BRENNEN

Industries / Mining and materials / Chemicals and petrochemicals / Hydrogen production

Safety and leaks

Keeping hydrogen from leaking, catching fire or exploding, and measuring how much escapes into the air, where it adds to warming.

5

Milestones achieved, from 1937 to 2023.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1937

    The Hindenburg burns

    On 6 May 1937 the German airship Hindenburg, lifted by about 7 million cubic feet of hydrogen, caught fire while landing at the Naval Air Station in Lakehurst, New Jersey. Of 97 people aboard, 13 passengers and 22 crew died, and one man of the ground crew. Airships.net, a history site, reports that the German and American inquiries of 1937 both concluded that an electrostatic spark ignited leaking hydrogen. Where the leak came from was never explained.

    LessonA leak too small to find, plus a spark, was enough to destroy the largest aircraft of its day.

    airships.net

    Credit

    the 1937 German and US inquiries

  2. 2012

    Incidents pooled in one database

    The Hydrogen Incidents and Accidents Database (HIAD) gathers reports of unwanted events involving hydrogen so that designers and regulators can learn from them; its design was described in the International Journal of Hydrogen Energy in 2012. It is now run by the European Commission's Joint Research Centre (JRC) with the Clean Hydrogen Partnership, which deployed version 2.5 in October 2026. A 2026 study of version 2.4 found 1,235 records with a known starting point: 66.8% began in hydrogen equipment and 25.2% in equipment that did not hold hydrogen.

    knowledge.clean-hydrogen.europa.eudoi.orgdoi.org

    Credit

    Cristina Galassi and colleagues (2012); the JRC

  3. 2019

    Oxygen in a hydrogen tank

    On 23 May 2019 three hydrogen storage tanks exploded at an experimental plant in Gangneung, South Korea, that made hydrogen with a 200 kW alkaline electrolyzer (a machine that splits water with electricity) run on solar power. Two people died and six were hurt. HIAD's summary of the investigation says oxygen in the hydrogen had passed 6%, the explosion threshold, and a static spark set it off. Readings above 3% earlier had not stopped the plant. The electrolyzer often ran below its minimum power as the sunlight varied, and an oxygen remover was left out of the final design. A court later ordered five organizations to pay more than 10 billion won in compensation.

    LessonAn electrolyzer run at low and varying power can let oxygen into the hydrogen, so gas purity needs monitors that stop the plant automatically.

    knowledge.clean-hydrogen.europa.eu

    Credit

    Korean investigators; summary by HIAD

  4. 2019

    A refueling station leak

    On 10 June 2019 hydrogen leaked from the high-pressure storage of the Kjørbo refueling station outside Oslo, Norway, and ignited. The maker, Nel, traced the leak to a plug on one tank whose bolts had not been torqued to specification, and recommended that operators of stations of the same design in Europe and the US put them on standby. HIAD adds that the fault passed commissioning pressure tests and outside certification and held for over two years at about 950 bar, and that Nel was fined NOK 25 million. Nel says no unit exploded and the hydrogen burned in open air; HIAD records an explosion, with windows damaged 65 m away. No one died; police reported three people with minor injuries.

    LessonA gasket can hide a loose bolt from every leak test, and HIAD notes the station's monitoring was not set to flag an abnormal loss of pressure.

    knowledge.clean-hydrogen.europa.eunelhydrogen.com

    Credit

    Nel; the safety consultancy Gexcon, which investigated; HIAD

  5. 2023

    Leaked hydrogen warms the climate

    Hydrogen is not itself a greenhouse gas, but in the air it slows the breakdown of methane and adds ozone and stratospheric water vapor. Using five atmospheric chemistry models, Maria Sand and ten colleagues estimated hydrogen's 100-year global warming potential at 11.6 ± 2.8: a kilogram leaked warms about 12 times as much as a kilogram of carbon dioxide over a century. A year earlier Ilissa Ocko and Steven Hamburg showed the effect is stronger in the first two decades: if 10% of it leaks, green hydrogen might only halve the warming of the fossil fuels it replaces over 20 years.

    LessonLeaks matter for the climate, not only for fire.

    doi.orgdoi.org

    Credit

    Sand and colleagues (Communications Earth and Environment, 2023); Ocko and Hamburg (Atmospheric Chemistry and Physics, 2022)

Watch for

  1. Measuring real hydrogen leaks

    A study led by the Environmental Defense Fund (EDF), with Utrecht University, the University of Rhode Island, West Virginia University and Cornell University, began measuring hydrogen emission rates in March 2025. It uses the first commercial fast, high-precision hydrogen analyzers, from Aerodyne Research, at sites run by Air Products, Air Liquide, Shell and TotalEnergies: steam methane reformers, pipelines, liquefaction plants, refineries, refueling stations and vehicles. EDF says it fills "a gap where little empirical data exist today," and that the data will be published in peer-reviewed journals and made public. The University of Rhode Island reported in June 2025 that a first field stage at about ten hydrogen production and transfer facilities in Louisiana and Texas was complete, and that field measurements would continue through early 2026; no results had been found published by November 2026.

    Achieved whenpeer-reviewed papers report leak rates per facility and the measurements are released for others to check.

    edf.orguri.eduuu.nl

    Claimed by

    EDF and partners, 2025; restated by the University of Rhode Island, June 2025

Open

  1. 01

    Recheck incident cause codes

    A test of whether HIAD's cause labels can be reproduced. Draw 300 events at random from HIAD. Two trained people and an AI model each read the event narratives and assign the initiating cause and root cause classes, without seeing the curators' labels. Publish how often each coder agrees with the curators and with each other.

    The brief
    Credit · unclaimed
  2. 02

    Open recordings of leak sounds

    An open set of ultrasonic recordings of small gas leaks from the kinds of joints used in hydrogen systems (tube fittings, threaded plugs, flanged joints), with the leak flow measured for each clip. Use compressed air and helium, a non-flammable stand-in for hydrogen, at low pressure on the bench. Add a baseline model that detects a leak and estimates its size.

    The brief
    Credit · unclaimed

Open 01 · the brief

Recheck incident cause codes

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A test of whether HIAD's cause labels can be reproduced. Draw 300 events at random from HIAD. Two trained people and an AI model each read the event narratives and assign the initiating cause and root cause classes, without seeing the curators' labels. Publish how often each coder agrees with the curators and with each other.

Why it is still open

A 2026 analysis of HIAD 2.4 notes that its cause fields "were assigned by the same curators from the same narratives," so its findings show patterns inside the database, not an independent check. Language-model studies of HIAD 2.1 learn from those same curator labels; a 2025 conference paper that predicts causes and consequences from HIAD narratives reports no recoding by other people.

doi.orgrpsonline.com.sg

Built with what exists

HIAD 2.5 records; a coding guide written from HIAD's own category definitions; an AI model that scores fixed-answer questions in bulk (see the Note "A thousand questions, one pass"); Python with scikit-learn to compute Cohen's kappa, a measure of agreement beyond chance that runs from 0 (chance) to 1 (perfect).

  1. First step

    Export 30 HIAD events and pilot-code their initiating cause with two people to test the coding guide.

  2. Done looks like

    300 events recoded blind by three coders; kappa reported for each field and each pair of coders, including the curators; every field with kappa below 0.6 flagged; all codes and the guide published.

Open 02 · the brief

Open recordings of leak sounds

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open set of ultrasonic recordings of small gas leaks from the kinds of joints used in hydrogen systems (tube fittings, threaded plugs, flanged joints), with the leak flow measured for each clip. Use compressed air and helium, a non-flammable stand-in for hydrogen, at low pressure on the bench. Add a baseline model that detects a leak and estimates its size.

Why it is still open

Open leak recordings exist, but none has what this needs. Fraunhofer IDMT's compressed-air set (5,592 clips, 2020) is in the audible range, labels leak types without measured flow and may not be used commercially. GPLA-12 (2021) holds 684 signals from artificial leaks on a gas pipe, sorted into 12 classes. A 2024 set from the University of Southern California with Oak Ridge and Pacific Northwest national laboratories records gas mixtures inside a sealed canister, with no leaks at all. Searches found acoustic hydrogen-leak studies that report detection accuracy but release no recordings.

idmt.fraunhofer.dearxiv.orgzenodo.org

Built with what exists

An ultrasonic microphone of the kind sold as a bat detector; a thermal mass flow meter and a needle valve to set and measure each leak; off-the-shelf tube fittings; Python with librosa for audio features and PyTorch for the model. A small open audio model can be fine-tuned for this on one graphics card (see the Note "One GPU to fine-tune").

  1. First step

    Build a stand with one tube fitting, a needle valve and a flow meter on compressed air, and record 100 clips at three flow rates.

  2. Done looks like

    At least 2,000 labeled clips at five or more measured flow rates on at least three joint types, with background noise recorded separately, published under an open license; the baseline model's detection rate and false alarm rate reported on joints it never saw in training; a second group adds recordings from its own microphone, and the model's scores on them are published.