Design, failure and safety: how engineering knowledge grows: The design cycle, testing and how knowledge is stored
All courses / Understanding Engineering / Module 12: Design, failure and safety: how engineering knowledge grows Part 4 of 5The design cycle, testing and how knowledge is stored
Design goes round a loop of defining, imagining, choosing, building and testing, and each lap leaves something behind: a better part, a measured number or a new rule. This part follows the loop, the tests that feed it, the simulations that save tests, and the codes where the lessons end up.
4.1 The design cycle is drawn as a line and used as a loop, and errors cost more at every later step
Design textbooks, Kosky and colleagues among them, set the work out as a sequence of steps. In outline: say what the thing must do in numbers you can test, sketch several ways of doing it, pick one on stated grounds, work out its details, have it examined by people who did not draw it, then make it and see whether it meets the numbers you started with. The sequence is printed as a line, but in practice it is a loop, because a failed test sends the team back to an earlier stage.
Take a hand pump for a village well as an illustration. The definition says lift water from 30 m, deliver 20 litres a minute, survive a decade of daily use and cost under a set price. Concepts include a lever pump, a rotary pump and a diaphragm pump, and a weighted table chooses one. Detailing fixes wall thickness, seal material and handle length. A review asks what happens when someone hangs a child's weight on the handle. A prototype is then run for hundreds of thousands of strokes, the seals leak at 80,000, and the seal is redesigned. The lap ends when the thing has passed the test it was defined against.
Two points are often missed. First, a design is not finished until it can be made. Vincenti's study of flush riveting in aircraft shows a kind of design knowledge that lives in the workshop: how to prepare the holes in thin skin so that rivet heads end up level with the surface. It was learned by trial in production, not by calculation. Second, errors get costlier at each stage, as estimate 8 shows. The lesson is to check early: review every change to a load-carrying detail, and test the riskiest detail first.
| Stage where the flaw is found | Cost of fixing it | Working |
|---|---|---|
| Drawing review | Rs 5,000 | engineer time to revise the drawing |
| Prototype test | Rs 50,000 (calculated) | x 10: new parts and a repeat of the test |
| After production tooling is cut | Rs 5 lakh (calculated) | x 10: new moulds or dies, scrapped stock |
| After 50,000 units are sold | Rs 5 crore (calculated) | 50,000 units x Rs 1,000 each to recall and repair |
The factor of ten per stage is a rule of thumb, not a law, and the last step is a hundred times because a per-unit cost is multiplied by the number sold. Check: 50,000 x Rs 1,000 = Rs 5 crore (Rs 50,000,000). All figures are illustrative.
Source: Kosky et al., Exploring Engineering (Academic Press, 2013), part 3, ch. 17 to 24 (design steps); the cost figures are illustrative and the author's own- Kosky et al., Exploring Engineering (Academic Press, 2013) · part 3: ch. 17, 'Introduction to Engineering Design'; ch. 20, 'Design Step 3. Evaluation of Alternatives'; ch. 22, 'Design Step 5. Design Defense'; ch. 23, 'Design Step 6. Manufacturing and Testing'; ch. 24, 'Design Step 7. Performance Evaluation'
- Vincenti, What Engineers Know and How They Know It (Johns Hopkins University Press, 1990) · ch. 6, 'Design and Production: The Innovation of Flush Riveting in American Airplanes'
- Petroski, To Engineer Is Human (Macmillan, 1985) · ch. 7, 'Design as Revision'
4.2 Tests form a ladder from coupons to full scale, and the most informative test is the one that breaks the thing
Tests sit on a ladder. At the bottom are small coupons of material, pulled apart in a machine in their hundreds. Above them are joints and details, then whole components, then assemblies, and at the top one or two full-size articles. Each rung has fewer tests at greater cost and catches a different surprise: coupons show the scatter of the material, joint tests find the weak detail, and the full-scale test finds interactions nobody modelled.
The most informative test is the one that breaks the thing. A proof test shows only that the part survived a load. A test to failure shows how much more it could take and how it gave way, with warning or without. In 1995 Boeing loaded the wing of its 777 on a rig to its ultimate load, 150 per cent of the highest load expected in service, to confirm that the structure did not fail below the requirement. Pressure vessels are filled with water and loaded above their working pressure before use. From 1 October 2023, Bharat NCAP has rated new cars in India partly on a frontal crash at 64 km/h into a deformable barrier offset to one side, with dummies inside.
The energy involved explains the design. At 64 km/h a car and dummies of 1,400 kg carry about 220 kJ, the energy of a fall from 16 metres. The front must soak this up over about half a metre of crumpling while the cabin stays intact. Estimate 9 follows the chain. The trade-off shows: a soft front protects people by lengthening the stopping distance, but a car cannot grow indefinitely long, and a rigid cabin costs mass. Tests to failure destroy costly articles, so few are made, which is why scatter factors exist.
metal folds and the aluminium honeycomb barrier crushes over about 0.5 m; average deceleration = 17.8^2 / (2 x 0.5) = 316 m/s^2, about 32 g (calculated)
2 Cabinstays nearly undeformed, so the belt and airbag can slow the occupants over their own short distance
Outputs Heat and permanently bent metalalmost all of the 221 kJ Same energy as a fall from221,000 / (1,400 x 9.81) = 16 m (calculated)Check: g x h = v^2 / 2 gives h = 17.8^2 / 19.6 = 16 m (calculated). To hold the average to 20 g the front would need 17.8^2 / (2 x 20 x 9.81) = 0.81 m of crush (calculated). Round illustrative figures; the 64 km/h offset test is from the Bharat NCAP rules.
Source: Autocar India, 'Bharat NCAP crash test rules revealed ahead of October 1 roll-out' (2023), 64 km/h frontal offset test into a deformable barrier; https://www.autocarindia.com/industry/bharat-ncap-crash-test-rules-revealed-ahed-of-october-1-roll-out-429113; energy arithmetic by the author- Full-scale wing one or two articles, loaded to ultimate load
- Major assemblies wing box, fuel tank, control surface
- Components spars, ribs, skin panels
- Details and joints rows of fasteners, cut-outs, lugs
- Material coupons hundreds of small specimens
- Details and joints rows of fasteners, cut-outs, lugs
- Components spars, ribs, skin panels
- Major assemblies wing box, fuel tank, control surface
0.5 x 1,400 kg x (17.8 m/s)^2 = about 221 kJ, equal to a fall of v^2 / (2g) = 16 m
2 Crumple zoneassume the front stops the car in 0.6 m: average deceleration = v^2 / (2 x 0.6 m) = 263 m/s^2, about 27 g; average force = 1,400 x 263 = about 370 kN
3 Cabin and restraintsbelt and airbag spread the load over the occupants' strongest bones; the cabin must hold its shape
Outputs Bent metal and heatabout 221 kJRound numbers; real crash pulses vary and the 0.6 m is an assumption. Check: 16 m is about five storeys. Doubling the crumple distance to 1.2 m would halve the average deceleration.
Source: Autocar India, 'Bharat NCAP crash test rules revealed ahead of October 1 roll-out' (2023), 64 km/h offset deformable barrier, programme start 1 October 2023; https://www.autocarindia.com/industry/bharat-ncap-crash-test-rules-revealed-ahed-of-october-1-roll-out-429113- Kosky et al., Exploring Engineering (Academic Press, 2013) · ch. 23, 'Design Step 6. Manufacturing and Testing'; ch. 24, 'Design Step 7. Performance Evaluation'
- Petroski, To Engineer Is Human (Macmillan, 1985) · ch. 5, 'Success Is Foreseeing Failure'
- Boeing, '777 quality information' · 1995 wing static test to ultimate load, 150 per cent; https://www.boeing.com/commercial/777/quality-information
- Autocar India, 'Bharat NCAP crash test rules revealed ahead of October 1 roll-out' (2023) · web
4.3 Simulation tries thousands of variants cheaply, but only tests make it trustworthy
Computer simulation is Vincenti's selection by stand-in taken to its limit. A model of a car body, split into millions of small elements, can be crashed thousands of times in a week, and each run shows where the metal folds and how hard the occupant is hit. A model of air flow can try a hundred wing shapes before one goes to the tunnel. The wind tunnel for bridge decks has been joined by flow software. Most variations are bad, and simulation kills them cheaply.
Simulation does not remove the need for testing. A model is a hypothesis in code, and it earns trust only after it has predicted the results of real tests on simpler cases. The Sleipner platform of Part 1 shows what happens when a model is too coarse at the one place where the stress peaks. Petroski worried that engineers raised on computers would lose the slide-rule habit of judging roughly what an answer should be; a result shown to many decimal places looks precise, though it can be no better than the guesses fed in. Good teams build a ladder of trust: hand checks, then models checked against component tests, then full-scale tests.
The simplest habit is the sanity check. Before believing a computed deflection, estimate it by hand with a formula whose accuracy is known. Estimate 10 does this for a steel cantilever. If the program and the hand estimate differ by a factor of two, one of them is wrong, and in a surprising number of cases it is the program's input: a unit, a support condition, or a load applied in the wrong direction. The module Thinking in estimates makes this a way of working.
| Quantity | Working | Result |
|---|---|---|
| Bar | 20 mm x 20 mm steel, 1.0 m long, fixed at one end | E = 200 GPa |
| Second moment of area | b x h^3 / 12 = 0.02 x 0.02^3 / 12 | 1.33 x 10^-8 m^4 (calculated) |
| Load | 100 N at the free end | about 10 kg hung from the tip |
| Tip deflection by formula | P x L^3 / (3 x E x I) = 100 x 1 / (3 x 2 x 10^11 x 1.33 x 10^-8) | 0.0125 m = 12.5 mm (calculated) |
| What a simulation should return | the same, within a few per cent | a result near 1.25 mm or 125 mm points to an error in units or supports |
Check: 3 x 2 x 10^11 x 1.33 x 10^-8 = 8,000, and 100 / 8,000 = 0.0125. The cantilever formula is a standard result of beam theory.
Source: Petroski, To Engineer Is Human (Macmillan, 1985), ch. 15, 'From Slide Rule to Computer: Forgetting How It Used to Be Done'; standard cantilever formula, calculation by the author- Petroski, To Engineer Is Human (Macmillan, 1985) · ch. 15, 'From Slide Rule to Computer: Forgetting How It Used to Be Done'
- Vincenti, What Engineers Know and How They Know It (Johns Hopkins University Press, 1990) · ch. 8, 'A Variation-Selection Model for the Growth of Engineering Knowledge'
4.4 Codes and standards store what failures and tests have taught, which is why they change after disasters
What the loop learns must go somewhere, or each generation of engineers pays for the same lessons again. The store is the code or standard: an agreed document that sets out the loads to assume, the strengths to use, the margins to apply, the details to avoid and the tests to run. A code is failure and success compressed into rules. In India the Bureau of Indian Standards (BIS) publishes the IS codes: IS 456 for concrete, IS 800 for steel, IS 875 for loads and IS 1893 for earthquakes, gathered for buildings in the National Building Code of India. In Europe the Eurocodes, with EN 1990 as the base, do the same job.
Codes show their reasoning if you read the numbers. IS 800 divides the strength of steel by 1.10 where the limit is yielding, and by 1.25 where the limit is rupture of the net section or a bolt. The larger factor sits on the limit that comes with less warning. EN 1990 states its target outright. For ordinary buildings it asks for a reliability index of 3.8 over 50 years, a failure probability of about 7 in 100,000 (calculated), and it calibrates its partial factors, commonly 1.35 on permanent loads and 1.5 on variable loads, to deliver that on average. It also sorts structures into consequence classes, so that a grandstand or bridge gets more care than a farm shed.
Codes change after failures. IS 1893 first appeared in 1962, and its fifth revision of 2002 followed the Bhuj earthquake of 2001, with the ductile detailing code IS 13920 made mandatory in the higher seismic zones. The table lists other cases. Codes have limits. They freeze what was known on the day of writing and lag new methods, and compliance is not a guarantee: a bridge can comply on paper and be altered afterwards. A code is a floor under good engineering, not a substitute for it.
| Event | What the inquiry found | What changed |
|---|---|---|
| Tacoma Narrows Bridge, 1940 | wind forces on a slender deck were not in the design theory | wind-tunnel tests of models became standard for long suspension bridges; later decks were streamlined or stiffened against twist |
| Comet airliners, 1954 | fatigue of a pressurised fuselage had not been tested on the whole structure | British rules came to require fatigue tests on a complete pressure cabin; later aircraft were designed so that cracks are found and stopped |
| Hyatt Regency walkways, 1981 | an unchecked change doubled the load on a weak connection | sharper attention to who checks changes to connections and shop drawings |
| Challenger, 1986 | a cold-weather joint failure after known erosion had come to be treated as normal | redesign of the booster joint and stronger independent safety oversight |
| Bhuj earthquake, 2001 | weak detailing in buildings of the region | IS 1893 revised in 2002; ductile detailing to IS 13920 made mandatory in higher seismic zones |
Summarised from the sources cited; each change had many causes and took years.
Source: Washington State Department of Transportation, 'Tacoma Narrows Bridge history: lessons from failure', Carmody Board, wind-tunnel testing; https://wsdot.wa.gov/TNBhistory/bridges-failure.htm; National Information Centre of Earthquake Engineering (NICEE) and Delhi Disaster Management Authority, 'Earthquake Tip 11', IS 1893 first published 1962, fifth revision 2002, IS 13920 mandatory in zones III to V after Bhuj; https://ddma.delhi.gov.in/sites/default/files/ddma/EQT/eqtip11.pdf- Vincenti, What Engineers Know and How They Know It (Johns Hopkins University Press, 1990) · ch. 8, 'A Variation-Selection Model for the Growth of Engineering Knowledge' (preserving and propagating knowledge in codes, handbooks and teaching)
- Bureau of Indian Standards, IS 800:2007, General construction in steel · cl. 5.4.1 and Table 5: 1.10 for yielding and buckling, 1.25 for ultimate stress and for bolts; https://law.resource.org/pub/in/bis/S03/is.800.2007.pdf
- Steel Buildings in Europe, via Infosteel · EN 1990 Table A1.2(B): 1.35 on permanent and 1.50 on variable actions; https://infosteel.be/images/Steel-Buildings-in-Europe/MSB03/14
- Joint Research Centre, European Commission, presentation on EN 1990, Basis of structural design · reliability index 3.8 for 50 years, consequence classes CC1 to CC3; https://eurocodes.jrc.ec.europa.eu/sites/default/files/2022-06/Moscow08_9_Gulvanessian.pdf
See it in the table
H₂O WaterSt SteelCheck your understanding
A 1,400 kg car hits a barrier at 64 km/h, carrying about 220 kJ. A fall from what height has the same energy?
Equal energy means g x h = v^2 / 2, so h = 17.8^2 / (2 x 9.81) = 16 m, about five storeys.Why does a test to failure tell an engineer more than a proof test?
A proof test records survival at one load. Pushing to failure shows how much more the part could take and whether it gave warning.IS 800 uses 1.10 for yielding and 1.25 for rupture. Which reading fits the idea of matching margin to warning?
Ductile steel stretches visibly before yielding, while net-section or bolt rupture gives little warning. A larger factor on the abrupt limit is consistent with codes setting margins by consequence.