Deep learning-based modeling of Sulfur-Extended Asphalt (SEA) properties: A novel computational framework
Abstract
This research presents a deep learning method for modeling the properties of sulfur-extended asphalt (SEA), a material that has the potential to enhance pavement durability and sustainability. Traditional methods may struggle to capture the complex behavior of modified materials, such as SEA. We developed four deep learning neural network (DLNN) models to predict key SEA performance parameters: maximum load (ML), fracture energy (FE), secant stiffness (SS), and flexibility index (FI). The Bayesian optimization algorithm was used to tune the model hyperparameters. Early stopping and dropout regularization techniques were implemented to prevent the models from overfitting. The models were trained on an interpolated dataset and demonstrated strong predictive capabilities, with R² values ranging from 0.929 to 0.971 for the training data and 0.730 to 0.955 for the unknown experimental data, indicating that the DLNNs effectively captured the intricate relationships between input parameters and SEA properties. Sobol global sensitivity analysis identified the most influential factors affecting SEA performance, providing valuable insights for optimizing mix designs. It revealed that binder content, binder type, aging, and mixing method are the most crucial factors in determining SEA’s ML, FE, SS, and FI. This AI-driven approach offers a promising avenue for improving the design and application of SEA in pavement construction, potentially leading to increased durability, reduced maintenance costs, and a more sustainable use of materials.