PSDI & Royce Materials Data Summit

Automated nanoparticle segmentation of TEM images and generation of synthetic databases with TEMPOS

29 Jul 2026, 15:00

3h

2A.034 (Nancy Rothwell Building)

Tutorial (invited) Tutorial Room 1 Tutorials

Speaker

Andrew Stewart (University College London)

Description

Motivation

Quantitative analysis of nanoparticle populations from transmission electron micrographs remains a bottleneck in materials characterisation: manual segmentation is slow, subjective, and poorly scalable to the large datasets demanded by modern synthesis workflows. TEMPOS (Transmission Electron Microscopy Particle Outline Segmentation) addresses this through an instance-segmentation pipeline built on Mask R-CNN, enabling automated detection and morphometric analysis of nanoparticles across a wide range of imaging conditions and material systems. A distinguishing feature of the package is its synthetic data engine, which generates physically plausible TEM micrographs with ground-truth annotations, providing an effective route to training data in domains where experimentally labelled images are scarce. This tutorial offers a hands-on introduction to both capabilities. Participants will install TEMPOS, run inference on reference and personal datasets, extract per-particle descriptors, equivalent diameter, aspect ratio, spatial distribution, and export results for downstream analysis. The second half of the session covers synthetic database generation: configuring substrate and particle parameters, sampling across imaging conditions, and using the resulting data to fine-tune the model for a target material system. No prior machine learning experience is assumed; the tutorial is aimed at electron microscopists who wish to automate and standardise nanoparticle analysis in their own workflows.

Tutorial

This hands-on tutorial introduces TEMPOS, an open-source platform for automated nanoparticle segmentation and morphometric analysis in TEM images. Participants will learn how to install and run TEMPOS, apply the pre-trained Mask R-CNN model to their own micrographs, interpret per-particle outputs (size, shape descriptors, spatial distributions), and fine-tune the model on a small custom dataset. The session combines short conceptual segments with guided live demonstrations and self-paced exercises, and is designed to leave attendees with a working installation, sample notebooks, and a clear path to integrating TEMPOS into their own analysis workflows.

Learning outcomes

By the end of the tutorial, participants will be able to:

1. Explain, at a conceptual level, how instance-segmentation networks (Mask R-CNN) differ from semantic segmentation and classical thresholding, and why this matters for touching or overlapping nanoparticles.
2. Install and configure TEMPOS in a reproducible Python environment (conda or container) and verify the installation against a reference dataset.
3. Run inference on TEM micrographs using the pre-trained model and adjust key parameters (confidence threshold, NMS, tile size) for their own image conditions.
4. Interpret TEMPOS output: extract per-particle masks, equivalent diameters, aspect ratios, and population-level distributions, and export results to standard formats (CSV, JSON, OME-TIFF).
5. Annotate a small dataset and fine-tune the model on a domain-specific particle type, evaluating performance with appropriate metrics (mAP, IoU).
6. Identify common failure modes (low contrast, beam damage artefacts, agglomeration) and apply targeted mitigations.
7. Locate documentation, the GitHub repository, and community resources for ongoing use.

Tutorial Structure

Please note that breaks are included in this 3-hour tutorial

Part 1 — Context and theory

• Motivation: bottlenecks in manual nanoparticle analysis, reproducibility, the autonomous-TEM context.
• Instance segmentation in a nutshell: bounding boxes, masks, Mask R-CNN architecture without the maths.
• What TEMPOS is and what it isn't: scope, supported imaging modes, current limitations.

Part 2 — Installation and first run

• Walk-through of installation via pip/conda and the containerised release; verifying the environment.
• Guided inference on a provided reference micrograph; reading the output structure.

Part 3 — Hands-on inference and analysis

• Participants run TEMPOS on a provided dataset (or, optionally, their own)
• Parameter sweeps: confidence threshold, tile size, overlap; effects on recall and precision.
• Extracting morphometrics, plotting size distributions, exporting for downstream analysis.

Part 4 — Fine-tuning on custom data

• Annotation workflow with a lightweight labelling tool; what makes a good training set.
• Launching a fine-tuning run on a small pre-annotated dataset; monitoring training.
• Evaluating the fine-tuned model and comparing against the baseline.

Part 5 — Failure modes, integration, and Q&A

• Common pitfalls and how to diagnose them; integration sketches (Jupyter, scripted pipelines, autonomous loops).
• Resources, roadmap, and Q&A.