Coupling machine learning with thermodynamic modelling to develop a composition-property model for alkali-activated materials

Alkali-activation is one of the most promising routes for utilisation of versatile aluminosilicate resources. However, the variations of chemical compositions in these resources have increased the challenge of designing alkali-activated materials (AAMs) with multiple sources, posing the demand for establishing composition-property correlations that can represent a wide range of AAMs. This study proposes a data-driven approach to develop such composition-property correlations combining machine learning with global sensitivity analysis and thermodynamic modelling. The strength performance of alkali-activated concretes was investigated for a benchmark study (196 data inputs). The impact of the five key chemical compositions, CaO–SiO₂–Al₂O₃–MgO–Na₂O, has been assessed. The results show that despite the use of different aluminosilicate precursors, there appear to be coherent connections between bulk binder chemical compositions, phase assemblages, and the performance of AAMs. The composition-property correlations established via machine learning can be used to facilitate the on-demand design of AAMs utilising varying aluminosilicate resources.