Explicit gradient boosting machine for accurate compressive strength prediction of metakaolin-based geopolymer concrete: feature importance and optimization

Understanding the compressive strength (CS) of metakaolin-based geopolymer concrete (MK-GPC) requires capturing the complex, nonlinear interactions among mix design parameters, precursor chemistry, and curing conditions, complexities that often elude traditional modeling approaches. In this study, a Gradient Boosting Machine (GBM) algorithm was deployed within a transparent machine learning framework to predict CS of MK-GPC with high accuracy, drawing on a curated dataset of 235 experimental mix designs. A comprehensive hyperparameter search across 1440 configurations identified the optimal model, comprising 950 decision trees, a learning rate of 0.1, and a maximum tree depth of two. This configuration yielded excellent predictive performance, with R 2 of 0.99 and 0.96 for training and testing sets, respectively, demonstrating strong generalization across a wide range of mix scenarios. Beyond predictive accuracy, the model emphasized interpretability through a combination of Shapley additive explanations, Gini feature importance, and Partial Dependence Plots, which together explained the relative impact and nonlinear influence of input variables. Five parameters: coarse-to-fine aggregate ratio (CF), water-to-NaOH ratio (WN), sodium hydroxide content (SH), extra water (EW), and silica-to-alumina ratio (SA) accounted for over 85 % of the model's explanatory power. Specific trends were identified, including peak CS at CF ≈ 2.12, SH ≈ 116.2 kg/m 3 , EW ≈ 6.1 kg/m 3 , SA ≈ 2.63, and WN within the range of 7.46–9.39. The model also captured critical thermal and temporal effects, with strength development peaking at ∼35 °C and plateauing beyond 28 days. By bridging data-driven modeling with chemical and materials science understanding, this study offers a rigorously validated and interpretable approach to optimizing MK-GPC formulations, which enables performance-based, sustainable mix design through evidence-led decision-making.