Comparative Study of Spatiotemporal Prediction Performance in 2D Cylinder Fluid Flow: Meshgraphnet vs Neural Implicit Representation

This study conducts a comparative analysis of the performance of MeshGraphNet (MGN) and Neural Implicit Representation (NIR) for predicting spatiotemporal dynamics in 2D cylinder flow problems. MGN leverages a graph neural network (GNN) to model spatial relationships by converting irregular mesh data into a graph structure consisting of nodes and edges. Using message-passing mechanisms, MGN captures the complex interactions between nodes, making it particularly effective for simulating nonlinear dynamics on irregular meshes. In contrast, NIR operates in a mesh-agnostic framework, representing data as continuous functions. NIR employs positional encoding to accurately capture high-resolution spatiotemporal patterns, allowing it to handle complex systems flexibly and efficiently and uses gaussian noise to enhance model’s generalization. We evaluate both models based on three key factors; prediction accuracy, computational efficiency, and generalization ability. The results of this analysis provide valuable insights into the strengths and weaknesses of each model, offering practical guidance for selecting the most appropriate method for spatiotemporal predictions in complex physical systems.

Korean Society of Mechanical Engineers (KSME 2024)