Machine learning and data science for materials discovery

Description

New materials hold great promise in addressing global challenges that our society faces today. For instance, green hydrogen as a sustainable transport fuel requires the development of materials for better electrolysers to accelerate the transition to clean transportation. Similarly, the realisation of better catalysts will minimise the impact of manufacturing the chemicals and materials that society needs. However, discovering these transformative materials is a challenging and time-consuming task that requires diverse expertise and is plagued by the need for trial-and-error. We need a better way, which means we need to develop better tools.

Ongoing automation of chemistry provide a principal way to overcome these challenges and accelerate material discovery. Nevertheless, it is very difficult for chemistry researchers to analyse the resulting multi-modal data and to select the subsequent experiments that ultimately will lead to new high-performing materials. This project is devoted to the development of such tools. You will work as an expert in computer science as part of an interdisciplinary team that includes chemists and materials scientists, who will use your tools in the development of new materials with the real-world impact. The PhD project is funded by Johnson Matthey for 3.5 years at the University of Liverpool.

Spectral data (or histograms) from powerful materials characterisation techniques such as X-ray diffraction or Raman spectroscopy provide the information on atomic-scale structure and composition needed to guide the discovery and optimisation of materials. However, these spectra can be noisy and are difficult for human researchers to assess at scale. In this project we will investigate self-supervised representation learning methods for spectral data as well as the broad set of deep learning tools in the context of automated discovery and data processing workflows. In particular, we will apply them to the problem of unmixing – the task of identification of constituent components of a given set of mixtures. This is a critical and generic problem for materials research. Mathematically, it can be seen as a noisy version of non-negative matrix factorisation with constraints, and by leveraging machine learning methods we aim to improve performance on this task.

This is a multi-disciplinary PhD that sits at the interface between Chemistry, Materials Science and Data Science. The student will be based at the Computer Science Department and in state-of-the-art laboratories in the Materials Innovation Factory (https://www.liverpool.ac.uk/materials-innovation-factory/) at the University of Liverpool, and will collaborate closely with researchers of Johnson Matthey.