ppak10/Peregrine-Dataset-v2023-11

Peregrine v2023-11 — Layer-wise L-PBF Imaging Dataset

A HuggingFace-formatted conversion of the Peregrine v2023-11 dataset released by Oak Ridge National Laboratory's (ORNL) Manufacturing Demonstration Facility (MDF). The dataset contains layer-wise in-situ imaging, anomaly segmentation masks, laser scan paths, and process sensor data from 5 Laser Powder Bed Fusion (L-PBF) builds of stainless steel 316L.

Original dataset: Layer-wise Imaging Dataset from Powder Bed Additive Manufacturing Processes for Machine Learning Applications (Peregrine v2023-11) Authors: L. Scime, V. Paquit, C. Joslin, D. Richardson, D. Goldsby, L. Lowe, D. Siddel, S. Baird, R. Duncan, F. Brinkley, C. Ledford Institution: Oak Ridge National Laboratory, Manufacturing Demonstration Facility Contact: Dr. Luke Scime — scimelr@ornl.gov

---

Dataset Summary

Machine	Concept Laser M2 (L-PBF)
Material	Stainless steel 316L
Builds	5 (TCR Phase 1, Builds 1–5)
Total layers	17,582
Image resolution	1842 × 1842 px (float32)
Segmentation classes	12 (DSCNN pixel-wise)
Tensile coupons	~6,299 SS-J3 specimens
Temporal sensors	19 per-layer scalar readings

Each row is one build layer. Full-build images are retained — users can derive per-part crops using the part_ids mask.

---

Builds

Build	Date	Layers
1	2021-07-13	3,575
2	2021-04-16	3,561
3	2021-04-28	3,297
4	2021-07-13	3,574
5	2021-08-23	3,575

Note: Build 3 has a scan path / slice layer count mismatch (3,297 slice layers, 3,186 scan layers). Layers without a scan path have scan_path = null.

---

Schema

Each parquet file contains one row (one build layer) with 42 columns:

Visual preview

Column	Type	Description
segmentation_preview	PNG bytes	Anomaly overlay on melt image. Colored masks per class, red bounding boxes per part with part ID label, legend of active anomaly classes.

Identifiers

Column	Type	Description
build	int	Build number (1–5)
build_name	str	Build name from HDF5 metadata
layer	int	Layer index (0-based)

Camera images

Column	Type	Description
image_after_melt	.npy bytes	Float32 visible-light image captured after laser melting (1842×1842)
image_after_melt_preview	PNG bytes	Normalized uint8 preview of image_after_melt
image_after_powder	.npy bytes	Float32 visible-light image captured after powder spreading (1842×1842)
image_after_powder_preview	PNG bytes	Normalized uint8 preview of image_after_powder

ID maps

Column	Type	Description
part_ids	.npy bytes	uint32 (1842×1842) — pixel-to-part-id map. 0 = background. Use to crop per-part images.
sample_ids	.npy bytes	uint32 (1842×1842) — pixel-to-sample-id map. 0 = background.

Segmentation masks (DSCNN)

One boolean .npy column per class (1842×1842 bool arrays):

Column	Class index	Description
segmentation_powder	0	Loose powder
segmentation_printed	1	Solidified material (normal)
segmentation_recoater_hopping	2	Recoater hopping anomaly
segmentation_recoater_streaking	3	Recoater streaking anomaly
segmentation_incomplete_spreading	4	Incomplete powder spreading
segmentation_swelling	5	Part swelling anomaly
segmentation_debris	6	Debris on powder bed
segmentation_super_elevation	7	Super elevation anomaly
segmentation_spatter	8	Spatter deposits
segmentation_misprint	9	Misprint
segmentation_over_melting	10	Over-melting
segmentation_under_melting	11	Under-melting

Scan path

Column	Type	Description
scan_path	.npy bytes or null	Float32 array (N×5): columns are [x_start, y_start, x_end, y_end, exposure_time]. Null where absent (some Build 3 layers).

Temporal sensors (19 float64 scalars)

absolute_image_capture_timestamp, actual_ventilator_flow_rate, bottom_chamber_temperature, bottom_flow_rate, bottom_flow_temperature, build_chamber_position, build_plate_temperature, build_time, gas_loop_oxygen, glass_scale_temperature, laser_rail_temperature, layer_times, module_oxygen, powder_chamber_position, target_ventilator_flow_rate, top_chamber_temperature, top_flow_rate, top_flow_temperature, ventilator_speed

---

Anomaly Statistics

Per-layer anomaly pixel counts across all 5 builds. Each plot shows the raw per-layer count (shaded) and a 30-layer rolling mean (solid line). Generated by info/anomaly_stats.py and info/plot_anomalies.py.

Build 1 — 3,575 layers

Build 2 — 3,561 layers

Build 3 — 3,297 layers

Build 4 — 3,574 layers

Build 5 — 3,575 layers

---

Benchmark Config Anomaly Distributions

Pixel and presence breakdowns for the three anomaly benchmark configs, generated by scripts/generate_anomaly_0250.py, scripts/generate_anomaly_0500.py, and scripts/generate_anomaly_1000.py. Each config shows two pie charts:

• By pixels — total anomaly pixel count per class (a layer contributes to multiple slices if multiple classes are active)
• By presence — number of layers where each class is present (pixels > 0)

anomaly_0250 — 50 layers/build · 250 layers total

By pixels	By presence

Build 1 — by pixels	Build 1 — by presence

Build 2 — by pixels	Build 2 — by presence

Build 3 — by pixels	Build 3 — by presence

Build 4 — by pixels	Build 4 — by presence

Build 5 — by pixels	Build 5 — by presence

anomaly_0500 — 100 layers/build · 500 layers total

By pixels	By presence

Build 1 — by pixels	Build 1 — by presence

Build 2 — by pixels	Build 2 — by presence

Build 3 — by pixels	Build 3 — by presence

Build 4 — by pixels	Build 4 — by presence

Build 5 — by pixels	Build 5 — by presence

anomaly_1000 — 200 layers/build · 1,000 layers total

By pixels	By presence

Build 1 — by pixels	Build 1 — by presence

Build 2 — by pixels	Build 2 — by presence

Build 3 — by pixels	Build 3 — by presence

Build 4 — by pixels	Build 4 — by presence

Build 5 — by pixels	Build 5 — by presence

---

Loading the Dataset

from datasets import load_dataset

ds = load_dataset("your-org/peregrine-v2023-11", split="train")

Or load a single build directly with pandas:

import pandas as pd
from pathlib import Path
import io
import numpy as np

df = pd.read_parquet("data/build/1/layer_0651.parquet")
row = df.iloc[0]

Deserializing .npy columns

All binary columns (images, masks, ID maps, scan paths) are serialized NumPy arrays. Load them with:

import io
import numpy as np

img = np.load(io.BytesIO(row["image_after_melt"]))   # float32, shape (1842, 1842)
part_ids = np.load(io.BytesIO(row["part_ids"]))       # uint32,  shape (1842, 1842)
mask = np.load(io.BytesIO(row["segmentation_spatter"]))  # bool, shape (1842, 1842)

Crop a part from the melt image

pid = 5  # example part ID
region = part_ids == pid
rows_idx = np.where(region.any(axis=1))[0]
cols_idx = np.where(region.any(axis=0))[0]
crop = img[rows_idx[0]:rows_idx[-1]+1, cols_idx[0]:cols_idx[-1]+1]

Rebuild the segmentation preview

See segmentation.py for a complete reference implementation:

uv run segmentation.py data/build/1/layer_0651.parquet

---

Segmentation Color Map

Fixed colors used in segmentation_preview, consistent across all builds and layers:

Class	Color	RGB
recoater_hopping	Cyan	(0, 230, 230)
recoater_streaking	Orange	(255, 140, 0)
incomplete_spreading	Yellow	(255, 220, 0)
swelling	Red	(220, 50, 50)
debris	Magenta	(230, 0, 230)
super_elevation	Lime green	(50, 220, 50)
spatter	Hot pink	(255, 80, 160)
misprint	Violet	(160, 80, 255)
over_melting	Dodger blue	(30, 144, 255)
under_melting	Teal	(0, 180, 180)

---

Attribution

This dataset is a reformatted version of the Peregrine v2023-11 dataset produced by Oak Ridge National Laboratory. All scientific credit belongs to the original authors.

If you use this dataset, please cite the original work:

@article{scime2020anomaly,
  title   = {Layer-wise anomaly detection and classification for powder bed additive
             manufacturing processes: A machine-agnostic algorithm for real-time
             pixel-wise semantic segmentation},
  author  = {Scime, Luke and Siddel, Daniel and Baird, Sean and Paquit, Vincent},
  journal = {Additive Manufacturing},
  volume  = {36},
  pages   = {101453},
  year    = {2020},
  doi     = {10.1016/j.addma.2020.101453}
}

Previous Peregrine dataset releases (also cite if relevant):

• Peregrine v2021-03: https://doi.org/10.13139/ORNLNCCS/1779073
• Peregrine v2022-10: https://doi.org/10.13139/ORNLNCCS/1896716

---

License

This dataset is distributed under the terms set by Oak Ridge National Laboratory. Please refer to the original dataset release and contact Dr. Luke Scime (scimelr@ornl.gov) for licensing details.