Versatile Rigid-Fluid Coupling for Incompressible SPH

Nadir Akinci∗ Markus Ihmsen∗ Gizem Akinci∗ Barbara Solenthaler† Matthias Teschner∗
University of Freiburg University of Freiburg University of Freiburg ETH Zürich University of Freiburg

Figure 1: Fluid-rigid interaction in a large-scale setting. A boat with ragdolls passes a bridge (left). A second boat with ragdolls collides
with the bridge due to an increased flow rate and the bridge is released (middle and right).

Abstract 1 Introduction
Lagrangian fluid simulation is a popular topic in computer ani-
We propose a momentum-conserving two-way coupling method mation. Its particle-based nature allows simulating small scale
of SPH fluids and arbitrary rigid objects based on hydrodynamic phenomena, while mass conservation is trivially satisfied. Recent
forces. Our approach samples the surface of rigid bodies with works have addressed the performance issues of incompressible
boundary particles that interact with the fluid, preventing deficiency simulations, turning SPH into a competitive tool to model hydro-
issues and both spatial and temporal discontinuities. The problem dynamic effects. For instance, [Solenthaler and Pajarola 2009] has
of inhomogeneous boundary sampling is addressed by considering presented a prediction-correction method to eliminate compression
the relative contribution of a boundary particle to a physical quan- artifacts efficiently. Other improvements in this context have been
tity. This facilitates not only the initialization process but also al- presented in [Bodin et al. 2011; Raveendran et al. 2011]. In ad-
lows the simulation of multiple dynamic objects. Thin structures dition, the efficiency of SPH can be further improved by includ-
consisting of only one layer or one line of boundary particles, and ing multi-scale concepts [Adams et al. 2007; Solenthaler and Gross
also non-manifold geometries can be handled without any addi- 2011] and elaborated data structures for neighbor queries [Ihmsen
tional treatment. We have integrated our approach into WCSPH et al. 2011].
and PCISPH, and demonstrate its stability and flexibility with sev-
eral scenarios including multiphase flow. The interesting fluid behavior, however, emerges when rigid ob-
jects are added to a simulation (see Figure 1). While the two-way
coupling of particle-based fluids and solids seems to be straightfor-
CR Categories: I.3.7 [Computer Graphics]: Three-Dimensional ward, there is no general agreement how this should be handled.
Graphics and Realism—Animation; On one hand, the coupling has to cope with particle deficiency is-
sues at the boundary in order to prevent spatial and temporal dis-
continuities of physical properties of the particles and sticking arti-
Keywords: physically-based animation, fluid simulation, facts [Ihmsen et al. 2010] (see Figure 2). On the other hand, lower
Smoothed Particle Hydrodynamics, fluid-solid coupling dimensional geometries, e.g. thin structures and non-manifold sur-
faces, as well as constrained rigid bodies should be supported. This
work addresses these open issues. In the remainder of this section,
Links: DL PDF V IDEO we first discuss existing methods for treating boundaries in SPH
(Section 1.1) as well as for rigid-fluid coupling (Section 1.2), and
then highlight our contribution (Section 1.3).
∗ e-mail: {nakinci,ihmsen,gakinci,teschner}@informatik.uni-freiburg.de
† e-mail: solenthaler@inf.ethz.ch 1.1 Boundary Handling in SPH

For SPH boundary handling, distance-based penalty methods with

boundary particles have been commonly used [Müller et al. 2004;
Monaghan 2005; Harada et al. 2007; Monaghan and Kajtar 2009].
These approaches, however, require large penalty forces that restrict
the time step. Furthermore, particles tend to stick to the boundary
due to the lack of fluid neighbors. In [Harada et al. 2007], this
problem is alleviated by employing a wall weight function for ap-
proximating density contributions for planar boundaries.
The sticking of particles is avoided with frozen and ghost particles
based models, e.g. [Libersky and Petschek 1991; Hu and Adams
2006]. Frozen particles are fluid particles which are constrained
to static positions, whereas ghost particles are mirrored across the
boundary on the fly. In order to guarantee non-penetration, either
more than one layer of frozen particles should be used [Dalrymple
and Knio 2001], or the positions of penetrating particles should be
corrected [Ihmsen et al. 2010]. Since this class of methods samples
the boundary with fluid particles, the relevant field variables can be
well approximated with SPH. This results in continuous pressure Figure 2: Pressure profile of a box filled with particles where the
gradients, which consequently prevents sticking of fluid particles front side is clipped for illustration purposes. The fluid particle
to the boundaries. However, handling the interaction of the fluid pressures are color-coded and proportional to their red satura-
with thin shells is problematic in these approaches since the ele- tion. Black particles denote boundary particles. While [Becker
vated density on one side of a boundary particle affects potential et al. 2009b] leads to pressure noise and sticking artifacts (left),
fluid particles on the other side. The interaction of fluids with de- our method avoids these problems (right).
formable thin shells has been demonstrated in [Lenaerts and Dutré
2008] with additional distance-based non-symmetric forces to pre-
vent the penetration of fluid particles.
Mosher et al. 2008]), as well as for 2D heightfield models (e.g.
Alternative to particles, boundaries can also be efficiently repre- [Chentanez and Müller 2010]). A thorough discussion of these
sented with triangles. In this case, however, it is challenging to methods is, however, beyond the scope of our paper.
handle discontinuous surface normals and non-manifold structures
that cause spatial and temporal discontinuities of the fluid proper-
ties. 1.3 Contributions

1.2 Two-Way Fluid-Rigid Coupling in SPH We present a novel, versatile method for the two-way coupling of
SPH fluids and rigid bodies. We use boundary particles to sam-
For the two-way coupling of SPH fluids and rigid bodies, only few ple the surface of rigid objects, which has several benefits. First,
approaches have been proposed so far. In [Clavet et al. 2005], the the use of particles allows us to derive a model that can cope with
fluid is considered as a collection of rigid spheres exchanging im- different shapes, including lower-dimensional rigid bodies consist-
pulses with surrounding rigid bodies. In [Oger et al. 2006; Keiser ing of one layer (referred to as thin shells) or one line of boundary
et al. 2006], the pressure at the boundary is taken into account for particles (referred to as rods), as well as non-manifold geometries.
two-way coupled fluid-rigid interaction. In those models, how- Second, the inclusion of boundary particles successfully alleviates
ever, dynamic forces, e.g. viscosity, are neglected. More recently, the particle deficiency problem of SPH near boundaries, preventing
an impulse-based approach for simulating the two-way coupling density (and consequently pressure) discontinuities at the boundary
of SPH fluids with particle-based rigid bodies has been proposed and particle sticking artifacts.
in [Oh et al. 2009]. This approach, however, is not purely based on Our model addresses the problem of inhomogeneous particle sam-
hydrodynamic forces, relies on normal information for rigid bodies, pling at the boundary by deriving new equations that consider the
and does not guarantee non-penetration for thin shells. relative contribution of a boundary particle to a physical quantity.
In [Becker et al. 2009b], direct forcing has been employed for This does not only facilitate the particle initialization at complex
both one- and two-way fluid-rigid coupling. This method uses a boundaries, but also enables the use of multiple dynamic objects
predictor-corrector scheme to compute forces that constrain particle where the boundary sampling in the neighborhood may change due
positions and velocities to specific values. Non-penetration is guar- to contacts. A friction model is additionally included to simulate
anteed by using position correction. Different slip conditions are various slip conditions and drag effects. All pressure and vis-
realized by including a non-symmetric friction model. The position cous forces that are applied between fluid and boundary particles
correction and the non-symmetric friction forces are, however, not are symmetric, conserving linear and angular momentum. The ap-
momentum-conserving. Another issue is that timestep-dependent proach is designed such that even very large density ratios between
operations are used that require careful parameter evaluation for fluids and rigid bodies can be handled.
each setup. Finally, it requires two additional neighbor queries for
two-way coupling, which is rather expensive. This method has not 2 Fluid-Rigid Coupling
yet been extended to handle the interaction of a particular fluid par-
ticle with multiple rigid bodies or the simultaneous contact among
the bodies. This section starts with a brief introduction of SPH in Section 2.1,
followed by a description of our scheme for a corrected density
Particle-based fluid and deformable object interaction has been pre- computation at the rigid-fluid interface in Section 2.2. Sections
sented in [Müller et al. 2004] where the coupling has been realized 2.3 and 2.4 describe novel pressure and friction forces for pairs of
with the Lennard-Jones potential, but without hydrodynamic forces. boundary and fluid particles, while Section 2.5 discusses the overall
Boundary particles are automatically generated per triangle based forces and the symmetry of these forces.
on Gaussian quadrature. Deformations can also be computed with
SPH as shown in e.g. [Solenthaler et al. 2007; Lenaerts and Dutré 2.1 SPH Concept
2008; Becker et al. 2009a]. This has been extended to simulate
porous flow in [Lenaerts et al. 2008]. Further, fluid-deformable in-
In SPH, a generic field variable A at position xi is approximated
teraction on GPUs has also been briefly discussed in [Allard et al.
using a finite set of sampling points xj located within a distance
2011].
kxi − xj k < h as
There exist impressive two-way coupling approaches for Eulerian X
and semi-Lagrangian schemes (e.g. [Carlson et al. 2004; Guendel- A(xi ) = Vj Aj W (xi − xj , h), (1)
man et al. 2005; Chentanez et al. 2006; Batty et al. 2007; Robinson- j
where Vj is the volume represented at xj and W (xi − xj , h), also
written as Wij , denotes a Gaussian like kernel function with sup-
port radius h. The derivatives of A can be computed by using the
derivatives of W in the interpolation.
In SPH simulations, density is one of the essential field variables
that is used to compute pressure and viscosity forces. For this pur-
pose, the density summation approach is commonly used, which
reads as X X
ρi = Vj ρj Wij = mj Wij . (2)
j j Figure 3: A frigate is sailing on wavy sea. The right image shows
the irregular sampling of boundary particles.
In order to compute the pressure from the density, we either em-
ploy PCISPH [Solenthaler and Pajarola 2009] or WCSPH [Mon-
aghan 2005; Becker and Teschner 2007]. The employed pressure where j denotes fluid particle neighbors. Applying this idea to the
and viscosity forces are based on [Monaghan 2005]. For generat- volume of a boundary particle results in
ing surface tension effects, we rely on [Becker and Teschner 2007].
Finally, for simulating multiphase fluids, we use [Solenthaler and m 1
Pajarola 2008]. Vbi = Pbi = , (4)
mbi k Wik δbi
P
2.2 Corrected Density Computation with δbi = k Wik . Finally, (4) implies that the volume of a
boundary particle gets smaller for densely sampled areas and larger
The density summation approach (2) approximates the density of a for sparsely sampled areas. We now derive the fluid density com-
fluid particle correctly only if a particle is spherically surrounded by putation based on the boundary particle volume Vbi .
particles with the same initial density. Therefore, densities of fluid Even though (3) addresses discontinuities for uniformly sampled
particles near the boundaries are underestimated. In order to allevi- particles, it does not account for a variable sampling of particles.
ate this underestimation, we apply the Shepard filter [Panizzo 2004] Therefore, since the homogeneity of rigid sampling is not guaran-
to the density field. Even though this simple correction scheme sig- teed, (3) causes fluid particles to get large contributions from overly
nificantly improves the situation, the density gradient still remains sampled regions. Those overestimated densities cause large pres-
discontinuous near the boundaries. Additionally, since the parti- sure forces and therefore stability issues. This is due to the fact that
cles near the boundaries do not have neighbors that spherically sur- the contribution of boundary particles in (3) does not consider the
round them, forces on such particles constrain their movements to volume of a particle. This contradicts with the SPH concept, where
the boundaries, which causes sticking artifacts. To avoid this prob- the contribution of a particle in the approximation of any field vari-
lem, we take the neighboring boundary particles into account when able should be governed by its volume (see (1)). Therefore, we
computing densities and forces for fluid particles, similar to [Ihm- write the contribution of a boundary particle to a fluid particle by
sen et al. 2010]. taking the volume of the boundary particle into account as
Since we focus on the interaction of fluids with non-deformable
rigid bodies without melting effects, particles do not necessarily Ψbi (ρ0 ) = ρ0 Vbi , (5)
need to be generated inside a rigid. Therefore, we generate par-
where ρ0 denotes the rest density of the fluid that the rigid is inter-
ticles as a single layer at the surface similar to [Bell et al. 2005;
acting with. Finally, the corrected density of a fluid particle can be
Becker et al. 2009a; Ihmsen et al. 2010]. This approach saves
written in the form
memory and improves performance. The particle representations
of rigid bodies in the framework are computed either directly (e.g. ρfi = mfi
X
Wij +
X
Ψbk (ρ0i )Wik , (6)
for analytical shapes) or from mesh representations. Particle rep-
j k
resentations of triangle meshes are generated based on [Bell et al.
2005], which permits placing particles at an arbitrary offset to the which computes the densities correctly regardless of the bound-
surface mesh and yields a quite homogenous sampling. However, ary particle sampling. Note that Ψb increases the contributions of
at high-curvature regions of the mesh, the particle distribution usu- boundary particles by the amount of the volume ratio of bound-
ally remains non-homogenous, resulting in a denser sampling in ary and fluid particles in a uniformly sampled case (by a factor of
such areas (e.g., see Figure 3). We observed similar issues using ∼ 1.4). Since Gaussian like kernels are commonly used for SPH
the remeshing algorithm of [Botsch and Kobbelt 2004] and placing simulations, the weight of the next layer of particles is significantly
particles at the vertex positions. Fortunately, neither of the algo- lower compared to the closer layer. Therefore, using a single layer
rithms results in undersampled regions. Now, we can say that each of boundary particles with (6) and taking the missing particles into
boundary particle bi represents a volume Vbi at the surface of a account in (5) is a decent approximation in practice. See Figure
rigid: 4 for an illustration of the particle contributions. Our approach up-
mbi mbi dates the contributions of boundary particles for changing boundary
Vbi = = P ,
ρ bi k m bk Wik configurations with a minimal influence on the fluid particles that
are in contact with the boundary. We experimentally verified that
where k denotes boundary particle neighbors. In [Solenthaler and even dynamically moving and overlapping boundaries can be han-
Pajarola 2008], it has been shown that the density summation ap- dled, which is illustrated in the video corresponding to Figure 5,
proach in (2) causes stability issues for large density ratios due to left.
erroneous density estimations for particles at the interface. There-
fore, the density of a fluid particle can be written as Even though boundary particles are precomputed, a boundary parti-
X X cle is included in the simulation only if it is in the neighborhood of
ρfi = mfi Wij + mfi Wik , (3) a fluid particle, similar to [Ihmsen et al. 2010]. For moving bound-
j k
ary particles and all neighboring boundary particles, the represented
 Figure 5: Handling of overlapping boundaries (left). The negli-
gible interaction of fluid particles on opposite sides of a thin shell
(right). Pressures on particles are proportional to their red satu-
ration, black particles denote boundary particles, front side of the
box is clipped to make the fluid visible.
Figure 4: Color-coded contributions of boundary particles for two
ragdolls with different poses. Black and white particles represent
this formulation eliminates sticking artifacts and prevents penetra-
large and small contributions, respectively. The contribution is
tion of fluid particles to the boundaries without using extra forces
defined by the particular sampling density. Please note, e.g., the
or position correction. It also eliminates the need for normal in-
smaller contribution of rigid parts in contact.
formation for our boundary particles. Additionally, densities and
pressures for boundary particles are not required.
particle volumes are recomputed for handling the case of overlap- 2.4 Boundary-Fluid Friction Force
ping object parts or objects in close proximity (e.g. for dynamic or
kinematic rigid bodies). Inspired by the viscosity-based friction model proposed in [Müller
et al. 2004], friction between interacting fluid and boundary parti-
2.3 Boundary-Fluid Pressure Force cles is generated by employing the laminar artificial viscosity model
used in [Monaghan 2005; Becker and Teschner 2007]. It is written
In SPH, the pressure force between two particles can be directly as  
derived as min(vij · xij , 0)
Πij = −ν , (11)
 
pj
Fpi←j = −mi mj ∇Wij , (7) |xij |2 + h2
ρi ρj
with the viscous factor ν = ρ2αhc s
, where α is the viscosity con-
where p denotes pressure of a particle [Monaghan 2005]. For purely i +ρj

incompressible flow one can say that, stant, cs denotes the speed of the numerical propagation, vij =
vi − vj , xij = xi − xj , and  = 0.01 is used to avoid singularities
lim (ρi − ρj ) = 0 and lim (pi − pj ) = 0 for |xij | = 0. Finally, the viscosity force between two particles can
η→0 η→0 be computed as
where η denotes the density fluctuation of the fluid. Therefore, for Fvi←j = −mi mj Πij ∇Wij . (12)
weakly compressible fluids, we can assume that ρi ≈ ρj and pi ≈
pj . Consequently, (7) can be approximated as Based on (12), we define the viscosity force from a boundary parti-
  cle to a fluid particle as
px
Fpi←j = −mi mj ∇Wij , (8)
ρ2x Fvfi ←bj = −mfi Ψbj (ρ0i )Πij ∇Wij , (13)

where x can be either i or j. A similar assumption has been also with the reformulated viscous factor
used in the derivation of PCISPH [Solenthaler and Pajarola 2009].
σhcs
In practice, the applied pressure from fluid to some region of the ν= , (14)
2ρfi
rigid does not have any kinematic influence on the nearby fluid par-
ticles. Based on this fact, we write the pressure force applied from where σ is the viscosity coefficient between fluid and rigid. From
a boundary particle bj to a fluid particle fi as (14), ρbi is eliminated based on the same assumption that was used
! when deriving (9). When computing the viscosity force, the ve-
p pfi locity of a boundary particle can be easily computed based on the
Ffi ←bj = −mfi Ψbj (ρ0i ) ∇Wij , (9) kinematic properties of the rigid body it belongs to.
ρ2fi
The symmetric friction force from a fluid particle to a boundary
by substituting mj with Ψbj (ρ0i ) as done in (6), and using the fluid particle can be written as
particle’s density and pressure only. The symmetric pressure force
from a fluid particle to a boundary particle is Fvbj ←fi = −Fvfi ←bj . (15)

Fpbj ←fi = −Fpfi ←bj . (10) (15) results in drag effects on the rigid (see Figure 6). This idea was
also presented in [Becker et al. 2009b]. However, in their work,
In (9) and (10), the idea is making use of a fluid particle’s own friction forces are not momentum-conserving.
pressure when computing the boundary force. Magnitudes of the
boundary forces are based on the pressure of the fluid particle, 2.5 Total Force and Force Symmetry
which increases as the particle gets closer to a boundary. Since the
pressure of a fluid particle near a boundary would result in a pres- Our boundary particles are transformed based on the position and
sure force to the boundary, that force can be counteracted by a force orientation returned by the rigid solver before computing all rele-
that is proportional to the pressure of the fluid particle. Therefore, vant forces. Based on the derived forces, the total boundary force
 Algorithm 1 Simulation update with our boundary handling model
1: while animating do
2: foreach moving-rigid-body i do
3: synchronize boundary particles with rigid body state
4: foreach fluid-particle i do
5: find fluid and boundary neighbors
6: activate neighboring boundary particles
7: foreach fluid-particle i do
8: compute density ρi (t)
9: compute pressure pi (t) (e.g. WCSPH, PCISPH)
Figure 6: Two spheres with different fluid viscosities (σ = 0 for 10: foreach fluid-particle i do
yellow and σ = 8 for red) are dragged differently by the vortex. 11: add fluid forces Fp,ν,c,ext (t)
i
Fluid particles are colored according to speed where blue denotes
slow particles. The curves visualize the trajectories of the spheres. 12: add forces exerted by boundary particles Ftotal
fi
13: foreach active-boundary-particle i do
total
14: add forces exerted by fluid particles Fbi
15: foreach rigid-body i do
acting on a fluid particle and the total force acting on a boundary 16: compute the total force exerted by fluids Frigidi
particle from its fluid neighbors can be written as 17: compute the total torque exerted by fluids τrigidi
18: foreach fluid-particle i do
X  19: update xi , vi
Ftotal
fi = Fpfi ←bj + Fvfi ←bj , 20: update rigid bodies (e.g. Bullet)
j
21: end while
X 
Ftotal
bi = Fpbi ←fj + Fvbi ←fj .
j

4 Results
Since the pairwise forces between particles are symmetric (i.e.
Fpfi ←bj + Fpbi ←fj = 0 and Fvfi ←bj + Fvbi ←fj = 0), the In this section, we demonstrate the versatility of our approach
total
P boundary Pand viscosity forces are symmetric as well, i.e. in various simulation settings. If not stated otherwise, we used
total
i Ffi = i Ftotal
bi . PCISPH [Solenthaler and Pajarola 2009] as the basic fluid simu-
lation model, where the maximum permissible degree of compres-
Afterwards, for each boundary particle i which belongs to a dy- sion was kept at 1%. In our simulations, we used different particle
namic rigid and has a fluid neighborhood, Ftotal
bi is converted to radii r for different scenarios. The SPH smoothing length was al-
total force and torque for the rigid body as ways chosen as 4r. All simulated fluids had low laminar viscosity
(α = 0.01) and surface tension (κ = 0.05), which were determined
experimentally to approach the behavior of water. The employed
X
Frigid = Ftotal
bi ,
i
adaptive time-stepping schemes produced time steps roughly be-
X tween 10−2 and 10−4 . For all scenes, the computation overhead of
τrigid = (xi − xcm ) × Ftotal
bi , the rigid-fluid coupling was mainly between 5-10%. The overhead
i varied based on the number of fluid-boundary particle pairs. The
overhead of the rigid body simulation was usually below 1%. Fluid
where xcm is the center of mass of the rigid body, and xi is the po- surfaces were generated using a parallelized implementation [Ak-
sition of a boundary particle. Finally, Frigid and τrigid are applied inci et al. 2012] of the method proposed in [Solenthaler et al.
to the rigid body. 2007]. Renderings were done using mental ray v3.9.4 [NVIDIA
ARC 2011]. The simulations and renderings were run on an In-
tel Xeon X5690 with 12 GB RAM. Average computation time per
3 Implementation Details frame (note that for one frame several simulation steps are com-
puted) was 1 second to 1.5 minutes depending on the complexity of
For the SPH interpolations, we employ the cubic spline ker- the presented scene. These timings exclude surface reconstruction
nel [Monaghan 2005]. We use the Euler-Cromer scheme for time and rendering.
integration. We further employ the adaptive time-stepping schemes
explained in [Monaghan and Kos 1999] for WCSPH, and [Ihmsen We firstly compare our approach to [Becker et al. 2009b] in a simple
kg
et al. 2010] for PCISPH, where the shock handling criteria in the setting where a cube with density 400 m 3 was dropped into a con-
latter is also added to the former. We also included the velocities of kg
tainer with 200K fluid particles that have a rest density of 1000 m 3.
boundary particles inside the time step estimation criteria so as to We used WCSPH in this example. In contrast to [Becker et al.
prevent fluid particles from passing through (i.e. tunnelling). Even 2009b], our pairwise forces are symmetric. One frame could be
though we presented all underlying equations based on Monaghan’s computed in 6 seconds on average with our model, compared to 36
pressure and viscosity terms [Monaghan 2005], the same assump- seconds with [Becker et al. 2009b]. The reasons are twofold; our
tions that we used in our derivations can be applied to different boundary handling allows to use larger time steps (in this experi-
formulations as well, e.g. the force terms in [Müller et al. 2003]. ment three times larger on average), and it does not require addi-
tional neighbor queries. This experiment is included in the video
We used Bullet [Coumans 2011] for simulating rigid bodies. How- corresponding to Figure 7, top-left.
ever, because of the clear fluid-rigid solver decoupling, any rigid
solver might be used as well. For finding neighboring particles, Our viscosity model can simulate drag effects. This is shown in
we employed compact hashing as proposed in [Ihmsen et al. 2011]. Figure 6 where two spheres with different fluid viscosities were
The application of our two-way coupling approach to SPH is pre- dropped into a whirlpool. The sphere with zero viscosity moved
sented in Algorithm 1. faster to the center of the vortex, while the sphere with high vis-
 Figure 8: Handling of lower dimensional objects. The right im-
age shows the underlying particles. Note that planes and rods are
modeled with single particle layers.

Figure 7: The simulation setting used for the comparison

to [Becker et al. 2009b], where a cube is dropped into water (top- Figure 9: From left to right, the density of a sphere was slowly
kg kg
left). The setting for testing different slip conditions (top-right). Six increased from 1 m 3 to 50000 m3 .
balls with different densities were included in a multi-fluid simula-
tion (bottom-left). Thin shells of different densities, each of them
represented by a single particle layer (bottom-right) . kg kg
600 m 3 and 1500 m3 , respectively. Due to their different densities,
primitives interacted differently with the flow that was simulated
with up to 2.5M particles (see Figure 11). Finally we show a
cosity was dragged by the velocity field of the fluid. In the video complex scenario with various rigid objects (see Figure 1). In this
corresponding to Figure 7, top-right, we further show in a simple scene, two boats filled with ragdolls were dropped into a river like
setting that the symmetric force can be used to generate friction on flow. While the second boat was floating, we collapsed the bridge
the fluid particles. by removing the constraints on both sides of the bridge. Up to 6M
fluid particles were used in this simulation.
One main advantage of our method is that the interaction with lower
dimensional rigid bodies can be simulated. Figure 8 shows an ex-
ample where rods and planes, sampled by a single layer of bound- 5 Discussion and Future Work
ary particles, were dropped into a fluid with 300K particles. Fur-
ther, we dropped ragdolls that are modeled by multiple capsules In our simulations, we generated boundary particles for all rigid
connected with different constraints. Even in such challenging sce- bodies including static planar objects (e.g. water containers). For
narios, our two-way coupling approach produces plausible results. such regions, it would be more efficient to use a wall weight func-
Related to the previous example, Figure 7, bottom-right, shows that tion to approximate the density and force contributions, as done in
our approach is able to interact with shell-like structures without [Harada et al. 2007]. For complex boundaries with varying triangle
interpenetration. In this scene, three cylindrical shells with differ- size or in non-manifold regions, however, defining a wall weight
ent densities (which were represented with one layer of boundary function is difficult. In those cases, the triangle mesh should be
particles) interacted with 200K fluid particles. remeshed to get a nearly isotropic triangle distribution. Our bound-
ary particles can be related to such an isotropic triangle mesh, with-
In order to show that our approach works with multiphase fluids, we
out topology and normal information.
simulated two fluids with a density ratio of 1:3 and several spheres
with different densities (see Figure 7, bottom-left). Our approach computes boundary forces based on the pressure of
Our approach can handle large density ratios between fluid and rigid
bodies. A scene is illustrated in Figure 9, where a sphere with vari-
able density interacted with 90K fluid particles. The density of the
kg kg kg kg
sphere was changed from 1 m 3 to 10 m3 , 100 m3 and 500 m3 . Note
that at this point of the sequence, exactly half of the sphere was be-
low the water surface. Finally, the sphere density was changed to
kg
50000 m 3.

A more complex ragdoll example is shown in Figure 10 where 2M

fluid particles were used. Due to the dynamics of the individual
rigid parts of the ragdoll, the boundary particle sampling can dy-
namically change in the neighborhood of a fluid particle. The sam-
pling density is, however, considered in our equations so that dis-
continuities and large forces are prevented. Furthermore, we show
a non-uniformly sampled frigate traveling on wavy sea that was Figure 10: Dynamically moving rigid bodies in close proximity can
simulated using 4M particles (see Figure 3). change the sampling density in neighborhood of a fluid particle.
In another high resolution example, two towers were modeled us- Since our method takes variable boundary sampling into account,
ing cubes and cylinders. Each cylinder and cube had a density of discontinuous forces are prevented.
 Figure 11: The approaching water collapses two towers that were modeled by individual cylinders and cubes of different densities.

fluid particles. While the boundary forces are appropriate for in- fluid-rigid coupling. While particle-based solvers offer the bene-
compressible and weakly compressible fluids, the forces might not fit that complex boundaries can be handled in a simple way, there is
be sufficient to prevent interpenetrations when used with compress- no general agreement how solid-fluid interaction should be handled.
ible fluids. This is particularly the case, if the computed pressure Compared to existing techniques like frozen or ghost particles, di-
based on a certain density is much smaller for a compressible fluid rect mesh interaction, or penalty forces, our method offers several
compared to an incompressible fluid. Further, our approach is lim- benefits: Sampling the solids with our proposed boundary parti-
ited in terms of a minimal object size that can be handled. The cles allows including thin and non-manifold geometries into simu-
diameter of rods and the thickness of shells cannot be smaller than lations, since normal information is not needed. Our method does
the diameter of a fluid particle. Thus, the minimum possible ob- not require a uniform boundary sampling, which facilitates the par-
ject size is defined by the fluid particle resolution. Another issue of ticle initialization, especially when dealing with complex geome-
our approach is the interaction of fluid particles on opposite sides tries. Our solution does not only account for the inhomogeneous
of thin boundaries. Due to the coupling of minimal thickness of sampling, but also considers density (and consequently pressure)
boundaries and particle resolution, these interactions hardly influ- discontinuities at the boundary as well as symmetry of the forces.
ence the behavior of the fluid as demonstrated in the video corre- Overall, our method adheres to the concept of SPH, is efficient to
sponding to Figure 5, right. For improperly large time steps (that compute, and allows versatile fluid-rigid coupling.
are larger than what is estimated by the employed adaptive time-
stepping schemes), fluid particle tunneling may occur. However,
this is a general problem that also exists for fluid-fluid interaction.
7 Acknowledgments

We generated the boundary particles such that they are completely This project is supported by the German Research Foundation
enclosed by the rigid. Generating boundary particles exactly at the (DFG) under contract numbers SFB TR8 and TE 632/1. We would
surface would cause stability issues when fluid particles stuck be- like to thank NVIDIA ARC GmbH for their support, Jens Cornelis
tween two layers of overlapping boundary particles. These issues for his various helps during the project and the reviewers for valu-
could be prevented by detecting such fluid particles and treating able suggestions that helped to improve the manuscript. The frigate
them differently. In our approach, fluid particles are not immedi- model is courtesy of www.thefree3dmodels.com.
ately updated after the collisions in the final rigid update of Al-
gorithm 1. However, in the next iteration, based on the updated References
position of the rigid, forces are generated, and the fluid particles
are updated. Although our approach outperforms the global ap- A DAMS , B., PAULY, M., K EISER , R., AND G UIBAS , L. 2007.
proach in [Becker et al. 2009b], we believe that the investigation of Adaptively sampled particle fluids. ACM Trans. Graph. (SIG-
alternative global approaches for a simultaneous coupling similar GRAPH Proc.) 26, 3, 48–54.
to [Chentanez et al. 2006] is a very promising direction for future
research in particle-based fluids. A KINCI , G., I HMSEN , M., A KINCI , N., AND T ESCHNER , M.
2012. Parallel surface reconstruction for particle-based fluids.
Our method could also be integrated into previously presented SPH Computer Graphics Forum, to appear.
frameworks. One way to employ our approach in [Solenthaler
and Gross 2011] would be using two boundary samplings, one A LLARD , J., C OURTECUISSE , H., AND FAURE , F. 2011. Implicit
for the low resolution simulation and another for the high resolu- FEM and fluid coupling on GPU for interactive multiphysics
tion simulation. Our approach could further be extended to model simulation. In SIGGRAPH Talks.
the two-way coupling of SPH and deformable objects. In case of
BATTY, C., B ERTAILS , F., AND B RIDSON , R. 2007. A fast varia-
large deformations, however, potential sampling gaps have to be
tional framework for accurate solid-fluid coupling. ACM Trans.
addressed. Existing unified SPH models such as [Solenthaler et al.
Graph. (SIGGRAPH Proc.) 26, 3, 100–106.
2007; Lenaerts and Dutré 2008] could be used in combination with
our model to simulate the interaction of fluid, rigid and deformable B ECKER , M., AND T ESCHNER , M. 2007. Weakly compress-
models, including thin shells such as cloth. ible SPH for free surface flows. In Proc. of the 2007 ACM
SIGGRAPH/Eurographics Symposium on Computer Animation,
209–217.
6 Conclusion
B ECKER , M., I HMSEN , M., AND T ESCHNER , M. 2009. Corotated
We presented a novel boundary handling method for incompress- SPH for deformable solids. Eurographics Workshop on Natural
ible SPH fluids that is applicable to both one-way and two-way Phenomena, 27–34.
B ECKER , M., T ESSENDORF, H., AND T ESCHNER , M. 2009. Di- L ENAERTS , T., A DAMS , B., AND D UTRÉ , P. 2008. Porous flow
rect forcing for Lagrangian rigid-fluid coupling. IEEE Transac- in particle-based fluid simulations. In SIGGRAPH ’08: ACM
tions on Visualization and Computer Graphics 15, 3, 493–503. SIGGRAPH 2008 papers, ACM, New York, NY, USA, 1–8.

B ELL , N., Y U , Y., AND M UCHA , P. J. 2005. Particle-based sim- L IBERSKY, L., AND P ETSCHEK , A. 1991. Smooth particle hy-
ulation of granular materials. In Proc. of the 2005 ACM SIG- drodynamics with strength of materials. Advances in the Free-
GRAPH/Eurographics Symposium on Computer Animation, 77– Lagrange Method Including Contributions on Adaptive Gridding
86. and the Smooth Particle Hydrodynamics Method 395, 248–257.

B ODIN , K., L ACOURSIERE , C., AND S ERVIN , M. 2011. Con- M ONAGHAN , J., AND K AJTAR , J. 2009. SPH particle boundary
straint fluid. IEEE Transactions on Visualization and Computer forces for arbitrary boundaries. Computer Physics Communica-
Graphics, 99, 1–12. tions 180, 10, 1811–1820.
M ONAGHAN , J., AND KOS , A. 1999. Solitary waves on a Cretan
B OTSCH , M., AND KOBBELT, L. 2004. A remeshing ap- beach. Journal of Waterway, Port, Coastal, and Ocean Engi-
proach to multiresolution modeling. In Proc. of the 2004 ACM neering 125, 145.
SIGGRAPH/Eurographics Symposium on Geometry processing,
185–192. M ONAGHAN , J. 2005. Smoothed particle hydrodynamics. Reports
on Progress in Physics 68, 8, 1703–1759.
C ARLSON , M., M UCHA , P., AND T URK , G. 2004. Rigid fluid:
animating the interplay between rigid bodies and fluid. ACM M ÜLLER , M., C HARYPAR , D., AND G ROSS , M. 2003. Particle-
Trans. Graph. (SIGGRAPH Proc.) 23, 3, 377–384. based fluid simulation for interactive applications. In Proc. of the
2003 ACM SIGGRAPH/Eurographics Symposium on Computer
C HENTANEZ , N., AND M ÜLLER , M. 2010. Real-time simulation Animation, 154–159.
of large bodies of water with small scale details. In Proc. of the
2010 ACM SIGGRAPH/Eurographics Symposium on Computer M ÜLLER , M., S CHIRM , S., T ESCHNER , M., H EIDELBERGER ,
Animation, 197–206. B., AND G ROSS , M. 2004. Interaction of fluids with deformable
solids. Computer Animation and Virtual Worlds 15, 34, 159–171.
C HENTANEZ , N., G OKTEKIN , T., F ELDMAN , B., AND O’B RIEN , NVIDIA ARC, 2011. mental ray 3.9 [software].
J. 2006. Simultaneous coupling of fluids and deformable bodies. http://www.mentalimages.com/products/mental-ray/about-
In Proc. of the 2006 ACM SIGGRAPH/Eurographics Symposium mental-ray.html.
on Computer Animation, 83–89.
O GER , G., D ORING , M., A LESSANDRINI , B., AND F ERRANT, P.
C LAVET, S., B EAUDOIN , P., AND P OULIN , P. 2005. Particle- 2006. Two-dimensional SPH simulations of wedge water entries.
based viscoelastic fluid simulation. In SCA ’05: Proceedings of Journal of Computational Physics 213, 2, 803–822.
the 2005 ACM SIGGRAPH/Eurographics symposium on Com-
puter animation, ACM Press, New York, NY, USA, 219–228. O H , S., K IM , Y., AND ROH , B. 2009. Impulse-based rigid body
interaction in SPH. Computer Animation and Virtual Worlds 20,
C OUMANS , E., 2011. Bullet physics library (version 2.78) [soft- 2-3, 215–224.
ware]. http://www.bulletphysics.org.
PANIZZO , A. 2004. Physical and numerical modelling of subaerial
DALRYMPLE , R., AND K NIO , O. 2001. SPH modeling of water landslide generated waves. PhD thesis, Universita degli studi di
waves. In Proc. Coastal Dynamics, 779–787. L’Aquila, L’Aquila.

G UENDELMAN , E., S ELLE , A., L OSASSO , F., AND F EDKIW, R. R AVEENDRAN , K., W OJTAN , C., AND T URK , G. 2011. Hybrid
2005. Coupling water and smoke to thin deformable and rigid smoothed particle hydrodynamics. In Proc. of the 2011 ACM
shells. ACM Trans. Graph. (SIGGRAPH Proc.) 24, 3, 973–981. SIGGRAPH/Eurographics Symposium on Computer Animation,
33–42.
H ARADA , T., KOSHIZUKA , S., AND K AWAGUCHI , Y. 2007. ROBINSON -M OSHER , A., S HINAR , T., G RETARSSON , J., S U , J.,
Smoothed particle hydrodynamics on GPUs. In Proc. of Com- AND F EDKIW, R. 2008. Two-way coupling of fluids to rigid and
puter Graphics International, 63–70. deformable solids and shells. ACM Trans. Graph. (SIGGRAPH
H U , X., AND A DAMS , N. 2006. A multi-phase SPH method for Proc.) 27, 46:1–46:9.
macroscopic and mesoscopic flows. Journal of Computational S OLENTHALER , B., AND G ROSS , M. 2011. Two-scale particle
Physics 213, 2, 844–861. simulation. ACM Trans. on Graphics (SIGGRAPH Proc.) 30, 4,
81:1–81:8.
I HMSEN , M., A KINCI , N., G ISSLER , M., AND T ESCHNER ,
M. 2010. Boundary handling and adaptive time-stepping for S OLENTHALER , B., AND PAJAROLA , R. 2008. Density
PCISPH. In Proc. of VRIPHYS, 79–88. Contrast SPH Interfaces. In Proc. of the 2008 ACM SIG-
GRAPH/Eurographics Symposium on Computer Animation,
I HMSEN , M., A KINCI , N., B ECKER , M., AND T ESCHNER , M. 211–218.
2011. A parallel SPH implementation on multi-core CPUs. Com-
puter Graphics Forum 30, 1, 99–112. S OLENTHALER , B., AND PAJAROLA , R. 2009. Predictive-
corrective incompressible SPH. ACM Trans. on Graphics (SIG-
K EISER , R., A DAMS , B., D UTRÉ , P., G UIBAS , L., AND PAULY, GRAPH Proc.) 28, 3, 1–6.
M. 2006. Multiresolution particle-based fluids. Tech. rep., ETH
Zurich. S OLENTHALER , B., S CHLÄFLI , J., AND PAJAROLA , R. 2007.
A unified particle model for fluid-solid interactions. Computer
L ENAERTS , T., AND D UTRÉ , P. 2008. Unified SPH model for Animation and Virtual Worlds 18, 1, 69–82.
fluid-shell simulations. In ACM SIGGRAPH 2008 posters, SIG-
GRAPH ’08, 12:1–12:1.