Shear behaviour of recycled asphalt pavement and natural aggregate: Large direct shear testing and machine learning based strength prediction

This study conducts a systematic series of large direct shear experiments to examine the effect of replacing natural aggregate (NA) with reclaimed asphalt pavement (RAP) in coarse (4.75–10 mm) blends on shear behaviour and evaluates the performance of data driven models for predicting shear strength. Eight mixtures of NA and RAP were tested in a large direct shear machine under normal stresses of 50, 100 and 150 kPa, generating 1440 records. Increasing RAP content consistently reduced peak shear stress and friction angle while raising cohesion, indicating a transition from friction dominated to binder-assisted resistance. Field emission scanning electron microscopy (FESEM) confirmed this mechanism by revealing angular, rough asperities on uncoated NA and smoother, bitumen sheathed surfaces on RAP, which diminish inter-particle interlock but add adhesive bonding. Four easily measured variables such as RAP percentage, complementary NA percentage, seating load and horizontal displacement were used to train artificial neural network (ANN), random forest (RF) and extreme gradient boosting (XGB) models on 80 % of the data while the remaining 20 % was used as an independent test set. All algorithms achieved R² > 0.99 on unseen data, with ANN delivering the lowest error and the narrowest training-testing gap. In conclusion, the study not only clarifies the mechanisms by which RAP alters shear strength but also introduces a combined experimental – computational framework that links FESEM evidence with high accuracy machine learning forecasts based on four basic test parameters, supporting the sustainable incorporation of RAP in performance based earthworks and pavement applications.