Basic fluids

767 views

Published on

Published in: Technology
1 Comment
0 Likes
Statistics
Notes
  • Be the first to like this

No Downloads
Views
Total views
767
On SlideShare
0
From Embeds
0
Number of Embeds
3
Actions
Shares
0
Downloads
36
Comments
1
Likes
0
Embeds 0
No embeds

No notes for slide

Basic fluids

  1. 1. F LUID S IMULATION Robert Bridson, UBC Matthias Müller-Fischer, AGEIA Inc.
  2. 2. Course Details <ul><li>www.cs.ubc.ca/~rbridson/fluidsimulation </li></ul><ul><li>Schedule: </li></ul><ul><ul><li>1:45 - 3:30 Basics [Bridson] </li></ul></ul><ul><ul><li>3:30 - 3:45 Break </li></ul></ul><ul><ul><li>3:45 - 4:45 Real-Time [M-F] </li></ul></ul><ul><ul><li>4:45 - 5:30 Extras [Bridson] </li></ul></ul>
  3. 3. The Basic Equations
  4. 4. Symbols <ul><li>: velocity with components (u,v,w) </li></ul><ul><li> : fluid density </li></ul><ul><li>p: pressure </li></ul><ul><li>: acceleration due to gravity or animator </li></ul><ul><li> : dynamic viscosity </li></ul>
  5. 5. The Equations <ul><li>Incompressible Navier-Stokes: </li></ul><ul><ul><li>“ Momentum Equation” </li></ul></ul><ul><ul><li>“ Incompressibility condition” </li></ul></ul>
  6. 6. The Momentum Equation
  7. 7. The Momentum Equation <ul><li>Just a specialized version of F=ma </li></ul><ul><li>Let’s build it up intuitively </li></ul><ul><li>Imagine modeling a fluid with a bunch of particles (e.g. blobs of water) </li></ul><ul><ul><li>A blob has a mass m, a volume V, velocity u </li></ul></ul><ul><ul><li>We’ll write the acceleration as Du/Dt (“material derivative”) </li></ul></ul><ul><ul><li>What forces act on the blob? </li></ul></ul>
  8. 8. Forces on Fluids <ul><li>Gravity: mg </li></ul><ul><ul><li>or other “body forces” designed by animator </li></ul></ul><ul><li>And a blob of fluid also exerts contact forces on its neighbouring blobs… </li></ul>
  9. 9. Pressure <ul><li>The “normal” contact force is pressure (force/area) </li></ul><ul><ul><li>How blobs push against each other, and how they stick together </li></ul></ul><ul><li>If pressure is equal in every direction, net force is zero… Important quantity is pressure gradient : </li></ul><ul><ul><li>We integrate over the blob’s volume (equivalent to integrating -pn over blob’s surface) </li></ul></ul><ul><li>What is the pressure? Coming soon… </li></ul>
  10. 10. Viscosity <ul><li>Think of it as frictional part of contact force: a sticky (viscous) fluid blob resists other blobs moving past it </li></ul><ul><li>For now, simple model is that we want velocity to be blurred/diffused/… </li></ul><ul><li>Blurring in PDE form comes from the Laplacian: </li></ul>
  11. 11. The Continuum Limit (1) <ul><li>Model the world as a continuum: </li></ul><ul><ul><li># particles  Mass and volume  0 </li></ul></ul><ul><li>We want F=ma to be more than 0=0: </li></ul><ul><ul><li>Divide by mass </li></ul></ul>
  12. 12. The Continuum Limit (2) <ul><li>The fluid density is  =m/V: </li></ul><ul><li>This is almost the same as the stanford eq’n (in fact, the form we mostly use!) </li></ul><ul><li>The only weird thing is Du/Dt… </li></ul>
  13. 13. Lagrangian vs. Eulerian <ul><li>Lagrangian viewpoint: </li></ul><ul><ul><li>Treat the world like a particle system </li></ul></ul><ul><ul><li>Label each speck of matter, track where it goes (how fast, acceleration, etc.) </li></ul></ul><ul><li>Eulerian viewpoint: </li></ul><ul><ul><li>Fix your point in space </li></ul></ul><ul><ul><li>Measure stuff as it flows past </li></ul></ul><ul><li>Think of measuring the temperature: </li></ul><ul><ul><li>Lagrangian: in a balloon, floating with the wind </li></ul></ul><ul><ul><li>Eulerian: on the ground, wind blows past </li></ul></ul>
  14. 14. The Material Derivative (1) <ul><li>We have fluid moving in a velocity field u </li></ul><ul><li>It possesses some quality q </li></ul><ul><li>At an instant in time t and a position in space x, the fluid at x has q(x,t) </li></ul><ul><ul><li>q(x,t) is an Eulerian field </li></ul></ul><ul><li>How fast is that blob of fluid’s q changing? </li></ul><ul><ul><li>A Lagrangian question </li></ul></ul><ul><li>Answer: the material derivative Dq/Dt </li></ul>
  15. 15. The Material Derivative (2) <ul><li>It all boils down to the chain rule: </li></ul><ul><li>We usually rearrange it: </li></ul>
  16. 16. Turning Dq/Dt Around <ul><li>For a thought experiment, turn it around: </li></ul><ul><li>That is, how fast q is changing at a fixed point in space (∂q/∂t) comes from two things: </li></ul><ul><ul><li>How fast q is changing for the blob of fluid at x </li></ul></ul><ul><ul><li>How fast fluid with different values of q is flowing past </li></ul></ul>
  17. 17. Writing D/Dt Out <ul><li>We can explicitly write it out from components: </li></ul>
  18. 18. D/Dt For Vector Fields <ul><li>Say our fluid has a colour variable (RGB vector) C </li></ul><ul><li>We still write </li></ul><ul><ul><li>The dot-product and gradient confusing? </li></ul></ul><ul><ul><li>Just do it component-by-component: </li></ul></ul>
  19. 19. Du/Dt <ul><li>This holds even if the vector field is velocity itself: </li></ul><ul><li>Nothing different about this, just that the fluid blobs are moving at the velocity they’re carrying. </li></ul>
  20. 20. The Incompressibility Condition
  21. 21. Compressibility <ul><li>Real fluids are compressible </li></ul><ul><li>Shock waves, acoustic waves, pistons… </li></ul><ul><ul><li>Note: liquids change their volume as well as gases, otherwise there would be no sound underwater </li></ul></ul><ul><li>But this is nearly irrelevant for animation </li></ul><ul><ul><li>Shocks move too fast to normally be seen (easier/better to hack in their effects) </li></ul></ul><ul><ul><li>Acoustic waves usually have little effect on visible fluid motion </li></ul></ul><ul><ul><li>Pistons are boring </li></ul></ul>
  22. 22. Incompressibility <ul><li>Rather than having to simulate acoustic and shock waves, eliminate them from our model: assume fluid is incompressible </li></ul><ul><ul><li>Turn stiff system into a constraint, just like rigid bodies! </li></ul></ul><ul><li>If you fix your eyes on any volume of space, volume of fluid in = volume of fluid out: </li></ul>
  23. 23. Divergence <ul><li>Let’s use the divergence theorem: </li></ul><ul><li>So for any region, the integral of is zero </li></ul><ul><li>Thus it’s zero everywhere: </li></ul>
  24. 24. Pressure <ul><li>Pressure p: whatever it takes to make the velocity field divergence free </li></ul><ul><li>If you know constrained dynamics, is a constraint, and pressure is the matching Lagrange multiplier </li></ul><ul><li>Our simulator will follow this approach: </li></ul><ul><ul><li>solve for a pressure that makes our fluid incompressible at each time step. </li></ul></ul>
  25. 25. Alternative <ul><li>To avoid having to solve linear systems, can turn this into a soft-constraint </li></ul><ul><ul><li>Track density, make pressure proportional to density variation </li></ul></ul><ul><ul><li>Called “artificial compressibility” </li></ul></ul><ul><li>However, easier to accelerate a linear solver… if the linear system is well-posed </li></ul>
  26. 26. Inviscid Fluids
  27. 27. Dropping Viscosity <ul><li>In most scenarios, viscosity term is much smaller </li></ul><ul><li>Convenient to drop it from the equations: </li></ul><ul><ul><li>Zero viscosity = “inviscid” </li></ul></ul><ul><ul><li>Inviscid Navier-Stokes = “Euler equations” </li></ul></ul><ul><li>Numerical simulation typically makes errors that resemble physical viscosity, so we have the visual effect of it anyway </li></ul><ul><ul><li>Called “numerical dissipation” </li></ul></ul><ul><ul><li>For animation: often numerical dissipation is larger than the true physical viscosity! </li></ul></ul>
  28. 28. Aside: A Few Figures <ul><li>Dynamic viscosity of air: </li></ul><ul><li>Density of air: </li></ul><ul><li>Dynamic viscosity of water: </li></ul><ul><li>Density of water: </li></ul><ul><li>The ratio,  /  (“kinematic viscosity”) is what’s important for the motion of the fluid… … air is 10 times more viscous than water! </li></ul>
  29. 29. The inviscid equations <ul><li>A.k.a. the incompressible Euler equations: </li></ul>
  30. 30. Boundary Conditions
  31. 31. Boundary Conditions <ul><li>We know what’s going on inside the fluid: what about at the surface? </li></ul><ul><li>Three types of surface </li></ul><ul><ul><li>Solid wall: fluid is adjacent to a solid body </li></ul></ul><ul><ul><li>Free surface: fluid is adjacent to nothing (e.g. water is so much denser than air, might as well forget about the air) </li></ul></ul><ul><ul><li>Other fluid: possibly discontinuous jump in quantities (density, …) </li></ul></ul>
  32. 32. Solid Wall Boundaries <ul><li>No fluid can enter or come out of a solid wall: </li></ul><ul><li>For common case of u solid =0: </li></ul><ul><li>Sometimes called the “no-stick” condition, since we let fluid slip past tangentially </li></ul><ul><ul><li>For viscous fluids, can additionally impose “no-slip” condition: </li></ul></ul>
  33. 33. Free Surface <ul><li>Neglecting the other fluid, we model it simply as pressure=constant </li></ul><ul><ul><li>Since only pressure gradient is important, we can choose the constant to be zero: </li></ul></ul><ul><li>If surface tension is important (not covered today), pressure is instead related to mean curvature of surface </li></ul>
  34. 34. Multiple Fluids <ul><li>At fluid-fluid boundaries, the trick is to determine “jump conditions” </li></ul><ul><ul><li>For a fluid quantity q, [q]=q 1 -q 2 </li></ul></ul><ul><li>Density jump [  ] is known </li></ul><ul><li>Normal velocity jump must be zero: </li></ul><ul><li>For inviscid flow, tangential velocities may be unrelated (jump is unknown) </li></ul><ul><li>With no surface tension, pressure jump [p]=0 </li></ul>
  35. 35. Numerical Simulation Overview
  36. 36. Splitting <ul><li>We have lots of terms in the momentum equation: a pain to handle them all simultaneously </li></ul><ul><li>Instead we split up the equation into its terms, and integrate them one after the other </li></ul><ul><ul><li>Makes for easier software design too: a separate solution module for each term </li></ul></ul><ul><li>First order accurate in time </li></ul><ul><ul><li>Can be made more accurate, not covered today. </li></ul></ul>
  37. 37. A Splitting Example <ul><li>Say we have a differential equation </li></ul><ul><li>And we can solve the component parts: </li></ul><ul><ul><li>SolveF(q,∆t) solves dq/dt=f(q) for time ∆t </li></ul></ul><ul><ul><li>SolveG(q,∆t) solves dq/dt=g(q) for time ∆t </li></ul></ul><ul><li>Put them together to solve the full thing: </li></ul><ul><ul><li>q* = SolveF(q n , ∆t) </li></ul></ul><ul><ul><li>q n+1 = SolveG(q*, ∆t) </li></ul></ul>
  38. 38. Does it Work? <ul><li>Up to O(∆t): </li></ul>
  39. 39. Splitting Momentum <ul><li>We have three terms: </li></ul><ul><li>First term: advection </li></ul><ul><ul><li>Move the fluid through its velocity field (Du/Dt=0) </li></ul></ul><ul><li>Second term: gravity </li></ul><ul><li>Final term: pressure update </li></ul><ul><ul><li>How we’ll make the fluid incompressible: </li></ul></ul>
  40. 40. Space <ul><li>That’s our general strategy in time; what about space? </li></ul><ul><li>We’ll begin with a fixed Eulerian grid </li></ul><ul><ul><li>Trivial to set up </li></ul></ul><ul><ul><li>Easy to approximate spatial derivatives </li></ul></ul><ul><ul><li>Particularly good for the effect of pressure </li></ul></ul><ul><li>Disadvantage: advection doesn’t work so well </li></ul><ul><ul><li>Later: particle methods that fix this </li></ul></ul>
  41. 41. A Simple Grid <ul><li>We could put all our fluid variables at the nodes of a regular grid </li></ul><ul><li>But this causes some major problems </li></ul><ul><li>In 1D: incompressibility means </li></ul><ul><li>Approximate at a grid point: </li></ul><ul><li>Note the velocity at the grid point isn’t involved! </li></ul>
  42. 42. A Simple Grid Disaster <ul><li>The only solutions to are u=constant </li></ul><ul><li>But our numerical version has other solutions: </li></ul>u x
  43. 43. Staggered Grids <ul><li>Problem is solved if we don’t skip over grid points </li></ul><ul><li>To make it unbiased, we stagger the grid: put velocities halfway between grid points </li></ul><ul><li>In 1D, we estimate divergence at a grid point as: </li></ul><ul><li>Problem solved! </li></ul>
  44. 44. The MAC Grid <ul><li>From the Marker-and-Cell (MAC) method [Harlow&Welch’65] </li></ul><ul><li>A particular staggering of variables in 2D/3D that works well for incompressible fluids: </li></ul><ul><ul><li>Grid cell (i,j,k) has pressure p i,j,k at its center </li></ul></ul><ul><ul><li>x-part of velocity u i+1/2,jk in middle of x-face between grid cells (i,j,k) and (i+1,j,k) </li></ul></ul><ul><ul><li>y-part of velocity v i,j+1/2,k in middle of y-face </li></ul></ul><ul><ul><li>z-part of velocity w i,j,k+1/2 in middle of z-face </li></ul></ul>
  45. 45. MAC Grid in 2D
  46. 46. Array storage <ul><li>Then for a nx  ny  nz grid, we store them as 3D arrays: </li></ul><ul><ul><li>p[nx, ny, nz] </li></ul></ul><ul><ul><li>u[nx+1, ny, nz] </li></ul></ul><ul><ul><li>v[nx, ny+1, nz] </li></ul></ul><ul><ul><li>w[nx, ny, nz+1] </li></ul></ul><ul><li>And translate indices in code, e.g. </li></ul>
  47. 47. The downside <ul><li>Not having all quantities at the same spot makes some algorithms a pain </li></ul><ul><ul><li>Even interpolation of velocities for pushing particles is annoying </li></ul></ul><ul><li>One strategy: switch back and forth (colocated/staggered) by averaging </li></ul><ul><li>My philosophy: avoid averaging as much as possible! </li></ul>
  48. 48. Advection Algorithms
  49. 49. Advecting Quantities <ul><li>The goal is to solve “the advection equation” for any grid quantity q </li></ul><ul><ul><li>in particular, the components of velocity </li></ul></ul><ul><li>Rather than directly approximate spatial term, we’ll use the Lagrangian notion of advection directly </li></ul><ul><li>We’re on an Eulerian grid, though, so the result will be called “semi-Lagrangian” </li></ul><ul><ul><li>Introduced to graphics by [Stam’99] </li></ul></ul>
  50. 50. Semi-Lagrangian Advection <ul><li>Dq/Dt=0 says q doesn’t change as you follow the fluid. </li></ul><ul><li>So </li></ul><ul><li>We want to know q at each grid point at t new (that is, we’re interested in x new =x ijk ) </li></ul><ul><li>So we just need to figure out x old (where fluid at x ijk came from) and q(x old ) (what value of q was there before) </li></ul>
  51. 51. Finding x old <ul><li>We need to trace backwards through the velocity field. </li></ul><ul><li>Up to O(∆t) we can assume velocity field constant over the time step </li></ul><ul><li>The simplest estimate is then </li></ul><ul><ul><li>I.e. tracing through the time-reversed flow with one step of Forward Euler </li></ul></ul><ul><ul><li>Other ODE methods can (and should ) be used </li></ul></ul>
  52. 52. Aside: ODE methods <ul><li>Chief interesting aspect of fluid motion is vortices: rotation </li></ul><ul><li>Any ODE integrator we use should handle rotation reasonably well </li></ul><ul><li>Forward Euler does not </li></ul><ul><ul><li>Instability, obvious spiral artifacts </li></ul></ul><ul><li>Runge-Kutta 2 is simplest half-decent method </li></ul>
  53. 53. Details of x old <ul><li>We need for a quantity stored at (i,j,k) </li></ul><ul><ul><li>For staggered quantities, say u, we need it at the staggered location: e.g. </li></ul></ul><ul><li>We can get the velocity there by averaging the appropriate staggered components: e.g. </li></ul>
  54. 54. Which u to Use? <ul><li>Note that we only want to advect quantities in an incompressible velocity field </li></ul><ul><ul><li>Otherwise the quantity gets compressed (often an obvious unphysical artifact) </li></ul></ul><ul><li>For example, when we advect u, v, and w themselves, we use the old incompressible values stored in the grid </li></ul><ul><ul><li>Do not update as you go! </li></ul></ul>
  55. 55. Finding q(x old ) <ul><li>Odds are when we trace back from a grid point to x old we won’t land on a grid point </li></ul><ul><ul><li>So we don’t have an old value of q there </li></ul></ul><ul><li>Solution: interpolate from nearby grid points </li></ul><ul><ul><li>Simplest method: bi/trilinear interpolation </li></ul></ul><ul><ul><li>Know your grid: be careful to get it right for staggered quantities! </li></ul></ul>
  56. 56. Boundary Conditions <ul><li>What if x old isn’t in the fluid? (or a nearest grid point we’re interpolating from is not in the fluid?) </li></ul><ul><li>Solution: extrapolate from boundary of fluid </li></ul><ul><ul><li>Extrapolate before advection, to all grid points in the domain that aren’t fluid </li></ul></ul><ul><li>ALSO: if fluid can move to a grid point that isn’t fluid now, make sure to do semi-Lagrangian advection there </li></ul><ul><ul><li>Use the extrapolated velocity </li></ul></ul>
  57. 57. Extrapolating u Into Solids <ul><li>Note that the inviscid “no-stick” boundary condition just says </li></ul><ul><ul><li>The tangential component of velocity doesn’t have to be equal! </li></ul></ul><ul><ul><li>So we need to extrapolate into solids, and can’t use the solid’s own velocity (result would be bad numerical viscosity) </li></ul></ul>
  58. 58. Extrapolation <ul><li>One of the trickiest bits of code to get right </li></ul><ul><li>For sanity check, be careful about what’s known and unknown on the grid </li></ul><ul><ul><li>Replace all unknowns with linear combinations of knowns </li></ul></ul><ul><li>More on this later (level sets) </li></ul>
  59. 59. Body Forces
  60. 60. Integrating Body Forces <ul><li>Gravity vector or volumetric animator forces: </li></ul><ul><li>Simplest scheme: at every grid point just add </li></ul>
  61. 61. Making Fluids Incompressible
  62. 62. The Continuous Version <ul><li>We want to find a pressure p so that the updated velocity: is divergence free: while respecting the boundary conditions: </li></ul>
  63. 63. The Poisson Problem <ul><li>Plug in the pressure update formula into incompressibility: </li></ul><ul><li>Turns into a “Poisson equation” for pressure: </li></ul>
  64. 64. Discrete Pressure Update <ul><li>The discrete pressure update on the MAC grid: </li></ul>
  65. 65. Discrete Divergence <ul><li>The discretized incompressibility condition on the new velocity (estimated at i,j,k): </li></ul>
  66. 66. Discrete Boundaries <ul><li>For now, let’s voxelize the geometry: each grid cell is either fluid, solid, or empty </li></ul>S S S E S S F S S E E F F F F F F F
  67. 67. What’s Active? <ul><li>Pressure: we’ll only solve for p in Fluid cells </li></ul><ul><li>Velocity: anything bordering a Fluid cell is good </li></ul><ul><li>Boundary conditions: </li></ul><ul><ul><li>p=0 in Empty cells (free surface) </li></ul></ul><ul><ul><li>p=? whatever is needed in Solid cells so that the pressure update makes (Note normal is to grid cell!) (Also note: different implied pressures for each active face of a solid cell - we don’t store p there…) </li></ul></ul>
  68. 68. Example Solid Wall <ul><li>(i-1,j,k) is solid, (i,j,k) is fluid </li></ul><ul><li>Normal is n=(1,0,0) </li></ul><ul><li>Want </li></ul><ul><li>The pressure update formula is: </li></ul><ul><li>So the “ghost” solid pressure is: </li></ul>S F
  69. 69. Discrete Pressure Equations <ul><li>Substitute in pressure update formula to discrete divergence </li></ul><ul><li>In each fluid cell (i,j,k) get an equation: </li></ul>
  70. 70. Simplified <ul><li>Collecting terms: </li></ul><ul><li>Empty neighbors: drop zero pressures </li></ul><ul><li>Solid neighbors: e.g. ( i-1, j, k) is solid </li></ul><ul><ul><li>Drop p i-1jk - reduce coeff of p ijk by 1 (to 5…) </li></ul></ul><ul><ul><li>Replace u i-1/2jk in right-hand side with u solid </li></ul></ul>
  71. 71. Linear Equations <ul><li>End up with a sparse set of linear equations to solve for pressure </li></ul><ul><ul><li>Matrix is symmetric positive (semi-)definite </li></ul></ul><ul><li>In 3D on large grids, direct methods unsatisfactory </li></ul><ul><li>Instead use Preconditioned Conjugate Gradient, with Incomplete Cholesky preconditioner </li></ul><ul><li>See course notes for full details (pseudo-code) </li></ul><ul><li>Residual is how much divergence there is in u n+1 </li></ul><ul><ul><li>Iterate until satisfied it’s small enough </li></ul></ul>
  72. 72. Voxelization is Suboptimal <ul><li>Free surface artifacts: </li></ul><ul><ul><li>Waves less than a grid cell high aren’t “seen” by the fluid solver – thus they don’t behave right </li></ul></ul><ul><ul><li>Left with strangely textured surface </li></ul></ul><ul><li>Solid wall artifacts: </li></ul><ul><ul><li>If boundary not grid-aligned, O(1) error – it doesn’t even go away as you refine! </li></ul></ul><ul><ul><li>Slopes are turned into stairs, water will pool on artificial steps. </li></ul></ul><ul><li>More on this later… </li></ul>
  73. 73. Smoke As per [Fedkiw et al.’01]
  74. 74. Soot <ul><li>Smoke is composed of soot particles suspended in air </li></ul><ul><li>For simulation, need to track concentration of soot on the grid: s(x,t) </li></ul><ul><li>Evolution equation: </li></ul><ul><ul><li>Extra term is diffusion (if k≠0) </li></ul></ul><ul><ul><li>Simplest implementation: Gaussian blur </li></ul></ul><ul><li>Boundary conditions: </li></ul><ul><ul><li>At solid wall: ds/dn=0 (extrapolate from fluid) </li></ul></ul><ul><ul><li>At smoke source: s=s source </li></ul></ul><ul><li>Note: lots of particles may be better for rendering… </li></ul>
  75. 75. Heat <ul><li>Usually smoke is moving because the air is hot </li></ul><ul><ul><li>Hot air is less dense; pressure vs. gravity ends up in buoyancy (hot air rising) </li></ul></ul><ul><li>Need to track temperature T(x,t) </li></ul><ul><li>Evolution equation (neglecting conduction) </li></ul><ul><li>Boundary conditions: conduction of heat between solids and fluids </li></ul><ul><ul><li>Extrapolate T from fluid, add source term… </li></ul></ul>
  76. 76. Boussinesq and Buoyancy <ul><li>Density variation due to temperature or soot concentration is very small </li></ul><ul><li>Use the “Boussinesq approximation”: fix density=constant, but replace gravity in momentum equation with a buoyancy force: </li></ul><ul><li>Constants  and  can be tuned </li></ul>
  77. 77. Open Boundaries <ul><li>If open air extends beyond the grid, what boundary condition? </li></ul><ul><li>After making the Boussinesq approximation, simplest thing is to say the open air is a free surface (p=0) </li></ul><ul><ul><li>We let air blow in or out as it wants </li></ul></ul><ul><li>May want to explicitly zero velocity on far boundaries to avoid large currents developing </li></ul>
  78. 78. Water
  79. 79. Where is the Water? <ul><li>Water is often modeled as a fluid with a free surface </li></ul><ul><li>The only thing we’re missing so far is how to track where the water is </li></ul><ul><ul><li>Voxelized: which cells are fluid vs. empty? </li></ul></ul><ul><ul><li>Better: where does air/water interface cut through grid lines? </li></ul></ul>
  80. 80. Marker Particles <ul><li>Simplest approach is to use marker particles </li></ul><ul><li>Sample fluid volume with particles {x p } </li></ul><ul><li>At each time step, mark grid cells containing any marker particles as Fluid, rest as Empty or Solid </li></ul><ul><li>Move particles in incompressible grid velocity field (interpolating as needed): </li></ul>
  81. 81. Rendering Marker Particles <ul><li>Need to wrap a surface around the particles </li></ul><ul><ul><li>E.g. blobbies, or Zhu & Bridson ‘05 </li></ul></ul><ul><li>Problem: rendered surface has bumpy detail that the fluid solver doesn’t have access to </li></ul><ul><ul><li>The simulation can’t animate that detail properly if it can’t “see” it. </li></ul></ul><ul><li>Result: looks fine for splashy, noisy surfaces, but fails on smooth, glassy surfaces. </li></ul>
  82. 82. Improvement <ul><li>Sample implicit surface function on simulation grid </li></ul><ul><ul><li>Even just union of spheres is reasonable if we smooth and render from the grid </li></ul></ul><ul><li>Note: make sure that a single particle always shows up on grid (e.g. radius ≥ ∆x) or droplets can flicker in and out of existence… </li></ul><ul><ul><li>If not: you must delete stray particles </li></ul></ul><ul><li>But still not perfect for flat fluid surfaces. </li></ul>
  83. 83. Level Sets <ul><li>Idea: drop the particles all together, and just sample the implicit surface function on the grid </li></ul><ul><li>In particular, best behaved type of implicit surface function is signed distance: </li></ul><ul><li>Surface is exactly where  =0 </li></ul><ul><li>See e.g. the book [Osher&Fedkiw’02] </li></ul>
  84. 84. Surface Capturing <ul><li>We need to evolve  on the grid </li></ul><ul><li>The surface (  =0) moves at the velocity of the fluid (any fluid particle with  =0 should continue with  =0) </li></ul><ul><li>We can use same equation for rest of volume (doesn’t really matter what we do as long as sign stays the same – at least in continuous variables!) </li></ul>
  85. 85. Reinitializing <ul><li>One problem is that advecting  this way doesn’t preserve signed distance property </li></ul><ul><ul><li>Eventually gets badly distorted </li></ul></ul><ul><ul><li>Causes problems particularly for extrapolation, also for accuracy in general </li></ul></ul><ul><li>Reinitialize: recompute distance to surface </li></ul><ul><ul><li>Ideally shouldn’t change surface at all, just values in the volume </li></ul></ul>
  86. 86. Computing Signed Distance <ul><li>Many algorithms exist </li></ul><ul><li>Simplest and most robust (and accurate enough for animation) are geometric in nature </li></ul><ul><li>Our example algorithm is based on a Fast Sweeping method [Tsai’02] </li></ul>
  87. 87. Set Up <ul><li>Start by identifying grid cells where  old changes sign (where the surface is) </li></ul><ul><li>Determine points and/or mesh elements to approximate surface </li></ul><ul><ul><li>Set nearby grid values to exact distance to closest surface element </li></ul></ul><ul><ul><li>Also store index of closest surface element </li></ul></ul><ul><li>Every other grid point is reset to  = ±L and index of closest surface element = UNKNOWN </li></ul><ul><li>We then propagate information out to these points </li></ul>
  88. 88. Fast Sweeping <ul><li>One sweep: loop i,j,k through the grid: </li></ul><ul><ul><li>Check neighbors for their closest surface element </li></ul></ul><ul><ul><li>If the closest of those is closer than |  ijk | then update  ijk and set its index to that point </li></ul></ul><ul><li>Idea: keep sweeping in different directions </li></ul><ul><ul><li>++i, ++j, ++k </li></ul></ul><ul><ul><li>--i, ++j, ++k </li></ul></ul><ul><ul><li>++i, --j, ++k </li></ul></ul><ul><ul><li>--i, --j, ++k </li></ul></ul><ul><ul><li>++i, ++j, --k </li></ul></ul><ul><ul><li>etc … </li></ul></ul>
  89. 89. Extrapolation <ul><li>We often need fluid values extrapolated outside of the fluid </li></ul><ul><ul><li>e.g. velocity must be extrapolated out from free surface and into solid </li></ul></ul><ul><li>With signed distance this is much simpler </li></ul>
  90. 90. Extrapolation (1) <ul><li>Interpolate fluid quantities onto stored surface points (generated in redistancing) </li></ul><ul><ul><li>Make sure to only use known quantities, not values outside the fluid that the physics doesn’t touch </li></ul></ul><ul><li>Then use index-of-closest-element to copy values everywhere needed </li></ul>
  91. 91. Extrapolation (2) <ul><li>Gradient of distance points parallel to ray to closest surface point </li></ul><ul><li>Phrase extrapolation as a PDE: </li></ul><ul><li>Then solve the PDE with finite differences </li></ul><ul><ul><li>Careful “upwind” scheme needed (want to get information from points with smaller distance, not bigger) </li></ul></ul>

×