Industries / Industrial equipment and services / Industrial services / Industrial hygiene
Air sampling
Air sampling is measuring how much of a dust, gas or vapor a worker breathes, usually with a small pump and filter worn near the face, and comparing the result with an exposure limit, the most a worker may breathe on average over a working day.
7
Milestones achieved, from 1919 to 1993.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1919
A tube that changes color
Arthur B. Lamb, a lieutenant colonel in the US Army's Chemical Warfare Service, and Charles R. Hoover patented a gas detector: a glass tube packed with layers of chemicals, through which the suspect air is pushed. Carbon monoxide turns the reagent from white to green to brown, and the patent reports a response from 1.0% down to 0.01% of the air (100 parts per million). The patent names mines and industrial plants as places to use it.
Credit
Arthur B. Lamb and Charles R. Hoover, US patent 1,321,062, filed 28 December 1918 and granted 4 November 1919
1946
ACGIH's first exposure limits
The National Conference of Governmental Industrial Hygienists, founded in Washington in 1938, renamed itself the American Conference of Governmental Industrial Hygienists (ACGIH) in 1946. That year it adopted its first list of 148 exposure limits, then called maximum allowable concentrations. It renamed them threshold limit values (TLVs) in 1956, and says the list now covers more than 700 chemical substances and physical agents.
Credit
ACGIH's Threshold Limit Values committee, set up in 1941
1960
A sampler worn by workers
R. J. Sherwood and D. M. Greenhalgh published "A personal air sampler" in the Annals of Occupational Hygiene. In personal sampling, a pump worn by the worker draws air through a filter, so the result shows what that person breathed, not what one fixed spot in the room held. In 2014 NIOSH researchers, citing this paper first, wrote that personal sampling pumps had been in use for more than 50 years.
LessonA pump's airflow pulses with each stroke, which can change how much dust a size-selecting sampler collects; the same NIOSH team found that the 10% pulsation limit in the European pump standards had been set "without validated technical justification.".
Credit
R. J. Sherwood and D. M. Greenhalgh
1971
Limits copied into law
On 29 May 1971 the new US Occupational Safety and Health Administration (OSHA) issued its general industry standards by adopting existing federal rules and consensus standards. Most of its permissible exposure limits (PELs) came from rules for government contractors under the Walsh-Healey Act, which had taken them from ACGIH's 1968 TLVs. OSHA says most have not been updated since, and that many are "outdated and inadequate for ensuring protection of worker health."
LessonA limit written into law can stay fixed for decades after the science behind it has moved on.
Credit
OSHA
1973
Sampling without a pump
E. D. Palmes and A. F. Gunnison described a personal monitor for gases that needs no pump. In the version they adapted for nitrogen dioxide in 1976, the gas diffuses up an open tube to a coated collector at the sealed end, and the authors called the devices accurate, light and simple to use. A sampler that works by diffusion needs no pump or battery, so it can be worn all day.
Credit
E. D. Palmes and A. F. Gunnison
1992
Court voids the new limits
OSHA's Air Contaminants rule of 19 January 1989 lowered 212 exposure limits and set limits for 164 substances that had none. On 7 July 1992 the US Court of Appeals for the Eleventh Circuit vacated the rule in AFL-CIO v. OSHA. The government did not appeal, and in 1993 OSHA stopped enforcing the 376 tighter limits and restored the old table.
LessonOSHA now prints limits from California's OSHA, NIOSH and ACGIH beside its own, as alternatives "that may serve to better protect workers.".
Credit
OSHA, overturned by the Eleventh Circuit
1993
Three sizes of dust
The European standard EN 481 (1993), followed by ISO 7708 (1995), defined three fractions of airborne dust by how far into the body they travel: inhalable (breathed in through the nose and mouth), thoracic (passing beyond the larynx) and respirable (reaching the lung's gas-exchange region). Samplers are judged against these definitions; the UK Health and Safety Executive (HSE) recommends the IOM sampler for inhalable dust and a Higgins-Dewell type cyclone, run at 2.2 liters a minute, for respirable dust.
Credit
the European Committee for Standardization (CEN) and ISO
Watch for
Silica read in real time
Trolex says its AIR XS silica monitor, model TX8100, "is the first and only product in the world to detect and distinguish respirable crystalline silica (RCS) in real time." SafeWork NSW, the work safety regulator of New South Wales, Australia, had sought such a device from 2018. Two experts who reviewed the evidence for the NSW Chief Scientist & Engineer could not reach definitive conclusions, and the summary report says the monitor "should not be solely relied upon" until its efficacy is shown.
Achieved whenfield results for the TX8100 are published beside filter samples analyzed by X-ray diffraction or infrared spectroscopy, across several dust mixtures, and agree near the NSW limit of 0.05 mg/m³ averaged over 8 hours.
Claimed by
Trolex, quoted in the NSW review of December 2025
A 149-gram dust monitor
Sensidyne launched the Gilian Dustlight, a worn monitor that reads dust continuously as PM1, PM2.5, PM4 (which it calls respirable) and PM10 (which it calls inhalable), and says that at 149 g it is the most compact personal dust monitor available. It is sold to complement filter sampling, not to replace it. Monitors of this kind measure light scattered by particles, and NIOSH researchers have found that their readings shift with the type of dust and need correcting against a filter. The workplace inhalable fraction also runs to particles of 100 µm, far larger than PM10.
Achieved whenan independent laboratory or field study publishes the Dustlight's readings beside filter samples from a respirable cyclone and an inhalable sampler for several named dusts.
Claimed by
Sensidyne, December 2025
Open
01
Label dust peaks from video
A program that takes body-camera video and a second-by-second dust monitor log from the same shift, finds the short peaks that drive the shift average, and names the task on screen at each peak from a fixed list, such as bagging, sweeping or driving with the window down.
The briefCredit · unclaimed02
Test AI on sampling methods
A public question set that asks an AI model, for each of 300 chemicals, which sampling medium, flow rate, air volume and laboratory method to use, scored against the answers OSHA publishes, and run against several current models.
The briefCredit · unclaimed
Open 01 · the brief
Label dust peaks from video
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A program that takes body-camera video and a second-by-second dust monitor log from the same shift, finds the short peaks that drive the shift average, and names the task on screen at each peak from a fixed list, such as bagging, sweeping or driving with the window down.
Why it is still open
NIOSH's Helmet-CAM method pairs worn video with real-time dust readings in its EVADE software, and a NIOSH script already flags readings above 10 times the mean, but a person still watches the video to say what caused each peak. The NIOSH authors wrote in 2020 that this suits a one-day study at one mine but "could become unmanageable and impractical" for monthly sampling across a dozen sites. Searches of papers and GitHub found no tool that names the task automatically; a 2021 video method for risk assessment uses a human observer's checklist.
Built with what exists
A body camera; a real-time light-scattering dust monitor (the NIOSH studies used a Thermo Scientific pDR-1500; a Plantower PMS5003 sensor will do for bench trials); a model that reads video, such as Google's Gemini models or the open-weight Qwen2.5-VL; NIOSH's 10-times-the-mean peak rule written in Python; a fixed-answer classifier to pick each task from the list (Notes: "A thousand questions, one pass").
First step
Record two hours of a scripted, low-hazard dusty task, such as scooping and bagging chalk (calcium carbonate) powder in a ventilated garage while wearing a dust mask, with a camera and a PMS5003 logging every 2 seconds, and label the peaks by hand.
Done looks like
On 20 hours of recorded work holding at least 200 peaks, labeled by two people who have not seen the program's output, the program names the same task as both people for at least 80% of the peaks, with the code, video and labels published.
Open 02 · the brief
Test AI on sampling methods
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public question set that asks an AI model, for each of 300 chemicals, which sampling medium, flow rate, air volume and laboratory method to use, scored against the answers OSHA publishes, and run against several current models.
Why it is still open
OSHA tells its own inspectors to check its Occupational Chemical Database before sampling, because the correct media and flow rates can change after a method is written. The closest published test of AI models in this field put GPT-3.5 and GPT-4o through 385 questions from safety certification exams (Sammour and colleagues, 2024) and found them weakest in science and mathematics. A University of Minnesota poster at AIHA Connect 2026 evaluated how accurately AI models answer quantitative occupational hygiene questions; no published results were found. Searches of papers and GitHub found no test of sampling-method choices.
Built with what exists
OSHA's Occupational Chemical Database and its sampling and analytical methods; the NIOSH Manual of Analytical Methods for a second answer key; a scoring script in Python; current models through their public interfaces, including at least one open-weight model (Notes: "First on three benchmarks").
First step
Copy 50 entries from OSHA's Occupational Chemical Database by hand, write each as a question with its answer and link, and run them through two models.
Done looks like
300 questions published with a sourced answer for each, and at least five models scored field by field (medium, flow rate, air volume, method), with the share of fully correct answers and a list of answers that would have put a sample on the wrong medium.