A Novel GNN-TCN Framework for Enhanced Turbulent Flow Predictions

In this paper, we present a new framework that combines Graph Neural Networks with Temporal Convolutional Networks (GNN-TCN) for turbulent flow modeling with Smoothed Particle Hydrodynamics (SPH). Our method employs dilated causal convolutions to learn temporal dependencies across sequences of graph representations, effectively capturing the multi-scale evolution of particle systems. The framework introduces a hierarchical modeling approach that progressively incorporates physical constraints while maintaining computational efficiency, where novel parameterized smoothing kernels enhance spatial representation while preserving physical consistency. We evaluate our approach to three canonical flow problems. Results demonstrate superior performance in both short-term prediction accuracy and long-term stability compared to SODA baselines, with significant improvements in different metrics. Our method successfully captures complex temporal dependencies in turbulent flows while maintaining physical consistency and computational efficiency that are suitable for practical applications.