A miniaturized computational spectrometer with optimum number of nanophotonic filters: Deep-learning autoencoding and inverse design-based implementation

Abstract

In the last decade, computational spectrometers emerged as a new paradigm in the miniaturization of optical spectrometers. These systems are composed of distinct filters or detectors, and the input spectrum is reconstructed from the spectral response of these encoding elements. For improving the performance of these systems, much work has been done in designing efficient and advanced algorithms for spectrum recovery. Still, the remaining challenge is finding the optimal basis of filters for the encoding operation. In this article, we propose a systematic approach based on deep learning for the intelligent design of spectral encoding elements. This approach consists of two steps: first, we utilize an autoencoder to find the optimal spectral responses with minimal correlation and then use a conditional generative adversarial network to inverse design the corresponding filter structures of those extracted spectral responses. With this approach, we design 15 optimal filters that can be used as meta-pixel in hyperspectral image sensors and computational spectrometers. The realized computational spectrometer with compact size can also be embedded in smartphones as an external gadget and utilize the available processing power of these devices.