acceptodds
Under review as a conference paper at ICLR 2027

Neural-PD: A Peridynamics-Inspired Neural Simulator for 3D Fracture Dynamics

Abstract

Predicting fracture in deformable objects remains challenging for neural dynamics models because cracks introduce localized, irreversible discontinuities. Peridynamics (PD) is a well-established continuum physical simulation framework whose interaction operators share the same structure as deep networks on point clouds. Inspired by Peridynamics, we present Neural-PD, a particle-based neural simulator that uses the interaction structure of PD as a physical inductive bias for learning deformation, contact, and fracture. Different from other particle-based neural dynamics models, Neural-PD follows Peridynamics in combining material forces and contact forces, while utilizing each type of force on a different neighborhood: material forces over the reference configuration and contact forces over the current one. An attention-based formulation inspired by PD explicitly models the strength of each bond, which helps predicting fracture and damage. Unlike traditional peridynamics, Neural-PD adopts a U-Net architecture that propagates force information across spatial scales, supporting long-range dynamics predic- tion at large time steps. Due to the lack of benchmark for evaluating neural dy- namics models in damage and fracture scenarios, we introduce a new benchmark spanning visco-elastic deformations and multiple crack and damage scenarios. On these benchmarks, Neural-PD achieves 38% − 93% lower error than other neural dynamics models, showcasing the significant potential of incorporating particle-based physics ideas such as PD in neural dynamics modeling.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Learned physical simulation, 3D fracture dynamics, Peridynamics, Neural Simulators, Pointcloud dynamics

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_7rLxOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents Neural-PD, a particle-based neural simulator for 3D deformation and fracture. Inspired by peridynamics, the method uses fixed reference-configuration neighborhoods for intra-object material interactions and dynamically recomputed current-configuration neighborhoods for inter-object contact. A recurrent, non-increasing bond-connectivity variable gates material attention to represent irreversible damage, while a point-cloud U-Net propagates interaction information across spatial scales before predicting force-density-like quantities and integrating the state with symplectic Euler. The paper also introduces a 3,000-trajectory benchmark containing Ball-Plate, Mug-Drop, Bottle-Bullet, and Cloth-Fall scenarios. Neural-PD generally improves long-horizon position error over several learned simulators and runs faster than the reported numerical methods.

Critical strengths
  1. A promising inductive bias for fracture dynamics. Separating persistent material connectivity in the reference configuration from transient contact in the current configuration is intuitive and well matched to fracture. The monotonic connectivity gate is also a useful mechanism for representing irreversible loss of material bonds.

  2. A potentially valuable benchmark and encouraging initial results. Learned simulation benchmarks with contact-driven 3D fracture are limited, and the four proposed scenarios cover several useful deformation and damage modes. Neural-PD obtains strong position-rollout results on most scenarios, and the ablation suggests that both the reference-frame graph and connectivity gate contribute. With clearer and more complete validation, this could become a useful contribution.

Critical weaknesses
  1. The experimental protocol and quantitative claims are insufficiently documented. The paper does not clearly report the trajectory count per subset, particle counts, train/validation/test split, whether one model or separate models are trained for each scenario, the exact geometry and material variations, or whether the test conditions are in-distribution or out-of-distribution. Important model details such as neighborhood sizes, contact radius, U-Net hierarchy, hidden dimensions, and number of random seeds are also absent. The efficiency table omits hardware and learned-baseline runtimes. Moreover, Table 1 contradicts the claim that Neural-PD is best on every scenario: on Mug-Drop, SGNN reports 1.94×10−21.94\times10^{-2} versus 1.980×10−21.980\times10^{-2} for Neural-PD, yet Neural-PD is highlighted as best. These omissions and inconsistencies make the reported improvements difficult to assess reliably.

  2. The evaluation does not directly establish accurate fracture prediction, and the visual evidence is limited. The only quantitative accuracy measure is mean particle-position error, which can hide localized errors in crack onset, crack location, connectivity, or fragment structure. Although the model explicitly predicts irreversible bond connectivity and the generated data include bond damage, the paper provides no connectivity accuracy, damage-field error, fracture-onset error, crack-surface metric, or fragment-level evaluation. The connectivity variable is trained only indirectly through position loss, so it is unclear whether it represents physical bond failure or merely acts as an attention gate. The few qualitative examples are too small and sparse to assess the diversity and fidelity of fracture outcomes. In addition, the final MLP predicts pointwise force-density-like outputs without pairwise antisymmetry or conservation constraints, so the method should be described as peridynamics-inspired rather than as a physically consistent peridynamic force model.

Questions for authors
  1. Can the authors provide complete dataset splits, particle counts, geometry/material ranges, model hyperparameters, seed counts, and hardware/runtime settings, and clarify whether a single model is trained across all four scenarios? Which test conditions specifically measure out-of-distribution generalization?

  2. Can the authors directly evaluate fracture prediction using ground-truth bond connectivity or damage, fracture-onset time, crack geometry, and fragment-level metrics? How well does the learned connectivity correlate with the simulator's failed bonds?

  3. Could the authors correct the Mug-Drop comparison and provide larger, more numerous qualitative rollouts covering diverse test conditions and failure cases? Learned-baseline FPS and scaling with particle count would also substantially strengthen the efficiency claim.

Reviewer_M5EnOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a neural network-based learning model for modeling multi-body peridynamics. The proposed model uses a U-Net architecture to learn from a collection of 3d particles.

Critical strengths
  1. The paper is well-written, especially for an ML audience. The underlying concepts in peridynamics are clearly explained with sufficient examples.
  2. I believe multi-body peridynamics is useful in developing future applications.
Critical weaknesses

I am not familiar with this topic. So I was unable to find any critical weaknesses. But I am a little bit surprised that there aren't any existing works that model multi-body collisions and fractures. The related works section mentions two works that use neural operators that focus on single-body peridynamics. Through a quick Google search, I found a few other papers that handle peridynamics, but I understood those works to focus on the single-body setting as well.

Questions for authors
  1. The quantitative results in the table clearly show an improvement over the baselines. But the qualitative results for Neural-PD differ significantly from the row labeled "peridynamics." For example, in Figure 2, the top row clearly shows the crack after the ball hits the plates. But the Neural-PD row shows only blue spots (i.e., some damage) around the regions where the crack appeared to terminate in the first row. Even the ball's position at step 30 in the Neural-PD row differs from that of other rows.
Reviewer_sNH7OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Neural-PD is a learned particle simulator for 3D deformation and fracture. Its architecture is modeled on the peridynamics equation of motion. Each object is a point cloud. A network predicts force density at each point, and symplectic Euler integration turns this into the next state. Training uses a multi-step position Huber loss on peridynamics trajectories. Contributions:

  1. Separate neighborhoods: reference-frame neighbors for material forces and current-frame neighbors for contact.
  2. A learned bond connectivity \omega between 0 and 1. A min-update makes \omega non-increasing, so bonds cannot heal. \omega multiplies the attention weights after the softmax, so broken bonds contribute nothing. Updates to ω start only after an object first makes contact.
  3. A benchmark of 3,000 peridynamics trajectories (383,000 frames) in four subsets. Three subsets involve fracture: Ball-Plate, Mug-Drop, and Bottle-Bullet. Cloth-Fall is viscoelastic with no fracture. The biggest strength of
Critical strengths

The biggest strength of this paper is the design of the irreversible, learned bond-connectivity gate in the material attention. It addresses a real limitation of existing learned simulators. A neighborhood rebuilt from current positions cannot tell a closed crack from intact material. A fixed mesh or a fixed object partition cannot form new fracture surfaces. Neural-PD keeps the reference-frame bond list and attaches a learned connectivity \omega to each bond. \omega can only decrease, so a crack stays open in the model’s interaction structure even when the two crack faces touch again. This is a direct and physically motivated transfer of the peridynamics bond-failure idea into attention.

Critical weaknesses

There is no metric for damage or crack accuracy. There is no supervision of \omega and no check that \omega matches peridynamics damage. There is no experiment on held-out geometries or loading cases, although the introduction claims generalization. There is no speed comparison with the learned baselines. The network time step is not reported relative to the solver time step. The ablation covers one dataset.

Questions for authors
  1. How does the connectivity gate act at the coarser U-Net levels?
  2. How is damage computed for Neural-PD, and does it match the reference?
  3. Why does the gate activate only after first contact, and what happens without this rule?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.