‘MetaNETs’ - Accelerated discovery and design of photonic metamaterials using deep learning Prajith Pillai TCS Innovation Labs, Tata Consultancy Services Limited, Bangalore, India, 560066 p.prajith@tcs.com Parama Pal TCS Innovation Labs, Tata Consultancy Services Limited, Bangalore, India, 560066 parama.pal@tcs.com Rinu Chacko TCS Innovation Labs, Tata Consultancy Services Limited, TRDDC, Pune, India, 411013 rinu.chacko@tcs.com Deepak Jain TCS Innovation Labs, Tata Consultancy Services Limited, TRDDC, Pune, India, 411013 deepak.jain3@tcs.com Beena Rai TCS Innovation Labs, Tata Consultancy Service Limited, TRDDC, Pune, India, 411013 beena.rai@tcs.com Abstract We propose an approach based on neural networks for approximating the electro- magnetic (EM) responses of mesoscale, split-ring resonator photonic metamaterials. We demonstrate that by treating the EM spectral data as time-varying sequences and the inverse problem as a single-input, multi-output model, we force our architecture to learn the geometry of the designs from the training data as opposed to abstract features thereby addressing both the forward and the inverse design problems with great promise. 1 Introduction Optical metamaterials (MMs) are artificially structured composites that demonstrate customizable electromagnetic (EM) properties that stem from morphological feature dimensions of the order of the interrogating wavelength. MM devices are composed of periodic 2D or 3D arrays of sub- wavelength conducting elements and exhibit light-matter interactions that enable the manipulation of the effective electrical permittivity and effective magnetic permeability, therefore opening up new paths for versatile, photonic devices [3]. Traditionally, the MM design process relies heavily on the knowledge and intuitive reasoning of the researcher as achieving a target spectral behavior involves iteratively tweaking the design till the outcome is satisfactory. Starting with an initial geometry, standard numerical techniques such as the finite-difference time domain method, boundary element method are employed for solving the Maxwell’s equations for obtaining the spectral response [6]. Of late, commercially available EM solver software packages have been widely adopted by the MM community for designing metamaterials but there too, the optimization of target design depends largely on intuition. Third Workshop on Machine Learning and the Physical Sciences (NeurIPS 2020), Vancouver, Canada.