Dresden 2020 – wissenschaftliches Programm

Die DPG-Frühjahrstagung in Dresden musste abgesagt werden! Lesen Sie mehr ...

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

MM: Fachverband Metall- und Materialphysik

MM 35: Topical Session: Data Driven Materials Science - Descriptors (joint session MM/CPP)

MM 35.5: Vortrag

Mittwoch, 18. März 2020, 12:45–13:00, BAR 205

Similarity descriptors for data-driven materials science — •Martin Kuban, Santiago Rigamonti, and Claudia Draxl — Humboldt-Universität zu Berlin

Learning from materials data is a topic of increasing importance in materials science. This task is supported by the availability of data through large online databases, like NOMAD [1]. For the application of artificial-intelligence (AI) methodology, materials must be characterized by a set of features that together build up descriptors. The success of AI tasks depends heavily on the quality of these descriptors, since they must contain all relevant information to map the input data onto the target property. Recent advances in the development of high-quality descriptors have allowed for both accurate predictions of material properties as well as highly interpretable models [2]. In this work, we develop a new type of descriptors based on the similarity of materials. To achieve this goal, we use both existing and newly developed descriptors to establish metrics that serve as quantitative similarity measures. These measures are combined into "similarity descriptors", which are then used for the construction of AI models. The performance of these models is optimized with respect to their predictive power. We demonstrate the applicability of our approach by predicting target properties for different classes of materials, including oxides and 2D systems.

[1] C. Draxl and M. Scheffler, MRS Bulletin, 43, 676, (2018).

[2] L. Ghiringhelli et al., PRL, 114, 105503, (2015).