SurfaceScience21 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 55: Poster Session IV: Poster to Mini-Symposium: Machine learning applications in surface science II
O 55.4: Poster
Dienstag, 2. März 2021, 13:30–15:30, P
Configurational polaron energies using machine learning — •Viktor Birschitzky, Michele Reticcioli, and Cesare Franchini — University of Vienna, Faculty of Physics
Polarons are quasiparticles formed by the coupling of excess charge carriers with the phonon field. Polarons form preferentially at surfaces and have a wide range of effects on the chemical and physical properties of the hosting material.1 First principles calculations of polarons conformational energies typically require large supercells and long molecular dynamics (MD) simulations, making the modeling of multipolaron system within reasonable timescales very challenging. Here, we propose a supervised machine learning scheme based on kernel-regression to solve this problem by learning single polaron energies for the prototypical oxygen-defective rutile TiO2−x(110) surface, where each oxygen vacancy provides two excess electrons. To achieve accurate predictions on an ab initio MD database of polaronic energies2 a descriptor has been developed, which embodies the interactions between polarons with defects and other localized charge carriers. Our results show that the proposed ML method is able to expand the DFT database with energetically more favorable polaron configurations – improving the convex hull construction – and that generalization at arbitrary polaron concentration and defect types is possible.
[1] C. Franchini et al., Polarons in Material, Nature Review Materials, (2021)
[2] M. Reticcioli et al., Formation and dynamics of small polarons on the rutile TiO2 surface, Physical Review B, (2018)