Mechanism · a length everyone measures from the wrong end

The Fire Too Small to Notice

A sprinkler head is a thermometer with one job, and it does that job alone. The length that decides whether it ever notices a fire is not the height of the room. It is the clearance from the top of whatever is burning up to the ceiling, and in a building nobody is going to rebuild, the fuel is the end of that length that actually moves.

In the film, the alarm goes off and the whole ceiling rains at once. Real wet-pipe systems almost never do that. In the United States fire-incident record for 2017 to 2021, as published by NFPA:

One sprinkler is usually enough to control a fire. In 77 percent of the structure fires where sprinklers operated, only one operated. In 96 percent, five or fewer operated.

Quoted from the report's key findings, in order and unedited. Four pages later the same report gives the figure a second time and splits it: 76 percent for sprinklers of any type, 77 percent for wet pipe. Neither number is doing any work below.

Read the first sentence again, because it is the explanation and not the puzzle. One head opens, water lands on the fire, the fire stops growing, the gas under the ceiling cools, and the neighbours never get their turn. That is a control loop, and it is why the number is 77 and not 7. The statistic is evidence that water works. It is not evidence about the thing this page is about, which is what happens in the seconds before any water falls, when each head is still nothing but a small piece of metal or glass hanging in moving air, deciding for itself.

That decision has a length in it, and almost everyone has the wrong length.

The ceiling, from above

Below is a flat unobstructed ceiling with sprinkler heads on a square grid, seen from above, and beside it the same room from the side. Move the fire. Each head computes its own answer from its own distance to the plume, so the one that opens is the one the fire happens to be nearest, and the rest are simply too far away to have heard about it yet.

You can also drag the fire on the plan view. The two sliders are the control; dragging just moves them.

Drag the fire to a head and then to the middle of four, and the whole of layer one is in that one gesture. At the settings the page opens with, a fire essentially under a head opens it after 52 s at 128 kW, and its neighbours are nowhere. The same fire at the worst point in the grid takes 122 s and 699 kW, and opens four heads within a second of each other, because four heads are the same distance away. Where the fire happens to sit in the grid is worth a factor of 5.47 in the size of the fire when the ceiling finally reacts, and nobody chooses where the fire sits.

What the element is actually measuring

Not the fire. A sprinkler head cannot see a fire. It sits in a thin sheet of hot gas that the plume spreads sideways under the ceiling, called the ceiling jet, and it measures that. Alpert fitted the temperature and speed of that sheet in the 1970s, and the fit is still what the industry calculates with:

Q is the fire's heat release rate in kilowatts, r is the distance from the plume centre to the head, and H is the one that matters. In the paper that quotes these correlations, H is defined in a single clause that is easy to read straight past:

and H is the height between ceiling and fuel surface, m.

Not the ceiling height. The height of the rise. The plume drags room air into itself all the way up, and that dilution is where the fifth-thirds power comes from: a plume with less room to rise has entrained less and arrives hotter. So the length in the equation is the clearance, and a building has as many clearances as it has things to burn on.

The element then lags the gas. A bulb or a solder link takes time to heat, and the standard model of that lag, which NIST attributes to Heskestad and Bill, is a first-order equation with the gas speed in it:

Give the fire the usual growing shape, Q = alpha t^2, and both of those substitutions turn out to be powers of t, which means the equation can be solved exactly rather than stepped. The page does that, and the solution is short:

Those two numbers are usually not equal, and that is worth noticing. The element's lag tends to a constant, but in a fast fire under a low ceiling the head has already opened while the transient is still running, so the trail at the moment of opening is smaller than the value it was heading for. In other words the element never catches up and never settles: it opens out of a gas that is already much hotter than the number stamped on it.

The curve below is that solution for the first head to open and for the next one out, with the gas temperature the first head is sitting in drawn faintly behind it. The gap between the two link curves is the head start the first head gets, and it is the reason a system that is only a set of independent thermometers still behaves, in the record, like a system that aims.

First head reaches its rating at 122 s, out of gas that is by then 134 C, which is 66 C hotter than the element reads. The next head out is still 25.9 C below its rating at that moment. When that head would open is outside this model: by then the flame is at the ceiling.

The fire too small to notice

Hold the fire steady instead of letting it grow and the question sharpens. The gas temperature under the ceiling settles at a value fixed by Q, H and r, so there is a heat release rate below which it simply never reaches the element's rating, no matter how long the fire burns. Invert the correlation and that threshold is a closed form with no data in it at all:

That second line is worth a moment. The H^(5/3) and the (r/H)^(2/3) partly cancel, so once the head is outside the plume's impingement patch, the threshold goes as the three-halves power of the clearance and the first power of the radius. The famous five-halves power is the other branch: it applies when the head is essentially directly over the fire, which is the luckiest case and not the one anybody designs for. This page quotes the worst case, and it says so in every table.

This table is frozen at the case the paragraphs around it quote: a 3.6 m grid, a 68 C element, 20 C air, fire midway between four heads. It does not follow your sliders, because the sentences either side of it are fixed text and a live table under fixed text is a contradiction waiting for your first slider move. For your own settings read the instrument at the top. Worst-case radius 2.55 m, from the 3.6 m grid above. Engine A is Alpert's 1972 fit; engine B is Alpert and Heskestad's later dimensionless one, which this page runs beside it precisely because they are not the same equation. The exponent columns are measured from the curve, not asserted. Rows marked out sit outside engine B's stated validity band of r/H below 4, and Alpert's own measurements mostly sat below r/H of about 1, so every row here is an extrapolation of one fit or the other. The table is the steady threshold only and carries none of the instrument's refusals: at half a metre of clearance the ceiling jet is thinner than the element hangs below the ceiling, and the instrument above declines that case outright.

Take the two ends of that column. At 2,820 kW, twelve metres of clearance is not a detector at all for anything a person would call a small fire; at two metres the same ceiling, the same head and the same spacing answer at 192 kW. And if you are lucky enough to have put the fire directly under a head, so that the plume-impingement branch applies at both ends, the five-halves law comes back in its pure form: 2,388 kW at twelve metres of clearance against 4.79 kW at one, a ratio of 498.8, which is 12^(5/2) to four figures. The worst case and the best case differ by exactly 5.567 times r/H, which is why the distinction matters most in a low room and hardly at all in a tall one.

Two warnings before you carry that threshold anywhere. It is a property of an unconfined ceiling: an open volume where the hot gas spreads and leaves. A closed room does not do that. It fills from the top down, and room tests cited by Zeng et al. have 57 to 68 C heads opening at 50 to 200 kW where this correlation would say nothing should happen. So the fire too small to notice is a fact about atria, hangars, car parks and high open warehouse volume, and not about your bedroom. And the film image is not universally false: deluge systems, in hangars and some process areas, are built with open heads and release every one of them at once, on purpose.

Yes, obviously. It is a tall room.

Here is the sentence a well-informed reader has ready by now, and it is worth putting at its strongest rather than paraphrasing it down:

Of course a fire is harder to detect in a big warehouse. The smoke has twelve metres to climb instead of three, and it is cold and spread out by the time it gets there. That is just distance. Tall building, slow detection.

The model inside that sentence is the building decides. It is a good model. It agrees with the clearance model on every fire that burns on the floor, which is most of the fires anybody pictures, and it agrees with Alpert's own experiments, because those were pool and spray fires burning at floor level under a high roof, where the two lengths are the same number. Nothing in that dataset can tell the two apart.

So put them somewhere they disagree. Below are four rooms. Two of them have a 12 m ceiling and two have a 2.7 m ceiling, and the fire, the head, the spacing and the air are identical in all four. Every figure is computed in your browser at the settings named under the table.

Fast growth, 3.6 m square spacing, fire midway between four heads, 68 C element, RTI 80, 20 C air, 1.5 m fire base. Engine A. The two highlighted rows are where the two models are forced apart; whether that convicts either of them is the next section.

Look at the two highlighted rows. A 12 m warehouse and a 2.7 m office, buildings whose heights differ by a factor of 4.44, produce the same threshold to every digit, the same activation time to every digit, and the same fire size at activation to every digit, because both have two metres of air above the thing that is burning. The model that says the building decides has to put those two rooms a factor of 9.37 apart. It cannot put them together, and it cannot be repaired to, because the building height is the only length it has.

And in the same 12 m shell, moving the fuel from the floor to the top of a 10 m rack drops the threshold from 2,820 kW to 192 kW, a factor of 14.70, which is exactly 6^(3/2). The second engine, which shares no constants with the first, puts the same collapse at 21.73. For the growing fire the effect is milder because the fire has to grow through the threshold anyway: the time to open falls by a factor of 2.31 and the fire size at opening by 5.34. The building never changed. Nobody moved a sprinkler. Somebody filled the racks.

That is a statement about the model, and a real rack fire will not hand it to you. The preset is labelled fire on top for a reason. FM Global's full-scale tests of almost exactly this geometry, 10.7 m of storage under a 12.2 m ceiling, ignite at the base: Igniters are prepared just prior to ignition and placed at the base of the bottom pallet-loads within the central transverse flue of the main array. The first sprinkler in that test opened at 4 min 38 s, and the report says what the fire was doing at that moment: The fire reached the top of the fifth tier (approximately 25 ft [7.6 m]) and flames were up to the third tier on the east face of the ignition array when the first sprinkler actuated. So the clearance above the burning surface at the moment of detection was about 4.6 m, not the 1.5 m above the top of the stack, and eight sprinklers opened before that test was done. In a real rack fire the clearance is a function of time that starts at the full ceiling height and collapses as the fire climbs. This page holds it fixed, which is a choice, and the number it produces is the clearance the fire would see if it started at the top.

The part where the other model contradicts itself

That is a disagreement, not yet a decision. Neither table can say which model is right, because the page has no fire test in it. But one of the two can be caught telling a story that cannot happen, and the reader already owns the observation that catches it: a lighter held under a smoke alarm sets it off, and the same lighter across the room does not. Same room, same ceiling height, two answers. The building-height model predicts one.

Here is the arithmetic version, which the page can run. Take the racked case. The building-height model says nothing opens until the fire reaches 3,735 kW. Ask Heskestad's flame-height correlation how tall a flame of that size is, and it answers 4.50 m, standing in 2.00 m of clearance: a factor of 2.25 more flame than there is room for. Sweep the fire base diameter across its whole range, from 0.3 m to 3 m, and the flame stays taller than the clearance the entire way, between 1.48 and 2.86 times. The model's own answer requires the fire to be burning through the roof before the head has noticed anything.

The clearance model does not do that. It opens the head at 699 kW, and flame reaches that particular ceiling at 980 kW, so its answer arrives while there is still a smoke ceiling jet to arrive in. The same test run on the floor-fire row comes back clean for both models, which is the point: a test that fired everywhere would be testing nothing.

It also does not fire everywhere it would need to. Run it across 560 settings of this instrument in which the two models genuinely differ, and it catches the building-height model in 384 of them, or 68.6 percent. The office row in the table above is one of the ones it misses: the building-height model says 881 kW there, which is a 1.85 m flame in 2.00 m of clearance, and nothing about that is impossible. So this is a test that convicts the other model where the ceiling is high and the fuel is stacked, and stays silent in a low room. That is a real limit on the argument and not a footnote to it.

Where the model stops, and says so

Push the rack up one more metre and the clearance model runs out too. At 11 m of racking in the same shell, engine A puts the head opening at 427 kW, and flame reaches that one-metre ceiling at 426 kW. The flame gets there first, by about a second, and from that instant this is not a smoke ceiling jet at all but a flame touching a ceiling, which Alpert's correlation does not describe. So the page refuses the number rather than print it, and says which way the error would have gone: a real head would open sooner, not later. Press the last preset button above and watch it decline.

Engine B, on the same case, does not refuse. It opens the head at 328 kW, comfortably under the flame-height bound, and so it answers where engine A will not. Two published correlations disagreeing about whether a model applies at all is the honest state of this problem at one metre of clearance, and it is printed rather than resolved.