EMSeek Dataset Overview

Background

The EMSeek dataset accompanies the paper Bridging Electron Microscopy and Materials Analysis with an Autonomous Agentic Platform (Chen et al., 2025). It supports the system's reference-guided, single-click segmentation workflows and provides curated examples for downstream lattice reconstruction, property prediction, and literature-grounded reasoning. The collection spans atomic-resolution and nanoscale transmission electron microscopy scenarios so that SegMentor, EM2CIF, MatProphet, and related agents can be trained and evaluated on a unified benchmark of heterogeneous materials tasks.

Directory Layout

Task-specific data are organized at the repository root:

atom_columns/
atomic_defects/
metal_alloy_defects/
nanoparticles/
single_atom_catalysts/
README.md

Each task directory contains one folder per material or specimen. Inside every material folder the files follow a consistent structure:

<task>/<material>/
├── image/        # Raw microscopy frames (.png)
├── label/        # Pixel-wise masks (.png)
├── bbox/         # YOLO-style annotations (.txt)
└── prompt.txt    # Optional textual prompt (one-word / concise / detailed)

Task Coverage

Atomic Column Segmentation (atom_columns)

Contains annular dark-field STEM frames of crystalline materials such as perovskites, transition-metal dichalcogenides, metals, silicon, and graphene. Images and masks are stored at 256×256 resolution to emphasize lattice periodicity, while accompanying bounding boxes mark atomic-column centroids used by the SegMentor agent.

Atomic Defect Identification (atomic_defects/WSe2)

Provides high-resolution WSe₂ micrographs featuring irradiation-induced point defects. RGB masks distinguish background lattice regions from localized defects, and bounding boxes highlight each defect center to support both semantic segmentation and detection workflows.

Metal Alloy Defects (metal_alloy_defects/FeCrAl)

Captures FeCrAl alloy foils imaged under differing magnifications. Binary masks delineate irradiation damage, and prompts summarize defect characteristics for conditioning language-aware agents. Image resolutions vary between 1,024×1,024 and 2,048×2,048 pixels to preserve mesoscale context.

Nanoparticle Analysis (nanoparticles)

Includes supported nanoparticle systems (for example Au@C, Au@Ge, Au@SiN, CdSe@C, Fe₃O₄, Fe₃O₄@SiO₂, Pd@C). Masks trace particle outlines for downstream size statistics and morphology analysis, while bounding boxes provide coarse localization for detection models. Large-format micrographs retain their native dimensions so users can perform multi-scale cropping when training.

Single-Atom Catalysts (single_atom_catalysts)

Reserved for future releases. The directory remains as a placeholder for upcoming single-atom catalyst annotations referenced in the paper.

Annotation Formats

• Images (image/*.png): Grayscale microscopy frames. Atomic column data are normalized to 256×256, whereas other tasks keep their original resolution.
• Masks (label/*.png): Typically encoded as binary grayscale (0 = background, 255 = target). The WSe₂ defect set uses an RGB palette to represent multiple semantic categories.
• Bounding boxes (bbox/*.txt): Provided in normalized YOLO format as <class> <x_center> <y_center> <width> <height>. Coordinates and extents are expressed between 0 and 1.
• Prompts (prompt.txt): Contain short, medium, and detailed textual descriptions that can be fed to language-conditioned agents to reproduce the EMSeek planning workflow.

Usage Notes

• Use image/ and label/ pairs to train segmentation models; bbox/ files can drive detection or assist with weakly supervised learning.
• For integration with the EMSeek pipeline, pass segmentation masks to EM2CIF for mask-aware lattice reconstruction and then to MatProphet for property inference, recreating the paper's pixels-to-properties loop.
• When handling large nanoparticle images, consider tile-based sampling, multi-scale augmentations, or down-sampling to fit GPU memory constraints. RGB masks in the defect task can be converted to index maps if a network expects single-channel labels.
• Some directories may contain system-generated files (for example ._*.png); filter them out during preprocessing to avoid loading errors.

License and Citation

• License: Apache License 2.0.
• When using this dataset, please cite the original work:

Chen, G., Yuan, W., & You, F. (2025). Bridging Electron Microscopy and Materials Analysis with an Autonomous Agentic Platform. Cornell University AI for Science Institute.

For the latest updates, visit the project repository: https://github.com/iCGY96/EMSeek.