FractureFields: Contact-Aware Binary Multi-Field Transfer for Fractured 3D Gaussian Simulation

Jianchen Wang; Runyang Qu; Fei Li

cs.CVAug 29, 2026

FractureFields: Contact-Aware Binary Multi-Field Transfer for Fractured 3D Gaussian Simulation

Authors: Jianchen Wang, Runyang Qu, Fei Li

Organizations: School of Computer Science and Technology, Xidian University

Abstract

Physics-integrated 3D Gaussian representations make it possible to simulate image-reconstructed assets directly as particles, but current Gaussia-MPM pipelines keep a single Eulerian velocity field even after fracture. When disconnected fragments share interpolation support, they still write to and read from the same grid nodes, producing cross-fragment momentum leakage that appears as residual adhesion and non-physical stretching. We present FractureFields, a topology-adaptive transfer for fractured 3D Gaussian objects. After a structural event assigns persistent fragment identities, FractureFields builds fragment-specific mass and momentum fields in a single P2G pass, advances each field independently, and performs a field-aware G2P update so particles only sample their own fragment's grid state. To handle re-contact, we add a momentum-conserving contact projection that applies equal and opposite normal impulses only when two fragment fields are approaching, preserving free separation otherwise. Experiments on reconstructed scenes and a controlled re-contact benchmark show that fragment-conditioned routing eliminates realized cross-fragment mixing by construction, while contact projection reduces interpenetration during collision without reintroducing residual coupling. Overall, we argue that post-fracture simulation should treat structural disconnection as a change in local dynamical state, not merely a change in constitutive stress.

Figures & tables

Figure 1: Qualitative comparison on the Ficus scene. PhysGaussian and our method are evaluated at matched time steps. After fracture, the shared-field baseline exhibits residual fragment adhesion, whereas our fragment-conditioned transfer enables persistent separation. The circular insets magnify the fracture region.

Figure 2: Overview of the proposed fracture-aware simulation pipeline. (a) A stress-guided event detector localizes the fracture interface and temporally latches persistent topology. (b) Fragment identities route particles into separate local mass and momentum fields during a single P2G pass. (c) The fields evolve independently during separation and activate contact handling only when approaching, after which field-aware G2P returns each particle from its corresponding field.

Figure 3: Contact-aware local multi-field update. (a) Persistent particle identities divide the fractured object into two fields across the latched interface. (b) Fragment-aware P2G constructs separate nodal velocities. Separating fields receive no impulse, while approaching fields receive equal and opposite normal impulses that preserve node-level linear momentum. (c) Field-aware G2P reads the field associated with each particle identity, preserving the post-fracture velocity discontinuity.

Figure 4: Qualitative comparison across reconstructed datasets. Rows correspond to different reconstructed scenes, while columns compare the shared-field baseline, Stress-Zero, and our full method at matched post-fracture time steps. Our fragment-conditioned transfer produces cleaner and persistent separation with less residual adhesion and fracture-band stringing.

Scene	Transfer	$E_{\mathrm{mix}}^{\mathrm{cf}}$ mean/max	$E_{\mathrm{mix}}^{\mathrm{real}}$ mean/max	Shared (%)
Ficus	Shared (Stress-Zero)	$5.156{\times}10^{-7}/7.637{\times}10^{-6}$	$5.156{\times}10^{-7}/7.637{\times}10^{-6}$	2.280
Ficus	Fragment-conditioned (Full)	$5.436{\times}10^{-5}/7.099{\times}10^{-4}$	$0/0$	2.520
Bread	Shared (Stress-Zero)	$1.637{\times}10^{-6}/6.721{\times}10^{-6}$	$1.637{\times}10^{-6}/6.721{\times}10^{-6}$	96.580
Bread	Fragment-conditioned (Full)	$3.424{\times}10^{-3}/5.346{\times}10^{-3}$	$0/0$	96.420

Table 1: Post-fracture transfer diagnostics. Mixing entries are post-latch mean/maximum over frames 23–43 for Ficus and 10–89 for Bread. For Full, realized mixing is a structural zero established by the P2G/G2P routing invariant. Mass residuals are zero at the exported precision.

Branch	Passes (P2G/G2P/frag)	Median (ms)	Relative	Grid-buffer footprint (MiB)
Shared Field	1/1/0	0.589	1.000	56.000
Global Fragment Pass	2/2/1	0.713	1.211	56.000
Local Multi-Field	1/1/0	0.707	1.202	112.000
FractureFields	1/1/0	0.726	1.232	112.000

Table 2: Transfer efficiency. Runtime is normalized to Shared Field and excludes rendering, probing, export, and diagnostic shadow fields. The memory column is the branch-specific grid-buffer footprint.

Scene	Method	Opening/ $\Delta x$	Reapp.	Stringing	Sep.
Ficus	Baseline	$-0.018$	0.000	0.060	No
Ficus	Probe+Latch	$-0.018$	0.000	0.060	No
Ficus	Stress-Zero	0.083	0.000	0.208	No
Ficus	Full	7.788	0.018	0.069	Yes
Bread	Baseline	$-0.423$	0.053	0.276	No
Bread	Probe+Latch	$-1.750$	0.086	0.258	No

Table 3: Real-scene component ablation. Opening is measured over frames 23–43 for Ficus and 10–89 for Bread, and Sep. uses the common

0.1\Delta x

opening threshold. Only Full combines stable topology, fracture-band stress release, and fragment-conditioned transfer.

Panel A: Ficus static-plane interaction (gravity-enabled, fragment–plane contact)
Method	$t_{c}^{0}/t_{c}^{1}$	Pen./ $\Delta x$	max $\|v_{t}\|$ near plane	global residual †	max post-contact $v_{n}$	Task Success
Stress-Zero	$0.181/0.470$	$0.000$	$0.799$	$1.173$	$3.495$	Yes
Full	$0.276/0.297$	$0.026$	$0.325$	$1.139$	$3.015$	No
Panel B: Controlled fragment re-contact (gravity-disabled, fragment–fragment contact)
Method	first-contact $t$	Pen./ $\Delta x$	max $\|\Delta v_{t}\|$	CP mom. residual	$v_{\mathrm{rel},n}^{+}$ at contact	Task Success
Multi-field, no contact	$0.095$	$2.833$	$0.000$	$0.000$	$0.600$	No

Table 4: Post-fracture interaction (dual-panel). Panel A tests fragment–scene response under gravity. † The global residual is gravity-compensated but includes the external plane contact impulse; it is a scene-level diagnostic, not a pure internal momentum-conservation error. Panel B tests fragment–fragment re-contact without gravity. Task Success requires

\max(\mathrm{penetration})/\Delta x<0.1

and no identity crossing. Panel-specific diagnostics must not be compared across panels.

Scene	Init.	$E$	$\rho$	$\sigma_{R}$	Latch	Opening/ $\Delta x$	Sep.	Valid
Ficus	VLM	$2.0{\times}10^{6}$	200	$5.5{\times}10^{4}$	24	$-2.045$	No	Yes
Ficus	Manual	$2.0{\times}10^{6}$	200	$5.0{\times}10^{4}$	23	$-2.101$	No	Yes
Bread	VLM	$2.5{\times}10^{3}$	200	55	9	78.432	Yes	Yes
Bread	Manual	$2.0{\times}10^{3}$	200	50	10	78.180	Yes	Yes

Table 5: Material-prior initialization. The VLM values are obtained from GaussianProperty outputs and anchor-guided simulator mapping; they are initial conditions rather than recovered real-world SI parameters. Opening/

\Delta x

is a signed, run-level proxy used only for within-scene VLM/manual comparison.

Explore similar work

Scene-Level Heterogeneous Physics Simulation with 3D Gaussian Splats
Jun 19, 2026Xiaoyang Liu, Shangzhe Wu, Kai Han3D Gaussian Splatting 3D Scene Representation
1The University of Hong Kong · University of Cambridge
VersaGauss: A Versatile Framework for Generating Multiphase Dynamics with 3D Gaussians
Aug 28, 2026Ruijie Su, Lingxiao Yang, Xiaohua Xie +13D Scene Generation Physics-Based Simulation
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · School of Systems Science and Engineering, Sun Yat-sen University, Guangzhou, China · Guangdong Province Key Laboratory of Information Security Technology, Sun Yat-sen University, Guangzhou, China
PersistGS: Differentiable Physics for Object Permanence in 4D Gaussian Splatting
Jun 2, 2026Adrian Ramlal, John S. Zelek3D Gaussian Splatting 4D Reconstruction
University of Waterloo

Jun 19, 2026J/K move · Enter open · S save

Scene-Level Heterogeneous Physics Simulation with 3D Gaussian Splats

Xiaoyang Liu, Shangzhe Wu, Kai Han

1The University of Hong Kong · University of Cambridge

3D Gaussian Splatting 3D Scene Representation Physics-Based Simulation Deformable Object Simulation

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) has achieved state-of-the-art photorealistic rendering, but the representation gap prevents these assets from being physically interactive. Production-grade physics engines do not understand the 3DGS representation, while prior physics-for-3DGS methods are monolithic silos. These prior works are fundamentally limited, demonstrating only object-centric physics in isolated environments, such as on an ideal plane. They are incapable of interacting with complex static collision geometry or heterogeneous assets. We propose a novel framework that, for the first time, bridges this gap by enabling 3DGS assets to participate in scene-level, heterogeneous, multi-solver physical simulations. Our core contribution is a Representation Abstraction Framework that translates all diverse assets, including 3DGS, virtual meshes, and fluids, into a unified physical particle set. This abstraction is key to enabling complex behaviors, such as the non-rigid deformation of 3DGS assets, within a unified physics pipeline. This particle set, along with the static scene collision boundaries derived from scene capture, is processed within a solver-agnostic physics kernel. The physical results are then mapped back to drive each asset's specific visual reconstruction. This architecture unlocks capabilities impossible with prior art. We demonstrate complex, two-way interactions between deformable 3DGS assets, standard CG assets such as fluids and meshes, and large-scale captured static environments, showcasing realistic coupled phenomena that were previously unattainable.

Abstract

Explore similar work

Scene-Level Heterogeneous Physics Simulation with 3D Gaussian Splats

VersaGauss: A Versatile Framework for Generating Multiphase Dynamics with 3D Gaussians

PersistGS: Differentiable Physics for Object Permanence in 4D Gaussian Splatting