Predicting battery applications for complex materials based on chemical composition and machine learning

Materials informatics uses machine learning to predict the properties of new materials, but generally requires extensive characterisation and feature extraction to describe the input data, which can be time consuming and expensive. Predicting properties or classes of materials based on minimal input information, such as a chemical formula, can be a useful first step to identify which materials are promising candidates before investing resources. This is particularly desirable when working with complex compounds containing a large variety of elements, such as materials for battery applications. In this paper we show how to classify battery compounds into either charge or discharge formulas, or identify suitable anode or cathode materials, based exclusively on the chemical formulas of materials available in online repositories. Without any structural information, we train high-performing classifiers that can be used to rapidly screen hypothetical materials and assign potential applications. The models are applied to a total of 471 materials from the literature, and deliver a 96% success rate over 80% probability. These methods are general and the workflow can be applied to any complex crystalline materials to predict end-uses in advance of synthesis or simulation, opening up the opportunity for machine learning to use used for research planning, in addition to prediction or inference. • Machine learning reproduces domain knowledge based on minimal information. • Accurate models of complex materials without extensive structural characterisation. • Predicting applications of materials in advance of synthesis or simulation.