Automated detection of overlapping and well-separated nanoparticles in transmission electron microscopy via deep learning.

Accurate detection of nanoparticles in transmission electron microscopy (TEM) images is essential for high-throughput characterization in chemistry and materials science. TEM micrographs of advanced photocatalysts often contain densely packed and overlapping nanoparticles, creating electron-dense regions where individual particle boundaries cannot be resolved. This challenges accurate particle counting, sizing, and morphological analysis, and motivates the development of reliable automated analysis workflows. Here, we present a deep learning workflow for robust nanoparticle detection that combines YOLO for fast and reliable localization with the Segment Anything Model (SAM) for pixel-level refinement of particle contours. The TEM dataset consists of newly developed photocatalysts, including undoped and noble metal (Ag, Pd, Ni), doped ZnO, TiO₂, Ni/g-C₃N₄, CdS, heterojunction copper(I/II) oxides, ZnO mixed metal oxides, and metal-organic frameworks. These novel materials produce distinctive TEM images with both well-separated and heavily overlapping nanoparticles. Multiple YOLO variants, including YOLOv8s, YOLOv8x, YOLOv11s, YOLOv12s, and YOLO26s, were benchmarked for nanoparticle detection. The YOLO-generated bounding boxes were subsequently refined using SAM to obtain accurate pixel-level segmentations of resolvable nanoparticles. Performance is evaluated on both the challenging TEM dataset and external literature images using standard metrics and confusion matrices. The proposed workflow was further benchmarked against widely used manual and semi-automated TEM analysis methods, demonstrating superior robustness, accuracy, and consistency, particularly in regions with severe particle overlap. Overall, this workflow provides an efficient and scientifically rigorous framework for high-throughput characterization of resolvable nanoparticles in complex TEM micrographs while avoiding unsupported measurements in regions where individual particle boundaries cannot be objectively determined.