Predicting Asphalt-Mixture Fatigue Life Under Multi-Peak Axle Loading: A Comparison of Machine Learning and Multilinear Regression Models

Fatigue cracking is a major concern in asphalt-pavement design and maintenance and is influenced by structural configuration, material properties, environmental conditions, and traffic loading. Full-scale fatigue evaluation is costly and time-consuming, while laboratory testing remains demanding, particularly for low-amplitude signals associated with long fatigue lives. This study compares five machine learning approaches—artificial neural networks, support vector regression, Gaussian process regression, random forests, and gradient boosting—with a recalibrated multilinear model for predicting asphalt-mixture fatigue life under axle-induced multi-peak loading. The models use four physically defined signal descriptors: maximum strain, number of positive peaks, normalized positive area, and average duration per positive peak. Development and evaluation employed 78 unique experimental observations and repeated nested cross-validation. The recalibrated multilinear model achieved the most favorable mean held-out performance, with a mean absolute error of 0.181, root mean square error of 0.224, coefficient of determination of 0.853 in logarithmic fatigue-life space, and a20-index of 32.1% on the original cycle scale. Gaussian process regression was the strongest nonlinear method and produced closely comparable results. However, fold-wise variability overlapped substantially, and a corrected paired comparison did not detect a statistically significant MAE difference between the multilinear model and GPR (p = 0.525). For this small, structured database, the results did not demonstrate a consistent predictive advantage from added model complexity. Physically meaningful descriptors combined with an interpretable model can therefore support preliminary assessment and help prioritize laboratory testing within the investigated domain.