Open problems in the rain simulation
After the first fortnight, the water still had real faults: streams that were chains of beads rather than ribbons, a fluid solver I chose not to finish, and a list of critiques still open. This piece covers only what was not fixed, as the project’s notes record it.
The streams were still beads
One of my reference clips has streams that are a continuous rippled ribbon. Ours at my 9 October handover notes were a chain of beads. On a dark background the middle of a stream was 1.85 times the glass against 1.5 in the footage, the peak 2.9 against 2.4, the darkest point 1.26 against 0.83. The note beside those numbers reads “No dark gaps between the bright parts”.
Smaller ones from the stream research: a Y where two streams join has a twiggy texture; small sharp kinks where a runner is pulled toward a neighbouring lane; and a wide standing stream over a warm part of the photograph reads as a reddish bar, because the cylinder lens squeezes that colour sideways. On 11 October one more change: beads shed from a stream start as a segment of it and draw round, with no noise in the stream body.
A fluid solver held drops but drew no streams, so I set it aside
On 10 October I wanted the water to react as it should to gravity, wind and other water sources. A rule for every behaviour is brittle, so a thin-film fluid solver (a simulation of the water as one thin continuous layer) was started, on the CPU and then the GPU, behind a beta switch in the lab, the site’s private settings page.
What holds: 2 to 6 µL drops hold, 10 µL creeps or holds, 15 to 25 µL run at 35 to 55 mm a second, touching drops merge, volume is conserved.
What does not: no visible trails or streams, because a runner’s rear keeps up with it and leaves no thread; drops are coarse at 0.4 to 0.5 mm cells; no splash tails; and the cost is about 200 substeps a frame, too heavy for phones and about 11 seconds a frame under the recording rig’s software rendering. I chose to finish the rule-based water for now. The fluid stays behind its switch.
Still open: density, shapes, brightness and speed
From the handover notes, standing critiques still to satisfy: density patchy with empty lanes, which the photograph count argues with (see the piece on the 117 photographs); misshapen drops; the laser visible across the glass; backlit water; words flat on the glass magnified by the drops. And every new behaviour is to arrive as a lab switch, never as a preview flag.
From the design: the fog fields and the Beysens growth (the research’s laws for how fog droplets grow) do not appear in the code, and wind-blown spray as a volumetric effect is out of scope. From the path-traced check: the site’s light layer is about ten times brighter than a lamp of that size would make it.
And speed: the research says the software renderer runs the page at under one frame a second and says nothing about cost on a real GPU. Whether I measured the rain on my own machine with the performance switch, the repository does not show.
On 9 October this work was on the development branch and not pushed, by my choice. By 11 October the project log shows it merged in three pull requests.
Sources and notes (6)
- Beysens 2006, dew nucleation and growth — condensation growth the design draws on but the code does not yet contain
- Water session handoff — the notes this was written from
- Streams research — the notes this was written from; also records his six rain clips, as recorded in the repo
- Rain system design — the notes this was written from
- Rain research, round 3 — the notes this was written from
- Water drops research — the notes this was written from
Water, rain & windAll 11 parts →
