Dresden 2017 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
HL: Fachverband Halbleiterphysik
HL 80: Electronic-Structure Theory: New Concepts and Developments in Density Functional Theory and Beyond - VII
HL 80.6: Vortrag
Donnerstag, 23. März 2017, 17:30–17:45, GER 38
Spectral property prediction with artificial neural networks — •Annika Stuke1, Milica Todorovic1, Kunal Ghosh2, Aki Vehtari2, and Patrick Rinke1 — 1Department of Applied Physics, Aalto University, Finland — 2Helsinki Institute of Information Technology, Department of Computer Science, Aalto University, Finland
The ability to efficiently design new and advanced optoelectronic materials is hampered by the lack of suitable methods to rapidly and accurately identify yet-to-be-synthesized materials that meet a desired application. To overcome such design challenges, a machine learning model based on a deep multi-task artificial neural network (ANN) is presented that can predict spectral properties of small organic molecules. The ANN is trained and validated on data generated by accurate state-of-the art quantum chemistry computations for diverse subsets of the GDB-13 and GDB-17 datasets [1,2]. The molecules are represented by a simple, easily attainable numerical description based on nuclear charges and cartesian coordinates and are mapped onto multiple excited-state properties simultaneously using a deep ANN trained by gradient descent and error backpropagation [3]. This on-demand prediction model can be used to infer spectral properties of various candidate molecules in an early screening stage for new optoelectronic materials at negligible computational cost, thereby completely bypassing conventional laborious approaches towards materials discovery.
[1] L. C. Blum et al., J. Am. Chem. Soc. 2009, 131, 8732, [2] R. Ramakrishnan et al., Scientific Data 2014, 1, 140022, [3] G. Montavon et al., New J. Phys. 2013, 15, 095003