BRENNEN

Industries / Mining and materials / Chemicals and petrochemicals / Specialty chemicals / Adhesives and sealants

Testing bonds

Testing bonds is the work of showing that a glued joint is as strong as it should be: breaking sample joints in a laboratory, inspecting finished parts without damaging them, and controlling each step of bonding where no test can see inside the joint.

5

Milestones achieved, from c. 1917 to 2019.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. c. 1917

    Aircraft glue tested wet

    Casein glue is made from milk protein mixed with lime and caustic soda. The US Forest Products Laboratory (FPL) in Madison, Wisconsin, says it was little used in the United States until about 1916 or 1917, when military aircraft needed a glue that could stand water, and that one of its standard formulas came from a 1918 patent by S. Butterman, formerly of the FPL, dedicated to the public. The FPL's tests also set out the limits: casein-glued plywood soaked for 48 hours kept 40 to 60% of its dry strength, joints soaked and dried again and again finally failed, and molds destroyed them, so the FPL counted such joints durable only where the wood stayed below about 18 to 20% moisture.

    LessonA glue's strength when dry says little about how it lasts in damp, so wet and ageing tests are separate checks.

    ir.library.oregonstate.edu

    Credit

    US Forest Products Laboratory; S. Butterman

  2. c. 2000

    Certified bonding for trains

    Around 2000 Germany's Federal Railway Authority (Eisenbahn-Bundesamt) set off DIN 6701, a standard for making adhesive bonds on rail vehicles. It certifies the company doing the bonding, not the glue: its people, equipment, work areas and organization. Andreas Groß of the Fraunhofer Institute IFAM, who chaired the DIN 6701 working groups from 2001, and the consultant Hartwig Lohse wrote in 2015 that it was the first binding DIN standard on quality assurance for bonded joints outside woodworking, and that 456 companies worldwide were certified to it by 31 March 2015. DIN 2304 (2016) carried the same approach to every industry: each joint is put in a safety class from S1 (failure endangers life) to S4 (no safety demand), a trained supervisor answers for the bonding work, and the loads on each joint must be shown to stay below what it can carry. Both rest on the ISO 9001 idea of a "special process", one whose result cannot be checked 100% without destroying it, so errors must be prevented rather than found.

    weiterbildung.ifam.fraunhofer.deifam.fraunhofer.de

    Credit

    Eisenbahn-Bundesamt; DIN working groups led by Andreas Groß (Fraunhofer IFAM)

  3. 2005

    A rudder comes apart

    On 6 March 2005 Air Transat Flight 961, an Airbus A310 carrying 262 passengers and 9 crew from Varadero, Cuba, to Quebec City, lost most of its rudder in flight; the crew regained control and returned to Varadero. The rudder was a sandwich panel: a honeycomb core of Nomex, an aramid paper, between carbon-fibre skins bonded to it. A one-time check of 408 Airbus aircraft by eye and by tap test (listening for a dull note when the surface is tapped lightly) found more disbonds, places where a skin had come away from its core. Transport Canada wrote that composites had been treated as having "no damage growth" and that the tap test might miss small disbonds; the Transportation Safety Board of Canada (TSB) pointed to "the recent discovery that disbonds could grow undetected". Airworthiness directives from the European Union Aviation Safety Agency (EASA) in October 2007 and January 2008 replaced the tap test with ultrasonic, thermography and X-ray inspections, repeated at intervals set from damage growth rates measured during the investigation.

    LessonA hidden disbond can grow, and listening to taps does not reliably find a small one.

    bst.gc.ca

    Credit

    Transportation Safety Board of Canada; Transport Canada; EASA; Airbus

  4. 2009

    Bonds inspection cannot prove

    The US Federal Aviation Administration's advisory circular AC 20-107B, dated 8 September 2009, gives three ways to show that a bonded joint whose failure would be catastrophic can carry limit load, the largest load expected in service: show by analysis or test that the structure holds with the largest disbond its design features allow; proof-load each joint on every production part; or use non-destructive inspection that reliably shows each joint's strength. It says the third way awaits the future, because inspection able to find "weak bonds", joints that look whole but lack chemical adhesion, "has not been reliably demonstrated at a production scale". It calls a "process control mentality" the most reliable way to ensure bond quality, says peel tests have proven more reliable than shear tests for spotting poor adhesion, and treats adhesion failure, a clean break between glue and surface, as unacceptable in all qualification tests.

    LessonBecause a finished bond's strength cannot be read from outside, aircraft rules rely on designs that tolerate disbonds and on control of every bonding step.

    faa.gov

    Credit

    US Federal Aviation Administration

  5. 2019

    A laser finds weak bonds

    Laser bond inspection fires a laser pulse at a bonded part to send a stress wave through the glue line, tuned to break joints weaker than intended and leave sound ones intact; ordinary ultrasound then shows which broke. Richard Bossi of Boeing proposed it, and Boeing took it to LSP Technologies of Dublin, Ohio, in 2001 to build the laser. CompositesWorld says Boeing acquired a copy of the first prototype in 2011; Laser Focus World reported the first sale and delivery to Boeing in May 2013. In a validation funded and led by the US Air Force Research Laboratory (AFRL) with Boeing, reported in 2019, normal carbon-fibre joints averaged 30.34 MPa against 11.03 MPa for joints made weak on purpose, and the method found all the weak-bonded stringers and 6 of 6 hidden weak areas tested so far on a demonstration panel. The authors wrote that its role in certifying safety-critical bonds "is not yet defined", and that it cannot check the bond between a honeycomb core and its skins.

    compositesworld.comcompositesworld.comlaserfocusworld.com

    Credit

    Richard Bossi and Marc Piehl (Boeing); LSP Technologies; Kara Storage and Andrea Helbach (AFRL)

Watch for

  1. A network reads broken bonds

    After a test joint is broken, a person judges by eye whether the break ran through the glue (cohesive failure), peeled cleanly off a surface (adhesive failure) or both. A team led by Cournault trained a convolutional neural network, a kind of image-recognition model, on 434 broken dental bonds whose failure type was set by surface-height scans checked under an electron microscope. Working from ordinary microscope photos, it scored a weighted F1 (a combined measure of correct and missed calls, where 1.0 is perfect) of 0.893, against 0.824 to 0.911 for three human experts; most of its errors were on pieces near the line between mixed and mostly adhesive failure.

    Achieved whenanother laboratory runs or retrains the method on its own broken specimens, dental or industrial, and publishes matching scores.

    doi.org

    Claimed by

    Cournault, Vedrenne, Roman, Moyaux, Jmal and Etienne in the Journal of Dental Research, May 2026

Open

  1. 01

    Read the break by camera

    A shared benchmark for reading broken test joints: scans of broken lap-shear pieces (two strips glued at an overlap and pulled apart) from at least two laboratories, each marked by several people with the areas of cohesive and adhesive failure, and a public table scoring every method that estimates the cohesive share from an image against those people.

    The brief
    Credit · unclaimed
  2. 02

    A log of bond failures

    An open, searchable record of glued and sealed joints that failed in service, drawn from public recall notices and accident reports, each coded by what went wrong (surface preparation, cure, wrong adhesive, design, ageing) and linked to its documents.

    The brief
    Credit · unclaimed

Open 01 · the brief

Read the break by camera

Credit · unclaimed

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

What to build

A shared benchmark for reading broken test joints: scans of broken lap-shear pieces (two strips glued at an overlap and pulled apart) from at least two laboratories, each marked by several people with the areas of cohesive and adhesive failure, and a public table scoring every method that estimates the cohesive share from an image against those people.

Why it is still open

Methods exist, but each has been tried only on its makers' own pieces. Gaëtan Le Goïc and colleagues (Engineering Failure Analysis, 2022) mapped failure modes on fracture surfaces with reflectance transformation imaging, photographs taken under light from many directions, and the dental network under Watch for learned from one laboratory's teeth. Scans of broken single-lap joints from a 2016 US Army Research Laboratory study by Robert Jensen, Daniel DeSchepper and David Flanagan are public and free of copyright (CC0), with people's visual failure-mode surveys, but with survey answers, not marked areas. Open lap-shear data from the University of West Bohemia (November 2025) include photos with no failure labels. The reading matters: the FAA treats adhesion failure as unacceptable in every bond qualification test, so the share of clean break decides whether a bonding process passes.

doi.orgacdc.alcf.anl.govzenodo.orgfaa.gov

Built with what exists

The Army Research Laboratory scans from the Materials Data Facility; new pieces made from aluminium and plastic strips cut to the common 25 mm-wide single-lap shape, two or three shop-bought two-part epoxies and a polyurethane, and a universal test machine at a university or makerspace; a flatbed scanner at 1,200 dpi, or a phone on a copy stand with a ring light; Label Studio or CVAT for marking areas; Meta's Segment Anything (SAM 2) to speed up outlines; a small open-weight vision model fine-tuned on one GPU (see the Note "One GPU to fine-tune").

  1. First step

    Download the Army archive, read its survey form, and have three people mark the areas on 30 scans to measure how far they disagree.

  2. Done looks like

    At least 300 Army scans and 200 new pieces, each marked by three people and published under CC BY; at least three methods scored on pieces from a laboratory they were not trained on; a published count of pieces where each method's cohesive share falls within 10 percentage points of the people's average, with the spread among the people reported beside it.

Open 02 · the brief

A log of bond failures

Credit · unclaimed

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

What to build

An open, searchable record of glued and sealed joints that failed in service, drawn from public recall notices and accident reports, each coded by what went wrong (surface preparation, cure, wrong adhesive, design, ageing) and linked to its documents.

Why it is still open

Aviation has been surveyed once, as a paper. Guilherme Momm of Brazil's National Civil Aviation Agency (ANAC), with Larry Ilcewicz and Cynthia Ashforth of the FAA and David Fleming of the Florida Institute of Technology, reviewed 72 bond-related events in certified civil aircraft from 13 countries and sorted their root causes (SAMPE, 2021); they wrote that before them "no survey of those events associated with bonding issues was found." For cars, buildings and consumer goods no such survey was found. The records exist but sit apart: a text search of the US National Highway Traffic Safety Administration's (NHTSA) recall file, run for this page on 18 November 2026, found 209 recall campaigns since 2010 whose defect or consequence text names an adhesive or glue.

digitallibrarynasampe.orgstatic.nhtsa.gov

Built with what exists

NHTSA's recall flat file; the Federal Register API for FAA airworthiness directives; TSB, US National Transportation Safety Board and US Chemical Safety Board reports; Python and pandas; an AI model that answers fixed-choice questions about each document, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"), with every coded row checked by a person (the Note "Seven more flagged" covers an AI reviewer of long documents).

  1. First step

    Pull the 209 NHTSA campaigns, read 50 by hand, and draft the list of cause codes, starting from the root-cause groups in the aviation survey.

  2. Done looks like

    At least 300 failures from at least three sectors outside aviation, plus the 72 aviation events coded the same way, each with cause codes and a link; two people code the same 100 independently and agree on at least 85% of codes; a published count of failures by cause and sector.