Over the past several years, we have introduced a family of combinatorially designed semi-local and hybrid density functionals. A small fraction of a large amount of validated ab initio reference data is used to train a very large number (many billions) of candidate density functionals. Candidates with unphysical characteristics or with numerical instabilities are eliminated, and the remaining functionals are assessed on their predictive performance acorss a range of chemically relevant energy differences to allow the selection of single form that is representative of the highest accuracy possible. We refer to this approach as "survival of the most transferable". In this talk, we will present the results of applying this procedure to the design of a double hybrid density functional (DHDF). A DHDF adds a single shot non-local correlation correction of second order perturbation theory form to correct the energy obtained by a self-consistent Kohn-Sham calculation using a functional that does not depend on the Kohn-Sham virtual orbitals. Well-known examples of this type include B2PLYP, XYG3, DSD-BLYP, and ωB97X-2. Previously reported tests on chemical datasets suggest that these DHDFs yield higher accuracy than widely used hybrid functionals. Can the survival of the most transferable approach enable the design of a DHDF of higher accuracy than the best existing options? Is the difference chemically significant? Can ideas from machine learning be applied to improve or accelerate the associated combinatorial search problem? These are a few of the main questions that will be addressed in this talk.

Combinatorial design and assessment of a new double hybrid density functional

255th National Meeting (2018)

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