Semi-supervised machine learning workflow for analysis of nanowire morphologies from transmission electron microscopy images

Abstract

In the field of materials science, microscopy is the first and often only accessible method for structural characterization. There is a growing interest in the development of machine learning methods...

Emerging Trends in Machine Learning: A Polymer Perspective

“…We refer the reader to an excellent general review on the topic . Within polymer science, transfer learning is increasingly being used. ,− For example, Li et al used it to reconstruct microstructures and generate structure–property predictions for nanocomposites. In this particular case, they use a deep convolutional neural net trained on a nonscientific corpus for their source domain .…”

“…Then they transferred this encoder to perform other tasks such as morphology classification and nanowire segmentation. Ultimately, for morphology classification, they found they needed less than 10 labeled images per class and, if the underlying distribution was known a priori, only a single labeled image per class was necessary . These results are particularly exciting, since manual labeling of data is time-consuming and error prone.…”

nanoNET: machine learning platform for predicting nanoparticles distribution in a polymer matrix

“…We aim to address this challenge via machine learning. Recently, machine learning has shown tremendous success in solving a wide range of materials problems, such as the identification of the local structure of metallic NPs on an oxide-support and defective graphene sheets; 25 the automated recognition of metal NPs deposited on pyrolytic graphite; 26 atomic structure classification and prediction; 27–29 nanostructure identification from X-ray, SEM (scanning electron microscopy), TEM (transmission electron microscopy), and SAS (small-angle scattering); 30–34 and classification and prediction of sequence-defined morphologies of copolymers. 35–37 Machine learning methods have also been used recently to predict the properties of PNCs, 38–41 such as their dielectric constant, rubbery modulus, and glassy modulus.…”

Pair-Variational Autoencoders for Linking and Cross-Reconstruction of Characterization Data from Complementary Structural Characterization Techniques

“…The richness of metadata associated with each microscopy image also implies limitation on the curation of large microscopy databases because a generic (universal) labeling scheme would be nearly impossible for microscopy images coming from diverse sources. Furthermore, a unique problem to the (nonbiological) materials field is the smaller dataset sizes for the microscopy images collected from samples due to the laborious sample preparation required to acquire each image; this problem restricts the use of deep learning models that require large training datasets like those available in biomedical fields. , To address this challenge with small datasets of materials’ microscopy images , Lu et al recently proposed a semi-supervised transfer learning workflow that successfully classified the materials’ morphologies from TEM images of protein/peptide nanowires despite being trained with fewer than 10 labeled images per morphology.…”