Unsupervised learning for materials discovery by quantifying materials analogies

What is it about?

We propose a new concept of analogical materials discovery, enabled by unsupervised machine learning. It is found that the material embeddings extracted by an unsupervised deep learning model contains information about crystal structure of materials, although the model is only provided with composition-based features. This phenomenon is capitalised to discover new materials that are analogous to existing high performing, yet hazardous or expensive materials. The framework is demonstrated for compositionally-disordered perovskite oxide type materials that are of great technological importance.

Featured Image

Photo by JJ Ying on Unsplash

Photo by JJ Ying on Unsplash

Why is it important?

Target-driven discovery of materials is non-trivial especially for compositionally complex materials such as disordered ceramics. Such materials can be extremely computationally demanding to be modelled using first principal methods like density functional theory (DFT). Therefore, a machine learning framework that can bypass the computational complexity of DFT could be vital in rapid discovery of new materials.