Yashar Kiarashinejad1 Sajjad Abdollahramezani1 Mohammadreza Zandehshahvar1 Omid Hemmatyar1 Ali Adibi1

1, Georgia Institute of Technology, Atlanta, Georgia, United States

In this paper, we introduce a new approach based on deep learning (DL) for understanding thefundamental phenomena of nanoscale electromagnetic (EM) waves and matter interactions. Thismethod builds on the dimensionality reduction (DR) technique in which autoencoders are leveragedto reduce the dimensionality of an light-matter interaction problem with the minimal imposed errorby using the strong correlation among its different features [1]. The DR approach highlights the mostimportant features (e.g., design parameters, material state) of any linear or nonlinear EM structure thataffects its functionality. This method provides priceless information about the role of nanostructureparameters responding to incident EM waves for a given functionality. Moreover, along with providingthe analytical formulas, the DR technique converts a large-size problem into a smaller space which cansignificantly facilitate and reduce the computation of brute-force optimization and design techniques. Toshow the applicability of the proposed approach, here we consider two sophisticated design problemsfor implementation of a reconfigurable multifunctional metadevice enabling dual-band and triple-bandabsorption in the telecommunication window [2,3]. The metadevice consists of an array of supercells,where each supercell has 4 design parameters (i.e., height, widths, pitches, and crystallization levels ofthe phase-change material), so that the total number of design parameters become 10. The simulationand experimental results are in a good agreement justifying the effectiveness of our proposed approachin predicting the most critical design parameters in the design problem.


[1] Kiarashinejad, Yashar, Sajjad Abdollahramezani, and Ali Adibi. ”Deep learning approachbased on dimensionality reduction for designing electromagnetic nanostructures.” arXiv preprintarXiv:1902.03865 (2019).
[2] Kiarashinejad, Yashar, Sajjad Abdollahramezani, Mohammadreza Zandehshahvar, Omid Hem-matyar, and Ali Adibi. ”Deep Learning Reveals Underlying Physics of Light-matter Interactions inNanophotonic Devices.” arXiv preprint arXiv:1905.06889 (2019).
[3] Abdollahramezani, Sajjad, Hossein Taghinejad, Tianren Fan, Yashar Kiarashinejad, Ali A. Eftekhar,and Ali Adibi. ”Reconfigurable multifunctional metasurfaces employing hybrid phase-change plas-monic architecture.” arXiv preprint arXiv:1809.08907 (2018).