---
dataset_info:
  features:
  - name: case_id
    dtype: string
  - name: file
    dtype: string
  - name: num_nodes
    dtype: int64
  - name: num_edges
    dtype: int64
  - name: metadata
    dtype: string
  splits:
  - name: train
    num_bytes: 171422
    num_examples: 1100
  - name: test
    num_bytes: 44934
    num_examples: 290
  download_size: 108913
  dataset_size: 216356
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---

# Laser Powder Bed Fusion (LPBF) Additive Manufacturing Dataset

As part of our paper **[FLARE: Fast Low-Rank Attention Routing Engine](https://huggingface.co/papers/2508.12594)** ([arXiv:2508.12594](https://arxiv.org/abs/2508.12594)), we release a new **3D field prediction benchmark** derived from numerical simulations of the **Laser Powder Bed Fusion (LPBF)** additive manufacturing process.

This dataset is designed for evaluating neural surrogate models on **3D field prediction tasks** over complex geometries with up to **50,000 nodes**. We believe this benchmark will be useful for researchers working on graph neural networks, mesh-based learning, surrogate PDE modeling, or 3D foundation models.

---

## Dataset Overview

In metal additive manufacturing (AM), subtle variations in design geometry can cause residual stresses and shape distortion during the build process, leading to part inaccuracies or failures. We simulate the LPBF process on a set of complex 3D CAD geometries to generate a benchmark dataset where the goal is to **predict the vertical (Z) displacement field** of the printed part.

| Split        | # Samples | Max # Nodes / sample |
|--------------|-----------|----------------------|
| Train        | 1,100     | ~50,000              |
| Test         | 290       | ~50,000              |

Each sample consists of:

- `points`: array of shape `(N, 3)` (x, y, z coordinates of mesh nodes)
- optionally connectivity: `edge_index` array specifying axis-aligned hexahedral elements
- 3D `displacement` filed: array of shape `(N, 3)`
- Von mises `stress` field: array of shape `(N, 1)`

---

## Source & Generation

- Geometries are taken from the **Fusion 360 segmentation dataset**.
- Simulations performed using **Autodesk NetFabb** with Ti-6Al-4V material on a Renishaw AM250 machine.
- Full thermomechanical simulation producing residual stress and displacement fields.
- We applied subsampling and aspect-ratio filtering to select ~1,390 usable simulations.
- The dataset focuses on **steady-state residual deformation prediction**.

---

## Benchmark Task

**Task**: Given the 3D mesh coordinates of a part, predict the Z-displacement at each node after the LPBF build process (final state).

This surrogate modeling task is highly relevant to the additive manufacturing field, where fast prediction of distortion can save time and cost compared to full-scale FEM simulation.

---

## Citation

If you use this dataset in your work, please cite:

```
@misc{puri2025flare,
      title={{FLARE}: {F}ast {L}ow-rank {A}ttention {R}outing {E}ngine}, 
      author={Vedant Puri and Aditya Joglekar and Kevin Ferguson and Yu-hsuan Chen and Yongjie Jessica Zhang and Levent Burak Kara},
      year={2025},
      eprint={2508.12594},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2508.12594}, 
}
```

---

## Future Work & Extensions

We plan to expand this dataset toward larger-scale **3D shape foundation models**, and potentially include dynamic time-history fields (stress, temperature, etc.) in future releases.

---

## License

MIT License

---

## Contact

For questions about the dataset or related research, feel free to reach out via email or the GitHub repository linked in the paper: [`https://github.com/vpuri3/FLARE.py`](https://github.com/vpuri3/FLARE.py).