Data-Driven design of sustainable LC³ for 3D printing with Omani clays

Developing low-carbon binders from abundant, locally available clays are essential to reducing the embodied CO₂ of 3D-printed concrete. In this study, we developed a predictive–optimization model for Limestone-Calcined Clay Cement (LC³) mix design by leveraging complex, multi-phase natural clays without energy-intensive purification of high-purity minerals. Kaolinite, illite, and montmorillonite—identified as the dominant clay types in Oman's major deposits—served as the basis for LC³ formulation. A systematic experimental dataset including 287 LC³ formulations was generated through factorial variation of kaolinite, illite, montmorillonite, and limestone contents. Each formulation was experimentally tested to measure compressive strength, flexural strength, and yield stress—key properties for evaluating 3D printability and mechanical performance. Several machine learning models—including hybrid SVR–ANN, CatBoost–ANN, and a customized ANN model—were developed to predict these target properties. A multi-objective optimization was then performed under realistic engineering constraints, such as raw material availability, required strength levels, and printability thresholds. The predictive performance of the best model was validated using two real-world natural clays: Clay A (kaolinite-rich) and Clay B (montmorillonite-rich). The composites were also assessed through a comprehensive suite of experimental evaluations, including high-temperature resistance, acid–chloride durability, rheological characterization (flow curve and stress growth tests), and microstructural analysis via XCT, SEM, and TGA. Additionally, a finite element simulation of a 3D-printed wall element was conducted to assess the structural feasibility of the optimized LC³ at large scale. The customized dual-path ANN model (ReLU stream for mechanical prediction, ELU stream for rheology) demonstrated the highest accuracy, achieving R² values of 0.961 (compressive strength), 0.956 (flexural strength), and 0.793 (yield stress), with an RMSE of 33.95 Pa. The optimized formulation—comprising 60 % kaolinite, 5 % illite, 5 % montmorillonite, and 30 % limestone—met all target constraints. Prediction errors were limited to 2.6 % for Clay A and 4.2 % for Clay B. The optimized LC 3 composite exhibited 34.3 MPa compressive strength, 4.1 MPa flexural strength, and 726 Pa yield stress. Acid–chloride exposure testing showed strength retention above 87 % after 120 days. Microstructural analyses revealed pore refinement, densification, and high thermal stability. FEM simulation validated the large-scale structural performance of the optimized composite. Overall, this work demonstrates the potential of data-driven mix design to produce low-carbon and eco-efficient, digitally printable LC³ materials using regionally available clays.