Regularized multivariate polynomial regression analysis of the compressive strength of slag-metakaolin geopolymer pastes based on experimental data

Abstract

The compressive strength of geopolymer pastes is a fundamental parameter in the research on the macromechanical performance of geopolymer concrete. However, the factors that affect the strength are complex. In this work, we developed a regularized multivariate polynomial regression (MPR) model to predict the compressive strengths of 40%/60% ground granulated blast furnace slag (GGBFS)-metakaolin (MK)-based geopolymer pastes. Moreover, by conducting experiments on 120 groups of specimens, a dataset that considers the effects of the alkali activator concentration, SiO2/Na2O molar ratio, and mass ratio of the alkali activator to aluminosilicate materials (i.e., the liquid–solid ratio) is created. Based on the framework of machine learning, the established dataset is used to train and test the MPR model. Notably, in the optimizing process of the MPR model, Lasso regularization is employed to prevent the overfitting problem caused by regression coefficients that are too large. The proposed MPR models show robust estimation capabilities, with R2 values of 0.946 and 0.927 for predicting the 7-day and 28-day compressive strengths in the testing processes. Unlike other black-box algorithms such as artificial neural networks, the proposed MPR model is much more transparent and easier to use by researchers. In addition, a partial dependence plot (PDP) is utilized to quantify the importance of influencing factors of the experimental dataset. The regularized regression method discussed in this work can help researchers map the nonlinear relationship between the experimental variables and results.