Deep learning-based design of broadband GHz complex and random metasurfaces

Abstract

We are interested in exploring the limit in using deep learning (DL) to study the electromagnetic (EM) response for complex and random metasurfaces, without any specific applications in mind. For simplicity, we focus on a simple pure reflection problem of a broadband EM plane wave incident normally on such complex metasurfaces in the frequency regime of 2–12 GHz. In doing so, we create a DL-based framework called the metasurface design deep convolutional neural network (MSDCNN) for both forward and inverse designs of three different classes of complex metasurfaces: (a) arbitrary connecting polygons, (b) basic pattern combination, and (c) fully random binary patterns. The performance of each metasurface is evaluated and cross-benchmarked. Dependent on the type of complex metasurfaces, sample size, and DL algorithms used, the MSDCNN is able to provide good agreement and can be a faster design tool for complex metasurfaces than the traditional full-wave EM simulation methods. However, no single universal deep convolutional neural network model can work well for all metasurface classes based on detailed statistical analysis (such as mean, variance, kurtosis, and mean-squared error). Our findings report important information on the advantages and limitations of current DL models in designing these ultimately complex metasurfaces.

SUTD-PRCM Dataset and Neural Architecture Search Approach for Complex Metasurface Design

“…If this goal is met, we will consider that the DL model is successful, which may be able to capture the underlying physics. Unfortunately, we have concluded in a recent study [44] that current CNN based DL models are insufficient to reach this optimal condition for which cross classes forward prediction shows deteriorating performance. We postulate that this is due to non-optimal neural architecture adopted, which motivates us to share our dataset in this paper with other researchers for future improvements.…”

“…Fourier transform [23,47], discrete cosine transform, wavelet transform and uniform downsampling are some popular options in constructing compressed representations. In our prior work [44], we reported that uniform down-sampling has produced good results to predict the magnitude. However, the phase spectra are oscillatory due to the periodicity of 2π.…”

“…The randomly generated samples are sufficiently complex that are suitable for forward prediction and inverse design in testing different DL models in a simple GPU. In our recent work [44], we have demonstrated that in using this SUTD-PRCM dataset tested with some existing DCNN based neural architectures that might not be the most optimal neural architecture yet. Thus this dataset is published here to share with the community for further testing.…”

“…Easy access to such standard dataset will allow a more quantitative and fair comparison between different neural architectures and training strategies in the research community. Inspired by how the standard datasets in CV community has advanced the state-of-the-art in their field, we are ready to share our dataset (SUTD-PRCM), which was first generated in a previous work and tested on different DL models [44]. This dataset is essentially a collection of numerical simulated results of EM wave reflection of randomly created metasurfaces.…”

“…For demonstration purpose, an application to a classification problem based on this dataset is considered. Note that the approach is general and it is likewise applicable to a regression problem [44]. Thus the objective of this paper is two folds.…”