The Polymer Society of Korea

Abstract

Real-time observation of soft matters in solution is crucial for understanding structural evolution and assembly pathways, as growth direction, binding orientation, and rotational motion directly influence morphology control and surface engineering. While, liquid-phase transmission electron microscopy (LP-TEM) enables nanoscale visualization of solution-phase behavior with high spatial, temporal resolution yet conventional TEM produces 2D projected images, limiting access to vertical information. This constraint becomes critical for soft matters with intrinsic low contrast, hindering analysis of particles approaching or growth parallel to the electron beam. Here, we report a deep learning-based strategy to enable 3D tracking of soft matters in real time. We extract point spread functions from focal stack datasets acquired at varying defocus to train the DL model that predicts height and segment x, y coordinates. With increasing volume of training dataset, we enabled distinguishing individual soft matters, simultaneously segmentation and pixel-wise height regression, enabling analysis of directional approach. This 4D microscopy capability combining 3D spatial coordinates with temporal resolution provides direct access to direction-dependent assembly, which would facilitate control over 3D nanostructures and surface engineering through solution-state assembly of soft matters.