This AI model finds low-emission cement in milliseconds

Researchers at the Paul Scherrer Institute (PSI) developed a machine learning model to optimize cement formulations, aiming to reduce carbon dioxide (CO₂) emissions while maintaining mechanical performance through a novel modeling approach.

The carbon footprint of cement

Cement production involves heating ground limestone to 1,400 degrees Celsius in rotary kilns to produce clinker, the primary raw material for cement. This process is energy-intensive, relying on combustion processes that release significant amounts of CO₂. The majority of these emissions, more than half, originate from CO₂ chemically bound within the limestone itself, which is released during its transformation in the high-temperature kilns, rather than solely from the combustion process.

Modifying the cement recipe by replacing a portion of the clinker with alternative cementitious materials represents a strategy for reducing these emissions. An interdisciplinary team in the Laboratory for Waste Management within PSI’s Center for Nuclear Engineering and Sciences focused on this approach. The research team, instead of relying exclusively on time-consuming experiments or complex simulations, developed a machine learning-based modeling approach.

A “digital cookbook” for climate-friendly cement

Romana Boiger, a mathematician and the study’s first author, stated:

“This allows us to simulate and optimise cement formulations so that they emit significantly less CO₂ while maintaining the same high level of mechanical performance.”

Boiger elaborated:

“Instead of testing thousands of variations in the lab, we can use our model to generate practical recipe suggestions within seconds – it’s like having a digital cookbook for climate-friendly cement.”

The novel approach enabled researchers to selectively filter cement formulations that met specific criteria. Nikolaos Prasianakis, head of the Transport Mechanisms Research Group at PSI, initiator, and co-author of the study, noted the extensive range of possibilities inherent in material composition, which ultimately dictate the final properties of cement.

Prasianakis stated:

“Our method allows us to significantly accelerate the development cycle by selecting promising candidates for further experimental investigation.”

The findings of this study were published in the journal Materials and Structures.

The search for alternatives

Current industrial practices incorporate by-products such as slag from iron production and fly ash from coal-fired power plants to partially substitute clinker in cement formulations, thereby reducing CO₂ emissions. However, the global demand for cement is substantial, making it impossible for these by-products alone to fulfill the entire need for clinker replacement. John Provis, head of the Cement Systems Research Group at PSI and a co-author of the study, emphasized the necessity for specific material combinations.

Provis stated:

“What we need is the right combination of materials that are available in large quantities and from which high-quality, reliable cement can be produced.”

Dr. Provis also described the global consumption of cement:

“To put it bluntly, humanity today consumes more cement than food – around one and a half kilograms per person per day.”

He further explained the implications of even minor improvements in emissions profiles:

“If we could improve the emissions profile by just a few percent, this would correspond to a carbon dioxide reduction equivalent to thousands or even tens of thousands of cars.”

The process of finding optimal combinations is complex. Provis explained that cement functions as a mineral binding agent, used with water and aggregates in concrete to create artificial minerals that bind the material. He characterized this as “doing geology in fast motion,” highlighting the intricate physical processes involved, which are computationally intensive to model, leading the research team to utilize artificial intelligence.

AlphaGenome reshapes how scientists interpret mutations

Artificial neural networks are computer models trained on existing data to accelerate complex calculations. During the training process, the network processes a known dataset, learning by adjusting the relative strength or “weighting” of its internal connections. This allows it to predict similar relationships quickly and reliably, serving as a rapid alternative to computationally intensive physical modeling. The PSI researchers employed such a neural network.

They generated the necessary training data using GEMS, an open-source thermodynamic modeling software developed at PSI. Nikolaos Prasianakis explained:

“With the help of the open-source thermodynamic modelling software GEMS, developed at PSI, we calculated – for various cement formulations – which minerals form during hardening and which geochemical processes take place.”

By integrating these findings with experimental data and mechanical models, the researchers established a reliable indicator for mechanical properties and, consequently, cement material quality. A corresponding CO₂ factor, a specific emission value, was applied to each component, enabling the determination of total CO₂ emissions.

Prasianakis described this as “a very complex and computationally intensive modelling exercise.” The generated data allowed the AI model to learn.

Boiger noted:

“Instead of seconds or minutes, the trained neural network can now calculate mechanical properties for an arbitrary cement recipe in milliseconds – that is, around a thousand times faster than with traditional modelling.”

The reverse approach

The research team employed a reverse approach to utilize AI in identifying optimal cement formulations, prioritizing minimal CO₂ emissions and high material quality. Instead of evaluating numerous formulations, the process was inverted to determine which cement compositions met predefined specifications for CO₂ balance and material quality. Both mechanical properties and CO₂ emissions are direct functions of the cement recipe.

Boiger explained:

“Viewed mathematically, both variables are functions of the composition – if this changes, the respective properties also change.”

The problem was formulated as a mathematical optimization task: identifying a composition that simultaneously maximizes mechanical properties and minimizes CO₂ emissions. Boiger elaborated:

“Basically, we are looking for a maximum and a minimum – from this we can directly deduce the desired formulation.”

To achieve this, the team integrated genetic algorithms, an AI technology inspired by natural selection, into their workflow. This integration enabled the selective identification of formulations that optimally combined the two target variables. This “reverse approach” eliminates the need for extensive trial-and-error testing of recipes, allowing for targeted searches that meet specific criteria, such as maximum mechanical properties with minimum CO₂ emissions.

The identified cement formulations include promising candidates. John Provis stated:

“Some of these formulations have real potential,” citing their viability in terms of CO₂ reduction, quality, and practical production feasibility.

The recipes will require laboratory testing to complete the development cycle. Nikolaos Prasianakis noted:

“We’re not going to build a tower with them right away without testing them first.”

The study serves as a proof of concept, demonstrating that promising formulations can be identified through mathematical calculation. Romana Boiger indicated that the AI modeling tool can be expanded to incorporate additional factors, such as raw material availability or the intended usage environment, for instance, marine or desert settings, which affect cement and concrete behavior.

Nikolaos Prasianakis emphasized the significant time savings offered by such a general workflow, positioning it as a promising approach for various material and system designs. The project’s success relied on an interdisciplinary team, including cement chemists, thermodynamics experts, and AI specialists, capable of integrating these diverse fields. Prasianakis highlighted the crucial exchange with other research institutions, such as EMPA, within the SCENE project. SCENE, the Swiss Centre of Excellence on Net Zero Emissions, is an interdisciplinary research program focused on developing scientifically sound solutions for reducing greenhouse gas emissions in industry and energy supply, under which this study was conducted.

Featured image credit