Prediction of crushed numbers and sizes of ballast particles after breakage using machine learning techniques

Modeling the breakage of ballast particles is a key task for modeling the plastic deformation of ballast subjected to several loading cycles in DEM simulations. In this study, a series of single-particle crushing tests were per -formed on 700 ballast particles to specify the numbers and sizes of ballast particles crushed after ballast breakage. Four classification algorithms (support vector machine (SVM), backpropagation neural network (BPNN), random forest (RF), and CatBoost (CB)) were used to predict the number of particles crushed after ballast breakage. In addition, the particle sizes were estimated using the RF, SVM, BPNN, and eXtreme Gradient Boosting (XGB) regression algorithms. The initial particle size, shape, material, and loading conditions were considered variables for regression and classification. The dataset was divided into 75% and 25% as the training and test sets, respectively. The results indicated that CB and RF-based classification, with an accuracy of more than 91%, and RF and XGB based regression, with a normalized root-mean-squared error (NRMSE) of less than 0.08, appropriately estimated the number and sizes of crushed ballast particle. Moreover, the DEM simulation of ballast breakage using the particle bonded model (PBM) validated the values predicted using the machine learning models. The proposed method can be used to simulate ballast as bonded clumps instead of bonded pebbles with an acceptable breakage modeling accuracy.