BRENNEN

Industries / Energy, water and waste / Environmental services / Asbestos and lead abatement

Air monitoring

Air monitoring is the work of pulling air through small filters during and after abatement, then counting the fibers or weighing the lead caught, to show that workers and neighbors stay below legal limits.

7

Milestones achieved, from 1971 to 2023.

3

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1971

    First US asbestos limit

    On 29 May 1971 the US Occupational Safety and Health Administration (OSHA) adopted a limit of 12 fibers per cubic centimeter of air (f/cc). An emergency standard on 7 December 1971 cut it to 5 f/cc, and in July 1976 it fell to 2 f/cc. In 1986 OSHA cut the limit tenfold to 0.2 f/cc and wrote a separate construction standard with rules for abatement and demolition jobs.

    LessonOSHA said the early limits were meant to prevent asbestosis, a scarring of the lungs, and gave only "some incidental degree of protection" against cancer, which needed far lower limits.

    osha.gov

    Credit

    OSHA, acting on a petition from the AFL-CIO's Industrial Union Department

  2. 1972

    Labs graded on shared filters

    From 1972 the Proficiency Analytical Testing (PAT) program, run by the American Industrial Hygiene Association (AIHA) with the National Institute for Occupational Safety and Health (NIOSH), sent the same asbestos filters to many labs and compared their fiber counts. A review of its first 20 years found that variability in counting chrysotile, the most common asbestos, had fallen by about 35% since 1975–1977 on the most lightly loaded filters. In a Canadian test, 58 analysts from 38 labs counted the same fields of view and still undercounted chrysotile by 25.0%. Britain's scheme, the Regular Interlaboratory Counting Exchanges (RICE), is run by the Health and Safety Executive (HSE), and UK labs must take part in it or a similar scheme.

    LessonCounting fibers by eye varies from person to person, so a lab's numbers are trusted only when it is tested on the same samples as other labs.

    ebi.ac.ukebi.ac.ukhse.gov.uk

    Credit

    AIHA and NIOSH (PAT); P. C. Schlecht and S. A. Shulman (review, 1995); T. W. Pang (Canadian test, 2000); HSE (RICE)

  3. 1993

    Lead limit for construction

    On 4 May 1993 OSHA cut the lead limit for construction workers from 200 to 50 micrograms per cubic meter of air (µg/m³) over an 8-hour day, with extra duties starting at 30 µg/m³. For tasks known to raise lead dust, such as abrasive blasting or torch cutting on lead-painted steel, the employer must treat workers as highly exposed until air samples show otherwise. Congress ordered the rule in Title X of the Housing and Community Development Act of 1992.

    LessonWhere a task is known to be dusty, the rule assumes high exposure until measurement proves it lower, which puts the burden of proof on the air sample.

    osha.govosha.gov

    Credit

    OSHA; US Congress

  4. 1994

    A limit at the edge

    On 10 August 1994 OSHA lowered the asbestos limit to 0.1 f/cc. It called this "at the limit of feasibility" for the hardest workplaces, set no lower action level because asbestos below the limit "cannot be reliably measured", and estimated that a working lifetime at 0.1 f/cc still carries 3.4 extra cancer deaths per 1,000 workers. Five days later NIOSH issued Method 7400, Issue 2: count fibers longer than 5 micrometers (µm) and at least three times as long as wide, under a phase contrast microscope (PCM). The method cannot tell asbestos from other fibers and misses fibers thinner than about 0.25 µm. In one NIOSH field study, counts of the same samples by different labs varied by 45% (relative standard deviation).

    LessonA legal limit can only be as low as the method that checks it.

    osha.govcdc.gov

    Credit

    OSHA; NIOSH

  5. 2001

    Ground Zero air claims

    On 18 September 2001, a week after the World Trade Center fell, the US Environmental Protection Agency (EPA) said the air in Lower Manhattan was "safe" to breathe. EPA's Inspector General found in August 2003 that EPA "did not have sufficient data and analyses" to say so. Sampling for several pollutants had not started, 24 of 69 asbestos filters collected by 17 September were too clogged with dust to read, and PCM counts only fibers longer than 5 µm while most asbestos fibers at the site were shorter. The White House Council on Environmental Quality had EPA add reassuring statements and delete cautionary ones. EPA and New York City later read almost 10,000 air samples by transmission electron microscope (TEM).

    LessonStatements about air safety should wait for data from a method that can see the fibers in question.

    epa.gov

    Credit

    EPA Office of Inspector General (report 2003-P-00012)

  6. 2012

    France counts by electron microscope

    A decree of 4 May 2012 required, from 1 July 2012, that workplace asbestos air samples in France be read by analytical TEM, which also counts very thin fibers that light microscopes miss. The limit, then 100 fibers per liter (0.1 f/cc), was to fall to 10 fibers per liter (0.01 f/cc) from 1 July 2015. The basis was a government campaign of 265 measurements on 29 kinds of removal job (November 2009 to October 2010), in which the national safety institute INRS found that PCM generally gave lower exposures than TEM. Accredited labs must enter every result in INRS's Scola database. By May 2021 it held 312,860 TEM results from July 2012 to December 2020, and INRS set aside 62% of them as unusable for its analysis.

    LessonA national database of samples is only as useful as the job details recorded with each one.

    inrs.frquestions.assemblee-nationale.frinrs.fr

    Credit

    French Ministry of Labor (Direction générale du travail); INRS; AFSSET, now ANSES

  7. 2023

    EU limit cut tenfold

    Directive (EU) 2023/2668 of 22 November 2023 cut the European Union's asbestos limit from 0.1 to 0.01 f/cc, to be in national law by 21 December 2025. From 21 December 2029 counts must be made by electron microscope, against either 0.002 f/cc for fibers 0.2 to 3 µm wide or 0.01 f/cc if thinner fibers are counted too. In May 2025, ten national institutes in the PEROSH research network compared PCM, scanning electron microscope (SEM) and TEM results and found no general factor to convert one method's results into another's: TEM gave the highest counts, then PCM, then SEM, and the gap changed with the material and the work method.

    LessonChanging the counting method breaks the link to past measurements, so exposure records must be rebuilt rather than converted.

    publications.europa.eupublications.tno.nl

    Credit

    European Parliament and Council; R. Franken, P. Tromp and colleagues, TNO and the PEROSH asbestos network

Watch for

  1. AI counting on real filters

    Researchers in Japan trained an AI image model (SegFormer) on PCM images of outdoor air and air near demolition sites, using 98 image sets counted by expert analysts (77 for training, 21 for testing). They report that the system found 78% of the fibers the analysts counted, and that 67% of its finds matched fibers the analysts had counted.

    Achieved whenan outside lab tests the system on new demolition-site filters against accredited human counts and publishes the agreement, or a regulator accepts it for official counts.

    api.crossref.org

    Claimed by

    Y. Iida, T. Yamamoto, K. Iwasaki and K. Yuki, Frontiers in Analytical Science, June 2025

  2. Continuous fluorescence monitor

    A team led by A. Kuroda rebuilt a commercial fine-dust air sampler so that it stains the asbestos on its membrane filter with fluorescent asbestos-binding proteins, photographs it under a fluorescence microscope, counts fibers with AI, and then advances the filter for the next reading. In tests with crushed asbestos materials, group results agreed with PCM and SEM counts to within 14.1%, with more scatter at low fiber levels.

    Achieved whenthe monitor runs beside standard samplers at real demolition or removal sites and an independent party publishes the comparison.

    ebi.ac.uk

    Claimed by

    A. Kuroda and colleagues, Sensors, May 2026

  3. Flow imaging for fibers

    NIOSH researchers tested flow imaging microscopy, in which particles suspended in liquid flow past a camera, as a faster way to count fibers and measure their lengths. With plastic bead standards, counting bias was under 10% at 10x magnification; with glass fibers, measured lengths were within about 17% of PCM on average. They note that PCM reads only about 0.2% of a filter.

    Achieved whenthe method is tested on asbestos samples from real work sites, by a lab other than NIOSH, against NIOSH 7400 counts.

    ebi.ac.uk

    Claimed by

    B. K. Ku and P. Kulkarni, NIOSH, Environmental Science: Processes & Impacts, January 2026

Open

  1. 01

    A shared fiber-count benchmark

    A public set of PCM images of real air filters from abatement jobs, where every field of view has been counted separately by at least three accredited analysts under the NIOSH 7400 or World Health Organization counting rules. Each image carries the per-analyst counts, fiber outlines, and the filter's air volume and area. Add a scoring script that grades any counting model against the spread between human counters, not against one person's answer.

    The brief
    Credit · unclaimed
  2. 02

    Can labs see the limit

    A free calculator and short report that works out, for each counting method (PCM under the WHO method or NIOSH 7400, SEM under ISO 14966, TEM under ISO 10312), how much air must be sampled, how many fields or images must be read, and how many analyst hours that takes, to show with 95% confidence that a result is below 0.1, 0.01 or 0.002 f/cc, for clean, average and dusty sites.

    The brief
    Credit · unclaimed
  3. 03

    Update the OSHA exposure record

    An open analysis, which anyone can rerun, of every asbestos and lead air sample in OSHA's Chemical Exposure Health Data from 1984 to 2026, with construction and remediation work split out, the share of samples above the limit in force at the time, and the number of samples taken each year.

    The brief
    Credit · unclaimed

Open 01 · the brief

A shared fiber-count benchmark

Credit · unclaimed

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

What to build

A public set of PCM images of real air filters from abatement jobs, where every field of view has been counted separately by at least three accredited analysts under the NIOSH 7400 or World Health Organization counting rules. Each image carries the per-analyst counts, fiber outlines, and the filter's air volume and area. Add a scoring script that grades any counting model against the spread between human counters, not against one person's answer.

Why it is still open

The public images found are not of this kind. A set of 1,331 images published under CC BY 4.0 in the Roboflow 100 benchmark, from Ahmad Rabiee, whose 2023 study with Italy's INAIL used lab-prepared filters, includes images marked as synthetic and labels four kinds of "marks" rather than countable fibers. The Japanese team above trained on 98 image sets from real samples but describes no public release, and a 2026 University of Strathclyde poster on automated counting says "very limited public data is available" and trains on simulated fibers instead. Pang's test shows why one annotator is not enough. Two hobby projects on GitHub show the gap: one trained a detector on the Roboflow set in November 2026 and reported leakage between its official train and test splits, and another, from October 2026, is a web page that asks reviewers whether to count uncertain marks in Rabiee's photos.

huggingface.cogithub.comgithub.comebi.ac.ukstrathprints.strath.ac.ukapi.crossref.org

Built with what exists

Images taken by an accredited lab of sealed, permanent slides it already keeps, such as retired RICE or PAT slides, so nobody handles new samples; a motorized microscope with a camera; CVAT or Label Studio for outlining fibers; baseline detectors such as YOLO or SegFormer, first trained on the Roboflow 100 asbestos set; a small open vision model fine-tuned on one graphics chip (see the Note "One GPU to fine-tune").

  1. First step

    Train a baseline on the Roboflow 100 asbestos set using the leakage-reduced split already published on GitHub, and write to two labs in the RICE or PAT schemes asking whether they would photograph retired permanent slides under a data-sharing agreement.

  2. Done looks like

    At least 2,000 fields from at least 40 archived filters, each counted by three or more analysts, published with the spread between them; a baseline model whose filter-level count falls inside the analysts' range on at least 80% of held-out filters.

Open 02 · the brief

Can labs see the limit

Credit · unclaimed

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

What to build

A free calculator and short report that works out, for each counting method (PCM under the WHO method or NIOSH 7400, SEM under ISO 14966, TEM under ISO 10312), how much air must be sampled, how many fields or images must be read, and how many analyst hours that takes, to show with 95% confidence that a result is below 0.1, 0.01 or 0.002 f/cc, for clean, average and dusty sites.

Why it is still open

Slides from the PEROSH asbestos network (December 2025) estimate that SEM checks at the new EU limits need more than 100 images (0.002 f/cc) or 300 images (0.01 f/cc with thin fibers) per sample, warn that sampling more air overloads filters on construction sites, and conclude that existing standard methods "are not validated" for the new limits. They give worked figures for a few set conditions; no public tool was found that lets a lab or contractor enter its own flow rate, sampling time and dust level.

perosh.eu

Built with what exists

Published method figures (NIOSH 7400's detection limit of 7 fibers per mm² of filter and its 100–1,300 fibers per mm² working range; HSE's rule that 5 fibers in 100 fields is the lowest reliable count); Poisson statistics in Python; a simple web page or Streamlit app.

  1. First step

    Code HSE's two worked examples from its analysts' guide, HSG248, and check that the calculator returns the same two numbers.

    hse.gov.uk

  2. Done looks like

    The calculator reproduces HSE's two worked examples (240 liters with 100 fields read gives about 0.01 f/cc; 480 liters with 200 fields gives 0.003 f/cc) and PEROSH's image counts to within 10%, and publishes a table of samples, fields and hours for each method at each limit.

Open 03 · the brief

Update the OSHA exposure record

Credit · unclaimed

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

What to build

An open analysis, which anyone can rerun, of every asbestos and lead air sample in OSHA's Chemical Exposure Health Data from 1984 to 2026, with construction and remediation work split out, the share of samples above the limit in force at the time, and the number of samples taken each year.

Why it is still open

The main published analysis of asbestos, by Cowan and colleagues in 2015, covers OSHA samples only through 2011. It found the data limited by gaps and entry errors, with 40% of records overlapping between two OSHA databases. For lead, a 2011 NIOSH analysis of OSHA's records ran from 1979 to 2008 and found construction samples higher than any other industry's. Neither has been brought up to date, and OSHA now posts yearly sample files through 2026.

ebi.ac.ukstacks.cdc.govosha.gov

Built with what exists

OSHA's yearly sample_data zip files; Python with pandas or DuckDB; industry codes (NAICS 23 for construction, 562910 for remediation services); a notebook published on GitHub.

  1. First step

    Download OSHA's 2024 sample file and count personal asbestos and lead samples by industry code.

  2. Done looks like

    The analysis reproduces Cowan's yearly asbestos sample counts for 1984–2011 to within 5%, then extends them to 2026 and adds lead, with a published table of sample counts, medians and the share above the limit for each year.