Fast Prediction of Droplet Impingement using GNN

Recently, machine learning and data-driven methodshave been extensively applied into CFD realm to serveas an optimization tool to reduce computational costs.The graph networks show strong stability androbustness to improve the efficiency of CFD solver andare consistent with flow visualization techniques.However, the capability to investigate the effect of localgeometry configurations in a simulation of multi-phaseflows is still underdeveloped, which is one of the mostcrucial aspects in spray applications that generatemillions of droplets from a small liquid volume. In thispaper, we train a special graph network,MeshGraphNets, to fast predict snapshots of 2Dvelocity and pressure fields for a water droplet impactson arbitrary smooth surfaces under the supervisedlearning framework. The datasets are generated byopen-source code, Openform, and are split into trainingand testing cases with different distributions. In the end,the model has the ability to predict droplet dynamics onunseen substrate materials and it is more efficient toassess than the CFD solver within finer resolutions. Weexpect that this approach will allow us to develop abuilding block for prediction of thermal spray particledeposition and be easily extended to a wide range ofinterface tracking mechanisms.