Statistical Characterization of the Morphologies of Nanoparticles through Machine Learning Based Electron Microscopy Image Analysis

TLDR

Transmission electron microscopy is a powerful technique for studying nanoparticle morphology, yet quantitative statistical analysis remains difficult, and the polydisperse distribution of nanoparticles critically influences accurate optical property estimation. The authors aim to provide a mass‑throughput method for nanoparticle morphology analysis that will enable large‑scale statistical studies and become a valuable tool for big‑data research in nanoscience. They apply a genetic algorithm to an image‑analysis pipeline, cluster particles with similar shapes into groups for statistical evaluation, and account for the number of TEM measurements and average particles per image to ensure representative sampling. The method analyzes over 150,000 nanoparticles with 99.75 % precision and a 0.25 % false‑discovery rate, and shows that at least 1,500 particles are required to represent the population within a 95 % credible interval.

Abstract

Although transmission electron microscopy (TEM) may be one of the most efficient techniques available for studying the morphological characteristics of nanoparticles, analyzing them quantitatively in a statistical manner is exceedingly difficult. Herein, we report a method for mass-throughput analysis of the morphologies of nanoparticles by applying a genetic algorithm to an image analysis technique. The proposed method enables the analysis of over 150,000 nanoparticles with a high precision of 99.75% and a low false discovery rate of 0.25%. Furthermore, we clustered nanoparticles with similar morphological shapes into several groups for diverse statistical analyses. We determined that at least 1,500 nanoparticles are necessary to represent the total population of nanoparticles at a 95% credible interval. In addition, the number of TEM measurements and the average number of nanoparticles in each TEM image should be considered to ensure a satisfactory representation of nanoparticles using TEM images. Moreover, the statistical distribution of polydisperse nanoparticles plays a key role in accurately estimating their optical properties. We expect this method to become a powerful tool and aid in expanding nanoparticle-related research into the statistical domain for use in big data analysis.