Diffusion Model–Driven Denoising and Temporal-Aware 3D Reconstruction of Nanoparticles in Liquid-Cell TEM

Imaging nanoparticles in liquid environments in situ using transmission electron microscopy (TEM) is essential for directly probing dynamic processes such as nucleation, growth, and structural reconfiguration. However, liquid-cell TEM is often limited in signal-to-noise ratio and depth information due to low dose rate and projection constraints. Here we present a two-stage reconstruction pipeline that enhances both imaging contrast and structural interpretability. First, a self-supervised diffusion model is trained to denoise raw TEM frames, recovering fine morphological features without paired ground truth. Second, we introduce a temporal-aware machine learning method that incorporates inter-frame correlations and motion constraints to infer time-resolved 3D nanoparticle structures from 2D projections. This combined approach significantly improves spatial resolution and temporal coherence in nanoparticle tracking, offering a new avenue for quantitative analysis of dynamic systems imaged under liquid-cell conditions.