Compressing physical properties of atomic species for improving\n predictive chemistry

Abstract

The answers to many unsolved problems lie in the intractable chemical space\nof molecules and materials. Machine learning techniques are rapidly growing in\npopularity as a way to compress and explore chemical space efficiently. One of\nthe most important aspects of machine learning techniques is representation\nthrough the feature vector, which should contain the most important descriptors\nnecessary to make accurate predictions, not least of which is the atomic\nspecies in the molecule or material. In this work we introduce a compressed\nrepresentation of physical properties for atomic species we call the elemental\nmodes. The elemental modes provide an excellent representation by capturing\nmany of the nuances of the periodic table and the similarity of atomic species.\nWe apply the elemental modes to several different tasks for machine learning\nalgorithms and show that they enable us to make improvements to these tasks\neven beyond simply achieving higher accuracy predictions.\n