Machine learning-based prediction of 90-day tensile strength in geopolymer concrete under varying input-variable scenarios

Accurate prediction of the long-term tensile strength (TS) of geopolymer concrete remains a complex challenge owing to the nonlinear and interactive nature of the governing parameters. This study develops and comparatively evaluates four machine-learning models linear regression (LR), decision tree (DT), random forest (RF) and support-vector regression (SVR) to predict 90-day tensile strength from a suite of physicochemical and mixture-related variables. Eleven systematically designed input scenarios were formulated to quantify the influence of different feature combinations on predictive performance. The results demonstrate a pronounced dependence of model accuracy on input configuration. Early scenarios that employed limited variable sets yielded only moderate performance (R² ≤ 0.885, with most models below 0.86), whereas the optimal configuration (S11), which incorporates water absorption (WA), ground-granulated blast-furnace slag (GGBS), nano-silica, pH and bulk density (BD), substantially enhanced predictive capability. Among the models, DT achieved the highest test accuracy (R² = 0.984, RMSE = 0.751) together with the smallest generalisation gap (training R² = 0.988), indicating strong predictive performance within the investigated dataset and limited evidence of overfitting under the applied validation procedures. RF and SVR also delivered high predictive performance (R² = 0.964 and 0.956, respectively), while LR exhibited comparatively lower accuracy, underscoring the importance of nonlinear modelling approaches. Permutation-importance sensitivity analysis identified WA as the dominant influential parameter (mean ΔR² = 0.911), followed by GGBS (0.588) and nano-silica (0.277), indicating that WA and GGBS were the strongest predictors of TS within the trained model, consistent with though not a direct proof of their established physical roles in porosity and geopolymeric gel formation. Collectively, the findings indicate that the integration of advanced tree-based learning algorithms with comprehensive input characterisation provides a promising and computationally efficient framework for modelling the long-term tensile behaviour of geopolymer concrete within the range of mixtures examined; broader industrial application will, however, require external validation on independent datasets and formal uncertainty quantification.