O. Badmos, A. Kopp, D. Hohs, R. Büttner, T. Bernthaler, G. Schneider
Hochschule Aalen,
Germany
Keywords: computer vision, microstructure, deep learning, convolutional neural networks, lithium-ion battery, sintered magnets, electrode defects, quality evaluation
Summary:
Microscopy methods have been successfully used for material characterization, quality assurance and evaluating the influence of production parameters in various field of material science (powder metallography, high performance ceramics and high strength steels). Recently, there has been a lot of research in the field of materials informatics, where information technology and data analytics are used to correlate and interpret material data in order to better understand the characteristics of known materials and to accelerate the discovery, design and development process of new materials. This work aims to leverage major advances of the past few years in computer vision and deep learning for the task of microstructure recognition. In the past, most pattern recognition tasks were performed using hand-crafted feature extraction followed by a classifier. A typically computer vision pipeline is a follows: Find interest points in an image, crop patches around them, represent each patch with a sparse local descriptor, and combine the descriptors into a representation of the image. Nowadays, state-of-the-art computer vision model perform these steps automatically using mainly convolutional neural networks (ConvNet) and approach human level performance on some image recognition task. In this presentation we demonstrate how deep learning, a subfield of machine learning can be used for the automated detection of microstructure defects in different types of material samples using two different learning techniques. In the first example we use state-of-the-art computer vision models for the automated quality assessment of large prismatic li-ion batteries from images of the micrograph. It has been shown in various studies that the performance of a li-ion battery is intrinsically linked to the electrode microstructure. Consequently, quantitative measurements of key structural parameters will enable the optimization as well as motivate systematic numerical studies for the improvement of the battery performance. The aim here is to evaluate various cell components (cathodes, anodes, and separators) in high-resolution from optical microscope images to automatically detect various types of defects such as metal particle contamination, cracks, layer deformation and non-uniform coating present in the battery microstructure. For this example we use a supervised machine learning technique, which involves learning a representation for mapping input data to known target values given a set of training examples (often labelled by humans). In the case of detecting defects in battery micrographs, this would involve a battery expert examining a large amount of images and labelling each one as either with defect or without defect. This of course can be time consuming since getting a substantial amount of images with defects enough for training a deep leaning algorithm can be quite a challenging task. As a result, such a problem can also be addressed through unsupervised learning, which consist of finding useful transformations of the input data without the help of any labels to better understand the correlations present in the data. Therefore, for the second example we use unsupervised learning (variational autoencoder and conditional generative adversarial networks) for the automated detection of defects in sintered magnet microstructure to demonstrate the approach.