Patch-Based Class-Conditioned Diffusion Augmentation of TEM Images: Effects of Training Patch Size on Dispersion-Relevant Morphology in Epoxy Nanocomposites

Abstract

Transmission electron microscopy (TEM) is widely used to evaluate nanoparticle dispersion in polymer nanocomposites, which strongly influences their dielectric and insulation performance. However, the limited availability of TEM images restricts data-driven analysis. In our previous work, we established a framework to extract dispersion-related descriptors from TEM images and relate them to three-dimensional dispersion states. This study investigates the feasibility of augmenting limited TEM datasets using a patch-based, class-conditioned diffusion model and clarifies the effect of training patch size on reconstruction fidelity. Three TEM images (640 × 640 pixels) with different filler sizes and dispersion states were used. Training datasets were generated by cropping patches of various sizes (16–512 pixels), followed by rotation and flipping. A class-conditioned denoising diffusion probabilistic model with a U-Net backbone was employed to generate synthetic images. Quantitative evaluation was conducted using particle area fraction, Feret diameter distribution and Delaunay-based interparticle distance. An integrated error metric revealed a nonmonotonic dependence on patch size for all three samples, with the largest errors at the smallest patch size and an interior minimum in every case. The error varied only slowly near the minimum in each case, so that the result is better described as a range of comparable patch sizes than as a single optimum; for the samples examined here the position of this range was larger for the sample with the larger modal Feret diameter of the segmented features. These results provide an initial, empirically derived basis for selecting the patch size in diffusion-based augmentation of TEM images, subject to the limited number of source images used here.