A Generic Framework for Neural Networks Based Modeling and Design of Free-Form Manufacturable Metasurfaces

Metasurface design methods mainly rely on full electromagnetic simulations. Although they provide high accuracy, the time and computational power requirements of these simulations become a limiting factor with increased structural complexity. In recent years, deep learning (DL) based simulation and design approaches are emerged as an alternative solution to the computational cost problem. However, there are several challenges to be overcome to achieve feasible and widely applicable DL based models. Addressing high degrees of design freedom while ensuring manufacturable designs one of the main challenges. Fabrication limitations are either addressed in low dimensional design space or not taken into consideration by most DL based approaches. Additionally, most of the existing DL based metasurface models are valid for limited operation conditions that are defined by the parameter space of the training data. This brings the insufficient generalizability of models as another fundamental problem.<br/><br/>Here, addressing these problems, we introduce a deep learning enabled framework to design free-form metasurface unit cells, where the target design space is subject to the constraints of top-down nanofabrication techniques. First, a shape generation method is developed to create a diverse library of fabrication-friendly meta-atoms. Then, applying our recently proposed wavelength normalization approach [1,2] to this dataset, a combined architecture of variational autoencoder and fully connected layers is trained as a forward (from structure to response) model. In addition to an unseen validation set extracted from our library, the forward model is also tested on cross-polarization, material dispersion, and spectral ranges outside our dataset, which the model wasn’t explicitly trained for. All in all, our forward model is proven to be a successful surrogate to electromagnetic simulations that provide high accuracy predictions for a wide span of design and problem settings, for ultra-low computational cost. In the final stage, we realized inverse design of free-form metasurfaces while ensuring manufacturability conditions.<br/><br/>[1] Tanriover, I.; Hadibrata, W.; Aydin, K. Physics-Based Approach for a Neural Networks Enabled Design of All-Dielectric Metasurfaces. <i>ACS Photonics</i> <b>2020</b>, <i>7</i> (8), 1957–1964.<br/>[2] Tanriover, I.; Hadibrata, W.; Scheuer, J.; Aydin, K. Neural Networks Enabled Forward and Inverse Design of Reconfigurable Metasurfaces. <i>Opt. Express</i> <b>2021</b>, <i>29</i> (17), 27219–27227.