Introduction
Dynamical systems are systems that describe how states evolve with time. For instance given the position and the velocity of a ball, it is possible to calculate how its position and speed will evolve with time. Some dynamical systems display a tendency to evolve towards certain states independent of their initial states, like the Lorenz system.
In this post I will show how I rendered and animated two 3D dynamical systems.
The Clifford Attractor
The system that inspired me to generate a 3D attractor was the Clifford attractor, which is defined as:
\begin{align*}
x_{n+1} &= \sin(a y_n) + c \cos(a x_n)\\
y_{n+1} &= \sin(b x_n) + d \cos(b y_n)\\
\end{align*}
If the constants are set to \(a=1.5, \; b=-1.8, \; c=1.6, \; d=0.9\), and \(x_0\), \(y_0\) to random values, and the equation is iterated and the \(x_n, \; y_n\) pixel value are incremented, an image like this (after some post-processing) can be generated:
Using the Clifford attractor as template, I created two 3D attractor equations.
A1 Attractor
I defined the first attractor (which I arbitrarily called “A1”) as:
\begin{align*}
x_{n+1} &= z \sin(a + y_n) + y \cos(d + z_n) \\
y_{n+1} &= x \sin(b + z_n) + z \cos(e + x_n) \\
z_{n+1} &= y \sin(c + x_n) + x \cos(f + y_n) \\
\end{align*}
This attractor, depending on its parameters can generate images like this:
To speed up code, instead of computing all the points from a single initial point, I computed multiple initial points and tracked how they evolved with time. The code to generate a single image can be summarized as the following sequence:
- Generate a point cloud of P \((x_0,\; y_0,\; z_0)\) points around a unit sphere centered at \((0,\; 0,\; 0)\).
- For every generated \((x_0,\; y_0,\; z_0)\) point, compute \(x_n,\; y_n,\; z_n\) for \(n < N\) and add those points to the point cloud.
- Remove points \((x_n,\; y_n,\; z_n)\) for \(n < S\) from the point cloud.
- 3D rotate the points.
- Perspective project 3D points onto 2D.
- Store in the framebuffer the number of times each 2D point “hits” a pixel.
- Colormap the framebuffer.
To generate a video sequence I modified the \(a,\; b,\; c,\; d,\; e,\;f\) paramaters and the rotation transformation at every frame the video.
A2 Attractor
I defined the second attractor (A2) as:
\begin{align*}
x_{n+1} &= z \sin(a + y_n x_n) + y \cos(d + z_n x_n) \\
y_{n+1} &= x \sin(b + z_n y_n) + z \cos(e + x_n y_n) \\
z_{n+1} &= y \sin(c + x_n z_n) + x \cos(f + y_n z_n) \\
\end{align*}
Using the same algorithm as in A1, but computing 128 times more points, the attractor equations produced the following image:
Again, the video sequence was generated the same way as in the A1 video.
Conclusion
Rendering synthetic images is relatively simple, but requires some trial and error to find equations and parameters that can generate interesting images.
Here I rendered images at 4K (3840 × 2160), the A1 and A2 videos, and also the A1 image (a single frame of the A1 video sequence) used 500 M points per image, and most of the times look “grainy”. To make grain disappear (while keeping the same resolution), more points need to be computed, and that is what I did when rendering the A2 image, where I used 64 G points instead.
The tradeoff of computing more points is computational time. Rendering images can get computationally expensive, both video sequences, made of 1800 frames, took ~24 h to render using 16 threads on a i9-9900K @ 3.6 GHz. The A2 image on the other side, took ~1 h.
The code is available on GitHub.