São Paulo WorldGlass Lab
Open the lab Grenade in the shop Armoured V8
Glass Dynamics · the physics

How the Glass Lab works

Nothing in the lab is animated by hand. Every crack, every falling piece and every cloud of tempered dice comes out of Glass Dynamics, a C++ fracture engine compiled to WebAssembly and running in your browser. This page walks through what it calculates, in the order it calculates it, with the equations and the numbers it uses — and says plainly which numbers are measured, which are fitted to test standards, and which are estimated.

published a value from a standard, a handbook or a paper   fitted a constant tuned so the model passes or fails the tests that real glass does   estimated a reasoned value without a direct measurement

01Overview

A strike is followed through the same steps every time, in time steps from 50 nanoseconds (inside a bullet strike) to milliseconds (while a pane rings down or pieces fall):

Contactthe object presses on the glass with a force from contact mechanics
Bendingthe pane bends and vibrates as a plate under that force
Stressbending and residual stress at every face of every glass ply
Flawseach patch of glass has its own strength; the first to be beaten breaks
Crackscrack tips run, turn, fork and stop, tip by tip
Piecesclosed-off regions become pieces with mass, spin and a shape
Fallpieces fly, land, break again; tempered dice rain

The page only draws what the engine returns and adds sound. The same engine couples to the São Paulo World driving and flight simulators for windscreens, side glass and aircraft windshields.

02The glass itself

Ordinary window glass is soda-lime-silica float glass. The engine uses these properties for it:

PropertyValueSource
Density2,500 kg/m³published
Young's modulus E72 GPapublished
Poisson's ratio ν0.22published
Fracture toughness KIc0.75 MPa√mpublished
Hardness5.5 GPapublished
Terminal crack speed1,520 m/spublished (high-speed photography)
Branching constant Ab2.1 MPa√mpublished (fracture-mirror relation)
Weibull modulus m, scale σθ7, 70 MPa (1 m², 60 s)published (EN 16612 range)
Edge strength0.8 × surfacepublished (EN 16612)
Stress-corrosion exponent n16published
Comminution energy40 MJ/m³fitted (UL 752)
Interlayer strain across a crack, PVB / ionoplast3 / 1.25 ×fitted (EN 356, ASTM E1996)

Annealed, heat-strengthened, fully tempered

The same glass is made stronger by locking stress into it. Thermal tempering cools the surfaces first: they end up squeezed in compression and the core held in tension. Across the thickness t the profile is a parabola,

σ(ζ) = CT · (1 − 3ζ²),   ζ = 2z/t − 1 the surface carries a compression of 2·CT, the stress crosses zero at 21 % of the thickness from each face, and the core carries the tension CT.
0 front faceback face ← compressiontension → surface −2·CT core +CT
Residual stress through a tempered pane (horizontal: stress; vertical: depth, drawn left to right here as front face to back face). A crack must first get through the compressed skin; once it reaches the core, the stored tension drives it on by itself — which is why tempered glass bursts all at once.

Fully tempered glass has a surface compression of about 100 MPa or more, heat-strengthened about 24–52 MPa, annealed none. Chemically strengthened glass (an aircraft windshield) gets a thin, very hard ion-exchanged skin instead: a linear layer over its depth, balanced by a small uniform core tension. When the engine checks whether a flaw breaks, it subtracts the surface compression from the bending tension; when a crack is already running through the core, the core tension pushes it.

Laminates and bullet-resistant glass

Laminated glass bonds glass plies with polymer interlayers: PVB in a car windscreen (shear modulus about 20 MPa at impact rates), ionoplast in hurricane glass (about 120 MPa), polyurethane (TPU) in armour, with a polycarbonate liner at the back of bullet-resistant panels to catch the spall.

03The pane as a plate

Whatever the layup, the pane is reduced to one thin (Kirchhoff) plate with a mass per area ρh and a bending stiffness D. For a laminate the stiffness sits between the plies sliding freely and the plies fully bonded, by how much shear the interlayer passes on (the Bennison–Wölfel coefficient Γ used by ASTM E1300):

D = Dlayered + Γ (Dmono − Dlayered) Γ is evaluated with each interlayer's stiffness at impact rates, so a laminate hit hard behaves almost like one thick pane, and the ply farthest from the impact carries the most tension.

The bending stress at depth z in glass ply i, for curvature κ:

σ = E/(1−ν²) · κ · [ (z − zi) + Γ (zi − zNA) ]

Modes, stepped exactly

The pane is simply supported at its edges. Its deflection is a sum of standing waves (Navier modes),

w(x, y, t) = Σm,n qmn(t) · sin(mπx/a) · sin(nπy/b)

and each mode is a damped oscillator advanced with its exact solution for a force held over the step. That is why the engine can take a 50-nanosecond step during a bullet strike and a 2-millisecond step while the pane rings, without ever going unstable. Up to 44 × 44 modes are used (fewer on slower devices); the fundamental of a 1.2 × 1.0 m pane of 6 mm glass comes out at about 25 Hz.

The near field

A finite set of modes cannot hold the sharp stress peak right under a small load. Within the series' resolution radius rm the engine adds the missing curvature analytically:

Δκ = F / (4π D) · ln( rreach / max(r, re) ),   rreach = min( rm, √(t · √(D/ρh)) ) re is Westergaard's equivalent radius for thick-plate contact. The second limit is how far a bending wave has travelled since the load began: a 10-microsecond stone strike cannot bend the plate out to rm. Checked against a 600 × 600-mode reference to 4 %.

Under a point load an infinite plate also behaves like a dashpot, v = F / (8√(Dρh)) (Zener); the contact sees that local impedance first, relaxing into the modes. Without it a steel ball on glass would reach about twice the real force.

04Contact: how hard the object pushes

ObjectForce model
Balls, hail, bags, headformsHertz: F = 4/3 E* √R δ1.5; past the object's own yield pressure it flattens (Thornton); energy loss by Flores et al. so it rebounds with the right restitution
Sharp tips (window punch, stone corners)Cone indentation F = π tan²ψ · H · δ² at the glass hardness
Flat ends (2×4, brick)Boussinesq punch 2E*aδ, capped at the crushing load of the object
BirdsWilbeck: the shock pressure ρv(c₀ + kv), then the steady flow ρv²A as the bird spreads like a fluid
Bullets, fragments, cannonballsA penetration model ply by ply — see Bullets

The object and the plate are stepped together at a tenth of the contact time, so a 40-microsecond ball strike gets hundreds of steps.

Impact energy calculator

Kinetic energy
Momentum
Same as dropping it from
Speed

05When glass breaks

Glass is almost never as strong as its chemistry: it breaks at microscopic surface flaws, and how bad the worst flaw is depends on luck and on how much glass is stressed. The engine covers each face of each glass ply with a lattice of flaw probes, and gives each its own strength drawn from a Weibull distribution for its area:

P(fail) = 1 − exp( −A/A₀ · (σ/σθ)m ),   σθ = 70 MPa at A₀ = 1 m², m = 7 a small patch is stronger than a big one (the size effect); edges get 0.8 × the strength (EN 16612). The draws come from the seed, so the same shot breaks the same way and another seed differently.

Glass is also stronger the faster it is loaded, because water at the crack tip needs time to weaken it (stress corrosion):

σstrength(t) = σ60 s · (60 s / t)1/(n+1),   n = 16, at most × 3

Then, every few microseconds while something is happening, every probe compares the largest principal tension at its face (bending, minus any surface compression) with its strength. Three special cases sit under the load itself:

How many radial cracks

A load that only just breaks the pane leaves a handful of radial cracks; a much harder blow leaves a sunburst. The engine sends out

n = 4 + 4 ln(σ/σstrength) + 1.5 ln(1 + E/100 J) + U(0, 3),   4 ≤ n ≤ 18 at uneven angles. The crushed zone at the load point is never a clean circle: its edge carries five random harmonics, about 14 % of its radius. estimated from photographs of broken panes.

06Running cracks

Each crack is a polyline grown tip by tip across a fine grid (2–6 mm cells depending on the device), with the time each point was reached, so slow motion shows the crack front racing out. For each tip, every step:

Why cracks are not ruler-straight

Real glass is not perfectly uniform, so real cracks meander. Each tip carries a slowly wandering offset from its stress-preferred direction (an Ornstein–Uhlenbeck process):

dθw = −θw ds / L + σw √(2 ds / L) · N(0,1),   σw = 0.12 rad, L = 4 cm The turn the stress asks for is also scaled by how directional the stress is, dev / (dev + residual): the locked-in tension of heat-strengthened glass pulls equally every way and should not steer the crack. A running crack bends no tighter than a 10 cm radius; sharp changes of direction come only as forks. estimated

Wedges and rings

Radial cracks leave wedges of glass loaded at their tips by the object. As cantilevers, their roots carry σ = 3F / (πh²) on the face toward the impact; when that beats the strength there, a concentric ring opens — the classic spider web. In a laminate the interlayer holds the wedges, so only 15 % of that reaches their roots, which is why windscreens show many rings and plate glass few.

07Tempered glass

When the core tension is high enough, every running crack forks again and again: the pane "dices". The size of the dice comes from an energy balance — the strain energy stored by tempering pays for the new crack surfaces:

U = t (1−ν) CT² / (5E)  →   a = 10 KIc² / ((1−ν) · η · CT²),   η = 0.5 η is the share of the stored energy that goes into new surface, fitted so 4–6 mm fully tempered glass gives the 40–400 pieces per 50 × 50 mm square that UN ECE R43 requires. Slower devices draw coarser dice; the results panel always reports the real size.

How a crazed pane comes down

A diced pane has no strength left, but it rarely falls at once: the dice wedge against each other and the frame. The engine models every contact between two dice as friction with its own grip (drawn from the seed), and asks for each die whether its neighbours can still carry its weight:

capacity = Σ grip × (1.0 below · 0.12 beside · 0.05 above) × (neighbours left / neighbours it had)² + frame grip A die falls when its capacity drops below its weight. Each fall rechecks its neighbours after the time it takes to slide clear (0.2–1.5 mm under gravity), so the pane unzips from the hole upward at the speed gravity allows, slabs of dice that grip each other fall together and burst on the floor, and a fringe stays in the frame. fitted to footage of panes emptying in 0.3–1 s.

08Pieces and dice

When a pane's cracks have been quiet for 150 µs, its crack grid is flood-filled into regions; each region is traced into a polygon (holes included) and becomes a rigid piece with its area, mass, centre and moments of inertia. If there are more regions than the device's piece budget, the smallest merge into a neighbour, so the pieces always cover the whole pane — coarser, never missing.

Glass a blast is still pushing on keeps breaking after its first cracks. A large pane of annealed or heat-strengthened glass usually fails under the room's gas pressure, at moderate stress, so its first cracks are few and long and would leave sheets a metre long. But each piece is still held by the frame or by the pieces round it while the gas pushes, and bends like a cantilever, so the engine cuts every piece longer than the length that can carry that pressure across its long axis, through cracks it only meets as dead ends and round any pieces it wraps, until none is left:

σ = 3 p L² / h²  ⇒  Lmax = h √(σf / 3p) p is the blast pressure on the pane, h its thickness, σf the glass's characteristic strength at that loading rate. 10 mm plate under the shop's ~25 kPa of gas: about 40 cm, so a storefront comes out as hundreds of shards and blades, the largest about 40 cm long. estimated: the cantilever loading of a held piece.

Each piece starts with the plate's local velocity and spin, or the object's velocity if it lies in its path. A piece stays if it sits on the bottom edge, is wedged across two edges, or the frame's grip on its edge holds its weight. Free pieces fly on a fixed 0.5 ms step:

m dv/dt = m g − ½ ρair Cd A |v| v,   floor: restitution 0.25, friction 0.45 A landing breaks a piece again along a random line when its energy beats 40 × the energy of the new crack surface. A tempered piece drawn coarser than its real dice comes apart into them 40–160 ms after it leaves the frame, or when it lands, and the page scatters the real dice from there.

09Bullets

A projectile is followed ply by ply. Each glass ply crushes a cone (68° half-angle) ahead of it, which stops spreading at about 6.7 projectile radii (3 cm for a 9 mm, so a .50 or a cannonball crushes a cone its own size); the energy it absorbs is the comminution energy (40 MJ/m³) times the cone's volume, the plug ahead takes its share of the momentum, and a lead nose flattens as it goes (a hard core does not). Interlayers tear, a polycarbonate backing bulges and absorbs, and the momentum lost each step bends the plate — so the spider web and its rings come from the same crack model as everything else.

line-of-sight thickness = t / cos θ,   ricochet beyond θ = 72° A round that arrives at an angle meets every layer at its slant thickness; past 72° from square it glances off. A round landing in an earlier crater pays only 15 % of the comminution energy for glass already crushed there, so armour hit repeatedly in one place does give way. fitted so representative UL 752 builds stop their rated rounds.

A shotgun shell fires each pellet as its own projectile, spread by the pattern (00 buckshot: nine 3.5 g pellets, 6.5 mrad rms) and strung out along the shot string. The .50 BMG is the UL 752 Level 10 round: a 708 gr (45.9 g) bullet at 856–942 m/s, 18.6 kJ, five times a 7.62 mm rifle round. It goes through the 52 mm Level 8 build at about 565 m/s; the 95 mm Level 10 build stops it about 75 mm in.

Heavy shot on bullet-resistant glass

Behind the crushed glass a bullet meets polycarbonate and interlayer much thicker than itself, and punches a plug out of them. A cannonball 10 cm across meets the same 13 mm of polymer as a thin sheet, and thin sheets do not plug: they dish (thin-target behaviour). The engine decides which by the polymer's total thickness against the projectile's diameter:

polymer thickness < 0.3 × projectile diameter  ⇒  the polymer is pulled as a membrane 9 mm into Level 3: 13.3 mm of polymer, 9 mm bullet, ratio 1.5, plugged. Cannonball into Level 3: ratio 0.13, a membrane. estimated threshold.

The membrane is the same one that catches a ball in a windscreen: its tension falls off as 1/r from the rim of the ball, so it is highest at the rim (where it tears) and still pulling on the frame at the edge. The frame holds the edge with its sealant bond or bolts:

edge pull = Nrim · a / dedge   >   frame hold  ⇒  the whole laminate slides out of its rebate a is the rim radius (the ball, as the membrane wraps it), dedge the distance to the nearest edge. Frame holds: about 1 MPa × the bite for a sealant bond (25 kN/m for bullet-resistant glass 25 mm deep in its frame), 60 kN/m for an armoured car's bonded screen behind a steel overlap, 100 kN/m for a bonded windscreen, 75 kN/m for hurricane glazing set in structural silicone on both faces, 300 kN/m for a bolted aircraft windshield. estimated

How much a broken laminate takes before it tears comes down to three things, which together make it pass and fail the standards real laminates do (EN 356, ASTM E1996; the Tests menu runs them):

Once the edge is out, the laminate goes with the ball as one sheet. However cracked its glass is, every piece is still bonded to the polymer, so nothing flies off on its own: the sheet is one body (its glass, its polymer and the ball wrapped in it), flown with gravity and the drag of its whole area, and it lands and slides as one. Its glass is cracked into pieces up to 15 cm across, and 6 cm where it bends round the ball. Over the time the ball takes to travel the sheet's own radius, it folds round the ball's nose for about a radian, and the rest trails back from there as a cone, like a canopy with the ball in its apex. The ball shares its momentum with the sheet, so a 4 kg ball dragging 22 kg of Level 3 leaves at about 65 m/s instead of 440. Dragging the rest of the sheet to that speed over about its own radius L takes a rim tension

Nneed = Msheet u² / (2 L) / (2π a) If the polymer's strength (thickness × tensile strength, summed over the plies) is less, it tears round the ball: the ball goes on wrapped in a disc of laminate twice its radius (folded up to 125° round it), and the rest of the window follows as its own sheet, dished round the hole, with what the polymer could give it before it tore. Level 3 and Level 8 go whole; the all-glass Level 1, held together by PVB alone, tears round the ball. The supercar's wide windscreen, bonded behind a steel overlap, keeps its edge: its polycarbonate bulges about 30 cm round the ball before it tears.

10Blast waves

A blast from far off loads the whole pane alike; a charge close by does not. The engine takes the pressure history as two parts: the shock, reflected off the glass, as a Friedlander pulse; and, for a charge inside a closed room, the slower pressure of its hot gas as the room vents.

p(τ) = pr (1 − τ/td) e−bτ/td + pg (1 − e−τ/tr) e−τ/τv τ is the time since the shock arrived; pr the reflected peak, td the positive phase, b the decay, pg the gas pressure rising over tr and venting with time constant τv. The Friedlander part dips below zero after td — the suction behind every shock.

The numbers come from the charge and the geometry. For a TNT-equivalent charge of W kg at distance R, everything scales with Z = R / W1/3 (Hopkinson–Cranz), and the Kinney–Graham fits give, in free air:

ps/P₀ = 808 [1 + (Z/4.5)²] / √( [1 + (Z/0.048)²] [1 + (Z/0.32)²] [1 + (Z/1.35)²] )
td/W1/3 = 980 [1 + (Z/0.54)10] / ( [1 + (Z/0.02)³] [1 + (Z/0.74)6] √(1 + (Z/6.9)²) ) ms
i/W1/3 = 0.067 √(1 + (Z/0.23)4) / ( Z² ∛(1 + (Z/1.55)³) ) bar·ms

Glass facing the blast sees the shock reflected, which more than doubles it (Rankine–Hugoniot for air, γ = 1.4), less at a glancing angle α:

pr = 2 ps (7P₀ + 4ps) / (7P₀ + ps),   p = ps + (pr − ps) cos²α

The decay b is solved so the pulse carries the right impulse; the shock's arrival is integrated outward at its own speed, U = c₀ √(1 + 6ps/7P₀), faster than sound close in. In a closed room the gas adds a quasi-static pressure (the UFC 3-340-02 fit) that vents through the openings:

pg = 2,410 kPa · (W/V)0.72,   τv ≈ V / (Avent · 150 m/s)

What the glass does

The pane bends under the pressure like under any load, and its flaw probes break it — about 5 ms after the shock reaches a 6 mm tempered window in the shop scene: glass has inertia, and bending takes time. Then, while the pressure is above 1 kPa, every loose piece whose pressure force beats its frame grip is torn out (a crazed tempered sheet does not hold at all: its friction is a few times a die's weight). Pieces still in the opening are driven out by the venting gas:

dv/dt = p(t) · A / m · (1 − d/0.6 m)   for pieces within 0.6 m of the pane That launches the dice at 15–40 m/s, and windows leave as whole sheets of dice that spread into clouds — as in footage of blast-tested glazing. estimated: the 1 kPa hold and the 0.6 m reach of the gas jet.

Grenades in the lab

Two of the lab's objects explode. The hand grenade is lobbed at the pane at 7 m/s and goes off 1.5 s after the throw, wherever it has come to rest (often at the foot of the pane, where the shock meets the glass at a glancing angle: a 6 mm tempered pane can survive that, which is what the formulas say). The 40 mm grenade is a 230 g round fired at 76 m/s with an impact fuse: it bursts on the glass, 0.04 kg TNT equivalent and about 300 fragments. Both use the same blast, fragment and gas formulas as the shop, with the pane loaded from the side the charge is on.

Blast calculator

Scaled distance Z
Side-on overpressure
On the glass (reflected)
Positive phase
Impulse on the glass
Arrives after
Pressure on the glass against time after the shock arrives (the Friedlander pulse the engine is given). 0.22 kg is the TNT equivalent of a hand grenade's filling.

11The scenes

Armoured V8 under fire

The Rosser V8 from the driving simulator sits in a piazza with a 32 mm armoured windscreen: four 6 mm glass plies on polyurethane with a polycarbonate liner. The windscreen is the flat that best fits the car's curved screen, raked 69.5° from vertical. Five gunmen fire bursts at the driver, each round aimed with scatter and muzzle climb; the page works out where each round meets the windscreen's plane and fires it into the engine at the exact engine time it arrives, with its real angle. At that rake the rounds strike at a glancing angle, see a far thicker barrier (t/cos θ) and sometimes skip off. Rounds that miss strike sparks off the body. The barrage is generated from the pane's seed, so Replay shows it again exactly.

Grenade in the shop

A one-room glass shop on a street at night. Its front is a storefront: four floor-to-ceiling panels of 10 mm fully tempered glass (1.1 × 2.55 m), a frameless 10 mm glass door and a 6 mm transom over it; each other wall has three 6 mm tempered windows. That is fifteen panes, every one its own engine pane, stepped on one shared clock.

The Glass menu reglazes the shop. Every pane takes the type's layup, cut to its opening:

GlassStorefront and doorWindows and transomA thrown grenade (400 g, 25 m/s)
Tempered10 mm fully tempered6 mm fully temperedthrough the storefront every time; dice
Annealed10 mm annealed plate6 mm annealedthrough; long blades and guillotine shards
Laminated5 / 0.76 PVB / 53 / 0.38 PVB / 3caught by the storefront's PVB (a 7 cm bulge); tears through the thinner windows
Hurricane6 heat-strengthened / 2.28 ionoplast / 6 heat-strengthenedbounces off
Bullet-resistant32 mm glass-clad polycarbonate (UL 752 Level 3 class)bounces off

Where the storefront would stop the grenade, the door is propped open and the doorway is the default target: the dummy lobs it through at 12 m/s. You can still throw at the glass; it bounces back and goes off outside, in the street with the dummy. Laminated glass cracks in the blast but stays on its interlayer and in its frame, so a laminated, hurricane or bullet-resistant shop vents its gas only through the door: it stays pressurised longer, and less of the blast reaches the dummy (taken as 35 % of the free-field shock instead of 60 %, estimated).

What happens to the dummy

The dummy is a 1.85 m Hybrid-style crash dummy standing where it threw from. Three things can reach it, each worked out from the same blast and fragments that hit the glass:

The band round its chest turns from green to yellow (tempered dice, or a small chance of eardrum damage), orange (a fragment, an annealed shard, or a likely ruptured eardrum) or red (a fragment over 79 J, or lung injury). Red marks show where steel struck it, blue where glass did.

Panes behind panes

In the lab, + Add a pane behind puts up to three more panes behind the first, each its own glass at its own gap (2 mm to 1.5 m). Every pane is its own engine pane, solved at the same time on one clock. The engine knows nothing of the others; the page carries things across the gaps:

The result panel lists each pane: whether it held, cracked or went through, and what hit it. Shards a pane throws back can hit the one in front.

Sound: breaking glass with a note

Every pane has natural notes: the modes of its plate, each with its own pattern of swinging and still lines. A simply supported pane of width a, height b and bending stiffness D rings at

fmn = (π / 2) (m² / a² + n² / b²) √(D / ρh) m and n count the half-waves across and up. 2 mm picture glass, 30 × 25 cm, rings at 135 Hz: inside a voice's range. A 6 mm window, 1.2 × 1 m, at 25 Hz, below it. published: Navier's solution

The Sound panel plays a steady note at the glass. A loudspeaker or a voice square on is a uniform pressure on the face, twice the sound's own pressure because the glass reflects it; a uniform push can only drive modes odd in both directions. A shaker pushes at one point instead (off every line of symmetry) and drives them all. On a mode, each cycle adds a little to the swing until the damping takes away as much as the note puts in:

swing = static deflection × 1 / (2ζ),   built up over τ = 1 / (ζ ω) ζ is the modal damping: 0.5–2 % for glass in a frame, 0.5 % taken for the lightly held picture glass, so it swings about 100 times further than the same pressure would push it, after about 0.24 s. The engine's test plays it and gets 102. published

The engine checks every flaw 24 times a cycle. A flaw grows only near the peaks of tension, so the strength it is checked against is the one for a tenth of the time the note has played, the share of each cycle spent near its peak for glass's stress-corrosion exponent (n ≈ 16). estimated: that equivalent time. The meter shows the flaw nearest to failing against its own strength.

What it gives for the picture glass: 130 dB at the glass never breaks it (an 8 mm swing); 135 dB breaks it in 0.8 s; 145 dB in 0.15 s. Read those as a lower bound. At the break the glass is swinging about 7 times its own thickness, and the plate here is linear: a real pane that far out stretches like a drumskin, stiffens, and drifts off the note, so it needs more than this, or a note that slides up after it. The panel says when the swing passes the thickness.

Sing takes the pitch of your voice (autocorrelation) and its loudness from the microphone. A phone or laptop microphone is not calibrated: full scale is taken as about 120 dB SPL, so the voice level shown is rough, and a voice near the glass gives 90–110 dB. The boost (up to +60 dB) is what a loudspeaker would have to add. Mode view draws the swing the way time-averaged holography shows a vibrating plate: the still lines shine, and each dark band is a step more swing.

Prince Rupert's drop

The drop is molten glass dripped into cold water. Its skin freezes at once; the inside, still hot, shrinks as it cools and pulls the skin into very high compression, leaving the core in tension. The engine models it as a body of revolution: a spherical head (5 mm radius) thinning smoothly into a 75 mm tail, with the frozen axial stress a function of the share of the radius u only:

σ(u) = σt − (σt + σc) uk,   k = 2σc / σt That k makes the force over every section zero. Skin compression σc = 500 MPa; measurements by integrated photoelasticity (Aben, Chaudhri and others, 2016) found 400–700 MPa in the head. published Core tension σt = 100 MPa, so the compressed skin is 0.82 mm deep. estimated

Snapping the tail opens a crack straight into the core: the tail is too thin for its skin to stop it. The core's stored energy drives a front of forking cracks back to the head at the glass's terminal crack speed, 1,520 m/s (measured on drops: 1.45–1.9 km/s, Chandrasekar & Chaudhri 1994), so a 50 mm tail is gone in 33 µs. Behind the front the glass is in pieces about 1.4 mm across, from the same energy balance as tempered dice; each leaves with the speed its share of the stored energy gives it, faster in the skin than the core (5 % of it as motion estimated), a few to 12 m/s.

Hitting the head. A crack can only reach the core by getting through the compressed skin. A hard object (a hammer) presses as a Hertz contact and drives a cone crack whose length grows with the force (Roesler); it gets through once

χ F / d3/2 > KIc + 1.12 σ̄c √(π d) d is the skin's depth, σ̄c its mean compression (186 MPa), χ = 0.0136. About 20 kN for the 5 mm head; compression tests have found heads taking loads of the order of 15 kN. estimated: Roesler's law with the skin's compression closing the crack.

A soft bullet flows instead: its face presses at the stagnation pressure of its metal, Y + ½ρv² (Tate), and pulls the glass just outside it with (1 − 2ν)/2 of that (Love). A crack there must open against the skin's surface compression all the way down: 1.12 (σ − σc) √(πd) > KIc, about 513 MPa of pull. A .38 Special at 230 m/s gives 95 MPa and splashes; a 9 mm at 380 m/s, 290; a 7.62 rifle round at 850 m/s, 1,170: the drop bursts from the head. The splash flies off sideways in flakes. estimated: the flowing-contact criterion.

Polariscope shows the drop between crossed polarisers: the light along each line of sight is retarded by C ∫ σ dl (C = 2.7 × 10⁻¹² Pa⁻¹ for soda-lime), drawn for red, green and blue as sin²(π N λ₀/λ). Only the axial stress is counted. estimated

A window that explodes on its own

Tempered glass sometimes shatters years after it was fitted with nothing touching it. The culprit is usually a stone of nickel sulfide, a few tenths of a millimetre across, left in the glass when it was made. Tempering's heat leaves it in a compact crystal form; over months or years it slowly turns into its larger room-temperature form, swells by a few per cent, and opens a crack its own size. In the core of a tempered pane that crack sits in tension, which drives it like a penny-shaped crack:

K = 2 σt √(a / π) > KIc  ⇒  dc = 2π (KIc / 2σt)² σt is the core tension, a the stone's radius. For the lab's 10 mm storefront panel, stones over 0.35 mm break it. Glass makers heat-soak tempered panes (EN 14179) to set the stones off in the factory. estimated: the swelling only opens a crack the stone's size; what drives it after that is the core tension.

Once it goes, it is the same as any tempered break: the ply crazes into dice from that point, with the classic butterfly of two small pieces at the origin, and comes down from there.

Certification verdicts

Scenarios that reproduce a published test say whether the glass passed, by that test's own criterion, and what of the test is not simulated: ECE R43's 2.26 kg ball (the ball must not go through), ASTM E1996 missile D (no penetration on impact; the 9,000 pressure cycles that follow are not simulated), 14 CFR 25.775 bird strike (the windshield must keep the cabin sealed), ANSI Z97.1 (it must not break, or break into dice) and UL 752 (the round must be stopped; one shot of the three).

Odds of breaking

Every pane has its own scattering of flaws, drawn from the Weibull population of its glass, so the same shot breaks one pane and not the next. Odds of breaking at every speed fires the selected object into ten fresh panes at each of ten speeds, at the lab's own physics, on copies of the engine running in the background (one per spare core), and plots the share that break and the share it goes through. Where those curves cross 50 % is the speed you would quote; how steep they are is the Weibull modulus at work.

Read the cracks

A running crack stops when it meets one that is already there (section 6). Forensic examiners use that to order the shots in a window: the first impact's cracks run out to the frame, the second's stop at the first's, and so on. In the puzzle, three stones hit a laminated windscreen while you look away (laminated glass keeps every crack in place); tap the impact points in the order you think they happened, then Replay to watch them land.

A glass cutter

A cutter's wheel leaves a scratch with a vent crack under it, typically 10–20 % of the way through (12 % taken). That vent is a ready-made crack, so the score gives way at the tension that drives it, not at the strength of the glass:

σscore = KIc / (1.12 √(π c)) c is the vent's depth: 0.72 mm in 6 mm glass gives 14 MPa, against 100+ MPa elsewhere. The vent fails at KIc itself, so no loading-rate factor applies. estimated: the vent depth.

A crack that meets a score follows it, needing 60 % of the usual toughness along it (the vent has done part of the work) estimated, and does not fork off it. Bending closes a score on the side you push and opens it on the far side, which is why glaziers snap a score by bending it away and a burglar knocks the scored circle from the other side (or pulls on a suction cup). A rubber mallet spreads its force over a wide, soft contact, so it bends the pane without starting a Hertzian cone (Auerbach's law is measured with hard balls; a soft one never reaches the tension a cone needs).

Share links

Share copies a link that rebuilds exactly what is on screen: the glass, every pane behind it and its gap, the random seeds of each pane's flaws, and every shot (object, speed, point, angle, side and time). The fracture is deterministic, so whoever opens it sees the same cracks and the same pieces.

12Time and slow motion

The engine has its own clock. Each frame the page asks it to advance by the real time that passed times the playback speed — 1/10 to 1/100,000 — and the engine takes whatever steps the physics needs to get there (50 ns inside a bullet strike, microseconds while cracks run, milliseconds while pieces fly). Slow motion is not interpolation: at 1/100,000 the engine is simply asked for 0.17 µs of physics per frame.

Bullet time (the Auto button) slows down by itself as something approaches the glass, holds near 1/5,000 while it is in contact and cracks are running, and eases back up as the pieces fall. It eases down fast and up gently, in powers of ten.

Replay rebuilds the same pane from the same seed and refires the same shots at the same engine times. The engine is deterministic: the same shot gives the same cracks whatever the device's frame rate (checked at 60 and 997 frames per second).

13Runs on anything

One physics for every device. Every device computes the same pane at the same resolution: 32 plate modes along the longer side, a 2 mm crack grid, up to 500,000 cells a ply and 6,000 pieces. The resolution can tip a shot that is right on the edge of breaking (a ball just fast enough), so it is not left to the device: a link opens on the same break on a phone and on a desktop. A slow device takes longer over it instead; the clock waits for the physics.

Its own core. Where the browser allows it (a cross-origin isolated page, in Chrome, Edge and Firefox) the engine runs in a worker: the page asks it for each frame's cracks and pieces through shared memory, and while the frame is drawn the worker steps on toward the clock. Elsewhere it runs on the page, the same engine. It is built twice, with and without WebAssembly SIMD, and the two give bit-identical results.

Drawing to suit the device. The page times a fixed piece of work when it loads and picks how to draw:

TierResolutionGlassShadowsReal dice shown
Potato1×one pass: tint and reflectionbaked12,000
Low1.25×one pass: tint and reflectionbaked16,000
Medium1.5×one pass: tint and reflectionbaked24,000
High2×transmission, clearcoatshadow map32,000
Ultra2×transmission, clearcoatshadow map45,000

The cheap glass does in one pass what clear glass does to what is behind it: each face lets through a tinted share of it (a blend by a constant colour) and adds its reflection of the room, so it looks like the transmission pass of the higher tiers at a fraction of the cost. Below High the floor is flat-shaded (light falls evenly on a flat floor anyway) with the frame's shadow baked onto it from the key light, and each object casts a soft spot that follows it. Every tier multisamples on a real GPU (phones resolve it on the chip), lowers its resolution while a burst of thousands of pieces slows the frames and raises it again after, and stops drawing frames while nothing on screen can change.

In a scene with many panes the page also gives each pane a level of detail from the camera, a few times a second: a pane small on the screen updates its pieces every other frame, and a pane out of the picture or hidden behind a wall or the roof (tested along five lines of sight to its centre and corners) is not drawn and catches up every eighth frame. The physics of every pane is the same whatever its detail, so what you see when you turn to look is exactly what happened. Scenes of many panes at once (the shop's windows, the car under fire) compute each pane a step or two coarser than the lab, the same on every device.

The same link gives the same break on every tier, because the tier only draws it. If a device cannot keep up, the clock slows down rather than skipping physics, and the bar says "device-limited". Without WebGL the lab draws a flat front view.

14Sound

There are no recordings: every sound is synthesized in the browser from what the engine reports. Breaking glass is mostly noise, not tone, so the voices are built from sharp contact clicks and high, inharmonic resonances that die in milliseconds:

In slow motion everything drops in pitch and stretches by the same factor, (playback speed)0.2, so time audibly winds down. Sound starts muted; press Sound or M.

15Validation

Each scenario is run against the test standard it follows. The panes are flat, representative builds, so a pass means the model behaves the way glazing built to that standard does — not a certification.

ScenarioStandard and criterionThe engine
227 g steel ball from 2 m on 4 mm tempered side glassUN ECE R43: must not breakHolds
Window punch on 4 mm tempered side glassECE R43: 40–400 pieces per 50 × 50 mmAbout 156 per square
2.26 kg ball from 4 m on a laminated windscreenECE R43 / ANSI Z26.1: must not passBoth plies break, the PVB holds (62 mm bulge)
45 kg bag from 48 in on 6 mm annealed / temperedANSI Z97.1 / 16 CFR 1201Annealed: large shards; tempered: holds or dices safely
2×4 at 15.2 m/s on hurricane laminateASTM E1996 missile D: no penetrationGlass cracks, interlayer holds
9 mm on a 30 mm laminateUL 752 Level 1Stopped 13 mm in
.44 Magnum on 32 mm glass-clad polycarbonateUL 752 Level 3Stopped 19 mm in
7.62 × 51 on the same panelNot rated: should go throughThrough at 561 m/s
7.62 × 51 on 52 mm rifle-rated panelUL 752 Level 8Stopped 35 mm in
4 lb bird at 250 kt on the airliner windshield14 CFR 25.775: no penetrationHolds, cabin intact
9 mm on the 32 mm armoured windscreenEN 1063 BR4 / VPAM 4 classStopped 12 mm in; ricochets past 72°
A room blast on 6 mm temperedBlast-tested tempered glazing fails and is thrown clearFails ~5 ms after the shock, out at 15–40 m/s as dice

The plate itself is checked against closed forms: the fundamental frequency exactly, the centre deflection under a point load (0.0116 F a²/D) and under pressure (0.00406 p a⁴/D) within 1 %, and the exact stepping conserves energy to one part in a billion.

16What is estimated, and what is missing

Being honest about a model is part of it. These are reasoned values, not measurements:

  • the wander of running cracks (0.12 rad over 4 cm), the tightest bend (10 cm), the ragged edge of crushed zones;
  • the crazed sheet's friction (fitted to footage, not measured), and how long its edges keep crumbling;
  • the blast's 1 kPa hold and the gas jet's 0.6 m reach; the grenade's 800 effective fragments and its 0.22 kg TNT equivalent; the shop's venting speed; the share of the blast that gets past the storefront to the dummy, the fragments' 55 m drag length and the share of a spray of dice that hits it;
  • the 15 % cost of re-crushing rubble in an old crater, and the 72° ricochet angle;
  • when polymer dishes instead of plugging (0.3 of the projectile's diameter), how hard frames hold a laminate's edge, how a pulled-out sheet folds and what breaks its glass; the Level 10 build was sized so the model stops the .50, not taken from a product;
  • the windscreens and armoured layups are representative builds, not certified products.

And what the engine does not do yet:

  • panes are flat and simply supported: curved glass is flattened, clamped or bonded edges are not modelled;
  • the plate does not soften as it cracks (except under the impactor), so a heavily cracked laminate is too stiff;
  • polycarbonate and acrylic plies do not crack or craze; thermal stress (sunlit panes, heated windshields) is not modelled;
  • falling pieces do not collide with each other (only with the next pane of a stack), and walls do not stop them;
  • the plate is linear: a pane bent further than its own thickness does not stretch and stiffen as a real one does, so very large swings (a note on resonance, a big pane under a blast) are generous;
  • a blast is a pressure history given to the pane; the air itself (reflections round a room, the jet through each window) is not simulated.

17References