Industries / Mining and materials / Cement and concrete
Testing strength
Testing strength is how makers and buyers check that cement and concrete will carry the loads they were designed for, mostly by crushing small samples made from the same batch, and more and more by measuring the concrete in place.
6
Milestones achieved, from 1918 to 2010.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1918
The water-cement ratio
Duff Abrams, at the Lewis Institute in Chicago, crushed 6 by 12 inch concrete cylinders at 28 days and found that their strength followed one smooth curve against the ratio of water to cement, whatever the mix or the size of the aggregate. He called water "the most important ingredient". He presented the work in December 1918 and published it as Bulletin 1 in May 1919, using the slump of a freshly molded cylinder to describe how wet each mix was.
Credit
Duff A. Abrams, Structural Materials Research Laboratory, run by the Lewis Institute with the Portland Cement Association
1940
Alkali-silica reaction
Thomas E. Stanton showed that cements high in alkali can react with certain minerals in some aggregates, such as shales, cherts and impure limestones found along the California coast, and make concrete swell and crack. He also described a test in which mortar specimens cured in sealed containers show the damage in a comparatively short time.
LessonConcrete can pass its strength test at 28 days and still crack years later, so aggregates need their own tests for this reaction.
Credit
Thomas E. Stanton
1949
Sound pulses through concrete
J. R. Leslie and W. J. Cheesman described an instrument they called the Sonicscope, which sends pulses of ultrasound through concrete and measures their speed. The speed does not depend on the size or shape of the piece, so it can be measured in mass concrete, slabs or lab samples, and it revealed internal cracks and deterioration. The stiffness (dynamic modulus) calculated from it agreed closely with established test methods.
Credit
J. R. Leslie and W. J. Cheesman (ACI Journal, September 1949)
1950
The rebound hammer
Ernst Schmidt, a Swiss engineer, published "Der Beton-Prüfhammer" (the concrete test hammer) in the Schweizerische Bauzeitung in 1950, describing a device to judge the quality of concrete already built into a structure. The hammer fires a spring-loaded mass at the surface, and a chart turns the rebound reading into an estimated strength. Proceq, which has made instruments in Switzerland since 1954, still sells Schmidt hammers under that name.
Credit
Ernst Schmidt; Proceq
1998
Strength from data
I-Cheng Yeh trained a neural network to predict concrete strength from the amounts of cement, blast-furnace slag, fly ash, water, superplasticizer and aggregates, and the age of the sample. His 1,030 results, given to the UCI Machine Learning Repository in 2007, became a standard practice set for machine learning: the repository shows more than 146,000 views, and Crossref counts more than 1,300 works citing the paper.
LessonThe set records no cement type or chemistry, and Meta's BOxCrete team wrote in 2026 that no new public single-lab strength set had appeared since, so models trained on it may not carry over to other materials.
Credit
I-Cheng Yeh
2010
Faked strength tests
In February 2010 a Manhattan jury convicted Testwell Laboratories, the largest concrete testing lab in the New York area, its owner V. Reddy Kancharla and vice-president Vincent Barone of enterprise corruption. The firm was believed to have filed false reports on 82 projects, including One World Trade Center, Yankee Stadium and the Second Avenue Subway. Prosecutors charged that it certified Freedom Tower concrete at 12,000 psi (about 83 MPa) when the Port Authority's own tests averaged about 10,000 psi (about 69 MPa), and issued 200 to 300 mix-design reports a year while running six to eight real tests.
LessonA strength report is only as good as the lab behind it; New York's Department of Buildings set up a concrete unit in 2009 that audits the field and lab work of every licensed testing lab, and it had carried out more than 130 audits by March 2010.
Credit
The Port Authority of New York and New Jersey and the New York Yankees, who raised the problems; the Manhattan District Attorney's office
Watch for
Sensors that skip calibration
Most embedded strength sensors use the maturity method (ASTM C1074), which turns the concrete's temperature history into strength through a relation worked out in advance for each mix. A pooled-fund study led by the Indiana Department of Transportation built REBEL, a sensor that reads strength from how a small piezoelectric element vibrates as the concrete around it stiffens. Caltrans's March 2026 summary says its results were comparable to cylinder breaks and maturity testing, on more than 800 sensors in over 100 projects, and that the final report is forthcoming. Giatec says its SmartRock Pro sensor calibrates itself as the mix changes.
Achieved whena lab or agency independent of the makers publishes a same-pour comparison of these sensors against cylinder or core strengths.
Claimed by
the TPF-5(471) research team, March 2026; Giatec, on a comparison page last updated July 2026
Open AI mix design
Meta released BOxCrete, an open-source model that forecasts strength from 1 to 28 days and proposes mixes, with a data set of more than 500 strength results from 123 mortar and concrete mixes made with Type IL cement. Its preprint reports R² of 0.94 and an error (RMSE) of 0.69 ksi, about 4.8 MPa, on its own data. Meta says a BOxCrete mix used in a slab at its Rosemount, Minnesota, data center reached full strength 43% faster than the original mix.
Achieved whenthe paper passes peer review and a producer outside the project publishes how the model's forecasts compared with its own cylinder breaks.
Claimed by
Meta with Amrize and the University of Illinois Urbana-Champaign, March 2026
Forecasts from 70,000 tests
A preprint by Md Asiful Islam, Md Ahmed Al Muzaddid, Afia Jahin Prema and Sreenath Reddy Vuske reports that a neural network trained on about 70,000 industry strength records, using water-cement ratio, cementitious content, slump, air content, temperature and placement conditions, predicts 28-day strength with a mean error of about 2.5%, which they say matches the normal spread of lab testing.
Achieved whenthe model is tested on another producer's records, or the records are released so others can check the result.
Claimed by
the authors, January 2026
Open
01
Sensor shootout on one pour
An independent, same-pour comparison of embedded strength sensors. Cast sensors from several makers side by side in the same test slabs, and break cylinders and cores from the same concrete at fixed ages, so every sensor's estimate is checked against the same breaks.
The briefCredit · unclaimed02
A second open strength set
A public lab data set like BOxCrete's that varies the materials as well as their amounts: at least three cements from different plants and several sources each of fly ash, slag and calcined clay, each with its chemistry by X-ray fluorescence, its particle sizes and its R3 reactivity, and strength at 1 to 28 days. Then measure how well BOxCrete predicts the new set as released and after retraining.
The briefCredit · unclaimed03
Open set of break photos
A public set of photos of broken test cylinders from several labs, each labeled with one of the six fracture types in ASTM C39, the US standard for the cylinder test, and a classifier trained on some labs and scored on a lab it has never seen.
The briefCredit · unclaimed
Open 01 · the brief
Sensor shootout on one pour
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An independent, same-pour comparison of embedded strength sensors. Cast sensors from several makers side by side in the same test slabs, and break cylinders and cores from the same concrete at fixed ages, so every sensor's estimate is checked against the same breaks.
Why it is still open
The comparison pages found for this brief were written by a sensor maker, Giatec, and describe how the products work rather than measuring their accuracy. The Caltrans summary of the REBEL study checks one system against breaks, and its final report had not appeared. A search for an independent test across brands found none.
Built with what exists
Commercial sensors such as Giatec SmartRock, Converge Signal, Saint-Gobain's Maturix, Hilti's Concrete Sensors, Transtec's COMMAND Center and WaveLogix's REBEL; a university materials lab with a compression machine for ASTM C39 cylinder tests and a core drill; Python for the tally.
First step
Ask a university lab that already casts test slabs to host one pour, and buy two sensors from each of three makers.
Done looks like
Sensors from at least five makers in at least 10 pours, with cylinder and core strengths at 1, 3, 7 and 28 days, and a published table of each sensor's error in MPa against the breaks.
Open 02 · the brief
A second open strength set
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public lab data set like BOxCrete's that varies the materials as well as their amounts: at least three cements from different plants and several sources each of fly ash, slag and calcined clay, each with its chemistry by X-ray fluorescence, its particle sizes and its R3 reactivity, and strength at 1 to 28 days. Then measure how well BOxCrete predicts the new set as released and after retraining.
Why it is still open
The BOxCrete authors write that no new public strength set from a single lab appeared between Yeh's 1998 set and theirs, that larger industry sets of about 10,000 and 40,000 results are not public, and that their model "may not work well on other materials" than the ones it was trained on.
Built with what exists
The BOxCrete code and data (github.com/facebookresearch/SustainableConcrete, MIT license); a university concrete lab with a compression machine; X-ray fluorescence and laser particle sizing; the R3 test (ASTM C1897); Python.
First step
Rerun BOxCrete on its released data to reproduce the published error, then write the mix plan and ask three cement plants for samples.
Done looks like
At least 60 mixes and 300 strength results published under an open license, with a table of BOxCrete's error in MPa on the new set before and after retraining.
Open 03 · the brief
Open set of break photos
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of photos of broken test cylinders from several labs, each labeled with one of the six fracture types in ASTM C39, the US standard for the cylinder test, and a classifier trained on some labs and scored on a lab it has never seen.
Why it is still open
The only classifier found, by Jose Manuel Palomino Ojeda and colleagues at the National University of Jaén in Peru (2023), was trained on 2,650 photos from one lab and reached 96% accuracy; the authors built their own set "given the lack of open access data", and the paper says its data are available only in the article. Searches of GitHub and Hugging Face for this brief found no labeled set of break photos.
Built with what exists
Phone cameras; Label Studio or CVAT for labeling; PyTorch or Keras with MobileNet, the network that did best in the Peru study.
First step
Write a photo protocol (two views, fixed distance, a scale in the frame) and ask three testing labs to photograph one week of breaks.
Done looks like
At least 3,000 photos from at least five labs in three countries, each labeled by two technicians with their rate of agreement reported, and the classifier's accuracy on a lab held out from training.