Learning and interpreting drag force models for dense particle suspensions using graph neural networks

The dynamic behavior of fluid-particle systems is largely dependent on interaction forces between the two phases. In this paper Graph Neural Networks (GNN) are trained to predict drag force at the individual particle level utilizing data generated by Particle Resolved Simulations (PRS) of freely evolving particle suspensions. A static model uses a Graph Convolution Network (GCN) and considers data for each particle at each time step in isolation and does not take into account any history, followed by a dynamic model that takes into account the history of quantities of interest that influence the drag force at the current time-step utilizing an Attention based Graph Neural Network (GAT) combined with a Transformer architecture. It is found that the directionality of graph message passing influences the accuracy of model predictions with outward and bi-directional messaging from the particle of interest to neighboring nodes proving superior. To investigate the effect of number of neighboring particles (nodes) on the drag force, curriculum learning, which incrementally increases the number of neighbors included in the model is used and found to be superior to including all neighbors all at once. The dynamic model performs significantly better than the static model in the aggregate as well as in predicting the drag force history on individual particles, especially at higher Reynolds numbers.