Self-supervised machine learning framework for high-throughput electron microscopy

Abstract

Transmission electron microscopy (TEM) is a crucial analysis method in materials science and structural biology, as it offers a high spatiotemporal resolution for structural characterization and reveals structure-property relationships and structural dynamics at atomic and molecular levels. Despite technical advancements in EM, the nature of the electron beam makes the EM imaging inherently detrimental to materials even in low-dose applications. We introduce SHINE, the Self-supervised High-throughput Image denoising Neural network for Electron microscopy, accelerating minimally invasive low-dose EM of diverse material systems. SHINE uses only a single raw image dataset with intrinsic noise, which makes it suitable for limited-size datasets and eliminates the need for expensive ground-truth training datasets. We quantitatively demonstrate that SHINE overcomes the information limit in the current high-resolution TEM, in situ liquid phase TEM, time-series scanning TEM, and cryo-TEM, facilitating unambiguous high-throughput structure analysis across a broad spectrum of materials.