SKM 2023 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 3: Development of Computational Methods: Evaporation, Growth and Oxidation – Density Functional, Tight Binding

MM 3.3: Vortrag

Montag, 27. März 2023, 10:45–11:00, SCH A 251

How to teach my deep generative model to create new RuO₂ surface structures? — •Patricia König, Hanna Türk, Yonghyuk Lee, Chiara Panosetti, Christoph Scheurer, and Karsten Reuter — Fritz-Haber Institute, Max-Planck-Society, Faradayweg 4-6, 14195 Berlin, Germany

Data-driven approaches to inversely design novel materials with desired properties constitute an emerging pillar in the exploration of energy conversion materials. In recent works, artificial neural networks were successfully used to create crystalline porous materials. Here, we present a related approach to tackle the problem of structure generation for nano-porous to partially amorphous surfaces. As a model system, we use the well-studied RuO₂ catalyst for oxidative conversion of CO to CO₂. To explore the chemical space of RuO₂ surface structures, we trained a Generative Adversarial Network (GAN) that is capable of cheaply generating diverse structural guesses for novel surface structures. For the training set, 28,903 RuO₂ surface terminations were created with a grand-canonical basin hopping method using ML potential energetics. The atomic positions of these structures were mapped to Gaussian densities on a three-dimensional grid to generate the GAN input. We demonstrate how realistic three-dimensional surface models with inferred lattice lengths and energy conditioning can be created and how these generated densities can be mapped back to atomistic structures as a basis for property calculations.