Enhancing asphalt mix design with transfer learning and hybrid artificial intelligence

Asphalt content (AC) in hot mix asphalt is a critical factor influencing pavement durability and performance. Traditional determination methods are labor-intensive, costly, and prone to human error, limiting scalability. Artificial intelligence (AI) and machine learning (ML) offer efficient alternatives that improve accuracy and reduce manual effort, but they typically require large training datasets. This study addresses that challenge by proposing a hybrid transfer learning model (RF-DNN-T) that leverages a comprehensive literature-based dataset of 848 mix designs for pre-training, followed by fine-tuning with only 96 target-specific samples. The framework combines the feature-selection capabilities of Random Forest (RF) with the nonlinear pattern recognition strength of Deep Neural Networks (DNN) to enhance predictive efficiency and reliability. By transferring knowledge from multiple laboratories, the model minimizes experimental effort while maintaining high accuracy. The RF-DNN-T model is benchmarked against conventional ML approaches using SHapley Additive exPlanations (SHAP) for interpretability. Results show the model achieves the lowest testing errors (MAE = 0.08, MSE = 0.02, MAPE = 1.61) and the highest R² (0.97), indicating excellent accuracy, generalization, and a strong correlation between predictions and experimental data. SHAP analysis identifies voids as the most influential feature, contributing 32 % to prediction impact. This work demonstrates that the RF-DNN-T model can accurately predict AC with minimal laboratory data, significantly reducing experimental workload while preserving precision in asphalt mix design.