Neural network embeddings based similarity search method for atomistic systems

Abstract

With the popularity of machine learning growing in the field of catalysis there are increasing numbers of catalyst databases becoming available. These databases provide us with the opportunity to search for catalysts with desired properties, which could lead to the discovery of new catalysts. However, while there are search methods for molecules based on similarity metrics, for solid-state catalyst systems there is not yet a straightforward search method. In this work, we propose a neural network embeddings based similarity search method that is applicable for both molecules and solid-state catalyst systems. We illustrate how the search method works and show search examples for the QM9, Materials Project (MP) and Open Catalyst 2020 (OC20) databases. We show that the configurations found present similarity in terms of geometry, composition, energy and in the electronic density of states. These results imply the neural network embeddings have encoded effective information that could be used to retrieve molecules and materials with similar properties.