The last few years have seen much success of machine learning (ML) models in physical chemistry. The accuracy of these models relies heavily on the choice of the molecular representation (descriptor). Among the various descriptors proposed by different groups, the Coulomb matrix (CM) and its derivative Bag of Bonds (BoB) are  prominent, however, both are still amenable to improvement. We present an improved representation based on expanding the total energy of the system in interatomic many-body (MB) interaction potentials including one-body (BoA), two-body (BoP), three-body (BoT) and higher order terms and treat every possible single MB term as a variable in one dimension. When considering only BoA+BoP, with BoA and BoP being the respective diagonal and off-diagonal elements in an atom by atom matrix, the resulting descriptor performs as well as the {\em ad hoc} BoB descriptor. Inclusion of the three-body BoT term systematically improves the learning rate, with predicted errors lowered across molecular properties of thousands of organic molecules. We will briefly discuss results obtained for varying functional forms of BoP and BoT, and we will conclude with an outlook on 4-body (BoQ) based representations. <br/>

Many-body representations for machine learning models of molecular properties

252nd National Meeting (2016)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.