Updating Paulings rules using a machine learning approach

Abstract

Oxides are an important family of materials that have an extremely wide range of applications in for example semiconductors, pigments and catalysis. It is therefore important to have a solid understanding of these ubiquitous materials. In 1929 Linus Pauling proposed five rules for oxide stability that are widely used. These rules are however not good enough to describe oxide stability as only a fraction of stable oxides fulfil them. In this project a machine learning approach was used to attempt to find better rules based on the composition of oxides. This was done by training a set of autoencoders and analysing the latent spaces of these models by sampling new compositions from the models. Three different autoencoders were trained and based on the results, three new rules of thumb are proposed; Oxides containing only reactive non-metals are in general unstable, metals favour stability and heavier cations favour stability.