Generalized Additive Model for Predicting ECBR of Stabilized Subgrades for Pavement

This study presents a Generalized Additive Model (GAM) to predict the effective CBR (ECBR) of soil subgrade stabilized with recycled plastic waste under repeated loading (RL) conditions. Univariate and bivariate GAM were explored to evaluate the influence of parameter interactions on ECBR predictions. The Bivariate GAM model demonstrated superior performance, achieving a cross-validation mean squared error of $9.2203\times 10^{-5}$ , mean absolute error (MAE) of 0.007, root mean squared error (RMSE) of 0.009, mean absolute percentage error (MAPE) of 0.077 and percentage bias (PBIAS) of 0.00026% indicating minimal deviation and high accuracy. Localized interpretability analysis yielded ECBR predictions closely aligned with the actual values, highlighting the model’s robustness. Further interpretability analysis through mutual information and partial dependence plots identified resilient modulus ( $\text{M}_{\mathrm {r}}$ ), optimum moisture content (OMC), maximum dry density (MDD) and content of plastic waste (CP) as key predictors influencing ECBR. Among these Mr demonstrated the highest mutual information score (1.00), emphasizing its critical role and direct association with RL conditions. The Predicted ECBR values, with a 1.130% average deviation from literature-derived values, confirm the bivariate GAM’s exceptional accuracy and reliability in handling the diverse, unseen datasets. These findings underscore the model’s potential as a trustworthy, interpretable and efficient alternative to traditional methods for road pavement applications.