Enhancing the predictive accuracy of marshall design tests using generative adversarial networks and advanced machine learning techniques

Experimentally determining the Marshall design test results for Air voids (Va), Marshall Stability (MS), and Marshall Flow (MF) in hot mixed asphalt (HMA) is often expensive, time-consuming, and requires skilled personnel. To address these challenges, various traditional machine learning (ML) models have been employed to optimize the mix design of HMA. However, their performance is significantly limited by the size and quality of the training dataset. To address these limitations, this study employed Generative Adversarial Networks (GANs) to augment the dataset, which consisted of 184 samples gathered from four construction projects in Pakistan. The augmented dataset was then used to train two advanced ML models: Gene Expression Programming (GEP) and ensemble learning with stacking (ELS). A thorough comparison of the models trained on both original and GAN-augmented datasets was conducted using a range of statistical metrics to evaluate their predictive performance. Additionally, sensitivity and parametric analysis were performed to assess the impact of input variables on the outputs. The results demonstrate that GAN-augmented data significantly improved model accuracy, with GEP and ELS achieving R² values exceeding 0.93 in all cases. Furthermore, GEP models provided interpretable equations for HMA predictions. Sensitivity analysis identified binder content (Pb%) as the most influential variable, contributing over 55 % to the variance in Va and MF predictions and 61.56 % in MS. In contrast, other inputs had minimal influence, which was consistent with the experimental findings. This study highlights the potential of advanced ML techniques and data augmentation in developing reliable predictive models for Marshall design test results, advancing efficient HMA design practices.