ioandanielc/sph_dataset

SPH Melt Pool Dataset for Single-Track Laser Powder Bed Fusion

Paper: A Simulation-Based Dataset for Melt Pool Dynamics in Single-Track Laser Powder Bed Fusion

Authors: Ioan-Daniel Craciun, Stefan Adami, Felix Dietrich Institution: Technical University of Munich (TUM) License: CC BY 4.0

---

Dataset Description

This dataset contains SPH-simulated melt pool dynamics for single-track laser powder bed fusion (LPBF) experiments on Ti-6Al-4V (TI64). It is designed to enable machine learning research on melt pool geometry, regime classification (conduction vs. keyhole), and generative modeling of subsurface melt pool cross-sections.

The dataset spans the full process parameter space from stable conduction regimes to hazardous keyhole formation, providing the ground truth needed to learn the boundary between safe and defect-prone process conditions.

---

Dataset Structure

Each experiment corresponds to one simulation run at a unique parameter point. Experiments are named by their parameter values:

P-{laser_power}_VX-{scan_speed}_LS-{spot_size}_ST-{substrate_temp}_M-TI64_..._H-{hash}/
├── png_files/          # Cross-sectional images (top, side, front views)
├── monitor/            # Scalar monitoring time series (.dat files)
├── gif_files/          # Animated GIFs of melt pool evolution
├── labels.csv          # Per-frame phase labels (timestep, label)
└── *.json              # Simulation configuration file

Images

Three cross-sectional perspectives per timestep, rendered in false color:

• 🔴 Red — melt pool (liquid TI64)
• 🟢 Green — solid substrate (TI64)
• 🔵 Blue — gas atmosphere (Argon)

Perspective	Resolution
Top (XY)	256 × 256 px
Side (XZ)	512 × 256 px
Front (YZ)	256 × 256 px

Phase Labels

Each frame is assigned one of four mutually exclusive labels:

Label	Description
Initial Emptiness	Domain prior to laser arrival
Forming Phase	Melt pool nucleation and growth
Convection	Steady-state Marangoni-driven flow
Keyhole	Deep vapor depression regime

⚠️ Labels are auto-initialized via a delta-z heuristic and pending human verification. The labels.csv file contains headers only for experiments not yet reviewed.

Scalar Monitoring Data

Per-timestep .dat files in monitor/:

File	Description
maximum-temperature_melt.dat	Peak melt pool temperature
average-temperature_melt.dat	Mean melt pool temperature
particle-number_melt.dat	Proxy for melt pool volume
position-bounds_melt.dat	Spatial extent (depth, width)
transferred-laser-power_melt.dat	Absorbed energy per timestep
kinetic-energy_melt.dat	Marangoni convection intensity
time.dat / dt.dat	Temporal metadata

---

Process Parameters

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 temperature	T_s	300.3	399.6	K

Sampling: uniform random across the 4D parameter space.

---

Intended Uses

• Melt pool regime classification (conduction vs. keyhole)
• Generative modeling of melt pool morphology
• Melt pool geometry regression (width, depth)
• Initialization for Bayesian optimization-based active learning

Out-of-Scope Uses

• Multi-track or powder bed LPBF (bare substrate only)
• Materials other than Ti-6Al-4V
• Direct experimental validation without sim-to-real transfer

---

Citation

@dataset{craciun2026sph,
  title     = {A Simulation-Based Dataset for Melt Pool Dynamics in Single-Track Laser Powder Bed Fusion},
  author    = {Craciun, Ioan-Daniel and Adami, Stefan and Dietrich, Felix},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/ioandanielc/sph_dataset},
  license   = {CC BY 4.0}
}