Chemistry-informed machine learning prediction of compressive strength for alkali-activated materials

Alkali-activated materials (AAMs) have emerged as promising alternative binders to curb carbon dioxide emissions from portland cement production and decarbonize concrete construction. Wide-ranging research has been undertaken to explore diverse AAMs formulations. However, predicting the engineering properties of AAMs using traditional empirical and statistical methods is still hampered by inaccuracy and incertitude. This study proposes a chemistry-informed machine learning (ML) model to estimate the compressive strength of AAMs based on their mixture proportions and the chemical compositions of their precursors and activators. Purposefully, a comprehensive dataset encompassing 676 mixture design examples was extracted from peer-reviewed published research studies. A chemistry-based feature engineering was implemented to elevate the prediction performance of four applied ML models, including support vector machine, random forest, extra trees, and gradient boosting. Consequently, accurate predictions were achieved with a low mean absolute error of 3.228 MPa. In addition, extensive computational experiments were performed using the best predictive model to unravel the impact of several mixture design variables on the compressive strength development of AAMs. Accordingly, the effect of the formulation of the precursor and activator on the compressive strength was broadly examined, and the results were thoroughly scrutinized.

Automated summary

This study introduces a machine learning model based on chemistry information to estimate the compressive strength of alkali-activated materials (AAMs). By utilizing chemical feature engineering and four applied ML models, accurate predictions were achieved with detailed analysis on the impact of mixture design variables on the compressive strength development of AAMs.