MatSciFig is a large-scale dataset of panel-level crops from scientific compound figures in materials science literature, paired with panel-specific subcaptions and extended summaries extracted from the source papers.

Compound figures — multi-panel images labeled A, B, C… — are the dominant figure type in scientific publications. MatSciFig breaks these figures into individual panels and links each panel to its corresponding textual description, enabling fine-grained vision-language research on scientific imagery.

Dataset Statistics

	Count
Source figures	~147,500
Panels with full metadata	391,285

By domain:

Domain	Panels
Ceramics	164,366
Additive manufacturing	75,178
Ni-alloy	70,024
Steel	36,631
Alloy	36,249
Composite	8,837

Cross-domain duplicate figures (the same source figure independently scraped into more than one domain pipeline) have been identified and removed, so each panel appears exactly once.

By visualization category:

Category	Panels
Microscopy	97,578
Generic Plot	76,676
Schematic/Diagram	46,937
Mechanical Test	35,393
Photograph	29,212
Simulation	27,904
Diffraction	27,089
Compositional Map	11,621
Spectroscopy	8,419
Tomography/3D	7,714
Electrochemistry	6,230
Machine Learning	4,862
Thermal Analysis	4,458
Phase/Equilibrium Diagram	2,481
Crystal Structure	1,790
Magnetic/Electronic	1,455
Optical/Photonic	798
Table	366
Other	302

Each panel is further labeled with a fine-grained visualization_subtype (259 distinct subtypes, e.g. SEM, Line Graph, EBSD Map, XRD Pattern), and a subtype_valid flag marking whether that subtype assignment passed automated validation (381,726 valid / 9,559 flagged).

Columns

Column	Type	Description
`image`	bytes	Panel crop in original format
`file_name`	string	Relative path to the panel image (`images/<image_id>_<panel_suffix>.jpg`)
`image_id`	string	Parent figure identifier
`panel_suffix`	string	Panel label (A, B, C … or `single`)
`visualization_category`	string	High-level category (e.g. Microscopy, Generic Plot, Photograph)
`visualization_subtype`	string	Fine-grained subtype (e.g. SEM, Contour Plot, XRD Pattern)
`subtype_valid`	bool	Whether the subtype assignment passed automated validation
`subcaption`	string	Panel-level caption from the source paper
`summary`	string	Extended description of the panel content
`doi`	string	DOI of the source publication
`year`	int	Publication year of the source paper
`citation_count`	int	Citation count of the source paper at time of collection

Example

from datasets import load_dataset

ds = load_dataset("subham2507/MatSciFig")
sample = ds["train"][0]

print(sample["panel_suffix"])            # "A"
print(sample["visualization_category"])  # "Microscopy"
print(sample["visualization_subtype"])   # "SEM"
print(sample["subcaption"])              # "SEM image of Fe2O3 nanorods..."
print(sample["summary"])                 # "The SEM image reveals..."
print(sample["doi"])                     # "10.1016/J.ENG.2017.03.005"
print(sample["year"])                    # 2017
print(sample["citation_count"])          # 54
sample["image"].show()                   # displays the panel crop

Data Collection

Compound figures collected from open-access materials science publications
Panels detected and segmented using a custom-trained YOLO-based compound figure detector
Subcaptions and summaries extracted and aligned from paper text using automated parsing
Each panel is assigned a visualization_category and visualization_subtype, with a subtype_valid flag from automated validation
Source papers are enriched with doi, publication year, and citation_count metadata
Only panels with complete metadata are included

Intended Use

Scientific figure understanding and captioning
Vision-language model training and evaluation
Materials science multimodal research
Panel classification by visualization type

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

You need to agree to share your Hugging Face information to access this dataset

MatSciFig