A Unified Particle System Framework for Multi-Phase, Multi-Material

Visual Simulations
TAO YANG, Tsinghua University
JIAN CHANG, Bournemouth University
MING C. LIN, University of North Carolina, Chapel Hill
RALPH R. MARTIN, Cardiff University
JIAN J. ZHANG, Bournemouth University
SHI-MIN HU, Tsinghua University

Fig. 1. Frying an egg on a hot pan, achieved by enabling the diffusion of phases but disabling the diffusion of concentrations.

We introduce a unified particle framework which integrates the phase-field 1 INTRODUCTION

method with multi-material simulation to allow modeling of both liquids Physically-based simulation of fluids and solids has recently gained
and solids, as well as phase transitions between them. A simple elasto-
much attention in computer graphics, as it is central to reproducing
plastic model is used to capture the behavior of various kinds of solids,
including deformable bodies, granular materials, and cohesive soils. States of
realistic visual effects for a wide range of real-world phenomena.
matter or phases, particularly liquids and solids, are modeled using the non- Researchers have developed physically-based modeling techniques
conservative Allen-Cahn equation. In contrast, materials—made of different using various grid- or particle-based frameworks. The smoothed-
substances—are advected by the conservative Cahn-Hilliard equation. The particle hydrodynamics (SPH) method [Monaghan 1992] is widely
distributions of phases and materials are represented by a phase variable and used as a particle-based method due to its mass-conservation prop-
a concentration variable, respectively, allowing us to represent commonly erty, and its flexibility in handling topological changes. The SPH
observed fluid-solid interactions. Our multi-phase, multi-material system method has been widely used to simulate fluid flow [Müller et al.
is governed by a unified Helmholtz free energy density. This framework 2003], and it has been further adapted to simulate the dynamics of
provides the first method in computer graphics capable of modeling a con- both deformable bodies [Gerszewski et al. 2009; Müller et al. 2004]
tinuous interface between phases. It is versatile and can be readily used in
and granular materials [Alduán and Otaduy 2011].
many scenarios that are challenging to simulate. Examples are provided to
demonstrate the capabilities and effectiveness of this approach.
Since multiple fluids and solids often co-exist in the real world,
multi-material methods have also been introduced. Previous work
CCS Concepts: • Computing methodologies → Physical simulation; can be broadly divided into two categories: methods handling fluid-
Additional Key Words and Phrases: Phase-Field, Multi-Material Simulation, solid interaction, and those only handling multiple fluids. Many
Elastoplastic, Smoothed Particle Hydrodynamics, Phase Transition approaches for fluid-solid interaction have been proposed, covering
ACM Reference format: fluid to rigid solid coupling [Akinci et al. 2012], fluids interacting
Tao Yang, Jian Chang, Ming C. Lin, Ralph R. Martin, Jian J. Zhang, and Shi- with elastoplastic objects [Keiser et al. 2005; Solenthaler et al. 2007],
Min Hu. 2017. A Unified Particle System Framework for Multi-Phase, Multi- fluids interacting with granular materials [Lenaerts and Dutré 2009],
Material Visual Simulations. ACM Trans. Graph. 36, 6, Article 224 (Novem- and fluids in porous materials [Lenaerts et al. 2008]. In other cases,
ber 2017), 13 pages. the entire system may comprise different miscible or immiscible
https://doi.org/10.1145/3130800.3130882 fluids. Solenthaler and Pajarola [2008] simulated fluids using density
Permission to make digital or hard copies of all or part of this work for personal or contrast. By adopting the concept of volume fraction [Müller et al.
classroom use is granted without fee provided that copies are not made or distributed 2005], Ren et al. [2014] and Yang et al. [2015] managed to capture a
for profit or commercial advantage and that copies bear this notice and the full citation wide range of multi-fluid flow phenomena with rich visual effects,
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, using a mixture model and Helmholtz free energy. They focused
to post on servers or to redistribute to lists, requires prior specific permission and/or a on interactions between multiple fluids, and solids were not con-
fee. Request permissions from [email protected].
sidered. Yan et al. [2016] extended the multi-fluid SPH framework
© 2017 Association for Computing Machinery.
0730-0301/2017/11-ART224 $15.00 to incorporate solids, achieving impressive results. However, their
https://doi.org/10.1145/3130800.3130882

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
224:2 • Yang, T. et al

method is based on drift velocity, which requires small time steps, 2 RELATED WORK
and cannot easily or intuitively capture the evolution of phenomena In computer graphics, multi-material fluid simulation has been ex-
based on energy considerations [Yang et al. 2015]. Furthermore, tensively investigated during the last decade. Again, here we classify
it cannot simulate phase-change phenomena such as melting and multi-material fluid simulations as those involving only liquids, and
solidification. those involving solids and liquids. We start by introducing particle-
Phase transitions are commonly observed in the physical world, based solid simulation, then providing an overview of the relevant
such as dissolving, melting and solidification. Previous particle- works closest to this paper.
based simulation methods model this process mainly using either a
concentration criterion [Yan et al. 2016] or a temperature criterion
[Stomakhin et al. 2014], so that a given particle changes its phase if
its concentration or temperature exceeds a pre-defined threshold. 2.1 Solid Simulation
Although this approach is straightforward and easy-to-implement, In computer graphics, particle-based methods are commonly used
it suffers from two problems. First, the criteria are inconsistent: it to simulate fluids [Monaghan 1992; Müller et al. 2003]; they also
is problematic to model phase change phenomena involving both work well for simulating solid materials, including rigid objects,
concentration- and temperature-related variables. Secondly, the deformable bodies, and granular materials. Gray et al. [2001] simu-
interfaces between phases are discontinuous, as a given particle can lated elastic objects using a linear model by defining the artificial
only be in a single phase. To address these problems, we represent stress based on the signs of the principal stresses. Although this
the phase separately using an extra variable to describe it. This gave plausible results, their model suffers from accumulated nu-
phase variable is governed by a unified energy density function merical error. To capture the dynamics of truly elastic materials
related to both concentration and temperature. or large elastic deformations, researchers later developed a variety
In computer graphics, the terms ‘fluid’, ‘phase’ and ‘material’ have of solutions, such as storing the initial state [Müller et al. 2004],
been widely used, but not always consistently or accurately. We calculating the deformation gradient tensor [Gerszewski et al. 2009],
formally define them here: fluids are substances with zero shear adopting an implicit framework [Zhou et al. 2013], or use of ‘embed-
modulus; phases are states of substances, and materials are different ded space’ [Jones et al. 2014]. Müller et al. [2004] animated elastic,
kinds of substances (which may exist in any state or phase). For plastic, and melting objects using a physical model derived from
simplicity, we only consider the three most commonly observed continuum mechanics. This method needs to store the initial state
phases of matter, i.e., solid, liquid and gas, and largely focus on the in order to calculate stress. Gerszewski et al. [2009] proposed a new
first two for demonstration of concepts in this paper. way to compute the deformation gradient, avoiding the need to
Our approach significantly extends the energy-based multi-fluid store the initial rest configuration. Zhou et al. [2013] extended this
approach proposed by Yang et al. [2015] to represent ‘solid’ phases, work, using an implicit numerical integrator to achieve more stable
enabling modeling of a variety of fluid-solid interactions. We treat simulation of elastoplastic materials. Jones et al. [2014] introduced
phases and materials independently, thereby the unified multi-material embedded space—the least-squares best fit of the material’s rest state
simulation framework can model a much wider range of real-world in 3D—and used it to handle extreme elastic and plastic deforma-
phenomena. It is straightforward to describe the state of a particle tions. Jiang et al. [2015a] animated deformable objects, as well as
with a phase variable and a concentration variable. For simplicity, collision handling, using an affine particle-in-cell method (APIC)
we assume that the phase variable acts on the whole particle which with a Lagrangian force model.
can be composed of multiple materials. Zhu and Bridson [2005] adopted an existing fluid solver for simu-
In summary, the main contributions of this work are as follows: lating granular materials, decomposing the spatial domain according
to the strain rate tensor. Bell et al. [2005] used non-spherical parti-
• A simple elastoplastic model is introduced to simulate vari- cles to represent discrete elements of simulated materials, such as
ous solids, including deformable objects, granular materials, sand and rigid bodies. Narain et al. [2010] solved both the internal
and cohesive soils. It is capable of capturing a variety of pressure and frictional stresses in granular materials, achieving two-
solid phenomena that have been largely ignored in the lit- way coupling between granular materials and solid bodies. Using
erature, such as collapsing granular columns with varying a friction and cohesion model, Alduán and Otaduy [2011] simu-
aspect ratios, landslides, and dry/wet sand. lated granular flows in a predictive-corrective incompressible SPH
• By treating materials and phases separately, a much wider (PCISPH) framework [Solenthaler and Pajarola 2009]; this work was
range of challenging real-world phenomena involving mul- extended by Ihmsen et al. [2012], using a two-scale framework to
tiple materials and phases can be simulated, including com- simulate granular materials with high-quality visual detail. Recently,
plex fluid-solid interactions, such as dissolving, melting. the material point method (MPM) has gained popularity as a hybrid
• Our phase-field method defines the phase boundary using method to capture a variety of fluid and solid phenomena. Klár
a phase variable, providing a continuous interface between et al. [2016] recreated a wide range of visual sand phenomena by
phases. combining Drucker-Prager plastic flow and Hencky-strain-based
• A unified Helmholtz free-energy density for both the non- hyperelasticity. Daviet et al. [2016] presented a continuum-based
conservative Allen-Cahn equation and the conservative method for simulating non-smooth viscoplastic granular rheology
Cahn-Hilliard equation is proposed to guide the evolution using a material point method.
of concentrations and phases, respectively.

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
 A Unified Particle System Framework for Multi-Phase, Multi-Material Visual Simulations • 224:3

2.2 Multi-Fluid Simulation

A multi-fluid system is composed of more than one miscible or im-
miscible fluids. Particle-based multi-fluid simulations have attracted
much attention in the last decade. To simulate immiscible fluids,
Solenthaler and Pajarola [2008] derived a modified density calcula-
tion method to deal with density contrast more precisely. Peer et
al. [2015] incorporated this approach into an implicit incompress- Fig. 2. Particle views of a landslide coloured according to total deviatoric
ible SPH (IISPH) [Ihmsen et al. 2014a] to model multiphase highly plastic strain, for different Young’s modulus and Poisson ratio. These lead
viscous fluids. Macklin et al. [2014] simulated immiscible fluids to different failure patterns.
with a density ratio using position based dynamics, capturing the
Rayleigh-Taylor instability. For miscible fluids, the volume fraction
[Müller et al. 2005] is widely used to represent the spatial distribu- 2004]. Keiser et al. [2005] animated solids and fluids, as well as phase
tion of different components. Liu et al. [2011] integrated the volume transitions, by combining the equations of solid mechanics with
fraction with an SPH solver and mixed components using concen- the Navier-Stokes equations. Stomakhin et al. [2014] introduced
tration differences. To capture the underlying interactions between a dilational /deviatoric splitting of the constitutive model for heat
components due to flow motions and force distributions, Ren et al. transport, melting and solidifying materials. However, dissolving of
[2014] took the drift velocity into consideration. By doing so, they solids, which also involves a phase change, is largely overlooked by
could simulate a wide range of multi-fluid phenomena, including such methods.
mixing/unmixing of miscible and immiscible fluids, and chemical Dissolving of soluble materials is commonly observed in daily life.
reactions. Park et al. [2008] showed that energy can be used to drive Jiang et al. [2015b] proposed an energy-based method for real-time
multi-fluid simulations, and adopted the Cahn-Hilliard equation simulation of such phenomena. Yan et al. [2016] extended the work
to handle both miscible and immiscible fluids using a lattice Boltz- of Ren et al. [2014] to model solid phases. The distribution and shapes
mann method (LBM). Their approach was later extended by Yang of both fluids and solids are uniformly represented by their volume
et al. [2015], who achieved impressive results for problems such as fraction function and are governed by the conservation of mass
extraction (i.e. separating a substance from a mixture) and mixing and momentum within different phases. The approach impressively
of an egg yolk and egg white. All these multi-fluid simulations only captures various fluid-solid interactions, but it cannot capture the
consider fluids, but do not model fluid-solid interactions. evolution of phenomena based on energy, or provide flexible fluid
control. Furthermore, it cannot model phase-change phenomena
such as melting and solidification.
2.3 Fluid-Solid Interactions In such work, solid particles are insoluble in fluid-solid coupling
Fluid-solid interactions have attracted much attention in computer scenarios. It is also possible to model solid wetting by considering
animation. Commonly observed fluid-solid interactions fall into wettable solid particles to be slightly soluble while they remain in a
four categories: fluid-solid coupling, solid wetting, phase transitions solid state, rather than by considering them to be porous materials
between solid and fluid, and dissolving of solids. [Lenaerts et al. 2008]. To simulate pure phase transitions, particles
Solenthaler et al. [2007] used a unified particle model to han- only change their states of matter. When solids dissolve, the solid
dle the interactions between fluids and various solids including particles change phase in the liquid.
deformable bodies. Ihmsen et al. [2010] focused on rigid-fluid cou- These four categories cover a wide range of real-world phenom-
plings and showed that adaptive timesteps are required for boundary ena, and almost all phenomena modeled by previous research. In
handling in a PCISPH framework. Akinci et al. [2012] proposed a contrast to the previous work, we adopt a concentration variable
versatile SPH-based approach for two-way fluid-solid coupling us- and a phase variable to describe the separate evolution of materials
ing per-particle volume correction. More recently, Shao et al. [2015] and phases, allowing us to handle all four categories of interaction
combined PCISPH and geometric lattice shape matching to achieve (see Table 1 later) together in a versatile way. While many previous
two-way fluid-solid coupling with large time steps. works have considered multi-material simulations for solids and
For solid wetting, previous methods mainly focus on porous ma- liquids, none can handle all four types of interaction simultaneously.
terials. Lenaerts et al. [2008] succeeded in capturing a fluid flowing
through a porous deformable material; the porous flow is governed 3 SOLID MECHANICS
by Darcy’s Law. This work was later extended to simulate interac- As a basis for our approach, we first introduce an elastoplastic
tions between sand and fluids by using a unified SPH framework model for particle-based solid simulation. To solve the problem of
[Lenaerts and Dutré 2009]. Recently, Tampubolon et al. [2017] pre- inconsistent pressure forces when handling interactions, we treat
sented a multi-species model to capture the interactions between the hydrostatic pressure of solids in a similar way to that of fluids. To
porous sand and water, using a continuum mixture theory where allow realistic simulation of granular materials, we adopt a tension
species individually obey conservative equations and are coupled cracking treatment and a stress scaling-back procedure to handle
through a momentum exchange term. the numerical errors in computational plasticity. This elastoplastic
Phase transitions are highly correlated with temperature or inter- model underpins the solid simulations in our approach, allowing
nal energy. SPH has been widely used to simulate melting objects modeling of deformable bodies, wet and dry granular materials,
[Becker et al. 2009; Chang et al. 2009; Keiser et al. 2005; Müller et al. cohesive soils, and various interactions.

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
224:4 • Yang, T. et al

(see the Appendix A for the derivation) where ‘˙’ denotes the deriva-
tive with respect to time, G is the shear modulus, ϵ˙ = (∇u + ∇uT )/2
denotes the total strain rate tensor, and ė = ϵ˙ − Tr(ϵ˙ )/3δ is the
deviatoric shear strain rate tensor. The scalar function λ̇ denotes
the rate of change of the plastic multiplier λ, and is given by:
1  √2G 
s : ė + 3KαTr(ϵ˙ ) , f > 0



λ̇ =  G + 9Kα β S ,

 (6)
f

 0,
Fig. 3. A particle view demonstrating the benefit of using Jaumann stress  ≤ 0
rate. Left: with it, particles belonging to an elastic cube stay attached during where K is the bulk modulus. The shear and bulk moduli are related
rotation. Right: without it, particles at the corner detach as the cube rotates. to the Young’s modulus E and Poisson ratio v by:
E E
K= , G= . (7)
3(1 − 2v) 2(1 + v)
In our implementations, the deviatoric shear stress s in Equation (2)
is obtained by integrating its time derivative ṡ in Equation (5). We
use the leapfrog algorithm for numerical integration; s and S on the
right-hand side of Equations (5) and (6) are obtained at the half-time
step.
Fig. 4. Deformable jelly. Left: a purely elastic jelly. Middle and Right: plastic In this work, we use a fixed coordinate system and integrate the
jellies, with different degrees of plasticity.
stress rate. As a result, the rotation of a material introduces changes
in Cauchy stress. To eliminate this consequence, an extra stress rate,
3.1 Constitutive Model invariant with respect to material rotation, is employed to describe
the material response. The Jaumann stress rate is implemented,
The momentum equation for solids can be expressed as: and works well for small strains and large material rotations (see
Du 1 Figure 3).
= ∇ · σ + fext , (1)
Dt ρ s˙J = ṡ + s · ω̇ − ω̇ · s, (8)
where D/Dt = ∂/∂t + u · ∇ is the substantial derivative, u is the where ω̇ = (∇u − ∇uT )/2 denotes the spin rate tensor; subscript J
velocity, ρ is the density, fext is the external force, and σ is the denotes the Jaumann rate.
Cauchy stress tensor of the solid, which is normally divided into Jaumann stress rate is a so-called objective stress rate, aiming to
two parts in SPH: hydrostatic pressure p and deviatoric shear stress eliminate the mechanical response of a material with respect to the
s, i.e., σ = −pδ + s. We rewrite Equation (1) as: frame of reference. It has been previously used in the graphics com-
Du 1 munity occasionally, e.g. by Yan et al. [2016]. They simply employ
= (−∇p + ∇ · s) + fext . (2)
Dt ρ such a term, but did not illustrate how it benefits simulation; our
For consistency, the hydrostatic pressure is determined for both flu- examples do so. As illustrated in Figure 3 (and the supplementary
ids and solids using Tait’s equation [Becker and Teschner 2007]. To video), the rotation of a deformable cube introduces extra changes
determine s, the Drucker-Prager model is adopted in our work. The of the Cauchy stress tensor due to the fixed frame of reference,
yield condition f and plastic potential function д are respectively which however should not be part of the mechanical response of a
given by: material, resulting in particles at the corner detaching. By adopting
1 the Jaumann stress rate, this issue could be largely prevented.
f = √ S − 3αp − γ , (3) To smoothly approximate the Mohr-Coulomb hexagon on the de-
2 viatoric stress plane (π plane), several strategies have been proposed
1
д = √ S − 3βp, (4) to determine the Drucker-Prager cone parameters [Wang and Sitar
2 2004] (see the Appendix B). In 3D, the parameters are determined
where S = ||s ||; α and γ are Drucker-Prager’s constants, which by [Chen et al. 2012]:
are related to ϕ (internal friction) and the Coulomb constant kc 2 sin ϕ 2 sinψ 6kc cos ϕ
(cohesion); the latter can be treated as a function of the accumulated α= √ , β= √ , γ = √ .
3(3 − sin ϕ) 3(3 − sinψ ) 3(3 − sin ϕ)
plastic strain to handle plastic hardening and softening. For simplic-
ity, we just set kc to a constant. β has a similar expression to α and ψ is set to 0 in our examples. The implementation is given in Algo-
is related to the dilation angle ψ of materials such as soils. The yield rithm 1 later.
condition determines where plastic deformation occurs; the plastic
potential function is used to obtain the degree of plasticity (see the 3.2 Granular Materials
Appendix A). For simplicity, we formulate the deviatoric shear rate The proposed elastoplastic model can be used to capture the dynam-
ṡ as in the works of Bui et al. [2008] and Chen et al. [2012]: ics of purely elastic objects or plastic ones, depending on whether
 s  or not the Drucker-Prager yield criterion is satisfied (see Figure 4).
ṡ = 2G ė − λ̇ √ , (5)
2S This model is also capable of simulating granular materials [Bui et al.

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
 A Unified Particle System Framework for Multi-Phase, Multi-Material Visual Simulations • 224:5

Tension cracking treatment: if the stress state moves beyond the

apex of the yield surface (point E in Figure 6), a numerical error
known as tension cracking occurs [Chen and Mizuno 1990]. This
error shares some similarities with SPH tensile instability, resulting
in unrealistic fracturing or particle clumps. Following Chen and
Mizuno [1990], if the stress state exceeds the apex of the yield
surface, it should satisfy the condition below:
3αp + γ < 0. (9)
To avoid tension cracking, it is crucial to shift the hydrostatic pres-
sure to the apex of the yield surface, so the treatment is as follows:
γ
σ̃yy = σyy + (p + ), (10)
3
where y ∈ {1, 2, 3}. If the cohesion coefficient kc is set to 0, γ equals
0. In this case, the condition in Equation (9) is satisfied only if the
hydrostatic pressure p is negative. The treatment in Equation (10)
Fig. 5. Comparison with an MPM-based solution by revisiting [Zhu and ensures the pressure is non-negative, which also solves the SPH
Bridson 2005]’s column collapse. The friction angle is 20◦ ; 8000 particles tensile instability problem.
were used in both cases. Our SPH-based approach is about 4 times faster Stress-scaling back procedure: when plastic deformation occurs,
than the MPM-based one. Left: MPM solution achieves 22 fps. Right: SPH- numerical errors during computation may lead to a stress state far
solution achieves 80 fps. Without treating tension cracking or using stress from the yield surface (the path AB in Figure 6). In such circum-
scaling-back, the collapse of the sand pile exhibits cohesive behavior. stances, a stress rescaling procedure may be used with the help of
a scaling factor r [Bui et al. 2008]. For the Drucker-Prager yield
criterion, this scaling factor is given by:
Stress path without scaling
√ 3αp + γ
√ r= 2 . (11)
S/ 2 B S
Tension cracking
Accordingly, using the Drucker-Prager yield criterion, when the
Inaccessible region stress state exceeds the yield surface, it corresponds to the following
C Yield surface condition:
α A 1
E 3αp + γ < √ S. (12)
2
Elastic region
D The scaling factor r acts on the deviatoric shear stress components
3p and the hydrostatic pressure is left unchanged, so:
γ /α s̃ = rs. (13)
We have discussed two kinds of errors as well as their return map-
Fig. 6. 2D illustration of the Drucker-Prager yield criterion and numerical ping algorithms (Equations (10) and (13)). Since the leapfrog algo-
errors in computational plasticity. rithm is used in this work, the above return mapping algorithms
are applied during the half-time step.
We have modeled granular materials using the proposed elasto-
2008; Chen et al. 2012]. However, the granular materials directly plastic model in conjunction with the return mapping algorithms,
simulated by this model are cohesive (even if the cohesion kc is set achieving plausible visual results. Our model can simulate the col-
to 0), acting like wet sands (see Figure 5). We have implemented lapses of dry granular columns with different friction angles; see
this model using SPH and MPM [Sulsky et al. 1995], using CUDA; Figures 7b–7d. It can further capture the varying patterns when gran-
our observation is that an SPH-based approach is around 4 times ular columns with different aspect ratios collapse. When H /R = 2.0,
faster than the MPM-based one. See Figure 5. the entire column starts to flow immediately. At first, the upper free
We have introduced an elastic-perfectly plastic model. Theoret- surface of the column remains undeformed and horizontal, but later,
ically, the stress state should not lie outside the yield surface (the the top deforms to form a dome: see Figure 7d. When H /R = 0.9, a
inaccessible region in Figure 6) when plastic deformation occurs. circular area in the upper surface of the column is preserved undis-
However, due to numerical errors during computation, the stress turbed: see Figure 7g. Figure 7 shows collapsing granular columns
state may leave the elastic region as shown in Figure 6. In this case, a with varying friction angles and aspect ratios. We also demonstrate
return mapping algorithm is usually used to numerically return the the benefit of incorporating the tension cracking treatment: see
stress state to the yield surface [Bui et al. 2008; Chen and Mizuno Figure 7e. Without this term, the sand forms clumps and exhibits
1990]. We consider two kinds of errors. cohesive behavior.

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
224:6 • Yang, T. et al

(a) (b) (c) (d) (e) (f) (g)

Fig. 7. Sand collapses. (a) initial dry granular column with aspect ratio H /R = 2.0, for (b)–(e). (b)–(d) final states of collapse for friction angles ϕ of 20◦ , 30◦ ,
and 40◦ . (e) final states of collapse without tension cracking treatment for a friction angle ϕ of 40◦ . (f) initial dry granular column for (g) with aspect ratio
H /R = 0.9. (g) final state of the collapse with friction angle of 40◦ .

4 PHASE-FIELD MULTI-MATERIAL SIMULATION equation [Allen and Cahn 1972], which appears to have not been
In this section, we first review the energy-driven multi-fluid simula- previously used in computer graphics.
tion proposed by Yang et al. [2015], and then show how to extend
this framework to handle other phases. 4.2 Governing Equations
We make use of the momentum equation for motion, the Cahn-
4.1 Phase-Field Method Hilliard equation for the evolution of the concentration variable,
The phase-field model was established mainly for solving interface and the Allen-Cahn equation for the evolution of the phase variable.
problems. The method has been used to describe solidification pro- We assume the system has M materials with N possible phases.
cesses, as well as microstructure evolution at the interfaces of solids Du 1
and liquids. Yang et al. [2015] used the Cahn-Hilliard equation [Cahn = (−∇p + ∇ · s) + fext , (17)
Dt ρ
and Hilliard 1958] to calculate the changes in the fraction of each Dcm
material. The rate of change of each material is given by: = ∇ · (LCH ∇µm ), (18)
Dt
Dc ∂F
= ∇ · (LC H ∇µ), (14) µm = − ξ 2 ∆cm , (19)
Dt ∂cm
where c represents the concentration of the material (we use mass Dηk ∂F
= L AC (ξ 2 ∆ηk − ), (20)
fraction like Yang et al.), LC H denotes its degenerate mobility, and µ Dt ∂ηk
is its chemical potential, related to the Ginzburg-Landau free energy where s is the deviatoric stress for a solid (see Section 3) or the
density. The latter is composed of a bulk part and an interface part, viscosity tensor for a fluid (when Equation (17) becomes the Navier-
and can be expressed as: Stokes equation). cm and ηk are respectively the m-th material’s
∂F concentration and the k-th phase of a given particle. For simplicity,
µ= − ξ 2 ∆c, (15) we set ξ = 0.01 regardless of the type of the material or the state
∂c
of phase. LCH , L AC are both material dependent. However, unlike
where F is the Helmholtz free energy density, and ξ is associated density and viscosity, the values of LCH , L AC are set constant during
with diffuse interfaces; it introduces an internal length scale (inter- simulation for simplicity. The external force fext includes gravity,
face thickness). Yang et al. [2015] integrated their model into the as well as interactive forces between materials [Yang et al. 2015],
position-based fluids approach, achieving real-time performance. etc. For a consistent labeling convention, we use subscripts k, l to
As explained in Section 1, we introduce an extra phase variable η represent phases, m for materials, and i, j for particles.
to handle multiple phases consistently and continuously. η is used The concentrations of the materials and the phase variables must
to describe the current state of each particle in a continuous manner. satisfy the constraints:
For example, we may take the value one to mean solid and zero for
liquid; across an interface, the value varies smoothly from one to M
X −1 N
X −1
zero. The introduced phase variable η is considered as the proportion cm = 1, ηk = 1, cm , ηk ∈ [0, 1]. (21)
of a state. Unlike the concentrations of materials, the phase-field m=0 k =0
variables are not conserved. Real-world phenomena evolve from To meet the constraints, we adopt a post-correction process, as
high-energy states to low-energy states. We assume the system do Ren et al. [2014], Yang et al. [2015] and Yan et al. [2016]. For
locally tends to minimize energy and to conserve concentrations of simplicity, we use c, η to represent the vector of materials and
materials at the same time. Therefore, we represent the evolution phases of a given particle in the rest of this paper. They are vectors
of the phase variable using: of three elements, whose entries with indices 0, 1, 2 represent the
Dη ∂F proportions of solid, liquid and gas respectively. Since we focus on
= L AC (ξ 2 ∆η − ), (16) the first two phases, η 2 is always 0 in this work.
Dt ∂η
Temperature is important when simulating phenomena involv-
where L AC is the mobility. For simplicity, we adopt the same ξ as ing phase change. To model heat flow, we use a similar evolution
in the Cahn-Hilliard equation. This is known as the Allen-Cahn equation to that used by Stomakhin et al. [2014], adding an extra

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
 A Unified Particle System Framework for Multi-Phase, Multi-Material Visual Simulations • 224:7

source term for the phase transition: used for interpolation, satisfying q(0) = 0 and q(1) = 1; it is chosen
dT dη according to the situation. For example, if we use q(ηk ) = ηk2 (3 −
CH +Z = ∇ · (CT ∇T ), (22) 2ηk ), k ∈ {0, 1, 2}, it satisfies q ′ (0) = q ′ (1) = 0, which ensures
dt dt
where C H is the heat capacity per unit volume, Z is the latent that the minima of the free energy are at ηk = 0 and ηk = 1 (i.e.
heat per unit volume, and CT is the thermal conductivity. For each pure phase states) regardless of the temperature. However, when
particle at a given time, the phase change is considered to occur simulating a single material’s phase transition, the phase-change
between two phases, e.g., phase k, l (in this work, solid and liquid); process will not be triggered if only one phase exists in the initial
the η in Equation (22) can be either ηk or ηl , while Z should change stage. We then instead use q(ηk ) = ηk .
its sign accordingly. The proposed model can be used to capture simulations with
Every particle is considered to (potentially) be composed of mul- multiple phases including solid, liquid and gas. In this paper, we
tiple materials and multiple phases, but unlike Ren et al. [2014] and simply focus on two-phase examples involving solid and liquid.
Yan et al. [2016], we simply take it as a whole. We thus compute If additional constraints are introduced [Macklin et al. 2014], our
aggregate values for every particle at the beginning of each time approach can be extended to consider gaseous phase, which is not
step, for each of the required quantities: density, viscosity, heat ca- the main focus of this work. Thus, N in Equation (25) is set to 2 in our
pacity per unit volume, latent heat per unit volume, and thermal examples. Then only (Z 0,1 )m and (T0,1 )m appear in Equation (25),
conductivity. For simplicity, we use weighted averages: to model phase change between liquid and solid.
M
X −1
A= cm Am , (23) 4.4 Continuous Interfaces between Phases
m=0
Materials in different phases obey different physical laws. For in-
N −1
X stance, in Equation (17), s denotes the deviatoric stress for solids, but
B= ηk Bk , (24)
the viscosity tensor for fluids. To obtain this value, the velocity gra-
k =0
dient is needed. When calculating the velocity gradient in standard
where m and k are respectively the material and phase indices. A SPH, only neighboring particles in the same state are taken into
and B are continuous variables; the former is mainly influenced by consideration. This is problematic if particles have mixed phase. To
the type of the material, and the latter by the state of the phase, calculate the aggregate deviatoric stress, deviatoric stresses for both
respectively. In this paper, only the deviatoric stress is of the latter liquid and solid phases are needed. Each depends on its own veloc-
type. ity gradient, while the particle only has a bulk velocity. To address
this issue, previous works simply set each particle to a particular
4.3 Unified Helmholtz Free Energy Density
phase according to its concentration fraction, temperature and so
As explained in Section 4.2, the evolutions of c and η are driven on. However, we adopt a continuous phase interface: a particle at
by the Helmholtz free energy density F . In contrast to the work of the interface can be in multiple phases. We resolve this issue by
Yang et al. [2015], our Helmholtz free energy density works on both computing the velocity gradient for phase k of particle i as follows:
c and η. We thus call it the unified Helmholtz free energy density. It
plays a key role in guiding phase-field multi-material simulations.
The unified Helmholtz free energy density must be related to X (ηk )i + (ηk )j
(∇ui )k = Vj (u j − ui )∇Wi j , (26)
both the concentration fraction c and phase fraction η. Since phase 2
j
change typically depends on temperature, it must be taken into
consideration. Garcke et al. [2004] defined the bulk free energies for
individual phases. Since our simulations involve multiple phases, where Wi j is a symmetric kernel function; we use a spiky kernel
with the help of the interpolating function q, we define the final free for gradient calculation following the work by Müller et al. [2003].
energy function for phase-field multi-material simulation to be: Vj is the volume of the neighboring particle j. (ηk )i denotes the
N
X −1 M X −1 
T − (T̃k,l )m   k-th phase variable for particle i. As we only consider solids and
F= cm (Z k,l )m q(ηk ) − q(ηl ) liquids, we use (∇ui )k to calculate the appropriate tensor for phase k.
k,l =0,k <l m=0 (T̃k,l )m Finally, as we assume that the deviatoric stress is mainly influenced
(25)
M −1  by phase, we use Equation (24) to obtain the aggregate deviatoric
X RG   
+ Tcm ln(cm ) − CV T ln(T ) − 1 , stress. The whole process is summarised in Algorithm 1.
m=0
VM The use of continuous interfaces is of benefit in the implementa-
where (Z k,l )m and (T̃k,l )m are respectively the latent heat per unit tion of a multi-phase, multi-material framework, as no extra param-
volume and the phase transition temperature of the pure material m eter is needed to label the phase (solid, liquid or gas). This feature
when a phase transition occurs from phase k to phase l. (Z k,l )m = also avoids the computational effort needed to update such labels.
−(Zl,k )m , (T̃k,l )m = (T̃l,k )m for simplicity. CV is the specific heat, The phases of particles can be mixed, thereby closer to real physical
which is assumed to be independent of c and η. The molar volume behavior. Although existing methods can produce similar visual ef-
VM is taken to be a constant. RG is the gas constant. If cm is smaller fects, our method offers a unified treatment to seamlessly and easily
than a threshold (10−4 in our examples), the particle is considered handle multi-phase, multi-material simulations including transitions
to be free of the m-th material. q is a monotone function on [0, 1] between phases.

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
224:8 • Yang, T. et al

Algorithm 1: Aggregate Deviatoric Stress Computation

Input: η, u, s from the last step
Output: s for the current step
1: for all particles do
2: for all phases k do
3: compute (∇u )k according to Equation (26)
4: compute s k using (∇u )k : if phase k is solid, use Equation (5), if
liquid, use the standard viscosity tensor
5: end for
6: compute aggregate s using Equation (24) Fig. 8. Concentration and phase evolution. Water is poured onto two piles of
7: end for soluble and insoluble grains. The water particles are invisible. Left: diffusion
state of concentration at some given time (colors: blue: water, red: soluble
granular material, white: insoluble but wettable granular material). Right:
Algorithm 2: Simulation Loop for Multi-Phase, Multi-Material Modeling diffusion state of phase at the same time (color: blue: liquid, red: solid).
Framework
1: for all particles do
Table 1. Controlling concentration and phase evolution for various phenom-
2: compute aggregate values using Equation (23)
ena.
3: end for
4: for all particles do
5: compute thermal diffusion according to Equation (22), and update Phenomena Concentration Evolution Phase Evolution
the temperature
6: compute pressure using Tait’s equation Solid-Fluid Coupling ✗ ✗
7: end for Phase Transition ✗ ✓
8: for all particles do
9: update c and η using the Cahn-Hillard Equation (18) and the Wetting ✓ ✗
Allen-Cahn Equation (20) respectively Dissolving ✓ ✓
10: end for
11: for all particles do
12: compute pressure force using standard SPH approach
13: compute deviatoric stress using Algorithm 1 speed of sound, and νi j , ρ i j are respectively the average viscosity
14: compute artificial viscosity using Equation (27) and stress coefficient and density of particles i and j. We precompute the
following [Monaghan 2000] aggregate viscosity and density of every particle at the beginning
15: compute external forces of every step, to avoid treating miscible and immiscible particles in
16: end for different ways [Ren et al. 2014; Yan et al. 2016].
17: for all particles do Tensile instability is commonly observed in SPH simulations for
18: advance the particle both liquids and solids. We overcome this problem by using the
19: end for artificial stress term derived by Monaghan [2000].
The final SPH formulation for the momentum equation in con-
junction with the artificial viscosity and stress of particle i is given
4.5 Artificial Viscosity and Stress by:
When simulating solids including deformable bodies and granular Dui X  σi σ j 
materials using particle-based methods, physical oscillations are = + 2 − Πi j + (Ri + R j ) ∇Wi j + ( fext )i , (28)
Dt j ρi2 ρj
often observed due to instability of the numerical solutions. They
appear mostly in the initial stages of simulation, particularly if the where Ri is the artificial stress of particle i. We adopt the approx-
Courant–Friedrichs–Lewy (CFL) condition is not satisfied (e.g., a imations of spatial derivatives for SPH simulations suggested by
particle travels more than a certain fraction of its support radius Ihmsen et al. [2014b].
in one time step). To improve numerical stability and to damp out
undesirable oscillations, a dissipative term may be introduced into 4.6 Overall Algorithm
the pressure term, as suggested by Yan et al. [2016], who pointed Pulling all of the above ideas together, the main loop of our simula-
out that viscosity plays an important role in preventing particle tion approach is as given in Algorithm 2.
penetration. We therefore also apply an artificial viscosity term to
handle fluid-solid coupling. This term is given by: 5 APPLICATIONS
 −ac s Ωi j + bΩi2j We have proposed use of a concentration variable c and a phase
ν δ, vi j · x i j < 0


i j variable η to describe how concentration and phase evolve sep-

Πi j =  ρi j ,

(27)
arately. Table 1 shows how these may be controlled to simulate

vi j · x i j ≥ 0

 0,

 various challenging real-world phenomena. For instance, to capture
where i, j are particle indices, Ωi j = hvi j · x i j /(|x i j | 2 + 0.01h 2 ), phase transitions such as melting and solidification, we enable the
x i j = x i − x j , vi j = vi − v j , a, b are viscosity coefficients, c s is the evolution of phase, while not allowing concentrations to evolve (i.e.

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
 A Unified Particle System Framework for Multi-Phase, Multi-Material Visual Simulations • 224:9

Table 2. Performance; particle numbers change in some examples as water of dry granular columns with different angles (see Figures 7b–7d).
is added; our method is about 3 times faster than [Yan et al. 2016] We also demonstrate that the behavior of the collapse of an initially
vertical cylinder of granular materials depends on the aspect ratio
Example materials phases particles ms/step H /R, where H and R are its initial height and radius respectively
Deformable Jelly 1 1 50k 9.6 [Lube et al. 2004] (see Figures 7d and 7g).
Dry Sand (H /R = 0.9) 1 1 48k 10.2 Melting and Solidification: our multi-phase, multi-material model
Dry Sand (H /R = 2.0) 1 1 100k 15.0
can be used to capture a wide range of phase-changing phenomena
including melting and solidification. Figure 10 presents the melting
Butter Bunny 1 2 276k 24.1
of a butter bunny on a hot pan. The simulation involves one material
Egg 2 2 180k 20.8 (butter) with two phases (solid and liquid). We capture melting by
Dissolving Bunny (our method) 2 2 110k–260k 36.2 enabling diffusion of the phase variable. Since the initial conditions
Dissolving Bunny ([Yan et al. 2016]) 2 2 110k–260k 96.4 are single-phase, we set the monotone function q(ηk ) = ηk . Figure 1
Soluble/wettable Granules 3 2 200k–400k 46.4 demonstrates the frying of an egg. This is a two-material (egg white
and egg yolk) system with two phases (solid and liquid). The egg
Bubble Tea 4 2 320k–600k 61.0
white and egg yolk do not mix in this case, so any given particle
contains only one material. We initialize the egg white and the egg
yolk with different values of thermal conductivity to ensure different
we set the mobility to 0). We enable both concentration and phase
speeds of spread of phase change [Yang et al. 2016, 2017].
to evolve when simulating solid dissolution.
Dissolving Bunny: Figure 9 demonstrates the dissolution process
To simulate phase-change phenomena, we set the original phase
of a bunny. We also provides a comparison with results by Yan
to 0, and the target phase to 1. The temperature T in Equation (25)
et al. [2016] (see the supplementary video). Our method achieves
triggers the evolution of phase. For cases including solid wetting
comparable results and is about 3 times faster (see Table 2), as we
and dissolution, the temperature does not play a key role, so we set
avoid the use of computationally expensive drift velocity and an
it to a constant. The evolution of concentration occurs at interfaces
hypoplastic model.
according to the Cahn-Hillard equation (18). We set the monotone
Soluble and Wettable Granular Materials: Figure 11 shows advan-
function q(ηk ) in Equation (25) to ηk2 (3 − 2ηk ), so the evolution of
tages of our model over previous methods. Two piles of soluble
phase starts at the interfaces of phases thanks to the Laplacian term coffee granules and sands interact with water from above poured
in Equation (20) (The SPH formulation of the Laplacian operator onto them. To simulate soluble coffee granules, we enable diffusion
contains a gradient term [Ihmsen et al. 2014b]). To capture phe- of both concentration and phase variables, but to simulate insoluble
nomena like dissolution of solids, the evolution of concentration sand grains, we disable the diffusion of the phase variable. (The
interacts with the evolution of phase, and both only occur at inter- density ratio of granular material to water is 2 : 1 in each case).
faces. Figure 8 shows a snapshot of the differences in concentration Here, we assume the wettable sand and water slightly mix with each
and phase evolution in the same scenario as Figure 11 at a given other until reaching a saturation point: we assume a sand particle
time. can contain at most 5% water.
Decoupling the concentration and the phase leads to a straight- Bubble Tea: to demonstrate challenging real-world phenomena
forward and flexible approach to capture a variety of multi-material with multiple materials and phases, as shown in Figure 12, we sim-
and multi-phase interactions as illustrated in Table 1. ulate the making of bubble tea, a drink that contains soluble cof-
fee granules, wettable elastoplastic tapioca pearls, and two liquids
6 RESULTS
(water and milk). Our method can capture a variety of complex
6.1 Examples interactions in a versatile way. (The density ratios of coffee granules
In this section, we present various examples to demonstrate the vs. tapioca pearls vs. water vs. milk are 2 : 1.5 : 1 : 1). In this
range of simulations that our method can handle; supplementary scenario, we use particle boundary; to prevent particle penetration,
videos are also provided. Key parameters used in the simulations larger viscosity and artificial viscosity are adopted. We add vorticity
are given here; further details of parameters can be found in the confinement [Fedkiw et al. 2001; Macklin and Müller 2013] to make
supplementary material. the fluid less viscous.
Deformable Jelly: the proposed elastoplastic model can simulate We implemented our algorithm using CUDA on an NVIDIA
both purely elastic and plastic object behavior, as shown in Figure 4. GeForce GTX1080 11GB GPU and 8 Intel Core i7-3770 CPUs; the
An elastic jelly (left) and two plastic jellies (middle and right) are CPUs are relatively unimportant. Performance for the examples
dropped onto a plane. The elastic jelly rebounds and recovers its described above is given in Table 2. The particle numbers shown
shape, while the plastic jellies deform irreversibly with different include both fluid and boundary particles. The timings presented
degrees of plasticity. are averaged running times over the entire simulations. The time
Dry and Wet Sand: Figures 5 and 7 shows that our elastoplastic step is set to around 10−3 s, which is sufficient to achieve stable
model is capable of simulating both dry, non-cohesive sand, and simulation and prevent particle penetration.
cohesive, wet sand by choosing whether to use return mapping
algorithms or not. Our SPH-based approach is about 4 times faster
than the MPM-based one. Our method can simulate the collapses

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
224:10 • Yang, T. et al

Fig. 9. Soluble bunny. A deformable bunny dissolves in the water pouring from the pipe, coloring the water red.

Fig. 10. Melting a butter bunny on a hotplate. Phenomena such as wave-like spread and free sliding behavior of the unmelted butter can be observed.

Fig. 11. Soluble and wettable granular materials. Piles of coffee granules (brown) and sands (yellow) interact with water poured from above onto them. Left: the
sand is merely wettable if phase evolution is disabled. Right: the coffee granules dissolve, achieved by enabling the evolution of both phase and concentration.

Fig. 12. Making of Bubble Tea. A granular coffee column collapses in the goblet, then dissolves in water (pouring from right) and milk (pouring from left).

6.2 Discussion is introduced to drive the evolutions of both concentration and

Yang et al. [2015] also adopted the Cahn-Hilliard equation. Their phase.
work focused on interactions among multiple fluids; no solids are Although our examples show some similarities with those of Yan
considered. Furthermore, they proposed a user-defined Helmholtz et al. [2016], the methods used to produce the results are completely
free energy density function to artistically control the interactions; different. Yan et al. use drift velocity to capture the diffusion of con-
while in this paper, a unified Helmholtz free energy density function centration; thus, in every particle, each concentration has its own

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
 A Unified Particle System Framework for Multi-Phase, Multi-Material Visual Simulations • 224:11

velocity. Our method however uses a unified energy to drive the dif- For instance, the given particle above could be that all material A is
fusion of both concentration and phase. Every particle is considered solid and all material B is liquid, or vice versa, or various combina-
as a whole; each concentration within a single particle shares the tions in between. In the future, we hope to investigate the degree to
same velocity. Our method is much simpler to implement and is also which it is beneficial to use a more precise representation.
more efficient, overall being about 3 times faster than Yan et al.’s Crystallization is also commonly observed in daily life, e.g. in
method (see Table 2). Furthermore, Yan et al. use an elastoplastic ice formation on a window. The process of crystal growth involves
model to simulate deformable bodies and a hypoplastic model for both concentration and phase evolution, so in principle, is within
granular materials, which adds complexity. By improving the elasto- the scope of our model. Further effort is needed in future to capture
plastic model to make it capable of modeling granular materials, this phenomenon.
we are able to use a single general formulation for both cases. The We use the weakly compressible SPH (WCSPH) in our work, as
proposed method models landslides with different failure patterns it can readily integrate into our method. However, it suffers from
as well as collapsing granular columns with varying aspect ratios, compressibility compared with the MPM solver (see Figure 5 and
which have been largely overlooked in the graphics community. All the supplementary video). The position based fluids (PBF) [Mack-
the multi-material, multi-phase examples presented could not pre- lin and Müller 2013] and MPM have gained popularity in recent
viously be achieved in such a straightforward, easy-to-implement, years. It would be promising to integrate our method to these two
fast and unified way. frameworks.

7 CONCLUSION AND FUTURE WORK ACKNOWLEDGMENTS

We have extended the energy-based multi-fluid model [Yang et al. This work was supported in part by the Natural Science Founda-
2015] to simulate a wider range of multi-phase, multi-material sys- tion of China (Project Number 61521002), Research Grant of Beijing
tem including both liquids and solids, by incorporating the phase- Higher Institution Engineering Research Center, EU FP7 project
field method. Our proposed method can distinguish between phases Dr.Inventor (FP7-ICT-611383) and FP7 Marie Curie IRSES project
and materials, and treats them independently. By doing so, it be- AniNex (612627), the U.S. National Science Foundation, and an EP-
comes simpler to capture challenging real-world phenomena involv- SRC UK travel grant.
ing multiple materials and multiple phases. Our examples demon-
strate the effectiveness of the proposed unified simulation frame- REFERENCES
work. Nadir Akinci, Markus Ihmsen, Gizen Akinci, Barbara Solenthaler, and Matthias Teschner.
2012. Versatile rigid-fluid coupling for incompressible SPH. ACM Trans. Graph. 31,
We use a linear elastoplastic model, as it is simple and versatile, 4 (Jul. 2012), 62:1–62:8.
and enables modeling of a large variety of solid behavior, including Iván Alduán and Miguel A. Otaduy. 2011. SPH granular flow with friction and cohesion.
deformable bodies, granular materials, and cohesive soils. How- Proceedings of the 2011 ACM SIGGRAPH/Eurographics Symposium on Computer
Animation (2011), 25–32.
ever, when animating purely elastic objects, this model is unstable S.M Allen and J.W Cahn. 1972. Ground state structures in ordered binary alloys with
to capture large elastic deformation using SPH, due to numerical second neighbor interactions. Acta Metallurgica 20, 3 (1972), 423–433.
error of stress-rate accumulation, as simulation continues. In our Markus Becker, Markus Ihmsen, and Matthias Teschner. 2009. Corotated SPH for
deformable solids. Proceedings of the Fifth Eurographics Conference on Natural
experiments, smaller time steps can ensure longer time of stable Phenomena (2009), 27–34.
simulations, but it cannot solve the problem. More investigation is Markus Becker and Matthias Teschner. 2007. Weakly compressible SPH for free surface
flows. Proceedings of the 2007 ACM SIGGRAPH/Eurographics Symposium on Computer
needed to address this issue. Animation (2007), 209–217.
The proposed model is in principle capable of capturing gaseous Nathan Bell, Yizhou Yu, and Peter J. Mucha. 2005. Particle-based simulation of granular
phenomena. However, special techniques are needed for smoke material. Proceedings of the 2005 ACM SIGGRAPH/Eurographics Symposium on
Computer Animation (2005), 77–86.
simulation in current SPH based solvers [Macklin et al. 2014; Ren Ha H. Bui, Ryoichi Fukagawa, Kazunari Sako, and Shintaro Ohno. 2008. Lagrangian
et al. 2015, 2014], and these are not compatible with the governing meshfree particles method (SPH) for large deformation and failure flows of geomate-
equations for liquids or solids, making it tricky to incorporate smoke rial using elastic-plastic soil constitutive model. Int. J. Numer. Anal. Meth. Geomech.
32 (2008), 1537–1570.
within our new framework. Since only two phases, e.g., liquid and John W. Cahn and John E. Hilliard. 1958. Free energy of a nonuniform system. I.
solid are considered in this work, it is possible to use a single value of Interfacial free energy. Journal of Chemical Physics 28, 2 (1958), 258–267.
Yuanzhang Chang, Kai Bao, Youquan Liu, Jian Zhu, and Enhua Wu. 2009. A particle-
η (see explanations in Section 4.1) rather than a vector η to represent based method for viscoelastic fluids animation. (2009), 111–117.
the state of materials. We use the vector of phase η to leave room W. Chen and E. Mizuno. 1990. Nonlinear Analysis in Soil Mechanics: Theory and Imple-
for future extensions. mentation. Elsevier, Amsterdam.
Wei Chen, Tong Qiu, and M. ASCE. 2012. Numerical Simulations for Large Defor-
In this work, we have adopted the simplifying assumption that mation of Granular Materials Using Smoothed Particle Hydrodynamics Method.
the phase variable acts on the whole particle. For instance, suppose International Journal of Geomechanics 12, 2 (2012), 127–135.
a given particle is determined as 50% material A and 50% material B, Gilles Daviet and Florence Bertails-Descoubes. 2016. A Semi-Implicit Material Point
Method for the Continuum Simulation of Granular Materials. ACM Trans. Graph.
and as 50% solid and 50% liquid; we do not distinguish how much 35, 4 (2016), 102:1–102:13.
of A or B is solid or liquid, but just describe this mixed particle as R. Fedkiw, J. Stam, and H. W. Jensen. 2001. Visual simulation of smoke. ACM Trans.
Graph. (2001), 15–22.
50% solid and 50% liquid overall. Our experiments show that this Harald Garcke, Britta Nestler, and Bjorn Stinner. 2004. A diffuse interface model for
simplification (compared to strictly using a separate phase variable alloys with multiple components and phases. SIAM J. APPL. MATH 64, 3 (2004),
for each material) does not prevent us from capturing many com- 775–799.
Dan Gerszewski, Haimasree Bhattacharya, and Adam W. Bargteil. 2009. A point-
monly observed real-world phenomena. However, phenomena with based method for animating elastoplastic solids. Proceedings of the 2009 ACM
multiple phases and multiple materials are actually more complex. SIGGRAPH/Eurographics Symposium on Computer Animation (2009), 133–138.

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
224:12 • Yang, T. et al

J. Gray, J. Monaghan, and R. Swift. 2001. SPH elastic dynamics. Computer Methods in Barbara Solenthaler, Jürg Schläfli, and Renato Pajarola. 2007. A unified particle model
Applied Mechanics and Engineering 190, 49C50 (2001), 6641–6662. for fluid-solid interactions. Comput. Animat. Virtual Worlds 18, 1 (Feb. 2007), 69–82.
Markus Ihmsen, Nadir Akinci, Marc Gissler, and Matthias Teschner. 2010. Boundary Alexey Stomakhin, Craig Schroeder, Chenfanfu Jiang, Lawrence Chai, Joseph Teran,
handling and adaptive time-stepping for PCISPH. VRIPHYS (2010), 79–88. and Andrew Selle. 2014. Augmented MPM for phase-change and varied materials.
Markus Ihmsen, Jens Cornelis, Barbara Solenthaler, Christopher Horvath, and Matthias ACM Trans. Graph. 33, 4 (Jul. 2014), 138:1–138:11.
Teschner. 2014a. Implicit Incompressible SPH. IEEE Transactions on Visualization & Deborah Sulsky, Shi-Jian Zhou, and Howard L. Schreyer. 1995. Application of particle-
Computer Graphics 20, 3 (2014). in-cell method to solid mechanics. Computer Physics Communications 87, 1-2 (1995),
Markus Ihmsen, Jens Orthmann, Barbara Solenthaler, Andreas Kolb, and Matthias 236–252.
Teschner. 2014b. SPH Fluids in Computer Graphics. In Eurographics 2014 - State of A. Pradhana Tampubolon, T. Gast, G. Klar, C. Fu, J. Teran, C. Jiang, and K. Museth. 2017.
the Art Reports (2014), 21–42. Multi-species simulation of porous sand and water mixtures. ACM Trans. Graph. 36,
Markus Ihmsen, Arthur Wahl, and Matthias Teschner. 2012. High resolution simulation 4 (2017), 105:1–105:11.
of granular material with SPH. VRIPHYS (2012), 53–60. Gang Wang and Nicholas Sitar. 2004. Numerical Analysis of Piles in Elasto-plastic Soils
C. Jiang, C. Schroeder, A. Selle, J. Teran, and A. Stomakhin. 2015a. The Affine Particle- Under Axial Loading. 17th ASCE Engineering Mechanics Conference (2004), 13–16.
In-Cell Method. ACM Trans. Graph. 34, 4 (2015), 51:1–51:10. Xiao Yan, Yun-Tao Jiang, Chen-Feng Li, Ralph R. Martin, and Shi-Min Hu. 2016. Multi-
Min Jiang, Richard Southern, and Jian J. Zhang. 2015b. A Particle-based Dissolution phase SPH Simulation for Interactive Fluids and Solids. ACM Trans. Graph. 35, 7
Model using Chemical Collision Energy. Proceedings of the 10th International Con- (Jul. 2016), 79:1–79:11.
ference on Computer Graphics Theory and Applications (2015), 285–293. Tao Yang, Jian Chang, Bo Ren, Ming C. Lin, Jian Jun Zhang, and Shi-Min Hu. 2015. Fast
Ben Jones, Stephen Ward, Ashok Jallepalli, Joseph Perenia, and Adam W. Bargteil. 2014. multiple-fluid simulation using helmholtz free energy. ACM Trans. Graph. 34, 6 (Oct.
Deformation Embedding for Point-Based Elastoplastic Simulation. ACM Trans. 2015), 201:1–201:11.
Graph. 33, 2 (Apr. 2014), 21:1–21:9. Tao Yang, Ming C. Lin, Ralph. R. Martin, Jian Chang, and Shi-Min Hu. 2016. Versatile
Richard Keiser, Bart Adams, Dominique Gasser, Paolo Bazzi, Philip Dutré, and Markus Interactions at Interfaces for SPH-Based Simulations. Proceedings of the 2016 ACM
Gross. 2005. A unified lagrangian approach to solid-fluid animation. Proceedings SIGGRAPH/Eurographics Symposium on Computer Animation (2016), 57–66.
of the Second Eurographics / IEEE VGTC Conference on Point-Based Graphics (2005), Tao Yang, Ralph R. Martin, Ming C. Lin, Jian Chang, and Shi-Min Hu. 2017. Pairwise
125–133. Force SPH Model for Real-Time Multi-Interaction Applications. EEE Transactions
Gergely Klár, Theodore Gast, Andre Pradhana, Chuyuan Fu, Craig Schroeder, Chenfanfu on Visualization & Computer Graphics 23, 10 (2017), 2235 – 2247.
Jiang, and Joseph Teran. 2016. Drucker-Prager Elastoplasticity for Sand Animation. Yahan Zhou, Zhaoliang Lun, Evangelos Kalogerakis, and Rui Wang. 2013. Implicit
ACM Trans. Graph. 35, 4 (Jul. 2016), 103:1–103:12. integration for particle-based simulation of elasto-plastic solids. Computer Graphics
Toon Lenaerts, Bart Adams, and Philip Dutré. 2008. Porous flow in particle-based fluid Forum 32, 7 (2013), 215–223.
simulations. ACM Trans. Graph. 27, 3 (Aug. 2008), 49:1–49:8. Yongning Zhu and Robert Bridson. 2005. Animating sand as a fluid. ACM Trans. Graph.
Toon Lenaerts and Philip Dutré. 2009. Mixing fluids and granular materials. Computer (2005), 965–972.
Graphics Forum 28, 2 (2009), 213–218.
Shiguang Liu, Qiguang liu, and Qunsheng Peng. 2011. Realistic simulation of mixing
fluids. The Visual Computer (Mar. 2011), 241–248. A DERIVATION OF THE ELASTOPLASTIC MODEL
Gert Lube, Hernert E. Huppert, R. Stephen J. Sparks, and Mark A. Hallworth. 2004.
Axisymmetric collapses of granular columns. J. Fluid Mech. 508 (2004), 175–199. In this work, we adopt an elastic-perfectly plastic model. The total
Miles Macklin and Matthias Müller. 2013. Position Based Fluids. ACM Trans. Graph. 32, strain rate tensor ϵ˙ is commonly composed of two parts, i.e., elastic
4 (2013), 104:1–104–12.
Miles Macklin, Matthias Müller, Nuttapong Chentanez, and Tae-Yong Kim. 2014. Unified
strain rate ϵ˙e and plastic strain rate ϵ˙p :
ϵ˙ = ϵ˙e + ϵ˙p .
Particle Physics for Real-Time Applications. ACM Trans. Graph. 33, 4 (2014), 153:1–
153:12. (29)
J. J. Monaghan. 1992. Smoothed Particle Hydrodynamics. Annual Review of Astronomy
and Astrophysics 30, 1 (1992), 543–574. The generalized Hooke’s law states that:
J. J. Monaghan. 2000. SPH without a tensile instability. J. Comput. Phys. 82, 1 (2000),
1–15. σ̇ = De : ϵ˙e = De : (ϵ˙ − ϵ˙p ). (30)
Matthias Müller, David Charypar, and Markus Gross. 2003. Particle-Based Fluid
Simulation for Interactive Applications. Proceedings of the 2003 ACM SIG- where De is the fourth-order elasticity tensor, which can be formu-
GRAPH/Eurographics Symposium on Computer Animation (2003), 154–159.
M. Müller, R. Keiser, A. Nealen, M. Pauly, M. Gross, and M. Alexa. 2004. Point Based lated as:
Animation of Elastic, Plastic and Melting Objects. Proceedings of the 2004 ACM De = 2GT + Kδ ⊗ δ, (31)
SIGGRAPH/Eurographics Symposium on Computer Animation (2004), 141–151.
Matthias Müller, Barbara Solenthaler, Richard Keiser, and Markus Gross. 2005. Particle- where T = I − 1δ
⊗ δ, is the fourth-order deviatoric tensor. The
Based Fluid-Fluid Interaction. Proceedings of the 2005 ACM SIGGRAPH/Eurographics 3
Symposium on Computer Animation (2005), 237–244. plastic potential function д specifies the direction to which the plas-
Rahul Narain, Abhinav Golas, and Ming C. Lin. 2010. Free-flowing granular materials tic strain develops. The plastic strain rate tensor may be computed
with two-way solid coupling. ACM Trans. Graph. 29, 6 (Dec. 2010), 173:1–173:10.
Jinho Park, Younghwi Kim, Daehyeon Wi, Nahyup Kang, Sung Yong Shin, and Junyong
by using the plastic flow rule:
Noh. 2008. A Unified Handling of Immiscible and Miscible Fluids. Computer
∂д
Animation and Virtual Worlds 19, 3-4 (Sept. 2008), 455–467. ϵ˙p = λ̇ . (32)
Andreas Peer, Markus Ihmsen, Jens Cornelis, and Matthias Teschner. 2015. An Implicit ∂σ
Viscosity Formulation for SPH Fluids. ACM Trans. Graph. 34, 4 (Jul. 2015), 114:1–
114:10. Thus, substituting Equation (32) into Equation (30) gives:
Bo Ren, Yuntao Jiang, Chenfeng Li, and Ming C. Lin. 2015. A simple approach for
∂д 
σ̇ = De : ϵ˙ − λ̇
bubble modelling from multiphase fluid simulation. Computational Visual Media 1,

. (33)
2 (2015), 171–181. ∂σ
Bo Ren, Chenfeng Li, Xiao Yan, Ming C. Lin, Javier Bonet, and Shi-Min Hu. 2014.
Multiple-fluid SPH simulation using a mixture model. ACM Trans. Graph. 33, 5 (Sept. During the plastic flow, the stress state must remain on the yield
2014), 171:1–171:11.
H.F. Schwiger. 1994. On the Use of Drucker-Prager Failure Criteria for Earth Pressure
surface (see Figure 6). Thus, the consistency condition must be
Problems. Computer and Geotechnics 16, 3 (1994), 223–246. satisfied:
X. Shao, Z. Zhou, N. Magnenat-Thalmann, and W. Wu. 2015. Stable and Fast Fluid-Solid ∂f
Coupling for Incompressible SPH. Comput. Graph. Forum 34, 1 (Feb. 2015), 191–204. df = : σ̇ = 0. (34)
Barbara Solenthaler and Renato Pajarola. 2008. Density contrast SPH Interfaces. Proceed- ∂σ
ings of the 2008 ACM SIGGRAPH/Eurographics Symposium on Computer Animation This equation ensures that the new stress state σ + dσ still satisfies
(2008), 211–218.
Barbara Solenthaler and Renato Pajarola. 2009. Predictive-Corrective Incompressible the yield criterion after loading:
SPH. ACM Trans. Graph. 28, 3 (Jul. 2009), 40:1–40:6.
f (σ + dσ ) = f (σ ) + d f = f (σ ). (35)

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.
 A Unified Particle System Framework for Multi-Phase, Multi-Material Visual Simulations • 224:13

We use N and Q to represent the partial derivatives of f and д (see Table 3 summarizes determination of model parameters with
Equations (3) and (4)) with respect to σ for convenience: respect to friction angle ϕ, cohesive coefficient kc , and dilatancy
∂f 1 ∂д 1 angle ψ [Wang and Sitar 2004].
N := = √ s + αδ, Q := = √ s + βδ . (36)
∂σ 2S ∂σ 2S Table 3. Model parameter determination.
Substituting Equation (33) into Equation (34) gives:
 
N : De : ϵ˙ − λ̇Q = 0, α β γ

(37) D-P approximation
2 sin ϕ 2 sin ψ 6k cos ϕ
Compression Cone √ √ √ c
and thus: 3(3−sin ϕ ) 3(3−sin ψ ) 3(3−sin ϕ )

N : De : ϵ˙ 1  √2G  Compromise Cone

2 sin ϕ
√
2 sin ψ
√
6kc cos ϕ
√
3 3 3 3 3 3
λ̇ = = s : ė + 3Kα Tr(ϵ˙ ) (38)
N : De : Q G + 9Kα β S Extension Cone √
2 sin ϕ
√
2 sin ψ
√
6kc cos ϕ
3(3+sin ϕ ) 3(3+sin ψ ) 3(3+sin ϕ )
Substituting Equation (38) into Equation (33), it gives: sin ϕ sin ψ 3kc cos ϕ
Internal Cone √ √ √
√ 3(3+sin2 ϕ ) 1/2 3(3+sin2 ψ ) 1/2 3(3+sin2 ϕ ) 1/2
 2G 
σ̇ = 2G ė + K Tr(ϵ˙ )δ − λ̇ s + 3K βδ . (39)
S Experiments show that different failure patterns arise using dif-
The deviatoric part of σ̇ is as follows: ferent approximations [Schwiger 1994]. The appropriate choice of
√
2G approximation is application-dependent. When using an elastoplas-
ṡ = 2Gė − λ̇ s. (40) tic model, elastic and plastic deformations often co-occur. In this
S
work, we use the compression cone model, as small elastic deforma-
B DRUCKER-PRAGER APPROXIMATIONS tion can be ignored in comparison with plastic deformation when
simulating granular materials.
σ1
Compression cone
Compromise cone

Mohr-Coulomb

σ3 σ2

Internal cone
Extension cone

Fig. 13. Drucker-Prager approximations.

The Drucker-Prager yield criterion is a pressure-dependent model

to estimate the stress state for determining whether a material has
failed or undergone irreversible plastic deformation. The Drucker-
Prager yield surface is a smooth version of the Mohr-Coulomb yield
surface. A variety of Drucker-Prager approximations have been
proposed to match the Mohr-Coulomb yield surface as shown in
Figure 13.
• Compression cone: matching Mohr-Coulomb yield surface
in triaxial compression;
• Extension cone: matching Mohr-Coulomb yield surface in
triaxial extension;
• Internal cone: inscribed inside Mohr-Coulomb yield surface;
• Compromise cone: an average between extension and com-
pression approximations.

ACM Transactions on Graphics, Vol. 36, No. 6, Article 224. Publication date: November 2017.