Eurographics/ACM SIGGRAPH Symposium on Computer Animation (2005) K. Anjyo, P.

Faloutsos (Editors)

Particle-based Viscoelastic Fluid Simulation

Simon Clavet, Philippe Beaudoin, and Pierre Poulin
LIGUM, Dept. IRO, Universit de Montral

Abstract We present a new particle-based method for viscoelastic uid simulation. We achieve realistic small-scale behavior of substances such as paint or mud as they splash on moving objects. Incompressibility and particle anti-clustering are enforced with a double density relaxation procedure which updates particle positions according to two opposing pressure terms. From this process surface tension effects emerge, enabling drop and lament formation. Elastic and non-linear plastic effects are obtained by adding springs with varying rest length between particles. We also extend the technique to handle interaction between uid and dynamic objects. Various simulation scenarios are presented including rain drops, fountains, clay manipulation, and oating objects. The method is robust and stable, and can animate splashing behavior at interactive framerates. Categories and Subject Descriptors (according to ACM CCS): I.3.5 [Computer Graphics]: Computational Geometry and Object Modeling Physically Based Modeling; I.3.7 [Computer Graphics]: Three-Dimensional Graphics and Realism Animation and Virtual Reality

1. Introduction The eld of computational uid dynamics has introduced many different algorithms to simulate the complex behavior of uids. Since liquids and gases play an important role in everyday life, several uid simulation methods have been successfully used for computer graphics applications. Two main categories of simulation methods exist: Eulerian grids and Lagrangian particles. Eulerian methods discretize the problem using a subdivision of the spatial domain and control uid ow in each cell. Such techniques have been able to achieve among the most realistic simulations of liquid types ranging from simple ows to highly viscous uids, with plastic and elastic behaviors. They easily enforce the incompressibility condition, but do not guarantee mass conservation for small features. Lagrangian methods discretize the uid mass using particles. By directly tracking chunks of matter as they travel through space, particle methods trivially guarantee mass conservation and provide a conceptually simple and versatile simulation framework.

Figure 1: Fountain simulation illustrating sheet, lament, and drop formation.

email: clavetsi | beaudoin | poulin@iro.umontreal.ca

c The Eurographics Association 2005.

Particles also alleviate the dependence on a specic reference frame. In an Eulerian approach, the distance a uid feature can travel during a simulation step is limited by the grid resolution. Therefore, for very rapid and detailed ows, tracking changes as they occur at a xed point in space appears less natural than tracking changes which occur along a particle trajectory. We revisit particle-based simula-

S. Clavet, P. Beaudoin, and P. Poulin / Particle-based Viscoelastic Fluid Simulation

tion and propose extensions and modications that increase their exibility and robustness. Our formulation is simple and intuitive, and it can be implemented within days. An important disadvantage of particle simulations is the difculty to represent a smooth surface with particles. Robustly handling surface tension effects for small features is also recognized as a difcult task. Surface tension forces exist at the surface of the uid, and depend on its curvature. The problem comes from the fact that particle simulations avoid modeling the surface explicitly, and thus robustly computing its curvature can be cumbersome. We propose a novel procedure for incompressibility and particle anti-clustering. We call this procedure double density relaxation. Loosely speaking, double density relaxation computes two different particle densities: one quantifying the number of neighbors, and another quantifying the number of close neighbors. The force between two particles depends on these two context-dependent measures, instead of depending only on the distance between them. Liquids animated with this method display a smooth particle surface, and surface tension effects such as those shown in Figure 1 emerge naturally. No curvature has to be evaluated, and the procedure is robust even in presence of small surface details. Numerical instabilities are often an issue with any physically-based simulations, requiring the use of prohibitively small timesteps. We use a conceptually intuitive prediction-relaxation scheme that remains stable even for large timesteps and fast splashing effects. We also present a simple method to simulate viscoelastic substances by insertion and removal of springs between pairs of particles. Various material behaviors can be obtained with simple spring rest length update rules. The intuitive parameters governing such a scheme, combined with our fast simulation, allow an artist to interactively experiment with the uid properties to obtain the desired results. Finally, we describe a simple way to integrate the uid solver with a rigid-body simulator, enabling two-way coupling between liquids and objects. We also propose a simple solution to the stickiness of liquids on objects. 2. Previous Work 2.1. Eulerian Simulation Grid-based methods have been quite popular for computer graphics applications. Foster and Metaxas [FM96, FM97] were the rst to propose solving the full 3D Navier-Stokes equations on a regular grid in order to re-create the visual properties of a dynamic uid. Stam [Sta99] produced dynamic gases using a semi-Lagrangian integration scheme that achieves unconditional stability at the expense of articial viscosity and rotational damping. Foster and Fedkiw [FF01] extended the technique to liquids, tracking the surface using both a level-set method and particles inside

the liquid. Enright et al. [EMF02] added particles outside the uid volume to improve surface tracking. They also proposed an extrapolation technique to assign velocities to air cells just outside the liquid surface, resulting in a smoother surface motion on a coarse grid. A number of researchers have developed methods for modeling highly viscous or viscoelastic uids. Carlson et al. [CMHT02] developed an Eulerian solver for very high viscosity liquids and were able to simulate melting objects. Fluids simulated by Goktekin et al. [GBO04] exhibit not only viscous but also elastic and plastic behaviors by integrating and advecting strain-rate throughout the uid. 2.2. Lagrangian Simulation Smoothed particle hydrodynamics (SPH) is an alternative approach, rst developed by Lucy [Luc77] and by Gingold and Monaghan [GM77], to tackle astrophysical problems. In SPH, space is non-uniformly sampled using particles. These particles maintain various eld properties, such as mass density or velocity, and are tracked during the simulation. The eld quantities can be evaluated anywhere in space using radially symmetric smoothing kernels. Reeves [Ree83] introduced particle systems to computer graphics. They quickly became a popular tool for portraying various effects such as re or waterfalls. Miller and Pearce [MP89] borrowed ideas from molecular dynamics to add basic particle interactions, resulting in limited simulations of liquids and melting solids. Terzopoulos et al. [TPF89] modeled melting thermoelastic materials using particles that apply various forces to their neighbors. In a solid material, the particles are connected with springs, which weaken and eventually disappear as the material melts. This model does not handle plasticity. Desbrun and Gascuel [DG96] applied SPH concepts to computer graphics in order to simulate highly deformable bodies. Recently, Mller et al. [MCG03] implemented interactive liquid simulation and rendering using SPH. They simulate surface tension using ideas from Morris [Mor00], implicitly dening an interface with the particles and applying a force proportional to curvature, computed as the divergence of the normal eld. For surface features represented by a small number of particles, such curvature evaluation can cause numerical problems. Premoe et al. [PTB 03] also obtained realistic looking uids using a Lagrangian method. They solve the NavierStokes equations using the Moving-Particle Semi-Implicit (MPS) method [KO96], which ensures a greater level of incompressibility than standard SPH. Mller et al. [MKN 04] proposed a particle-based method for animating volumetric objects with material properties ranging from stiff elastic to highly plastic. Their technique is derived from continuum mechanics and therefore
c The Eurographics Association 2005.

S. Clavet, P. Beaudoin, and P. Poulin / Particle-based Viscoelastic Fluid Simulation

allows for direct specication of well-dened physical properties. Although they do not try to simulate a owing liquid, they obtain very realistic results for the material types supported. Steele et al. [SCED04] presented a particle-based method for simulating viscous liquids. They dene the adhesion properties of different types of particles using general distance-dependent forces between particles. Their iterative relaxation scheme to conserve volume is similar to ours, but the linear density kernel and the hard anti-penetration constraints limit their method to highly viscous uids. 3. Simulation Step Our uid is represented by particles evolving through space and time. A typical particle simulation goes through the following steps. First, various forces are computed and accumulated for each particle. These forces then modify the velocities, which are nally used to update particle positions. Such an explicit scheme tends to be unstable unless very small timesteps are used.

1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20.

Algorithm 1: Simulation step. foreach particle i

// apply gravity // modify velocities with pairwise viscosity impulses applyViscosity // (Section 5.3) foreach particle i // save previous position // advance to predicted position

vi vi + tg

xi

prev

xi

// add and remove springs, change rest lengths adjustSprings // (Section 5.2) // modify positions according to springs, // double density relaxation, and collisions applySpringDisplacements // (Section 5.1) doubleDensityRelaxation // (Section 4) resolveCollisions // (Section 6) foreach particle i // use previous position to compute next velocity

xi xi + tvi

vi (xi xi

prev

)/t

the particles according to their velocities. Lines 12 to 15 modify spring rest lengths and apply spring forces as particle displacements. Volume conservation, anti-clustering, and surface tension are then enforced (line 16). Finally, collisions between particles and static/dynamic bodies are resolved, and particle velocities are recomputed. Figure 2: Prediction-relaxation scheme. Our method avoids explicit force integration by using a prediction-relaxation approach. It is similar in concept to a more involved implicit scheme, but it is fast and straightforward to implement. Particles are moved according to their velocities and then their positions are relaxed, subject to positional constraints (Figure 2). At the end of the timestep, velocities are recomputed by subtracting previous positions from relaxed position. Because of this velocity recomputation, the relaxation displacements are equivalent to impulses applied to the velocity at the beginning of the timestep. These impulses being computed near the end position (on a path between the predicted position and the end position), the prediction-relaxation scheme bears some similarities with an implicit scheme in which the force exists at the unknown end conguration. Intuitively, using forces that exist further in time prevents instabilities by predicting difcult situations and reacting before they actually occur. Mathematical considerations regarding the prediction-relaxation scheme can be found in the Appendix. Our simulation step is detailed in Algorithm 1. First we update particle velocities according to gravity and viscosity (lines 1 to 5). Then we save the previous positions and move
c The Eurographics Association 2005.

4. Double Density Relaxation Our double density relaxation is a simplied and extended formulation of the SPH paradigm. The impulses exchanged between particles depend on two different measures of their neighbor density. 4.1. Density and Pressure In an SPH framework, the global goal of minimizing compressibility translates into a local constraint to maintain constant density. The density at particle i is approximated by summing weighted contributions from each neighbor j. We choose the density at particle i to be i =
jN(i)

(1 ri j /h)2

(1)

where ri j = |ri j |, ri j = x j xi , and N(i) denotes the set of neighboring particles that are closer than the interaction radius h. This form of density is not a true physical property; it is simply a number quantifying how the particle relates to its neighbors. We tried various other density kernel shapes, such as the unbounded (r/h)2 or the bell shaped polynomial (1 (r/h)2 )3 , but we obtained our best results with the quadratic spike (1 r/h)2 .

S. Clavet, P. Beaudoin, and P. Poulin / Particle-based Viscoelastic Fluid Simulation

We dene a pseudo-pressure Pi proportional to the difference between the current density i and a rest density 0 Pi = k(i 0 ) (2) where k is a stiffness parameter. To keep the formulation as simple as possible, we combined the normalizing constant, that usually scales the density in the SPH literature, with the parameters k and 0 . We also consider that all the particles have the same mass. This way, mass cancels out and does not have to be included in the equations. 4.2. Incompressibility Relaxation The incompressibility relaxation is implemented as a global loop over each particle i. An iteration of this loop is composed of two passes on each neighbor j of particle i. The rst pass estimates the local density at particle i by summing the weighted contributions of its neighbors (Equation 1). If this density is higher than the rest density 0 , neighbors will be pushed away. If it is lower, they will be pulled closer. Pushing and pulling neighbors is done in the second pass on the neighbors of particle i, in which each pair i j exchanges a displacement. The density relaxation displacement between two particles is proportional to the pseudo-pressure, weighted by the linear kernel function: Di j = t 2 Pi (1 ri j /h) i j r (3)

We solve the clustering problem by adding instead a second context-dependent pressure term. This new force does not simply depend on the distance between particles, but depends on a new density the near-density computed with a different kernel function. We dene near-density as near = i
jN(i)

(1 ri j /h)3

(4)

which uses a kernel with a sharper spike than our simple density. Again, we tried various other spike shapes, but this simple cubic gave good results in our tests.

In order to make the corresponding force exclusively repulsive, we dene near-pressure as Pinear = knear near i (5)

which is analogous to Equation 2, but with vanishing rest density. As for simple pressure, near-pressure produces a displacement for each particle pair in the second pass of the relaxation. Augmented with this new term, Equation 3 becomes Di j = t 2 Pi (1 ri j /h) + Pinear (1 ri j /h)2 ri j (6)

where ri j is the unit vector from particle i to particle j. Since force is integrated twice in time to get displacement, a factor t 2 is included in Equation 3 to preserve timestep duration independence (see the Appendix). We apply this displacement directly by modifying the predicted position of particle j. An equal and opposite displacement is applied to particle i, enforcing the action-reaction principle and thus linear momentum conservation. The displacement being directed along ri j , angular momentum is conserved. The linear ramp (1 r/h) scales the magnitude of the displacement depending on the distance between particles. Note that this pressure kernel is proportional to the derivative of the associated density kernel. This particular relation provides a consistent algorithm, as it is theoretically justied in the SPH literature (for example in [MCG03]). 4.3. Near-Density and Near-Pressure As described, this relaxation procedure does not prevent particle clustering. A particle can reach rest density by strongly pulling a small number of neighbors. The uid then separates into a collection of independent clusters, exhibiting no coherency. In preliminary tests, we followed an approach similar to Steele et al. [SCED04] and tried to prevent clustering by adding a simple distance-dependent repulsion force, but it caused visible artifacts when simulating low viscosity uid.

where the quadratic kernel denes how near-repulsion is applied to neighbors. Again, the near-pressure kernel is proportional to the derivative of the near-density kernel. Algorithm 2 summarizes our approximate volume conservation and anti-clustering method. For each particle, density and near-density are computed (Equations 1 and 4). Pressure and near-pressure are then evaluated (Equations 2 and 5). Finally, the second loop on neighbors applies the displacements to the particles (Equation 6). Neighbor particles j are immediately moved, but we sum the displacements of particle i and apply them only at the end of its relaxation step. This prevents any bias that could result from a xed neighbor processing order. The order in which particles i are relaxed is randomized but is not modied throughout the simulation. The result of double density relaxation is a coherent uid representation in which particles tend to be at equal distance from all immediate neighbors. However, directly enforcing a constant immediate neighbor distance with a Lennard-Joneslike force can lead to undesirable artifacts due to particles aligning in a rigid pattern. Here, near-density minimization smoothly restricts how the target density can be approached, and no rigid patterns appear.
c The Eurographics Association 2005.

S. Clavet, P. Beaudoin, and P. Poulin / Particle-based Viscoelastic Fluid Simulation

Algorithm 2: Double density relaxation. 1. foreach particle i 2. 0 3. near 0

4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 20. 21. // compute density and near-density foreach particle j neighbors( i ) if q < 1

will tend to move more since they are not affected much by near-pressure. The smooth repulsion of the nearest particles causes an indirect long-distance attraction. This contextaware attraction leads to smooth and stable particle structures. 5. Viscoelasticity Viscoelastic behavior is introduced in our model through three sub-processes: elasticity, plasticity, and viscosity. Elasticity is obtained by inserting springs between particles, plasticity comes from the modication of the spring rest lengths, and viscosity is introduced by exchanging radial impulses determined by particle velocity differences. 5.1. Elasticity To simulate elastic behavior, we add springs between pairs of neighboring particles. Springs create displacements on the two attached particles. The displacement magnitude is proportional to L r, where r is the distance between the particles and L is the spring rest length. It is also scaled with the factor 1 L/h, which gradually reduces to zero the force exerted by long springs. The process is detailed in Algorithm 3.
1. foreach spring i j 2. 3. 4.

q ri j /h + (1 q)2 near near + (1 q)3

P k( 0 ) Pnear knear near dx 0 foreach particle j neighbors( i ) q ri j /h if q < 1

// apply displacements

// compute pressure and near-pressure

D t 2 (P(1 q) + Pnear (1 q)2 ) i j r x j x j + D/2 dx dx D/2 xi xi + dx

4.4. Surface Tension We observed that besides reducing particle clustering, our method provides another important uid behavior: surface tension effects. As illustrated in Figure 3 and in the accompanying video [VID], particles group into structures such as sheets, laments, and drops.

Algorithm 3: Spring displacements. D t 2 kspring (1 Li j /h)(Li j ri j ) i j r xi xi D/2 x j x j + D/2

5.2. Plasticity A perfectly elastic substance always remembers its xed rest shape, and ghts external forces to recover it. On the other hand, a perfectly plastic substance considers its current shape as its rest shape. In general, plasticity can be thought of as the rate at which a substance forgets its past. This intuitive view of elasto-plasticity leads to our dynamic rest length spring scheme. At each timestep, springs are added or removed, and their rest lengths change depending on their current lengths. The rate of rest length change of a linearly plastic spring is proportional to its deformation: L = t (r L) where is the plasticity constant. A linearly plastic material slowly ows until all forces reach equilibrium. To model substances such as clay, which change shape under the signicant pressure given by ones ngers but resists small forces such as gravity, a non-linear plasticity model is needed. (7)

Figure 3: Oscillating drop. Surface tension is physically caused by attraction between molecules. Inside the uid, this attraction cancels out, but for molecules near the surface, asymmetry in neighbor distribution results in a non-zero net force towards the uid. Furthermore, this asymmetry changes depending on surface curvature. In view of these physical considerations, surface tension is usually considered to be an external force oriented towards the negative surface normal and with magnitude proportional to the surface curvature. We can visualize how the combined effect of pressure and near-pressure can produce surface tension. Suppose that density at a particle position is smaller than rest density. Neighboring particles will be pulled by pressure with an impulse proportional to the linear kernel, and then pushed by near-pressure with an impulse proportional to the sharpspiked quadratic kernel. Neighbors that are further away
c The Eurographics Association 2005.

S. Clavet, P. Beaudoin, and P. Poulin / Particle-based Viscoelastic Fluid Simulation

Algorithm 4: Spring adjustment. 1. foreach neighbor pair i j, (i < j) 2. q ri j /h 3. if q < 1 4. if there is no spring i j 5. add spring i j with rest length h
6. // tolerable deformation = yield ratio * rest length 7. d Li j 8. if ri j > L + d // stretch 9. Li j Li j + t (ri j L d) 10. else if ri j < L d // compress 11. Li j Li j t (L d ri j ) 12. foreach spring i j 13. if Li j > h 14. remove spring i j

Figure 5: Various plastic behaviors.

particles. These impulses modify particle velocities at the beginning of the timestep, before moving them to their predicted positions.
1. foreach neighbor pair i j, (i < j) 2. q ri j /h 3. if q < 1 4. // inward radial velocity 5. u (vi v j ) ri j 6. if u > 0 7. // linear and quadratic impulses 8. I t(1 q)(u + u2 ) i j r 9. vi vi I/2 10. v j v j + I/2

Our plastic spring model is inspired by the von Mises condition [Fun65], which states that plastic ow should occur only if the deformation is large enough. Translated into the spring system, this means that L should be changed only if |r L| is larger than some fraction of L (Figure 4). This fraction is specied by the yield ratio, denoted , for which we typically choose a value between 0 and 0.2 . The rest length increment can then be written L = t sign(r L) max(0, |r L| L) which becomes identical to Equation 7 when = 0. (8)

Algorithm 5: Viscosity impulses.

Algorithm 5 shows how viscosity changes particle velocities. We measure how fast particle j is moving towards parti cle i by projecting the velocity difference on ri j (line 5). For non-viscous uid, viscosity is only used to handle collisions, and we therefore apply impulses only if particles are running into each other. The impulse dependence on distance is captured by the linear kernel (1 ri j /h), and the factor (u + u2 ) controls the viscositys linear and quadratic dependences on velocity. This formulation is inspired by usual SPH techniques [DG96]. If a highly viscous behavior is desired, can be increased. For less viscous uids, only should be set to a non-zero value. The quadratic term prevents particle interpenetration by removing high inward velocity, but avoids smoothing the interesting features of the velocity eld. Viscosity impulses are applied sequentially to particle pairs. The ordering of particle pairs can theoretically bias the solution, but no visible artifacts were observed in our tests. 6. Interactions with Objects We have integrated our uid simulation into a rigid-body system, enabling interesting simulation scenarios such as oating objects and liquid sticking on surfaces. During the
c The Eurographics Association 2005.

Figure 4: Rest length change as a function of current length. As the uid moves, springs must be added and removed. A spring is added between two particles if their distance becomes smaller than the radius of interaction h, and is later removed if its rest length becomes larger than h. Pseudocode for our complete spring adjusting procedure is given in Algorithm 4. Additional control on viscoelastic behavior can be gained by using separate values of and for compression and stretching. For example, a stickier material can be simulated by setting compress to zero while letting stretch take some non-zero value. 5.3. Viscosity Viscosity has the effect of smoothing the velocity eld. It is applied as radial pairwise impulses between neighboring

S. Clavet, P. Beaudoin, and P. Poulin / Particle-based Viscoelastic Fluid Simulation

collision stage, the uid is considered to be an assembly of rigid spheres exchanging impulses with surrounding objects. 6.1. Collisions For each object, a signed distance eld is sampled and stored in a grid structure. For each particle we obtain the interpolated distance value d. A collision is identied when d is smaller than the particle collision radius R. For the collid ing particle we obtain the object normal n using the distance eld gradient. Our rigid-body solver is impulse-based, as in [GBF03]. In this kind of framework, object penetrations are resolved by sequentially applying impulses between bodies. This method for particle-body collisions leads to instabilities and makes it impossible to simulate oating objects. We therefore propose the three step method illustrated in Figure 6.

Algorithm 6: Particle-body interactions.

1. foreach body 2. save original body position and orientation 3. advance body using V and 4. clear force and torque buffers 5. foreach particle inside the body 6. compute collision impulse I 7. add I contribution to force and torque buffers 8. foreach body 9. modify V with force and with torque 10. advance from original position using V and 11. resolve collisions and contacts between bodies 12. foreach particle inside a body 13. compute collision impulse I 14. apply I to the particle 15. extract the particle if still inside the body

where is a friction parameter enabling slip ( = 0) or noslip ( = 1) boundary conditions. In some difcult situations, such as when a particle is already inside an object at the beginning of the timestep, removing the penetration velocity is not sufcient. We solve this issue by using the distance eld to extract the offending particles at the end of the collision stage (line 15). As it is common with any distance eld-based collision detection, collisions can be missed if the object is too thin with respect to the velocity of the particles. 6.2. Stickiness As they are now described, particles will detach and fall from objects. We implemented a simple method to make them stick, even underneath a horizontal surface.

Figure 6: Particle-body interactions. In the rst step, corresponding to lines 1 to 7 of Algorithm 6, bodies are advanced and impulses due to penetrating particles are accumulated into force and torque buffers. In the second step (lines 8 to 11), the velocity V and angular velocity of each body is updated using the accumulated force and torque. The bodies are advanced again from the initial conguration according to these new velocities, but the positions of the colliding particles are not yet modied. The usual body-body collisions and contacts are then resolved, and the bodies reach their nal positions and velocities. The third step (lines 12 to 15) updates the particle positions to remove penetration velocities. In this nal step, the bodies are considered to have innite inertia. Impulses due to particle-body collisions are based on a zero-restitution collision model with wet friction. This model requires the current particle velocity vi , computed using the difference between the current and previous particle positions. We then compute the particle relative velocity v = vi v p , where v p is the body velocity at the contact point. This velocity is separated into normal and tangential components: vnormal = ( n)n v v
tangent

Figure 7: Particles sliding underneath a sphere, before forming a drop. The idea is to modify the impulse computed at Equation 11 by adding an attraction term for particles that are close to an object. Let di be the distance between the object and the particle collision surface. An attraction impulse occurs when di is smaller than a xed distance d stick . To avoid artifacts such as particles jumping to the surface, d stick

(9) . (10)

= vv

normal

The impulse computed at lines 6 and 13 cancels the normal velocity and removes a fraction of the tangential velocity: I = vnormal vtangent
c The Eurographics Association 2005.

(11)

S. Clavet, P. Beaudoin, and P. Poulin / Particle-based Viscoelastic Fluid Simulation

should stay small enough with respect to the particle interaction range h. We use the following formulation: Istick = t kstick di 1 stick d di n (12)

7.3. Results The results, shown in the accompanying video sequences [VID] and in Figures 8 to 12, took an average of 2 seconds per frame to simulate, with typically 20000 particles. A 10 fps interactive session can simulate 1000 particles, which is enough to create relatively complex splashing effects. These timings include surface extraction with the marching cube algorithm and OpenGL display. Surface extraction usually takes from 10% to 60% of the computation time. High quality renderings were generated ofine with the raytracer Pixie [PIX].

which is maximal at distance d stick /2. Lines 5 and 12 of Algorithm 7 must now consider all particles in the attraction range. 7. Implementation and Results 7.1. Neighbor Search Finding all particles that are within the interaction radius of a particle is a frequent task in any particle-based method. Since the interaction radius h is constant and identical for each particle, it makes sense to use a simple regular spatial hashing grid with cube size h. Each cube stores a list of enclosed particles, which is updated each time a particle moves from one cube to another. All neighbors reside in the particles cube and its 26 neighboring cubes. Furthermore, we avoid using a xed cube lattice by only instantiating cubes that contain particles. The cubes are stored in a hash table, indexed with their 3D index. The animation scenario can thus take place in a virtually innite domain, and no information on the future location of the uid is needed at the beginning of the simulation. This spatial hashing method is similar to Teschner et al. [THM 03]. 7.2. Rendering For fast previsualization, particles are simply rendered as polygonized spheres. High quality rendering uses a surface mesh extracted with the marching cube algorithm [LC87]. The marching cube extracts an isosurface of a scalar function dened by the particles. We use the function (x) =

Figure 8: Viscous rain on a character.

(1 r/h)
j

1 2

(13)

where r = |x x j |. The square root ensures that the function behaves like a distance function, i.e., has an almost constant slope around the extracted isosurface. This is important for moving isosurfaces because the location of the surface in a cube is linearly interpolated. If the function had a changing slope around the isovalue, temporal artifacts would appear as particles move. Simply using a sum of cones could achieve this goal, but the result would not be smooth enough. As in the case of neighbor search, the cube grid used for isosurface extraction is sparse and dynamic. The cube size hsurface is independent from the particle interaction radius h, and is typically smaller if a smooth surface mesh is desired. A spatial hash table is used to store the non-zero grid points as particles add their contributions to the implicit function.

Figure 9: Pouring liquid in a tank containing a heavy object.

Figure 10: A oating object.

c The Eurographics Association 2005.

S. Clavet, P. Beaudoin, and P. Poulin / Particle-based Viscoelastic Fluid Simulation

deformable 3D models and rain drops. The integration of our uid solver into a rigid-body system has enabled oating objects, and a simple solution to the stickiness of liquids on objects has further extended the simulation possibilities. Our method is specically designed for fast simulation of rapidly moving uids, but is not as well adapted for simulation of low viscosity liquids such as water. We could adapt the method to simulate water by using an explicit high order accurate integration scheme, and simulating numerous timesteps per frame. The increased computation times would, however, contradict our initial goals of interactivity. Figure 11: Pouring liquid on the Stanford Bunny. Several other avenues for future work can be identied. Volume conservation could be improved by coupling our method with a more sophisticated technique such as the Moving-Particle Semi-Implicit method [KO96]. We would also like to consider several types of particles to animate the interaction between different substances. Simulating air would signicantly increase animation realism by enabling bubble formation. We believe that the simplicity, stability, speed, and versatility of our method should prove very useful for many applications. Acknowledgements The authors thank all the LIGUM members for their support. Luc Leblanc deserves a special mention for his help with the renderings. We thank the reviewers for their helpful comments. Financial support from NSERC is acknowledged. References [CMHT02] C ARLSON M., M UCHA P. J., H ORN R. B. V., T URK G.: Melting and owing. In SIGGRAPH/Eurographics Symposium on Computer Animation (2002), pp. 167174. [DG96] D ESBRUN M., G ASCUEL M.-P.: Smoothed particles: A new paradigm for animating highly deformable bodies. In Computer Animation and Simulation 96 (1996), pp. 6176. [EMF02] E NRIGHT D., M ARSCHNER S., F EDKIW R.: Animation and rendering of complex water surfaces. In SIGGRAPH (2002), pp. 736744. [FF01] F OSTER N., F EDKIW R.: Practical animations of liquids. In SIGGRAPH (2001), pp. 2330. [FM96] F OSTER N., M ETAXAS D.: Realistic animation of liquids. Graphical Models and Image Processing 58, 5 (1996), 471483. [FM97] F OSTER N., M ETAXAS D.: Modeling the motion of a hot, turbulent gas. In SIGGRAPH (1997), pp. 181 188. [Fun65] F UNG Y. C.: Foundations of Solid Mechanics. Prentice-Hall, 1965.

Computing many timesteps per frame can greatly improve volume conservation and overall simulation quality. We used up to 10 timesteps per frame for very fast moving liquid or when volume conservation was critical, but most of the time, only one timestep per frame was computed. Despite the fact that many parameters control the simulation, tweaking uid behavior is intuitive and easy. Our typical parameter values for the most important parameters are t = 1 (where the time unit is 1/30 second), 0 = 10, k = 0.004, knear = 0.01, kspring = 0.3, and = 0.3. The interaction range h can be specied independently to set the uid resolution. 8. Conclusion We have presented a particle-based method to interactively simulate the complex behavior of viscoelastic uids. The main contributions of our work are : 1. A simple and exible method to simulate viscoelasticity with varying rest length springs. Inserting, removing, and modifying the rest length of springs between particles provide an intuitive control of viscoelasticity. Complex non-linear plastic effects such as clay or gel manipulation can be achieved by simple rest length update rules. 2. A new scheme for the robust simulation of surface tension in particle-based systems, without requiring computation of curvature over the liquid surface. In contrast to typical particle systems, it can produce a smooth surface, yet enables the formation of coherent features such as liquid drops, laments, and sheets. 3. A stable method for long timesteps. Numerical instabilities that often plague physically-based simulations are signicantly reduced, and fast animations can be obtained by computing a single timestep per frame. The application of our method was geared mostly towards interactive uid jets, but we demonstrated its exibility in a number of different situations, such as plastic and elastic
c The Eurographics Association 2005.

S. Clavet, P. Beaudoin, and P. Poulin / Particle-based Viscoelastic Fluid Simulation

[GBF03] G UENDELMAN E., B RIDSON R., F EDKIW R.: Nonconvex rigid bodies with stacking. In SIGGRAPH (2003), pp. 871878. [GBO04] G OKTEKIN T. G., BARGTEIL A. W., OB RIEN J. F.: A method for animating viscoelastic uids. In SIGGRAPH (2004), pp. 463468. [GM77] G INGOLD R., M ONAGHAN J.: Smoothed particle hydrodynamics theory and application to nonspherical stars. Monthly Notices of the Royal Astronomical Society 181 (1977), 375. [KO96] KOSHIZUKA S., O KA Y.: Moving-particle semiimplicit method for fragmentation of incompressible uid. Nuclear Science Engineering 123 (July 1996), 421 434. [LC87] L ORENSEN W. E., C LINE H. E.: Marching cubes: A high resolution 3D surface construction algorithm. In SIGGRAPH (1987), pp. 163169. [Luc77] L UCY L.: A numerical approach to the testing of the ssion hypothesis. Astronomical Journal 82 (1977), 1013. [MCG03] M LLER M., C HARYPAR D., G ROSS M.: Particle-based uid simulation for interactive applications. In SIGGRAPH/Eurographics Symposium on Computer Animation (2003), pp. 154159. [MKN 04] M LLER M., K EISER R., N EALEN A., PAULY M., G ROSS M., A LEXA M.: Point based animation of elastic, plastic and melting objects. In SIGGRAPH/Eurographics Symposium on Computer Animation (2004), pp. 141151. [Mor00] M ORRIS J. P.: Simulating surface tension with smoothed particle hydrodynamics. International Journal for Numerical Methods in Fluids 33, 3 (2000), 333353. [MP89] M ILLER G., P EARCE A.: Globular dynamics: A connected particle system for animating viscous uids. Computers & Graphics 13, 3 (1989), 305309. [PIX]

M UELLER M., P OMERANETS D., G ROSS M.: Optimized spatial hashing for collision detection of deformable objects. In Vision, Modeling, and Visualization (2003), pp. 4754. [TPF89] T ERZOPOULOS D., P LATT J., F LEISCHER K.: Heating and melting deformable models (from goop to glop). In Graphics Interface (1989), pp. 219226. [VID] www.iro.umontreal.ca/labs/infographie/papers

Appendix Prediction-Relaxation Scheme The integration scheme we use can be written as x = xn + t vn 1

2

x x + t 2 F1 (x )/m x x + t 2 F2 (x )/m x x + t 2 F3 (x )/m ... xn+1 = x vn+ 1 = (xn+1 xn )/t.

2

This sequential application of force components is analogous to operator splitting often used in Eulerian simulations [Sta99]. Each stage of force application uses the previously updated position, which can lead to greater stability. For example, it is intuitively correct to compute collisions using positions that have already been modied by other constraints. Allowing a particular constraint to adjust to what happened in previous stages often prevents overshoots and erroneous constraint responses. If only one force is used, the integration scheme can be simplied to vn+ 1 = vn 1 + t F(xn + t vn 1 )/m xn+1 = xn + t vn+ 1
2 2 2 2

sourceforge.net/projects/pixie

[PTB 03] P REMOE S., TASDIZEN T., B IGLER J., L EFOHN A., W HITAKER R. T.: Particle-based simulation of uids. Computer Graphics Forum 22, 3 (2003), 401410. [Ree83] R EEVES W. T.: Particle systems a technique for modelling a class of fuzzy objects. In SIGGRAPH (1983), pp. 359376. [SCED04] S TEELE K., C LINE D., E GBERT P. K., D IN J.: Modeling and rendering viscous liquids. Journal of Computer Animation and Virtual Worlds 15, 3-4 (2004), 183192.
ERSTEIN

which is similar to the usual leap-frog scheme, but with the force computed at the predicted position. Testing the canonical example F = k x leads to vn+ 1 = vn 1 + t (k xn k t vn 1 )/m
2 2 2

in which the actual force being computed is augmented with the term k t vn 1 , an articial viscosity.
2

[Sta99] S TAM J.: Stable uids. In SIGGRAPH (1999), pp. 121128. [THM 03] T ESCHNER M., H EIDELBERGER B.,
c The Eurographics Association 2005.

S. Clavet, P. Beaudoin, and P. Poulin / Particle-based Viscoelastic Fluid Simulation

Liquid jet on a dynamic character.

Fast detailed splash.

A polygonal mold is lled with liquid, springs are added, and then the mold is removed. The resulting elastic bunny is stepped on, but recovers its rest shape.

Then plasticity is increased and the foot leaves its print. Finally, springs are deleted and the bunny ows away. Figure 12: Various illustrations of our particle-based fast uid simulation.

c The Eurographics Association 2005.