DualFluidNet: an Attention-based Dual-pipeline Network for FLuid Simulation

TL;DR

DualFluidNet tackles SPH-based 3D fluid simulation by learning with two complementary pathways: a Main Path powered by CConv for broad fluid dynamics and a Constraint-guided Path via ASCC to enforce momentum conservation, fused through an Attention-based Feature Fusion Module. It introduces a Type-aware Input Module to distinguish fluid and solid particles, improving fluid-solid coupling, and adds Tank3D to stress-test complex scenes. Quantitative results show superior accuracy and distribution fidelity, while qualitative analysis demonstrates faithful adherence to physical laws and robust generalization, including rotation invariance. Overall, the approach offers a practical, physics-consistent, near-real-time solution for complex fluid scenarios with strong generalization across diverse terrains and obstacle configurations.

Abstract

Fluid motion can be considered as a point cloud transformation when using the SPH method. Compared to traditional numerical analysis methods, using machine learning techniques to learn physics simulations can achieve near-accurate results, while significantly increasing efficiency. In this paper, we propose an innovative approach for 3D fluid simulations utilizing an Attention-based Dual-pipeline Network, which employs a dual-pipeline architecture, seamlessly integrated with an Attention-based Feature Fusion Module. Unlike previous methods, which often make difficult trade-offs between global fluid control and physical law constraints, we find a way to achieve a better balance between these two crucial aspects with a well-designed dual-pipeline approach. Additionally, we design a Type-aware Input Module to adaptively recognize particles of different types and perform feature fusion afterward, such that fluid-solid coupling issues can be better dealt with. Furthermore, we propose a new dataset, Tank3D, to further explore the network's ability to handle more complicated scenes. The experiments demonstrate that our approach not only attains a quantitative enhancement in various metrics, surpassing the state-of-the-art methods but also signifies a qualitative leap in neural network-based simulation by faithfully adhering to the physical laws. Code and video demonstrations are available at https://github.com/chenyu-xjtu/DualFluidNet.

Figures (11)

• Figure 1: Armadillo in the flowing water, with 138k particles.
• Figure 2: The architecture of our network. It consists of two pathways: the Main Path and the Constraint-Guided Path. These two pathways share a similar structure, with the Main Path utilizing CConv as its convolutional kernel (the orange rectangle), and the Constraint-Guided Path employing ASCC as its kernel (the blue diamond). At each layer, both pathways pass through a module for feature fusion before feeding the fused features into the next layer.
• Figure 3: Anti-symmetric kernel is obtained by negating and mirroring the trainable variables by the center point.
• Figure 4: The architecture of the Attention-based Feature Fusion Module. Features from both pathways pass through two layers of CConv and are then concatenated. The concatenated features are subsequently input into a Particle Selector to determine the respective weights for fusion. The original features from both pathways are then combined with weighted fusion.
• Figure 5: The architecture of the Type-aware Input Module in the Constraint-guided Path, with ASCC as the kernel. In contrast, the Main Path's Type-aware Input Module replaces the ASCC kernel with CConv. The Particle Selector is the same as Figure \ref{['fig:Particle Selector']}.