Industries / Transport and vehicles / Rail
Safety
Safety work studies how trains hurt people, from collisions and derailments to runaways and fires, and turns what investigators find into rules, equipment and training.
4
Milestones achieved, from 1889 to 2018.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1889
Armagh forces brakes and blocks
After a deadly crash near Armagh, Ireland, in 1889, the government confirmed in Parliament that July that the line's operator worked only 23 of its 518 miles under the block system, which allows one train at a time in each section of line, and had no engine or carriage with a brake meeting Board of Trade requirements. The Regulation of Railways Act 1889 let the Board of Trade order railways to adopt the block system, interlock points and signals, and fit continuous brakes, which work along the whole train, to all passenger trains.
Credit
the UK Parliament; the Board of Trade
1893
Brakes and couplers by law
US freight crews once stepped between moving cars to couple them and rode the cars to work the brakes. Around 1887 George Westinghouse adapted his air brake to long freight trains and Eli Janney developed an automatic coupler, and the Safety Appliance Act of 1893 required both. The historian Mark Aldrich calls it "the first federal law intended primarily to improve work safety", and writes that by 1900 "risks to trainmen had fallen dramatically".
Credit
US Congress; George Westinghouse; Eli Janney
2013
Lac-Mégantic runaway
In the early hours of 6 July 2013 an unattended train of 72 tank cars holding 7.7 million liters of crude oil rolled about seven miles downhill into Lac-Mégantic, Quebec; 63 cars derailed, fire followed and 47 people died. The Transportation Safety Board of Canada (TSB) found the hand brakes too few to hold the train, between 17 and 26 were needed, and that the test of them had been done with the locomotive air brakes still on. It also found a weak safety culture at the railroad and gaps in Transport Canada's follow-up.
LessonA brake test run with another brake still applied proves nothing.
Credit
the TSB
2018
Lessons fade at Waterloo
On 15 August 2017 a passenger train leaving London Waterloo hit a stationary engineering train at 13 mph, with no injuries, after wiring added outside the design process left points in the wrong position. In its 2018 report RAIB wrote that "some of the lessons from the 1988 Clapham Junction accident are fading from the railway industry's collective memory" as people who remember it retire.
LessonKnowledge held only in people's heads leaves with them.
Credit
RAIB
Watch for
Low-cost crossing warnings
About 75% of Australia's 23,000 level crossings are passive, with signs but no barriers, lights or bells. Swinburne University of Technology's Smart-LX uses AI and solar power to warn road users of trains, and was reported on 31 August 2026 to have detected every train in a three-month trial on the Swan Hill line south of Kerang, Victoria.
Achieved whenVictoria's Department of Transport and Planning or the national rail safety regulator publishes the trial data or approves the system for routine use.
Claimed by
Swinburne University of Technology, as reported in August 2026
Open
01
Find the earlier warnings
A search tool that, given the first facts of a new accident, finds earlier investigation reports from any country that share its causes. Score it on RAIB's own reports, each of which records a review of earlier RAIB investigations relevant to the accident: give the tool only the facts known at the start, count how many of RAIB's cited reports it ranks in its top ten, and count the relevant reports from other countries it adds, as judged by two reviewers.
The briefCredit · unclaimed02
Heat buckles, tested on history
Rails expand in heat, and if the track cannot hold them they can bend sideways, a buckle railroaders call a sun kink, and derail a train. Build an open model that turns weather forecasts into a daily buckle-risk map, and test it against real buckled-track derailments in the FRA's public accident records and the weather on those days.
The briefCredit · unclaimed
Open 01 · the brief
Find the earlier warnings
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A search tool that, given the first facts of a new accident, finds earlier investigation reports from any country that share its causes. Score it on RAIB's own reports, each of which records a review of earlier RAIB investigations relevant to the accident: give the tool only the facts known at the start, count how many of RAIB's cited reports it ranks in its top ten, and count the relevant reports from other countries it adds, as judged by two reviewers.
Why it is still open
RAIB warned in 2018 that the lessons of Clapham were fading. HazardMap (Hong, Clifton and Nelson, 2024) groups the text of reports from Britain's RAIB, Australia's ATSB, the US NTSB and Canada's TSB by topic to map hazards, but this research found no tool scored on finding the earlier reports that investigators themselves cite.
Built with what exists
Public reports from RAIB, the NTSB, the TSB and the ATSB; an AI model that reads long documents, and a second model that flags weak matches for a person to check (see the Note "Seven more flagged"); open-source search libraries.
First step
Take 20 recent RAIB reports, list the earlier investigations each one cites, and test a plain keyword search as the baseline.
Done looks like
A public benchmark of at least 100 RAIB reports with the earlier reports each one cites, and a results table giving the share found in the top ten and the number of added foreign reports accepted by both reviewers.
Open 02 · the brief
Heat buckles, tested on history
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Rails expand in heat, and if the track cannot hold them they can bend sideways, a buckle railroaders call a sun kink, and derail a train. Build an open model that turns weather forecasts into a daily buckle-risk map, and test it against real buckled-track derailments in the FRA's public accident records and the weather on those days.
Why it is still open
A 2024 study by Wongkaew and colleagues trained its buckling model on computer simulations, citing "the inadequacy of field data collection on track buckling" and noting that data linking buckles to track conditions "is scarce". This research found no published model tested on a national record of real buckling derailments.
Built with what exists
The FRA's public train accident database, which records a cause for each accident; a public archive of historical weather; Python; a coding agent to build the data pipeline.
First step
Download ten years of FRA accident records, select those caused by buckled track, and pull the day's maximum temperature for each from a public weather archive.
Done looks like
A back-test over at least ten years, published with code, giving the share of buckling derailments that fell on days and places the model rated high-risk, and the share of all track-days it rated high-risk, as a measure of false alarms.