Graph convolutional networks applied to unstructured flow field data

Many scientific and engineering processes produce spatially unstructured data. However, most data-driven models require a feature matrix that enforces both set number order of features for each sample. They thus cannot be easily constructed an dataset. Therefore, graph based model to perform inference on fields defined mesh, using Graph Convolutional Neural Network (GCNN) is presented. The ability the method predict global properties from irregular measurements with high accuracy demonstrated by predicting drag force associated laminar flow around airfoils scattered velocity measurements. network can infer field samples at different resolutions, invariant in which within sample are GCNN method, inductive convolutional layers adaptive pooling, able this quantity validation $R^{2}$ above 0.98, Normalized Mean Squared Error below 0.01, without relying spatial structure.