MIT used language models to read 88,000 papers in search of cement substitutes

May 17, 2025 — An MIT team used large language models to pull the chemistry of 14,000 materials out of 88,000 academic papers, then trained a neural network to predict which could stand in for part of the cement in concrete. Here's what happened, what it cost, and what went wrong.

Cost and outcome (as of June 2, 2025)

Outcome: A screening framework and a map of candidate materials, published open access. Its scenario: current supply of construction and demolition waste and municipal incinerator ash could substitute half of global cement production, cutting global greenhouse gas emissions by 3%. That's a modeled estimate. (Mahjoubi et al., Communications Materials, May 17, 2025)
Cost: No cost figures in the paper or MIT's release.
Next step: The team plans to validate some of the best candidates experimentally. (MIT News, June 2, 2025)

The problem

Cement production accounts for more than 6% of global greenhouse gas emissions, the paper says, driven by energy-intensive clinker production and limestone calcination. The usual substitutes, coal fly ash and blast-furnace slag, face supply limits: the paper says their availability has fallen 37% over two decades, from 25% to 17% of total cement production, partly because of less coal power and more steel recycling. (Mahjoubi et al., Communications Materials, May 17, 2025)

"The challenge that the team discovered wasn't a lack of candidates; the problem was that there were too many to sort through," MIT News wrote. Lead author Soroush Mahjoubi, a postdoc, said sorting "hundreds of thousands of pages of scientific literature" by hand "would have taken many lifetimes of work." Concrete is the world's most widely used building material, at an average of three tonnes per person per year, the paper notes. (MIT News, June 2, 2025; Mahjoubi et al., Communications Materials, May 17, 2025)

What the model does

Language models extracted chemical compositions and material types, and a multi-headed neural network predicts heat release, calcium hydroxide consumption and bound water from composition, particle size, specific gravity and amorphous or crystalline content. Together those give a unified view of reactivity and pozzolanicity. (Mahjoubi et al., Communications Materials, May 17, 2025)

MIT News says the team analyzed the literature and over 1 million rock samples and sorted candidates into 19 types, from biomass to mining byproducts to demolition materials, and found many could be used just by grinding. The paper reports that roughly 5-25% of 20 rock types, including pumice, shale and rhyolite, are reactive, with heat release above 200 J/g. (MIT News, June 2, 2025; Mahjoubi et al., Communications Materials, May 17, 2025)

Old tiles and pottery

"Some of the most interesting materials that could replace a portion of cement are ceramics," Mahjoubi said. "Old tiles, bricks, pottery — all these materials may have high reactivity." Senior author Elsa Olivetti said AI tools "have gotten this research far in a short time." The work ran through the MIT Concrete Sustainability Hub, supported by the Concrete Advancement Foundation, with funding from the MIT-IBM Watson AI Lab. Co-author Randolph Kirchain, the hub's director, said the aim is to help industry "build more sustainably, without compromising on strength, safety, or durability." (MIT News, June 2, 2025)

What went wrong

It's a model. The paper itself says "experimental validation remains essential"; computational screening only narrows the field. (Mahjoubi et al., Communications Materials, May 17, 2025)
The big number is conditional. The half-of-global-cement and 3% figures are the paper's scenario, "subject to performance constraints," not a measured outcome. (Mahjoubi et al., Communications Materials, May 17, 2025)
Supply is regional. The paper notes that regional variation in where substitute materials are generated limits clinker substitution. (Mahjoubi et al., Communications Materials, May 17, 2025)
What we don't know yet: which candidates pass lab tests, and what grinding and logistics would cost against cement.

Our follow-up question for The MIT Concrete Sustainability Hub

Which candidate materials have been tested in the lab since the paper, and how did their measured strength compare with the model's predictions?

Sources

Mahjoubi, S., Venugopal, V., Manav, I.B., et al., “Data-driven material screening of secondary and natural cementitious precursors,” Communications Materials 6, 99, published May 17, 2025. nature.com
MIT News (Andrew Paul Laurent, MIT Concrete Sustainability Hub), “AI stirs up the recipe for concrete in MIT study,” June 2, 2025. news.mit.edu

Disclosure: [not in copy] Editor placeholder