Machine learning model for classification of light microscopic images of bainitic steel samples under different etching conditionsMittwoch (16.09.2020) 17:38 - 17:42 Uhr
Machine learning model for classification of light microscopic images of bainitic steel samples under different etching conditions
Amit Kumar Choudhary, Andreas Jansche, Gaby Ketzer-Raichle, Timo Bernthaler, Gerhard Schneider
Materials Research Institute, Aalen University, Aalen, Germany
The microstructural characterization of materials is challenging as the microstructure contains information related to its processing history, constituents and is responsible for its physical and chemical properties. The subjectivity of human experts leads to variability in microstructure classification when performed manually due to the presence of multiple phases and its complex underlying structures. Further, the appearance of the microstructure differs based on parameters such as heat treatment, etching condition, alloying elements, etc. Therefore, a robust and accurate classification of these microstructures is very important while hard to achieve.
Recently, the field of materials informatics has aimed to improve and accelerate the process of quality assurance where information technology and data are used to analyze material data to better understand materials characteristics and the relationship with its microstructure.
This work aims to leverage the role of machine learning and computer vision for the task of automatic microstructure classification of different bainite samples etched under varying conditions. We demonstrate how machine learning and feature engineering can be used for the automated and reproducible microstructure classification of bainite samples. The model extracts texture information from the optical light microscopic image of bainite samples to train a supervised learning model. This involves a materials expert for labeling the acquired images to generate a training dataset for the classification task. The tests for automatic microstructure classification were performed on two different tool steel samples having varying bainitic structures and performance assessment was done by comparing results against the expert based manual classification. The performance of the trained model for microstructure classification of new or reproduced bainite samples as well as the importance of feature engineering to improve the robustness and performance of the model will also be discussed.