Developing Deep Neural Network-based Denoising Techniques for Time-Resolved In Situ TEM of Catalyst Nanoparticles

Abstract

Recent advancements in the realization of highly efficient shot noise-limited direct detectors now enable atomically-resolved in situ TEM image time-series to be acquired with temporal resolutions in the millisecond (ms) regime [1]. Many catalysts exhibit turnover frequencies on the order of 10 – 10 sec, so the opportunity to visualize atomic behavior with high time resolution holds much promise for understanding the chemical transformation processes occurring on catalyst surfaces. Unfortunately, acquiring in situ TEM time-series with ~ms temporal resolution necessarily produces datasets severely degraded by shot noise [2]. For typical atomicresolution in situ TEM imaging conditions, at high frame rates the average dose in each frame can be < 1 e per pixel. Following Poisson statistics, counted images with an average dose < 1 e per pixel necessarily have a signal-to-noise ratio less than unity, and consequently, ascertaining the structure in the image becomes a major obstacle. There is a pressing need for sophisticated noise reduction techniques that both (1) preserve the temporal resolution of the image time-series and (2) facilitate the retrieval of atomic-level structural features at the aperiodic catalyst surface.