Developing a "microstructural fingerprint" of titanium alloys - metallurgy in the information age

Abstract

Microstructure characterisation plays a crucial role in metallurgy for many applications, such as mechanical property prediction and quality control procedures. However, traditional methods often involve limited metrics applied to image data that neglect vast amounts of information and are specific to the type of microstructure in question. This results in microstructural descriptors that potentially mask some features when comparing microstructures and can result in non-comparable features in some cases. Developing general, concise and quantitative representations of microstructure, known as ``microstructural fingerprints'', will enable more robust characterisation, utilising all available image data. This is a critical step towards harnessing data-centric materials science and capitalising on the large volumes of data that can now be collected in a relatively small timeframe. Advances in computer vision and machine learning (ML) have provided a new set of tools for extracting feature information from image data, but this has yet to be fully exploited within materials science - particularly when regarding industrial applications. Suitable fingerprints could be leveraged to enable more rigorous, traceable quality control and vastly reduce alloy development time. Here, we explore a range of ML-based methods for feature extraction, including keypoint-based methods and convolutional neural networks (CNNs), and assess their capability to describe microstructure via a variety of classification tasks. The embedded feature space of CNNs is explored through the use of variational autoencoders (VAEs) and an industrial application is presented to demonstrate the effectiveness of fingerprinting towards more robust quality control.

Date of Award	20 May 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Mia Maric (Supervisor) & Philip Withers (Supervisor)