Evaluation of strength development of one-part alkali-activated materials through machine learning and thermodynamic modelling

One-part alkali-activated materials (AAMs) have emerged as one of the sustainable alternatives to conventional cement formulations. Although intensive research has been conducted on one-part AAMs, their formulation remains challenging due to the diverse chemistries of precursors available from different sources. Machine learning offers a data-driven approach that can link binder chemistries to the performance properties (e.g. strength). However, existing studies applying machine learning have often lacked explainability and mechanistic understanding of how model inputs drive predictions, which are essential for ensuring model reliability and supporting rational formulation strategies. This study presents an integrated chemistry-informed machine learning (CIML) framework that combines machine learning with thermodynamic modelling to link binder chemistry (precursors and solid activators), phase assemblages, and compressive strength in one-part AAM formulations. Strength prediction model was trained and tested using a compiled dataset from the literature. To address the ‘black-box’ nature of machine learning models, Shapley Additive exPlanations (SHAP) were used to identify and interpret the relative influence of each input feature on strength predictions. To improve explainability and support validation, thermodynamic modelling was employed to simulate phase assemblages and porosity changes, offering a mechanistic grounding specifically for the chemical features driving the strength development of one-part AAMs. This integrated framework delivers an explainable and validated approach that reveals how specific chemical features govern strength development, supporting more rational formulation strategies for one-part AAM systems. • Chemistry-informed machine learning framework links binder chemistry to strength. • Model maps binder composition and mix design to compressive strength. • SHAP improves model interpretability by quantifying feature influence. • Thermodynamic modelling provides mechanistic insights into predicted strength. • Framework enables interpretable, data-driven mix design for one-part AAMs.