Neural Polarimeter and Wavemeter

Abstract

Numerous optical devices can be conveniently described in terms of a transfer function matrix formalism. An important example is the intensity-division Stokes polarimeter where four device outputs can be related to the four parameters of the Stokes vector using a linear 4 × 4 matrix transformation. In the present paper, we demonstrate how the functionality of such devices can be substantially enhanced by increasing the number of outputs and employing deep neural networks instead of the traditional linear algebra approach to establish correlations between device outputs and inputs. Specifically, we employ a neural network calibration of a metasurface-based intensity-division Stokes polarimeter with six outputs to accurately measure the four parameters of the Stokes vector of the input light across a much wider wavelength range than is afforded by a canonical linear transfer matrix model. Furthermore, the neural network model allows the device to determine the input wavelength from the measured data. We arg...