DecurtainNet, A Deep Learning Model and Large-Scale Dataset for Fast Curtaining Artefact Feature Removal in Electron Microscopy

Abstract

Focused Ion Beam (FIB) milling enables electron microscopy (EM) studies of mate rial across diverse applications. However, the use of FIB on composite or beam-sensitive materials can cause curtaining artefacts compromising EM data quality. Existing mit igation approaches, such as optimised FIB acquisition parameters and Fast Fourier Transform (FFT), require costly instrument adjustments and time-intensive tuning, lim iting high-throughput use. Here, we present DecurtainNet, a UNet-based model that automatically removes curtaining artefact features from EM images rapidly (∼0.03 s per image). Developed using an in-house dataset of over 22,000 manually processed images from diverse EM techniques and samples, DecurtainNet demonstrates superior general isability and performance. It ensures the preservation of images otherwise discarded due to severe artefact appearance and enhances the clarity and depth of downstream EM tasks, assisting EM material property analysis. Additionally, the individually, manually processed dataset, derived from real-world experiments, holds substantial potential to advance the development of EM decurtaining tools for the community