Physics-Informed Neural Networks for High-Fidelity Dynamic Cloth Simulation: Parameter Estimation and Differentiable Simulation
This repository provides an implementation of a dual-driven physics-informed neural network (PINN) framework for accurate dynamic cloth simulation. The approach integrates physical modeling with data-driven learning, achieving precise parameter estimation and high-fidelity simulations for cloth behavior. The framework consists of two key components:
- K-PINN: Reduces density estimation error from 1.8% to 1.2% compared to existing methods.
- DP-PINN: Outperforms traditional physics-based methods and data-driven models in simulation speed (6.1 ms for two-point fixed cloth) and physical consistency, providing stable and realistic dynamic cloth behaviors.
- High-Fidelity Simulation: Achieves real-time and accurate cloth simulation with complex deformations like stretching, bending, and wrinkling.
- Dual-Driven Framework: Combines physics-based modeling and data-driven learning, offering both simulation accuracy and computational efficiency.
- Differentiable Physics Engine: Enables seamless optimization of physical parameters and network weights through backpropagation.
- Realistic Cloth Deformations: Suitable for virtual reality, digital humans, and virtual try-on applications.
To install the dependencies, run:
pip install -r requirements.txt- Find the path to the training script.
ncs/train.py/
python main.py
(1) Stretch parameter prediction and error variation (left), shear parameter prediction and error variation (center), bend parameter prediction and error variation (right).
| Method | Density Error (%) ↓ | Stretch Error (%) ↓ | Shear Error (%) ↓ |
|---|---|---|---|
| K-PINN (Full) | 1.2 ± 0.03 | 40 ± 23 | 47 ± 17 |
| Fixed weights | 1.5 ± 0.1 | 45 ± 30 | 52 ± 20 |
| Pure data-driven (w/o physical) | 10.1 ± 1.8 | 68 ± 45 | 72 ± 38 |
| Model | Avg. ositional Error (%) ↓ | Energy Conservation (%) ↑ | Material Parameter Error (%) ↓ |
|---|---|---|---|
| PINN (baseline) | 4.72 ± 0.63 | 68.2 | 0.035 |
| +Constitutive | 3.55 ± 0.41 | 79.5 | 0.045 |
| +Momentum | 2.94 ± 0.37 | 88.3 | 0.042 |
| K-PINN | 2.14 ± 0.31 | 95.1 | 0.027 |
(2) Results of ablation experiment.
(3) Detail comparison of ribbed fabric simulation results using Physics-Based, MGN, and DPPINN methods from left to right.
(4) From left to right are the comparison results of clothing simulation details using the physical method, the MGN method, and the DP-PINN model.
