbf{n}_{i,j}\right)p_{1/1}(\mathbf{x}_j\mid\mathbf{x}_i)\,\mathrm{d}\mathbf{S}_2, \tag*{(25)}

where p1/1(xj∣xi)p_{1/1}(\mathbf{x}_j\mid\mathbf{x}_i) is the probability density for finding a particle at xj\mathbf{x}_j given that there is a particle at xi\mathbf{x}_i, and n=N/Vn=N/V is the number density of particles in the suspension (where VV and NN are the system volume and particle number respectively). The integral is over the surface of contact S2S_2 of two touching particles.

To compute this function we sum for each particle the diadic product of its unit vector with each of its neighbours within a thin shell Δ=0.05ai\Delta=0.05a_i, so that for a given configuration the Brownian contribution to the stress is [30]:

ΣB=−kbTV∑iaiΔ∑j∈Δ(ni,j⊗ni,j).(26)\boldsymbol{\Sigma}^{B}=-\frac{k_bT}{V}\sum_i\frac{a_i}{\Delta}\sum_{j\in\Delta}\left(\mathbf{n}_{i,j}\otimes\mathbf{n}_{i,j}\right). \tag*{(26)}

The stress measured by this approach is not added to the hydrodynamic and contact stresses computed in our model, but is available to provide insight into the role of Brownian motion in setting the overall material response.

Additional simulation details

We simulate O(103)\mathcal{O}(10^3) spherical particles of radius aa and 1.4a1.4a (mixed approximately equally by volume) in a cubic periodic simulation box of length LL. For each set of flow conditions we carried out between 10 and 800 realisations in order to obtain satisfactory ensemble averages. The principle particle properties (these set the length, mass and time scales) are the characteristic particle radius aa [length], the particle density ρ\rho [mass/length3^3] (taken throughout to be equal to the fluid density so that the particles are neutrally buoyant), and the particle normal stiffness knk_n [mass/time2^2] (this has a tangential counterpart ktk_t). With respect to these quantities, 1 time unit corresponds to the inverse frequency of a mass m=a3=1m=a^3=1 on a linear spring with stiffness kn=1k_n=1. The remaining material properties to be defined are the fluid viscosity η\eta [mass/(length×\timestime)] and the particle-particle friction coefficient μ\mu [dimensionless], relevant for micron sized (and larger) particles. The thermal energy scale in the system is set by kbTk_bT.

The simulation box is deformed according to a specified ∇U∞\nabla\mathbf{U}^{\infty}. For instance, when the only nonzero element of ∇U∞\nabla\mathbf{U}^{\infty} is an off-diagonal (say γ˙\dot{\gamma}), shearing is applied by tilting the triclinic box (at fixed volume) according to Lxy(t)=Lxy(t0)+Lγ˙tL_{xy}(t)=L_{xy}(t_0)+L\dot{\gamma}t. When the strain (γ=γ˙t\gamma=\dot{\gamma}t, with tt the time for which the simulation has run) reaches 0.5 in this example, the system is remapped to a strain of -0.5. This has no effect on the particle-particle forces or on the stress, and is simply a numerical tool to permit unbounded shear deformation while preventing the domain from becoming elongated in one axis [31]. Reported in the following is the relative viscosity of the suspension ηr=Σxy/ηγ˙\eta_r=\Sigma_{xy}/\eta\dot{\gamma}, with Σxy\Sigma_{xy} the shear component of the stress tensor, γ˙\dot{\gamma} the shear rate and η\eta the fluid viscosity.

The timescales that appear in the simulation

The full list of dimensional parameters taken as inputs to the model is then aa, LL, tt, ρ\rho, knk_n, kBTk_BT, η\eta and γ˙\dot{\gamma}. Taking a/L≪1a/L \ll1 and γ˙t≫1\dot{\gamma}t \gg1, dimensional analysis dictates that we require three non-dimensional groups to fully characterise this system. In other words, a measured non-dimensional quantity e.g. the reduced viscosity ηr=Σxy/ηγ˙\eta_r = \Sigma_{xy}/\eta\dot{\gamma}, can be a function of at most three non-dimensional control parameters. This is in addition to non-dimensional inputs viz. the volume fraction ϕ\phi and the friction coefficient μ\mu. Central to our work will be the study of viscosities as a function of Peclet number, since this latter quantity will control the colloidal to granular crossover. It is desirous to choose the remaining two non-dimensional control parameters such that particles are effectively hard and non-inertial. To obtain an appropriate set of non-dimensional control parameters, we consider the following list of timescales present in the model (in which we only include dimensional elements for simplicity):

τC=ρa3kn,τI=ρa2η,τB=ηa3kBT,τS=1γ˙.\tau_C = \sqrt{\frac{\rho a^3}{k_n}}, \qquad \tau_I = \frac{\rho a^2}{\eta}, \qquad \tau_B = \frac{\eta a^3}{k_BT}, \qquad \tau_S = \frac{1}{\dot{\gamma}}.

The contact time τC\tau_C is a characteristic time spent by two particles in contact (assuming contacts are describable as linear springs), in the absence of other forces playing a role. It is obtained by solving the following equation of motion for the overlap δ\delta between contacting particles: ρa3(d2δ/dt2)=knδ\rho a^3(d^2\delta/dt^2) = k_n\delta. The inertial relaxation time τI\tau_I is the characteristic time taken for the velocity of a particle to reach that of the background fluid in the absence of other forces. It is obtained by solving the following equation of motion for the velocity vv of a particle: ρa3(dv/dt)=ηav\rho a^3(dv/dt) = \eta av. The Brownian time τB\tau_B is the characteristic time take for a particle to diffuse by a distance equal to its own radius under thermal motion in the absence of other forces. The convective timescale τS\tau_S is simply the inverse of the shear rate. To resolve each of these timescales accurately within the simulation we chose the numerical timestep to be substantially smaller than the smallest of the timescales listed above. The Peclet number (PePe) described above is given by 6πτB/τS=6πηa3γ˙/kBT6\pi\tau_B/\tau_S = 6\pi\eta a^3\dot{\gamma}/k_BT, and we vary this quantity across a broad range from 0.01 to 100000, aiming to explore the colloidal to granular transition.

The contact time τC\tau_C should be chosen to be sufficiently small that overlaps between particles are orders of magnitude smaller than the particle radii, such that particles be considered hard spheres. To do this we ensure throughout that τC\tau_C is at least an order of magnitude smaller than the next smallest timescale. The role of particle inertia can be expressed via (i) a particle Reynolds number τI/τS=ρa2γ˙/η\tau_I/\tau_S = \rho a^2\dot{\gamma}/\eta, and (ii) an inertia-diffusion ratio τI/τB=ρkBT/η2a\tau_I/\tau_B = \rho k_BT/\eta^2a. Below we explore how small each of these quantities need to be set in order to ensure inertia plays no significant role in the measured results.

Results: interactions and diffusion

Two-particle simulations measuring the effective potential

To evaluate the net pairwise potential resulting from the particle-level forces described above, we carried out O(103)O(10^3) simulations of two particles with radii aa in a cubic periodic box of length 4a4a (see snapshot in Figure 2(a) Inset) subject to all of the forces described above, and with U∞=0U^\infty= 0. We calculate the radial distribution function g(r)g(r) with r=∣rij∣r = |\mathbf{r}_{ij}| and averaged this across timesteps in the steady state and across all realisations (Figure 2(a)), then obtained the potential of mean force as U(r)/kBT=−ln⁡(g(r))U(r)/k_BT = -\ln(g(r)), Figure 2(b). The result confirms that there is no net potential acting between particles when they are not in contact (i.e. when r>2ar > 2a), so the lubrication and Brownian forces do not introduce an overall repulsion or attraction. When particles are in contact (r/(ai+aj)<1r/(a_i+a_j) < 1) there is a steep repulsive potential that, as expected, is related to the stiffness of our contacts defined above as U(r)/kBT=0.5knδij2U(r)/k_BT = 0.5k_n\delta_{ij}^2. The model thus approximates a suspension of colloidal hard spheres, in which the particle-particle interaction is zero and infinite for non-contacts and contacts respectively.

Six-panel evaluation of the potential of mean force and diffusion properties

Figure 2. Evaluating the potential of mean force and the diffusion properties that arise from the particle-level forces described above, in the absence of shear flow. (a) The radial distribution function g(r)g(r) (with r=∣ri,j∣r = |\mathbf{r}_{i,j}|) computed from a two particle simulation [Inset: snapshot of simulation]; (b) Potential of mean force U(r)U(r), showing measured result (points) and the input particle stiffness (solid line); (c)-(d) Mean squared displacement as a function of elapsed time for (c) three values of the timescale ratio τI/τB\tau_I/\tau_B at ϕ=0.001\phi= 0.001; (d) three values of ϕ\phi at τI/τB=1.7\tau_I/\tau_B = 1.7; (e) Diffusion coefficient as a function of elapsed time for a range of ϕ\phi at τI/τB=1.7\tau_I/\tau_B = 1.7. The solid line in (c)-(e) represents the predictions of Equation (27); (f) Long time diffusion coefficient at a broad range of ϕ\phi and τI/τB\tau_I/\tau_B.

Mean square displacement

We next verify that our simulated particles follow statistically the anticipated trajectories by computing their mean squared displacement (MSD) under various conditions. An isolated particle with motion governed by the single body drag and Brownian forces described above is expected to follow a trajectory with a short-time ballistic part and a long-time diffusive part that leads to an overall MSD given by [32, 33]:

⟨x2⟩=2kbTmγ2(γmt−1+e−γmt),(27)\langle x^2 \rangle= 2k_bT\frac{m}{\gamma^2}\left(\frac{\gamma}{m}t - 1 + e^{-\frac{\gamma}{m}t}\right), \tag*{(27)}

with m=(4/3)πρa3m = (4/3)\pi\rho a^3 and γ=6πηa\gamma= 6\pi\eta a. This expression gives ⟨x2⟩∼t2\langle x^2 \rangle\sim t^2 and ⟨x2⟩∼t\langle x^2 \rangle\sim t at small and large times respectively. It can equivalently be written in terms of our characteristic timescales defined above as:

⟨x2⟩/a2=227πτIτB(4.5tτI−1+e−4.5tτI),(28)\langle x^2 \rangle/a^2 = \frac{2}{27\pi}\frac{\tau_I}{\tau_B}\left(4.5\frac{t}{\tau_I} - 1 + e^{-4.5\frac{t}{\tau_I}}\right), \tag*{(28)}

Shown in Figure 2(c) are MSDs for a dilute sample with ϕ=0.001\phi= 0.001 in which pairwise particle-particle interactions are absent. In terms of our model timescales, we set τS=∞\tau_S = \infty (i.e. no shear); τC=10−3\tau_C = 10^{-3}; τI=10−1\tau_I = 10^{-1}; and we vary τB\tau_B to explore the behaviour at different temperatures. We measure the elapsed time in units of τI\tau_I, so that the crossover from ballistic to diffusive behaviour begins in each case at t/τI∼1t/\tau_I \sim1. As expected based on the expression above, increasing temperature (which decreases τB\tau_B) while keeping all other variables constant simply shifts the MSD result vertically with ⟨x2⟩∼kbT\langle x^2 \rangle\sim k_bT.

We next calculate the MSD for a series of larger ϕ\phi, with results shown in Figure 2(d)-(Figure 2e). In all cases the particles follow a ballistic trajectory at short times that is roughly independent of ϕ\phi. The longer time behaviour shows a decreasing diffusion coefficient (D=d/dt(⟨x2⟩)\mathcal{D} = d/dt(\langle x^2 \rangle)) with increasing ϕ\phi, a consequence of pairwise hydrodynamic and contact interactions resisting particle motion. For all volume fractions below jamming D\mathcal{D} approaches a constant at long time scales, confirming the presence of a diffusive regime.

In order for inertia to play a negligible role in our model, it is important for the diffusive timescale to be longer than the inertial relaxation one. In other words, the time taken for a particle velocity to relax to that of the background fluid should be much shorter than the time taken for the particle to diffuse by its own radius. To understand quantitatively how to achieve this, we measured D\mathcal{D} for varying τI/τB\tau_I/\tau_B across a broad range of ϕ\phi. The normalized long time diffusion coefficient (D(ϕ)/D0)(\mathcal{D}(\phi)/\mathcal{D}_0) is shown in Figure 2(f), with D0=kBT/πηa(=a2/τB)\mathcal{D}_0=k_BT/\pi\eta a(=a^2/\tau_B). Our result shows that when τI/τB\tau_I/\tau_B is smaller than 0.17, D(ϕ)/D0\mathcal{D}(\phi)/\mathcal{D}_0 becomes independent of temperature and follows a linearly decreasing trend. This suggests a criteria for the maximum value of τI/τB\tau_I/\tau_B, which we check under shearing conditions in the following.

Results: rheology

In the following we first describe the need for substantial ensemble averaging, especially when Brownian motion dominates, and we demonstrate the convergence of the measured rheology with the size of the sampling window. We next go on to expose the role of particle inertia in our model under shear, and establish the parameter range in which it can be assumed negligible. We then present rheology data showing ηr\eta_r as a function of PePe, highlighting the breakdown of the individual contributions (hydrodynamic, contact and Brownian) and their variation with volume fraction. We finally demonstrate the role of particle contact friction and a short-ranged repulsive potential.

Averaging method

All of the rheology simulations described in the following were carried out with O(103)\mathcal{O}(10^3) particles, comprising an approximately equi-volume mixture of those with radius aa and 1.4a1.4a. Given the comparatively small number of particles (compared to a real experimental system, for instance) and the random nature of the Brownian forces added to the system, the stress signals output by a single simulation are extremely noisy, especially at low PePe. (The same is true for inertia-free simulations [34], though the error bars are rarely reported.) Thus the number of realisations that must be averaged over to obtain smooth data and reliable estimates of the true rheology increases as PePe is reduced.

Shown in Figure 3(a) is the range of measured ηr\eta_r as a function of the number of steady state snapshots averaged over, for 3 different PePe. At low PePe one must sample the system ≈108\approx10^8 times to obtain a measurement of ηr\eta_r with standard deviation less than 10%, whereas for large PePe 10510^5 samples are sufficient. Importantly, the time taken to reach steady state also differs drastically with PePe. For systems dominated by thermal fluctuation (i.e. low PePe) the

Computing the suspension viscosity under sheared and non-sheared conditions, and the scaling of computational run time with system size

Figure 3. Computing the suspension viscosity ηr\eta_r under sheared and non-sheared conditions, and the scaling of computational run time with system size. Shown in (a) is the convergence of the measured ηr\eta_r as a function of the number of snapshots averaged over, for Pe=0.01Pe=0.01 (red), 1 (green) and 100 (blue). The noisy stress signal when Brownian motion dominates necessitates large numbers of realisations. In (b) is ηB,GK\eta_{B,GK} measured via the Green-Kubo relation taking the autocorrelation of the Brownian shear stress as input, plotted as a function of the correlation time; (c) Simulation run time versus number of particles for ϕ=0.5\phi=0.5, Pe=1Pe=1, when running a serial compilation of LAMMPS on a single processor. We show data for a short simulation comprising 10710^7 timesteps.

approach to steady state is set by the passage of Brownian time as opposed to the accumulated strain, with systems at ϕ=0.54\phi= 0.54 and below taking 3–4 Brownian times to reach steady state at Pe=0.01Pe = 0.01. For larger ϕ\phi this timescale is stretched rapidly, likely due to the proximity of glassy physics. At very large PePe, meanwhile, steady states are reached for strains γ˙t\dot{\gamma}t of 1–2 [35].

Brownian stress at zero shear rate

To obtain the Brownian contribution to the viscosity in the limit of zero shear rate, we apply the Green-Kubo method [36] by calculating the time autocorrelation function of the shear stress, taking as input the Brownian stress computed as described above, for unsheared simulations. The Brownian viscosity is written as:

ηB,GK=VkBT∫0∞⟨ΣxyB(t+Δt)ΣxyB(t)⟩ dΔt,(29)\eta_{B,\mathrm{GK}} = \frac{V}{k_\mathrm{B}T}\int_{0}^{\infty}\langle\Sigma^{B}_{xy}(t+\Delta t)\Sigma^{B}_{xy}(t)\rangle\,d\Delta t, \tag*{(29)}

where ΣxyB\Sigma^{B}_{xy} is the shear component of the Brownian stress tensor. The stress correlation decreases exponentially with increasing Δt\Delta t so that the Brownian viscosity can be modelled as ηB,GK(t)=ηϕ(1−e−Δt/τϕ)\eta_{B,\mathrm{GK}}(t) = \eta_{\phi}(1-e^{-\Delta t/\tau_{\phi}}). As shown in Figure 3(b), the correlation time τϕ\tau_{\phi} is short and weakly varying for ϕ<0.5\phi< 0.5, so that ηϕ\eta_{\phi} can be measured using readily accessible data for which Δt/τϕ\Delta t/\tau_{\phi} is large. For ϕ>0.5\phi> 0.5, however, τϕ\tau_{\phi} grows quickly and we estimate ηϕ\eta_{\phi} by extrapolation. The rapid growth of the correlation time τϕ\tau_{\phi} is likely indicative of a nearby glass transition, though we defer detailed analysis of this behaviour to future work. By this approach we obtain an estimate of the Brownian contribution to the viscosity at zero shear rate, which we discuss further in the following.

Given the large quantity of data required for obtaining smooth results, it is worth considering the scaling of the simulation run time with the system size. To estimate the scaling of the run time we carried out simulations with N=101N = 10^1, 10210^2, 10310^3, 10410^4 particles with ϕ=0.5\phi= 0.5, running a serial build of LAMMPS on one core for 10710^7 timesteps. The result shown in Figure 3(c) confirms that our simulation has complexity O(N)O(N).

The role of inertia

The particle-particle contact timescale τC\tau_C is set sufficiently small that it does not compare to any other timescale in the system under any conditions, so that particles can always be considered to be hard. We verify this in Figure 5(b) by showing that the relationship between ηr\eta_r and PePe measured under different values of τC\tau_C does not vary. It is, however, crucial that in varying PePe one maintains acceptable values of τI/τS\tau_I/\tau_S and τI/τB\tau_I/\tau_B. To determine sufficiently small values of these two ratios so that inertia may be neglected, we carried out two sets of simulations. In the first we simulate shear flow with ϕ=0.55\phi= 0.55 and kbT=0k_bT = 0 (so we don’t need to consider the Brownian timescale τB\tau_B), while varying the dimensionless shear rate τI/τS\tau_I/\tau_S from 5×10−35 \times10^{-3} to 10. To be in the limit in which inertia is negligible, we require a linear relation between the shear stress and the shear rate i.e. a Stokes flow. In other words, we are correctly simulating an inertia-free flow if ηr\eta_r is independent of τI/τS\tau_I/\tau_S. From Figure 4(a) we can observe that this holds for τI/τS≲10−1\tau_I/\tau_S \lesssim10^{-1}. In what follows, we therefore ensure that this inequality holds for all parameter sets. Our result here is qualitatively consistent with prior simulations [17] and experiments [37, 38], though the value of the Stokes number at the crossover is apparently highly sensitive to system details.

Suspension viscosity versus timescale ratios for particle inertia

Figure 4. Establishing the parameter range in which particle inertia can be neglected. (a) Suspension viscosity ηr\eta_r (rescaled by the low τI/τS\tau_I/\tau_S value) plotted against the timescale ratio τI/τS=ργ˙a2/η\tau_I/\tau_S = \rho\dot{\gamma}a^2/\eta (the Stokes number) for kBT=0k_\mathrm{B}T = 0 (so that τB=∞\tau_B = \infty) and ϕ=0.55\phi= 0.55, showing that our rheology results are rate-independent and therefore inertia-free for τI/τS≲0.1\tau_I/\tau_S \lesssim0.1; (b) ηr\eta_r (rescaled by the low τI/τB\tau_I/\tau_B value) plotted against the timescale ratio τI/τB=ρkBT/ηa2\tau_I/\tau_B = \rho k_\mathrm{B}T/\eta a^2 for τI/τS<0.01\tau_I/\tau_S < 0.01 and ϕ=0.45\phi= 0.45. Shown are various values of PePe.

Rheology and microstructure of dense suspensions at the transition from Brownian to non-Brownian flow

Figure 5. Rheology and microstructure of dense suspensions at the transition from Brownian to non-Brownian flow. Shown in (a)-(c) are the suspension viscosity ηr\eta_r as functions of PePe, for frictionless particles with (a) ϕ=0.45\phi= 0.45; (b) ϕ=0.5\phi= 0.5 (shown also are results for two additional values of knk_n); (c) ϕ=0.55\phi= 0.55, showing the contributions from contacts, hydrodynamics and Brownian forces. In (d) is the total viscosity as a function of PePe and ϕ\phi. In (e)-(g) are slices through the three-dimensional radial distribution function g(ri,j)g(\mathbf{r}_{i,j}) showing the flow-gradient (xyxy) plane under steady state simple shearing conditions for ϕ=0.5\phi= 0.5 and (e) Pe=0.01Pe = 0.01; (f) Pe=1Pe = 1; (g) Pe=10000Pe = 10000.

In the second we simulate shear flow with τI/τS<0.01\tau_I/\tau_S < 0.01, ϕ=0.45\phi= 0.45 and at a range of PePe, exploring the relative importance of inertia by varying the timescale ratio τI/τB\tau_I/\tau_B. This control parameter essentially sets the distance a particle will typically cover under ballistic motion. In order for inertia to be negligible in the model, we expect that this distance should be at least an order of magnitude smaller than the particle size, so that a typical Brownian kick to a particle does not lead it to collide with a distant neighbour. From our result in Figure 2(b) we find that a ballistic to diffusive crossover occurs at ⟨x2⟩/a2=0.01\langle x^2\rangle/a^2 = 0.01 for τI/τB=O(10−2)\tau_I/\tau_B = O(10^{-2}). Our shear simulations (Figure 4(b)) similarly show that ηr\eta_r is a function of τI/τB\tau_I/\tau_B only when the latter quantity is >0.01> 0.01. Therefore, in what follows we carry out simulations with τI/τB<0.01\tau_I/\tau_B < 0.01 and τI/τS<0.1\tau_I/\tau_S < 0.1.

Flow curves

Our main rheology results are presented in Figure 5. We simulated a broad range of PePe (10−2−10410^{-2} - 10^4), focussing on three different volume fractions ϕ\phi (Figure 5(a)-(Figure 5c)) and adhering to the constraints on τI\tau_I obtained above. To achieve this range of PePe it was necessary to vary both the shear rate γ˙\dot{\gamma} and the thermal energy kbTk_bT. We present in Table I a full list of the parameters used to generate the result in Figure 5(a). In each rheology figure we break the overall viscosity down into its contributions from hydrodynamic, contact and Brownian stresses. The stresses obtained by taking the outer product of the pairwise vectors and forces evaluated during the simulation run are the hydrodynamic one, the contact one, and the ‘instantaneous’ Brownian stress. The latter (not shown in Figure 5), as described earlier, averages to zero so does not lead to a viscosity contribution. The total stress (shown in black in Figure 5(a)-(Figure 5c)) is therefore just the sum of the hydrodynamic and contact parts.

PePeγ˙\dot{\gamma}kbTk_bTaaρ\rhoknk_nη\eta
0.010.000090.01710.1100000.1
0.0460.000420.01710.1100000.1
0.220.00190.01710.1100000.1
10.0090.01710.1100000.1
4.60.0090.003710.1100000.1
220.0090.0007910.1100000.1
1000.0090.0001710.1100000.1
4600.0090.00003710.1100000.1
22000.0090.000007910.1100000.1
100000.0090.000001710.1100000.1

Table I. The parameters used to generate each data point in Figure 5(a).

As a post-processing step we make an estimate of the effective Brownian stress (approximating the one that would be measured in a Stokesian Dynamics simulation), following the calculation based on structural anisotropy described earlier. This gives us the red lines in Figure 5(a)-(Figure 5c). Interestingly the Brownian stress maps quite closely to the contact stress for low PePe, indicating that the surge in contact stress observed in this range may be due to short-lived contacts induced by the Brownian kicks. Indeed the formulation of the Brownian stress is similar to that of the

contact stress, differing only in the presence of the contact overlap δi,j\delta_{i,j} appearing in the latter.

Overall we find that the predicted rheology corresponds well with canonical results, both in the experimental literature [39, 40] and those obtained by Stokesian Dynamics simulation [41] and similar numerical methods [34]. At all volume fractions there is a shear thinning region for Pe<1Pe < 1 that gives way to shear thickening beyond Pe>1Pe > 1, with the ηr\eta_r values at large PePe tending towards those reported for non-Brownian suspensions under a very similar numerical framework [23]. For ϕ=0.45\phi= 0.45 and ϕ=0.5\phi= 0.5 we observe a low PePe plateau, whereas at ϕ=0.55\phi= 0.55, ηr\eta_r apparently continues to increase with decreasing PePe. The latter effect is perhaps an artefact of proximity to a glass transition, though we defer a more detailed study of this effect to future work due to the diverging timescales involved. The hydrodynamic stress increases weakly with increasing PePe, whereas the contact stress qualitatively follows the overall stress in its shape. The increase in contact viscosity at high PePe may be attributed to the onset of contact force chains as the system approaches the non-Brownian limit and can be considered granular [42], while at low PePe it is related to the Brownian forces as described above.

Shown in Figure 5(d)-(Figure 5f) are slices through the three dimensional radial distribution function g(ri,j)g(\mathbf{r}_{i,j}), showing the flow-gradient (xyxy) plane at Pe=0.01Pe = 0.01, Pe=1Pe = 1 and Pe=104Pe = 10^4. The general shape of the pairwise distributions is consistent with literature data [43], showing increased anisotropy with increasing PePe and a sharpening of the peaks at a+aa + a, a+1.4aa + 1.4a and 1.4a+1.4a1.4a + 1.4a.

Viscosity variation with volume fraction

In order to understand better the limiting behaviour at small PePe, we determine the behaviour of ηr\eta_{\mathrm{r}} as a function of ϕ\phi. To do so we first evaluate the Brownian contribution to the zero shear viscosity using the Green-Kubo method described above. To obtain an estimate of the full viscosity, we take the value of the hydrodynamic viscosity at the smallest (non-zero) measured PePe, and add this to the Green-Kubo prediction of the Brownian stress (assuming the latter to be a good proxy for the contact stress, as was assumed by Brady [29] and is supported by our simulation data in Figure 5). Doing so at a range of ϕ\phi, and comparing the result to the minimum ηr\eta_{\mathrm{r}} measured at Pe=4.6Pe = 4.6 for each ϕ\phi as well as the large PePe limit, we obtain Figure 6(a).

In both the low and high PePe limits, we find that ηr\eta_{\mathrm{r}}, particularly at large ϕ\phi, can be fit relatively well with a simple relation as ηr≈(1−ϕ/ϕJ)−λ\eta_{\mathrm{r}} \approx(1-\phi/\phi_J)^{-\lambda}, with ϕJ(Pe→0)=0.587\phi_J(Pe \to0) = 0.587 and ϕJ(Pe=104)=0.642\phi_J(Pe = 10^4) = 0.642 (and λ≈1.5\lambda\approx1.5, similar to Mari et al. [44]). At intermediate PePe, ηr\eta_{\mathrm{r}} is reduced relative to its value in the the non-Brownian limit, and the value of ϕJ\phi_J is marginally increased. The large PePe value of ϕJ\phi_J will be highly sensitive to details of the particle-particle contact interaction, especially the presence of a static friction coefficient as we have reported elsewhere [23, 45]. In particular, for large friction coefficients the large PePe value of ϕJ\phi_J (usually denoted ϕm\phi_m) will likely drop below the low PePe value. In this scenario one expects flow curves near jamming to be diverging at both low and high PePe, with finite ηr\eta_{\mathrm{r}} at intermediate PePe. We leave this complexity to be explored in future work, and in the following we examine the role of friction for a small range of ϕ\phi.

Role of particle-particle friction and short-ranged repulsion

In the context of experimental work by Guy et al. [6], it is important to consider the role of particle friction at the colloidal-to-granular transition. Since granular particles are large, micron size objects they will likely have a static friction coefficient, which may constitute both sliding and rolling components [45, 46]. So far we have only considered a model system of frictionless particles. It is well-established that the presence of static sliding friction means that each particle-particle contact will constrain more than one degree of freedom of each particle, so that for large friction coefficients (in practice μ≳0.5\mu\gtrsim0.5) a rigid packing can be obtained with a per particle contact number of ≈4\approx4 (as opposed to 6 for frictionless spheres), with limiting volume fraction ϕm≈0.57\phi_m \approx0.57. In Figure 6(b) we report rheology predictions from simulations of suspensions with a range of particle-particle friction coefficients μ\mu (black data), demonstrating that the presence of friction leads to a dramatic increase in ηr\eta_{\mathrm{r}} at large PePe. This behaviour, and its sensitivity to ϕ\phi demonstrated in Figure 6(c) (black data), is qualitatively consistent with the large literature on friction-driven shear thickening e.g. Mari et al. [44]. Notably, ηr\eta_{\mathrm{r}} at lower PePe is unaffected by friction, suggesting that Brownian forces suppress the mobilisation of static friction for all μ\mu, at least at ϕ=0.5\phi= 0.5. In this respect the Brownian forces act

analogously to a weak repulsive potential, inhibiting the formation of sustained particle contacts and rendering the suspension effectively frictionless even when μ>0\mu> 0. This leads to a bulk viscosity with rate dependence qualitatively similar to that of shear thickening suspensions with load-activated friction describable by the canonical model of Wyart and Cates [47].

Importantly, though, is it not clear that the shear thickening transition, when controlled by Brownian motion, is governed by a single stress scale. In particular, the range of PePe over which the transition happens in Figure 6(b) (black data) is rather broad (occurring over 4-5 orders of magnitude in PePe), especially when compared to Mari et al. [44] in which the transition takes at most 2 orders of magnitude in shear rate. To explore this we introduce a short ranged repulsive force defined by

Fi,jR=Aκexp⁡((ai+aj)−∣ri,j∣κ)ni,j,(30)F_{i,j}^{\mathrm{R}} = \frac{A}{\kappa} \exp\left(\frac{(a_i+a_j)-\lvert\mathbf{r}_{i,j}\rvert}{\kappa}\right)\mathbf{n}_{i,j}, \tag*{(30)}

with κ=0.01(ai+aj)\kappa= 0.01(a_i+a_j). We show results of this model for A/kbT=0,1,102,103A/k_bT = 0, 1, 10^2, 10^3 in Figure 6(b) and for several ϕ\phi at μ=0.5\mu= 0.5 in Figure 6(c). Introducing a sufficiently large repulsive force scale (in practice we required A/kbT≈100A/k_bT \approx100) narrows the range of PePe over which shear thickening occurs, and shifts the transition to larger PePe. This result suggests not only an additive effect of Brownian and repulsive forces as reported by Mari et al. [34], but rather a qualitative change in the functionality of ηr\eta_r with PePe when the onset of contacts is set by the magnitude of Brownian or repulsive forces. Examining the subtly in more detail is a promising area in which our model might be deployed. Thus with the introduction of particle-particle friction and a short ranged repulsive force we can control in our model the position and extent of shear thickening, providing a flexible starting point from which to make predictions of the rheology in more specific contexts.

Concluding remarks

In conclusion, we have implemented a minimal numerical model for the rheology of dense suspensions that incorporates sufficient microscopic physics to predict the colloidal to granular crossover as a function of PePe. The model is implemented in LAMMPS [21] so that its run time scales linearly with the number of particles. The Brownian component of our model differs from that in Stokesian Dynamics in that we resolve the fluctuations at a much shorter, inertial timescale. The naively-calculated Brownian stress therefore averages to zero over realisations and instead we compute an estimation of the Brownian contribution to the stress based on the structural statistics measured from the simulation. This stress follows closely the contact stress that we measure directly from the pairwise forces and relative positions. The model predicts shear thinning at low PePe, with a low PePe plateau (in some cases) that increases with volume fraction. At larger PePe a Brownian regime gives way to a contact dominated regime in which particle-particle interactions proliferate and friction (if present) becomes important. In this latter regime shear thickening is observed even for zero particle friction, though its extent increases with increasing friction coefficient. We finally introduced into our model a short range repulsive force, a crucial prerequisite for shear thickening in the paradigmatic model of non-Brownian suspensions [44]. This keeps particles separated and inhibits the contact contribution to the stress, thus broadening the intermediate PePe viscosity plateau (as observed by Cwalina and Wagner [48]) or equivalently shifting the value of PePe at which particle contacts become important.

We have focussed in this article on steady, simple shear rheology. Broadening the work to inhomogeneous conditions (such as those described by Gillissen and Ness [49]) and to dynamic simple shear to measure the frequency-dependent (and indeed amplitude dependent [50]) response are promising lines of future research that will provide additional scope for constitutive model development and for validation against experimental data.

In future we anticipate deploying our code in mixed systems, in which one population of particles are Brownian and another are non-Brownian [48]. This is motivated by numerous real world examples such as geophysical flows and many scenarios in chemical engineering and manufacturing. In such systems the small, Brownian particles (in some industries these are referred to as superplasticizers) will simultaneously contribute a Brownian stress but improve the efficiency of packing, so that their overall effect on the rheology is non-trivial and likely to be non-monotonic and PePe-dependent. Mapping out this complexity as functions of the small and large particle sizes and their relative numbers requires a tractable numerical model, and will likely rely on the implementation of more advanced neighbour listing algorithms such as those by Shire et al. [51]. Extending this further to systems with continuous, broad size distributions remains on open challenge [52].

Acknowledgements

Codes and scripts necessary to reproduce the results reported in this article are available on request. C.N. acknowledges support from the Royal Academy of Engineering under the Research Fellowship scheme and from the Leverhulme Trust under Research Project Grant RPG-2022-095. We thank John Brady, Jeff Morris, Aleksander Donev, Emanuela Del Gado, Abhay Goyal, Anthony Ge, Romain Mari and Ryohei Seto for useful discussions.

References

References

  1. [1]C. Ness, R. Seto, and R. Mari, Annual Review of Condensed Matter Physics 13, 97 (2022).
  2. [2]J. J. Stickel and R. L. Powell, Annual Review of Fluid Mechanics 37, 129 (2005).
  3. [3]A. R. Jambrak, Z. Herceg, D. Šubarić, J. Babić, M. Brnčić, S. R. Brnčić, T. Bosiljkov, D. Čvek, B. Tripalo, and J. Gelo, Carbohydrate Polymers 79, 91 (2010).
  4. [4]R. Roussel, A. Lemaître, R. J. Flatt, and P. Coussot, Cement and Concrete Research 40, 77 (2010).
  5. [5]R. Kostychnick, H. Matinpour, S. Pradeep, S. Haber, A. Sauret, E. Meiburg, T. Dunne, P. Arratia, and D. Jerolmack, Proceedings of the National Academy of Sciences 119, e2209109119 (2022).
  6. [6]B. Guy, M. Hermes, and W. C. Poon, Physical Review Letters 115, 088304 (2015).
  7. [7]J. J. J. Gillissen, C. Ness, J. D. Peterson, H. J. Wilson, and M. Cates, Journal of Rheology 64, 353 (2020).
  8. [8]R. Seto, R. Mari, J. F. Morris, and M. M. Denn, Physical Review Letters 111, 218301 (2013).
  9. [9]J. Richards, B. Guy, E. Blanco, M. Hermes, G. Poy, and W. Poon, Journal of Rheology 64, 405 (2020).
  10. [10]L. Lobry, E. Lemaire, F. Blanc, S. Gallier, and F. Peters, Journal of Fluid Mechanics 860, 682 (2019).
  11. [11]J. F. Brady and G. Bossis, Annual Review of Fluid Mechanics 20, 111 (1988).
  12. [12]A. J. Banchio and J. F. Brady, Journal of Chemical Physics 118, 10323 (2003).
  13. [13]D. L. Ermak and J. A. McCammon, Journal of Chemical Physics 69, 1352 (1978).
  14. [14]A. Sierou and J. F. Brady, Journal of Fluid Mechanics 448, 115 (2001).
  15. [15]A. M. Firoozabadi and J. W. Swan, Journal of Fluid Mechanics 878, 544 (2019).
  16. [16]P. A. Cundall and O. D. Strack, Geotechnique 29, 47 (1979).
  17. [17]M. Trulsson, B. Andreotti, and P. Claudin, Physical Review Letters 109, 118305 (2012).
  18. [18]C. Ness and J. Sun, Physical Review E 91, 012201 (2015).
  19. [19]E. J. Hinch, Journal of Fluid Mechanics 72, 499 (1975).
  20. [20]We note that in LAMMPS the skім argument of the neighbour command has units of [length].
  21. [21]S. Plimpton, Journal of Computational Physics 117, 1 (1995).
  22. [22]F. Boyer, É. Guazzelli, and O. Pouliquen, Physical Review Letters 107, 188301 (2011).
  23. [23]O. Cheal and C. Ness, Journal of Rheology 62, 501 (2018).
  24. [24]R. Ball and J. R. Melrose, Physica A: Statistical Mechanics and its Applications 247, 444 (1997).
  25. [25]R. V. More and A. M. Ardekani, Journal of Rheology 64, 67 (2020).
  26. [26]S. Kim and S. J. Karrila, Microhydrodynamics: principles and selected applications (Courier Corporation, 2013).
  27. [27]R. Radhakrishnan, https://doi.org/10.5281/zenodo.1137305 (2018).
  28. [28]G. Bossis and J. Brady, Journal of Chemical Physics 91, 1866 (1989).
  29. [29]J. F. Brady, Journal of Chemical Physics 99, 567 (1993).
  30. [30]N. Y. C. Lin, M. Bierbaum, P. Schall, J. P. Sethna, and I. Cohen, Nature Materials 15, 1172 (2016).
  31. [31]C. Ness, Computational Particle Mechanics , 1 (2023).
  32. [32]D. S. Lemons and A. Gythiel, American Journal of Physics 65, 1079 (1997).
  33. [33]A. P. Hammond and E. I. Corwin, Physical Review E 96, 042606 (2017).
  34. [34]R. Mari, R. Seto, J. F. Morris, and M. M. Denn, Proceedings of the National Academy of Sciences 112, 15326 (2015).
  35. [35]C. Ness and J. Sun, Physical Review E 93, 012604 (2016).
  36. [36]J.-P. Hansen and I. R. McDonald, Theory of simple liquids: with applications to soft matter (Academic press, 2013).
  37. [37]F. Tapia, M. Ichihara, O. Pouliquen, and É. Guazzelli, Physical Review Letters 129, 078001 (2022).
  38. [38]Y. Madraki, A. Oakley, A. Nguyen Le, A. Colin, G. Ovarlez, and S. Hormozi, Journal of Rheology 64, 227 (2020).
  39. [39]C. G. de Kruif, E. M. F. van Iersel, A. Vrij, and W. B. Russel, Journal of Chemical Physics 83, 4717 (1985).
  40. [40]H. M. Laun, Die Angewandte Makromolekulare Chemie: Applied Macromolecular Chemistry and Physics 123, 335 (1984).
  41. [41]D. R. Foss and J. F. Brady, Journal of Fluid Mechanics 407, 167 (2000).
  42. [42]N. Y. C. Lin, B. M. Guy, M. Hermes, C. Ness, J. Sun, W. C. Poon, and I. Cohen, Physical Review Letters 115, 228304 (2015).
  43. [43]D. R. Foss and J. F. Brady, Journal of Rheology 44, 629 (2000).
  44. [44]R. Mari, R. Seto, J. F. Morris, and M. M. Denn, Journal of Rheology 58, 1693 (2014).
  45. [45]A. Singh, C. Ness, R. Seto, J. J. de Pablo, and H. M. Jaeger, Physical Review Letters 124, 248005 (2020).
  46. [46]M. Blair and C. Ness, Journal of Fluid Mechanics 948, A48 (2022).
  47. [47]M. Wyart and M. E. Cates, Physical Review Letters 112, 098302 (2014).
  48. [48]C. D. Cwalina and N. J. Wagner, Journal of Rheology 60, 47 (2016).
  49. [49]J. J. J. Gillissen and C. Ness, Physical Review Letters 125, 184503 (2020).
  50. [50]C. Ness, Z. Xing, and E. Eiser, Soft Matter 13, 3664 (2017).
  51. [51]T. Shire, K. J. Hanley, and K. Stratford, Computational Particle Mechanics 8, 653 (2021).
  52. [52]P. M. Mwasame, N. J. Wagner, and A. N. Beris, Journal of Rheology 60, 225 (2016).

Paper details

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

views · 8

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents