Where my fracture simulator fails on pressed glass
My simulated glass fails at about a fifth of real glass’s strain (how far it stretches before it breaks). Struck panes still break in the order real glass follows, but a pane pressed slowly stores a 27th of the real energy and breaks into a sparse web.
The fracture simulator is the offline half of the broken-glass system: a Julia script that strikes a framed pane with a hammer using Peridynamics.jl, an analysis script that measures the result the way the sixty forensic test panes were measured, and an export that turns a run into a file the site can play back. By 2 and 3 October there were 14 plain breaks: ten 100 mm panes, hit at 6.5 to 16 m/s, with a stone, or pressed; and four full-size 203 mm panes.
The first strike broke in the order real glass does
The earliest test was a 100 × 100 × 4 mm pane at 0.8 mm point spacing, the outer 5 mm held, hit off centre by a 0.5 kg, 16 mm steel sphere at 6 m/s: 78,125 points, 454 seconds. At about 180 microseconds there is an X of radials; at 275, radials across about 70% of the pane; at 390 a larger crushed centre and a circular arc on the struck face. That is the order real glass follows.
Two mistakes in the first runs
Two things in the first runs were mistakes of mine. A regular grid of points gave kaleidoscope patterns in the cross-check with LAMMPS, a second simulator, so the points are now jittered. And a rigid clamp made false cracks along the frame line; the design calls for a springy glazing bead (the strip that holds a pane in its frame) instead.
Each pane’s point strength is random (Weibull), because real glass’s strength is set by random flaws.
Pressed panes break too sparsely, because the glass is too weak
The harder problem was the slow press. The simulated glass fails at about a fifth of real glass’s strain, because the peridynamic horizon (the distance over which each point is bonded to its neighbours) sets the critical stretch, how far a bond may stretch before it breaks. That stretch is proportional to one over the square root of the horizon: at a 3 mm horizon it is 1.9 × 10-4 against a real 1 × 10-3. A pane pressed to failure therefore stores a 27th of the real energy and breaks into a sparse web.
The fix was to put the missing energy in. A point-load pre-bend gives a dense centre web with radials to the corners, as in one of the full-size pressed runs. The window is narrow: half a millimetre of bend gives the web, while 0.75 mm dices the pane into diagonal ladders, like tempered glass, which I had ruled out.
A press driven at fixed speed grows the web generation by generation on a 100 mm pane, but at full size the long hinged pieces follow the ram without cracking again.
Still open
At full size, the forensic study’s pressed density of about 178 pieces is still to find, and a shallow bowl under the point-load bend is the next try. At 1 mm resolution the patterns may read too smooth close up, so the design allows a 0.5 mm library later, which needs a bigger machine.
The design plays the cracks back in the browser about a thousand times slower than real ones, which run at about 1.5 km/s.
A break is accepted when the same code that measured sixty real panes measures it. Where the simulator fails, a fifth of the strain and ladders where there should be a web, is recorded beside the results.
Sources and notes (7)
- Peridynamics.jl — the solver the hammer-strike runs use
- LAMMPS PERI — the offline cross-check that showed kaleidoscope patterns on a regular grid
- Wang et al. 2024, Int. J. Fract. (NSF manuscript) — horizon-dependent strength, which is why the simulated glass fails early
- NIJ 241445, forensic study of 60 broken panes — the real pressed-pane density the simulator is measured against
- Fracture simulator README — the notes this was written from
- Broken-glass system design — the notes this was written from
- Prior art in fracture simulation — the notes this was written from
Light & shadowAll 19 parts →
