A mineralogy-driven machine learning framework for optimizing LC 3 -based 3D-printed cementitious composites with an integrated, end-to-end python-based design tool
Abstract This study presents a framework for developing low-carbon, printable limestone–calcined clay cement (LC 3 ) using locally available clays in Oman that primarily contain kaolinite, illite, and montmorillonite. A dataset of 287 formulations using varying the proportions of these components and limestone, with each mix experimentally tested for compressive strength, flexural strength, and static yield stress to link mechanical performance with printability. Two supervised machine learning models: Gaussian process regression (GPR) and a shape-constrained generalized additive model, were t...
This study presents a framework for developing low-carbon, printable limestone–calcined clay cement (LC 3 ) using locally available clays in Oman that primarily contain kaolinite, illite, and montmorillonite. A dataset of 287 formulations using varying the proportions of these components and limestone, with each mix experimentally tested for compressive strength, flexural strength, and static yield stress to link mechanical performance with printability. Two supervised machine learning models: Gaussian process regression (GPR) and a shape-constrained generalized additive model, were trained on the dataset. GPR showed superior performance, achieving R 2 values of 0.966 (compressive strength), 0.958 (flexural strength), and 0.736 (yield stress). A multi-objective optimization approach identified an optimal LC 3 composition of 60% kaolinite, 5% illite, 5% montmorillonite, and 30% limestone. This mix yielded predicted properties of approximately 35 MPa compressive strength, 4.0 MPa flexural strength, and 720 Pa static yield stress. Experimental validation confirmed improved rheological behavior, with peak shear stress reaching about 950 Pa compared to 670 Pa for an OPC control, while maintaining shear-thinning characteristics. Mechanical testing of 3D-printed elements confirmed anisotropy due to layer interfaces, with 28 d strengths of 33.80 MPa (perpendicular) and 24.00 MPa (parallel), compared to higher values in the control mix. Durability tests using 1% HCl over 120 d showed better residual strength for LC 3 (≈24 MPa) than the control (≈22 MPa). Strength degradation followed a Kohlrausch–Williams–Watts (KWW) model ( R 2 = 0.91), enabling structural parameterization. A validated finite-element model of a 3D-printed wall predicted a 47% reduction in load capacity and 35% stiffness loss after acid exposure. Overall, the study demonstrates a practical, reproducible pathway for sustainable, printable LC 3 systems using regional materials. All codes and inputs are provided in Appendix for full reproducibility.
Imported from Crossref · View original