Prediction and modelling marshall stability of modified reclaimed asphalt pavement with rejuvenators using latest machine learning techniques

Abstract

Abstract The primary problem with the experimental evaluation of Marshall stability (MS) of reclaimed asphalt pavement (RAP) is the inherent complexity and variability involved in the process. Traditional experimental methods for predicting MS can be time-consuming, labor-intensive, and costly. In the present research, an effort has been made to assess the most appropriate machine learning model for the prediction of MS of RAP. The study addresses the problem of accurately predicting MS by using a variety of input parameters derived from experimental work. The data for models was split in 7:3 for training and testing of models. Bitumen content (BC %), virgin binder percentage (VB %), virgin binder performance grade (VB-PG), RAP percentage (RAP %), RAP binder percentage (RAPB %), RAP binder PG (RAPB-PG), rejuvenator type (Rej type) and rejuvenator percentage (Rej %) were applied as input parameters for MS prediction. Several machine learning models including random tree (RT), M5P, Gaussian process (GP), support vector machine (SVM), and random forest (RF) were utilized for determining the most appropriate prediction model. Seven metrics were used for assessing the performance of these models, such as CC, MAE, RMSE, RA, RRSE, WI, and NSE. Based upon these metrics, the RF model is found to outperform the other applied models with the values of CC = 0.9959 and 0.9763, MAE = 0.3129 and 0.7847, RMSE = 0.3976 and 1.0492, RAE = 9.0062 and 21.8247, RRSE = 9.3624 and 23.6832, WI = 0.998 and 0.984 and NSE = 0.991 and 0.944 for training and testing stages, respectively. Also, box plots and sensitivity analysis confirm the superiority of the RF model over other models. Finally, the sensitivity analysis suggests the importance of bitumen content in the prediction of MS of reclaimed asphalt pavement modified with rejuvenators.