2026-10-12T10:01:27Z https://www.scidb.cn/oai

10.57760/sciencedb.nbsdc.00281 2026-09-04T10:07:54Z

2026-09-04 A benchmark dataset and baseline methods for four-dimensional STEM diffraction patterns doi:10.57760/sciencedb.nbsdc.00281 en Four-dimensional scanning transmission electron microscopy (4D-STEM) provides spatially resolved reciprocal-space information with exceptional sensitivity to crystal structure, orientation, strain and specimen heterogeneity. However, the large size of 4D-STEM datacubes, detector-specific formats, limited availability of standardized public datasets and strong correlations between neighbouring scan positions continue to restrict data reuse and the reliable development of machine-learning methods.We present 4D-ImageNet, a standardized benchmark dataset containing 174,000 diffraction patterns, including 145,000 experimental patterns selected from 29 independent 4D-STEM acquisitions and 29,000 multislice simulations generated from 13 crystal structures. The experimental component covers Ag, Au, mixed Au–Ag, CoO, Pd and ZnO specimens acquired under multiple fields of view, scan dimensions, camera lengths and exposure conditions. Each acquisition contributes 5,000 quality-ranked patterns together with source scan coordinates, acquisition metadata and pattern-selection metrics, enabling every retained diffraction pattern to be traced back to its original scan position.Beyond providing diffraction arrays, we introduce machine-readable, pattern-level Bragg-disk annotations generated using a set-prediction architecture combining a ConvNeXt V2 backbone with multi-scale deformable attention and learned disk queries. The resulting records contain predicted direct-beam coordinates, Bragg-disk centres and confidence scores without modifying the underlying experimental intensities. We clearly identify these annotations as model-derived pseudo-labels and retain the model configuration, preprocessing procedure and confidence threshold required for reproducible filtering or regeneration.We further establish baseline methods for evaluating the usability of the experimental and simulated components. A mixed-domain masked-reconstruction U-Net is trained using both domains and evaluated with acquisition-level and simulation-file-level splits to minimize information leakage. We additionally quantify simulation coverage and the remaining experiment–simulation domain gap through Euler-angle statistics, approximate low-index beam-direction distributions, radial-intensity profiles and frozen feature-space analysis. These evaluations document both the capabilities and limitations of the dataset rather than presenting the simulations as paired ground-truth counterparts of the experimental patterns. Electron Microscopy; 4D-STEM; diffraction pattern; scanning transmission electron microscopy; deep learning Guan Yuyan Wang Hong Zeng Xiaoqin Xie Yujun PUBLIC https://creativecommons.org/licenses/by/4.0/ dataset Science Data Bank