MIT researchers have developed a machine learning framework that can identify sustainable natural and industrial materials to partially replace cement, offering a scalable path to significantly reduce global carbon emissions from construction.
Study: Towards net zero by data-driven discovery of sustainable cement alternatives. Image Credit: Sirsendu Gayen/Shutterstock.com
A recent study published in Communications Chemistry details how the team screened over 14,000 materials drawn from scientific literature and over one million rock specimens to pinpoint viable, lower-carbon alternatives to traditional cement ingredients.
Background
Cement production alone accounts for more than 6 % of global greenhouse gas emissions, primarily due to the energy-intensive processes involved in limestone calcination and clinker production. One of the most effective ways to curb these emissions is to replace clinker—the key binding phase in cement—with alternative materials.
Historically, industrial byproducts like blast furnace slag and coal fly ash have served as supplementary cementitious materials (SCMs). But access to these conventional substitutes has declined by 37 % in the past two decades, driven by rising steel recycling and the closure of coal-fired power plants.
This shift has created an urgent need to identify new materials that can offer similar reactivity in cement systems. With its ability to sift through vast and complex datasets, machine learning offers a promising path forward.
Methods
Led by researchers Elsa Olivetti and Soroush Mahjoubi at the Massachusetts Institute of Technology (MIT), the team developed a comprehensive framework that merges natural language processing (NLP) with predictive modeling.
Using NLP, the team mined more than 5.7 million scientific papers to extract the chemical compositions of over 14,000 candidate materials referenced across 88,000 publications. These materials were then categorized into 19 groups using fine-tuned large language models.
To evaluate their potential, a multi-headed neural network was trained to predict three critical indicators of cementitious reactivity: heat release, bound water content, and calcium hydroxide consumption. Inputs included chemical composition, specific gravity, particle size, and crystallinity. The model achieved a coefficient of determination (R2) greater than 0.85, allowing the researchers to assess and map reactivity across a broad material spectrum.
Results and Discussion
The findings revealed a wide range of previously underutilized materials with promising reactivity. For example, construction and demolition waste—such as crushed concrete and ceramics—demonstrated heat releases up to 450 J/g, comparable to conventional pozzolanic materials.
Ashes from municipal solid waste incineration and agricultural byproducts like rice husk, wood, and sugarcane bagasse also showed significant pozzolanic behavior. Additionally, mine tailings, especially from zinc and copper operations, emerged as promising secondary SCMs. Collectively, these industrial byproducts could replace up to 68 % of global cement production.
Yet not every region has access to industrial waste streams, making natural alternatives essential. To explore this, the researchers applied their predictive model to a global geochemical database of more than one million rock specimens. They identified 25 rock types that could be mechanically activated to exhibit cement-like reactivity. Silicic and ignimbrite rocks had the highest reactive-to-total sample ratios, while more abundant rocks like andesite and rhyolite offered better global accessibility despite lower reactivity.
These reactive rocks are typically found in tectonically active zones like the Pacific Ring of Fire, the Andes, and the Great Rift Valley, making them viable regional substitutes in areas lacking industrial byproducts.
One of the major challenges the team tackled was incomplete and inconsistent data, particularly for key attributes like amorphous content. To manage this, they developed a specialized neural network architecture capable of intelligently filling in missing data and making accurate predictions based on available physical and chemical properties. This innovation allowed them to navigate the messy, often fragmented data landscape typical of materials science research.
Conclusion and Future Directions
This study demonstrates the power of machine learning to uncover viable, lower-carbon alternatives to cement. If adopted at scale, the identified materials could reduce global CO2 emissions by up to 3 %—roughly equivalent to removing 260 million cars from the road.
Many of these materials require only mechanical grinding, bypassing the high energy costs associated with thermal processing. And because the identified natural materials are geographically diverse, they offer a path to more equitable access to sustainable construction solutions worldwide.
Going forward, the most promising candidates will need to be validated through experimental testing. Incorporating cement hydration kinetics into the model is also a key next step to further improve prediction accuracy and real-world applicability.
Journal Reference
Evans, J. D. (2025). Towards net zero by data-driven discovery of sustainable cement alternatives. Communications Chemistry, 8(1). DOI: 10.1038/s42004-025-01608-w. https://www.nature.com/articles/s42004-025-01608-w
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.