Refinement and Processing of Steel Microstructure Images Facilitating Automated Heat Treatment Process Prediction

Abstract

Steels are alloys of iron and carbon (with up to 2.1 weight percent), widely used in construction, electronics, automotive and transport, energy, packaging, containers, appliances, etc. The variety of properties that the different grades of steel (used in these different industries) possess, is high. Different applications and various industries require different physical, mechanical and chemical characteristic features of steels, which necessitates making of different grades of steel. The application specific properties or features of steels are obtained by subjecting the metallic alloy to thermomechanical, or heat treatment procedures. Majority of these heat treatment procedures in the steel industry generally involve a lot of human resources as automation in this sector still requires more advancement. Also, these processes are highly resource intensive and time consuming. If because of these manual errors, the desired grade/quality is not achieved in the final product, it leads to a lot of wastage of resources. To curb this, automated or computerized simulations of these heat treatment procedures are increasingly being considered as an alternative. In this thesis the challenge of digitizing plain carbon steel microstructure images is addressed, which would aid in the automation of steel heat treatment processes in the metallurgy industry. This kind of automation will help in quicker novel material discovery and facilitate easier inverse design of materials, i.e., designing materials given a set of desired properties. Multiple neural network architectures for the refinement (denoising and segmentation) of raw steel microstructure images are developed, suitable for the above purpose. During the process of building optimal deep learning architectures, it was found that generative models are better suited for handling the variable types and amounts of noise found in plain carbon steel microstructure images. Also, using generative models for microstructure image refinement tasks like denoising and segmentation, helps us bypass deploying very deep convolutional neural network models, which have high computational demands, and take more time to converge/train successfully. Finally, in this thesis, a machine learning framework is elucidated, for the accurate prediction of a suitable thermomechanical treatment process, in order to attain a target grade of steel from a given initial steel sample. Conventional image featurebased or texturebased tools, when combined with machine learning techniques prove to be very useful in microstructure classification/prediction. Deep learning based classifiers have also been used, and the conventional machine learning techniques report comparable performance with these stateoftheart deep learning models.