How to write an Eulerian Fluid Simulator

with 200 lines of JavaScript code

Matthias Müller, Ten Minute Physics
 For the code and the demo see:

www.matthiasmueller.info/tenMinutePhysics
a few remarks…
Fluid =

or

liquid gas

Similar physical structures → same simulation method

Eulerian
𝑒𝑒 𝑖𝑖𝜋𝜋 = −1

Swiss Italian

Leonhard Euler (1707-1783) Joseph-Louis Lagrange (1736-1813)

grid-based grid-free (e. g. particles)

From 2d to 3d

straight forward - left as an exercise

Incompressible Fluid

• We assume incompressibility

• Water is very close to being incompressible

10,000 kg / cm2 → 3% compression!

• Reasonable assumption for free gas as well

→ compressible simulation in upcoming tutorial

Inviscid Fluid

• We assume zero viscosity

• Good approximation for gas and water

inviscid viscous
→ viscous simulation in upcoming tutorial
the method…
Fluid as a Velocity Field on a Grid
𝑢𝑢
• Velocity is a 2d vector 𝐯𝐯 = (vectors are written in boldface)
𝑣𝑣

𝐯𝐯𝑖𝑖,𝑗𝑗
𝑣𝑣𝑖𝑖,𝑗𝑗
ℎ
𝑢𝑢𝑖𝑖,𝑗𝑗 𝑢𝑢𝑖𝑖,𝑗𝑗
𝑣𝑣𝑖𝑖,𝑗𝑗

ℎ
collocated grid staggered grid
Fluid Simulation

Modify velocity values (e.g. add gravity)

Make the fluid incompressible (projection)

Move the velocity field (advection)

Update Velocity

for all 𝑖𝑖, 𝑗𝑗

𝑣𝑣𝑖𝑖,𝑗𝑗 ← 𝑣𝑣𝑖𝑖,𝑗𝑗 + ∆𝑡𝑡 � 𝑔𝑔

• Gravity 𝑔𝑔: −9.81 𝑚𝑚/𝑠𝑠 2

1
• Timestep ∆𝑡𝑡: ( e. g. 𝑠𝑠 )
30
Divergence (Total Outflow)
𝒅𝒅 ← 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 − 𝒖𝒖𝒊𝒊,𝒋𝒋 + 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 − 𝒗𝒗𝒊𝒊,𝒋𝒋

𝑣𝑣𝑖𝑖,𝑗𝑗+1

𝑢𝑢𝑖𝑖𝑖𝑖 𝑢𝑢𝑖𝑖+1,𝑗𝑗

𝑣𝑣𝑖𝑖𝑖𝑖

• Positive • Negative • Zero

• Too much outflow • Too much inflow • Incompressible
Forcing Incompressibility

Too much inflow Changing one velocity Push all velocities outward
A fluid cannot do that! by the same amount
 Forcing Incompressibility
𝑣𝑣𝑖𝑖,𝑗𝑗+1
𝒅𝒅 ← 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 − 𝒖𝒖𝒊𝒊,𝒋𝒋 + 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 − 𝒗𝒗𝒊𝒊,𝒋𝒋

𝒖𝒖𝒊𝒊,𝒋𝒋 ← 𝒖𝒖𝒊𝒊,𝒋𝒋 + 𝒅𝒅/𝟒𝟒

𝑢𝑢𝑖𝑖,𝑗𝑗 𝑢𝑢𝑖𝑖+1,𝑗𝑗
𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 ← 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 − 𝒅𝒅/𝟒𝟒
𝒗𝒗𝒊𝒊,𝒋𝒋 ← 𝒗𝒗𝒊𝒊,𝒋𝒋 + 𝒅𝒅/𝟒𝟒
𝑣𝑣𝑖𝑖,𝑗𝑗
𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 ← 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 − 𝒅𝒅/𝟒𝟒
Obstacles / Walls
𝑣𝑣𝑖𝑖,𝑗𝑗+1
𝒅𝒅 ← 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 − 𝒖𝒖𝒊𝒊,𝒋𝒋 + 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 − 𝒗𝒗𝒊𝒊,𝒋𝒋

𝒖𝒖𝒊𝒊,𝒋𝒋 ← 𝒖𝒖𝒊𝒊,𝒋𝒋 + 𝒅𝒅/𝟒𝟒

𝑢𝑢𝑖𝑖,𝑗𝑗 𝑢𝑢𝑖𝑖+1,𝑗𝑗 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 ← 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 − 𝒅𝒅/𝟑𝟑
𝒗𝒗𝒊𝒊,𝒋𝒋 ← 𝒗𝒗𝒊𝒊,𝒋𝒋 + 𝒅𝒅/𝟑𝟑
𝑣𝑣𝑖𝑖,𝑗𝑗 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 ← 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 − 𝒅𝒅/𝟑𝟑

𝑢𝑢𝑖𝑖,𝑗𝑗 ≠ 0
• Moving object
• Turbine pushing fluid (e. g. wind tunnel)
General Case

𝒅𝒅 ← 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 − 𝒖𝒖𝒊𝒊,𝒋𝒋 + 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 − 𝒗𝒗𝒊𝒊,𝒋𝒋

𝒔𝒔 ← 𝒔𝒔𝒊𝒊+𝟏𝟏,𝒋𝒋 + 𝒔𝒔𝒊𝒊−𝟏𝟏,𝒋𝒋 + 𝒔𝒔𝒊𝒊,𝒋𝒋+𝟏𝟏 + 𝒔𝒔𝒊𝒊,𝒋𝒋−𝟏𝟏
𝑠𝑠𝑖𝑖−1,𝑗𝑗 = 0
𝑠𝑠𝑖𝑖,𝑗𝑗 = 1 𝒖𝒖𝒊𝒊,𝒋𝒋 ← 𝒖𝒖𝒊𝒊,𝒋𝒋 + 𝒅𝒅 𝒔𝒔𝒊𝒊−𝟏𝟏,𝒋𝒋 /𝒔𝒔
𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 ← 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 − 𝒅𝒅 𝒔𝒔𝒊𝒊+𝟏𝟏,𝒋𝒋 /𝒔𝒔
𝒗𝒗𝒊𝒊,𝒋𝒋 ← 𝒗𝒗𝒊𝒊,𝒋𝒋 + 𝒅𝒅 𝒔𝒔𝒊𝒊,𝒋𝒋−𝟏𝟏 /𝒔𝒔
𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 ← 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 − 𝒅𝒅 𝒔𝒔𝒊𝒊,𝒋𝒋+𝟏𝟏 /𝒔𝒔
Solving the Grid
for 𝒏𝒏 iterations • Gauss-Seidel method
for all 𝒊𝒊, 𝒋𝒋 • On the boundary we access
𝒅𝒅 ← 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 − 𝒖𝒖𝒊𝒊,𝒋𝒋 + 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 − 𝒗𝒗𝒊𝒊,𝒋𝒋 cells outside of the grid!

𝒔𝒔 ← 𝒔𝒔𝒊𝒊+𝟏𝟏,𝒋𝒋 + 𝒔𝒔𝒊𝒊−𝟏𝟏,𝒋𝒋 + 𝒔𝒔𝒊𝒊,𝒋𝒋+𝟏𝟏 + 𝒔𝒔𝒊𝒊,𝒋𝒋−𝟏𝟏 • Add border cells

𝒖𝒖𝒊𝒊,𝒋𝒋 ← 𝒖𝒖𝒊𝒊,𝒋𝒋 + 𝒅𝒅 𝒔𝒔𝒊𝒊−𝟏𝟏,𝒋𝒋 /𝒔𝒔 • Set 𝑠𝑠𝑖𝑖,𝑗𝑗 = 0 for walls

𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 ← 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 − 𝒅𝒅 𝒔𝒔𝒊𝒊+𝟏𝟏,𝒋𝒋 /𝒔𝒔 • or copy neighbor values
𝒗𝒗𝒊𝒊,𝒋𝒋 ← 𝒗𝒗𝒊𝒊,𝒋𝒋 + 𝒅𝒅 𝒔𝒔𝒊𝒊,𝒋𝒋−𝟏𝟏 /𝒔𝒔
𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 ← 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 − 𝒅𝒅 𝒔𝒔𝒊𝒊,𝒋𝒋+𝟏𝟏 /𝒔𝒔
Measuring Pressure
𝑝𝑝𝑖𝑖,𝑗𝑗 is physical pressure
for all 𝒊𝒊, 𝒋𝒋
𝒑𝒑𝒊𝒊,𝒋𝒋 ← 𝟎𝟎 𝝆𝝆 density
ℎ grid spacing
for 𝒏𝒏 iterations
for all 𝒊𝒊, 𝒋𝒋 ∆𝑡𝑡 time step
𝒅𝒅 ← 𝒖𝒖𝒊𝒊+𝟏𝟏,𝒋𝒋 − 𝒖𝒖𝒊𝒊,𝒋𝒋 + 𝒗𝒗𝒊𝒊,𝒋𝒋+𝟏𝟏 − 𝒗𝒗𝒊𝒊,𝒋𝒋 Not necessary for simulation
𝒔𝒔 ← 𝒔𝒔𝒊𝒊+𝟏𝟏,𝒋𝒋 + 𝒔𝒔𝒊𝒊−𝟏𝟏,𝒋𝒋 + 𝒔𝒔𝒊𝒊,𝒋𝒋+𝟏𝟏 + 𝒔𝒔𝒊𝒊,𝒋𝒋−𝟏𝟏
…
𝒅𝒅 𝝆𝝆 𝒉𝒉
𝒑𝒑𝒊𝒊,𝒋𝒋 ← 𝒑𝒑𝒊𝒊,𝒋𝒋 + �
𝒔𝒔 ∆𝒕𝒕
Overrelaxation
• Gauss-Seidel is very simple to implement
• Needs more iterations than global methods
• Acceleration: Overrelaxation – Magic!!

𝑑𝑑 ← 𝒐𝒐 𝑢𝑢𝑖𝑖+1,𝑗𝑗 − 𝑢𝑢𝑖𝑖,𝑗𝑗 + 𝑣𝑣𝑖𝑖,𝑗𝑗+1 − 𝑣𝑣𝑖𝑖,𝑗𝑗

• Multiply the divergence by a scalar 1 < 𝑜𝑜 < 2

• I use 𝑜𝑜 = 1.9 in the code
• Increases convergence dramatically
• The computed pressure value is still correct!
Last step: Advection

𝐯𝐯𝑡𝑡 𝐯𝐯𝑡𝑡+∆𝑡𝑡

𝑝𝑝𝑖𝑖 𝑢𝑢𝑖𝑖,𝑗𝑗
𝑝𝑝𝑖𝑖
moving particle
𝑣𝑣𝑖𝑖,𝑗𝑗

• In a fluid the velocity state is carried by the particles

• Particles move, grid cells are static
• We need to move the velocity values in the grid!
Semi-Lagrangian Advection

𝑢𝑢𝑡𝑡+∆𝑡𝑡 ← 𝑢𝑢𝑡𝑡

𝑢𝑢𝑡𝑡

• Which fluid “particle” moved to the location where 𝑢𝑢 is stored?

• Set the new velocity 𝑢𝑢𝑡𝑡+∆𝑡𝑡 to the velocity 𝑢𝑢𝑡𝑡 at the previous position
Semi-Lagrangian Advection
𝐯𝐯(𝐱𝐱)
𝐱𝐱

𝐱𝐱 − ∆𝑡𝑡 � 𝐯𝐯(𝐱𝐱)

• Compute 𝐯𝐯 at position 𝐱𝐱 where 𝑢𝑢 is stored

• The previous location can be approximated as 𝐱𝐱 − ∆𝑡𝑡 � 𝐯𝐯(𝐱𝐱)
• Assuming a straight path introduces viscosity!
Can be reduced with vorticity confinement
Get the 2d Velocity
𝑣𝑣𝑖𝑖−1,𝑗𝑗+1 𝑣𝑣𝑖𝑖,𝑗𝑗+1

𝑣𝑣̅ 𝐯𝐯 = [𝑢𝑢𝑖𝑖,𝑗𝑗 , 𝑣𝑣]̅

𝑢𝑢𝑖𝑖,𝑗𝑗

𝑣𝑣𝑖𝑖−1,𝑗𝑗 𝑣𝑣𝑖𝑖,𝑗𝑗

• Average surrounding vertical velocities

𝑣𝑣̅ = (𝑣𝑣𝑖𝑖,𝑗𝑗 + 𝑣𝑣𝑖𝑖,𝑗𝑗+1 + 𝑣𝑣𝑖𝑖−1,𝑗𝑗 + 𝑣𝑣𝑖𝑖−1,𝑗𝑗+1 )/4
General Grid Interpolation
𝑣𝑣𝑖𝑖,𝑗𝑗+1 𝑣𝑣𝑖𝑖+1,𝑗𝑗+1

𝑣𝑣̅ 𝑤𝑤00 = 1 − 𝑥𝑥/ℎ 𝑤𝑤01 = 𝑥𝑥/ℎ

𝑥𝑥
𝑤𝑤10 = 1 − 𝑦𝑦/ℎ 𝑤𝑤11 = 𝑦𝑦/ℎ
𝑦𝑦
𝑣𝑣𝑖𝑖,𝑗𝑗 𝑣𝑣𝑖𝑖+1,𝑗𝑗

𝑣𝑣̅ = 𝑤𝑤00 𝑤𝑤10 𝑣𝑣𝑖𝑖,𝑗𝑗 + 𝑤𝑤01 𝑤𝑤10 𝑣𝑣𝑖𝑖+1,𝑗𝑗 + 𝑤𝑤01 𝑤𝑤11 𝑣𝑣𝑖𝑖,𝑗𝑗+1 + 𝑤𝑤00 𝑤𝑤11 𝑣𝑣𝑖𝑖+1,𝑗𝑗+1
Smoke Advection

• Store density value at the center of each cell

• Advect it like the velocity components
Streamlines

𝐱𝐱 ← start position
𝑠𝑠 ← step size 𝑠𝑠 � 𝐯𝐯(𝐱𝐱 𝟐𝟐 )
𝐱𝐱 𝟑𝟑
𝐱𝐱 𝟐𝟐

for 𝒏𝒏 steps
𝑠𝑠 � 𝐯𝐯(𝐱𝐱 𝟏𝟏 )
𝐯𝐯 ← sampleV(𝐱𝐱)
𝐱𝐱 𝟏𝟏
𝐱𝐱 ← 𝐱𝐱 + 𝑠𝑠𝐯𝐯
Let’s look into the code…