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

MM: Fachverband Metall- und Materialphysik

MM 10: Poster Session 1

MM 10.2: Poster

Montag, 5. September 2022, 18:00–20:00, P2

Transfer learning on organic/inorganic interfaces for different substrates — •Elias Fösleitner, Johannes Cartus, Lukas Hörmann, and Oliver T. Hofmann — Graz University of Technology, Graz, Austria

Performing structure search of organic molecules on metallic surfaces requires finding the structure with the lowest energy. Using conventional density functional codes this proves to be a time-consuming task, since the number of possible configurations is large and individual calculations are expensive. For all-electron approaches, this becomes even more problematic when calculating molecules on metal substrates of higher nuclear charge number, e.g. on gold. To circumvent the computation of all possible configurations, machine learning techniques such as Gaussian process regression proved to be a useful tool to reduce the amount of calculated data.

In our work we further reduce the data requirements by using transfer learning from one substrate to another. To this end, we first train the system on substrate A, and use this information to accelerate the learning process of the system on another substrate B. This is done by using the energy predictions of substrate A as a prior for the machine learning model imposed on system B. By doing so, we can reduce the data requirements for the training of expensive systems to an extent that makes the investigation computationally feasible.