A hybrid framework of attention-based tabular network and ensemble learning with generative adversarial network for stiffness modulus prediction

The dynamic modulus is a fundamental property in pavement engineering, quantifying the viscoelastic stiffness of asphalt mixtures under cyclic loading and critical in mechanistic-empirical design frameworks for modeling pavement responses to traffic loads and temperature variations. However, its in-situ measurement remains costly, time-consuming, and technically demanding, creating a pressing need for accurate and efficient predictive models. Traditional machine learning approaches have been explored to estimate the dynamic modulus, but their predictive capabilities are often constrained by scarce, heterogeneous datasets, limiting generalization. To address these challenges, this study proposes a hybrid learning framework that integrates Attention Based Tabular Network (TabNet)-Extreme Gradient Boosting (XGB) with Conditional Tabular Generative Adversarial Networks (CTGAN) for synthetic data augmentation. The framework leverages TabNet’s attention-based deep feature representation and XGB’s boosting capabilities to enhance predictive accuracy and robustness under data-scarce conditions. A comprehensive evaluation using different statistical metrics demonstrated that the proposed model trained on CTGAN-synthesized data consistently outperformed models trained solely on real data and those using alternative augmentation methods, achieving a superior coefficient of determination (R²) of 0.98. SHapley Additive exPlanations analysis further identified Marshall Stability (28.7 %), Flow (18.8 %), and asphalt content (13.4 %) as the most influential features in dynamic modulus prediction. The CTGAN-augmented proposed hybrid model reduced Mean Absolute Erro by 67.30 %, improved R² by 13.27 %, and decreased Standard Error Ratio by 59.38 %, reflecting significant improvements across all evaluation metrics. These findings highlight the transformative potential of GAN-based data augmentation to advance predictive modeling and promote more efficient data-driven practices in pavement engineering. • A novel TabNet-XGB hybrid model is proposed to predict the stiffness modulus of asphalt mixtures with high accuracy and generalization. • CTGAN-based data synthesis effectively addressed data scarcity by generating statistically consistent synthetic datasets. • The CTGAN-augmented TabNet-XGB framework and achieves superior accuracy, cutting MAE by 67.30 %, improving R² by 13.27 %, and lowering SER by 59.38 %. • SHAP analysis identifies Marshall Stability, Flow, and asphalt content as the most influential features affecting stiffness predictions. • The framework minimizes reliance on labor-intensive testing, enables scalable pavement modeling, and supports efficient pavement design.