A combined data-driven and discrete modelling approach to predict particle flow in rotating drums

Abstract

This work developed a data-driven model combined with the discrete element method (DEM) to predict the features of the particle flow inside a drum. The SVR (Support Vector Machine for Regression) method was adopted to predict two important properties of particle flow, angle of repose and collision energy. The model was trained and tested using 142 sets of data generated from the DEM simulations. The Kennard-Stone (K-S) method, due to its advantages over random selection method, was adopted to select training data. The optimal values of the parameters in the SVR model were determined by the grid-search method. Results showed the robust SVR model was able to predict angle of repose and collision energy under different conditions, such as changing drum size, rotation speed, particle-wall sliding friction and filling level, reasonably well with R2 values of 0.92 and 0.86, respectively. The relatively less accurate prediction on collision energy was discussed. The study showed that this approach can be implemented to link off-line DEM simulation with rapid prediction of particle behaviour in various industrial applications.