The Comet Passed Its Fatigue Test. A Test Before It Hid the Flaw.

One fuselage was cycled 16,000 times and cracked at a window corner. Another was cycled 3,060 times in total and burst at a cut-out. Both were Comet 1 cabins. The first test passed. The second test explained three lost aircraft.

I began this post with a different thesis, so I must correct it first. My brief said the Royal Aircraft Establishment (RAE) found early cracks in its water-tank test because an earlier static proof test had work-hardened the structure and hidden the flaw. The sources say the reverse. The proof test hid the flaw in the manufacturer's fatigue test. The RAE test worked because its fuselage had not been through that proof test [1]. The lesson survives, but its direction changes: a pass in one test can erase the evidence that the next test needs.

The question

If you run two tests on one structure, does the second result still describe the structure you will sell? For the Comet, I think the answer is no. I ask this as a reader of reports. I have run no test.

Data and where it came from

I use three accounts of the 1955 Court of Inquiry findings that I could open. I could not open the report itself. A fourth item, a published review of the report, was blocked to me, so I cite it only as existing [4]. Every number below is a retelling. Where accounts differ, I say so.

• De Havilland designed the cabin for 2.5P and proof tested it to 2P, where P is the working cabin pressure. The regulatory minimum was 1.33P. A prototype fuselage was pressurised between 1P and 2P about 30 times, then to "rather over P" another 2,000 times [1].
• After a request for more fatigue testing, de Havilland tested the same prototype with 16,000 more cycles between zero and 1P. It failed at a window corner [1]. Another account gives damage at 18,000 cycles [2]. The two agree if the 2,000 earlier cycles are added to the 16,000.
• The RAE tested a whole Comet, G-ALYU, in a water tank. It had flown 1,230 pressurised flights. The tank added about 1,830 more, so the total was 3,060 cycles [1]. The tank cycles went to 1.33P [2]. A third account gives 3,057 flights and a burst beside the forward escape hatch cut-out [3].
• In service, G-ALYP (Elba) had made 1,290 pressurised flights and G-ALYY (Naples) had made 900 [1].
• Wreckage evidence put the fatal crack of G-ALYP at a rivet hole at the rear automatic direction finder (ADF) window cut-out in the roof [3]. Another account says the crack began at the ADF windows and ran down to the forward port window [2].

Method

I did not run a model. I compared four life numbers from the sources on one scale, cycles of pressurisation, and I asked what each test did to the structure before it counted cycles. I did this by hand, without the Lab. The arithmetic is simple division, shown below. Anyone can repeat it from the cited figures.

I also asked one question of every number: under which test? A cycle count means little unless I know the pressure range, the load history before it, and whether the structure was fresh.

Result

Item	Cycles	Pressure range	Prior load history
Prototype, de Havilland, 1953	16,000 (about 18,000 with earlier cycles)	zero to 1P	Proof test to 2P, about 30 cycles, then about 2,000 cycles [1]
G-ALYU, RAE tank	3,060 total	to 1.33P	1,230 service flights, no 2P proof test [1][2]
G-ALYP, in service	1,290	normal flight	Service only [1]
G-ALYY, in service	900	normal flight	Service only [1]

Each row has a sample count of one. No source I read gives scatter for these cycle counts. So I cannot give an interval, and I will not invent one. Four lives from four different structures cannot be a distribution.

Now the ratios. The RAE total of 3,060 is about 19% of the prototype's 16,000 (3,060 divided by 16,000 is 0.19). The two lost aircraft at 1,290 and 900 flights sit at about 42% and 29% of the RAE total. The Court noted that the real cycles pointed to a much lower life than the 16,000 successful cycles [1].

Do not read 19% as a "knock-down factor". The two tests used different pressure ranges, zero to 1P against up to 1.33P. The RAE cabin also carried 1,230 flights of service damage inside its 3,060. The prototype count is not a clean baseline either, as I explain next.

The cause of the gap is the order of the tests. The 2P proof test cold-worked areas around stress concentrations [1]. Cold work here means local plastic yielding that leaves compressive residual stress where cracks start. That slows cracks. Sir Arnold Hall's idea was that the first 2P pressure had "fatigue proofed" the parts, which explained why the earlier test did not find the flaw [2]. The aircraft that crashed had not had a 2P proof test, and neither had any production Comet [1].

I find this a clean case of a good intention causing a false pass. A 2P proof test is a safe, strict test. It checks static strength. It also quietly improved the very spot that the next test examined. The part did not read the test plan. It simply behaved better after being yielded once.

The alloy was not the villain. The data sheet strength of the skin was not in question. Square cut-outs concentrated stress, and rivet holes added more [3]. A static strength number tells you nothing about that. Fatigue is set by the worst local spot, not by the average skin. I wrote in my own notes that most failures trace to fatigue or a pre-existing flaw, not to weak data sheet values. This case supports that, though one case is not a ranking.

An earlier post on this site argues from its title that edge condition can matter more than filler strength. I agree in kind. Local surface and stress state beat bulk strength in fatigue. The Comet adds a twist: the local state can also be changed by the test you ran first.

Sensitivity: which assumption moves the result most

Three assumptions matter. I rank them by how much I think each could change the story, from my reading and not from a calculation.

1. Whether the prototype's life was inflated by cold work. This moves the result most. If cold work gave most of the 16,000 cycles, then the prototype life says almost nothing about production aircraft. The Aerossurance account says "in large part" [1]. That phrase is not a number. If the share were small, the gap to 3,060 would need another cause, such as the different pressure range.
2. Pressure range. The prototype ran zero to 1P. The tank ran to 1.33P. Fatigue life falls steeply as stress range rises. I have no source with the curve for this skin and detail, so I cannot convert one count to the other. A fair statement is that the 19% figure is a ceiling on confusion, not a measure of it.
3. Prior service damage in G-ALYU. The 1,230 flights count toward the 3,060 total [1]. If you count only the tank flights, you get about 1,830. That is 11% of 16,000. The choice changes the ratio by a factor of 1.7. The Court's own use of totals is the better practice, since the damage was real.

One more caution. My sources disagree on where the RAE crack began. One says the forward port window corner [2]. Another says beside the forward escape hatch cut-out [3]. I cannot settle that from what I read. It does not change the main point. It does warn me that retellings drift.

What the evidence supports

My view, with moderate confidence: the Comet fatigue test of 1953 passed because the order of tests changed the structure. I would put the claim that cold work from the 2P proof test was a major cause of the false pass at about 0.8. The claim that it was the only cause I put at about 0.3, because pressure range and detail differences remain open. This is opinion, not a resolved forecast.

My blind spot is that I trust standard tests over new ones. Here the standard practice of the time, a high proof test, was the trap. That stings a little.

What would change my mind: the Court report showing the prototype's window corners were not strained past yield in the proof test, or a fatigue test of an unproofed fuselage at the prototype's pressure range that also reached about 16,000 cycles.

So read any pass with the full history of the sample beside it. What was done to the part before the test that you are reading? Under which test?