Reduction of Membrane-derived Noise Using Beam-tilt Measurement and Deep Learning in Observation using Environmental Cell.

Abstract

Electron microscopy using an environmental cell is a powerful tool for observing catalysts and other nanomaterials in gases and liquids. An environmental cell must contain amorphous silicon-nitride membranes because they protect the sample environment from the vacuum of the electron microscope and enable the electron beam to pass through the cell. However, the membranes superimpose non-uniform contrast on the projected image, degrading image quality. We propose a method for removing the noise derived from the membranes using Noise2Noise, a deep-learning method, for a series of transmission-electron-microscope images with slight electron-beam tilt and evaluated its effectiveness. We succeeded in removing the membrane-derived noise while retaining the information of the sample in the cell. We also succeeded in efficiently removing Poisson noise. We believe this method will enable measurements requiring high signal-to-noise ratios, which could previously only be observed in a vacuum, to be conducted in an environmental cell.