---
license: mit
pipeline_tag: other
tags:
- seismology
- geophysics
- time-series
---

# TimesNet-Gen: Deep Learning-based Site Specific Strong Motion Generation

TimesNet-Gen is a time-domain conditional generator designed for site-specific strong ground motion synthesis from accelerometer records. It employs a latent bottleneck with station identity conditioning to capture local site influences on ground motion characteristics.

- **Paper:** [TimesNet-Gen: Deep Learning-based Site Specific Strong Motion Generation](https://huggingface.co/papers/2512.04694)
- **GitHub Repository:** [brsylmz23/TimesNet-Gen](https://github.com/brsylmz23/TimesNet-Gen)
- **Demo:** [Hugging Face Space](https://huggingface.co/spaces/Barisylmz/TimesNet-Gen)

## Installation

To use the model, first clone the repository and install the required dependencies:

```bash
git clone https://github.com/brsylmz23/TimesNet-Gen.git
cd TimesNet-Gen
pip install -r requirements.txt
```

## Sample Usage

After downloading the pre-trained checkpoints (as described in the [GitHub README](https://github.com/brsylmz23/TimesNet-Gen#2-download-pre-trained-model)), you can generate samples using the provided script:

```bash
# Generate 50 samples per station for the default stations
python generate_samples.py --num_samples 50

# Generate for specific stations (e.g., station IDs 0205 and 1716)
python generate_samples.py --stations 0205 1716 --num_samples 100
```

The output includes generated waveforms in NPZ format along with HVSR curves and f₀ distribution comparisons.

## Citation

If you find this work useful, please cite:

```bibtex
@misc{yilmaz2025timesnetgendeeplearningbasedsite,
      title={TimesNet-Gen: Deep Learning-based Site Specific Strong Motion Generation}, 
      author={Baris Yilmaz and Bevan Deniz Cilgin and Erdem Akagündüz and Salih Tileylioglu},
      year={2025},
      eprint={2512.04694},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2512.04694}, 
}
```