Smart Data Machine Learning for Surface Structure Search

Hörmann, L. (Redner/in), Jeindl, A. (Beitragende/r), Alexander Egger (Beitragende/r), Michael Scherbela (Beitragende/r), Hofmann, O. (Beitragende/r)

Aktivität: Vortrag oder PräsentationPosterpräsentationScience to science


Monolayers of organic molecules on inorganic substrates often show rich polymorphism with diverse structures in differently shaped unit cells. Determining the different commensurate structures from first principles is far from trivial due to the large number of possible polymorphs. We introduce the SAMPLE approach[1], which employs physically motivated, coarse-grained modeling in combination with Bayesian linear regression to efficiently explore the potential energy surface.

SAMPLE works as follows: First, we determine the adsorption geometries isolated molecules adopt on the substrate. Then, we arrange these geometries in a large number of different ways to generate possible, commensurate polymorphs. For a subset of these polymorphs, we determine the adsorption energies using dispersion-corrected density functional theory (DFT). This subset serves as training data for an energy model, consisting of inter-molecule and molecule-substrate interactions. Bayesian linear regression enables determining these interactions and thereby allows to not only gain meaningful physical insight but also to predicting the adsorption energies for millions of possible polymorphs.

We demonstrate these capabilities for naphthalene on Cu(111), TCNE on Cu(111) and benzoquinone on Ag(111).

[1] Scherbela et al., Phys. Rev. Mat. 2, 043803 (2018)
Zeitraum12 Jul 2018
EreignistitelInterfacing Machine Learning and Experimental Methods<br/>for Surface Structures (IMPRESS)
OrtPetersgasse 16, 8010 Graz , Österreich