Domain-specific simulated data enhances knife-mark noise suppression in microscopy images of materials.

Accurate quantitative analysis of material microstructures from images is often hindered by noise and artifacts generated during sample preparation. While deep learning is a promising approach for this challenge, preparing the large amount of 'supervised data' (labeled real images) required for training poses a significant barrier in material science. This study proposes and validates a simulation-driven learning paradigm where a deep learning model is trained exclusively on simulated images that mimic the key features of target structures and noise, serving as a powerful solution to this data scarcity problem. As a specific case study, we applied this paradigm to the removal of 'knife-mark noise' from cross-sectional images of rubber materials to enable accurate filler region segmentation. In evaluations using simulated data, the proposed method showed superior performance across all the metrics (PSNR, SSIM, and MAE) compared with conventional methods such as the median filter and TV reconstruction, as well as a U-Net model trained on general-purpose Gaussian noise. More importantly, the model also performed effectively on real images, despite being trained solely on simulated data. It successfully removed both knife-marks and material-derived background textures, which demonstrates the viability of simulation-driven learning to overcome the need for manually annotated datasets. This work highlights the power of task-specific simulations as a practical alternative to manual data annotation in quantitative materials analysis.