Mapping the complex landscapes of materials and molecules
© 2016 EPFL
With atomistic simulations of materials reaching unprecedented levels of complexity, it becomes crucial to rationalize the relations between different compounds and phases, and to unravel structure-property relations. Can machine-learning algorithms help us build maps to navigate such complex landscapes?
In the recently-published article "Comparing molecules and solids across structural and alchemical space", we introduce a metric to determine the similarity between different molecules that encodes physical-chemical intuition on the relations between different structures and compositions. This metric is based on the combination of local descriptors (the SOAP kernels introduced by Gábor Csány, Albert Bartók and collaborators) that are matched between the two structures being considered. We show that this approach, combined with non-linear dimensionality reduction techniques provides insightful representation of the structural and compositional landscape of materials and molecules, and when used together with a kernel ridge regression algorithm makes it possible to predict with unprecedented accuracy the properties of molecules, circumventing expensive quantum chemical calculations.
Dr. Sandip De was supported by the NCCR MARVEL during the preparation of this article.