Property Prediction\nfor Complex Compounds Using Structure-Free\nMendeleev Encoding and Machine Learning

Predicting the properties for unseen materials exclusively\non the\nbasis of the chemical formula before synthesis and characterization\nhas advantages for research and resource planning. This can be achieved\nusing suitable structure-free encoding and machine learning methods,\nbut additional processing decisions are required. In this study, we\ncompare a variety of structure-free materials encodings and machine\nlearning algorithms to predict the structure/property relationships\nof battery materials. It was found that the physical units used to\nmeasure the property labels have an important impact on the predictive\nability of the models, regardless of the computational approach. Property\nlabels with respect to weight give excellent performance, but property\nlabels with respect to volume cannot be predicted with confidence\nusing only chemical information, even when the underlying physical\ncharacteristics are the same. These results contrast with previous\nstudies of unsupervised learning and classification, where structure-free\nencoding excelled, and highlight how the structural features or property\nlabels of materials are represented plays an important role in the\npredictive ability of machine learning models.