KalininGroup/microscopy-hackathon-data

Microscopy Hackathon Data — Kalinin Group

A multi-modality, multi-instrument STEM corpus assembled from 10 published Zenodo deposits plus one plasmonic EELS dataset, normalized into ML-ready HDF5 with full provenance retained.

Layout

processed/                              # decoded acquisitions, one .h5 + .json sidecar each
  haadf_2d/                             # 64  HAADF-STEM images
  eels/                                 # 13  EELS spectra and spectrum images
  stem_edx_si/                          #  2  EDX spectrum image survey cubes
  stem_4d/                              #  3  4D-STEM diffraction stacks
  plasmonic_eels/                       # 12  plasmonic EELS spectrum images (split from one source)
metadata/
  inventory.parquet                     # full file index with DOI, modality, shape, dtype, pixel size
  inventory.csv                         # same, CSV
  missing_payloads.md                   # provenance ledger covering all 10 source DOIs
  batch_decode_log.json                 # what was decoded, deduplicated, or flagged partial
  acquisition_parameter_coverage.csv    # per-HAADF coverage of 7 standard acquisition params
DOI_to_Raw_Folder_Mapping.md

How to load

import h5py, json
from pathlib import Path

f = Path("processed/haadf_2d/from_AH__graphene__dset2__0045_-_20240404_STEM_2.72_nm.emd.h5")
with h5py.File(f, "r") as h:
    image = h["data"][...]                # (H, W, T) for HAADF
    pixel_nm = h.attrs.get("pixel_size_nm")

prov = json.loads(f.with_suffix(".json").read_text())
print(prov["doi"], prov["dataset_name"], prov["modality"])

Each .json sidecar carries: source_rel_path, source_size_bytes, doi, dataset_name, modality, decoder, conversion_timestamp, pixel_size_nm, dwell_time_s, frame_time_s, shape, dtype, n_frames, and the complete vendor metadata in a vendor_metadata field:

• For Velox EMDs: a list of per-frame metadata dicts (one per frame in the acquisition), with all top-level Velox sections preserved (Core, Instrument, Acquisition, Optics, EnergyFilter, Stage, Scan, Vacuum, Detectors, BinaryResult, Sample, GasInjectionSystems, CustomProperties).
• For Gatan DM3/DM4: rsciio's full metadata and original_metadata dicts for the primary signal, plus shape/metadata for any secondary signals in the file.
• For MRC: full header struct.

The same content is also stored inside each .h5 file as a vendor_metadata_json string dataset, so consumers don't need the sidecar to access it.

Provenance

Dataset	DOI	n files
Plasmonic EELS	10.5281/zenodo.20131814	1 source → 12 acquisitions
Graphene Dataset 2	10.5281/zenodo.20040988	16
Graphene Dataset 1	10.5281/zenodo.19965490	5
aBN-encapsulated Graphene	10.5281/zenodo.20131420	26
aBN-encapsulated Oxidised Graphene	10.5281/zenodo.20131691	20
aBN-encapsulated MoS2	10.5281/zenodo.20131712	53
SnSe HAADF	10.5281/zenodo.19963421	2
MoS2 Monolayer Y-implanted	10.5281/zenodo.19963879	7
CoAg	10.5281/zenodo.20131365	8
SiN-Au	10.5281/zenodo.20131796	2

32 additional files from a 2024-10-05 session (aBN-MoS2 unsintered + 4D-STEM + EELS) are included without a DOI; see metadata/missing_payloads.md.

Decoding notes

• Velox EMD pixel pitch is read from BinaryResult.PixelSize + BinaryResult.PixelUnitX. The N nm / N µm token in Velox filenames is the field of view, not the pitch.
• Gatan .dm3 / .dm4 are read via rsciio.digitalmicrograph. Files whose signal_type reads "EELS" are routed to processed/eels/ regardless of filename (inventory_modality in the sidecar preserves the original guess).
• MRC read via mrcfile. The single 4D-STEM acquisition is shape (782, 256, 256) int16.
• Kevin's plasmonic NPY is a pickled dict of 12 sub-acquisitions, one per top-level key, decoded with allow_pickle=True.
• Velox EMDs were deduplicated by the image dataset UUID under Data/Image/<uuid>/Data; duplicate source paths are preserved as duplicate_source_rel_paths on the canonical output.
• Acquisitions with image dims smaller than 128 px are flagged partial: true in the sidecar.

License

Each underlying Zenodo deposit retains its original license. See the DOI on each file's sidecar for the source. Code in scripts/ is provided under the same terms as the parent repository.

Acquisition-parameter coverage for HAADF

metadata/acquisition_parameter_coverage.csv reports, for each HAADF acquisition, which of seven standard acquisition parameters (pixel size, dwell time, convergence angle, beam current, gain, offset, inner collection angle) were recoverable from the source file. Useful when sub-selecting samples for any analysis that needs a complete parameter vector.