SKM 2023 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 12: Poster I
MM 12.14: Poster
Montag, 27. März 2023, 18:15–20:00, P2/OG1+2
Efficient graph neural networks for accurate interatomic potentials between surfaces and adsorbed atoms — •Nian Wu1, Fabio Priante1, Eric Kramer Rosado1, and Adam S Foster1,2 — 1Department of Applied Physics, Aalto University, 00076, Espoo, Finland — 2WPI Nano Life Science Institute (WPI-NanoLSI), Kanazawa University, Kakuma-machi, Kanazawa 920-1192, Japan
Various advanced microscopy techniques have been developed to resolve molecular systems on surfaces with increasing resolution. However, 2D images representing 3D structures generally lead to missing information, making the interpretation of images challenging. Density functional theory (DFT) with explicit 3D structures offer a route to understanding through analyzing multiple molecular conformations, but the DFT is extremely computationally demanding. To address this issue, in this work we train a neural network potential to predict the interatomic interactions between adsorbates and surfaces as an efficient substitute for DFT calculations with decent accuracy. Firstly, we established a dataset of 2000 organic molecules adsorbed on the Cu(111) surface either in equilibrium or distorted, which includes total energy, atomic force, electrostatic potential, and charge obtained by DFT calculations. And then, based on the small training dataset, we further developed a model on a basis of the NequIP framework, to predict forces. With the auxiliary information from charge and electrostatic potential, our modified model could reach state-of-the-art performance with forces mean absolute errors of ~1.0 meV/atom, comparable to DFT accuracy.