Effect of number and surface area of the aggregates on machine learning prediction performance of recycled hot-mix asphalt

• Published in: Construction & building materials Vol. 445; p. 137788
• Main Authors: Atakan, Mert, Valentin, Jan, Yıldız, Kürşat
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 27.09.2024
• Subjects: Aggregate surface area; Asphalt mixture; Machine learning; Marshall properties; Number of aggregate particles; Random forest; Stiffness modulus; Virtual aggregates; Number of aggregate particles; Stiffness modulus; Virtual aggregates; Aggregate surface area; Machine learning; Asphalt mixture; Random forest; Marshall properties
• ISSN: 0950-0618
• DOI: 10.1016/j.conbuildmat.2024.137788

This study addresses the challenge of designing hot-mix asphalt by evaluating the impact of aggregate surface area (ASA) and the number of aggregates (NA) in machine learning (ML) models. A dataset of 107 asphalt mixtures containing 0–50 % reclaimed asphalt pavement (RAP) was analyzed. Virgin aggregates and RAP particles were counted and measured via digital photography to calculate NA and ASA, with specific surface areas determined in a physics engine environment. Then, measured average aggregate particle weights were calibrated using 13 specimens. Various ML models were developed with the random forest algorithm, incorporating different input feature sets (IFS), including NA, ASA, and other basic features of the mixtures. Results revealed that including NA and ASA did not significantly improve model performance compared to using only gradation percentage inputs. Consequently, IFS-4, which includes only gradation inputs, was recommended for simplicity. The most crucial features were found to be gradation-related, with R² values around 0.90 and above achieved for stiffness modulus (ITSM), air voids, Marshall stability (MS), and theoretical maximum density (Gmm). Specifically, the test R² values for air voids, ITSM, and MS were 0.96, 0.89, and 0.87, with a mean absolute percentage error (MAPE) of 5.8 %, 5.4 %, and 5.6 %, respectively. Predictions for Gmm demonstrated the highest performance across all metrics with an R² value of 0.99. Air void content predictions performed better than those for ITSM, MS, and MF regarding R² and mean squared error (MSE) values, although their MAPE values were similar. These findings suggest that while NA and ASA provide additional details, gradation features are the most critical inputs for accurate ML model predictions in asphalt mixture design. [Display omitted] •The number and surface area of aggregates were calculated in different asphalt mixtures.•The number and surface area of aggregates did not make a significant difference in ML models.•Aggregate gradation had the highest feature importance in the ML model.•R2 values around 0.90 and above were obtained for air voids, ITSM, Marshall stability, and Gmm.