Optimising sustainable alkali-activated mortar: experimental work and machine learning predictions

This pioneering research involved an in-depth experimental evaluation of the mechanical properties of ambient-cured alkali-activated mortar (AAM), while assessing an innovative machine learning (ML) driven solution for sustainable construction. A comprehensive dataset was used, comprising 635 compressive strength and 94 flexural strength data points, including data from previous studies. The performance of six ML algorithms in predicting the compressive and flexural strengths of AAM was evaluated. Hyperparameter optimisation was performed with Optuna and ten-fold cross-validation. Multi-objective optimisation aimed to maximise compressive strength while minimising the carbon dioxide footprint. The findings highlight the significant impact of ground granulated blast-furnace slag (GGBS) content on strength, with higher GGBS improving compressive and flexural strengths but reducing workability. The highest compressive strength was 56.28 MPa at 28 days, for the AAM with 100% GGBS. The highest flexural strength was 0.580 MPa at 28 days, with 75% GGBS. Extreme gradient boosting was found to be the most reliable model in predicting the compressive strength, achieving a coefficient of determination (⁠$R2$⁠) of 98.1% on training data and 86.8% on testing data. Extra tree regression showed high accuracy in predicting the flexural strength of the AAM, achieving $R2$= 90% on the testing dataset. A user-friendly interface was developed for predicting the mechanical properties of AAMs.