Machine learning-accelerated discovery of green cement recipes

Cement production, a significant contributor to global CO₂ emissions (about 8%), requires strategies to reduce environmental impact while maintaining material integrity. This study uses advanced modeling techniques, machine learning, and multi-objective constrained optimization to explore and identify cementitious blend compositions that are optimal with respect to material performance and CO₂ footprint. Given a cement recipe (clinker and supplementary cementitious materials), thermodynamic modeling is used to predict the corresponding hydration products that can then be utilized to calculate homogenized mechanical properties at the paste scale, e.g. elastic moduli. This supports a robust training database. An Artificial Neural Network is then trained to predict elastic moduli directly from the cement recipe, greatly accelerating the evaluation of the cement properties. The optimization targets are set to minimize CO₂ emissions associated with the recipe, and maximize the bulk modulus of the resulting hydrated cement paste. The results show a sequence of optimal cement recipes, lying on a Pareto front, showcasing bulk modulus values of hydrated cement paste from 14.61 to 23.63 GPa and CO₂ equivalents (in kg per kg cement paste) ranging between 0.39 to 0.76. The developed methodology highlights the potential to reduce CO₂ emissions while maintaining or improving the mechanical properties of cementitious materials, enhancing sustainability in construction practice.

Read more: https://doi.org/10.1617/s11527-025-02684-z