Deep Learning Boosts Quality of Millisecond-Scale STEM Imaging
Researchers have developed a deep learning–based noise-filtering framework that enables high-speed scanning transmission electron microscopy (STEM) imaging with significantly improved quality, even at millisecond-per-frame acquisition rates. Rapid STEM scans typically suffer from low signal intensity, statistical noise, and unidirectional blurring due to hardware limitations, making real-time observation of dynamic processes challenging. The team trained a convolutional neural network (U-Net) on distortion-corrected datasets to learn and remove both noise and scanning artifacts that conventional filters fail to address. The new method restores high-fidelity images without requiring hardware modifications, enabling fast, high-quality STEM imaging suitable for in situ studies of materials dynamics such as defect motion. This approach could open new possibilities for materials science studies that rely on high temporal and spatial resolution electron microscopy.

Journal: Scientific Reports

DOI: 10.1038/s41598-022-17360-3