Prediction of asphalt binder elastic recovery using tree-based ensemble bagging and boosting models

Abstract

The Multiple Stress Creep Recovery (MSCR) test has been proposed and implemented over the last two decades as an alternative to traditional tests, especially the elastic recovery (ER) test, to better capture rutting resistance. This paper presents a study on predicting asphalt binder ER using MSCR test results through ensemble learning methods, based on a database of more than 700 points for mostly polymer-modified binders. The ensemble approach involved tree-based Bagging (Random Forest and Extra Trees) and Boosting (Adaboost, gradient boosted decision tree, XGBoost, LightGBM, and CatBoost). The findings revealed the effectiveness of ensemble learning methods in capturing complex relationships within the data. Extra Trees and XGBoost emerged as the most accurate models based on coefficient of determination, mean absolute error, and root-mean-square error, compared to the other mentioned models and Neural Networks. These models offer precise predictions of asphalt binder ER from MSCR test results, surpassing the ER-DSR test proposed in the literature. Recovery at stress levels of 0.1 and 3.2 kPa were identified as the top influential features. In addition, clustering methods were employed to detect patterns within the dataset. The outcomes highlighted difficulties in discerning binders based on existing MSCR parameters, suggesting opportunities for refining MSCR analysis, and exploring additional tests to improve the distinction among asphalt binders. This investigation provides practical implications for states still utilizing PG-Plus, promoting the adoption of MSCR specifications for asphalt binder characterization.