Data-driven investigation and multi-objective optimization design of alkali-activated concrete

This study employs a Bayesian-optimized XGBoost algorithm to develop machine-learning models for predicting the 28-day cubic compressive strength, porosity, and CO2 emission of alkali-activated concrete. Utilizing a dataset of 483 samples with 13 extracted input features, the models demonstrate high predictive accuracy, evidenced by robust R² values and low MAE, MAPE, and RMSE metrics on both training and testing datasets. The SHAP algorithm is employed to interpret the prediction process, revealing that the mass percentage of fly ash in the precursor significantly influences compressive strength and porosity. Other critical factors include curing temperature, superplasticizer-to-precursor ratio, and alkali activator-to-precursor ratio for compressive strength, while precursor dosage and water content critically affect porosity. The study suggests optimal conditions for minimizing porosity, including precursor content below 500 kg/m³, mass percentages of fly ash in precursor and the coarse aggregate content in total aggregate less than 60%, high content of Na2SiO3 in alkali activators. A superplasticizer-to-precursor ratio above 3% and a water-to-precursor ratio below 25% are suggested to balance workability and porosity. In addition, a multi-objective optimization framework is developed, with the target of achieving desired compressive strength of alkali-activated concrete, as well as reduced CO2 emission and porosity.