Abstract: Predictions of granular flows in the tumbling mill are one of the main challenges in the grinding process. Discrete Element Method (DEM) has been widely used for better understanding mechanisms of granular materials. However, this method cannot be directly applied in the real industry due to unaffordable computational cost associated with detecting and computing contacts. In the work, we propose a physics-informed machine learning model based on continuous convolution neural network (CCNN) to replace the direct calculation of particle–particle and particle-boundary collisions. The DEM simulation was used to generate the training and testing dataset at different rotation speeds. The data was used to train the model and test the prediction results. A loss function based on distance was instructed to guide model learning. The modelling of a lab scale ball mill demonstrated the accuracy and efficiency of the machine learning in comparison with DEM in the simulation of granular flows.