Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 33: Topical session (Symposium MM): Big Data Analytics in Materials Science
MM 33.3: Vortrag
Donnerstag, 4. April 2019, 11:00–11:15, H43
The NOMAD 2018 Kaggle Competition: Tackling Materials-Science Challenges through Crowd Sourcing — •Christopher Sutton1, Luca M. Ghiringhelli1, Takenori Yamamoto2, Xiangyue Liu1, Angelo Ziletti1, and Matthias Scheffler1 — 1Fritz-Haber-Institut der Max-Planck-Gesellschaft, Berlin, Germany — 2Institute for Mathematical and Computational Sciences, LLC Yokohama, Japan
Machine learning (ML) promises to accelerate the discovery of novel materials by screening candidate compounds at significantly lower computational cost than traditional electronic-structure approaches. However, it is often a priori unclear which ML models are suitable for a given problem and optimizing a model can be a time-consuming endeavor. Crowd sourcing allows for comparing several ML models by identifying a key problem and challenging the community to solve it. To this end, the Novel Materials Discovery (NOMAD) Centre of Excellence together with Kaggle - one of the most well known hosting platforms - organized an open data-science competition to predict two key properties of transparent conducting oxides (TCOs): band gap energy (for transparency) and formation energy (for stability). Although these materials are crucial for optoelectronic devices, only a small number TCOs are currently known. In this contribution, we present the winning model out of nearly 900 participants based on a novel crystal-graph representation and an analysis of the relative importance of representation vs regression model for the performance of several ML approaches.