A physics-informed neural network for predicting the drag coefficient of irregular particles utilizing computational fluid dynamics

The drag coefficient of particles is a critical physical parameter in multiphase flow systems. Despite the publication of numerous empirical models for drag prediction, these models exhibit limitations as they are often applicable only under specific experimental conditions. Consequently, an accurate and universally applicable formula of drag coefficient that encompasses all particle shapes in high Reynolds number flows remains elusive at present. Based on the existing empirical formula and the prediction ability of neural network, this paper proposes a physics-informed neural network (PINN) to predict the drag coefficient. In addition to sphericity and Reynolds number, the input of model also includes shape description symbol that significantly influence drag coefficient corroborated by the existing literature. A computational fluid dynamics model was developed to calculate the drag coefficients for a series of particles. These particles are solid stones of irregular shape. The shape characteristics of particles and the drag coefficient together form the dataset. Different particles were selected to conduct the pipe sedimentation experiment and develop the corresponding computational fluid dynamics–discrete element method model. The predicted drag coefficient from the PINN model was coupled into the model to verify the practicality of the prediction results. The PINN model is suitable for irregular particles and provides a new feasible method for predicting drag information with enhanced generalizability and reliability.