Research map / 3D vision / 3D generation
DEL: Discrete Element Learner for Learning 3D Particle Dynamics with Neural Rendering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DEL, which embeds learnable graph-network kernels into the classic Discrete Element Analysis framework to learn 3D particle dynamics purely from multi-view 2D images via inverse neural rendering with the GPF point renderer.
Problem
Learning 3D particle dynamics from 2D image observations without per-particle 3D correspondences is ill-posed because different 3D particle distributions can yield similar 2D images; prior GNN simulators need 3D particle tracks and black-box models generalize poorly.
Approach
Use graph networks only as learnable kernels that regress magnitudes of specific mechanical operators (normal/tangential contact, bond, and dissipation forces) inside a DEA framework whose force directions are precomputed, rather than learning the whole dynamics mapping; trained end-to-end through a differentiable point renderer with normal-force and gradient-direction losses.
Key results
- In particle-view evaluation, DEL reaches CD 7.54 and EMD 7.10 on Plasticine versus 16.20/14.61 for the best prior baseline (EGNN*), and CD 8.48 on Multi-Objs versus 13.21 for EGNN*.
- In render-view evaluation on Plasticine, DEL achieves PSNR 28.09 and LPIPS 0.091, beating VPD (27.06, 0.101).
- Ablation on SandFall shows removing the gradient loss raises CD from 1.73 to 3.95 and removing the tangential force term raises it to 2.26.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DEL instead of VPD | particle dynamics learned from 2D images · Plasticine | CD: 16.96 -> 7.54 | Helped |
| use DEL instead of the strongest baseline | particle dynamics learned from 2D images · SandFall | CD: 1.99 -> 1.73 | Helped |
| use DEL instead of the strongest baseline | particle dynamics learned from 2D images · Multi-Objs | CD: 13.21 -> 8.48 | Helped |
| use DEL instead of the strongest baseline | particle dynamics learned from 2D images · FluidR | CD: 2.58 -> 1.72 | Helped |
| use DEL instead of VPD | particle dynamics learned from 2D images · Bear | CD: 3.41 -> 3.54 | Hurt |
| remove gradient loss L_g (ablation) | DEL · SandFall | CD: 1.73 -> 3.95 | Hurt |
| remove gradient loss L_g (ablation) | DEL · Multi-Objs | CD: 8.48 -> 29.7 | Hurt |
| remove tangential force component f_ij^t (ablation) | DEL · SandFall | CD: 1.73 -> 2.26 | Hurt |
| remove tangential force component f_ij^t (ablation) | DEL · Multi-Objs | CD: 8.48 -> 11.4 | Hurt |
| replace DEA force-direction priors with predicted force directions and magnitudes (ablation) | DEL · SandFall | CD: 1.73 -> 3.15 | Hurt |
| replace DEA force-direction priors with predicted force directions and magnitudes (ablation) | DEL · Multi-Objs | CD: 8.48 -> 16.5 | Hurt |
| remove normal-force rendering loss L_n^r (ablation) | DEL · SandFall | CD: 1.73 -> 2.65 | Hurt |
Surprising findings
- Given 3D ground-truth labels, SGNN performs roughly on par with DEL, while under 2D supervision SGNN degrades sharply, attributed to 2D-to-3D ambiguity.
- The method is reported to underperform SGNN on sand and viscous liquid materials, which the authors attribute to bond forces constraining particles of the same material.
Limitations
- All experiments are on synthetic MPM-generated datasets; real multiview dynamic videos were not evaluated.
- Relies on bond forces linking particles of the same object, which may hurt simulation of particulate matter like sand and liquids.
Experimental setup
Datasets: Plasticine Multi-Objs SandFall FluidR Bear Fluids
Benchmarks: Plasticine SandFall Multi-Objs FluidR Bear Fluids
Metrics: Chamfer distance Earth Mover's distance PSNR SSIM LPIPS
Baselines: SGNN* EGNN* VPD 3DIntphys NeRF-dy Neurofluid
Builds on: Discrete Element Analysis Generalizable Point Field EGNN SGNN Neurofluid
Closest papers by meaning
- Learning 3D Particle-based Simulators from RGB-D Videos (ICLR 2024 · similarity 0.90)
- NeuMA: Neural Material Adaptor for Visual Grounding of Intrinsic Dynamics (NeurIPS 2024 · similarity 0.88)
- Latent Intuitive Physics: Learning to Transfer Hidden Physics from A 3D Video (ICLR 2024 · similarity 0.86)
- DreamPhysics: Learning Physics-Based 3D Dynamics with Video Diffusion Priors (AAAI 2025 · similarity 0.86)
- MaNGO — Adaptable Graph Network Simulators via Meta-Learning (NeurIPS 2025 · similarity 0.85)
- Diffusion-Based Hierarchical Graph Neural Networks for Simulating Nonlinear Solid Mechanics (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.