Assessment of short and long-term pozzolanic activity of natural pozzolans using machine learning approaches

This investigation introduces the optimal performance models for predicting the compressive strength (CS) and pozzolanic activity index (PAI) by comparing the machine learning models. The machine learning models, i.e., multilinear regression (MLR), support vector machine (SVM), gaussian process regression (GPR), decision tree (DT), random forest (RF), and gene expression programming (GEP) have been trained (TRN) and tested (TST) by 28 and 7 data points. For the first time, the SiO2, Al2O3, Fe2O3, SiO2 +Al2O3 +Fe2O3, reactive SiO2, Blaine specific surface area, and specific gravity have been used as input variables to compute the CS, and 28 days PAI (28PAI), and 90 days PAI (90PAI) of the natural pozzolans. The multicollinearity analysis showed the SiO2, Al2O3, Fe2O3, SiO2 +Al2O3 +Fe2O3, reactive SiO2, and specific gravity have problematic multicollinearity (variance inflation factor – VIF > 10). Therefore, the root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (R), performance index (PI), and variance accounted for (VAF) metrics have been implemented to evaluate the model's performance and multicollinearity impact. From the comparison of models, it has been recorded that model GPR outperformed the MLR, SVM, DT, RF, and GEP models in predicting CS (PI = 1.29, VAF = 71.31, R = 0.8473, MAE = 0.9390 MPa), 28PAI (PI = 1.87, VAF = 94.88, R = 0.9744, MAE = 0.7295 %), and 90PAI (PI = 1.72, VAF = 88.11, R = 0.9393, MAE = 1.2444 %) in the TST phase, close to ideal values. The score, generalizability. Wilcoxon test, uncertainty analysis, Anderson-daring test, and accuracy metrics have confirmed the superiority of GPR models in predicting CS, 28PAI, and 90PAI of natural pozzolans.