Fly ash and slag cement have a proven track record as partial replacements for clinker-based cements, which are carbon-intensive to produce. But to meet greenhouse-gas reduction goals set by the concrete and cement industries, it will be necessary to broaden the menu of options for lower-carbon reactive clinker replacements (RCRs).
Developing RCRs, however, is a time- and labor-intensive process. The first barrier has traditionally been discovering materials that could become viable RCRs. To help overcome this barrier, our team at the MIT Concrete Sustainability Hub (MIT CSHub) and Olivetti Group developed AI tools that helped us first identify and then vet more than 14,000 prospective RCRs.
AI Leads the Way
The large language models returned 14,000+ prospective materials, categorized into 19 types.
The first hurdle we confronted was the need to sift through millions of known compounds to pinpoint the ones that showed potential for replacing clinker. Reviewing scientific literature is a critical step in any research project; however, a truly comprehensive search for potential RCRs required tools capable of scanning the literature at a scale previously not possible. We developed fine-tuned large language models (LLMs) capable of quickly "reading" more than 10,000 academic papers to identify substances with the right chemical profiles and material classifications.
The LLMs returned 14,000+ prospective materials, categorized into 19 types:
- Lime species
- Cements
- Clinkers
- Slags
- Other minerals
- Construction and demolition wastes
- Other by-products and wastes
- Glasses
- Biomass ashes
- Silica fume
- Natural pozzolans
- Granite
- Clays
- Bottom ashes
- Fly ashes
- Calcined clays
- Metakaolin
- Other ashes, and
- Mine tailings
With this list of thousands of potential materials, we began to narrow down the results. Since a central obstacle in finding new RCRs is understanding how they will react when combined with water in a concrete mixture, we focused on examining each of the 19 material types for characteristics that would influence reactivity. Our goal was to identify candidates likely to have sufficient reactivity to be commercially viable.
Data our LLMs extracted from scientific literature provided information on the chemical compositions of the material types. Some, such as silica fume, showed narrow compositional ranges, indicating a somewhat consistent chemistry. Others, such as biomass ashes and construction and demolition wastes, had broad compositional ranges, reflecting their heterogeneous sources and processing conditions, and indicating that they might exhibit large variations in reactivity.
Construction and demolition wastes, incineration ash from biomass and various types of slags emerged as high-potential options.
Since reactivity is typically measured by laboratory tests that take weeks to perform, it was impractical for us to physically test all the materials our LLMs proposed. Therefore, we turned to AI-enabled processes again. We trained a machine learning model using experimental data from more than 300 materials to create a tool capable of making reactivity predictions for the various potential RCRs. Validation tests we performed confirmed that the model's predictions closely matched those of physical experiments.
While laboratory reactivity tests typically monitor heat release, calcium hydroxide consumption and the amount of water that is chemically or physically trapped within the hydration products, our machine learning model made its predictions on these metrics based on chemical composition, particle size, specific gravity and amorphous content. Output from the model indicated which potential RCRs were highly reactive and hydraulic, and which were less reactive and pozzolanic.
With reactivity data in hand, we moved on to mapping out priority RCR replacement candidates, examining materials within two broad groups: industrial by-products and natural resources (i.e., rock formations).
An important finding was that these natural RCRs are not randomly distributed but form distinct geological clusters, especially in tectonically active regions.
Of the industrial by-products, construction and demolition wastes, incineration ash from biomass and various types of slags emerged as high-potential options. Used as RCRs, the three groups together could replace up to 40 percent of global clinker production. Such a shift would reduce annual greenhouse gas emissions by roughly 1.2 billion tons, equivalent to removing 260 million cars from U.S. roads. These materials not only offer a lower-carbon alternative to clinker but create opportunities for transforming waste streams into valuable construction inputs.
Natural rock formations were also found to serve as viable RCRs. Using our AI tools, we screened more than one million rock samples from around the world using global geochemical data. Twenty volcanic and sedimentary rock types showed consistent promise, including ignimbrites, silicic tuffs, pumice, shales and rhyolites.
An important finding was that these natural RCRs are not randomly distributed but form distinct geological clusters, especially in tectonically active regions. These concentrations of RCRs could serve as locally available substitutes in areas where industrial by-products are scarce. Significant concentrations appear across northern and central Europe (Baltic Shield, Scafell Pike), Asia (Japan, Taurus and Zagros Mountains, Indian cratons, Tibetan Plateau), Africa (Great Rift Valley), Oceania (Great Dividing Range and ancient cratons) and South America (Andes). In the U.S., the Rocky Mountains and Cascade Volcanic Arc in the West, as well as the Appalachians along the East Coast, appear to have a particularly high availability of potentially reactive rocks.
What This Means for a Lower-Carbon Future
For years, the concrete and cement industries have worked to reduce carbon emissions. A large share of these emissions comes from producing clinker, the primary constituent of cement. The production process requires heating limestone to temperatures above 1,300° C, usually in fossil-fuel-fired kilns. The industry has therefore established a target to reduce the global clinker-to-cement ratio from the current 76 to 52 percent by 2050.
A vast landscape of low-carbon clinker replacements exists and is ready to be systematically uncovered and assessed.
Decades of successful use of slag and fly ash have shown that RCRs can effectively substitute for a portion of clinker. The materials we identified point to additional streams that could broaden the pool of available RCRs, strengthening the cement industry supply chain and reducing dependence on clinker. Our study also highlights naturally occurring, locally available candidates that could be especially valuable in regions where industrial by-products are limited, such as the U.S. West Coast.
Realizing this potential at a commercial scale will require further testing and validation, as well as activation strategies for less reactive materials. Areas for evaluation include fresh-state workability, hardened performance, long-term durability, and environmental impacts. Still, the implications of our research are clear. A vast landscape of low-carbon clinker replacements exists and is ready to be systematically uncovered and assessed.
Beyond our findings on clinker replacement, our study methodology offers a useful template for other material-intensive sectors. By combining artificial intelligence, large-scale data mining and predictive modeling, researchers can rapidly identify promising alternatives, evaluate their defining characteristics and assess their potential applications. Applied more broadly, this approach could help accelerate carbon-reducing discoveries across industries and support the scientific community in the wider effort to lower atmospheric carbon.
