README.md

---
datasets:
- BGLab/FlowBench
license: mit
pipeline_tag: time-series-forecasting
library_name: pytorch-lightning
---

# Time-Dependent DeepONet for FlowBench (FPO)

This repository hosts pre-trained **time-dependent DeepONet** checkpoints used in the paper:

> **Predicting Time-Dependent Flow Over Complex Geometries Using Operator Networks**

These models are trained on the **FlowBench FPO** data and learn to predict unsteady flow over complex 2D geometries.

---

## Associated Resources

- **Paper (arXiv):** https://arxiv.org/abs/2512.04434  
- **Paper page on Hugging Face:** https://huggingface.co/papers/2512.04434  
- **Dataset (FlowBench on Hugging Face):** https://huggingface.co/datasets/BGLab/FlowBench  
- **Code (model implementation & training scripts):** https://github.com/baskargroup/TimeDependent-DeepONet  
- **Interactive demo (Hugging Face Space):** https://huggingface.co/spaces/BGLab/DeepONet-FPO-demo

---

## Checkpoints

All checkpoints are stored under the `checkpoints/` directory:

- `time-dependent-deeponet_1in.ckpt`  – model trained with input sequence length `s = 1`
- `time-dependent-deeponet_4in.ckpt`  – model trained with input sequence length `s = 4`
- `time-dependent-deeponet_8in.ckpt`  – model trained with input sequence length `s = 8`
- `time-dependent-deeponet_16in.ckpt` – model trained with input sequence length `s = 16`

Each checkpoint contains the weights for the time-dependent DeepONet used in the paper. For the exact architecture, data preprocessing, and training details, please refer to the GitHub repository.

You can download any checkpoint using `huggingface_hub`:

```python
from huggingface_hub import hf_hub_download
import torch

from models.geometric_deeponet.geometric_deeponet import GeometricDeepONetTime

REPO_ID = "BGLab/DeepONet-FlowBench-FPO"
filename = "checkpoints/time-dependent-deeponet_4in.ckpt"  # choose 1in / 4in / 8in / 16in

# 1) Download checkpoint file locally
ckpt_path = hf_hub_download(REPO_ID, filename)

# 2) Load the Lightning model from checkpoint
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = GeometricDeepONetTime.load_from_checkpoint(ckpt_path, map_location=device)
model = model.eval().to(device)
```
For full, reproducible training and evaluation, including data loading and post-processing, please see:

> https://github.com/baskargroup/TimeDependent-DeepONet

---

## Citation

```bibtex
@article{rabeh2025predicting,
  title={Predicting Time-Dependent Flow Over Complex Geometries Using Operator Networks},
  author={Rabeh, Ali and Murugaiyan, Suresh and Krishnamurthy, Adarsh and Ganapathysubramanian, Baskar},
  journal={arXiv preprint arXiv:2512.04434},
  year={2025}
}
```