Machine Learning to Reveal Nanoparticle Dynamics from Liquid-Phase TEM Videos

Area of Science:

• Materials Chemistry
• Nanotechnology
• Machine Learning Applications

Background:

• Liquid-phase transmission electron microscopy (TEM) offers nanometer-resolution insights into dynamic processes.
• Quantitative analysis of liquid-phase TEM videos is hindered by noise and heterogeneity.
• Automated analysis methods are lacking for extracting physical and chemical parameters.

Purpose of the Study:

• To develop and integrate a novel machine learning framework for analyzing liquid-phase TEM videos.
• To enable quantitative extraction of nanoscale parameters from challenging imaging conditions.
• To investigate the dynamics of colloidal nanoparticles, including diffusion, reaction kinetics, and assembly.

Main Methods:

• Integration of liquid-phase TEM imaging with a customized analysis framework.
• Application of a U-Net neural network, trained on simulated TEM images with ground truth.
• Real-time, real-space analysis of colloidal nanoparticle systems.

Main Results:

• Successfully mapped diverse nanoscale properties of anisotropic nanoparticles (nanoprisms, nanorods, nanocubes).
• Revealed anisotropic interactions, curvature-dependent etching profiles, and a first-order chaining assembly law.
• Demonstrated U-Net's superior performance in nanoparticle segmentation from noisy TEM images compared to existing methods.

Conclusions:

• The developed framework significantly enhances the quantitative capabilities of liquid-phase TEM.
• Provides high-throughput, statistically significant insights into nanoscale dynamics of synthetic and biological nanomaterials.
• Overcomes limitations of traditional analysis methods for noisy and heterogeneous TEM data.