Machine Learning to Describe Structure-Property Relationships Within a Large Cluster Expansion Approach

The cluster expansion (CE) approach is used to describe and model the configurational site disorder of alloy materials systems and how changes in the local arrangement can affect materials' properties. When increasing the cell size used in the expansion, the number of non-degenerate configurations exponentially increases thus manually sampling the resulting configurational space becomes expensive when using first-principles density functional theory (DFT). In this work, we demonstrate the use of machine learning to model the structure-property relationship of 36-atom Be_(1-x)Zn_xO clusters by uniformly selecting 5% of all configurations and their dielectric functions from DFT as the training set. The model is used to predict the remaining 95% of total configurations and its accuracy is demonstrated by using thermodynamic averages compared to results from smaller supercells. We then discuss the best practices and considerations when applying this approach to other multicomponent systems like dataset size and curation followed by descriptor quality when making predictions.