Crystal Group Prediction for Lithiated Manganese Oxides Using Machine Learning

Abstract

This work aimed to predict the crystal structure of a compound starting only from the knowledge of its chemical composition. The method was developed to select new materials in the field of lithium-ion batteries and tested on Li-Fe-O compounds. For each testing compound, the correspondence with respect to the training compounds was evaluated simply by calculating the Euclidean distance existing between the stoichiometric coefficients of the elements constituting the two compounds. At the compound under test was assigned the crystal structure of the training compound for which the distance value was minimum. The results showed that the model can predict the crystalline group of the test compound with an accuracy higher than 80% and a precision higher than 90%, for a cut-off distance higher than four. The method was then used to predict the crystalline group of manganese-based compounds (Li-Mn-O). The analysis conducted on twenty randomly selected compounds showed an accuracy of 70%. Out of ten valid predictions, nine were true positives, with a precision of 90%.

Keywords

• crystal structure prediction
• machine learning
• K-nearest neighbours
• lithium-ion battery
• cathodes
• iron