Regensburg 2019 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 12: Poster session I

MM 12.22: Poster

Montag, 1. April 2019, 19:15–20:45, Poster C

Smart-data machine learning for surface polymorph search — •Andreas Jeindl, Lukas Hörmann, Alexander T. Egger, and Oliver T. Hofmann — Institute of Solid State Physics, NAWI Graz, Graz University of Technology, Austria

The major challenge of surface structure search is the large number of possible polymorphs. The SAMPLE approach [1,2] can help circumvent this problem. It combines a coarse-grained potential energy surface with Bayesian linear regression to efficiently predict the adsorption energies of an exhaustive set of commensurate organic monolayers.

In this contribution we present the SAMPLE approach. The first step is finding all local minima for a single molecule on the surface. We then create all possible combinations of these local minima up to a certain unit cell size. With the help of ’experimental design’ theory, we select a subset of maximally diverse structures. This subset is evaluated with dispersion-corrected density functional theory and used as training set for a Bayesian linear regression model. The linear regression uses an energy model based on the assumption that the main interactions on the surface are governed by molecule-substrate interactions and pairwise molecule-molecule interactions. Thus, we can not only predict the adsorption energies of millions of possible polymorphs, but also gain meaningful physical insight.

We use three complementary molecules on coinage metals to showcase the capabilities of our approach.

[1] Scherbela et al., Phys. Rev. Materials 2, 043803

[2] Hörmann et al., arXiv:1811.11702