---
language:
- en
license: apache-2.0
pipeline_tag: other
tags:
- physics
- PDEs
- surrogate
- FoundationModels
- AI4Sci
---

<p align="center">
<a href="https://github.com/tum-pbs/Tadpole">
  <img src="https://huggingface.co/thuerey-group/Tadpole/resolve/main/assets/images/tadpole_colorful.png" width="100"/>
</a>
</p>
<h4 align="center">Autoencoders as Foundation Models for 3D PDEs with Online Learning</h4>
<p align="center">
<a href="https://arxiv.org/abs/2605.15284">
  <img src="https://img.shields.io/badge/arXiv-2605.15284-b31b1b?logo=arxiv" alt="Read on arXiv"/>
</a>
<a href="https://github.com/tum-pbs/Tadpole">
  <img src="https://img.shields.io/badge/Github-Tadpole-181717?logo=github" alt="Github project"/>
</a>
</p>

## About

This repository contains pre-trained weights of **Tadpole**, a foundation model for three-dimensional partial differential equations (PDEs) introduced in the paper [Tadpole: Autoencoders as Foundation Models for 3D PDEs with Online Learning](https://huggingface.co/papers/2605.15284). 

Tadpole distinguishes itself from existing PDE foundation models in three key aspects: 

(1) **Autoencoder Pre-training**: Tadpole is pre-trained to learn the inherent representation of PDE solutions, which is more generalizable than the traditional paradigm of training models directly on temporal dynamics.

(2) **Online Learning**: It utilizes a GPU-based solver to generate diverse data on-the-fly, avoiding storage and I/O bottlenecks associated with massive 3D PDE datasets.

(3) **Multi-functionality**: Tadpole can be applied to multiple downstream tasks, including autoencoding, dynamics prediction, and generative modeling.

## Installation and Loading Pre-trained Weights

Tadpole can be installed via pip:

```bash
pip install git+https://github.com/tum-pbs/Tadpole
```

The pre-trained weights are named as `tadpole_{SIZE}_{TYPE}.safetensors`, where `{SIZE}` can be `S`, `B`, or `L` indicating the model size, and `{TYPE}` can be `encoder` or `decoder`. Weights can be loaded through `weight_{TYPE}` arguments in Tadpole model classes:

```python
from huggingface_hub import hf_hub_download

# Example: loading the B-size model
ae = TadpoleAutoencoder(
    size="B",
    weight_encoder=hf_hub_download(repo_id="thuerey-group/Tadpole", filename="tadpole_b_encoder.safetensors"),
    # or you can also download the weights from Hugging Face and load it locally
    weight_decoder=hf_hub_download(repo_id="thuerey-group/Tadpole", filename="tadpole_b_decoder.safetensors"),
    )
```

Please refer to the [GitHub repository](https://github.com/tum-pbs/Tadpole) for more details and tutorials.

Note: Currently, pre-trained weights for the B-size model are provided; S- and L-size models will be released in the future.

## Citation

```latex
@article{Liu2026Tadpole,
  author = {Qiang Liu, Felix Koehler, Benjamin Holzschuh, and Nils Thuerey},
  title = {Tadpole: Autoencoders as Foundation Models for 3D PDEs with Online Learning},
  eprint={2605.15284},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2605.15284} 
}
```