An Information Oriented Approach to Materials Informatics

Machine learning and artificial intelligence (AI) in recent years have revolutionised materials sci- ence and the world at large. Through systematic and efficient retrieval of collective knowledge and, as some recent advances suggest, the ability to synthesise novel ideas from this knowledge, AI has ushered in a fourth industrial revolution. In materials science, machine learning has been deployed for various applications, from predicting material properties to discovering new materi- als, offering unprecedented high-throughput capabilities based on existing data. Machine learning is powerful, but there are fundamental limits to the performance of a model. The model is fun- damentally limited by the constraints of the architecture and the available data. A linear model can never model non-linear relationships, two inputs that map to the same representation can never deterministically give a different output, and garbage in will always mean garbage out. This thesis approaches machine learning in the spirit of these fundamental limits. This work begins by using existing representations of the composition of the materials to assist in the search for novel mate- rials. From there, the key focus is on the problem of representing crystal structures for machine learning, which led to the development of the cubic supercell representation and the Site-Net trans- former architecture. These innovations address the need for crystal machine learning pipelines that are unambiguous, invariant to symmetry, and information-complete. Further work addresses the challenge of creating compact vector representations of crystals with these same properties through self-supervised learning with the Deep InfoMax methodology that explicitly maximises mutual in- formation between crystal representations and a vector space. This information-driven approach to machine learning lays the groundwork for future advancements in materials informatics.