Regensburg 2022 – wissenschaftliches Programm

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

DY: Fachverband Dynamik und Statistische Physik

DY 45: Poster Session: Nonlinear Dynamics, Pattern Formation, Data Analytics and Machine Learning

DY 45.9: Poster

Donnerstag, 8. September 2022, 15:00–18:00, P2

Characterizing quasicrysalline patterns with neural networks — •Jonas Buba and Michael Schmiedeberg — Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany

There exist multiple ways to generate two dimensional quasicrystal patterns such as superpositions of plane waves [1-3] or projection methods [4], each using different parametrisations. A deep convolutional Autoencoder is utilized to determine whether an artifical neural network is capable of finding sufficient parametrisations for these patterns.

Furthermore another neural network was used to classify them by symmetry. Finally, deep dream methods [5] are employed to analyze which features the artificial intelligence bases its symmetry classification on. Currently we are trying to find out how the deep dream approach can be applied to modify existing patterns by enhancing features corresponding to specific symmetries.

[1] D.S. Rokhsar, N.D. Mermin, and D.C. Wright, Acta cryst. A44, 197 (1988). [2] S.P. Gorkhali, J. Qi, and G.P. Crawford, J. Opt. Soc. Am. B 23, 149 (2005). [3] M. Schmiedeberg and H. Stark, J. of Phys.: Cond. Matter 24(28), 284101 (2012). [4] M. Duneau and A. Katz, Phys. Rev. Lett. 54, 2688 (1985). [5] A. Mordvintsev, C. Olah, and M. Tyka, "Inceptionism: Going Deeper into Neural Networks", Google AI Blog (2015).