A Bi-Path Continuous Convolutional Network for Efficient Lagrangian Fluid Simulation

Data-driven Lagrangian fluid simulators provide a fast and approximate solution to fluid dynamics problems, offering greater convenience compared to traditional numeri-cal methods. However, current approaches often overlook the importance of fluid-solid coupling and struggle to balance high accuracy with efficiency. To address this issue, we propose a bi-path continuous convolutional network that processes fluid and obstacle features independently on separate paths. The lightweight attention-gated fusion module is seamlessly embedded between the paths to selectively fuse these features, emphasizing critical particle collisions. Moreover, we introduce continuous central difference convolutions and historical velocity inputs to provide more effective features for predicting accurately. Experimental results on multiple datasets show that our approach strikes an excellent balance between prediction accuracy and inference time while achieving great generalizability.