FROM PENGUINS TO DUST CLOUDS
How unconventional thinking helped us visualize interstellar dust in virtual reality
By Benjamin Bairrington – 15 July 2026
A SILLY QUESTION ABOUT PENGUINS
Let us begin with a deliberately silly question: How many penguins live natively on Cape Canaveral?
Cape Canaveral is best known as a launch site for NASA, not as a penguin habitat. The answer, therefore, is zero—at least for now.
The answer itself is unimportant. What matters is the brief mental detour required to reach it. Even if you knew immediately that Cape Canaveral has no native penguin population, you still had to connect two normally unrelated ideas, test the connection against what you know, and decide whether it made sense.
That small exercise illustrates a valuable approach to problem-solving. When familiar lines of reasoning lead nowhere, deliberately combining unrelated ideas can reveal assumptions we did not realize we were making—and sometimes uncover a path we would otherwise never consider.
THE FAMILIAR PATH IS NOT ALWAYS THE SHORTEST
This kind of conceptual leap is easy for humans to make, but it can be less natural for generative AI systems. Models such as ChatGPT and Claude generate responses by drawing on patterns learned from large collections of training data. As a result, they are often very good at producing conventional solutions: approaches that appear frequently, follow established practice, or seem most likely to succeed.
Most of the time, this is exactly what we want. When asking an AI system to check grammar, explain a standard method, or help structure a program, conventional reasoning is useful. However, the most familiar solution is not necessarily the best one.
Consider the simplified network below. The conventional route from A to E follows alphabetical order: A → B → C → D → E. Each step has a cost of 0.5, giving the route a total cost of 2.0. The alternative route begins with the apparently less efficient step from A to F, which costs 0.7. From F, however, reaching E costs only another 0.2. The unconventional route therefore has a total cost of just 0.9.
The first step looks worse, but the complete route is much better.
Two possible routes from A to E. Although the alphabetical route begins with a cheaper step, its total cost is 2.0. The unconventional route through F begins with a more expensive step but has a total cost of only 0.9.
This is why the wording of a question matters. Asking for the “optimal” solution may reinforce familiar assumptions about what an acceptable solution should look like. Asking for an unconventional solution creates permission to explore paths that initially appear wrong—and occasionally leads to a genuine breakthrough.
THE CHALLENGE: BRINGING A DUST MAP INTO VIRTUAL REALITY
This brings us to the problem faced by our group at the Technion. We study the three-dimensional distribution of interstellar dust: tiny solid particles spread throughout the space between stars. For this project, each sampled point was described by four values—its x, y, and z coordinates, together with its dust density.
The three-dimensional dust map produced by Edenhofer and collaborators contains this information at enormous scale, with source files several gigabytes in size. We wanted to transform those numerical data into an environment that could be explored from within using a Meta Quest 3 virtual-reality headset.
That ambition immediately created a performance problem. Our Unity prototype represented each sampled location as a separate visual object. Although this was straightforward to implement, rendering thousands of partially transparent objects simultaneously placed a heavy burden on the headset’s limited graphics hardware.
We applied a series of conventional optimizations. We retained only the densest five percent of the data, randomly selected 20,000 points from that subset, removed dynamic lighting and shadows, simplified the color palette, and moved data preprocessing into Python so that Unity could load a prepared CSV file. The individual dust samples were rendered as fuzzy spheres.
These changes helped, but not enough. The simulation still suffered from a low and inconsistent frame rate, making it uncomfortable to explore in virtual reality.
We then showed the C# implementation to both ChatGPT and Claude and asked how it could be optimized. Both systems recommended several established techniques, including asynchronous loading. These suggestions improved how the data entered the scene, but they did not substantially reduce the cost of rendering it once it was there.
“Instead of asking for the optimal solution, we asked for an unconventional one.”
The resulting suggestion was surprisingly simple: replace the three-dimensional spheres with two-dimensional billboards. In a game engine, a billboard is a flat image that continually rotates to face the camera. Billboards are commonly used to represent objects such as distant vegetation, particles, or clouds without constructing full three-dimensional geometry.
For our purposes, that apparent shortcut was exactly what we needed. Each dust sample now required only a flat sprite rather than a three-dimensional object. The scene retained the appearance of a volumetric cloud, while rendering performance improved dramatically.
Billboard sprites demonstrated in the Wareware game engine. As the camera rotates, the two-dimensional characters rotate with it, ensuring that they always face the viewer. The same principle allowed our dust samples to appear three-dimensional while remaining inexpensive to render.
MAKING A CLOUD FROM CLOUDS
The billboard solution resolved the performance problem, but it introduced a new visual one. When each dust sample was represented by a soft circular sprite, the combined structure resembled an out-of-focus image rather than a cloud of interstellar material. Looking more closely only revealed the individual blurred circles.
To improve the illusion, we drew inspiration from the approximately self-similar structure of natural clouds. A small region within a cloud often retains some of the irregular texture and complexity of the larger formation. This quality is commonly associated with fractal geometry, in which similar patterns appear across multiple spatial scales.
A fuzzy circle does not possess that kind of internal structure. A cloud-shaped sprite, however, does. We therefore created a 256 × 256-pixel sprite using Krita’s Essential Cloud and Water brush. Each billboard still represented a single sample from the dust map, but it now contained an irregular, cloud-like pattern.
When thousands of these sprites overlapped, their smaller structures combined into a larger and more convincing cloud. The improvement did not come from adding more geometric complexity. It came from choosing a two-dimensional image whose visual structure reinforced the appearance of the whole.
Two representations of the same dust-map data. Left: each sampled point is rendered as a fuzzy circular sprite, producing a blurred appearance. Right: each point is rendered with a textured cloud sprite, creating a sharper and more convincing volumetric structure.
A DIFFERENT QUESTION, A BETTER ROUTE
The final prototype transformed a massive three-dimensional dust map into an environment that could be explored in virtual reality. Achieving that result required more than making the original approach incrementally faster. It required reconsidering what each dust sample needed to be.
Replacing spheres with billboards reduced the rendering cost. Replacing fuzzy circles with textured cloud sprites improved the visual result. Together, these changes produced a simulation that was both more efficient and more convincing.
The experience also demonstrated a useful—but limited—role for generative AI in technical problem-solving. ChatGPT and Claude did not build the simulation or determine which suggestion would work. Their initial recommendations largely reflected established optimization practices. The breakthrough came only after we reframed the prompt and explicitly requested an unconventional approach. Human judgment was still required to recognize the promising idea, implement it, test it, and refine its appearance.
That brings us back to Cape Canaveral. There may be no penguins living there, but briefly connecting those unrelated ideas illustrates the habit that moved this project forward: when the familiar route reaches a dead end, it can be worth asking a question that initially seems absurd.
Sometimes the better path is the one we would normally dismiss.
AUTHOR NOTE
Benjamin Bairrington is a graduate student in the Star Formation and Interstellar Medium Lab at the Technion.