FluidFormer : Transformer with continuous convolution for particle-based fluid simulation.

Area of Science:

• Computer Graphics
• Computational Physics
• Artificial Intelligence

Background:

• Traditional fluid simulation methods like Navier-Stokes solvers are computationally expensive.
• Existing learning-based fluid simulation methods, often based on Smoothed Particle Hydrodynamics (SPH), suffer from instability due to reliance on local particle interactions and error accumulation.

Purpose of the Study:

• To introduce FluidFormer, a novel neural network architecture for efficient and stable fluid simulation.
• To address the instability issues in current learning-based fluid simulators by incorporating a hierarchical local-global modeling paradigm.

Main Methods:

• Developed FluidFormer, a novel architecture featuring a Fluid Attention Block (FAB).
• FAB combines continuous convolution for local interactions and self-attention for global hydrodynamic phenomena.
• Employed a dual-pipeline network to integrate physical biases with global reasoning.

Main Results:

• FluidFormer achieves state-of-the-art performance in fluid simulation tasks.
• Demonstrated significantly improved stability and generalization capabilities in complex fluid scenes.
• Validated the effectiveness of the hierarchical local-global modeling paradigm.

Conclusions:

• FluidFormer offers a robust and efficient solution for complex fluid simulation.
• The proposed architecture overcomes limitations of existing neural methods by addressing error accumulation through global reasoning.
• FluidFormer shows promise as a powerful tool for simulating complex physical systems.