Automating Nanoparticle Analysis in Transmission Electron Microscopy with Synthetically Trained Segmentation Models
Year: 2026
Type: preprint
Abstract: Nanoparticle morphology governs the functional behaviour: plasmonic, catalytic, magnetic, electronic, of colloidal and supported nanomaterials, yet distributions of single-particle morphology are seldom reported. Published distributions typically aggregate ensemble averages or a few dozen manually segmented transmission electron microscopy (TEM) images, leaving population-level descriptors statistically underpowered. The rate-limiting step is instance segmentation: convolutional neural networks require extensive pixel-level annotations that are prohibitive to curate for every new particle system. Here we report TEMPOS, a deep-learning instance-segmentation package trained on procedurally generated synthetic TEM micrographs which simulate arbitrary particle distributions, substrate, noise, and imaging aberrations. TEMPOS outputs per-particle shape descriptors to a FAIR-compliant relational database and renders image archives searchable by morphological query. Distributed with a graphical interface for nondomain experts, the package reduces large-dataset analysis from weeks of manual segmentation to minutes of active work, enabling routine statistical morphometry across nanomaterials research.
Source: ChemRxiv
Authors: Natalia Sá, Andrew Stewart
Institution: University College London
Language: English
Subfield: Structural Biology
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7221768201