Some engineering failures need books. Some can scare the hell out of you in 29 lines of HTML.
Read the Sleipner A disaster page
Douglas N. Arnold’s page summarizes the sinking of the first concrete base structure for the Sleipner A offshore platform during a controlled ballasting operation near Stavanger, Norway, on August 23, 1991.
The structure developed a leak, failed and sank.
Nobody needs a dramatic soundtrack after that sentence.
The numerical error was not academic
The post-accident investigation traced the failure to a wall in a critical triangular section called a tricell. According to the page, the design problem came from a serious error in the finite-element analysis combined with insufficient reinforcement anchorage.
The numerical model underestimated shear stresses by about 47 percent.
That is the entire moral of computational engineering in one number.
A later, more careful analysis predicted failure at a depth of roughly 62 meters. The actual failure occurred at about 65 meters.
The first structure was lost completely. Arnold’s summary puts the economic loss at roughly $700 million and notes that the collapse produced a seismic event measuring about 3.0 on the Richter scale.
The platform deck involved tens of thousands of tons of structure and equipment. The failed concrete base ended up as debris hundreds of meters below the surface.
This was not a rounding error in a homework problem.
Why this page still works
Arnold’s page is part of a broader collection on disasters attributable to bad numerics. The Sleipner example is especially effective because the chain is so clean:
A numerical model produced stresses that were too low. The design followed the model. The concrete wall was not adequate. Reality ran the test suite.
The page links the summary back to investigation material and structural-engineering articles, and it uses a few diagrams and photographs instead of burying the event under interface chrome.
It was last modified in 2009 and still reads perfectly well because the information architecture is basically indestructible: title, facts, explanation, references.
There is a lesson there for both engineers and web developers.
Models are not reality, and presentation layers are not knowledge.
CacheRat’s 1,967 Ancient Web Domains research list exists to surface pages like this—small technical documents that have survived because their usefulness is much larger than their visual footprint.
You can read the Sleipner page in a couple of minutes.
You may remember the 47 percent for the rest of your career.