Learning Fluid Simulation with Dynamic Boundaries via Temporal Awareness

Lagrangian fluid simulation plays a significant role in virtual reality, film production and digital twins. Recently, data-driven deep learning methods have shown great potential in enhancing simulation efficiency. However, most existing studies focus on static scenes, limiting their ability to generalize to dynamic scenarios. To better adapt our model to complex dynamic boundary conditions in practical applications, we propose a continuous convolution method based on temporal awareness. Our approach employs a dual-branch continuous convolution network as the backbone, using signed distance field (SDF) values to represent boundaries. The network introduces a Transformer-style Temporal-Aware Guidance module (TAGFormer) to improve the network's temporal modeling of dynamic boundary interactions. Additionally, a Gated Feature Fusion module (GFF) integrating Temporal Context adaptive Kernel convolution (TCKconv) is also designed to dynamically aggregate contextual information and assess the impact of dynamic boundaries on fluid flows. To validate the effectiveness of our fluid simulation approach, we construct multiple datasets containing scenes with dynamic boundaries. Extensive experimental results show that our method achieves state-of-the-art results on tasks of fluid simulation in dynamic scenarios, in terms of both accuracy and generalization.