A.K. Choudhary, A. Jansche, O. Badmos, T. Bernthaler, G. Schneider
Aalen University,
Germany
Keywords: machine learning, correlative microscopy, material characterization, microstructure analysis, EBSD
Summary:
Material characterization is one of the major challenges faced in the field of materials research. The general approach is the assessment of quantitative properties, which are dependent on the use of destructive/non-destructive techniques. Of these, some non-destructive methods are based on the analysis of the micrographs obtained from the microscopy are very popular and used extensively. Conventional methods require the user to manually assess the micrographs obtained to identify the microstructural features followed by physical tests to quantify properties and obtain an overall characterization. This approach requires experienced personnel for effective and complete results, is also tedious, time-consuming and does not yield consistent results. Thus, the automation of this process, based on digital image processing and analysis of micrographs has been studied and shows good first results. However, this approach lacks robustness to give consistently good results. A recent development in this area is the use of the concept of machine learning (ML) and deep learning (DL) in image analysis. Research over the years in this area has resulted in the development of stable, robust and reliable systems, which yield consistently good results. These systems are also capable of handling and analyzing large data sets associated with complex materials. Quantitative image analysis for the measurement of properties and controlling process parameters is vital for new material development and improving quality. Using advanced methods of image analysis such as machine learning to find the relationship between microstructural characteristics and material properties is crucial. Due to complex microstructural features of some materials, it is not a feasible option to use traditional approaches of image analysis for extracting quantitative information from micrographs. For materials such as magnets, which play a vital role in advancement of scientific growth and has wide applications like for data storage, power generation and transmission, sensors etc. One of the techniques to study the domain structure of magnetic materials is magneto-optical imaging which has the advantage of being a relatively inexpensive, non-invasive and its ability to handle a large range of magnetic samples. Because magnetic domain images from kerr microscopy have a lot of information that has a direct impact on the properties of the materials, the application of machine learning here could assist materials researchers in solving some of the complex fundamental tasks. The microstructure classification of micrographs obtained from kerr microscopy using traditional image analysis approach fails due to the presence of optically similar regions and varying domain patterns. Using higher dimensional feature set, a machine learning model is trained to classify different microstructures from kerr images of Nd-Fe-B sinter magnet. Later the output from trained model is used for quantifying grain sizes and to predict domain orientation in the each grains using correlative microscopy (kerr microscopy and EBSD data). The obtained results are compared with reference data, which is produced manually by experts.