DualFluidNet: An attention-based dual-pipeline network for fluid simulation.

Area of Science:

• Computational fluid dynamics
• Machine learning for physics simulations
• Deep learning for scientific computing

Background:

• Traditional numerical methods for fluid simulation are computationally expensive.
• Machine learning offers potential for faster, near-accurate physics simulations.
• Existing neural network approaches struggle to balance global control and physical constraints.

Purpose of the Study:

• To develop an efficient and accurate 3D fluid simulation method using deep learning.
• To address the trade-off between global fluid dynamics control and adherence to physical laws.
• To improve the handling of fluid-solid interactions in simulations.

Main Methods:

• Utilized an Attention-based Dual-pipeline Network architecture.
• Integrated an Attention-based Feature Fusion Module for enhanced processing.
• Developed a Type-aware Input Module for adaptive particle recognition and feature fusion.
• Introduced the Tank3D dataset for complex scene simulations.

Main Results:

• Achieved quantitative improvements across various metrics, surpassing state-of-the-art methods.
• Demonstrated a qualitative leap in neural network-based fluid simulation accuracy.
• Successfully balanced global fluid control with strict adherence to physical laws.
• Showcased improved handling of fluid-solid coupling through the Type-aware Input Module.

Conclusions:

• The proposed Attention-based Dual-pipeline Network offers a superior approach to 3D fluid simulations.
• This method achieves a better balance between simulation efficiency, accuracy, and physical fidelity.
• The Type-aware Input Module effectively addresses fluid-solid coupling challenges.
• The Tank3D dataset facilitates further research into complex fluid dynamics simulations.