Data-driven approaches are particularly useful for computational materials discovery and design as they can be used for rapidly screening over a very large number of materials, thus suggesting lead candidates for further in-depth investigations. A central challenge of such approaches is to develop a numerical representation, often referred to as a fingerprint, of the materials. Inspired by recent developments in cheminformatics, we propose a class of hierarchical motif-based topological fingerprints for materials composed of elements such as C, O, H, N, F, etc., whose coordination preferences are well understood. We show that these fingerprints, when representing either molecules or crystals, may be effectively mapped onto a variety of properties using a similarity-based learning model and hence can be used to predict the relevant properties of a material, given that its fingerprint can be defined. Two simple machine-learning-based procedures are introduced to demonstrate that the learning model can be inverted to identify the desired fingerprints and then to reconstruct molecules which possess a set of targeted properties.

• Received 26 March 2015

DOI:https://doi.org/10.1103/PhysRevB.92.014106