SurfaceScience21 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 9: Poster Session I: New methods I
O 9.2: Poster
Monday, March 1, 2021, 10:30–12:30, P
How to train a Gaussian Approximation Potential for metal oxides: A three stage process towards fast surface sampling — •Carsten Staacke1,2, Jakob Timmermann1,2, Yongyhuk Lee1,2, Christoph Scheurer1,2, and Karsten Reuter1,2 — 1Fritz-Haber-Institut der MPG — 2Technische Universität München
Surface orientation and termination play a decisive role for the chemical and physical behavior of catalysts and functional materials. Ab initio thermodynamics comparing the stability of different density-functional theory computed trial structures has emerged as a standard tool to determine surface structures under working conditions. However, limited by the high computational cost, these trial structures are up to now typically derived manually, reflecting the researcher’s ability to imagine possible structural models. Efficient and accurate machine-learned (ML) interatomic potentials promise to replace this state-of-the-art by systematic global optimization methods. To this end, we have developed a protocol to train short-ranged ML potentials for metal oxides based on the Gaussian Approximation Potential and Smooth Overlap of Atomic Positions (SOAP) descriptors to capture atomic environments.[1,2] We will present a systematic three-stage training process. For various rutile type metal oxides the role of all three stages will be exemplified and critically evaluated. [1] J. Timmermann et al., Phys. Rev. Lett. 125, 206101 (2020). [2] A.P. Bartók et al., Phys. Rev. Lett. 104, 136403 (2010).