Abstract
Simulating incompressible, inviscid fluids remains a challenge in computer graphics due to numerical dissipation, which prematurely dampens turbulent structures and energy. We present CO-FLIP, a novel hybrid Eulerian-Lagrangian fluid simulation method that leverages geometric mechanics to preserve energy and circulation. By reformulating fluid motion as a Hamiltonian system evolving on coadjoint orbits, CO-FLIP maintains intricate vortical structures even at low grid resolutions where traditional simulators fail.
Introduction
The simulation of inviscid fluids, governed by the Euler equations, is essential for creating realistic smoke, fire, and water effects. However, standard discretization techniques often introduce "artificial viscosity"—a numerical artifact where energy leaks out of the system and vortices disappear. While hybrid methods like Fluid Implicit Particles (FLIP) mitigate this by using particles to carry momentum, they often suffer from instability, noise, and a failure to conserve the topological "twistiness" (circulation) of the flow.
CO-FLIP addresses these issues by bridging the gap between high-level geometric theory and practical numerical algorithms. The method ensures that the fluid state remains on its theoretical "coadjoint orbit," a subspace where circulation is mathematically guaranteed to be preserved.
Methodology
CO-FLIP introduces several key innovations to the standard simulation loop, focusing on structural preservation through Hamiltonian mechanics.
Velocity Covectors and Lie Advection
Unlike traditional methods that store velocity vectors on particles, CO-FLIP stores "velocity covectors" (also known as impulses). This distinction allows the method to employ Lie advection, a mathematically rigorous framework for moving momentum that naturally preserves circulation. This approach ensures that the underlying Poisson structure of the fluid—the rules governing how swirls move—remains intact regardless of grid resolution.
Mimetic B-Spline Interpolation
To prevent common artifacts like particle clumping, the authors utilize a family of mimetic B-spline interpolants. These are designed to be "divergence-consistent," meaning that if the grid data is divergence-free, the resulting continuous velocity field is pointwise divergence-free everywhere. This eliminates the need for the complex pressure projection hacks often required by other high-order methods.
Exact Pseudoinverse P2G
In standard FLIP, transferring data from particles to the grid (P2G) typically involves simple weighted averaging. CO-FLIP instead defines P2G as the exact mathematical pseudoinverse of the grid-to-particle (G2P) interpolation. This is solved efficiently using a Preconditioned Conjugate Gradient (PCG) solver, ensuring that the transfer process does not introduce additional dissipation.
Geometric Time Integration
The system is integrated using a Lie group trapezoidal rule. This geometric integrator, combined with a minimal energy-based correction step, ensures that kinetic energy remains constant. This compensates for minor errors introduced by finite particle density and keeps the simulation stable over long durations.
Results and Evaluation
CO-FLIP was evaluated against standard fluid dynamics benchmarks, demonstrating significant improvements in both accuracy and stability.
- Convergence: The method achieves third-order convergence in space when using quadratic B-splines, outperforming the first- or second-order accuracy typical of traditional simulators.
- Conservation: In 2D leapfrogging vortex tests—a classic measure of vorticity preservation—CO-FLIP maintained core structures for over 500 seconds. In contrast, competing methods like PolyPIC and PolyFLIP saw vortices dissipate or merge within 100 seconds.
- Stability: CO-FLIP is remarkably stable, requiring "global resets" of particle data only once every 200 frames, compared to every 20 frames in state-of-the-art neural flow maps.
- Visual Fidelity: The method produces high-resolution-style turbulent details on surprisingly coarse grids. For example, complex pyroclastic clouds and smoke plumes were successfully simulated on grids as low as $96 \times 192 \times 96$, capturing structures that would normally require much finer discretization.
Discussion and Limitations
While CO-FLIP offers superior preservation of fluid structures, it comes with a higher computational cost per time step. The requirement for a PCG solve during P2G transfer and the use of fixed-point iterations for implicit integration result in approximately 50% more overhead than an explicit midpoint solver.
Additionally, the current formulation is specialized for inviscid fluids. While ideal for gases and smoke, extending the framework to include controlled viscosity for liquids or handling free-surface flows (like splashing water) remains an area for future research. Specifically, managing the Galerkin Hodge star near boundaries in free-surface scenarios presents a significant theoretical challenge.
Conclusion
CO-FLIP demonstrates that by discretizing the Hamiltonian of a fluid system first, one can derive a structure-preserving integrator that excels at maintaining turbulent detail. By focusing on the geometric properties of the flow, the method allows for high-fidelity visual results at lower resolutions, providing a robust alternative for high-end fluid animation.
Summary updated