Learning Simulatable Models of Cloth with Spatially-varying Constitutive Properties

Guanxiong Chen, Shashwat Suri, Yuhao Wu, Etienne Voulga, David I. W. Levin, Dinesh K. PaiarXiv 2025 · DOI · arXiv

Introduction

Simulating cloth is a fundamental challenge in computer graphics, typically requiring a trade-off between computational speed, physical accuracy, and model complexity. Traditional Finite Element Methods (FEM) provide high accuracy but are computationally intensive and suffer from "membrane locking"—a numerical artifact where coarse meshes exhibit artificial stiffness. Conversely, neural-network-based approaches like Graph Neural Networks (GNNs) or Neural ODEs offer speed but often lack long-term stability and require massive datasets.

Furthermore, real-world textiles are rarely uniform. Variations in stitching, padding, or material composition result in spatially-varying constitutive properties. Most existing models fail to capture this heterogeneity, assuming uniform stiffness across the entire garment.

The SpringTime Framework

The authors propose SpringTime, a framework that learns efficient mass-spring surrogate models directly from motion data. Unlike black-box neural networks, SpringTime utilizes a physically-grounded mass-spring system where stiffness and damping parameters are learnable for every individual spring.

SpringTime captures spatially-varying material properties, allowing for the simulation of heterogeneous textiles.

Core Methodologies

  1. Force-and-Impulse Loss: To overcome the noise inherent in position-based losses, SpringTime employs a novel loss function inspired by PID controllers. It combines a force term to penalize instantaneous errors and an impulse term to prevent error accumulation over time.
  2. Dual-Pass Training: The model uses a curriculum-based approach to disentangle elasticity from friction. It first learns stiffness from low-velocity data (where elastic forces are dominant) and subsequently learns damping parameters.
  3. Resolution Independence: SpringTime can learn from high-resolution source data (e.g., dense FEM simulations) and map this behavior onto coarser surrogate meshes via barycentric interpolation, effectively decoupling the simulation resolution from the data resolution.

Key Insights: Stability and Artifact Mitigation

A significant finding is SpringTime's ability to mitigate numerical artifacts. When trained on low-resolution FEM data plagued by membrane locking, the SpringTime surrogate effectively "filters" these errors, producing results that align more closely with high-resolution FEM ground truths.

Moreover, because the underlying architecture is a physical system, SpringTime maintains indefinite stability. In contrast, purely neural approaches often suffer from structural disintegration during long-duration rollouts.

Comparison of simulation stability: SpringTime maintains structural integrity while non-physical baselines often disintegrate over time.

Results

SpringTime demonstrates superior performance in both efficiency and accuracy:

  • Efficiency: The model converges in approximately 120 minutes, whereas MeshGraphNet (MGN) requires 48 hours to reach similar performance levels.
  • Accuracy: In 8-second rollouts, SpringTime achieved a Motion RMSE of 0.64, vastly outperforming Neural ODEs (RMSE 6.34), which failed to preserve the cloth's structure.
  • Generalization: The model successfully generalizes to unseen collision geometries, such as draping learned cloth over complex meshes not present during training.

Conclusion

SpringTime bridges the gap between high-fidelity offline simulation and real-time application requirements. By embedding physical constraints into the learning process, it achieves a balance of speed, stability, and the ability to represent complex, heterogeneous materials. Future work will explore learning optimal network topologies and extending the model to more exotic non-linear materials.

Summary updated

Abstract

Materials used in real clothing exhibit remarkable complexity and spatial variation due to common processes such as stitching, hemming, dyeing, printing, padding, and bonding. Simulating these materials, for instance using finite element methods, is often computationally demanding and slow. Worse, such methods can suffer from numerical artifacts called ``membrane locking'' that makes cloth appear artificially stiff. Here we propose a general framework, called SpringTime, for learning a simple yet efficient surrogate model that captures the effects of these complex materials using only motion observations. The cloth is discretized into a mass-spring network with unknown material parameters that are learned directly from the motion data, using a novel force-and-impulse loss function. Our approach demonstrates the ability to accurately model spatially varying material properties from a variety of data sources, and immunity to membrane locking which plagues FEM-based simulations. Compared to graph-based networks and neural ODE-based architectures, our method achieves significantly faster training times, higher reconstruction accuracy, and improved generalization to novel dynamic scenarios. Codebase for the paper can be found at https://github.com/ericchen321/springtime.
BibTeX
@misc{chen2025learning,
  author = {Guanxiong Chen and Shashwat Suri and Yuhao Wu and Etienne Voulga and David I. W. Levin and Dinesh K. Pai},
  title = {{Learning Simulatable Models of Cloth with Spatially-varying Constitutive Properties}},
  year = {2025},
  eprint = {2507.21288},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.21288},
}
Figure