enriching and accelerating the material point...
TRANSCRIPT
Introduction Rigid Body with Controllable Friction
Acceleration: AsyncMPM and IMEX-MPM
Improving Particle Distribution
s
The Material Point Method (MPM) is a numerical method capable of simulating
the behavior of liquids, deformable solids, and granular materials. In this work, we
propose novel techniques to make it even more versatile and efficient:
1. A surface tension model to MPM liquid.
2. Improved particle distribution using an anisotropic position correction model.
3. Rigid body simulation in MPM, where friction is modeled as plasticity.
4. Two new ways to speed up MPM: Asynchronization, and implicit-explicit mixed
time integration.
5. Investigation on divergence-free particle-grid transfer on APIC and PolyPIC.
Figure 3: A 3D water surface scene simulated using the CSS model. The total particle number is
88374. (Simulated with JIXIE, rendered by Houdini.)
Enriching and accelerating the Material Point Method Yuanming Hu1 Yu Fang1 Ziheng Ge2 Ziyin Qu3 Andre Pradahana3 Chenfanfu Jiang3
1Tsinghua University 2University of Science and Technology of China 3University of Pennsylvania
The Continuum Surface Force Model (CSF):
The Continuum Surface Stress Model (CSS):
Normal vector: 𝑛 =𝛻𝐶
𝛻𝐶
Curvature: 𝜅 = 𝛻 ∙ 𝑛
Surface force: 𝑓 = 𝛾𝜅𝑛
Calculate
on grid
Surface stress: 𝜎 = 𝛾(1
𝑑𝛻𝐶 𝐼𝑛 −
(𝛻𝐶)𝑇𝛻𝐶
𝛻𝐶)
The Smoothed Particle Hydrodynamics Model (SPH):
Cohesion force:
𝑓 𝑖𝑗 = −𝛾𝑚𝑖𝑚𝑗𝐶( 𝑥𝑖 − 𝑥𝑗 )𝑥𝑖−𝑥𝑗
𝑥𝑖−𝑥𝑗
There are some choices of kernel C(r),
and we use the one shown in blue plot.
𝛻𝐶 𝑝 = Σ𝑖𝛻𝑤𝑖𝑝𝑚𝑖
𝛻𝐶 𝑖 = Σ𝑝𝛻𝑤𝑖𝑝𝑚𝑝
𝜅𝑖 = Σ𝑗 𝛻𝑤𝑗𝑖𝑇𝑛𝑗
Calculate on particles
Figure 2: A 2D dam break simulation.
Figure 1: An overview of the new Material Point Method. Our contributions are highlighted in red.
AsyncMPM: a novel asynchronous time integration scheme for MPM which offers
temporal adaptivity when objects of different stiffness or velocity coexist. We show
via multiple test scenes that our asynchronous MPM leads to 6x speed up in Table 1.
IMEX-MPM: We propose an effective approach to use implicit and explicit time
integragtion scheme in different regions of MPM. The regions are automatically
detected based on material stiffness.
Simulating rigid bodies in MPM is an open problem. To this end, we Soft
Rigidification, which preserves energy and coupled naturally with other parts of
MPM. In short, this is achieved by penalizing non-rigid position of rigid body
particles before particle position update.
In order to model rigid surface friction, we use a tailored representation of
deformation for rigid bodies (left), as well as a plasticity model for objects with
different surface friction, by changing the amount of shearing allowed (right).
Constitutive Model: Constitutive model relating material stress to material state
(such as deformation gradient F) is needed for governing the material responses
under deformations. Commonly used hyper-elastic constitutive models include Neo-
Hookean Model and Fixed Corotated Model.
For liquid simulation, J-based constitutive model is used, where 𝐽 is the determinant
of deformation gradient 𝐹. Water is modeled to be weakly incompressible with
partial stress
𝜎𝑤 = −𝑃𝑤𝐼, 𝑃𝑤 = 𝑘1
𝐽𝑤𝛾− 1
The pressure 𝑃𝑤 is designed to stiffly penalize liquid volume change, which is
characterized in terms of the determinant of the liquid deformation gradient
𝐽𝑤 = det (𝐹𝑤). Here 𝑘 is the bulk modulus of water, and 𝛾 is a term that more stiffly
penalizes large deviations from incompressibility.
Weakly Compressible Liquids with Surface Tension
Surface tension arises as a result of attractive forces between molecules in a fluid.
As a force acting only on the surface, it balances radially inward inter-molecular
attractive force with radially outward pressure gradient force across the surface. Figure 4: Comparison of dam break without position correction(left) and with position
correction(right).
An inherent problem of MPM is uneven spatial particle distributions during
simulation, which in turn can cause “holes” due to missed fluid cells and lead to
artifacts such as bumpy surfaces. To prevent this, we use anisotropic position
correction to improve particle distribution. The anisotropic displacement vector ∆𝑃
is computed by
∆𝑃𝑖= −∆𝑡𝛾𝑠𝑑∑𝑟
𝑟𝑊𝑠𝑚𝑜𝑜𝑡ℎ(𝑟, 𝑑)
𝑟 =1
2𝑘𝑠𝐶𝑖
−1 + 𝐶𝑗−1 𝑝𝑗 − 𝑝𝑖 ,
After we obtain ∆𝑃, we recompute our new velocity by
𝑢 𝑥 =∑𝑖𝑚𝑖𝑢𝑖𝑊𝑠ℎ𝑎𝑟𝑝(𝑝𝑖 − 𝑥, 𝛼𝑢𝑑𝑖)
∑𝑖𝑚𝑖𝑊𝑠ℎ𝑎𝑟𝑝(𝑝𝑖 − 𝑥, 𝛼𝑢𝑑𝑖)
Divergence-Free Particle-Grid Transfer
The particle-grid transfer in MPM usually leads to velocity mode loss, since there are more
degrees of freedom (DoFs) on the grid than those on the particles. Such loss is commonly
considered to be harmful and dissipative. However, thinking positively, if we intentionally
drop the divergence part of the velocity field, such “loss” can actually be considered as
pressure projection, making fluid more incompressible.
We tried this idea with APIC while it turns out APIC itself is too dissipative, making the
fluid viscous. Therefore, we investigate PolyPIC, which is indeed lossless and thereby
promising for future research on divergence-free transfer.
Initial PPIC 20 iter. APIC 20 iter. Initial PPIC 20 iter. APIC 20 iter.
Platforms with
increasing
friction from
left to right
We focus on the following three models for surface tension: (𝛾 is a constant
coefficient for surface tension strength.)