There is currently no way of digitally capturing the essence of a microstructural fingerprint: a set of descriptors that encapsulate those features in a microstructural image that result from mesoscale (i.e. larger than nano- and smaller than micro-scale) structure in a material, and on which properties of the material depend. This project will survey, develop, and implement machine learning methods for data compression, feature extraction and dimension-reduction on populations of images of material microstructure from the Henry Royce Institute. The aim will be to digitally fingerprint microstructure images such that these fingerprints can be incorporated into materials processing and property optimisation methods, using machine learning or other computational science tools.
Explaining the science
The primary mathematical methods employed in this project will generally fall into the categories of (i) factor models such as principal and independent component analysis (P/ICA), and their nonlinear counterparts via kernel embeddings, and (ii) deep learning methods, such as autoencoders, and their probabilistic generative counterparts.
The project will also explore the potential for topological data analysis and other emerging methodologies. This will identify key opportunities, challenges, and open questions which are fit to the purpose of digitally fingerprinting the data sets being generated at the Henry Royce Institute.
In medicine, geology and materials a picture may well be worth a thousand words. This explains why companies, life scientists, medics, earth and materials scientists invest in microscopes, X-ray scanners, etc. Despite massive advances in the tools used to record images there is currently no way of digitally capturing the essence of a microstructural fingerprint.
Great progress is being made in digital recognition in histology (the study of the microscopic structure of tissues) but a larger multi-disciplinary cross council effort could transform the digital interpretation of images across many sectors.
Being unable to describe a microstructure digitally is a major roadblock for the storage, analysis and exploitation of data across the life, materials and earth sciences. If a language can be established to describe the features making up a microstructure, this would enable researchers to take full power of digital techniques such as machine learning to:
- Compactly store the state of a material digitally
- Provide a framework for new data to be generated, reported and made publicly available
- Be able to digitally monitor materials supply chains
- Make links between features and certain properties, processes, or outcomes
- Learn which features are associated with certain behaviours to optimise existing materials properties and processes, and invent new ones
The exploratory nature of this project will open up the pathway to design new innovative methods which are fit to the purpose of digital fingerprinting of microstructures from the latest types of materials images being generated at the Henry Royce Institute.
The need to turn images into (condensed) information is core to many imaging modalities. The methodology will be broadly applicable well beyond materials science and engineering, to chemistry, biology, medicine, petroleum engineering and beyond.