The affine particle-in-cell method

Chenfanfu Jiang, Craig A. Schroeder, Andrew Selle, Joseph Teran, Alexey StomakhinACM Trans. Graphics 2015 · DOI

TL;DR

This paper introduces the Affine Particle-In-Cell (APIC) method, a novel simulation technique for fluids and granular materials that significantly improves upon existing Particle-In-Cell (PIC) and Fluid Implicit Particle (FLIP) methods. APIC achieves greater stability and reduces dissipation by representing particle velocities with locally affine transformations, leading to more realistic and energetic simulations with better conservation of angular momentum.

What Problem Does This Paper Address?

Computer graphics simulations of natural phenomena like fluids and granular materials often rely on hybrid Lagrangian-Eulerian methods. These methods use particles to track material properties and topology changes, while an Eulerian grid handles force calculations and collision responses. Traditional PIC methods are stable but highly dissipative, leading to unrealistic damping of motion. FLIP methods reduce dissipation but introduce noise and instability. Blending PIC and FLIP offers a compromise but can still suffer from dissipation or require manual tuning. The core problem is finding a method that is both stable and non-dissipative, accurately conserving physical quantities like angular momentum, without the drawbacks of existing approaches.

Methods

The paper introduces the Affine Particle-In-Cell (APIC) method. The key innovation lies in how particle velocities are represented and transferred between the particle and grid representations.

  • Standard PIC: Transfers a single, constant velocity value from the grid to each particle. This filtering process is stable but introduces significant dissipation.
  • FLIP: Transfers velocity increments from the grid to particles, reducing dissipation but potentially leading to instability and noise.
  • APIC: Augments each particle with a locally affine description of its velocity (represented by a 3x3 matrix). This allows for a more accurate representation of the velocity field at the particle's location, capturing not just translation but also shearing and rotation.
  • Transfer Mechanism: APIC uses a novel particle-to-grid transfer that preserves these affine velocity fields. This transfer is designed to be stable and to conserve angular momentum more accurately than previous methods. The paper also discusses a Rigid Particle-In-Cell (RPIC) method as an intermediate step, which conserves angular momentum but not shearing.

Datasets

Not discussed in the paper. The paper uses synthetic examples and simulations of various phenomena (wine pour, granular materials, fountain, lava flow, collision scenarios) to demonstrate the method's capabilities.

Key Insights and Ideas

  • Dissipation stems from information loss: The paper argues that the dissipation in PIC arises from forcing a complex velocity field onto a single constant value per particle during grid-to-particle transfer.
  • Affine representation is key: By representing particle velocities as locally affine transformations, APIC captures more of the underlying velocity field, leading to reduced dissipation and better conservation of physical quantities.
  • Angular momentum conservation is crucial for realism: The paper highlights how poor angular momentum conservation leads to artifacts like objects unnaturally stopping their rotation, which APIC significantly mitigates.
  • APIC offers a "best of both worlds" scenario: It combines the stability of PIC with the non-dissipative nature and energetic behavior of FLIP, without introducing the noise and instability of FLIP.
  • Generalizability: The APIC method is shown to be applicable to both incompressible fluid simulations and Material Point Method (MPM) simulations for solids and granular materials.

Key Results

  • Improved Stability and Reduced Dissipation: APIC simulations exhibit significantly less dissipation than PIC and are more stable than pure FLIP, as demonstrated in various examples like the wine pour, granular materials, and lava flows.
  • Accurate Angular Momentum Conservation: APIC shows superior conservation of angular momentum compared to PIC and FLIP, leading to more realistic rotational behavior in simulations (e.g., in collision scenarios).
  • Energetic and Lively Simulations: APIC produces more dynamic and energetic simulations, with more detailed splashing and flow patterns, without sacrificing stability.
  • Comparable Performance: While APIC introduces some overhead for storing and processing the affine matrices, its performance is comparable to or slightly slower than FLIP in many cases, and significantly faster than unstable FLIP simulations. The memory overhead is also manageable.

Limitations and Future Work

  • Ringing Instability: The paper acknowledges that APIC does not improve or exacerbate the "ringing instability" (a phenomenon where particle velocities can become large but are zero when transferred to the grid). This is a limitation inherited from the underlying PIC/FLIP framework.
  • Extra Storage and Computation: APIC requires storing an additional matrix per particle and performing more complex calculations during transfers, which adds some computational and memory overhead.
  • Desirable Artifacts: The authors note that while APIC removes "undesirable" artifacts like dissipation, some artifacts (like those introduced by FLIP) can be visually desirable for certain effects (e.g., smoke solvers). They suggest that artists would prefer to control these effects deliberately rather than have them imposed by the simulation method.
  • Future Work: The paper suggests that the affine representation could be extended to higher-order representations for even greater accuracy. The method's applicability to other simulation domains beyond fluids and granular materials could also be explored.

Summary updated

BibTeX
@article{jiang2015affine,
  author = {Chenfanfu Jiang and Craig A. Schroeder and Andrew Selle and Joseph Teran and Alexey Stomakhin},
  title = {{The affine particle-in-cell method}},
  journal = {ACM Trans. Graphics},
  year = {2015},
  doi = {10.1145/2766996},
}
Figure