Turning a forensic glass study into an acceptance test
A forensic study broke 60 identical panes, and no overall pattern repeated. I turned its statistics into the test a simulated break must pass before it goes into the site’s library.
Sixty identical panes broken in a forensic study gave no two overall patterns alike. Their statistics became the test for my own broken glass: a simulated break goes into the library only if it matches them.
The research plan of 1 October had listed broken glass as parked. On 2 October I stopped all other work on it until there had been thorough research with as many references as could be found, and an existing system to build on, chosen as the best available rather than the easiest to find. The scope was plain and laminated glass (two sheets bonded together), no tempered, with the tiny shards at the edges, the sides and through the pane.
My first breaks read as computer-made
The tempered pane was an even honeycomb, which plain glass never makes. The pieces were Voronoi-like, made by splitting the pane into cells around scattered points.
Voronoi cells average six corners with three cracks meeting at every point, while real cracks stop against each other in T-junctions and give pieces averaging four corners. That alone makes a pattern read as computer-made, and it is why the standard game and animation tools look alike. And the cracks were lines drawn on the glass, where real ones are faces through its thickness.
Sixty real panes, and no pattern repeated
The best evidence was a forensic study, NIJ 241445. Sixty identical 8-inch panes, nominally an eighth of an inch thick, were broken under controlled conditions: ten per series, struck by a 965 g weight with a round, sharp or blunt tip, or pressed slowly in a wooden frame. Across 1,770 pairwise comparisons, no overall pattern was duplicated.
The authors’ own summary, which my research notes re-checked at the source: “The fracture patterns produced using dynamic impact were much simpler than the fracture patterns produced using static pressure.”
My measuring code, run on the study’s hand-traced patterns, gives these medians: a dropped round tip leaves about 7 or 8 pieces and crosses 6 or 7 cracks on a circle 30 mm round the impact; a pressed pane has 130 to 180 pieces and 61 to 72 crossings.
The first extraction in the physics note gives slightly different medians: 168 pieces for the pressed round tip where the later tool gives 178, and 162 against 159.5 for the dropped blunt tip. They agree on the shape, not the last digit.
A break must match the real panes’ statistics
The system design turns this into an acceptance test. A break goes into the library only if it matches measured radial counts and their spread; ring bands 6 to 15 plate thicknesses out, and only for harder blows; piece areas and elongation (25 to 29% of pieces in dense breaks are more than eight times longer than wide); the share of T-junctions; and an average of about four corners per piece.
It must also sit beside the reference photographs of its category, 203 of them in 17 categories, and be compared with path-traced renders (pictures computed by following light physically).
The breaks are simulated with peridynamics
The method is physical: peridynamics, which models a material as points joined by bonds that can break, and which the glass-impact literature uses and checks against photographs of the actual plates. Peridynamics.jl, a library in the Julia language, is MIT-licensed. I ran a 25-line script of a hammer striking a framed pane, and it took 7.6 minutes.
A paper corrected my notes
The physics note had inferred that the radial cracks in a laminated pane’s two plies (its two sheets) do not line up. The 2014 PLOS ONE paper measured the opposite: they overlap. The design follows the paper.
Sources and notes (8)
- NIJ 241445, forensic study of 60 broken panes — no two patterns alike, and the piece and crack statistics
- PLOS ONE 2014, laminated glass fracture — 15 to 112 radial cracks across repeats, and the ply overlap correction
- Domokos et al., polygonal fracture patterns (arXiv 1912.04628) — T-junctions and four-cornered pieces versus Voronoi
- Peridynamics.jl — the MIT-licensed solver chosen for the simulation
- Broken-glass research README — the notes this was written from
- Fracture physics — the notes this was written from
- Broken-glass system design — the notes this was written from
- NIJ statistics — the notes this was written from
Light & shadowAll 19 parts →
