Abstract
For successful construction, three‐dimensional concrete printing (3DCP) demands tailored fresh‐state properties balancing pumpability, extrudability, and buildability. In this work, machine learning was applied to predict testability and printability of Portland‐limestone cement–metakaolin (PLC‐MK) mixtures; testability was defined as the ability to form a homogeneous mixture, while printability represented the quality of the 3DCP forms. It was hypothesized that printability could be related to storage modulus, viscosity, flow onset stress, and open time, and a dataset of 54 PLC‐MK mixtures with a water–binder ratio of 0.30–0.45, metakaolin fraction up to 30% by weight, and polycarboxylate superplasticizer dosage up to 0.4% by weight of binder was generated to explore the dependence of composition on rheology. From such data, a hierarchical machine learning framework incorporating physicochemical domain knowledge was developed to classify testability and identify mixtures with rheological parameters associated with printability. Such mixtures had a water–binder ratio of 0.30–0.34 with either high metakaolin and superplasticizer content (21%–30% and 0.35%–0.37%, respectively) or low metakaolin and superplasticizer content (6%–12% and 0.17%–0.18%, respectively). To test the transferability of paste‐scale rheological descriptors to mortar‐scale printing behavior, two candidate mixtures were selected for validation. A low metakaolin–low superplasticizer mortar exhibited filament tearing during printing, while a high metakaolin–high superplasticizer mortar was difficult to extrude and displayed behavior consistent with apparent shear thickening. Although the hierarchical machine learning framework successfully predicted the targeted oscillatory rheological properties, the validation experiments revealed important limitations in translating those paste‐scale descriptors directly to mortar‐scale printing performance. The results indicate that oscillatory rheology, while informative for mixture characterization, does not fully capture the extrusion‐related phenomena governing printability. Nevertheless, hierarchical machine learning successfully modeled the composition–property relationships in PLC‐MK blends and provides a framework for future mixture design as characterization methods for extrusion and printing continue to develop.