Structure Prediction of Molecules on Noble Metals
Andreas Jeindl
Institute of Solid State Physics, TU Graz
11:15 - 12:15 Wednesday 29 November 2017 PH01150 The type of surface structure that forms on a metal surface determines many properties of the thin films, such as their solubility or conductivity.
A crucial step towards designing functional organic-inorganic interfaces is therefore to understand and predict polymorphs that form at these interfaces. Unfortunately, finding low energy structures via ab-initio methods is far from trivial due to the vast number of possible polymorphs.
With the SAMPLE approach we tackle this problem with a combination of coarse-graining the PES and machine learning.
To that end, we first determine isolated adsorption geometries and combine them to a set of possible "guess polymorphs".
We then use a small subset of these polymorphs to learn the molecule-slab and molecule-molecule interactions.
After learning these interactions we can predict the formation energies of all polymorphs within our discretized system. Compared to "brute force" screening of the possible polymorphs, this can reduce the required time by several orders of magnitude.
Yet even this approach is pushed to the limits for large, technologically interesting molecules.
In this talk, I will explain the challenges for structure search and discuss several improvements, that allow to extend the range of our method to realistic molecules. Specifically, I improved three stages of our approach:
The search of local geometries for larger molecules was enabled by efficient search on a subspace of the PES.
The computational cost was reduced by introducing a mixed-basis set slab.
Our prior used for the pairwise molecule-molecule interaction was improved by calculating the electrostatic and Van der Waals contributions.
|
Institute of Solid State Physics
