The surge in interest in nanomaterials in the past decade is ascribable, in large part, to the specialized properties due to size, shape, and structure at the nanoscale. Catalysts are a good example of this almost atom-by-atom dependence, with the number and coordination of each atom within a particle impacting the performance. A recent study even has shown that the simple rearrangement of 25 gold atoms from a spherical to cylindrical shape can significantly impact the efficacy of that catalyst [1]. The length scale of such materials makes characterization of these materials more difficult. Most studies must rely on ensemble measurements such as powder x-ray diffraction (XRD) or small angle x-ray scattering (SAXS) and assume a homogenous population such that the ensemble measurement is representative of each particle. This means that information on structural heterogeneity in samples is completely lost. The only technique which can give local structural information on a nanoparticle by nanoparticle basis is high-resolution transmission electron microscopy (HRTEM) or high resolution scanning transmission electron microscopy (HR-STEM). HRTEM is generally extremely low throughput due to constraints on data analysis. How to make HRTEM a more high-throughput process has become a goal for the TEM community and has large implications for the understanding of structure-property relationships in nanomaterials. Currently, no method exists to analyze HRTEM data in a high throughput manner across a large range of samples. The recognition of defects and shape require semantic segmentation (partitioning of the image on a pixel by pixel basis) which can be extremely difficult due to the low signal to noise ratio in most HRTEM images. Recent advances in convolutional neural networks (CNNs) for computer vision have made advanced segmentation problems solvable [2-5]. However, only in the last year have CNNs started to be applied to electron microscopy. Both Ziatdinov et
Paper
Full text
Machine Learning for High Throughput HRTEM Analysis
Semantic Scholar · Materials Science · 2019
Abstract
The surge in interest in nanomaterials in the past decade is ascribable, in large part, to the specialized properties due to size, shape, and structure at the nanoscale. Catalysts are a good example of this almost atom-by-atom dependence, with the number and coordination of each atom within a particle impacting the performance. A recent study even has shown that the simple rearrangement of 25 gold atoms from a spherical to cylindrical shape can significantly impact the efficacy of that catalyst [1]. The length scale of such materials makes characterization of these materials more difficult. Most studies must rely on ensemble measurements such as powder x-ray diffraction (XRD) or small angle x-ray scattering (SAXS) and assume a homogenous population such that the ensemble measurement is representative of each particle. This means that information on structural heterogeneity in samples is completely lost. The only technique which can give local structural information on a nanoparticle by nanoparticle basis is high-resolution transmission electron microscopy (HRTEM) or high resolution scanning transmission electron microscopy (HR-STEM). HRTEM is generally extremely low throughput due to constraints on data analysis. How to make HRTEM a more high-throughput process has become a goal for the TEM community and has large implications for the understanding of structure-property relationships in nanomaterials. Currently, no method exists to analyze HRTEM data in a high throughput manner across a large range of samples. The recognition of defects and shape require semantic segmentation (partitioning of the image on a pixel by pixel basis) which can be extremely difficult due to the low signal to noise ratio in most HRTEM images. Recent advances in convolutional neural networks (CNNs) for computer vision have made advanced segmentation problems solvable [2-5]. However, only in the last year have CNNs started to be applied to electron microscopy. Both Ziatdinov et