Research map / 3D vision / Face and shape
NeuralClothSim: Neural Deformation Fields Meet the Thin Shell Theory
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 NeuralClothSim, a quasistatic cloth simulator in which equilibrium deformation is a neural deformation field (SIREN MLP) trained by minimizing the non-linear Kirchhoff-Love thin-shell potential energy.
Problem
Mesh- and point-based cloth simulators have fixed resolution, large memory cost, resolution-dependent folds, and are hard to differentiate and integrate into neural architectures.
Approach
Encode cloth equilibrium as a coordinate-based neural deformation field whose weights are optimized directly on the non-linear Kirchhoff-Love shell energy, using periodic sine activations for higher-order derivatives, hard Dirichlet constraints via a distance-based multiplier, and input-space periodic mapping for seams.
Key results
- On the Scordelis-Lo roof, the full model gives 0.3018 max vertical displacement versus analytical 0.3024 and Bastek et al. 0.297.
- Pinched rigid-end cylinder: 1.81e-5 for the full model versus 1.825e-5 analytical; removing periodicity drops it to 3.6e-9.
- Memory of the NDF weights stays constant with spatial resolution, while FEM mesh memory grows linearly with vertex count.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use sine activations instead of GELU (ablation) | NeuralClothSim · cloth simulations requiring folds and wrinkles | high-frequency deformation representation | Helped |
| Set SIREN frequency parameter to 1 instead of 30 (ablation) | NeuralClothSim · cloth simulations requiring folds | fold formation | Hurt |
Surprising findings
- Removing periodic boundary conditions causes large errors on the free-end cylinder (3.13e-6 vs 4.52e-4 analytical), and GELU activation fails to capture high-frequency folds.
Limitations
- No collision or contact handling and no friction.
- Quasistatic only; no inertia or damping in the current formulation.
- Linear isotropic and nonlinear orthotropic material models only; material conditioning limited to low-dimensional parameters.
Experimental setup
Datasets: Belytschko obstacle course analytical solutions
Benchmarks: Belytschko obstacle course (square plate, Scordelis-Lo roof, pinched cylinders)
Metrics: maximum displacement vs analytical solution memory vs resolution discretisation consistency
Baselines: DiffARCSim DiffCloth Bastek et al. Guo et al. ARCSim
Builds on: SIREN Kirchhoff-Love shell theory Clyde et al. nonlinear orthotropic material model Bastek et al. PINN for shells
Closest papers by meaning
- Thin-Shell-SfT: Fine-Grained Monocular Non-rigid 3D Surface Tracking with Neural Deformation Fields (CVPR 2025 · similarity 0.88)
- Learning 3D Garment Animation from Trajectories of A Piece of Cloth (NeurIPS 2024 · similarity 0.86)
- Physics-guided Shape-from-Template: Monocular Video Perception through Neural Surrogate Models (CVPR 2024 · similarity 0.86)
- Implicit-ARAP: Efficient Handle-Guided Neural Field Deformation via Local Patch Meshing (NeurIPS 2025 · similarity 0.85)
- Bayesian Differentiable Physics for Cloth Digitalization (CVPR 2024 · similarity 0.84)
- UniPhy: Learning a Unified Constitutive Model for Inverse Physics Simulation (CVPR 2025 · similarity 0.84)
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.