README.md

---
license: cc-by-4.0
language:
  - en
pretty_name: SPH-Simulated LPBF Melt-Pool Dataset
size_categories:
  - 100K<n<1M
task_categories:
  - image-classification
  - image-segmentation
tags:
  - additive-manufacturing
  - laser-powder-bed-fusion
  - smoothed-particle-hydrodynamics
  - melt-pool
  - keyhole
  - physics-simulation
  - scientific-machine-learning
---

# SPH-Simulated LPBF Melt-Pool Dataset

Single-track laser powder bed fusion (LPBF) melt-pool simulations for Ti-6Al-4V,
produced with the LAMAS smoothed-particle-hydrodynamics solver. 241 simulations
sampled uniformly i.i.d. over a 4D process-parameter cube (laser power, scan
speed, laser spot radius, substrate temperature), spanning conduction, transition,
and keyhole regimes.

Companion to the NeurIPS 2026 Evaluations & Datasets Track submission
*A Simulation-Based Dataset for Melt Pool Dynamics in Single-Track Laser Powder
Bed Fusion*.

See examples folder in order to reproduce the results of the toy models.

Maintained by Ioan-Daniel Crăciun, Chair for Physics-Enhanced Machine Learning,
Technical University of Munich.

## Stats

| | |
|---|---|
| Simulations | 241 |
| Frames | 65,472 |
| Images (3 perspectives) | 196,416 |
| Top / front resolution | 256 × 256 px |
| Side resolution | 512 × 256 px |
| Material | Ti-6Al-4V |
| Atmosphere | Argon |

**Regime label distribution:** Initial Emptiness 5.0% / Forming 15.8% / Conduction 67.0% / Keyhole 0.8%.
The keyhole share is a lower bound — manual relabeling is ongoing.

## Labels

Two label types per frame:

- **Per-pixel phase** (image RGB): red = melt, green = solid substrate, blue = gas.
- **Per-frame regime** (4 classes): Initial Emptiness, Forming, Conduction, Keyhole.

Regime labels are seeded by a depth-based heuristic and verified through a
two-round human review with majority-vote consolidation.

## Process-parameter ranges

| Parameter | Symbol | Min | Max | Unit |
|---|---|---|---|---|
| Laser power | P | 76.70 | 249.99 | W |
| Scan speed | v | 0.204 | 0.998 | m/s |
| Laser spot radius | r_s | 45.10 | 89.70 | µm |
| Substrate temp. | T_s | 300.3 | 399.6 | K |

Fixed: particle spacing 4 µm, track length 1200 µm, end time 2.1 ms,
laser absorptance 0.2, wall temperature 300 K.

## Layout

```
experiments/<hash>/
├── images/{top,side,front}/   # PNGs per timestep
├── monitoring/*.dat            # scalar time series
├── labels.csv                  # per-frame regime labels
└── metadata.json               # process parameters
index.json                      # hash → parameters
```

## Usage

```python
from huggingface_hub import snapshot_download
path = snapshot_download(repo_id="ioandanielc/sph_dataset", repo_type="dataset")
```

## Limitations

Single-track on bare substrate (no powder bed, no multi-track). Fixed laser
absorptance. 2D cross-sections, not full 3D volumes. Keyhole class sparse by
design — the dataset is the Stage-1 initialization for an active-learning loop.
Not yet validated against experimental measurements.

## Citation

```bibtex
[bibtex placeholder — fill in after acceptance]
```

## License

CC BY 4.0.

## Contact

Ioan-Daniel Crăciun · TUM · <email-or-HF-handle>