Explicable AI-based modeling for the compressive strength of metakaolin-derived geopolymers

Abstract

Geopolymers (GPs) are produced using a variety of alumina and silica-rich ingredients, including natural sources, such as calcined clays, as well as waste by-products, like fly ash and slag. It is difficult to measure the effect of each parameter on the strength of GP composites via experimental research. In this regard, this research aimed to build artificial intelligence (AI)-aided estimation models to quantify the impact of mix proportions on the compressive strength (CS) of MK-based GP mortar. Gene expression programming (GEP) and multi-expression programming (MEP) tools were used for models' development due to their advantage of yielding model equations for future predictions. Hyperparameters in GEP and MEP were systematically adjusted to enhance the models' optimal predictability. The regression and error analysis proved the superiority of MEP models over GEP models. In comparison to the GEP model, which had R2, MAE, MSE, RMSE, and objective function values of 0.90, 4.02, 25.3, 5.03, and 0.08, respectively, the MEP model demonstrated greater accuracy with values of 0.96, 3.24, 16.4, 4.05, and 0.02 respectively. Furthermore, SHapley Additive exPlanations (SHAP) analysis was conducted to determine the effect of various parameters on the CS of MK-based GP mortar. The established models' equations may be exploited to assess the CS of MK-based GP composites with various input parameters, hence reducing the need for further laboratory testing. Implementing this approach not only improves the efficiency of material design but also encourages the sustainable utilization of locally accessible resources in the manufacturing of geopolymers.

Keywords

• Geopolymers
• Metakaolin
• Compressive strength
• Modeling
• Machine learning
• Prediction models