Prediction of the rutting factor of sulfur-modified bitumen using an Aquila-optimized gradient boosting model

Predicting the rutting factor (RF) of sulfur-modified bitumen from published experiments requires accounting for nonlinear relationships among binder condition, modifier dosage, and aging-related variables while recognizing differences among material systems. This study develops an Aquila Optimization-tuned Gradient Boosting (GBoost-AO) regression model and compares it with XGBoost, Random Forest, support vector regression (SVR), and an Initial-RF identity baseline. The primary analytical dataset contains 221 experimental records from seven published studies. Using five fixed model-selection folds within 185 development records, GBoost-AO achieved a mean coefficient of determination (R 2 ) of 0.8899 and a pooled out-of-fold R 2 of 0.9040. GBoost-AO had the highest mean fold R 2 among the tested configurations, although Random Forest and the identity baseline had lower mean absolute errors. Performance decreased in a retrospective assessment of 36 records from another source study, yielding R 2 = 0.1825. An independent assessment using 23 digitized points from a published 40%-modifier curve yielded a nominal GBoost-AO R 2 of −8.2120 and substantial sensitivity to digitization uncertainty. Direct partial-dependence analysis identified Initial RF as having the largest modeled response range, followed by sulfur/modifier dosage and Aging Type; Sample Age did not affect predictions of the fitted model. Nested assessment across all seven sources also showed source dependence, and a two-record refit shifted the dosage-PDP maximum from 40% to 30%. Favorable development performance therefore coexisted with poor transfer and substantial errors relative to low RF measurements. The study provides a comparative framework for investigating these limitations rather than a validated replacement for binder testing or a general-purpose material-screening tool.