Integrating advanced data imputation techniques and machine learning models to develop a predictive model for compressive strength of geopolymer concrete

Abstract Geopolymer concrete (GPC) offers an environmentally sustainable alternative to Portland cement, reducing carbon emissions and utilizing fly ash. However, accurately predicting GPC’s compressive strength remains challenging due to complex interactions among the chemical, physical, and mineralogical properties of fly ash, alkali activators, and curing conditions. This study presents a novel predictive model that integrates advanced data imputation techniques and machine learning models—Artificial neural network (ANN), Random Forest (RF), and Extreme Gradient Boosting (XGB)—to address missing data and material variability. The final ANN model, which incorporates the chemical composition, surface area, particle size, and amorphous content of fly ash, achieved a testing accuracy of 90% in predicting the compressive strength of GPC, with values ranging from 10 to 80 MPa. The ANN model for compressive strength prediction of GPC outperformed XGB and RF by 38% and 29%, respectively, highlighting its superior ability to generalize to new data. SHAP analysis identified total water content, Fe 2 O 3 , and BET surface area as the most influential parameters, with optimized values leading to enhanced compressive strength. Additionally, ANN-based imputation consistently outperformed traditional methods, namely MICE, SVD and KNN, improving data integrity and achieving superior model performance with the highest R 2 and lowest RMSE and MAE for both train and test data sets. This research provides a robust methodology for predicting GPC compressive strength, offering valuable insights for mix design and addressing the challenges of material variability and missing data.