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

O: Fachverband Oberflächenphysik

O 67: Frontiers of Electronic Structure Theory: Focus on Artificial Intelligence Applied to Real Materials 3

O 67.2: Vortrag

Donnerstag, 8. September 2022, 10:45–11:00, S054

Symmetry and completeness in machine-learning models for atomistic simulations — •Sergey Pozdnyakov and Michele Ceriotti — EPFL, Switzerland

During the last decade, machine learning methods have drastically changed atomistic simulations. On the one hand, they scale linearly with the size of the system and thus, are significantly faster than the quantum mechanical calculations. On the other, they provide a functional form that is much more flexible than so-called classical force fields such as the Lennard Jones potential or embedded atom models. From one point of view, incorporating rotational symmetry is important for ML since it can make models more data-efficient and robust, but can also lead to incompleteness, limiting the ultimate accuracy of the model. I will discuss some examples of this and compare different types of models to show how one can find an optimal balance of the two effects.