PHYSICAL REVIEW FLUIDS 5, 034307 (2020)

Interaction network analysis in shear thickening suspensions

Marcio Gameiro,1,2,* Abhinendra Singh,3,4,5,† Lou Kondic ,6,‡
Konstantin Mischaikow ,2,§ and Jeffrey F. Morris 3,7,
1
Instituto de Ciências Matemáticas e de Computação, Universidade de São Paulo,
Caixa Postal 668, 13560-970, São Carlos, SP, Brazil
2
Department of Mathematics, Rutgers, The State University of New Jersey,
Piscataway, New Jersey 08854, USA
3
Benjamin Levich Institute, CUNY City College of New York, New York, New York 10031, USA
4
Pritzker School of Molecular Engineering, University of Chicago, Chicago, Illinois 60637, USA
5
James Franck Institute, University of Chicago, Chicago, Illinois 60637, USA
6
Department of Mathematical Sciences, New Jersey Institute of Technology, Newark, New Jersey 07102, USA
7
Department of Chemical Engineering, CUNY City College of New York, New York, New York 10031, USA

(Received 13 March 2019; accepted 18 February 2020; published 23 March 2020)

Dense frictional particulate suspensions in a viscous liquid undergo increasingly strong

continuous shear thickening as the solid packing fraction, φ, increases above a critical
volume fraction, and discontinuous shear thickening is observed for even higher packing
fractions. Recent studies have related shear thickening to a transition from mostly lubri-
cated to predominantly frictional contacts with the increase in stress, with the transition
determined by overcoming a repulsive force. The rheology and networks of frictional
forces from two- and three-dimensional simulations of shear-thickening suspensions are
studied. These are analyzed using measures of the topology of the network, including tools
of persistent homology. We observe that at low stress, the frictional interaction networks are
predominantly quasilinear along the compression axis. With an increase in stress, the force
networks become more isotropic, forming loops in addition to chainlike structures. The
topological measures of Betti numbers and total persistence provide a compact means of
describing the mean properties of the frictional force networks, and provide a link between
macroscopic rheology and the microscopic interactions. A total persistence measure de-
scribing the significance of loops in the force network structure, as a function of stress and
packing fraction, shows behavior similar to that of relative viscosity, and displays a scaling
law near the jamming fraction for both two- and three-dimensional systems considered.
The total persistence measures for both dimensions are found to be very similar.

DOI: 10.1103/PhysRevFluids.5.034307

I. INTRODUCTION
In suspensions of solid particles in viscous liquid, an increase in viscosity with increasing shear
rate or imposed stress is well-known, and is called shear thickening. If the particles are highly
concentrated (or “dense”), such that the conditions approach the maximum flowable solid fraction

*
[email protected]
†
[email protected]
‡
Corresponding author: [email protected]
§
[email protected]

[email protected]

2469-990X/2020/5(3)/034307(21) 034307-1 ©2020 American Physical Society

 MARCIO GAMEIRO et al.

(φA or φ in two or three dimensions, respectively), a very sharp increase in viscosity over a narrow
range of shear rate may occur. This is termed “discontinuous shear thickening.”
Recent simulation studies [1–5] have employed a fluid-mechanics-based approach as a model
leading to strong shear thickening. Within this approach, a breakdown of lubrication films between
particles is responsible for the shear thickening behavior; phenomenological theories [6–8] contain-
ing the same basic elements are successful in predicting the flow behavior. It should be noted that
alternative models not based on lubrication forces have been considered as well [9,10].
Recent studies describing a stress-induced frictional rheology found a relationship between
strong shear thickening and shear jamming, in the sense that thickening may be a precursor to
jamming: an increase of either stress or solid fraction from a shear thickening condition may lead
to shear jamming. To be specific, shear jamming is the phenomenon of generating a solid structure
under shear, but with this solid being fragile in the sense that it loses its ability to bear a shearing
load (stress) without deformation if the load direction is changed sufficiently [11]. This loss of
solidity is a temporary material failure, as rearrangement of the structure, and in particular of the
contact network, should once again result in jamming in the new direction of motion [12,13].
Furthermore, one may wonder whether it is the contact network that is relevant to the process of
shear thickening and shear-induced jamming, or whether one needs to consider not only contacts,
but also their strength. If the strength of the contacts is considered as well, the concept of a
weighted interaction network is introduced. Examples include weak or strong interaction networks
considered commonly when analyzing dry granular matter, with so-called “force chains” being the
most commonly considered structures.
In shear-thickening suspensions, large changes in both the viscosity and the normal stress
response have been associated with the formation of a contact network driven by increasing shear
stress. This prompts us to focus on the connectivity properties of this network, and its relation to
stress transmission. Essential questions relate to how the connectivity arises, how it depends on
the interaction force strength, what are the geometrical and topological properties of the connected
structures, and how these structures evolve with both the solid fraction and the imposed stress. All
of these issues must be examined for a specific model of microscopic physics.
We apply a discrete-particle simulation model that, using a minimal set of ingredients, has been
found to closely approximate behavior observed in experiments [1–4]. This model assumes that
particle surface separation is maintained by a repulsive force, and is thus lubricated by the liquid
film that remains, until the applied stress overwhelms this repulsive force [1,2]. In practice, many
suspensions have steric or electrostatic forces, repulsive in either case, between particles to stabilize
against surface contact, reducing aggregation or contact interactions. This leads to better flowability,
i.e., lower viscosity and less of a tendency to jam. Such a repulsive force, of magnitude F0 , gives
rise to a stress scale σ0 = F0 /6π a2 for particles of radius a, and within our simulation model or
the theory of Wyart and Cates [7], this stress level marks the crossover from lubricated to contact
frictional interactions between particle surfaces. At low stress, σ  σ0 , particle interactions are
lubricated (and hence frictionless) and the viscosity is relatively low, diverging at random close
packing (RCP) at a solid fraction that we denote by φJ0 , with the subscript J indicating jamming.
On the other hand, at high stress, σ  σ0 , most close interactions are frictional contacts, and
the viscosity is both much larger and diverges at a solid fraction φJμ < φJ0 ; φJ0 depends on the
interparticle friction [2]. With an increase in stress, the transition between these two states results in
shear thickening. Based on the value of packing fraction φ relative to the frictional jamming point
φJμ , different forms of shear thickening can be observed. For φ  φJμ , the relation between shear
rate γ̇ and shear stress σ is monotonic, leading to continuous shear thickening (CST), while for
larger φ more abrupt discontinuous shear thickening (DST) may occur.
To gain better insight into the lubricated-to-frictional rheology transition, we apply tools of
topology, with a focus on the network theoretical methodology of persistent homology, which has
previously been applied to the closely related question of jamming in dry granular systems [14–16].
Alternate approaches [17–19] have been applied to elucidate the role of strong and weak networks in
dry granular systems under shear. Persistent homology provides precisely defined and quantitative

034307-2
 INTERACTION NETWORK ANALYSIS IN SHEAR …

measures of the global interaction network, which is built on top of the contact network, as discussed
in Sec. II B. In the present context, we use these tools to characterize flowing steady states, and
thus the statistical properties of the interaction networks formed under different conditions in shear
thickening-suspensions.
We point out two features of the analysis of interaction networks based on persistent homology
that distinguish it from alternative approaches. First, this type of analysis naturally includes
information about the forces between particles, and allows for quantification and comparison of
the interaction networks, both between different states of a system and between different systems.
Second, this approach applies to both two- and three-dimensional (2D and 3D) systems. This
is convenient, since the 2D case allows for development of intuition and visual demonstration
of the tools applied here, while 3D results are in accord with physical experiment. The ability
of the persistent homology tools to be applied without a change of definition to 3D contact
networks is an advantage for this method over complementary approaches to the exploration of
networks; see [20] for a review. We develop this approach here, demonstrating that measures of the
structure determined by persistent homology show sharp changes where the rheology of the material
undergoes abrupt shear thickening.

II. BACKGROUND AND METHODS

A. Suspension flow simulation
We simulate simple-shear flow of an assembly of non-Brownian frictional spheres (a monolayer
in the two-dimensional case) immersed in a Newtonian fluid under an imposed stress using Lees-
Edwards periodic boundary conditions [21]. To avoid ordering, bidisperse particles with radii a and
1.4a are mixed at equal volume fractions. The particles interact through near-field forces of fluid
mechanical origin (i.e., lubrication forces) and frictional contacts. This simulation scheme has been
shown to reproduce many features seen experimentally in shear thickening of dense suspensions
[1,2]. We consider the motion to be inertialess (at zero Reynolds number) so that the equation of
motion reduces to force balance between lubrication (FH ) and contact (FC ) forces on each particle,
0 = FH (r, u) + FC (r), (1)
where u and r are written here as the many-body position and velocity vectors (u ≡ ṙ). A similar
torque balance applies. To allow contact, the lubrication resistance singularity is cut off at a surface
separation h = 10−3 a. The contact interactions are modeled using the approach of Cundall and
Strack [22]. The tangential force between two particles is limited according to the Coulomb friction
law to be no larger than the friction coefficient μ times the normal force, i.e., |FtC |  μ|FCn |, for
compressive normal forces.
To incorporate rate dependence, we employ a critical load model (CLM), in which an interparticle
normal force threshold F0 must be applied to activate friction between particles [1,2,5]. In this study,
we use an interparticle friction coefficient of μ = 1.0. The apparent viscosity of the suspension is
defined as η = σ /γ̇ , where γ̇ is the shear rate of the simple-shear flow, i.e., dux /dy. The viscosity is
presented through its form relative to the fluid viscosity ηr = η/η0 , where η0 is the pure fluid value.
A large increase in ηr with increasing shear rate is found, and this is the shear thickening of interest.
The force F0 sets a stress scale, σ0 = F0 /6π a2 , for the onset of shear thickening. In the remainder
of this work, we report quantities in terms of σ̃ = σ /σ0 and the scaled strain rate γ̇ /γ̇0 with
γ̇0 ≡ F0 /6π η0 a2 .
We normalize the force by σ0 a2 . We use N = 2000 particles for 2D simulations and N = 500
particles for 3D simulations in a unit cell; to test the size dependence, N = 1000 and 2000 were
considered in 3D for a few cases.
From the simulations, we determine particle positions, normal and tangential contact forces,
and noncontact lubrication forces. We perform simulation at a given condition for 20 strain units
and record the data at strain steps of 0.01. Figures 1(a) and 1(b) show snapshots from simulation
with σ̃ = 1 for packing fractions φA = 0.78 and φ = 0.56 in 2D and 3D, respectively. In these

034307-3
 MARCIO GAMEIRO et al.

FIG. 1. (a),(b) Snapshots of the force network at σ̃ = 1 for (a) φA = 0.78 in 2D and (b) φ = 0.56 in 3D.
Line thickness is proportional to the normalized force F with red, green, and blue lines showing frictional
contact, frictionless contact, and lubrication interactions, respectively. (c),(d) Relative viscosity, ηr , as a
function of dimensionless shear rate γ̇ /γ̇0 in stress-controlled simulations in (c) 2D and (d) 3D for several
packing fractions.

images, we show the frictional (red solid line) and frictionless contacts (dashed green lines), and
lubrication interactions (blue dashed lines). The thickness of each frictional segment is proportional
to the normalized force level F acting at the contact. As we will see, the development of the contact
network with increasing stress is responsible for significant changes in the rheological response of
the suspension.
If only the forces of fluid mechanical origin are relevant, the suspension would be expected to
exhibit a rate- or stress-independent relative viscosity, with close interactions lubricated. Similarly,
lubrication interactions together with frictional contacts, but without the repulsive force, result in
a rate-independent rheology. The rheological rate dependence thus arises due to the shear force
overwhelming the repulsive forces, modeled here by the CLM approach [2], to cause a change in
the dominant stress generation mechanism. In our simulations, we use a controlled shear stress
protocol [3] in which the shear rate γ̇ fluctuates. The relative viscosity is given by ηr = σ̃ /γ̇ /γ̇0 ,
where angular brackets imply time average.
To present the results of topology-based analysis in a visually transparent manner, we consider
2D before turning to 3D simulations, for which comparisons of the rheology to experimental work
have been made [2,4]. Figures 1(c) and 1(d) show ηr as a function of shear rate for four packing
fractions in 2D and 3D, respectively. For the lowest packing fractions, namely φA  0.76 in 2D
and φ  0.54 in 3D, ηr increases continuously between two rate-independent values: this is CST.
At packing fractions of φA = 0.78 in 2D and φ = 0.56 in 3D, the viscosity increases at almost

034307-4
 INTERACTION NETWORK ANALYSIS IN SHEAR …

FIG. 2. Flow-state diagram in the shear stress vs 2D packing fraction (σ̃ , φA ) plane for μ = 1; for the 3D
version, see [8]. The red dashed line shows the locus of points for DST (where ∂ γ̇ /∂σ = 0) and the leftmost
point on this curve is the minimum packing fraction, φA,C , at which DST is observed. The blue solid curve
shows packing-fraction-dependent maximum flowable stress above which the suspension is shear-jammed.
μ
Dashed and dotted-dashed black lines represent frictional (φA,J = 0.795) and frictionless (φA,J
0
= 0.855)
jamming points in 2D, respectively.

constant shear rate over a brief interval, indicating that these are approximately the packing fractions
for onset of DST. For higher packing fractions, φA = 0.79 and φ = 0.57, the viscosity function
becomes sigmoidal under imposed stress, indicating that the suspension is fully in the DST state
[8]. Note that if the simulations at the larger fractions are performed at a fixed rate, ηr (γ̇ /γ̇0 ) is
observed to change discontinuously rather than following the S-shaped curve [3,7,8].
Figure 2 shows a flow-state diagram, based on simulation results and a recently proposed model
[8]; this particular diagram is for 2D, with the 3D version in the noted reference showing the same
features. As discussed above, for the lower packing fractions φA < φA,C , CST is observed. For
μ
packing fractions φA,C  φA < φA,J , DST is observed between two flowing states. For this range
μ
of φA , a curve shows the locus of DST points, i.e., d γ̇ /dσ = 0. For φA > φA,J , DST is observed
between a flowing state (at low stress) and a shear-jammed state and is termed DST-SJ. The stress
required to observe DST as well as shear jamming decreases with an increase in packing fraction.
Both stresses vanish upon the approach to the frictionless jamming point. We do not consider the
DST-SJ state further in this work, as the growth of the contact network with increasing stress is
conceptually similar to the “pure” DST regime, while the subtleties of the network structure at
jamming would expand beyond our desired scope. The goal here is to demonstrate the utility of
topological data analysis as a tool for describing the underlying features of the interaction networks
resulting in the rheology of shear-thickening suspensions.
Recent studies have linked shear thickening with a growing number of frictional contacts as
shear stress increases [1,2,5,7]. Only limited work has explored the network properties of sheared
suspensions [23,24]. In the following, we focus on the evolution of the network structure formed by
the frictional forces. These are readily visualized in 2D, and Fig. 3 shows how the contact interaction
network builds in 2D for stresses varying through the DST transition at φA = 0.79. At low stress
σ̃ = 0.1, frictional contacts are absent, so that the suspension is in the low-viscosity frictionless
state. For σ̃ = 0.5 and 1.0, the suspension begins to shear-thicken, and frictional forces appear as
roughly linear structures (force chains) along the compression axis, i.e., along y = −x of the simple
shear flow ux = γ̇ y. With an increase in stress to σ̃  2.0, the suspension enters the DST regime,
and frictional interaction networks begin to form loops. With a further increase in stress, the number
of loops increases and saturates for σ̃ > 10. Combined with the rheological curves in Fig. 1(c), this

034307-5
 MARCIO GAMEIRO et al.
034307-6

FIG. 3. Interaction network of all frictional contacts (normal force above the critical value, F0 ) corresponding to a snapshot in steady state obtained in simulations
with σ̃ = (a) 0.1, (b) 0.5, (c) 1.0, (d) 2.0, (e) 5.0, (f) 10, (g) 50, and (h) 100 in 2D. Line thickness is proportional to the normalized force, F .
 INTERACTION NETWORK ANALYSIS IN SHEAR …

illustrates the correlation between strong shear thickening and frictional interaction networks. Note
that all structures described are transient, continuously forming, flowing, and breaking as the result
of the bulk shearing motion.

B. Overview of topological measures: From Betti numbers to persistent homology

To introduce the relevant concepts, consider Fig. 4, which shows the interaction networks as a
function of the interaction level, an interaction network filtration or simply a filtration; here the
interparticle force is scaled by σ0 a2 . This figure illustrates the portions of the contact network
experiencing force larger than the associated threshold or filtration level. At a very large force, only
the most highly loaded contacts and the particles they connect are shown, and these are isolated
objects. As the threshold is reduced, these components “grow” and eventually merge with others.
The information contained in such figures for a set of filtration levels will be used in what follows
to formulate the relevant topological measures. Note that there is a visual similarity between the
filtration at decreasing force threshold in Fig. 4 with the contact networks formed at increasing
stress level in Fig. 3: although structural change takes place with increasing stress, the implication
of this similarity is that the types of structures that are able to achieve the necessary force to generate
a contact at low stress generate contacts that have the largest forces at elevated stresses where a fully
ramified network forms.
To gain better insight into the structures that are formed, we consider Betti numbers, which
specify the number of structures of a particular type. The number of connected structures of any
sort, from a contacting pair to an extended branched chain, is given by the zeroth Betti number, β0 ,
while the number of structures that form a closed loop is given by the first Betti number, β1 . The
structures formed by connected particles enclosing a volume is given by the second Betti number,
β2 , which is relevant in 3D, but not in 2D. We find no such structures in our 3D simulations (i.e.,
β2 = 0), and thus we only consider β0 and β1 in the following.
The manner of defining the structures is associated with a decreasing force threshold. As the force
threshold F decreases, “birth” of a connected structure of a given Betti number is associated with
the force threshold in a filtration for which the structure appears, i.e., it is the largest force within the
structure. As the threshold is further decreased, the structure may grow in size as additional contacts
are added. Consider a β0 structure, formed at F1 , and a second that formed at F2 < F1 ; if these
two structures merge, the resulting union results in the “death” of one structure, and by definition
it is the younger structure, which formed “later” in the filtration (at lower force in a sequence of
progressively reduced force threshold), that is lost, i.e., the structure with birth at F2 dies upon the
merging. The persistence of structures through the reduction of force threshold gives rise to the term
“persistent homology.”
Figure 5 shows the Betti numbers for the configuration and network filtration of Fig. 4; the
red points in Fig. 5 correspond to the threshold levels shown in Fig. 4. We see from the β0 plot
that the maximum contact force level is slightly below F = 6. As the force threshold is decreased,
β0 increases monotonically until F ≈ 2.5. Thus we see that connected structures are formed by
contacts carrying different levels of force. The fact that there is a rapid increase in β0 until F ≈ 2.2
indicates that there are many such components, with considerable variation in the force level.
Figure 5(b) shows that loops are absent until F ≈ 1.5 and appear in significant numbers only
for F < 1, which is well below the value at which the number of components starts decreasing
rapidly. This implies that there is significant force heterogeneity, even on the level of single loops.
To understand why, consider the formation of a loop at a given force level. Typically the loop forms
as a single edge is added to a connected structure consisting of edges with higher forces. From
Figs. 5(a) and 5(b) we infer that most loops form after many components of varying force intensity
have merged. Thus, these loops are made up of components with a wide range of force levels.
Furthermore, the fact that β1 grows to much larger values than β0 shows that the larger connected
structures must contain numerous loops. This may be thought of as similar to the meshwork structure
of a net, although it lacks the regularity of such a structure as shown by Fig. 4; a single connected

034307-7
 MARCIO GAMEIRO et al.
034307-8

FIG. 4. Filtration of the interaction network of all frictional contacts for φ = φA = 0.76 and σ̃ = 10 corresponding to a snapshot in steady state for several values
of the threshold, in order from left to right and top to bottom: F = (a) 4, (b) 3, (c) 2, (d) 1.6, (e) 1.3, (f) 1, (g) 0.5, and (h) 0. Here, black dots show particle centers,
and the thickness of the red line corresponds to the strength of the normal force at the contact.
 INTERACTION NETWORK ANALYSIS IN SHEAR …

FIG. 5. Betti numbers corresponding to the filtration in Fig. 4 for components (a) and loops (b). The (red)
dots are at the force threshold levels corresponding to the values illustrated in Fig. 4. Note the difference in the
axes range for (a) and (b).

piece of the mesh may contain from no loops to many loops, and the number and size of the loops
is clearly mechanically relevant. Furthermore, the fact that the ratio β1 /β0 decreases significantly as
the force threshold is increased beyond F = 1 clearly illustrates a transition from structured network
(with many loops) to more “chainy” structure, as visualized also in Fig. 4. Further discussion of
the insights that could be reached by considering Betti numbers for different threshold values and
different stress values can be found in Sec. IV C.
While Betti numbers provide useful information regarding the number of components and loops,
their utility is somewhat limited by the fact that they depend on the chosen force threshold, and
therefore one needs to consider a large number of force threshold values to reach a reasonably
complete understanding of the system. In addition, Betti numbers provide only the number of

034307-9
 MARCIO GAMEIRO et al.

FIG. 6. Persistence diagrams corresponding to the filtration in Fig. 4 for components (a) and loops (b).

components and loops; they do not provide information about how these components and loops are
connected. In principle, the system could even change its topological properties without a change
of Betti numbers, e.g., a component could be born at essentially the same force level at which
another component dies, and therefore the topology would change but β0 would remain the same.
Persistence diagrams allow us to deal with these concerns.
Figure 6 shows an example of such diagrams for β0 and β1 . Here, for each structure, a point is
placed at the (birth, death) coordinates (b, d ). If a structure is still present at the minimum force
threshold of the filtration, it is plotted (in red) with the coordinate pair (b, −0.1) to illustrate that
it was born at the threshold level b, and that it never dies. The (arbitrary) choice of −0.1 takes the
corresponding points outside of the physically allowable values in order to impart the information
that these structures never die. In Fig. 6(a) the red point with b ≈ 6 shows the component that formed

034307-10
 INTERACTION NETWORK ANALYSIS IN SHEAR …

first (at the largest threshold value), and this component by definition never dies. The red dots with
b ≈ 0 are specific to the system considered, and they show the existence of separate components that
form at very low threshold values. Such structures are possible due to fluid mechanical forces; in dry
granular systems where only contact forces are present, such structures cannot form for a system
in force balance, since they would imply unbalanced forces on the participating particles. We also
note that our network filtrations [Fig. 4(b)] are formed by the particles and the edges representing
the force interaction between them, and hence once a loop forms it never dies (loops cannot merge).
The approach that we use here is slightly different from the one implemented elsewhere [14–16],
where “trivial” loops (formed by three adjacent particles) were not considered. In the present work,
we find that including trivial loops influences the results only marginally, so we keep them in our
calculations for simplicity.
Persistence diagrams provide a significant data reduction, expressing properties of complex
weighted networks by a collection of points. However, these diagrams are essentially point clouds,
and one needs to decide how to extract relevant information. While different approaches have
been considered in earlier works [14–16,25,26], here we will focus on a single measure, the total
persistence, which essentially quantifies how “mountainous” is the interaction network, with larger
distance from peaks (where a structure is born) to a valley (where structures merge and one dies)
contributing more to the total persistence measure. The total persistence τ , = 0 for components
and = 1 for loops, is defined as the sum of the distances of all points from the diagonal of a
persistence diagram. More precisely,

τ = |d − b|,
(b,d )∈PD

where we use the value d = 0 for the points in the persistence diagram PD that never disappear,
shown as (b, −0.1) pairs in Fig. 6. In superficial terms, one could think of τ as a measure of the
structure of the interaction network; larger total persistence means more structure. Conveniently,
choosing τ as a measure reduces this complex structure to a pair of scalars. Clearly, considering
total persistence instead of persistence diagrams leads to additional data reduction and loss of
information. However, we will show that consideration of total persistence provides insight into
the connection of topological properties of the interaction networks and rheology of the suspension.
After discussing total persistence in what follows, we will return to persistence diagrams and Betti
numbers to discuss how these relate to properties of interaction networks as the applied stress is
varied, in both 2D and 3D.

III. TOTAL PERSISTENCE AS AN ORDER PARAMETER

Figure 7 shows the total persistence for 2D simulations at four packing fractions as a function
of imposed stress. The data are obtained by averaging over 200 samples of the interaction network
in the steady state at each (φA , σ̃ ) condition; we show the mean and the standard deviation. The
total persistence for the connected structures (τ0 ) rises strongly in the transition region, from the
low-viscosity state at σ̃ < 1 to the fully thickened one at σ̃ > 10. The transition is even more distinct
for the case of the loop (τ1 ) structures, as the total persistence is essentially zero for σ̃ < 1 and rises
rapidly for σ̃ > 1, again saturating at σ̃ ≈ 10; this saturated total persistence for β1 is considerably
larger for higher φA , with the value for φA = 0.79 roughly double that at φA = 0.76. The clear
message is that the loop structures are more prevalent for the larger φA , and once formed many of
these structures persist for all lower force thresholds: this is as expected on physical grounds, as
the maximum force regions can be separated from one another, but stress continuity in the material
requires that they be connected by lower force contacts.
The increase of the total persistence τ with both σ̃ and φA is directly correlated with the increase
in the relative viscosity (and as shown elsewhere [2], there is a coincident increase in the normal
stresses). The correlation between τ and ηr makes the total persistence useful as an order parameter
that is indicative of a transition in the state of the material. This transition is manifested through the

034307-11
 MARCIO GAMEIRO et al.

FIG. 7. Total persistence for 2D data as a function of σ̃ for components (a) and loops (b), for several values
of packing fraction φA .

growth of total persistence as a function of imposed stress for an interaction network of frictional
contacts that form more abruptly and become stronger at larger φA .
This observation is consistent with the picture emerging from the interaction networks displayed
as a function of increasing stress in Fig. 3, or as a function of decreasing force threshold in the
filtration illustrated by Fig. 4. While the 2D geometry allows ready visual identification of the
network structure, the main interest is ultimately in 3D simulations. It is particularly valuable to
have quantitative measures in 3D, where visualization is more difficult. The simulation approach
used in this study has been shown to closely replicate experimental behavior [4]. The difficulty of
identifying the networks in 3D flows is illustrated in Fig. 8 by snapshots of the contacting particles
(shown as dots) and the forces (shown as connecting lines whose thicknesses indicate the magnitude
of the force). The configurations are from simulations at φ = 0.56 at varying imposed stress from
σ̃ = 0.5 to 10, taken after sufficient strain, of O(1), has been imposed to allow development of the
structure. Unlike the equivalent network structures for 2D shown in Fig. 3, here we do not show the
particles. Regardless of whether they are shown in a transparent form, the particles obscure the view,
which even without the particles is clearly less enlightening than the equivalent 2D visualization.
We thus turn to the persistent homology measures obtained from our 3D simulations, considering
both the similarities with and differences from 2D. Figure 9 shows the total persistence for several φ.
As stress increases, there is a rapid rise in τ0 between σ̃ = 0.5 and 1, with near-saturation at σ̃ = 1,
while the loop structure measured by τ1 begins to increase at σ̃ ≈ 1 and grows up to σ̃ = 100,

034307-12
 INTERACTION NETWORK ANALYSIS IN SHEAR …

FIG. 8. Interaction networks for the 3D data for φ = 0.56, for σ̃ = 0.5 (a), σ̃ = 1 (b), σ̃ = 2 (c), and
σ̃ = 10 (d), each showing a single instant from the simulation once the structure has developed.
with a slowing of the growth apparent above σ̃ = 10. We thus find similarity in the measure of the
interaction network structure, represented by the total persistence, and the viscosity transition seen
in Fig. 1. In 3D, the variation of τ up to σ̃ = 100 indicates that the interaction network structure
continues to develop with increasing σ̃ well above the initial transition to the higher viscosity region.
To this point, all results we have presented for 3D were obtained from simulations with N = 500.
One obvious question is how the results (for the total persistence in particular) vary with N. To
answer this question, we have carried out additional 3D simulations with different numbers of
particles. Figure 10 displays τ0 and τ1 for 3D simulations at φ = 0.56 scaled by the value of N
used for that case. We observe the data to collapse within the error bars; however, there does
seem to be a weak trend of τ0 decreasing as N increases, and the opposite trend for τ1 . This
trend is explained by the boundary effects. The simulations are periodic in all directions, but this
periodicity is not considered when computing the persistence diagrams. For this reason we miss the
“across boundary” particle interactions. This has the effect of considering some components that
are supposed to be connected “across the boundary” as two separated components, and hence it
produces additional components (points in the diagrams in dimension 0). Additionally, some loops
that are formed by connections “across the boundary” are not counted as loops, and hence we miss
some loops (points in the diagrams in dimension 1). This effect is more noticeable for smaller
systems, where the ratio of boundary area to volume increases.
Figure 11 displays correlation in both 2D and 3D simulations between relative viscosity
and total scaled persistence corresponding to loops, τ1 /N. This figure shows that the viscosity
increases with τ1 /N, and appears to diverge at some maximum persistence. The dashed line shows

034307-13
 MARCIO GAMEIRO et al.

FIG. 9. Total persistence for 3D data as a function of σ̃ for components (a) and loops (b), for several values
of packing fraction φ.

ηr ∝ [(τ1J − τ1 )/N]−α , where τ1J /N = τ1 (φ = 0.57, σ̃ = 100)/N = 0.34 (for 3D), τ1 (φA = 0.79,
σ̃ = 100)/N = 0.42 (for 2D), and α = −1.8. The higher–φ data, in which DST is seen, follow a
single scaling law, while for packing fraction φ = 0.5 and φA = 0.72, the data do not agree with the
proposed correlation. The structural implication is that the thickened state of DST is associated with
a topology of the frictional interaction network similar to that in the jammed state. The observed
correlation suggests that the viscosity is solely given by the microstructure, described by the total
persistence. The persistence intrinsically is governed by the volume fraction (or distance from the
jamming point) and the applied stress. It is important to highlight that a given combination of
(φ, σ̃ ) that has the same τ1 at the mesoscale results in the same macroscopic behavior, i.e., relative
viscosity.

IV. STRUCTURE OF INTERACTION NETWORKS

Here we consider insight into the structure of the interaction networks that can be gained from
consideration of persistence diagrams and Betti numbers.

A. Interaction structure variation with stress

We address the question of how the structure of the interaction network changes as stress
increases at fixed φ (or φA ), and to answer this we turn to the persistence diagrams. Total persistence,

034307-14
 INTERACTION NETWORK ANALYSIS IN SHEAR …

FIG. 10. Scaled total persistence for 3D data for φ = 0.56 as a function of σ̃ : (a) τ0 and (b) τ1 .

τ , compresses all available information into a pair of numbers, and thus we seek insight from
the additional information in the persistence diagrams from which these numbers are determined.
Figure 12 shows the superposition of all β0 persistence diagrams at each of seven values of imposed
stress for the 2D simulations at φA = 0.79. We observe that as σ̃ increases, the density of the points
on the persistence diagrams increases. This is consistent with Fig. 7(a), and we see that once the
thickened state is reached at σ̃ ≈ 5, the density of the points shows little further variation. We also
note that for σ̃ > 1, there is an increased number of points at the lower right side of the diagrams,
suggesting that the corresponding networks include more disconnected parts that live longer. In
more general terms, we deduce that as the suspension reaches the thickened state, its primary
structural properties are essentially unchanging, and thus the interaction forces on the contacts
simply increase proportionally with the applied stress. This conclusion is supported by considering
β1 persistence diagrams (figure not shown for brevity).

B. Interaction structure variation with solid fraction

Here we consider whether the structure changes as one progresses from CST to DST at fixed
stress by modifying the packing fraction. Note that if there are significant changes in the structure
of the interaction network, then we expect differences in the corresponding persistence diagrams.
Examination of plots such as those shown in Fig. 12 at different packing fraction shows only minor
differences, consistent with total persistence results. Essentially, the only influence of the packing
fraction is a slight shift of the points in persistence diagrams to lower values for smaller packing

034307-15
 MARCIO GAMEIRO et al.

FIG. 11. (a) Relative viscosity ηr vs scaled total persistence (dimension 1) τ1 /N for different packing
fractions for two- and three-dimensional data. Solid and empty symbols are used for 2D and 3D, respectively.
The dashed lines represent ηr ∝ [(τ1J − τ1 )/N]−α with τ1J /N = 0.34 and 0.42 for 2D and 3D, respectively, and
α = −1.8. Red and blue symbols represent two- and three-dimensional data. (b) The same data on a log-log
scale showing ηr vs (τ1J − τ1 )/N. The dashed line shows the power law of −1.8.

fractions. Therefore, if the system is already in the thickened state, we do not find qualitative
differences in the structure of interaction networks for the systems that are brought to the thickened
state by increasing either applied stress or packing fraction.

C. Role of physical dimensionality

We address finally whether and how the structure of interaction networks depends on the number
of physical dimensions. We note that resolving the influence of dimensionality by considering
purely topological measures is difficult, since the significant data reduction involved in computing
homology may also remove relevant information about the structure. Nevertheless, to gain some
basic insight, we default first to Betti numbers. Figure 13 shows the Betti numbers for the 2D and
3D simulations, scaled by the number of particles. We find that the Betti number curves are similar
for 2D and 3D, especially for β0 , which is strikingly similar for the 2D and 3D at large stress, where
the network structure has saturated; note that the value of peak shifts to larger F for 2D. The finding
that Betti numbers are so similar for 2D and 3D is surprising, given the difference in the number
of near neighbors in 2D (about six) and 3D (about 12). If we now consider the relationship of total
persistence to relative viscosity shown in Fig. 11, we find a similar scaling law for either 2D or 3D

034307-16
034307-17
INTERACTION NETWORK ANALYSIS IN SHEAR …

FIG. 12. (a) Relative viscosity ηr plotted as a function of stress σ̃ for area fraction φA = 0.79. Density plots of all the β0 persistence diagrams superimposed for
the given parameter values for the 2D simulations colored according to the density of points. The color bar indicates the density normalized across all the plots. The
plots correspond to φA = 0.79 and for (b) σ̃ = 0.5, (c) σ̃ = 1, (d) σ̃ = 2, (e) σ̃ = 5, (f) σ̃ = 10, (g) σ̃ = 50, and (h) σ̃ = 100.
 MARCIO GAMEIRO et al.

FIG. 13. Betti numbers for several values of σ̃ , scaled by the number of particles N for 2D simulations
(a),(b) for φA = 0.79, and for 3D simulations (c),(d) for φ = 0.56. Parts (a),(c) show β0 , and (b),(d) show β1 .
Here we have analyzed only the frictional contacts, i.e., contacts for which force is larger than the critical
normal force F0 and is normalized with the applied stress. Since there are no frictional contacts for F < F0 ,
the Betti number plateaus for small forces, and due to the normalization with applied stress, the onset of the
plateau decreases with an increase in stress.

on approach to jamming. This is again surprising and intriguing, as the total persistence provides a
compact description of the connectivity of the network, and this would a priori be expected to differ
depending on dimensionality. Such expectations arise from noting that a mechanically stable static
or jammed packing depends systematically on dimensionality D, varying from Z = 2D to D + 1
as the friction coefficient varies from μ = 0 to μ → ∞ [27]. Thus, we have identified similarities
across dimensionality in both the Betti numbers for the high-stress, or “fully networked” conditions,
and for the near-jamming total persistence of sheared suspension contact networks. This similarity
between measures of 2D and 3D networks points to a need for a more thorough understanding of the
relation of the network structure to the mechanical properties, as it may reveal principles of force
organization in amorphous materials.
As a final note regarding Fig. 13, the flat portion of the β0 and β1 curves at smaller values
of F and small σ̃ are due to lubrication forces that allow for the existence of a nonzero number of
components. For large values of σ̃ these separate components are not present since the applied stress

034307-18
 INTERACTION NETWORK ANALYSIS IN SHEAR …

is strong enough to make all components connect. A similar argument explains constant values of
β1 for the same values of the force threshold.

V. CONCLUSION AND PERSPECTIVE

In this paper, we have shown that there is a clear connection between the local properties of
the interaction networks that develop in sheared suspensions and their global rheological response.
This finding suggests that a proper understanding of the properties of the interaction networks may
be the key to quantifying their rheological response not only for relatively simple flow geometry
considered here, but in other more complex settings as well. It is encouraging that a relatively
simple measure, the total persistence describing the structure of loops in the interaction networks,
appears as an order parameter that describes well the rheology. This finding is important, since it
suggests that the details of the interaction networks may be absorbed into a compact measure for
the purpose of understanding the rheology. On the other hand, it should be noted that the total
persistence includes (in compressed form) the information about interaction network properties
for all interaction strengths at once, and therefore understanding and quantifying particle-particle
interactions may be needed for the purpose of describing the material response.
One particularly interesting finding is that the total persistence of loops in interaction networks is
the measure that correlates best with the rheological properties (viscosity). This result suggests that
future research should focus in more detail on the connectivity of the interaction networks in order
to understand better the rheology. It remains to be seen whether the fact that the total persistence of
loops in particular appears as an order parameter describing the system behavior remains valid for
more complex systems.
Our results show that the considered topological measures, both the simple ones (Betti numbers)
and more complex ones (total persistence), appear very similar for the two- and three-dimensional
suspension flows studied. Since the number of contacts and the geometry of particle packing are
strongly dependent on the number of physical dimensions, one wonders whether there is some
additional not yet discovered property of the interaction networks that leads to their dimension-
independent properties.
The interaction networks clearly play an important role in quantifying the rheological response
of sheared suspensions. While these networks are complex, particularly when considered for
all interaction strengths at once, persistent homology provides a way toward their analysis and
quantification, removing a significant amount of complexity and isolating rather straightforward
measures that could be used to quantify the rheology. Our results suggest that a similar type of
analysis may be of utility for the purpose of describing many other systems such that the interaction
between the basic building blocks can be described in terms of interaction networks. We also note
the potential for using persistent homology to analyze the temporal evolution of evolving networks.
Established measures for computing correlations between networks at different times exist, and we
expect that significant insight could be gained by applying such measures and comparing them with
evolving rheological properties.

ACKNOWLEDGMENTS
The work on particle simulations carried out in this work was supported, in part, under National
Science Foundation Grants No. CNS-0958379, No. CNS-0855217, and No. ACI-1126113 and
the City University of New York High Performance Computing Center at the College of Staten
Island. J.F.M. was supported by NSF 160528. L.K. was supported by NSF DMS-1521717 and
ARO W911NF1810184. M.G. was supported by FAPESP Grants No. 2016/08704-5 and No.
2016/21032-6 and by CNPq Grant No. 310740/2016-9, Brazil. K.M. and M.G. were supported
by NSF DMS-1521771, DMS-1622401, DMS-1839294 and DARPA contracts HR0011-16-2-0033
and FA8750-17-C-0054.

034307-19
 MARCIO GAMEIRO et al.

[1] R. Seto, R. Mari, J. F. Morris, and M. M. Denn, Discontinuous Shear Thickening of Frictional Hard-
Sphere Suspensions, Phys. Rev. Lett. 111, 218301 (2013).
[2] R. Mari, R. Seto, J. F. Morris, and M. M. Denn, Shear thickening, frictionless and frictional rheologies in
non-Brownian suspensions, J. Rheol. 58, 1693 (2014).
[3] R. Mari, R. Seto, J. F. Morris, and M. M. Denn, Nonmonotonic flow curves of shear thickening
suspensions, Phys. Rev. E 91, 052302 (2015).
[4] R. Mari, R. Seto, J. F. Morris, and M. M. Denn, Discontinuous shear thickening in Brownian suspensions
by dynamic simulation, Proc. Natl. Acad. Sci. (USA) 112, 15326 (2015).
[5] C. Ness and J. Sun, Shear thickening regimes of dense non-Brownian suspensions, Soft Matter 12, 914
(2016).
[6] I. A. Bashkirtseva, A. Yu. Zubarev, L. Yu. Iskakova, and L. B. Ryashko, On rheophysics of high-
concentrated suspensions, Colloid J. 71, 446 (2009).
[7] M. Wyart and M. E. Cates, Discontinuous Shear Thickening Without Inertia in Dense Non-Brownian
Suspensions, Phys. Rev. Lett. 112, 098302 (2014).
[8] A. Singh, R. Mari, M. M. Denn, and J. F. Morris, A constitutive model for simple shear of dense frictional
suspensions, J. Rheol. 62, 457 (2018).
[9] C. Heussinger, Shear thickening in granular suspensions: Inter-particle friction and dynamically correlated
clusters, Phys. Rev. E 88, 050201(R) (2013).
[10] J. Dong and M. Trulsson, Analog of discontinuous shear thickening flows under confining pressure,
Phys. Rev. Fluids 2, 081301 (2017).
[11] M. E. Cates, J. P. Wittmer, J.-P. Bouchaud, and P. Claudin, Jamming, Force Chains and Fragile Matter,
Phys. Rev. Lett. 81, 1841 (1998).
[12] R. Seto, A. Singh, B. Chakraborty, M. M. Denn, and J. F. Morris, Shear jamming and fragility in dense
suspensions, Gran. Matter 21, 82 (2019).
[13] A. Singh, S. Pednekar, J. Chun, M. M Denn, and J. F Morris, From Yielding to Shear Jamming in a
Cohesive Frictional Suspension, Phys. Rev. Lett. 122, 098004 (2019).
[14] M. Kramár, A. Goullet, L. Kondic, and K. Mischaikow, Quantifying force networks in particulate systems,
Physica D 283, 37 (2014).
[15] M. Kramár, A. Goullet, L. Kondic, and K. Mischaikow, Persistence of force networks in compressed
granular media, Phys. Rev. E 87, 042207 (2013).
[16] M. Kramár, A. Goullet, L. Kondic, and K. Mischaikow, Evolution of force networks in dense particulate
media, Phys. Rev. E 90, 052203 (2014).
[17] E. Azéma and F. Radjai, Internal Structure of Inertial Granular Flows, Phys. Rev. Lett. 112, 078001
(2014).
[18] D. Cantor, E. Azéma, P. Sornay, and F. Radjai, Rheology and structure of polydisperse three-dimensional
packings of spheres, Phys. Rev. E 98, 052910 (2018).
[19] E. Azéma and F. Radjai, Force chains and contact network topology in sheared packings of elongated
particles, Phys. Rev. E 85, 031303 (2012).
[20] L. Papadopoulos, M. Porter, K. Daniels, and D. Bassett, Network analysis of particles and grains,
J. Complex Netw. 6, 485 (2018).
[21] A. W. Lees and S. F. Edwards, The computer study of transport processes under extreme conditions,
J. Phys. C 5, 1921 (1972).
[22] P. A. Cundall and O. D. L. Strack, A discrete numerical model for granular assemblies, Geotechnique 29,
47 (1979).
[23] L. E. Edens, S. Pednekar, J. F. Morris, G. K. Schenter, A. E. Clark, and J. Chun, Global topol-
ogy of contact force networks: Insight into shear thickening suspensions, Phys. Rev. E 99, 012607
(2019).
[24] J. E. Thomas, K. Ramola, A. Singh, R. Mari, J. F Morris, and B. Chakraborty, Microscopic Origin of
Frictional Rheology in Dense Suspensions: Correlations in Force Space, Phys. Rev. Lett. 121, 128002
(2018).

034307-20
 INTERACTION NETWORK ANALYSIS IN SHEAR …

[25] L. A. Pugnaloni, C. M. Carlevaro, M. Kramár, K. Mischaikow, and L. Kondic, Structure of force networks
in tapped particulate systems of disks and pentagons I: Clusters and loops, Phys. Rev. E 93, 062902
(2016).
[26] L. Kondic, M. Kramár, L. A. Pugnaloni, C. M. Carlevaro, and K. Mischaikow, Structure of force networks
in tapped particulate systems of disks and pentagons II: Persistence analysis, Phys. Rev. E 93, 062903
(2016).
[27] A. J. Liu and S. R. Nagel, The jamming transition and the marginally jammed solid, Annu. Rev. Condens.
Matter Phys. 1, 347 (2010).

034307-21