The Explosive Lab

Six materials matter here, and only three of them are charges. The other three are the supporting cast that carries fire from one charge to the next — the reason a domino line works at all.

Material	Id	Role	Distinguishing behaviour
FIRECRACKER	21	charge	Fast fuse, hard shell. Wants to be well cracked and well pressurised before it lets go.
REACTIVE_BURN	25	charge	Deflagration, not detonation. Cracks earliest, pushes least, and its gas comes from pyrolysis rather than a fuse.
SEALED_CHARGE	62	charge	The keg. Highest pressure, fastest fragments, and the only charge whose ignition temperature is randomised per particle.
FLAMMABLE_FLUID	16	fuel	A spreading pool that relights itself at ~465 K while oxygen holds. This is how fire crosses a floor.
TINDER / PEAT	— / 89	fuse	Ignite at 340 K and 360 K. Cheap, slow-burning bridges between stations.
EMBER	8	residue	Burns at 350 K and stays hot for a long time afterwards. Turns a spent station into a delayed igniter.

One sentence

That single inequality is the entire chain-reaction mechanism. There is no propagation code, no event queue, no scripted delay. A charge heats the air around it, the air heats the next charge, and the next charge crosses its own threshold whenever it happens to cross it.

Why the timing looks organic

Because it is not timed. Each particle carries a per-particle seed (base_params.w, called shell_seed in the shader, with core_seed = 1 - shell_seed) and that seed perturbs ignition rate, gas yield, shell strength and fragment speed independently. A sealed charge goes further and seeds the ignition temperature itself across a 48 K band, so a single keg does not light all at once — it lights from wherever the hot side is.

Three lanes, borrowed

A charge needs three pieces of persistent per-particle state: how far the fuse has burnt, how much gas is inside, and how broken the shell is. It does not get three new buffers for that. The P2G shader sits at exactly sixteen std430 blocks — the driver cap — so there is no room for another one.

Instead the charges reuse three lanes that already exist for other purposes:

Lane	Normally holds	Holds for a charge
damage[idx]	continuum damage / burn progress	fuse or cook progress, 0–1
phase_data[idx]	phase fraction	gas fraction inside the shell, 0–1
Jp_data[idx]	plastic volume change	crack, 0–1

The Jp borrow is the one with a visible consequence. Because Jp no longer means volume, all three charges appear in material_uses_zero_jp() in mpm_solver.cpp and spawn with Jp = 0 instead of the usual 1. A charge that spawned at Jp = 1 would be born fully cracked and would vent on its first warm frame.

The four stages

The loop that decides everything

Crack reduces sealed, which reduces pressure, which reduces crack growth. That negative feedback is drawn dashed in the figure and it is what separates a bang from a fizzle. A charge that cracks slowly bleeds its pressure away through the seal term and never reaches the vent gate; a charge that stays sealed until it is fully pressurised crosses both smoothsteps at once. The three charges differ mostly in how fast that race runs.

The margin is 152 K

The highest ignition threshold anywhere in the family is 368 K, at the top of the sealed charge’s seeded band. The lowest vent floor anywhere in the family is 520 K, on reactive burn. The gap between them is 152 K, and it is one-directional: every vent clears every threshold, with room to spare.

This is why chains are robust rather than fragile. A domino line does not depend on tuning each station against the next one; it depends on a single inequality that holds for all nine pairs at once. The only thing that varies from station to station is when, not whether.

Where the delay comes from

Nothing in the shader waits. The delay between stations is transport: the vent dumps heat into the air grid, the air solve advects and diffuses it, and the next charge samples its own temperature some number of substeps later. Move the stations further apart and the chain slows down on its own. Put a draught across the line and it bends.

Why the sealed charge’s threshold is a band

Every other material on the axis is a single tick. The sealed charge is a 48 K bar, because its threshold is mix(320.0, 368.0, shell_seed) — drawn independently for every particle in the body.

The consequence is worth stating plainly, because it is the difference between a charge that reads as an object and one that reads as a texture. With a single threshold, a heated keg crosses it everywhere at once and the whole body converts in the same frame: a uniform, synthetic-looking flash. With a seeded threshold, the particles nearest the heat cross first, their vents heat their neighbours, and the conversion travels through the body as a front. The keg tears open from the hot side.

Four things happen at once

The vent block is short, and each of its four lines exists to prevent a specific failure.

Line	What it does	What breaks without it
vel += burst_dir * burst_speed	Kicks the particle along a uniformly-sampled angle, vec2(cos(a), sin(a)) with a = hash × 2π.	Sampling a random vector and normalising it clusters fragments toward the diagonals. Sampling the angle is uniform on the circle by construction.
gas *= 1 - vent·k	Dumps the gas that just did the work.	The pressure term would still be high next substep and the same particle would vent again, forever.
C *= 1 - vent·k	Damps the affine velocity matrix.	APIC would carry the pre-vent local spin into a fragment that no longer has anything to spin about, and the debris would corkscrew.
temp = max(temp, floor)	Forces the particle to at least 520/680/760 K.	This is the entire chain mechanism, and also the entire signal the CPU-side detector watches. Without it a detonation is invisible to everything downstream.

And then the shell gives up

After the vent block, all three charges run the same closing move: stress is multiplied down by the crack (to as little as 2–5% of its value), and past a crack threshold the deformation gradient is blended toward an isotropic mat2(sqrt(J)).

That second part is the one that makes fragments look like fragments. Blending F toward isotropic discards the accumulated shear history — the particle forgets which way it was being stretched. What is left is a lump with momentum, which is what a piece of a burst casing is. Keeping the full F would leave every fragment still trying to satisfy an elastic shape it no longer has any business having.

The blend rates differ, and they read differently on screen: the firecracker blends hardest (0.22) and produces the most uniform debris; the sealed charge blends least (mix(0.18, 0.08, shell_seed)) and keeps more structure, so its fragments stay chunkier.

The problem

The blast effect stack — shockwave rings, explosion details, debris, foam, scorch and crack decals, blast lights, screen shake, the detonation flash and the volumetric god rays — all hung off a just_ruptured hook that only fires for pressure vessels, and pressure vessels were only ever created by the projectile-firing path.

Scene-authored charges detonate per-particle inside the P2G vent branch, entirely on the GPU. They produced no CPU event at all. The measured consequence: 35 seconds of a scene built around chained explosions reported blast_lights = 0, flash = 0.000, trauma = 0.000 for the whole run. Roughly eight effects were invisible in exactly the scenes authored to show explosions off.

Why rate of change does not work

The obvious detector — watch for a big frame-to-frame jump in air temperature — was built first and found nothing. Detection runs per physics substep, and a vent’s heat diffuses into the air grid over many of them, so the per-step delta is only 1–8 K even while the grid is reaching 1000 K. The first model measured 1.3–8.2 K against a 220 K threshold: it never fired once.

What does work

What separates a detonation from a steady fire is not how fast the temperature changes but how far it gets from its own recent average. A fire’s baseline catches up with it. A detonation outruns its baseline.

So each bin of the air grid keeps a slow running average (τ = 0.85 s) and fires when the current temperature excurses far enough above it and ends above an absolute floor. Measured on the chain scene, detonations peak at 243–259 K of excursion and steady fire flicker peaks at 108–136 K. The default threshold of 165 K sits in the empty gap with margin on both sides.

The air temperature grid is already downloaded to the CPU every frame — to be re-uploaded as a texture — and then discarded. Retaining that copy is the entire cost of the feature: no new readback, no new GPU work, and no hook inside the physics path.

The four authored thermal channels

Every charge reads a vec4 of thermal coupling, fetched per particle from a texture and falling back to a per-material default when unauthored. The four channels are named after what they do, and each charge uses a different subset:

Channel	Name in shader	Effect	Used by
.x	batch_outgas	Scales the gas target directly.	sealed charge only
.y	batch_heat	Adds to the self-heating rate.	all three
.z	batch_cooling	Scales the ignition rate.	all three
.w	batch_loft	Adds 100–120 K to the vent temperature floor.	all three

batch_cooling is the confusingly-named one: it does not cool anything in these branches, it modulates how quickly the fuse takes. It is shared with the baking materials, where the name makes more sense.

Pseudo-2.5D enclosure

With the pseudo-2.5D system enabled, the enclosure knob produces pseudo_enclosure_hold = 1 + enclosure × (0.14 + 0.86 × core_seed), and the sealed charge folds it straight into its cavity term:

cavity_hold = (1.46 + pseudo_enclosure_hold) × mix(2.30, 0.88, shell_seed)

With pseudo-2.5D off that resolves to 2.46 × mix(2.30, 0.88, shell_seed) — between 2.16 and 5.66 depending on the seed. Turning enclosure up raises the whole band, so a buried keg genuinely holds more pressure before it lets go, and lets go harder when it does.

Of the three charges, only the sealed charge reads this. The firecracker and reactive burn ignore it entirely, which is intentional: a firecracker is a small thing in open air and burying it should not turn it into a bomb.

Where to look in the app

The Scene Blast Events checkbox and its four sliders live in the effects section. The blast-model selector (Legacy per-frame push versus Sedov Front) is a separate, independent toggle that governs how the shockwave is drawn, not how the charge behaves.

Honest limitations

• There is no detonation physics. Nothing here models a reaction front, a Chapman–Jouguet state or an equation of state for hot product gas. What exists is a scalar gas fraction that inflates a hyperelastic solid until it tears. It is a pressure vessel, and it is honest about being one.
• The burst direction is uniformly random, not radial. Fragments do not fly outward from a centre — each particle picks its own angle. The outward appearance comes from the fact that the outside of the charge vents first and the inside is still solid. A charge that vents from its middle outward will look wrong.
• Fragment speed is applied per substep while the gate is open, so total delivered impulse depends on how long the gate stays open, which depends on how fast gas bleeds. The quoted 28–84 m/s is a per-substep ceiling, not a muzzle velocity.
• The detector cannot distinguish two charges in the same bin. Bins are 8 grid cells on a side with a 0.45 s cooldown, so a tight cluster reports one event, not five.
• Terrain never moves. The SDF pushes on particles; particles never push back. A charge against a wall cannot crater it. See the coupling graph on the Relationship Map — that missing return edge is drawn there explicitly.

Things that look like bugs and are not

• A charge that heats slowly may never go off. That is the seal-bleed feedback working as designed: crack grows, sealed falls, pressure never reaches the gate. Heat it faster and it detonates.
• A sealed charge tears open asymmetrically. Intended — the ignition threshold is seeded per particle across 48 K, so the hot side converts first.
• Reactive burn makes a lot of noise and little shrapnel. Also intended: it cracks earliest and pushes least of the three. It is the deflagration in the family.
• Nothing happens below 405 K with reactive burn even though it is visibly charring. Charring starts at 350 K but gas generation does not begin until 405 K. There is a 55 K window where it looks like it is doing something and physically is not.