# Model Card for DeepFracture

## Model Description

- **Model type**: 3D Fracture Pattern Prediction via impulse-code conditional VQ-VAE models
- **Language(s)**: Python
- **License**: MIT
- **Finetuned from model**: Custom architecture

### Model Sources

- **Repository**: https://github.com/nikoloside/TEBP
- **Paper**: https://doi.org/10.1111/cgf.70002
- **Demo**: [WIP]

## Uses

### Direct Use

These models are designed for:
- Computer graphics applications
- Game development
- Virtual reality environments
- Paper code release

### Out-of-Scope Use

- Medical applications
- Safety-critical systems
- High-precision engineering simulations

## Bias, Risks, and Limitations

### Bias

The models are trained on synthetic simulation data and may not generalize well to real-world scenarios with different material properties or environmental conditions.

### Risks

- Models may produce unrealistic deformations under extreme conditions
- Performance may significantly different from training data
- No guarantees for physical accuracy in safety-critical applications

### Limitations

- Limited to the specific one target shape and object categories in the training dataset
- Be able to reach real-time inference in network, but need resconstruction within 2-10 seconds

## Training Details

### Training Data

- **Dataset**: Break4Model dataset
- **Training samples**: Varies by model (277-433 samples per category)
- **Validation samples**: 20% of training data
- **Data preprocessing**: Normalized impact conditions and geometry of GS-SDF

### Training Procedure

- **Training regime**: Supervised learning
- **Optimizer**: Adam
- **Learning rate**: 1e-3 - 1e-5
- **Batch size**: 1
- **Training epochs**: 1000
- **Hardware**: NVIDIA RTX 3090 GPU
- **Training time**: ~24 hours per model

### Training Results

(metrics and performance)[https://doi.org/10.1111/cgf.70002]

## Evaluation

### Testing Data

- **Dataset**: Break4Models
- **Metrics**: Geometric accuracy, Fragment size, Inference time, Fragment distribution, Surface normals, MPCD (Multiple-Phase Charmfer Distance)

### Results

The models achieve high accuracy in predicting fracture patterns while maintaining nearly real-time performance suitable for interactive applications.

## Environmental Impact

- **Hardware Type**: GPU
- **Hours used**: ~24 hours per model
- **Dataset**: Break4Model dataset
- **Framework**: PyTorch
- **Optimizer**: Adam
- **Loss Function**: L2 Loss
- **Training Time**: ~24 hours per model on NVIDIA RTX 3090

## Technical Specifications

### Model Architecture

- **Encoder**: SIREN Layer
- **Decoder**: CNN
- **Parameters**: ~2M parameters per model
- **Input**: Impact conditions (position, velocity, impulse strength) from Bullet3
- **Output**: GS-SDF (Geometrically-Segmented Signed Distance Fields)

### Compute Requirements

- **Training**: NVIDIA RTX 3090 or equivalent
- **Inference**: CPU or GPU
- **Memory**: 8GB RAM minimum
- **Storage**: ~200MB per model

## Citation

```bibtex
@article{huang2025deepfracture,
author = {Huang, Yuhang and Kanai, Takashi},
title = {DeepFracture: A Generative Approach for Predicting Brittle Fractures with Neural Discrete Representation Learning},
journal = {Computer Graphics Forum},
pages = {e70002},
year = {2025},
keywords = {animation, brittle fracture, neural networks, physically based animation},
doi = {https://doi.org/10.1111/cgf.70002},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1111/cgf.70002},
eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1111/cgf.70002}
}
```

## Data Card Authors

Huang Niko(nikoloside), Fang Chaowei(fangsunjian)

## Data Card Contact

https://github.com/nikoloside/TEBP