colonelwatch/pi-sph-fluid

Fluid simulation toy on a Raspberry Pi 4 with tilt input, using smoothed particle hydrodynamics (SPH)

pi-sph-fluid

This project uses an SSD1306 OLED module, an MPU6050 accelerometer, and a Raspberry Pi 4 to realize a fluid simulation toy that lets you toss around an ocean of water in your hand. (That is, it simulates fluid dynamics at a scale larger than the actual screen.) You can see it in action on Youtube. This can be thought of as a "free-surface flow" problem, and it's solved by a "smoothed particle hydrodynamics" (SPH) technique.

Running

The only dependencies are OpenMP and the ssd1306 driver by lexus2k. Thanks to the SDL emulation of the SSD1306 offered by the ssd1306 library, this project offers a choice of two executables: desktop_sph_fluid and pi_sph_fluid.

1. desktop_sph_fluid generates and renders a circular drop falling on a dry surface, and it can be compiled (on Linux) with the commands

git clone https://github.com/colonelwatch/pi-sph-fluid
cd pi-sph-fluid
git submodule update --init ssd1306
make desktop_sph_fluid

./desktop_sph_fluid

2. pi_sph_fluid generates the same but with gravity determined by reading the MPU6050 accelerometer. I2C and the driver for accelerometer need to be enabled by configuring /boot/config.txt. All the necessary commands are

sudo apt install git

# configure /boot/config.txt (you can also write these lines in manually)
sudo raspi-config nonint do_i2c 0 # call raspi-config non-interactively for turning on i2c correctly
sudo raspi-config nonint set_config_var dtparam=i2c_arm on,1000000 /boot/config.txt # set i2c speed to 1MHz
echo dtoverlay=mpu6050 | sudo tee -a /boot/config.txt # pull up device tree overlay for mpu6050
sudo reboot

git clone https://github.com/colonelwatch/pi-sph-fluid
cd pi-sph-fluid
git submodule update --init ssd1306
make pi_sph_fluid

# limit the rate at which the MPU6050 is being polled
echo 10 | sudo tee /sys/bus/iio/devices/iio:device0/sampling_frequency

./pi_sph_fluid

What's implemented?

Besides the ssd1306 driver, this project is just under 750 lines of C! Here's what this project implements in that many lines:

1. Linked-list neighbors search, a common technique but outlined well in [1]
2. Weakly-compressible smoothed particle hydrodynamics (WCSPH), completely described in [2] and [3]
   • Artificial viscosity and momentum-preserving pressure, described in the same
   • Ordinary advection (not XSPH)
   • Working around negative density error by clamping negative pressure to zero, mentioned in the IISPH, DFSPH, and PBF papers ([4-6] respectively) along with other papers
3. Boundary handling offered by [7]
4. Fake surface tension effects using artificial pressure, mentioned in [6]
5. Rendering using metaballs, following the original implementation in [8]
6. OpenMP acceleration

To be exact: expect a writeup on this or even some cleaning up, but the following equations are implemented.

Wendland C2

Given two-dimensional coordinates written as $\vec{x}_i = (x_i, y_i)$, the two-dimensional Wendland C2 kernel found in [9]:

$$W(\vec{x}_i - \vec{x}_j, h) = \begin{cases} \frac{7}{4 \pi h^2} (1 - 0.5 q)^4 (1 + 2 q) & q < 2 \\ 0 & q \geq 2 \end{cases}$$ $$\frac{d W}{d q} = \begin{cases} \frac{7}{4 \pi h^2} (-5 q) (1 - 0.5 q)^3 & q < 2 \\ 0 & q \geq 2 \end{cases}$$

where $q = \left\Vert \vec{x}_i - \vec{x}_j \right\Vert / h$, and the derivatives $dW/dx_i$ and $dW/dy_i$ found by the chain rule

$$\frac{dW}{dx_i} = \frac{dW}{dq} \frac{dq}{dx_i} = \left( \frac{dW}{dq} \right) \left( \frac{x_i - x_j}{\left\Vert \vec{x}_i - \vec{x}_j \right\Vert h} \right)$$ $$\frac{dW}{dy_i} = \frac{dW}{dq} \frac{dq}{dy_i} = \left( \frac{dW}{dq} \right) \left( \frac{y_i - y_j}{\left\Vert \vec{x}_i - \vec{x}_j \right\Vert h} \right)$$

Also note:

• $W(\vec{x}_i - \vec{x}_j, h)$ is typically rewritten as $W_{i j}$
• The typical expression of the gradient can be rewritten in terms of the above as

$$\nabla_i W_{i j} = \left( \frac{d W}{d x_i} , \frac{d W}{d y_i} \right)$$

Boundary using rigid particles

The pseudo-mass of which is defined in [7]:

$$\psi_{b_i} = \frac{\rho_0}{\sum_{b_j} W_{b_i b_j}}$$

where $\rho_0 = 1000$ is the reference fluid density

Fluid density with rigid correction

Found in [7]:

$$\rho_{f_i} = m_{f_i} W ( \vec{0} ) + \sum_{f_j} m_{f_j} W_{f_i f_j} + \sum_{b_j} \psi_{b_j} W_{f_i b_j}$$

Also note:

• Calculating the density seems to be the one exception where a particle can neighbor itself (see calculation of density in SPlisHSPlasH)
  • However, a particle doesn't seem to neighbor itself in the general case (see the comparison against brute-force search in the neighbors search library used in SPlisHSPlasH)
• Here, the self-contribution is written out as an explicit term
• This may or may not be the interpretation taken in [7] as well

WCSPH explicit pressure

Stated first in the context of fluid simulation with SPH in [2], plus the clamping at zero (as employed in incompressible SPH formulations in [4, 5]) used as a hack:

$$p_{f_i} = \begin{cases} \frac{c_0^2 \rho_0}{7} \left( \left( \frac{\rho_{f_i}}{\rho_0} \right)^7 - 1 \right) & \rho_{f_i} > \rho_0 \\ 0 & \rho_{f_i} \leq \rho_0 \end{cases}$$

where $c_0 = 400$ is the numerical speed of sound (see [2] or [3]) and $\rho_0 = 1000$ is the reference fluid density

Acceleration

Including momentum-preserving pressure gradient from [2], artificial viscosity from [2], and artificial pressure from [6], and rigid correction from [7]:

$$\begin{align*} \frac{d \vec{v}_i}{dt} & = - \sum_{f_j} m_{f_j} \left( \frac{p_{f_i}}{\rho_{f_i}^2} + \frac{p_{f_j}}{\rho_{f_j}^2} + \Pi_{{f_i} {f_j}} + s_{f_i f_j} \right) \nabla_{f_i} W_{f_i f_j} \\ & \qquad -\sum_{b_j} \psi_{b_j} \left( \frac{p_{f_i}}{\rho_{f_i}^2} + \Pi_{f_i b_j} + s_{f_i b_j} \right) \nabla_{f_i} W_{f_i b_j} \\ & \qquad + \vec{g} \end{align*}$$

where $\vec{g}$ is gravitational acceleration, and the artificial viscosities (original and corrected) being

$$\Pi_{f_i f_j} = \begin{cases} -\frac{\alpha c_0}{\bar{\rho}_{f_i f_j}} \frac{h \, \vec{x}_{f_i f_j} \cdot \vec{v}_{f_i f_j}}{\left\Vert \vec{x}_{f_i f_j} \right\Vert^2 + \epsilon h^2} & \vec{x}_{f_i f_j} \cdot \vec{v}_{f_i f_j} \leq 0 \\ 0 & \vec{x}_{f_i f_j} \cdot \vec{v}_{f_i f_j} > 0 \end{cases}$$ $$\Pi_{f_i b_j} = \begin{cases} -\frac{\alpha c_0}{\rho_{f_i}} \frac{h \, \vec{x}_{f_i b_j} \cdot \vec{v}_{f_i b_j}}{\left\Vert \vec{x}_{f_i b_j} \right\Vert^2 + \epsilon h^2} & \vec{x}_{f_i b_j} \cdot \vec{v}_{f_i b_j} \leq 0 \\ 0 & \vec{x}_{f_i b_j} \cdot \vec{v}_{f_i b_j} > 0 \end{cases}$$

where $\alpha=0.01$ and $\epsilon=0.01$, and the artifical pressure being

$$s_{i j} = k_1 \left( \frac{W_{i j}}{W(k_2 h \, \vec{e})} \right)^4$$

where $k_1 = 0.1$, $k_2 = 0.2$, and $\vec{e}$ is just a unit-length vector

Also note:

• The mean rho is written as $\bar\rho_{f_i f_j} = (\rho_{f_i} + \rho_{f_j})/2$
• The position difference is written as $\vec{x}_{i j} = \vec{x}_i - \vec{x}_j$
• The velocity difference is written as $\vec{v}_{i j} = \vec{v}_i - \vec{v}_j$

Metaballs rendering

Given $\vec{x}_{p_i}$ is the center of a pixel and $\Delta x_p$ is the width of a pixel in the domain, metaballs rendering:

$$g(\vec{x}_{p_i}) = \sum_{f_j} \frac{W_{p_i f_j}}{W(0.5 \, \Delta x_p \,\, \vec{e})}$$

where $\vec{e}$ is just a unit length vector and pixel $p_i$ should light up if $g(\vec{x}_{p_i}) \geq 1$

Leapfrog integration

$$\begin{align*} \vec{v}_{i+1/2} & = v_i + 0.5 \, \Delta t \left. \frac{d \vec{v}}{d t} \right\rvert_{\vec{x}_i} \\ \vec{x}_{i+1} & = x_i + \Delta t \, \vec{v}_{i + 1/2} \\ \vec{v}_{i+1} & = v_{i+1/2} + 0.5 \, \Delta t \left. \frac{d \vec{v}}{d t} \right\rvert_{\vec{x}_{i+1}} \end{align*}$$

What's not implemented?

This project really sits on the ground floor of SPH, and some next steps include:

1. Z-order or some other -order sort of particles for a higher cache hit rate, as described in [1]
   • Ideally, this sort should be multithreaded
2. Inferring velocity from accelerations taken from the MPU6050 and assigning that velocity to the boundary
   • This would add a bit more realism to the fluid simulation
3. GPU acceleration or an implementation of one of the incompressible schemes (PCISPH, IISPH, DFSPH, PBF, etc)
   • There's a failed attempt at IISPH in the IISPH branch
   • Some of the most interesting fluid phenomena don't arise in the small number of particles that CPU WCSPH can handle without going unstable (currently 431)

References

[1] J. M. Domínguez, A. J. C. Crespo, M. Gómez-Gesteira, and J. C. Marongiu, “Neighbour lists in smoothed particle hydrodynamics,” International Journal for Numerical Methods in Fluids, vol. 67, no. 12, pp. 2026–2042, 2011, doi: 10.1002/fld.2481.

[2] J. J. Monaghan, “Smoothed particle hydrodynamics,” Reports on progress in physics, vol. 68, no. 8, p. 1703, 2005.

[3] J. J. Monaghan, “Simulating Free Surface Flows with SPH,” Journal of Computational Physics, vol. 110, no. 2, pp. 399–406, 1994, doi: 10.1006/jcph.1994.1034.

[4] M. Ihmsen, J. Cornelis, B. Solenthaler, C. Horvath, and M. Teschner, “Implicit Incompressible SPH,” IEEE Transactions on Visualization and Computer Graphics, vol. 20, no. 3, pp. 426–435, Mar. 2014, doi: 10.1109/TVCG.2013.105.

[5] J. Bender and D. Koschier, “Divergence-free smoothed particle hydrodynamics,” in Proceedings of the 14th ACM SIGGRAPH / Eurographics Symposium on Computer Animation, Los Angeles California: ACM, Aug. 2015, pp. 147–155. doi: 10.1145/2786784.2786796.

[6] M. Macklin and M. Müller, “Position based fluids,” ACM Trans. Graph., vol. 32, no. 4, pp. 1–12, Jul. 2013, doi: 10.1145/2461912.2461984.

[7] N. Akinci, M. Ihmsen, G. Akinci, B. Solenthaler, and M. Teschner, “Versatile rigid-fluid coupling for incompressible SPH,” ACM Trans. Graph., vol. 31, no. 4, pp. 1–8, Aug. 2012, doi: 10.1145/2185520.2185558.

[8] J. F. Blinn, “A Generalization of Algebraic Surface Drawing,” ACM Trans. Graph., vol. 1, no. 3, pp. 235–256, Jul. 1982, doi: 10.1145/357306.357310.

[9] D. Violeau, Fluid mechanics and the SPH method: theory and applications. Oxford University Press, 2012.

[10] J. J. Monaghan, “SPH without a tensile instability,” Journal of computational physics, vol. 159, no. 2, pp. 290–311, 2000.