Flow field reconstruction of cylinder water entry using graph neural networks

This study proposes a graph neural network (GNN)-based method for reconstructing the water entry flow field of a cylinder, which generates datasets containing different initial angles and Froude numbers through numerical simulations, and combines unstructured mesh topology with physical field data to train the model. The results show that the GNN model can learn the implied phase interface evolution laws from the known velocity and pressure fields and predict the unknown water volume fraction. Under the case of Froude number 15.65 and initial angle of 0 °, the trained GNN model successfully predicts the global process of the formation, development, and pinch-off of the cavity, with the maximum error occurring near the free surface and the flow field boundary. The prediction error of the water volume fraction is positively correlated with the Froude number and the initial angle, and the ability of the GNN model to capture the flow details decreases with the increase in the Froude number and the angle. The error can be reduced by adding similar cases to the training dataset. This study provides a new idea for the rapid reconstruction and prediction of water entry flow fields. In future research, the prediction accuracy and generalization can be further improved by adding physical constraints and optimizing models.